Spectrochemical Analysis Using Infrared Multichannel Detectors

Spectrochemical Analysis using Infrared Multichannel Detectors Edited by Rohit Bhargava Department of Bioengineering and Beckman Institute for Advanced Science and Technology, University of Illinois at Urbana-Champaign Urbana, IL 61801 USA And Ira W. Levin Laboratory of Chemical Physics, NIDDK National Institutes of Health Bethesda, MD 20892 USA

Blackwell Publishing

Spectrochemical Analysis using Infrared Multichannel Detectors

Analytical Chemistry Series Series Editors: John M. Chalmers and Alan J. Handley A series which presents the current state of the art in chosen sectors of analytical chemistry. Written at professional and reference level, it is directed at analytical chemists, environmental scientists, food scientists, pharmaceutical scientists, earth scientists, petrochemists and polymer chemists. Each volume in the series provides an accessible source of information on the essential principles, instrumentation, methodology and applications of a particular analytical technique. Titles in the series: Inductively Coupled Plasma Spectrometry and its Applications Edited by S.J. Hill Extraction Methods in Organic Analysis Edited by A.J. Handley Design and Analysis in Chemical Research Edited by R.L. Tranter Spectroscopy in Process Analysis Edited by J.M. Chalmers Gas Chromatographic Techniques and Applications Edited by A.J. Handley and E.R. Adlard Chemical Analysis of Contaminated Land Edited by K.C. Thompson and C.P. Nathanail Atomic Spectroscopy in Elemental Analysis Edited by M. Cullen Pharmaceutical Analysis Edited by D.C. Lee and M. Webb Environmental Toxicity Testing Edited by K.C. Thompson, K. Wadhia and A.P. Leibner Spectrochemical Analysis using Infrared Multichannel Detectors Edited by R. Bhargava and I.W. Levin

Spectrochemical Analysis using Infrared Multichannel Detectors Edited by Rohit Bhargava Department of Bioengineering and Beckman Institute for Advanced Science and Technology, University of Illinois at Urbana-Champaign Urbana, IL 61801 USA And Ira W. Levin Laboratory of Chemical Physics, NIDDK National Institutes of Health Bethesda, MD 20892 USA

Blackwell Publishing

© 2005 by Blackwell Publishing Ltd Editorial Offices: Blackwell Publishing Ltd, 9600 Garsington Road, Oxford OX4 2DQ, UK Tel: +44 (0)1865 776868 Blackwell Publishing Professional, 2121 State Avenue, Ames, Iowa 50014-8300, USA Tel: +1 515 292 0140 Blackwell Publishing Asia Pty Ltd, 550 Swanston Street, Carlton, Victoria 3053, Australia Tel: +61 (0)3 8359 1011 The right of the Author to be identified as the Author of this Work has been asserted in accordance with the Copyright, Designs and Patents Act 1988. All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, except as permitted by the UK Copyright, Designs and Patents Act 1988, without the prior permission of the publisher. First published by Blackwell Publishing Ltd Library of Congress Cataloging-in-Publication Data Spectrochemical analysis using infrared multichannel detectors/edited by Rohit Bhargava & Ira Levin. p. cm. Includes bibliographical references and index. ISBN-13: 978-1-4051-2504-8 (acid-free paper) ISBN-10: 1-4051-2504-7 (acid-free paper) 1. Infrared spectroscopy. 2. infrared imaging. 3. Spectrum analysis. I. Bhargava, Rohit, 1973– II Levin, Ira, 1935– QD96.I5.S64 2005 535.8’42–dc22 2005009290 ISBN-13: 978-1-4051-2504-8 ISBN-10: 1-4051-2504-7 British Library Cataloguing-in-Publication Data: A Catalogue record for this title is available from the British Library Set in 10/12 pt Times by Newgen Imaging Systems (P) Ltd, Chennai, India Printed and bound in India by Replika Press, Pvt Ltd., Kundli The publisher’s policy is to use permanent paper from mills that operate a sustainable forestry policy, and which has been manufactured from pulp processed using acid-free and elementary chlorine-free practices. Furthermore, the publisher ensures that the text paper and cover board used have met acceptable environmental accreditation standards. For further information on Blackwell Publishing, visit our website: www.blackwellpublishing.com

Contents

Contributors

xi

Preface

xv

1

Fourier transform mid-infrared spectroscopic imaging

1

Rohit Bhargava and Ira W. Levin

2

1.1 1.2

Introduction Fundamentals of FTIR spectroscopy 1.2.1 Interferometer characteristics 1.3 FTIR microspectroscopy using a single-element detector 1.3.1 IR microscopes and point spectroscopy 1.3.2 FTIR mapping 1.3.3 Limitations of FTIR point mapping 1.4 FTIR imaging with multichannel detectors 1.4.1 Imaging with large format array detectors 1.4.2 Interfacing an interferometer to large array detectors 1.4.3 The SNR of imaging spectrometers 1.4.4 The evolving detector array technology 1.5 Raster scanning with linear array detectors 1.5.1 Choice of either small or large detector arrays 1.6 Conclusions References

1 1 2 8 8 11 11 13 13 15 16 19 20 21 22 23

Near-infrared spectral imaging with focal plane array detectors

25

E. Neil Lewis, Linda H. Kidder, Eunah Lee and Kenneth S. Haber 2.1 2.2

Background: single-point near-infrared spectroscopy Development of NIR spectral imaging 2.2.1 History of spectral imaging 2.2.2 FPAs – specifications 2.2.3 Implementation of NIR imaging 2.2.4 Data processing 2.2.5 Comparison of vibrational spectroscopic imaging modalities 2.2.6 Safety in numbers

25 27 27 28 29 31 33 35

vi

CONTENTS

2.3

Examples of NIR spectral imaging capabilities 2.3.1 Sample statistics and FOV 2.3.2 High-throughput applications 2.3.3 Statistics, morphology, abundance – using an internal reference 2.4 Conclusions References 3

Multichannel detection with a synchrotron light source: design and potential

37 37 42 43 51 52

56

G. Larry Carr, Oleg Chubar and Paul Dumas

4

3.1 3.2

Introduction Comparisons of thermal and SR sources 3.2.1 Blackbody radiation 3.2.2 SR as an IR source 3.3 The IR microspectrometer: instrumentation and optical analysis 3.3.1 Microspectrometer system components 3.3.2 Performance: imaging at the diffraction limit 3.3.3 The FPA microscope system 3.4 Combining SR with an FPA microspectrometer 3.4.1 FPA microspectrometer for PSF image deconvolution 3.4.2 SR as an extended IR source 3.5 Summary Acknowledgements References

56 58 59 59 68 68 72 77 80 80 81 82 83 83

Multivariate analysis of infrared spectroscopic image data

85

Scott W. Huffman and Chris W. Brown

5

4.1 4.2

Introduction Preprocessing hyperspectral images 4.2.1 Data compression 4.2.2 Smoothing spectra 4.2.3 Noise in hyperspectral images 4.3 Processing hyperspectral images 4.3.1 Feature extraction 4.3.2 Concentration image maps 4.4 Conclusions References

85 85 86 90 92 101 101 109 113 113

FTIR imaging of multicomponent polymers

115

Jack L. Koenig 5.1

Introduction

115

CONTENTS

5.2 5.3

Imaging requirements for polymer characterization Polymer sampling for FTIR imaging 5.3.1 Transmission measurements 5.3.2 Reflection FTIR imaging measurements 5.3.3 ATR FTIR imaging 5.4 FTIR image analysis 5.4.1 Selection of characteristic spectral stains for each component 5.4.2 Construction of contour plots 5.4.3 Histograms 5.5 Applications of FTIR imaging to complex polymer systems 5.5.1 FTIR imaging of polymer laminate films 5.5.2 Chemical morphology of multi-component polymeric materials 5.5.3 Immiscible polymer blends 5.5.4 Crosslinking-induced phase separation of elastomers 5.5.5 Semicrystalline polymer systems 5.5.6 Semicrystalline polymer blends 5.6 Summary and conclusions References 6

7

vii 115 116 116 118 119 121 122 122 123 126 126 126 132 135 137 139 140 140

Combinatorial approaches to catalyst development with multichannel detectors

143

Christopher M. Snively and Jochen Lauterbach 6.1 Introduction – combinatorial materials development 6.2 Array detection schemes for high-throughput analysis 6.3 FTIR imaging as a high-throughput technique 6.4 Applications 6.4.1 Application I: resin-supported ligands 6.4.2 Application II: adsorbates on catalyst surfaces 6.4.3 Application III: reactor effluent quantification 6.5 Data management 6.6 Summary References

143 145 146 148 148 149 150 151 155 156

Materials analysis systems based on real-time near-IR spectroscopic imaging Martin Kraft, Raimund Leitner and Herwig Mairer 7.1 Introduction 7.2 Data acquisition 7.2.1 Image acquisition 7.2.2 Sample–radiation interaction 7.3 Instrumentation

158 158 158 158 161 162

viii

8

9

CONTENTS

7.4

Real-time data analysis 7.4.1 Pre-processing 7.4.2 Spectral data evaluation 7.5 Integrated image processing 7.6 Material analysis applications 7.6.1 Industrial waste classification and sorting 7.6.2 Surface coating inspection 7.6.3 Food control 7.6.4 Mineralogical material analysis References

164 165 166 168 169 169 172 172 172 173

Industrial applications of near-IR imaging Anthony E. Dowrey, Gloria M. Story and Curtis Marcott

175

8.1 8.2 8.3

Introduction Experimental Application using NIR spectroscopic imaging 8.3.1 Water migration on fabrics 8.3.2 Spray nozzle patterns 8.3.3 Surfactant deposition on a nonwoven substrate 8.3.4 Flavored chips 8.3.5 Lotion distribution on nonwoven paper 8.4 Conclusions Acknowledgements References

175 177 178 178 179 179 181 182 184 187 188

IR spectroscopic imaging

189

Max Diem, Melissa J. Romeo, Susie Boydston-White and Christian Matthäus 9.1 9.2

Introduction: definition and goals of spectral mapping Experimental 9.2.1 Instrumental aspects: PE Spotlight 300 9.2.2 Samples 9.2.3 Spectral maps of individual cells 9.2.4 Spectral maps of ‘smears’ 9.2.5 Spectral maps of tissues 9.2.6 Mathematical analysis 9.3 Results and discussion 9.3.1 Spectral histopathology of lymph nodes 9.3.2 Spectral maps of individual cells 9.3.3 Spectral maps of ‘cell smears’ 9.4 Conclusions Acknowledgements References

189 190 190 191 191 192 192 193 194 194 197 200 202 202 202

CONTENTS

10 FPA imaging and spectroscopy for monitoring chemical changes in tissue

ix

204

Bayden R. Wood and Don McNaughton 10.1 Introduction 10.2 Applications of FTIR tissue imaging to cervical cancer 10.2.1 History of FTIR spectroscopy applied to cervical cancer diagnosis 10.2.2 FTIR point-to-point mapping of cervical tissue 10.2.3 FTIR focal plane array imaging of cervical tissue 10.3 FPA imaging and spectroscopy for monitoring chemical changes associated with collagen-induced arthritis 10.4 Application of FTIR 3D imaging to histology 10.5 Conclusions Acknowledgements References 11 Infrared microscopy and imaging of hard and soft tissues

204 205 205 206 207 224 229 230 231 231 234

Richard Mendelsohn, Adele L. Boskey and Nancy P. Camacho 11.1 Introduction 11.2 IR imaging protocols 11.3 Applications of FTIR microscopy and imaging to tissues 11.3.1 Bone 11.3.2 Skin 11.3.3 Cartilage Acknowledgements References

234 235 235 235 243 249 257 257

12 Mid-infrared imaging applications in agricultural and food sciences 261 Douglas L. Elmore, Carrie A. Lendon, Sean A. Smith and Chad L. Leverette 12.1 Introduction 12.2 Spatially resolved chemical and physical information 12.3 Chemical infrared imaging of protein, carbohydrates and fat in agri-food mixtures 12.4 Sampling 12.5 Chemometrics 12.6 Applications 12.7 Complementary imaging techniques 12.8 Conclusions References

261 264 266 268 270 272 277 278 279

x

CONTENTS

13 Applications of near-infrared imaging for monitoring agricultural food and feed products Vincent Baeten and Pierre Dardenne 13.1 Introduction 13.2 Use of NIR imaging for remote control and monitoring in agriculture 13.2.1 The problem 13.3 NIR imaging for food analysis 13.3.1 The problem 13.4 NIR imaging for feed analysis 13.4.1 The problem 13.5 Conclusion References Index

283 283 284 284 288 288 294 294 297 297 303

Contributors

Dr Vincent Baeten

Département Qualité des Produits Agricoles, Centre Wallon de Recherches Agronomiques, 24 Chaussée de Namur, 5030 – Gembloux, Belgium

Professor Rohit Bhargava

Department of Bioengineering and Beckman Institute for Advanced Science and Technology, University of Illinois at Urbana-Champaign, 405 North Mathews Avenue, Urbana, IL 61801, USA

Dr Adele L. Boskey

Mineralized Tissues Laboratory, The Hospital for Special Surgery, 535 E 70th Street, New York, NY 10021, USA

Dr Susie Boydston-White

Department of Chemistry and Biochemistry, Hunter College and Graduate School, City University of New York, New York, NY 10021, USA

Professor Chris W. Brown

Department of Chemistry, University of Rhode Island, Pastore Hall, Kingston RI 02881, USA

Dr Nancy P. Camacho

The Musculoskeletal Imaging and Spectroscopy Laboratory, The Hospital for Special Surgery, 535 E 70th Street, New York, NY 10021, USA

Dr G. Larry Carr

National Synchrotron Light Source (NSLS), Brookhaven National Laboratory, 75 Brookhaven Avenue, Bldg 725B, Upton, NY 11973-5000, USA

xii

CONTRIBUTORS

Dr Oleg Chubar

LURE and Synchrotron SOLEIL, L’Orme des Merisiers, Saint-Aubin - BP 48, 91192 Gif-Sur-Yvette Cedex, France

Dr Pierre Dardenne

Département Qualité des Produits Agricoles, Centre Wallon de Recherches Agronomiques, 24 Chaussée de Namur, 5030 – Gembloux, Belgium

Professor Max Diem

Department of Chemistry and Biochemistry, Hunter College and Graduate School, City University of New York, New York, NY 10021, USA

Mr Anthony E. Dowrey

The Procter & Gamble Company, Miami Valley Laboratories, Cincinnati, OH 45253-8707, USA

Dr Paul Dumas

LURE and Synchrotron SOLEIL, L’Orme des Merisiers, Saint-Aubin - BP 48, 91192 Gif-Sur-Yvette Cedex, France

Dr Douglas L. Elmore

Scientific Resources Center, Cargill Incorporated, 7101 Goodlett Farms Parkway, Cordova, TN 38016, USA

Dr Kenneth S. Haber

Spectral Dimensions Inc, 3416 Olandwood Court, Suite 210, Olney, MD 20832, USA

Dr Scott W. Huffman

Laboratory of Chemical Physics, NIDDK, Building 5, B1-38, National Institutes of Health, Bethesda, MD 20892-0520, USA

Dr Linda H. Kidder

Spectral Dimensions Inc, 3416 Olandwood Court, Suite 210, Olney, MD 20832, USA

Professor Jack L. Koenig Department of Macromolecular Science, Case Western Reserve University, Kent Hale Smith Building, Room 212, 2100 Adelbert Road, Cleveland, Ohio 44106-7202, USA

CONTRIBUTORS

xiii

Dr Martin Kraft

Carinthian Tech Research AG, Europastrasse 4/1, A - 9524 Villach / St. Magdalen, Austria

Professor Jochen Lauterbach

Department of Chemical Engineering, University of Delaware, Newark, DE 19716, USA

Dr Eunah Lee

Spectral Dimensions Inc., 3416 Olandwood Court, Suite 210, Olney, MD 20832, USA

Dr Raimund Leitner

Carinthian Tech Research AG, Europastrasse 4/1, A - 9524 Villach / St. Magdalen, Austria

Dr Carrie A. Lendon

Scientific Resources Center, Cargill Incorporated, 7101 Goodlett Farms Parkway, Cordova, TN 38016, USA

Dr Chad L. Leverette

Department of Chemistry and Physics, University of South Carolina Aiken, Aiken, SC 29801-6399, USA

Dr Ira W. Levin

Building 5, B1-32, Laboratory of Chemical Physics, NIDDK, National Institutes of Health, Bethesda, MD 20892-0520, USA

Dr E. Neil Lewis

Spectral Dimensions Inc., 3416 Olandwood Court, Suite 210, Olney, MD 20832, USA

Dr Herwig Mairer

Carinthian Tech Research AG, Europastrasse 4/1, A - 9524 Villach / St Magdalen, Austria

Dr Curtis Marcott

The Procter & Gamble Company, Miami Valley Laboratories, Cincinnati, OH 45253-8707, USA

Dr Christian Matthäus

Department of Chemistry and Biochemistry, Hunter College and Graduate School, City University of New York, New York, NY 10021, USA

Professor Don McNaughton

School of Chemistry, PO Box 23, Monash University, Melbourne, Victoria 3800, Australia

xiv

CONTRIBUTORS

Professor Richard Mendelsohn

Department of Chemistry, Rutgers University, 73 Warren St, Newark, NJ 07102-1811, USA

Dr Melissa J. Romeo

Department of Chemistry and Biochemistry, Hunter College and Graduate School, City University of New York, New York, NY 10021, USA

Dr Sean A. Smith

Scientific Resources Center, Cargill Incorporated, 7101 Goodlett Farms Parkway, Cordova, TN 38016, USA

Dr Christopher M. Snively

Department of Materials Science and Engineering, 201 Dupont Hall, University of Delaware, Newark, DE 19716-3106, USA

Ms Gloria M. Story

The Procter & Gamble Company, Miami Valley Laboratories, Cincinnati, OH 45253-8707, USA

Dr Bayden R. Wood

School of Chemistry, PO Box 23, Monash University, Melbourne, Victoria 3800, Australia

Preface A tremendous growth in the utilization of multichannel detectors for infrared (IR) spectroscopy has been observed over the past decade. In some cases, the incorporation of multichannel detectors has significantly changed the practice of IR spectroscopy; while in others, it has provided new opportunities for spectroscopic analyses. Since the underlying premise of spectroscopic analyses is to uncover chemical information, we have devoted this volume to examining the new insights and potential applications enabled by technology based on multichannel detectors. Instead of focusing on the science and technology of detectors or of IR spectroscopy, the chapters illustrate how these detectors are incorporated into spectroscopic instrumentation, the varied numerical techniques required for data analysis and the applications of multichannel-detector-enabled spectroscopy in a variety of fields. This compilation attempts to present the material in a manner that is accessible to readers without extensive background in spectroscopy or of the applications, yet the depth is sufficient to serve as a ready reference to seasoned practitioners in the field. We have not covered in-depth some recently emerging instrumentation and techniques that are likely to become more prevalent. The focus is, instead, on technologies that are mature enough to provide practitioners the tools to undertake spectrochemical analyses. Mindful that a compilation of this nature is always at danger of becoming outdated soon, the contents of this volume seek to provide the reader with an appreciation of the current state of the art as well as a perspective that allows an appreciation of future developments. Multichannel detectors for spectroscopy have moved rapidly from the laboratory to practical applications. Hence, we have also attempted to balance contributions from the ‘laboratory’ and from the ‘field’, providing a taste of both fundamental developments and ‘real-world’ applications to the reader. The first section (Chapters 1–4) provides an overview of instrumentation and data analysis techniques that form the foundation for practical applications. Chapter 1 outlines the development and capabilities of microspectroscopy in the mid-IR spectral region facilitated by multichannel detectors. Chapter 2 provides an overview of technology for the near-IR region and the unique contribution of imaging to near-IR spectral analyses. Utilization of a unique light source, the synchrotron, is discussed in Chapter 3, which details the theoretical aspects of enhanced microscopy and design considerations for future developments. A crucial bridge between instrumentation and applications is the mathematical analysis of large mathematical datasets. Chapter 4 presents several multivariate analysis techniques that are useful for deriving information from datasets.

xvi

PREFACE

The second section of the book details application of instrumentation and numerical analysis to spectroscopic analyses in a number of fields. The applications cover fields such as materials science (Chapters 5–8), biomedical science (Chapters 9– 11) and agricultural and food sciences (Chapters 12 and 13). Chapter 5 details the application of mid-IR FTIR spectroscopic imaging to multicomponent polymeric systems, salient features of data analysis for these systems, and a number of examples. Chapter 6 describes the utility of multichannel detectors to catalyst development and provides examples to demonstrate the translation of laboratory concepts to viable industrial catalysis. Chapter 7 provides an overview, and examples, of the application of near-IR imaging systems to the ‘real world’ in ‘real time’. Issues in the industrial design and analysis of several commercial products are detailed in Chapter 8. The application of mid-IR imaging systems based on multichannel detectors to cells, cell ensembles, and soft tissues are discussed in Chapter 9. Chapter 10 describes the application of focal plane array systems, which contain a very large multichannel advantage, to the rapid analysis of tissue for histopathologic changes and disease diagnosis in soft tissue. Chapter 11 examines the biochemical changes in both soft and hard tissue using IR microscopy based on different multichannel detectors. Chapter 12 provides an overview of the applications of mid-IR systems in food sciences, including the prospects of detailed biochemical descriptions of food through imaging. Chapter 13 provides an overview of the requirements and possibilities of monitoring agricultural materials through near-IR imaging. We would like to acknowledge the contributions of several colleagues over the years to our research efforts in the areas of applying multichannel detectors to spectroscopic analyses. We would also like to acknowledge the superb job by Blackwell (David McDade and Graeme Mackintosh) and Newgen (Mohan Kumar). Last, but not the least, we would like to thank all authors for their enthusiastic efforts contributing to the timely publication of this volume. Rohit Bhargava Ira W. Levin

1

Fourier transform mid-infrared spectroscopic imaging Microspectroscopy with multichannel detectors Rohit Bhargava and Ira W. Levin

1.1 Introduction Infrared (IR) spectroscopy is commonly employed for a range of spectrochemical analyses.1,2 A variety of technical advances over the last 30 years have resulted in the ability to rapidly record IR spectra using interferometers in relatively straightforward configurations. The scope of applications available to this technique has been enhanced by the versatility inherent in its instrumentation and multiplicity of sampling techniques which, when coupled to available data analysis approaches and spectral databases, provide insights in the characterization of molecular properties. IR spectroscopy, however, had been primarily a bulk material technique, since obtaining spectral information from microscopic sample volumes has often proved difficult. The convergence of optical microscopy and IR spectroscopy, however, now permits the high-throughput recording of spatially resolved spectral information, thus providing an attractive approach to assessing the component properties of complex materials. Although the first efforts in IR microscopy occurred more than 50 years ago,3 progress in making microspectroscopy routinely accessible to the larger vibrational spectroscopic community, in general, was slow. Considerable activity over the last 15–20 years, in particular, has resulted in the capability of routinely collecting IR spectra from narrowly delineated sample regions through a combination of IR microscopes, IR interferometers, sensitive detectors and powerful computers.4 These developments have resulted in the emergence of a Fourier transform infrared (FTIR) microspectroscopy subdiscipline.5–11 In the present discussion, we briefly review the evolution of spatially resolved IR spectroscopy in the mid-IR spectral region and the revolutionary changes in practice that are facilitated specifically by the integration of multichannel array detectors. Several applications of FTIR imaging, including its relevance to biological and polymeric materials analyses, are discussed elsewhere in this volume.

1.2 Fundamentals of FTIR spectroscopy Fourier transform infrared microspectroscopy couples both interferometry and microscopy into an integrated instrument. Since interferometry is an important

2

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

component of the technology, it is instructive to appreciate the salient features of conventional FTIR spectroscopy in the context of microspectroscopy and imaging. Since the fundamentals of FTIR spectrometry, including spectroscopic applications, are summarized in detail elsewhere,12 we briefly describe here the mathematical basis of interferometer data acquisition, the integration of an interferometer with a microscope and the performance metrics useful in spectrochemical analysis.

1.2.1 Interferometer characteristics The Michelson interferometer employed in FTIR spectroscopy utilizes a single source, recombines split radiation beams to realize interference and then measures the interference pattern of a broadband spectrum. Figure 1.1 presents a schematic demonstrating the concept of scanning interferometry. A thin, nonabsorbing beamsplitter divides a beam from a broadband radiation source into two perpendicular directions. These beams, after being reflected by mirrors along the normal, undergo interference at the beamsplitter where one component of the combined beam is reflected back to the source, while the second component is directed either to a detector or to a microscope as modulated light. A change of the relative position of the mirrors, by, for example, repositioning one mirror with the second remaining stationary, results in altering the path differences between the two interfering beams. The intensity of light at the detector due to interference at a single frequency may be written in terms of the wavenumber, ν, ¯ as I  (δ) = 0.5I (¯ν)[1 + cos(2πδ ν)] ¯ I

(1.1)

where, is the intensity directed to the detector, I is the original intensity of the incident beam at the beamsplitter and δ is the mirror retardation. The resultant intensity at the detector consists of a component that is invariant of the path difference and a second sinusoidally varying component. The invariant component, termed the direct current (d.c.) component, results in the addition of a constant value to Fixed mirror M1 M2

From source To source

B

To detector Beam splitter Figure 1.1 Schematic diagram of a Michelson interferometer.

Moving mirror

FOURIER TRANSFORM MID-IR IMAGING

3

the detector readout signal, which reduces the available dynamic range for large array detectors. The d.c.-component does not contribute, however, to the spectral signal and may be subtracted. The component of the detected signal that changes with mirror retardation is termed the interferogram. In practice, beamsplitters do not reflect and transmit exactly 50% of light; their recorded performance, which varies with wavelength and the detector response, including that of the associated electronics, affects the intensity of the recorded interferogram. Both the prefactor in Equation (1.1) and various instrumental efficiencies are included in a simple equation for the interferogram that relates the observed intensity at the detector to the broadband emission profile of the source, B, as a function of the retardation by  +∞ B(¯ν ) cos(2πδ ν¯ )dν¯ (1.2) I (δ) = −∞

Radiation path differences may be achieved in an interferometer by either a continuous motion of the mirror(s) or by incremental steps that sequentially move a mirror to a specific retardation and then, after a time delay, rapidly move the mirror to the next optical retardation. In this manner, a range of path differences are obtained. It is also instructive to visually examine the interferogram for a broadband spectral source as shown in Fig. 1.2. The large amplitude signal, or the centerburst, contains the majority of spectral frequencies that have been added constructively. Regions away from the centerburst are small in magnitude and are termed wings. Mathematically, the other half of the cosine pair derived from the above equation is given by the integral of the even function I (δ) such that  +∞ I (δ) cos(2π ν¯ δ)dδ (1.3) B(¯ν ) = 2 0

Hence, a spectral profile may be specified completely by the measured interferometric signal at a known sequence of interferometer mirror retardations. The spectral profile is then computationally recovered by measuring the interferogram from zero mirror retardation to an infinitely long retardation at infinitesimally small increments of retardation. It is clear that computational limitations and instrumental considerations require that the interferogram be sampled over a finite number of optical retardations. This practical consideration dictates that the sampling interval determines the measured spectral range while the finite number of mirror retardation steps serve to provide a measure of the spectral resolution achieved. While these details and their relationships are explained elsewhere,12 it suffices for spectroscopic analyses to understand that extended measurements are required to record either high spectral resolution or large spectral bandwidths. Bandpass filters coordinate the spectral output of the source to the measured spectral bandwidth to prevent spurious spectral features through mathematical aliasing.

1.2.1.1 Continuous-scan interferometers In the most common commercial implementation of interferometric measurements, the moving mirror is scanned at a constant velocity v. Consequently, the mirror

4

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Signal counts (a.u.)

9000 8000 7000 6000 5000 4000 0

500

1000

1500

2000

Retardation point 4200

Signal (a.u.)

3500 2800 2100 1400 700 0 500 1000 1500 2000 2500 3000 3500 4000 Wavenumber (cm–1)

Absorbance (a.u.)

0.5 0.4 0.3 0.2 0.1 0.0 1000 1500 2000 2500 3000 3500 4000 Wavenumber (cm–1)

Figure 1.2 Interferogram recorded by a d.c.-coupled detector in which the signal counts can vary from 0 to 16 384 (top). Fourier transformation of the recorded interferogram profile yields a single-beam spectrum (middle). Single-beam spectra from a sample can be ratioed point-by-point in the spectral domain to single-beam spectra acquired without a sample in the beam path, yielding absorbance spectra (bottom). The absorbance features in a spectrum can be correlated to the molecular properties of the sample (dark profile), while a featureless spectrum (light profile) denotes the lack of sample in the beam path.

FOURIER TRANSFORM MID-IR IMAGING

5

travels a distance given by vt in a time t. Since the path difference for radiation is twice of the difference in distance between the mirrors, the retardation can be expressed in terms of the velocity and time as δ = δ0 + 2vt

(1.4)

where δ0 is the retardation at the beginning of the observation time. Since the time is set arbitrarily, the point of reference can be taken to be the point of zero retardation, resulting in the optical retardation being set equal to 2vt. Analogous to velocity of the mirror, a rate of change of optical retardation may be termed the optical velocity, which is twice the velocity of the moving mirror. The expression for the interferogram, in terms of these directly measurable processing parameters, reduces to  ∞ B(¯ν) cos(4πvt ν)d ¯ ν¯ (1.5) I (t) = −∞

where t is the time after zero retardation. By comparing the sinusoidal variation of the interference pattern of a single wavenumber to a standard expression for a sinusoidal wave [cos(2πf t)], the characteristic frequency of the signal due to that wavenumber is determined. This characteristic frequency of the interferogram corresponding to a wavenumber, ν, ¯ is given by fν¯ = 2v ν¯

(1.6)

This frequency within the interferogram is termed the Fourier frequency corresponding to that wavenumber; consequently, the spectrum is encoded by the interferometer at different frequencies. The Fourier frequency is a function of the scanning mirror velocity and can be employed to increase the signal-to-noise ratio (SNR) of the commonly employed a.c.-coupled HgCdTe (MCT) detectors. When the Fourier frequency is greater than 1 kHz, the mode of interferometry is termed ‘rapid scan’. For the commonly recorded 4000–650 cm−1 range in the mid-IR spectral region, the rapid-scanning regime necessitates mirror speeds of at least 0.01 cm s−1 . Since the desired spectral range and resolution determine the data recording characteristics and the distance traveled by the interferometer’s moving mirror, the number of data points and their recording interval are easily specified with respect to the mirror motion. These quantities are critical since the detector’s data acquisition speed becomes important in the practice of FTIR spectroscopy when utilizing rapid-scanning interferometers. When mirror velocities are lower than the rapid-scan conditions, the mode of interferometer operation has been termed ‘slow scan’. At low velocities, however, control over mirror motion is difficult in commercial interferometers designed for rapid scanning, resulting in the measurement being susceptible to noise problems. Hence, continuous-scan interferometers generally operate above the rapid-scan limit.

6

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a)

(b)

d

d

t

t

Figure 1.3 Modes of the mirror scanning in interferometry (a) continuous-scan mode and (b) step-scan mode.

1.2.1.2 Step-scan interferometers In step-scan interferometers, the interferogram intensity is recorded at a constant retardation and the interferometer is subsequently stepped quickly to the next retardation. One mirror may be moved by itself while the other is held at a constant position in the interferometer, or the two mirrors may be moved in tandem to yield a constant retardation. The advantage of moving the second (usually stationary or fixed) mirror is twofold. First, the primary moving mirror can continue to scan, as in the rapid-scan mode, with no separate feedback, control or hardware modifications being required in comparison to a rapid-scan interferometer. Second, the second mirror may now be moved a short distance in a precise manner. Since control mechanisms are easier to implement and are more accurate for motion over short distances, independent control over the mirrors allows for greater flexibility in rapidly attaining a desired retardation. This configuration affords a greater flexibility in recording data since the retardation and the resultant Fourier frequencies imposed on the interferogram are decoupled from the time domain of data acquisition. Hence, unlike the rapid-scanning interferometry, any detector may be employed with the signal being recorded for any specific time period at each retardation step. Conceptually, in contrast to the rapid-scan implementation, the time period for recording spectral data need not be equal at every retardation. Although this concept has not yet been implemented, longer data recording times could be employed when recording data in the wings of the interferogram in contrast to the time expended in recording the centerburst. While an intensity profile at the detector as a function of retardation may be acquired in a step-scan mode, two major drawbacks affect this method of interferogram acquisition. First, the mirror(s) requires stabilization times with mirror inertia and time constants of the control loop determining this parameter in achieving a given optical retardation. Second, additional hardware and control mechanisms need to be incorporated into the spectrometer, thus increasing instrument cost and complexity. In certain cases, however, the utility of a step-scan instrument justifies this additional expense. Historically, the step-scan approach was favored with slow detectors. With the advent of fast detectors and electronics, step-scan interferometry became

FOURIER TRANSFORM MID-IR IMAGING

7

less popular. Step-scan interferometry, however, presents another advantage. While monitoring transient events that occur repetitiously, the mirror retardation can be held constant as the event is recorded as a function of time; the retardation is then changed with the event being triggered again. In this manner, the temporal evolution of signal from a repeated transient event can be obtained as a function of retardation. This procedure forms the basis of time-resolved FTIR spectroscopy and has been widely employed over the last decade.13–15

1.2.1.3 Performance figures of merit: the SNR The sensitivity and detection limits of an analytical technique are determined by the SNR of the measurement, an important metric for assessing both the instrumental performance and analytic limits of the spectral measurement. Following typical analytical practices, 3 and 10 times the noise have been suggested as limits of detection and of quantification for IR spectroscopy, respectively. The performance of interferometers in the continuous-scan mode, which is simpler compared with that of the step-scan mode, has been analyzed well. The SNR of a spectrum measured using a Michelson interferometer is given by12 SNR =

¯ ∗ξ √ Uν¯ (T )(ν)D t √ AD

(1.7)

where U is the spectral energy density at wavenumber ν¯ from a black-body source maintained at a temperature T . The specific detectivity of the detector, D ∗ , that has a sensing area denoted by AD measures the signal for a specific period of time, t. The throughput, , and efficiency of radiation transmission, ξ , of the interferometer also affect the quality of resulting data. The well-known ‘trading rules’ of FTIR spectroscopy16 provide for an easy understanding of the performance of spectrometers in terms of the experimental configuration and measurement variables. It must be emphasized that the trading rules are valid strictly for rapid-scanning interferometers, but their understanding is helpful in analyzing the performance of other modes of interferometry, as well. For example, an extension of the same principles to understand the complex dependence of SNR on measurement variables for stepscan interferometry is reported elsewhere.17 Most importantly, the SNR depends on the square root of the data acquisition time. Although the SNR of acquired data may be increased in a straightforward manner, an increase in the SNR by a factor, for example, of k, requires that the acquisition time be increased by a factor of k 2 . When the spectral resolution of recorded data (ν¯ ) is increased by a factor of k, the acquisition time must be increased by a factor of k 2 to maintain the SNR of the acquired data. If the throughput is decreased by a factor of k, the SNR decreases by the same proportion. This effect of throughput decrease on the SNR can be compensated by an increase in the data acquisition time by a factor of k 2 . (Though we have assumed that the specific detectivity and the dimensions of the detector change independently in this discussion and in a later discussion in the chapter, the assumption is not strictly valid. The characteristics of the detector change with size and the

8

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

wavelength sensitivity of the detector. As a first approximation for the limited range of sizes of detectors considered in our analyses, the characteristics of the detector may be considered constant.)

1.3 FTIR microspectroscopy using a single-element detector The first reports of a spectrometer interfaced to a microscope18 appeared more than 50 years ago with a commercial instrument being introduced within 6 years of that report. These configurations employed single-beam dispersive spectrometers for spectral measurements and all-reflecting microscope optics for localizing the beam. Single-beam spectra were plotted using a strip-chart recorder and absorbance information was generated by manual calculations of the relative background and sample scans. Clearly, this approach was ill suited to widespread use for rapid analysis of spatially small regions and, as a consequence, gained little acceptance as a major analytical tool. The microspectroscopy technique made no significant advances until the coupling of an interferometer and microscope to a digital computer provided modern Fourier transform spectroscopy, time averaging and mapping. (Though a microcomputer controlled microspectrometer was realized as early as 1978 (NanoSpec 20-IR from Nanometrics, Sunnyvale, CA), the first true FTIR coupling to a microscope was achieved a few years later by Digilab in 1983 based on a microscope built by Spectra-Tech – now, Thermo Electron Corp.) Since the light throughput is used more effectively in an interferometer and since the spectral response is precisely measured with respect to the wavelength, a significant increase is achieved in the SNR results. The SNR increase was also aided by the development of more stable, sensitive, fast-response cryogenic detectors that yielded reproducible measurements, further spurring interest in microspectroscopy. The use of a computer-controlled stage and digital spectral processing allowed for recording data from numerous sampling points on specimens, which facilitated mapping the composition of samples through sequential measurements. These units became commercially available in the 1980s and are, to date, popular analytical tools. Since the state-of-the-art instrumental microscopy configurations using multichannel detectors are derived from these systems, it is instructive to examine their concept and performance.

1.3.1 IR microscopes and point spectroscopy Although the IR microscope is similar to the optical microscope, it differs with respect to its physical construction and the source employed. Since the refractive glass optics of the typical optical microscope are not suitable for IR microscopy, as glass optics do not transmit wavelengths longer than ∼4.5 μm, microscopes for wide-band IR spectroscopy incorporate all-reflecting optics and aspherical reflecting surfaces in a Cassegrain-type configuration for minimizing optical aberrations. Refractive elements, if required for special needs, are constructed from IR transmitting materials that are resistant to moisture (e.g. CaF2 ). Although refracting

FOURIER TRANSFORM MID-IR IMAGING

9

elements may be incorporated, we note that typical IR microscopes operate over a significantly larger bandwidth (>12 μm) than optical microscopes (<0.5 μm). The all-reflective Schwarzchild objectives employed for IR microscopy perform adequately, in terms of optical aberrations, over this large bandwidth, which provides an additional disincentive for refractive optics. The central obscuration of the reflecting lens (∼25%) results, however, in a reduction of throughput and redistribution of energy from the beam waist at the focal plane. Commonly available microscopes are of the ‘infinity-corrected’ design, in which the output of the objective is collimated, allowing error-free incorporation of polarizers, filters or beamsplitters. Typical magnifications that are available for IR microscopes include 6X, 15X and 32X, while the numerical aperture of the objective ranges from 0.3 to 0.7. The second difference between an IR spectral and an optical microscope is the source. The typical interferometer employed for conventional FTIR spectroscopy can be utilized for microspectroscopy with the modulated beam being directed to a microscope instead of the spectrometer’s sampling chamber. No modifications are required to the interferometer and an attached computer can initiate data acquisition from the interferometer at specific sample spatial positions. A rapid-scan interferometer is generally used as a source of radiation for SNR considerations, although the technique is not modulator specific. The IR microscope also conveniently serves as a visible light microscope for the purpose of sample alignment. As there is no means to visualize the IR image in a single detector configuration, the sample is first viewed using optical microscopy to delineate the region for recording data. To obtain IR spectra from these defined regions, apertures are employed to selectively limit the wide field of view (FOV) of the optical signal. Apertures may be circular (for point measurements or measurements on small samples) or they may be rectangular (for mapping). The rectangular apertures consist physically of four independently adjustable knife blades that are manipulated through software to provide precise square or rectangular shapes for light transmission. Apertures were conventionally coated with highly absorbing carbon black to eliminate stray radiation but now consist of IR absorbing glass that transmits visible radiation, allowing for uninterrupted viewing of the brightfield optical image. IR spectra are subsequently acquired through the same configuration, requiring the optical paths for both IR and visible radiation to be parfocal and collinear. Since this optical configuration allows a spectrum to be recorded from a small sample region, the approach is often termed ‘point microspectroscopy’. The aperture–microscope configuration allows for the routine examination of small samples in the picogram range. Since both distortions due to the sample being smaller than the beam diameter can be avoided and contributions from the surrounding material can be eliminated, the sensitivity of the detection can be substantially increased. Radiation incident on a sample results in radiation interacting with the sample to be transmitted, reflected, refracted, absorbed or emitted. Almost any of these modes19 may be used for examining the absorbance characteristics in a microscopic configuration. Since microscopy places unique sampling demands upon an

10

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

CCD visible detector

Single element infrared detector

Aperture

Sample

Precision stage Aperture

Turning mirror Rapid-scan interferometer

Visible light source

Microscope

CCD visible detector

Multichannel infrared detector

Sample Precision stage

Turning mirror Rapid-scan interferometer

Visible light source

Microscope

Figure 1.4 Layout and principle of operation of IR microscopes when single-channel detectors are employed (top) and when multichannel detectors are employed (bottom). Single-channel detection requires the use of apertures. Multichannel detectors typically range from a linear 16-element detector to large 65 536-element (256 × 256) detectors.

investigator, certain techniques are more popular than others. Specifically, transmittance and reflectance approaches have been the most utilized due to the ease of sample preparation and the conduct of experiments. Light transmission through the sample allows for easy sample positioning, good light throughput and yields spectra that require little processing effort. Sample preparation becomes more involved, however, as most materials are strongly absorbing. Therefore, the preparation of thin samples usually requires experience and expertise in determining optimal thicknesses. Microtoming a sample is often necessary, and appropriate

FOURIER TRANSFORM MID-IR IMAGING

11

embedding media20 are sometimes used in preparing small samples. While spectra may be obtained by flattening a sample21 or by using a diamond cell22 to press the samples into thin layers, undesirable effects arise, which include the loss of orientation, interference fringes and polarization scrambling. Once an appropriate sample is obtained, quantification of recorded data is straightforward using Beer’s law. Reflectance spectroscopy is commonplace for samples supported on a reflecting substrate but suffers generally from poor data fidelity. Attenuated total reflection (ATR) has been combined with the microscope to perform microspectroscopy.23,24 An alternative approach has been developed that uses cartridges with hemispherical ATR elements and can be used in any microscope with reflectance capabilities.25 Adding to the versatility of the technique, temperature control may be achieved using a standard microscopy cell with IR transmitting windows.26 There are no limitations to employing constant humidity and pressure (and vacuum) cells if the need arises.

1.3.2 FTIR mapping Single-point examinations are often of limited use since statistical analyses are usually precluded in the examination of multicomponent systems. By moving the sample in a known, predetermined manner relative to the aperture, a point-by-point examination of contiguous areas of any size can be conducted. Plots of the absorbance of a specific vibrational mode yield a ‘chemical map’ of the visible image of the sample; this mode of data acquisition is termed ‘point mapping’. While optical microscopes use differences in refractive index, selective phase staining or polarized light for imaging, the vibrational spectroscopic signature of a material provides the contrast image in FTIR mapping. Unlike the wide-field observation of a visible light camera in optical microscopy, each spatial element in a map is individually measured. A transition from a point microscopy configuration to a mapping configuration requires the additional incorporation of a sample positioning stage to align the sample precisely and reproducibly. The computer recording the spectral data is also employed to control the sample stage and to register the recorded data to correspond to the optical image. Since spectral information from small sample regions is obtained by restricting the area illuminated by the IR beam using opaque apertures of predefined size, the dimensions of the truncated beam at the sample plane nominally determines the lateral resolution of the optical configuration. An advanced discussion regarding the spatial resolution in mapping systems is provided in Chapter 3 of this volume. The pathlength through the sample determines the absorbance of the sampled volume and depends nominally on the sample thickness.

1.3.3 Limitations of FTIR point mapping The speed of data acquisition in point mapping protocols generally precludes the examination of large sample regions at high spatial and spectral resolution. In addition to the usual trading rules of FTIR spectroscopy, there are tradeoffs in data

12

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

acquisition in mapping that stem from the use of apertures. A reduction in the dimensions of the aperture to obtain higher spatial resolution leads to a loss (with the condensing optics unchanged) in throughput. When the aperture dimensions approach the wavelengths of light, there is an additional decrease in the throughput due to diffraction effects. The decreased throughput results, in turn, in a proportional decrease in the SNR of the acquired data. The SNR decrease necessitates a compensation by time averaging to maintain the data quality, thereby increasing the data acquisition time in a quadratic manner. Thus, the three interdependent parameters in an experiment, namely, spatial resolution, SNR and data acquisition time, couple strongly to limit the effectiveness of the technique as a mapping tool to examine rapidly either extended samples or large numbers of samples. For small samples or for single-point measurements, however, this instrumental configuration remains the mode of choice due to lower detector costs and easy availability of versatile instrumentation.27 Since apertures decrease throughput, an ideal design would incorporate condensing optics to reduce the output beam of the interferometer (typically 5–10 mm) to the aperture size at the sample plane. In practice, however, the aperture size usually cannot be specified. Due to the tradeoffs discussed above, the aperture size is the most commonly, and easily, modifiable adjustment in experiments. Since spectra from specimen regions ranging in size from ∼1 mm down to 5 μm are commonly acquired by microscopy methods, the illuminated area at the sample plane must be fairly large. Similarly, the detector size and radiation throughput cannot be matched using only simple optics. Hence, a general design is employed as a compromise between the need to match throughput and instrumental versatility. The beam spot size at the sample plane is ∼750 μm (limited by both the condensing optics and the relay mirrors of the microscope) while the detector size is usually ∼250 μm × 250 μm. The detector size is problematic for mapping measurements where the aperture size may be 5–50 μm. Assuming unity magnification between the sample plane and the detector, the lack of size matching results in the detection of an interferometric signal by only a small fraction of the sensing area of the detector while the entire detector area contributes noise to the data. The deleterious effect on SNR of the recorded data is most apparent when mapping specimens at small aperture sizes for attaining high spatial fidelity. Newer microscope designs incorporate magnification optics between the sample and the detector, but it is obviously difficult to design optimal optics when the aperture size can be continuously changed. Smaller detectors (25–100 μm in size) are available and may be employed for mapping studies. Apertures are the greatest source of diffraction in the IR microscope and can lead to the detector sampling light from outside the apertured region. The detector also samples the secondary lobes of the diffraction pattern. Thus, spectral information from the delineated area is spread over a larger area than the aperture. A case study revealed the effects of this stray light sampling as many as 40 μm away from the sample.28 Consequences of this stray light on spectra were studied under conditions of different aperturing modes and sizes. This problem can be circumvented by using

FOURIER TRANSFORM MID-IR IMAGING

13

a second aperture to delineate the same region (redundant aperturing). The spectral purity is now increased at a cost to the amount of light allowed through the system. Hence, spectral quality (SNR) is degraded. With dual aperturing in the mid-IR range, the practical resolution limit is close to 15 μm. This lower limit may be further increased to ∼10 μm, but at a substantial cost in time. Near-field acquisition techniques have been suggested to improve spatial resolution and to mask stray light. One such study29 analyzed the acquisition of a physically masked sample (i.e. no mapping was possible). A near-field aperture turret to allow for mapping has also been proposed.30 While diffraction effects have been a major source of frustration, Siedel aberration may also serve to limit mapping fidelity.31 In particular, centrally obscured reflecting optics are prone to spherical aberration, a situation that requires adjustments to the optics after mapping small sample areas and which limits the FOV in wide-field analyses, if the sample is not to be disturbed. In addition, the central obscuration reduces light throughput, further starving the detector of already scarce radiation. Numerical apertures of reflecting systems are limited (typically to <0.7) utilizing a magnification of no more than 35X. In general, these factors are not particularly important for routine FTIR microscopy in the mid-IR spectral region, and despite these limitations prove to be better suited – scientifically and economically – than corrected refractive optics utilizing complex lens designs. If the required wavelength range for IR analysis allows for the use of materials (e.g. glass or CaF2 ) that may prove to be economically feasible, refractive optics that provide higher numerical apertures and spatial resolutions may be readily employed. However, chromatic aberration would have to be corrected by appropriate microscope design.32

1.4 FTIR imaging with multichannel detectors Fourier transform infrared spectroscopic imaging is the made possible by replacing the single-element detector in IR microscope–interferometer assemblies by multichannel array detectors. Although multichannel detectors for IR spectroscopy33 and for interferometric imaging had been reported earlier,34 applications of large format array detectors to IR microspectroscopic analyses have now permitted extensive structural descriptions at length scales relevant to materials science, biomedicine and forensics.

1.4.1 Imaging with large format array detectors Large format array detectors consist of numerous, small detection elements arranged in a two-dimensional grid. Array detectors are akin to the ubiquitous modern digital camera in that they are passive recording devices that simply record the incident intensity of the entire FOV at a specified time. Hence, they can be incorporated

14

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

into interferometric detection schemes in much the same manner as conventional IR detectors that are integrated in mapping instruments. The large number of individual detectors in an array configuration presents some unique complications, however, which prevent their straightforward utilization. These complications stem from the poorer detection characteristics of the miniaturized detectors (individual pixels) compared with monolithic single-element detectors and from the need to organize and digitally record the intensity values from several thousand pixels of the array. Here, we do not present a detailed review of array detectors and their performance,35,36 but present the unique performance parameters that are relevant to IR spectral analyses. The time required to acquire a frame corresponding to an image, the shortest time between sequential acquisitions of data by a pixel in an FPA, is termed the frame time; the number of frames acquired per second is termed the frame rate. The time taken to record a frame includes both the time required to acquire the signal and the time required to digitally record it. The acquisition time is the time period in which the signal from the photon flux is accumulated in terms of the resultant electrical charge. This period of time is smaller than the frame time and is termed the integration time. The integration time can usually be adjusted to provide optimal utilization of the dynamic range of the detector. For low flux conditions, an increase in the integration time allows for a larger interferometric signal to be detected. Since larger FPA detectors are d.c.-coupled, the thermal background contribution to the flux is also increased, thus reducing the effective dynamic range for the interferometric measurement. Optical designs should include the insertion of an optimized cold shield37 for rejecting the background thermal signal. The d.c.-component of the interferogram also contributes to the signal, further limiting the effective dynamic range of the detector. Array detectors may be operated in a ‘snapshot’ or in a ‘rolling’ mode based on the underlying readout electronics. In the snapshot mode, all pixels record intensities at the same time. In the rolling mode, a few rows of pixels record data at a time. Data readout and storage to the computer may, for example, be two rows of the array at a time followed by the next two and so on. In this manner, the recording of data ‘rolls’ along the array. While a small part of the array is acquiring data, the remainder of the array is idle. Hence, this mode of data recording is inefficient. In contrast, snapshot mode arrays provide opportunities for both efficient and temporally precise imaging and, consequently, are being increasingly employed for FTIR imaging. An interesting side-effect involving the use of large array detectors results in a simplification of microscope design. For point mapping experiments, the visible and IR optical trains must be parfocal (form images at same planes) and collinear (follow same paths) to allow preliminary identification by visible light microscopy and data acquisition by apertured spectroscopy. Since the FPA can directly form an image based on the scattering from the sample (as for brightfield microscopy) and absorption of all vibrational modes in the IR spectral region, an optical microscopy image is not required for reference.38 As an example, an IR brightfield image is compared with an IR absorbance image in Fig. 1.5.

15

FOURIER TRANSFORM MID-IR IMAGING

(a)

(b)

(d)

(e)

(c)

High

200 mm

Low Figure 1.5 (a) IR brightfield image of a human skin sample, (b) nucleic acid–amide III absorbance area (1120–980 cm−1 ) corresponding to the absorbance features, (c) amide II absorbance area (1592– 1480 cm−1 ), (d) OH–NH absorbance area (3670–3130 cm−1 ) and (e) orthogonalized Gram–Schmidt (G–S) intensity images for the tissue sample. While brightfield images provide differences based on scattering and absorbance, chemical images provide contrast based on specific vibrational modes and G–S processing of tissue can provide rapid visualizations of structure based on differences in entire spectra. The intensity grayscale denotes regions of low intensity by shades closer to black and those of high intensity closer to white for photon flux in the in the brightfield image in (a), for the absorbance magnitude for modes described in (b)–(d) and for the G–S intensities in (e).

1.4.2 Interfacing an interferometer to large array detectors Although continuous-scan interferometry is the most common mode of data acquisition in commonly available FTIR spectrometers, it was not the method of choice for the first commercial imaging configurations for wide-bandwidth IR microscopy. As noted previously, since the optical retardation is coupled to the time domain in continuous-scan interferometry, signal detection must be sufficiently rapid to prevent data acquisition errors. The rates of data acquisition of early FPA detectors did not permit the acquisition of high fidelity data in a rapid-scanning configuration. The decoupling of mirror retardation from the time domain in the step-scan mode of interferometric measurements, in contrast, allows the acquisition of data at specific retardations without regard to the speed of data recording. Hence, the coupling of a step-scan interferometer to a microscope equipped with an FPA detector allowed the acquisition of FPA frames at specific retardations using detectors of any frame rate. Since the inefficiency of the data acquisition process is well-known for step-scan interferometry, several other attempts were made to increase data fidelity. These modifications in the data acquisition, which can be better understood in the

16

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

theoretical context of increasing the SNR of acquired data, are discussed in the next section. Although the first imaging spectrometers employed continuous scanning with narrow-band detectors,34 continuous-scan imaging measuring the usual bandwidth of FTIR spectroscopy for chemical analysis was accomplished by using a low mirror velocity.39 Since the study employed a rolling mode detector, the minimum sampling interval between sequential measurements of the interferogram intensity could only occur after the entire array had completed a recording cycle. The mirror displacement in the interim was required to be minimal, leading to low mirror scanning rates. Since signal integration in snapshot FPAs occurs simultaneously for all pixels, the interferograms for all the pixels depend only on the sample characteristics, making phase corrections simpler. The simplest method for employing large FPA detectors in a rapid-scan configuration is to increase the frame rate through better electronics. We employed a detector, developed commercially in a slow-scan approach, for rapid-scan imaging. (Bhargava, R., Huffman, S. W., Levin, I. W. and Wang, S. Q., Unpublished results, 2002.) The interferometer scan speed was 0.4 cm s−1 and the resulting Fourier frequencies at the long wavelength cutoff (950 cm−1 ) was 760 Hz. Advances in detector technology are enabling faster readouts, and imaging is likely to be conducted at interferometer mirror speeds commonly encountered in conventional rapid-scan recording using single-element detectors. Since the noise in recorded data is primarily determined by the FPA, a tremendous economical advantage is realized if the interferometers employed for imaging can be the same as those used in low-end, mass-market scanning FTIR spectrometers. In an alternate approach to employing large detectors for rapid-scan measurements, a multipass approach has been suggested.40 In this approach, the mirror scanning velocity is similar to that of a typical rapid-scan device but the FPA measures a highly undersampled interferogram. Since IR interferometers are stable enough for noise in measurement to be dominated by FPA detection noise, the interferometer is allowed to scan while undersampled interferograms are acquired. By synchronizing the start of data collection at each pass to be offset from that of the previous pass, an entire interferogram can be reconstructed. This detection scheme allows FPA detectors of almost any readout characteristics to be coupled to commonly employed rapid-scan interferometers.

1.4.3 The SNR of imaging spectrometers The SNR analysis for a single-element detector in a point mapping configuration was adapted to analyse the SNR of a pixel in an imaging FPA detector as42  0.12πA(1 − 1 − (NA)2 )Uν¯ (T )¯νD ∗ 1/2 SNR = t (1.8) √ AD where NA is the numerical aperture and A the area of the sample imaged onto the pixel. The correction due to the decrease in throughput arising from apertures is not required and the interferometer and transmission efficiencies are taken to be

FOURIER TRANSFORM MID-IR IMAGING

17

commonly encountered values. The experimentally measured values were found to be significantly lower that the theoretically predicted performance. The understanding of the SNR in imaging systems is improved further by incorporating the detector characteristics in the analysis. The minimum collection time to obtain one interferogram data cube is given by t = ns td + nf tf + tr 

(1.9)

where ns is the number of spectral resolution elements in the interferogram, tf is the time required to acquire each frame, nf is the number of frames, td is the delay before data acquisition in each step of a step-scan operation (zero for rapid-scan operation) and tr is the time required for readout, coaddition (if done during data collection) and storage. As discussed previously, the integration time is adjusted to provide maximum utilization of the dynamic range while the delay and readout times depend on the characteristics of the interferometer and the array detector, respectively. Thus, the collection time depends on both the number of spectrometer steps, which is a function of spectral resolution, and range, and also on the detector characteristics and number of frames used to obtain the average frame data. The SNR equation for an imaging-detector-equipped interferometer is then 1/2  n f tI Uν¯ (T )νξ ¯ SNR = t 1/2 (1.10) NEP td + nf tf + tr where ξ incorporates the combined efficiency of the spectrometer and additional optics. Clearly, the scaling of the SNR with number of coadded frames for a given experimental time is not simple. The effect of changing the number of frames on the SNR can be understood through the acquisition ratio, εC = √ [(nf tI )/(td + nf tf + tr )]1/2 , which scales as nf in the low frame number limit and approaches a constant value in the large frame number limit. The efficiency can never be larger than a limiting acquisition ratio, which is dictated by the system electronics, that is, the square root of the ratio of integration time to frame time. The acquisition ratio is a measure of the efficiency of the data acquisition protocol and can be used as a stand-alone comparison metric for different data acquisition protocols. The first generation of imaging systems employed detectors with tI /tf ratios of ∼0.01, which accounts largely for the discrepancy between the predicted and observed data quality. Newer detectors are approaching ratios of 0.99, providing a significant increase in the capabilities of imaging spectrometers. Clearly, frame rates that are matched to integration times will provide the best performance. Continuousscan instruments require neither a delay time nor a readout time. Hence, they offer enhancements in efficiency compared to step-scan systems. It must be noted that the expression provides an accurate prediction for a small number of coadded frames. A deviation from predicted values for coadding more than ∼20 frames arises likely from low frequency components of FPA noise.41 Frame coaddition in FPA detectors and spectral coaddition, as employed conventionally in FTIR spectroscopy, may be combined to achieve a net higher SNR. Since the time available for an experiment is usually specified, a data acquisition

18

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

strategy may be devised to attain the highest possible SNR in that time period. A plot of SNR achievable against experimental time for different collection strategies, termed characteristic plots, has been proposed to determine the optimal operating point.42 While the idea was proposed for step-scan data acquisition, the principle is valid for in-scan coaddition in a continuous-scan acquisition mode and may be employed for multipass modes of data acquisition. A number of alternate approaches have been suggested to increase the SNR of imaging data based on the theoretical understanding of the SNR expression. These approaches rely on changing the data acquisition configuration43 or on changing the parameters of the FPA during data acquisition.44 The unique noise sources and behavior of FPA detectors prevent a straightforward implementation of some techniques employed to increase the SNR of FTIR spectral data. For example, the gain ranging concept can be employed to increase the SNR for an FTIR spectrometer by selectively increasing the amplitude of the response by a multiplicative gain factor for selected regions of the interferogram.44 Hence, data around the centerburst are collected at a lower gain in comparison with the rest of the interferogram. The selectively modified portions of the interferograms are then renormalized by the inverse of the gain factor before Fourier transformation to the wavenumber-intensity domain. The gain ranging process results in a multiplicative improvement factor in SNR given by a sum of amplification weights for every point in the interferogram, namely,45  Sr =

2rp + 1 np



 +

1 Rg

2 

2rp + 1 1− np

 −1/2 (1.11)

where SI is the ratio of the SNR of the interferogram collected using gain ranging to the SNR of the interferogram collected without gain ranging for the same detector under equivalent data acquisition conditions; Rg is the ratio of the amplification factor for the signal at the high gain setting to the signal at a low gain setting; rp is the number of points on either side of the centerburst collected at a lower gain and np is the total number of interferogram data points. The benefits of gain ranging were limited in experimental implementation by the nonlinear noise increase with increasing gain of the miniaturized amplifiers in the FPA. Hence, the expression was modified to incorporate the unique effects of noise as a weighted factor in a manner similar to that of the gain factor. Gain ranging does not involve any increase in data acquisition times and involves only a negligible increase in computational time, while providing an SNR increase of ∼45%. Additional attempts at SNR improvements through filtering the interferogram,46 spectral correlations to reduce noise47 in a transformed space48 or through spatial image processing routines are also available.49 While many routes for high fidelity imaging data have enjoyed varying degrees of success and limitations, any method that requires neither expensive hardware modifications nor involves major changes in collection parameters or data acquisition times is attractive.

FOURIER TRANSFORM MID-IR IMAGING

19

1.4.4 The evolving detector array technology Infrared sensitive array detectors are employed in a number of diverse fields including spectroscopic imaging, remote sensing, astronomy and thermal imaging. Array detector development and technology are separate areas of active research. A detailed description of detector theory, technology and current developments are provided elsewhere.50 We limit the discussion here to the manner in which the evolving detector technology has impacted spetrochemical analyses. The first FTIR imaging approaches34,51 that incorporated indium antimonide (InSb) array detectors employed relatively long signal recording times (>4 ms per frame) but were large in size (128 × 128 pixels).52 The limited wavelength sensitivity of InSb detectors in the mid-IR, however, limited the examination bandpass to wavelengths shorter than ∼5 μm and precluded recording of data in the important ‘fingerprint region’ of the spectrum. A limited wavelength range in the fingerprint region of the spectrum (8–12 μm) was accessible through a gallium–silicon (Ga : Si) FPA.53 Interfacing an arsenic-doped silicon (Si : As) focal plane array in a manner similar to that used previously for InSb arrays allowed recording of spectral data in an extended 4000–400 cm−1 spectral range incorporating the fingerprint region.54 The detector operated at liquid helium temperatures (4 K) and recorded data from a modest 1024 pixels (64 × 16), limiting the potential of this technology. The introduction of HgCdTe (MCT) arrays, the detection material of choice in conventional FTIR spectroscopy, allowed the recording of spectral information from the fingerprint regions while maintaining the detection elements at manageable cryogenic temperatures (77 K).55 MCT detectors, although having higher specific detectivity (D ∗ ) and an extended wavelength sensitivity, offered smaller array formats and slower readout rates. The initial readout rates (180 Hz) of 4096 pixels (64 × 64) lead to increasingly higher readout rates (315 Hz, 430 Hz and, finally, ∼600 Hz) for the same array detectors by improving the readout electronics. The increased readout rates of the detector also resulted in increased electrical noise. A number of detector arrays and their applications in a variety of fields are summarized elsewhere.56 Large MCT array detectors (256 × 256 pixels), although operating at lower frame rates (114 Hz) compared with the smaller format detectors, resulted in a greater number of pixels recording data per unit time due to their large multichannel detection advantage. This new generation of arrays employed a snapshot mode of data recording for the first time and allowed for true recording of rapid-scan data. Detectors incorporating faster electronics using moderate sized arrays (128 × 128 format at 1600 Hz, 64 × 64 at 3774 Hz and smaller arrays) were introduced a few years ago. The introduction of very small array detectors (16 × 1) with very high readout rates (>16 kHz) resulted in dramatic changes in imaging technology (next section). We have recently utilized a moderately large (128 × 128 format) array reading out at high frame rates (∼16 000 Hz), (Bhargava, R., Bartick, E. G., Schwartz, R., Peters H. and Levin I. W., Unpublished results) incorporating the fast readout of small arrays with the multichannel advantages of the larger arrays.

20

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Although the readout rates of the very small linear array and the 16 384 pixel array are nearly the same, the architecture of the larger array results in a significantly higher noise per pixel. Advances in manufacturing technology and the increasing level of sophisticated applications have resulted in a rapidly improving quality of available detectors. We anticipate that improvements in detector technology will result in continuously better noise characteristics at kHz readout rates similar to single-element MCT detectors. Since there is little advantage in detection gained by increasing the modulation frequency of FTIR spectroscopy beyond 1 kHz, we anticipate that the speed of data recording will become less important while the need for lower noise readout electronics will become apparent for the small integration times at these high frame rates.

1.5 Raster scanning with linear array detectors While large format array detectors have greatly impacted the practice of IR microspectroscopy, their developing sophistication and lack of straightforward handling makes it difficult to translate conventional IR spectroscopy to familiar platforms. The utilization of several key features of conventional FTIR spectroscopy, as, for example, a.c.-coupling, high frequency modulation and filtering, have been impossible. Small linear array detectors have been developed as a compromise between the multichannel advantages of the array detectors and the high fidelity features of FTIR spectroscopy. The design principle underlying these detectors is that they may be employed in a manner similar to conventional single-element detectors but that the spatial resolution would be determined by the optics of the system and would not require apertures, thus trading the combined advantages of imaging and rapid-scan spectroscopy against the multichannel detection advantages of large detectors. Individual elements of the linear array detector are smaller than the typical detector employed for mapping; in contrast to mapping, the optics permit a very limited number of magnifications. For example, a commercial system employs a 16-element detector array and records data from specimen areas of either 6.25 μm × 6.25 μm or 25 μm × 25 μm. A 4X optical magnification allows for convenient change between the two available nominal spatial resolutions. Since the number of detectors is small, their uniformity is relatively high and circuitry can be independent. Further, the readout electronics do not need to be miniaturized at the expense of quality and provide, pixel-for-pixel, significantly superior performance compared with large focal plane array detectors. Another commercial implementation utilizes a two-column detector, providing for an increased multichannel detection advantage. Analogous to point mapping systems, linear array detectors require an associated optical microscope and a high precision stage for specimen positioning. The image is built one linear element at a time, while the sample is moved to raster across the entire area of interest; hence, these systems are termed ‘raster scanning’ imaging spectrometers.

FOURIER TRANSFORM MID-IR IMAGING

21

1.5.1 Choice of either small or large detector arrays Small array detectors clearly demonstrated advantages over the prevalent large arrays when they were first commercially introduced. The development led to other vendors adopting the paradigm in small square arrays (16×16) or proposing other rectangular array formats (16 × 1 or 16 × 2). The small detector arrays afford many advantages: simpler rapid-scan interferometry, a.c.-coupling of detection and software and hardware filtering. Since the field of view is limited, the sample must be raster scanned to image a large area. The inefficiency of stepping the stage sequentially to record data is partly offset by the gain in efficiency in Fourier transforming the data during the stepping period. Thus, the rastering time of the stage is offset by the processing time for large format arrays in comparing resulting data in spectral, rather than interferometric, format. Just as point microscopy systems have an advantage over imaging systems in recording data from a single point, smaller linear detectors clearly have an advantage for data acquisition from small spatial areas. Similarly, there is no substitute for large array detectors if data from large sample areas are to be acquired rapidly. The comparison between linear array and two-dimensional array detectors is more interesting when data are acquired from large sample areas and data quality, not time for data acquisition, is a constraining variable. We have utilized a performance metric that is suitable for comparing the imaging performance for large spatial areas of two different systems. The figure of merit we employ is the number of pixels per minute (pixpm) that can be recorded for specified data quality (SNR of absorbance spectra) at defined spectral and spatial resolution, apodization and wavelength bandpass. Since the SNR √ of recorded data can be anticipated to scale with the data acquisition time (as t), the adjustment allows for easy comparisons of detectors with different sizes and performance. Consider the comparison between two detectors in terms of the ratio of their pixpm output as    SNR2 2 n2 t1 (1.12) R21 = n1 t 2 SNR1 where R21 is the ratio of the pixpm for detector 2 to detector 1, ni is the number of pixels measured in a single interferometer sweep, in time ti , by detector i. The SNR of a pixel for the detectors is indicated by SNRi . Hence, the relative performance of the detectors is dictated by a linear dependence on the number of pixels and time for data acquisition but by the square of the SNR. Smaller linear arrays that can rapidly record data and record data at higher SNR due to their various advantages discussed above may overcome the multichannel detection advantage of very large array detectors. The relatively poor SNR of large array detectors arises from both the high noise characteristics of the miniaturized electronics of the array pixels, the lack of modulation advantages of rapid scanning as well as the need to distribute source energy over a wider spatial area on the specimen. Larger arrays are being manufactured and with improvements in FPA electronics making data acquisition faster and increasing the SNR, larger arrays are likely to dominate future instrumentation.

22

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

The comparison in the future will likely be determined by the need to distribute source intensity for larger arrays compared with the need to raster scan for small array detectors.

1.6 Conclusions Fourier transform infrared spectroscopic imaging, utilizing multichannel detectors, has considerably expanded the scope of IR spectrochemical analyses. The challenges of incorporating array detectors in imaging spectrometers have been actively addressed and many configurations now provide high quality data that enable sophisticated molecular analyses with a structural underpinning. The direction in this field is clearly shifting from efforts enabling high fidelity data to be recorded to approaches employing the recorded data for sophisticated analyses. The scope of analyses is being extended further, for example, by the recent development of timeresolved imaging57,58 (Fig. 1.6) and by the use of the ATR configuration59,60 to enhance temporal and spatial resolution capabilities, respectively. We anticipate that the evolving array technology will soon provide both large detectors with frame rates in the range of several tens of kHz leading to, for example, video rate spectroscopic imaging and cheaper arrays that allow for the introduction of imaging systems that (a)

(b)

FPA sampling Droplet bulk FOV average Defect Domain boundary Matrix

0.45

Time (ms)

0.40 I=I(=i, )

0.35

0.30

0.25 10

100 Time (ms)

Figure 1.6 (a) In a time-resolved imaging configuration, the interferogram is obtained by observing the magnitude of signal at a constant retardation in the step-scan mode. Cyclic events are excited by a stimulus and the intensity of the interferogram as a function of time is recorded for every optical retardation. The stimulus and response can be observed in the figure. The characteristic profile of events may be measured by stimulating multiple times to obtain a composite profile that contains a sufficient number of measurements to accurately reproduce the reponse using a multiple-pass approach to accurately record the data. (b) Average absorbance profiles during the relaxation of orientation for the different regions of a composite demonstration differences in the response of dispersed droplets, defects, domain boundaries and the embedding matrix. A nonspatially resolved measurements of the composite would provide a profile similar to the average of the FOV. The best fit lines are superimposed for all except the average of the FOV.

FOURIER TRANSFORM MID-IR IMAGING

23

are only incrementally more expensive than their point mapping counterparts. We also anticipate the introduction of instrumentation in which FTIR imaging is integrated with other forms of microscopy, as, for example, fluorescence techniques, to provide comprehensive, detailed characterizations of complex materials.

References [1] Chalmers, J. M. and Griffiths, P. R. (eds) (2002) Handbook of Vibrational Spectroscopy, John Wiley & Sons, Chichester. [2] Painter, P. C., Coleman, M. M. and Koenig, J. L. (1982) The Theory of Vibrational Spectroscopy and its Application to Polymeric Materials, John Wiley & Sons, New York. [3] Barer, R., Cole, A. R. H. and Thompson, H. W. (1949) Nature 163, 198. [4] Harthcock, M. A. and Atkin, S. C. (1998) Appl. Spectrosc. 42, 3. [5] Kwiatkoski, J. M. and Reffner, J. A. (1987) Nature 328, 837. [6] Reffner, J. A., Martoglio, P. A. and Williams, G. P. (1995) Rev. Sci. Inst. 66, 1298. [7] Jamin, N., Dumas, P., Monicut, J. et al. (1998) Proc. Natl. Acad. Sci. 95, 4837. [8] Wetzel, D. L. and LeVine, S. M. (1999) Science 285, 1224. [9] Chalmers, J. M., Everall, N. J., Hewitson, K. et al. (1998) Analyst 123, 579. [10] Bhargava, R., Wang, S. Q. and Koenig, J. L. (2003) Adv. Polym. Sci. 163, 137. [11] Diem, M., Romeo, M., Boydson-White, S., Miljkovic, M. and Matthaus, C. (2004) Analyst 129, 880. [12] Griffiths, P. R. and de Haseth, J. A. (1986) Fourier Transform Infrared Spectrometry, WileyInterscience, New York. [13] Sakai, H. and Murphy, R. E. (1978) Appl. Opt. 17, 1342. [14] Palmer, R. A., Manning, C. J., Rzepiela, J. A., Widder, J. M. and Chao, J. L. (1989) Appl. Spectrosc. 43, 193. [15] Uhmann, W., Becker, A., Taran C. and Siebert, F. (1991) Appl. Spectrosc. 45, 390. [16] Griffiths, P. R. (1972) Anal. Chem. 44, 1909. [17] Bhargava, R. and Levin, I. W. (2001) Anal. Chem. 73, 5157. [18] Burch, C. R. (1947) Proc. Phys. Soc. 59, 41. [19] A general discussion of sampling techniques for microscopy may be found in Allen, T. J. (1992) Vib. Spectrosc. 3, 217. [20] Augerson, C. C. (1998) Appl. Spectrosc. 52, 1353. [21] Tungol, M. W., Bartick, E. G. and Montaser, A. (1993) Appl. Spectrosc. 47, 1655. [22] Lang, P. L., Katon, J. E., Schiering, D. W. and O’keefe, J. F. (1986) Polym. Mater. Sci. Eng. 54, 381. [23] Harrick, N. J., Milosevic, M. and Berets, S. L. (1991) Appl. Spectrosc. 45, 944. [24] Gentner, J. M. and Wentrup-Byrne, E. (1999) Spectrochim. Acta. A-Mol. Biol. 55, 2281. [25] Lewis, L. and Sommer, A. J. (1999) Appl. Spectrosc. 53, 375. [26] Mirabella, F. M., Jr. (1987) Applications of microscopic Fourier transform infrared spectrophotometry sampling techniques for the analysis of polymer systems. In The Design, Sample Handling and Applications of Infrared Microscopes (P. B. Rousch, ed.), American Society for Testing and Materials, Philadelphia, PA, pp. 74–83. [27] For example, an application of a concept integrating optical microscopy and spectroscopy for microspectroscopic analyses can be found in Norman, M. L., Gagnon, A. M., Reffner, J. A., Schiering, D. W. and Allen, J. D. (2004) Proc. SPIE 5269, 143. [28] Sommer, A. J. and Katon, J. E. (1991) Appl. Spectrosc. 45, 1663. [29] Messerschmidt, R. G. (1987) Photometric considerations on the design, sample handling and applications of infrared microscopes. In The Design, Sample Handling and Applications of Infrared Microscopes (P. B. Rousch, ed.), American Society for Testing and Materials, Philadelphia, PA, pp. 12–26. [30] Sahlin, J. J. and Peppas, N. A. (1997) J. Appl. Polym. Sci. 63, 103. [31] Born, M. and Wolf, E. (1980) Principles of Optics, 6th edn, Pergamon Press, Elmsford, New York. [32] Heimann, P. A. and Urstadt, R. (1990) Appl. Opt. 29, 495.

24

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

[33] Treado, P. J., Levin, I. W. and Lewis, E. N. (1994) Appl. Spectrosc. 48, 607; Lewis, E. N. and Levin, I. W. (1995) Appl. Spectrosc. 49, 672. [34] Bennett, C. L., Carter, M., Fields, D. and Hemandez, J. (1937) Imaging Fourier transform spectrometer. In Proceedings of the Imaging Spectrometry of the Terrestrial Environment, 14–15 April 1993, Orlando, FL, SPIE 191. [35] Dereniak, E. L. and Boreman, G. D. (1996) Infrared Detectors and Systems, John Wiley & Sons, New York. [36] Rogalski, A. (2002) Infrared Phys. Technol. 43, 187. [37] Bhargava, R., Fernandez, D. C., Schaeberle, M. D. and Levin, I. W. (2000) Appl. Spectrosc. 54, 1743. [38] The scattering contribution of the image results in a baseline offset in absorbance spectra, as detailed in Bhargava, R., Wang, S. Q. and Koenig, J. L. (1998) Appl. Spectrosc. 52, 323, which must be corrected for quantitative spectroscopy. [39] Snively, C. M., Katzenberger, S., Oskarsdottir, G. and Lauterbach, J. (1999) Opt. Lett. 24, 1841. [40] Huffman, S. W., Bhargava, R. and Levin, I. W. (2002) Appl. Spectrosc. 56, 965. [41] Snively, C. M. and Koenig, J. L. (1999) Appl. Spectrosc. 53, 170. [42] Bhargava, R. and Levin, I. W. (2001) Anal Chem. 73, 5157. [43] Bhargava, R., Schaeberle, M. D., Fernandez, D. C. and Levin, I. W. (2001) Appl. Spectrosc. 55, 1079. [44] Hirschfeld, T. (1979) Appl. Spectrosc. 33, 525. [45] Bhargava, R., Fernandez, D. C., Schaeberle, M. D. and Levin, I. W. (2001) Appl. Spectrosc. 55, 1580. [46] Bhargava, R. and Levin, I. W. (2002) Anal. Chem. 74, 1429. [47] Bhargava, R., Ribar, T. and Koenig, J. L. (1999) Appl. Spectrosc. 53, 1313. [48] Bhargava, R., Wang, S. Q. and Koenig, J. L. (2000) Appl. Spectrosc. 54, 486. [49] Bhargava, R., Wang, S. Q. and Koenig, J. L. (2000) Appl. Spectrosc. 54, 1690. [50] Rogalski, A. (2003) Prog. Quant. Electron. 27, 59. [51] Lewis, E. N., Gorbach, A. M., Marcott, C. and Levin, I. W. (1996) Appl. Spectrosc. 50, 263. [52] Lewis, E. N., Treado, P. J., Reeder, R. C. et al. (1995) Anal. Chem. 67, 3377. [53] Carter, M., Bennett, C. L., Fields, D. J. and Lee, F. (1995) Proc. SPIE 2480, 380. [54] Lewis, E. N., Kidder L.H., Arens, J. F., Peck, M. C. and Levin, I. W. 1997 Appl. Spectrosc. 51, 563. [55] Kidder, L. H., Levin, I. W., Lewis, E. N., Kleiman, V. D. and Heilweil, E. J. (1997) Opt. Lett. 22, 742. [56] Colarusso, P., Kidder, L. H., Levin, I. W., Fraser, J. C., Arens, J. F. and Lewis, E. N. (1998) Appl. Spectrosc. 52, 106A. [57] Bhargava, R. and Levin, I. W. (2003) Appl. Spectrosc. 57, 357. [58] Bhargava, R. and Levin, I. W. (2003) Macromolecules 36, 92. [59] Sommer, A. J., Tisinger, L. G., Marcott, C. and Story, G. M. (2001) Appl. Spectrosc. 55, 252. [60] Chan, K. L. A. and Kazarian, S. G. (2003) Appl. Spectrosc. 57, 381; Chan, K. L. A., Hammond, S. V. and Kazarian, S. G. (2003) Anal. Chem. 75, 2140.

2

Near-infrared spectral imaging with focal plane array detectors E. Neil Lewis, Linda H. Kidder, Eunah Lee and Kenneth S. Haber

2.1 Background: single-point near-infrared spectroscopy The analytical near-infrared (NIR) region spans the approximate range of 700–2500 nm, where absorptions arising from overtones and combination bands −H, N− −H and C− −H stretching and bending fundamentals are found. The relof O− atively weak absorptivity of these overtones and combination bands compared with the fundamentals in the mid-infrared (MIR) region is one of the primary reasons that NIR spectroscopy has become a workhorse analytical technique in a variety of fields. In the MIR region, the strong absorptions of the fundamental vibrational bands demands rigorous sampling techniques to limit the amount of material interacting with the incoming radiation. The necessity for thin sectioning, KBr pellet preparation or the use of special sampling accessories, such as ATR, implies that the quality of the MIR data will have at least some dependence on the skill of the person preparing the samples or taking the measurement. In the NIR region, where sample absorptivities are 1–2 orders of magnitude less, intact samples can be characterized using transmittance or diffuse reflectance. In contrast to the MIR where thick samples are often impenetrable, NIR excels at characterizing intact agricultural samples, whole pharmaceutical tablets, powders and other native solid state or liquid samples with little to no sample preparation. The strength of NIR absorption bands is not constant across the spectral range, but rather increases from shorter to longer wavelengths, creating an inverse relationship between penetration depth and wavelength. For shorter NIR wavelengths, where absorptivities are relatively weak, the penetration depth of the NIR radiation is significant. As a rule of thumb, these wavelengths are used when increased penetration depth is required and decreased specificity can be tolerated, such as for protein and moisture determination in grain1,2 and noninvasive in vivo clinical imaging.3 Where greater molecular specificity is more important than penetration depth, however, longer wavelengths that probe the first overtones and combinations are preferred. In addition to being significantly less intense, NIR spectral features tend to be broad and overlapped. Historically, this made band assignments more difficult than in the MIR, and slowed the adoption of NIR relative to this technique. Despite these difficulties, however, as early as 1922–9, researchers at UCLA and Johns Hopkins

26

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

studying the NIR spectra of organic compounds were able to assign spectral bands to functional groups.4 The US Bureau of standards also worked on making band assignments,5 and by 1945 enough information had been gathered to make the technique of interest in the chemical industry. Willis in the ICI Plastics division worked in the NIR region in the mid- to late 1940s, initially with the same type of equipment used in the 1920s – a spot galvanometer with photographic recording.5 In the mid-1950s the Grubb–Parsons grating spectrometer became available, and was used widely for industrial analysis in the NIR spectral region.5 A review by Whetsel highlights NIR spectroscopy during this early period.6 The era of NIR spectroscopy from the 1920s through the 1970s has been described as the ‘classical analysis’ period, where the analytical approach was to monitor the behavior and appearance of specific spectral features assigned to specific functional groups. This approach is familiar to practitioners of MIR imaging and Raman vibrational spectroscopy, and has significant merit and applicability for techniques with sharp and specific spectral features. This analytical methodology however, does not draw on the strengths of NIR spectroscopy, and as such, it did not give the technique any distinct advantages over other analytical approaches, and with the advent of high performance liquid chromatography (HPLC) in the 1970s, NIR spectroscopy was ‘temporarily’ displaced as an industrial analytical tool. At about this time, NIR spectroscopy was first being investigated for agricultural applications in seminal work by Karl Norris.7–9 Initial attempts to quantify agricultural samples were based on the analysis of solvent extractions of sample components, an approach that is obviously destructive, and requires significant operator skill. Quickly, however, the merits of NIR spectroscopy as a direct, nondestructive analytical technique were recognized,10 and the development of ‘correlation-based NIR analysis’,5 the methodology familiar to modern practitioners of the technique, took root and blossomed. As testament to the ease of operability, early quantitative analyses of grain samples collected with crude instrumentation operated by unsophisticated users were able to provide results comparable in accuracy to wetchemical methods. In 1975, the Canadian Grain Commission adopted NIR as an approved method for protein quantification, based on the work by Williams.11 The significant change from classical analysis to correlation-based analysis was enabled by a confluence of interrelated advances in hardware, software and applications knowledge. An iterative cycle of instrumental and computational advances, development of sophisticated chemometric algorithms and an increased understanding of applications particularly well suited to the technique worked in concert to further the applicability of the method in the agricultural area. The credibility demonstrated by the establishment of NIR analyses in the agricultural field, with continuing development in instrumentation, software and chemometric approaches, aided the rapid proliferation of NIR spectrochemical analyses to other fields of study. As an example, applications of NIR spectroscopy have expanded rapidly in the pharmaceutical industry in the last decade and it is now considered a standard analytical technique in the industry.12 Additional application areas include fine chemicals and chemical production, food and beverages, textiles, polymer science, biotechnology,

NIR SPECTRAL IMAGING WITH FPA DETECTORS

27

earth sciences and mineralogy, medicinal and clinical chemistry, petroleum and fuels research, environmental sciences, and paper and pulp. Workman has compiled an extensive review of NIR applications in these areas.13 In summary, NIR spectroscopy is currently a rugged, laboratory and process technique, delivering realtime analyses across a variety of application areas. The limited sample preparation requirements and the development of automated and rapid mathematical analysis enhance the quality of spectrochemical analyses and alleviate the need for analysis by highly trained personnel, making NIR spectroscopy an extremely useful tool.

2.2 Development of NIR spectral imaging In the last 16 years spectral imaging has evolved from a specialized analytical technique employing the point by-point synthesis of spectral maps to a mix of technologies in routine use in industrial, academic and government settings.14–19 Paralleling the development of single-point NIR spectroscopy as described above, imaging instrumentation has become more economically accessible because of iterative advances in hardware components, computational capabilities, electronics and chemometric algorithms. With increased accessibility and capability, the utility of imaging approaches to specific types of sample characterization has become more apparent and potential targets for imaging studies are more easily classified as such.

2.2.1 History of spectral imaging Collecting spatially resolved spectral information has evolved from sequential point mapping approaches to global imaging, and this evolution has been enabled by the availability of two-dimensional (2D) detectors and optical quality wavelength filters. One of the first 2D detectors available to analytical scientists was the CCD, with sensitivity between 180 and 1100 nm. Although developed at about the same time as the CCD, the infrared (IR) sensitive analog, the focal plane array (FPA) detector was not commercially available until much later because its development was driven primarily by strategic military purposes and therefore classified. Even before 2D detectors had been developed though, IR spectroscopists were looking for means to collect spatially localized spectra. In 1945, Barer et al. published results from the first IR microscope.20 The next significant step toward collecting IR spectral images or maps took place over forty years later in 1988 when Harthcock and Atkin published a description of a sequential point mapping approach using a traditional single-detector element, and scanning the sample point by point.21 The development of IR microscopy and the advantages of multichannel detectors is reviewed further in Chapter 1. The introduction of 2D detectors, starting with CCDs, into the scientific sector enabled the collection of spatially resolved spectral information simultaneously rather than via this point-by-point approach. The earliest work employed silicon

28

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

CCDs with spectral response to 1100 nm, either coupled with narrowband interference filters providing multispectral capability22 or with optical quality wavelength filters to perform hyperspectral imaging.23 This latter implementation employed a silicon CCD and acousto-optic tunable filter (AOTF) to record NIR spectra and images in the visible and short wavelength NIR region, from 800 to 1100 nm. The wavelength range was subsequently extended to 1900 nm using a monochromator and NIR sensitive camera.24 The first laboratory-based NIR spectral imaging instrument to encompass the entire NIR spectral range (from 1000 to 2500 nm) coupled a 128 × 128 indium antimonide (InSb) FPA to an AOTF.25 Laboratory-based MIR imaging implementations employing tunable dielectric filters and step-scan interferometers as wavelength selection devices were developed at about the same time.26,27 The first commercially viable, laboratory-based Fourier transform infrared (FTIR) spectral imaging measurements were performed with a 128 × 128 pixel InSb detector optically windowed from 3 to 5 μm and a step-scan FTIR spectrometer.28 Within a year of the publication of the first experimental results using a step-scan interferometer, a commercial unit was available (Bio-Rad Stingray, now Digilab/Varian). As with single-point NIR spectroscopy, NIR spectral imaging has been commercialized relatively slowly, despite being demonstrated earlier.25 Perhaps because NIR spectroscopy was traditionally utilized more for bulk identification or as a process stream monitor, the potential applications for an imaging implementation were not immediately apparent. Its utility did not go entirely unnoticed though, and several government and academic researchers were actively engaged in the development of NIR spectral imaging instrumentation.29–31

2.2.2 FPAs – specifications There are a variety of FPA detectors available that are sensitive in the NIR spectral region. The optimal choice of detectors depends on several factors: desired wavelength range, whether the application will be laboratory based or part of a process environment, the sensitivity needed to adequately differentiate sample spectra and price. The figure of merit most often used to describe detector performance is specific detectivity or D ∗ , which is the inverse of noise equivalent power (NEP), normalized for detector area and unit bandwidth. NEP is defined as the radiant power that produces a signal-to-dark-current noise ratio of unity. Indium gallium arsenide (InGaAs) arrays have a combination of attributes that make them very desirable for short wavelength applications. Detectors based on these arrays have the highest detectivity (D ∗ = ∼6 × 1012 cm Hz0.5 W−1 ) and operating temperature of available arrays. They require only modest thermoelectric cooling, making them particularly well suited for on-line or at-line settings as they do not require liquid nitrogen or other more expensive cooling implementations. These characteristics make them some of the least expensive detectors operative in this wavelength range. Arrays in the 320×256 format are abundant and economical, and even relatively large arrays (640 × 512) are readily available. There is no perfect solution, of course, and the wavelength range across which standard InGaAs is

NIR SPECTRAL IMAGING WITH FPA DETECTORS

29

responsive does not cover the entire NIR spectral range, but only ranges from ∼0.9 to 1.7 μm. There have been efforts to develop an extended range InGaAs array, but these prototypes have not yet demonstrated sufficiently satisfactory performance for spectral imaging applications. Indium antimonide (InSb) array detectors extend to cover the long wavelength NIR spectral range, and are sensitive from ∼1.0 to 5.4 μm wavelength range. Although InSb detectors have higher detectivity (D ∗ > 4×1011 cm Hz0.5 W−1 ) than platinum–silicide (PtSi), lead-sulfide (PbS) or mercury–cadmium–telluride (MCT) arrays, they must be cooled to liquid nitrogen temperatures or below. Stirling coolers can be used on InSb cameras for applications where liquid nitrogen cooling would be inconvenient. Because their sensitivity increases out to 5 μm, use in the short wave IR requires the use of an optical filter that blocks the longer wavelength light, a so-called cold filter. Large arrays, up to 1024 × 1024, have been fabricated, but are prohibitively expensive for commercial implementations. The bandgap and therefore the wavelength response of MCT detectors can be tuned by adjusting the ratio of mercury vacancy sites to cadmium donor sites according to the following stoichiometry – Hg1−x Cdx Te. A common doping scheme results in a detector that is responsive between ∼0.8 and 2.5 μm, with some of the highest detectivities for MCT achieved in this spectral range (D ∗ ∼ = 3 × 1011 cm Hz0.5 W−1 ). Cooling requirements are not as severe as for InSb arrays, and the shorter wave cutoff makes a cold filter unnecessary. Large format arrays can be made from this material. PtSi arrays are sensitive over the 1.0–5.0 μm wavelength range. These are silicon-based arrays in which the platinum–silicon junction in each pixel forms a Schottky barrier diode. Large, exceptionally uniform arrays can be made from this material, but the quantum efficiency is <1% (D ∗ = 3 × 1010 cm Hz0.5 W−1 ), making such arrays unsuitable for low light level applications. Photoconductive lead sulfide (PbS) arrays with 320 × 240 pixels are also available for operation in the ∼1.0–3.0 μm range, with reported D ∗ values up to ∼3 × 1011 cm Hz0.5 W−1 if cooled to 200 K.

2.2.3 Implementation of NIR imaging Most NIR imaging instruments are based on one of the following wavelength filtering options: tunable filters, dispersive spectrometers or interferometers. Tunable filters used for NIR imaging include liquid crystal tunable filters (LCTFs), AOTFs, Fabry–Perot tunable filters, linear and circularly variable filters and filter wheels. Additionally, NIR imaging instruments can be categorized by whether they rely on filtering either the optical source or the image. Commercial, interferometer-based mapping and imaging systems exclusively employ source filtering, in which the radiation from the source is modulated by the interferometer before impinging on the sample. Tunable-filter-based systems however, typically illuminate the sample with broadband radiation, and perform the filtering on the image before it is focused onto the detector. While it is theoretically possible to perform image filtering using an interferometer, the optical design needed to recombine the image

30

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

without introducing artifacts would add significant complexity. Similarly, tunable filters could be coupled with the appropriate lamps to construct a tunable source for illumination of a sample. Image filtering employed by tunable filter systems, however, provide a significant amount of flexibility in terms of the field of view (FOV) that can be illuminated. Acousto-optical tunable filters use Bragg diffraction in piezoelectric crystals under variable radiofrequency (rf) excitation to pass light of the desired wavelength. AOTF filters have fast response times, high efficiency (although they are polarization specific) and function throughout the short-wave infrared.32 Since they rely on wavelength-dependent angular dispersion of the spectral components, however, there must be a relatively large distance between the filter and the detector (or between the lamps and the sample if the filter is used in source filtering mode), complicating the optical design. Further, such filters may introduce astigmatism into the image, as their optical resolution along the two spatial axes is different.33 Filter wheels holding a selection of bandpass filters can be used for spectral imaging.22,34 Such a solution is relatively inflexible, and while it is possible to tune the bandpass of a dielectric filter by tilting it, this will also change the width of the bandpass and may shift the image on the detector. Without tilting the filters, however, one is left with only the limited wavelength range corresponding to the filter selection. Furthermore, tuning speed will be limited by the mechanical operation of translating the filters in and out of the optical path. Typical implementations of filter-based systems are used for remote sensing applications.35 Circular variable filters, in which the bandpass center wavelength varies with filter rotation, have also been used for spectral imaging in the NIR.26,36 In another scheme, linear variable filters have been used in a hybrid imaging mode. Image acquisition using this approach requires that the sample be translated relative to the filter (or vice versa), and the resulting sequence of images must then be reorganized to separate spatial and spectral information. This scheme precludes random access wavelength coverage, since the entire FOV must be translated past the portion of the filter corresponding to the desired wavelengths. However, when the linear variable filter is directly bonded to the FPA itself, it produces the simplest and most compact imaging system possible, which may make it potentially usefully for process monitoring.37 Fabry–Perot filters have been used for high spectral resolution imaging over a narrow wavelength range in the NIR.38 This type of device can be used for astronomical imaging but is generally not suitable for NIR chemical imaging, where sample absorption bands have intrinsic line widths larger than the tuning range of such filters. The most widely used tunable filter for commercial NIR chemical imaging systems is the LCTF. An LCTF is a Lyot filter in which optical retardation is electronically tunable.39 Filters are now available with a wavelength range spanning 1000–2450 nm and a bandpass of 6 nm at 1600 nm. The tuning time between discrete wavelengths is ∼50 ms, independent of the size of the wavelength step. The two major advantages of such filters are truly random wavelength access and

NIR SPECTRAL IMAGING WITH FPA DETECTORS

31

excellent image quality. The main drawback is that they limit the numerical aperture of the detection system, and have less than ideal transmission characteristics over their wavelength tuning range, from <5% to ∼30%. Like the AOTF, these filters are also polarization selective. Spectral imaging using dispersive spectrometers proceeds by scanning a sample past an optical system that images a line perpendicular to the scanning direction onto the entrance slit of an imaging spectrograph.24,40,41 Spectra from each point on the line are dispersed across an array, and the image cube is built up one y-λ plane at a time. As in the case of LVF imaging described earlier, random wavelength access is not possible, and the sample must be scanned in its entirety across the slit in order to obtain the full three-dimensional (3D) image cube. Finally, NIR imaging can be performed using a Michelson interferometer and an imaging array. An image is acquired at each step of the moving mirror as it translates through the normal sequence of positions, and then the spectrum of each point on the sample is recovered via a Fourier transform. For high spectral resolution images this is the method of choice. However, if the desired information is contained in only a few wavelengths, the interferometer is at a disadvantage, since the mirror must be scanned over its full range of motion to obtain a spectrum. An alternate approach to imaging that does not require the use of a 2D detector employs a Hadamard encoding mask. Implementations incorporate an interferometer, although a digital mirror array coupled to a dispersive grating has also been developed.42–44 We can conclude that for small samples requiring large wavelength extent, high spectral resolution and which are in a laboratory environment, interferometric imaging is an attractive option. Where speed is important, or for larger samples, or those which can be chemically characterized by images at a relatively small number of wavelengths or for samples in an environment hostile to precision motion, tunable filter imaging is preferable. Because of its imaging quality, lack of moving parts, random access tuning capability and optical flexibility the LCTF is currently the preferred filter for this mode of imaging.

2.2.4 Data processing A spectroscopic image at a single wavelength is a convolution of spectral absorbance, scattering, diffraction effects, instrument line shape, etc. Although looking at an uncorrected image at a single wavelength can yield a chemical image that highlights the spatial distribution of sample components, other optical effects can have a confounding effect. Just as spectral preprocessing (dividing through by a background dataset, baseline subtraction, multiplicative scatter correction, derivatives, etc.) is critically important for the correct interpretation of single-point spectra, it is equivalently necessary when processing NIR spectroscopic imaging datasets. For implementations that rely upon d.c. signal detection, such as the LCTF NIR imaging system, subtraction of the so-called dark response is critical. There are two components of this response, one that is wavelength independent, the detector response when no photons are impinging on the array, a characteristic of the

32

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

array and its associated electronics, and the second that is wavelength dependent, due to a residual component resulting from ambient and scattered light impinging on the array. The second component appears very similar to the single-beam background response of the instrument, but is significantly weaker. The effect of this dark response is only significant in low-light level experiments or in situations where the instrument is being operated at the extremes of its wavelength range and the signal from the sample is small. Careful correction to account for this effect is very important to yield correct spectral profiles. This contribution can be corrected by subtracting it from both the sample and background data, typically by collecting a separate data cube that has the same experimental parameters used for the sample, but with nothing in the FOV of the system. Recently, an implementation has been designed that enables the sample, dark and background data to be collected in an interleaved manner by mounting very thin ceramic and mirrored disks in a holder that automatically puts them into the FOV of the experiment when needed.45 In the LCTF-based NIR system the illumination comes from sources mounted at oblique angles to the sample, and a mirrored surface acts as a blank from which a dark cube can be collected. In this configuration, data collection is totally automatic, and sample, background and dark reference data are collected automatically at each wavelength, without operator intervention. Because the reference and sample data are collected close in time, it effectively modulates the data collection, and instrument performance is enhanced relative to data collected in a noninterleaved fashion. It could be a particularly important innovation for process measurements for several reasons, background data cannot become separated from sample data or be collected under difference experimental conditions, it speeds up the acquisition time by reducing the number of wavelength tuning steps by a factor of 3, it makes the idea of random wavelength access a much more viable option since ‘selected wavelengths’ are measured in self-contained triplets and are therefore not dependent on old background data. It adds the capability of automatically inserting reference materials without accessing the instrument. Additionally, other materials can be added to the holder for automatic wavelength or linearity calibration. Once a dataset is appropriately preprocessed, single wavelength images can yield highly specific chemical images, showing the distribution of chemical species in a sample. This obviously works particularly well for samples that contain species that are spectrally very distinct. As highly overlapping bands from relatively weak overtones and combination bands predominates in the NIR spectral region, advanced multivariate techniques are often required to clarify subtle spectral differences. Additionally, it is a massive underutilization of the data to focus on spectral and spatial information separately. The most powerful analytical approaches take into consideration both spectral and spatial information, and provide solutions to problems that cannot be solved with single-point spectroscopy. Robust multivariate classification can require large numbers of samples in order to characterize the covariance of the component distributions. One of the major advantages of spectral imaging over single-point spectroscopy is the ready

NIR SPECTRAL IMAGING WITH FPA DETECTORS

33

availability of a statistically meaningful ensemble of spectra in the former. A single NIR spectral image usually contains more than 80 000 spectra, so it becomes trivial to accumulate vast libraries of pure component spectra, enabling the application of a number of powerful multivariate classification techniques. In addition to standard statistical pattern recognition methods, such as Mahalanobis distance, linear discriminant analysis, N-nearest neighbor classifiers, etc., regression methods such as partial least squares, principal component regression, and multilinear regression can be trained on pure component spectra and used to generate classification score images. More details on chemometrics and data processing algorithms are reviewed in Chapter 5 of this book.

2.2.5 Comparison of vibrational spectroscopic imaging modalities The three complementary vibrational spectroscopy techniques, MIR, NIR and Raman, all have commercial mapping and imaging implementations. Mapping implementations employ either single-point detectors or, in a relatively new approach, linear array detectors, in which the sample is translated along at least one axis and usually both spatial axes to build up spatial information.46 Global imaging implementations acquire 2D image data simultaneously. As with their single-point counterparts, each technique has particular strengths and capabilities. A comparison of MIR imaging approaches is discussed in Chapter 1, the emphasis here is on the differences between MIR and other vibrational spectroscopic approaches to imaging. The reflective Cassegrainian optics required to prevent chromatic aberration across the broad spectral range of the MIR provide limited FOV flexibility,47 and there is a tradeoff between image fidelity and time spent to collect the data. As previously mentioned, FTIR imaging is a source filtering implementation, and since the light source for this implementation is the interferometer, currently used common source configurations do limit the FOV. The most commonly accessible and affordable arrays that are sensitive across the MIR spectral range tend to be small format, and in some cases, are not as robust as other arrays types. Both MIR mapping and imaging operating with the longest wavelength have the lowest spatial resolution due to diffraction effects, although ATR imaging using high refractive index materials can significantly improve on this.48 These factors suggest that typical MIR imaging systems have the lowest image fidelity of the three imaging techniques. MIR imaging has also not shown itself to be particularly successful in diffuse reflectance applications; most published examples are transmittance examples, which often require sample preparation, such as thin sectioning. Raman mapping and imaging systems have extremely high achievable spatial resolution coupled with high chemical specificity. Raman mapping techniques typically provide high spectral resolution, but suffer from the same speed issues of mapping techniques in general. Global Raman imaging techniques allow random wavelength access, and if a complete high quality Raman spectrum at each pixel is not necessary, this can be a significant time saver. Global Raman images from these kinds of samples can rapidly produce high quality images with high spatial

34

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

resolution. However, both Raman imaging and mapping suffer from inherently low intensities of the Raman scattered signal and interference from fluorescence. Raman imaging suffers from some additional limitations that are not shared by MIR, NIR or Raman mapping. Because the illumination source is a laser and is power limited, going to larger spot sizes decreases the available power density by a function of the area to be explored. Changing the spot size from 10 × 10 μm2 to 100 × 100 μm2 requires 100 times the laser power to achieve equivalent Raman scattering intensities. For interrogation of large FOVs, this rapidly becomes a limiting factor. Also, because of the weak signal, high NA objectives are used to collect light most efficiently. These two issues make it generally difficult to collect Raman spectra in global imaging mode across large sample areas. Also, the high NA objectives have very small depth of focus, and surface roughness across the viewing area can affect the sample focus and therefore usable Raman signal.49 Raman mapping systems can employ autofocusing systems to overcome this limitation. As with other spectroscopic techniques, NIR mapping and imaging inherit attributes and issues of the corresponding single-point spectroscopy. For example, both mapping and imaging require minimal sample preparation and generate high signal-to-noise ratio data from a variety of samples. NIR spectra, however, provide less chemical specificity than either MIR or Raman spectroscopy, and as a result it works best with classes of samples that are well characterized and differentiable with single-point NIR. Common commercially available NIR mapping implementations are exclusively FT-based, and therefore benefit from high spectral resolution. Due to the broad spectral features, however, high spectral resolution often is not necessary to perform chemical characterization in the NIR. High spectral resolution may instead result in excessively large data files, especially when considering the larger format of NIR array detectors. Contributing to extremely large data files for FT systems is the fact that the multiplexed signal (interferogram) is always recorded, even if only a few wavelengths in the broad wavelength range are desired. The undersampling schemes commonly used for MIR spectral imaging to limit the size of datasets, using optical filters to eliminate source radiation outside the computational bandpass, cannot be employed with the short wavelengths of the NIR. Despite the fact that the LCTF has some limitations in optical throughput relative to the FT spectrometer, this approach to global NIR imaging enables tremendous flexibility in setting the wavelength range of the experiment; any combination of wavelength ranges and/or single wavelengths is allowed and readily controlled by the data acquisition software. The bandpass of the LCTF filter is generally well suited to the nominal linewidths of a typical NIR absorption band. A detection scheme incorporating LCTFs also benefits from the large format arrays available in the NIR. The combination of tunable filter and 2D array detector can produce the best image fidelity of all the IR mapping and imaging systems currently available. In addition, because of abundant signal and bright optical sources, refractive optics with relatively low numerical apertures enable large FOVs to be easily implemented, and curved or rough samples accommodated.

NIR SPECTRAL IMAGING WITH FPA DETECTORS

35

2.2.6 Safety in numbers Spectrochemical imaging is making the transition from producing ‘pretty pictures’ based on the spectral properties of the sample to solving problems, albeit specialized ones, in the analytical/R&D laboratory. Enabling this transition is the understanding among practitioners of what applications are well suited to spectroscopic imaging analysis, and what applications are better left to traditional bulk spectroscopic or microspectroscopic analyses. Spectral imaging techniques enable the simultaneous collection of tens of thousands of spatially distinct vibrational spectra, and provide information on what chemical species are present, how much of each is present, and what is the distribution of the components across the sample. This information all contributes to provide a more precise understanding of the many characteristics that ultimately determine attributes of heterogeneous samples. An interesting, and sometimes overlooked, characteristic of these kinds of datasets is its statistical properties. In many cases a chemical imaging experiment returns data that contains heterogeneity in terms of spatial organization and can be interpreted as a picture. However, heterogeneity is not always spatially organized, and the spectra can be viewed simply as chemical distributions or populations. This representation has significant implications and provides tremendous analytical potential. By simply analyzing the frequency distribution of a single absorption band as an intensity histogram, one can, in some cases, quickly determine the number of chemical species present in a sample as well as their relative abundance. For example, in a case where the coating on a sample is of interest, a uniform coating may be characterized by a narrow, normal distribution of spectral intensities relating to the chemical composition of the coat, thereby indicating a relatively homogenous coating thickness. Deviations from this type of univariate distribution quantified by statistical skew, kurtosis or simply a large standard deviation would indicate a nonuniform distribution of chemical species and therefore a nonuniform coating thickness. In these instances quantitative data about the spatial distribution of the chemical species in a particular sample is determined without even looking at the image, and reliable data is obtained because of the large number of spectral data points available.18 The example above is obviously relatively simple but one can also analyze data derived from multivariate processing schemes in a similar fashion. In addition, data that has been processed into binary images displaying individual particles can be further analyzed using particle size distribution statistics. Similar statistical concepts can be applied to investigate deviations from a mean particle size, a particle size distribution and/or to determine the presence of several different types of particle populations within a single sample. These types of analyses are truly quantitative assessments of the spectral imaging data, and in the absence of these approaches chemical images can only be interpreted in a qualitative manner. In fact, rigorous mathematical analysis incorporating spectral parameters is the only means by which meaningful data can be extracted from chemical images. Visual examination of the images, while helpful in terms of pattern recognition, provides only

36

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

minimal insight to the underlying numerical data. With the advent of digital data processing systems the concept of visually interpreting conventional spectral data was supplanted by numerical approaches and by analogy, when processed correctly, there is very little about a chemical imaging dataset that is ‘subjective’ and open to interpretation. Just like conventional spectroscopic approaches, mathematical analyses are highly amenable to automation. Another interesting ‘bonus’ of multichannel spectral imaging is that, because the spectra collected are generally spatially independent, very small particles embedded in a sample can be measured and identified without the dilution effects that would be encountered if the same sample were measured using a ‘bulk’ single-channel NIR spectral measurement. In some cases, the question of detection limits becomes dependent on particle statistics rather than weight percentage, implying that NIR imaging can be used to look for components at significantly lower concentrations than is possible with single-point NIR spectroscopy. In theory, a particle as small as a single pixel would result in a measurable and identifiable spectrum. In the case of a standard NIR imaging experiment this would represent a single spectrum in 81 920, overing a 10 μm × 10 μm spatial location. The obvious assumption in this type of thinking is that the component is heterogeneously dispersed, which for many samples, and particularly contaminants, is a reasonable one. If the ‘contaminant’ is homogeneously dispersed, then chemical imaging derives no tactical advantage over the single spectrum approach since each pixel measures the sample at equivalent contamination ‘dilution’. Single-point mapping techniques can also be used to look for spatially localized impurities with equivalent sensitivity, but they do not have the ability to perform this measurement rapidly. Using the above strategy it is possible to perform high speed screening for impurities within a single sample, or across multiple samples.50 Another valuable and interesting aspect of array-based multichannel detection in spectroscopy is the capability to simultaneously view and spectrally measure both samples and pure components in a single measurement. In other words, numerous samples are arranged in a single FOV such that calibration standards, pure components or reference materials are measured simultaneously with one or more ‘unknowns’. The experimental output can be several chemical images of one or more unknown samples or a series of individual analytical assays. For example, if an imaging system of this type analyzed a multiwell-type sample, some of the wells could contain calibration standards while others might contain unknown samples. A single measurement generates both the calibration curve and the data on the unknown samples. Creating datasets that contain both the sample and the reference materials can in many ways be considered ‘self-calibrating’ and can simplify many aspects of a standard quantitative spectroscopic measurement. First, the dataset is ‘internally’ consistent because the reference and sample data are measured simultaneously by the same instrument – temporal variations in instrument response or differences between instruments can be automatically accounted for. If there is a wavelength inaccuracy in the instrument, this is reflected in all the data, including the calibration data. Further, since the relevant data for any given measurement

NIR SPECTRAL IMAGING WITH FPA DETECTORS

37

always remain together in a single-image data file, the data and reference samples are always processed together in identical ways, and cannot be separated. This helps to minimize data processing errors and inconsistencies and promotes good data housekeeping practices. For subsequent measurements the reference samples may be left in place when new ‘unknown’ samples are introduced, and therefore, a new experiment simultaneously results in new reference spectra. Any changes that affect instrument response or data collection parameters result in those changes also being reflected in the reference sample spectra. These concepts may have significant implications for chemometric research efforts devoted to instrument calibration transfer. In essence, the imaging instrument is converted into the ultimate ratio recording spectrometer where any combination of hundreds of different spectra can be used independently or simultaneously used to correct for instrument response, qualify performance, generate library spectra or produce calibration curves. Many of these concepts are discussed elsewhere.51

2.3 Examples of NIR spectral imaging capabilities The purpose of the following examples is to highlight the unique capabilities of global NIR imaging. Detailed application studies of NIR imaging and mapping can be found in other chapters in this book and other references.18,52–59 Unless otherwise noted, all data was acquired using the Sapphire NIR chemical imaging system (Spectral Dimensions, Inc., Olney, MD) and all data processing was performed with ISys (Spectral Dimensions, Inc.) data processing software specifically designed for handling chemical imaging datasets.

2.3.1 Sample statistics and FOV The statistical information contained within the 81 920 individual NIR spectral that comprise the dataset, as well as the ability to easily change the FOV enables NIR imaging to identify sample components, determine relative abundance of these components, and visualize component distribution across a variety of samples. This example explores NIR imaging results for an OTC pain reliever, Excedrin, which is ∼12 mm in diameter, and comprises three active pharmaceutical ingredients, aspirin, acetaminophen and caffeine. The optics are configured so that each pixel on the array samples 9.7 × 9.7 μm on the tablet. An array with 320 × 256 pixels, therefore, measures an FOV of ∼3.1 × 2.5 mm. This optical configuration can view approximately one quarter of the entire tablet surface. Data were collected over the spectral range 1200–2400 nm, with an increment of 10 nm. At each wavelength, 16 image frames were collected and coadded. To collect the entire sample scan comprising 81 920 unique NIR spectra takes ∼5 min. Background data cubes were collected using Spectralon™ (Labsphere, Inc., North Sutton, NH) as the reference. Additionally, the dark response from the instrument was corrected by subtracting a dark scan from both the sample and background datasets. A dark scan is acquired

38

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

by collecting data with the same collection parameters, but without any sample illumination. As mentioned in Section 2.2.4, careful dark correction is important, especially at the extreme scanning range of the instrument. The resulting image cube was then converted to log10 (1/R). To remove baseline effects, and normalize the variations in NIR signal strength for the various components, a Savistisky–Golay second derivative was performed on all 81 920 spectra, and then the data was mean centered and normalized to unit variance. Figure 2.1 shows representative second-derivative normalized spectra from the sample compared with pure component spectra of acetaminophen, aspirin and caffeine. The single-pixel spectra are from 9.7 × 9.7 μm areas on the surface of the tablet, each with signal to noise greater than 5000 : 1, which is comparable to a standard NIR spectrometer. This comparison between pure component spectra and single-pixel spectra ensures that the assignment of the component pixel spectra from the tablet is accurate. The spectra of these three components show spectral

(a)

(b)

1 mm 1200 1400

1600 1800 2000 Wavelength (nm)

2200

2400

Figure 2.1 (a) The figure shows normalized second-derivative NIR spectra of the three pure components contained within the sample: acetaminophen (solid line), aspirin (dashed line) and caffeine (dash-dot line). Beneath these pure component spectra are single-pixel spectra of the tablet taken from the NIR chemical imaging dataset. Comparison between the pure components and single-pixel spectra confirm the presence of the three active ingredients in the tablet. (b) The figure is a RGB summary image in which the acetaminophen, aspirin and caffeine component distributions are assigned to red, blue and green channels, respectively. The three channels are independently scaled and then overlaid to summarize the spatial distribution of the three components. The image encompasses an area of ∼3.1 × 2.5 mm2 , about 25% of the surface of the tablet sample.

NIR SPECTRAL IMAGING WITH FPA DETECTORS

39

features that uniquely characterize them: an image at 2030 nm will highlight the acetaminophen component, aspirin can be separated by an image at 1630 nm and the caffeine distribution can be visualized at 2250 nm. Because the contrast in a single wavelength image is based on the relative strengths of the spectral signatures at that wavelength, these unique spectral features or ‘marker bands’ present images with chemically relevant contrast. The distinguishing spectral feature at 2030 and 2250 nm are both negative, so multiplying these images by (−1) will make them more intuitive by mapping higher brightness to increasing amounts of the sample component of interest. If these three single-channel marker images are each assigned to a separate color channel, red, blue and green respectively, the spatial distribution of these three components can be summarized in a single RGB image. The image in Fig. 2.1(b) demonstrates this capability. Qualitatively, the casual observer can determine that the caffeine component (shown in green) is the least abundant component in this image, followed by aspirin (in blue) and acetaminophen (in red). Because of sample heterogeneity and the statistics afforded by the massively parallel approach, each pixel in this image can be crudely assigned to belong to one component or another by setting an intensity threshold, thereby providing the user with an approximation of the abundance of each component. The threshold value is normally determined from a histogram of spectral intensities and not from the image itself. Figure 2.2(a) shows a histogram representation of the image at 1630 nm, where the pixel populations are plotted as a function of intensity, the ‘bright’ pixels, those with values above a threshold of 1.46, are assigned to the aspirin component. In this example, the threshold is chosen by visual inspection of the histogram and by selecting pixels corresponding to a single population distribution. While this is only an approximation, in other cases the histograms can be further analyzed using a least-squares fitting procedure to more exactly determine the area under each population. A binary image as shown in Fig. 2.2(b) can be created that visualizes the effect of changing the threshold on the areas of the tablet assigned to aspirin, pixels above the threshold are set to ‘1’ and those below are set to ‘0’. For both the acetaminophen and caffeine components, the threshold values are negative, as the spectral markers that differentiate these components have negative intensity in the second-derivative data. The acetaminophen population in the 2030 nm image is set to a threshold below −2.0 (Fig. 2.2(c)) and the caffeine population as determined with the 2250 nm image is set below −2.1 (Fig. 2.2(d)). Dividing the number of pixels whose threshold lies above (or below for acetaminophen and caffeine) the given values by the total number of pixels in the tablet gives an estimation of component abundance as follows: acetaminophen – 45%, aspirin 13% and caffeine 11%. The reported weight/weight concentrations for these components are 37%, 37%, 10%, respectively. The remainder of the tablet is made of nonactive ingredients (excipients). The estimation overestimates the acetaminophen and caffeine abundance and underestimates the aspirin abundance. One reason for this, as described earlier, is the manner in which the relative population distributions were determined but another is that the FOV incorporates only a quarter of the sample surface. Given the significant

40 (a)

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS 2500

(b)

Numbers of pixels

2000

1500

1000

500

0 0

1

2

(c) 1200

(d) 2500

Numbers of pixels

Numbers of pixels

1000 800 600 400

2000

1500

1000

500

200 0 –3

–2

–1

0

0 –4

–3

–2

–1

0

Figure 2.2 (a) The figure is a histogram representation of the image at 1630 nm, a ‘marker band’ for the aspirin component. Bright pixels, those with values above a threshold of 1.46 are assigned to the aspirin component. This threshold is chosen by selecting pixels approximately corresponding to a distinct population within the histogram, and this population is marked with black bars in the histogram. (b) The figure is a binary image in which pixels above this threshold are set to ‘1’ and those below are set to ‘0’. This image clearly highlights the areas in the tablet that are being assigned to aspirin based on the chosen threshold. (c) and (d) The figures show the histogram of the image planes at 2030 nm, highlighting the acetaminophen distribution, and 2250 nm, corresponding to caffeine. Because the spectral markers that differentiate these components have negative intensity in the second derivative the threshold values are negative: below −2.0 for acetaminophen, and below −2.1 for aspirin. These populations are marked with black bars in the corresponding histograms.

sample heterogeneities, and large particles apparent in this image, the whole sample should be considered. By changing the optics to obtain ∼38 μm × 38 μm per pixel, an FOV ∼12.2 × 9.7 mm can be obtained, which is large enough to sample the entire tablet surface. All other data collection parameters were held constant, so the larger FOV dataset is collected in ∼5 min, which is the same amount of time as the higher magnification data. The same data processing steps were followed, and the resulting normalized second-derivative spectra compared with pure component spectra are presented in Fig. 2.3(a). The RGB image of the whole tablet is shown in Fig. 2.3(b). Looking at the whole tablet, and performing the identical thresholding technique as described above, the quantitative abundance values for the sample components now give estimates of 35%, 40% and 12%, closer to the label values of 38%, 38% and 9%

NIR SPECTRAL IMAGING WITH FPA DETECTORS

41

(a)

(b)

1 mm 1200

1400

1600 1800 2000 Wavelength (nm)

2200

2400

Figure 2.3 (a) The figure shows normalized second derivative NIR spectra of the three pure components contained within the sample: acetaminophen (solid line), aspirin (dashed line) and caffeine (dash-dot line). Beneath these pure component spectra are single pixel spectra of the tablet taken from the NIR chemical imaging dataset. Comparison between the pure components and single pixel spectra confirm the presence of the three active ingredients in the tablet. (b) The figure is a RGB summary image in which the acetaminophen, aspirin and caffeine component distributions are assigned to red, blue and green channels, respectively. The three channels are independently scaled and then overlaid to summarize the spatial distribution of the three components. The image encompasses an area of ∼12.2 × 9.7 mm2 , the entire tablet surface is captured in this FOV.

for acetaminophen, aspirin and caffeine, respectively. The speed with which NIR imaging data can be collected ensures that multiple samples can be examined, and reasonable sample statistics generated. While data collection speed is not the only consideration in the value of a chemical imaging system, the ability to generate data on a statistically significant portion of one or more samples may have significant impact on the accuracy of data interpretation. This becomes particularly true as applications for the technology evolve from R&D to process measurements. Conventional single-point mapping implementations, even fast mapping, do not enable large numbers of samples to be studied in a timely fashion. There is, however, a constraint on the types of samples that will give reasonable results for this abundance estimation approach. For example, in this particular analysis, there is no presumption of spatial organization or sample orientation. A measurement along any axis is assumed to be equivalent, resulting in an equivalent ‘random’ component distribution. For example, this analysis would not deliver sensible results for samples that may have drug cores, coatings or other timerelease mechanisms and other approaches would be required. Even with a randomly

42

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

heterogeneous sample, the concept of what constitutes chemical heterogeneity is implementation specific. To generate reasonable results, sound statistical sampling is critical. For imaging systems, this parameter is influenced by overall system FOV as well as the number of pixels used to image the sample. As an extreme example to illustrate the point, if the FOV of a chemical imaging system is equivalent to the area of single particle in a pharmaceutical tablet, the data could easily indicate that the sample only comprised that single component. And an imaging system that sampled the whole tablet with a large number of pixels would produce a different, more representative result. However, if only a limited number of pixels were used to examine the sample, yet a third result would be gathered. In summary, the spatial extent, magnification and sampling density of an imaging system must be considered when deriving chemical image data from a given sample. These characteristics must be considered relative to anticipated particle size and component distributions within the sample if reliable conclusions are to be derived.

2.3.2 High-throughput applications One of the unique capabilities of NIR global imaging is the ability to perform highthroughput applications, using the parallel capabilities of the technique to evaluate multiple samples simultaneously. For example, we report here a sample containing microspheres in a time-release capsule. The cross-section area of the microspheres is ∼0.5 mm2 , and the FOV (12.8 × 10.2 mm) encompasses 135 of them. Data was collected over the spectral range 1200–2400 nm with a 10 nm data increment. At each wavelength, 16 images were coadded to produce a data cube in ∼5 min of data collection time. The raw data was dark and background corrected and converted to absorbance as described previously. Because this particular dataset includes spatial regions that contain both the sample and empty regions, the first step in preparation toward further preprocessing is to spatially mask data outside of the microspheres themselves. Only spectra from the spheres are included in further data processing steps. This step is performed through an automated routine in the analysis software. To remove baseline effects, and normalize the variations in NIR signal strength for the various components, a Savistisky–Golay first derivative was performed, and the data mean centered and normalized to unit variance. Figure 2.4 shows a single wavelength image at 1800 nm showing no notable contrast, effectively equivalent to an optical image for this sample. Figure 2.5(a) shows overlaid single-pixel spectra from multiple microspheres, highlighting spectral differences at 2050 nm and 2130 nm, indicating that the spheres are not chemically identical. In this particular instance the two populations are the coated and uncoated versions of the same microsphere. By looking at the single-channel images at these wavelengths in Fig. 2.5(b) and (c), respectively, it is possible to visualize the distribution of these different species. For ease of identification, the bright particles highlighted at 2050 nm are uncoated and those highlighted at 2130 nm are coated. Table 2.1 summarizes the particle statistics for the two types of microspheres. There are approximately twice as many coated particles as there are uncoated, and the

NIR SPECTRAL IMAGING WITH FPA DETECTORS

43

1800 mm 0 2

(mm)

4 6 8 10 0

2

4

6 (mm)

8

10

12

Figure 2.4 The figure shows an image at 1800 nm of microspheres that are the constituents of a timerelease capsule. Because there is no notable spectral difference between the spheres at this wavelength, there is no contrast between the individual spheres. This image is comparable to a visible image of the samples. The FOV (12.8 × 10.2 mm) encompasses 135 of the microspheres.

mean particle diameter for the coated is slightly larger. From the particle statistics it is also possible to conclude that there is slightly more variation is size across the coated microspheres. While the chemical identity of each microsphere could be obtained through 135 successive NIR measurements using single-point spectroscopy, one for each particle, this would be a time-consuming and difficult exercise. With an imaging approach using multichannel detectors, individual sample size or placement is not a significant factor, and multiple samples can be randomly distributed in the same FOV. Even for a sample as complex as this one, both the chemical composition of each individual particle and the statistics describing the entire population can be combined with the size and shape information for each particle, providing extensive charcaterization. These data provide a much more complete description of the sample than could be obtained from conventional analytical approaches commonly employed, for example, HPLC or NIR point spectroscopy.

2.3.3 Statistics, morphology, abundance – using an internal reference We illustrate the utility of chemical imaging next in an example in which neither bulk spectroscopy nor microspectroscopy can provide efficient characterization. It demonstrates the concept of the inclusion of pure components in the same FOV and also includes aspects of the proceeding two examples, namely sampling a large FOV to perform high-throughput analysis of multiple samples in combination with a variety of statistical and morphological approaches. The goal is to demonstrate the variety of analyses that can be performed with a single hyperspectral imaging

44

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS (a)

First derivative data

0.4

0.2

0

–0.2 1200

1400

1600 1800

2000

2200

2400

Wavelength (nm) 2050 mm

2130 mm

2

2

4

4

(mm)

(c) 0

(mm)

(b) 0

6

6

8

8

10

10 0

2

4

6 8 (mm)

10

12

0

2

4

6 (mm)

8

10

12

Figure 2.5 (a) The figure shows first-derivative single-pixel spectra from the microspheres. The spectra clearly indicate differences at 2050 nm and 2130 nm. The single-channel images at these wavelengths, that is, (b) 2050 nm and (c) 2130 nm, respectively, show the spatial distribution of the two types of microspheres. The bright particles highlighted at 2050 nm are uncoated, and those highlighted at 2130 nm are coated. Single-pixel spectra from the uncoated spheres are graphed with solid lines, and the coated spheres are graphed with a dashed line in (a).

Table 2.1 Particle statistics of microsphere samples

Number of particles Mean particle volume (mm3 ) STD particle volume (mm3 ) Mean diameter (mm) STD diameter (mm)

Uncoated microspheres

Coated microspheres

Overall

44 0.185 0.008 0.779 0.275

91 0.221 0.016 0.826 0.348

135 0.209 0.015 0.811 0.337

45

NIR SPECTRAL IMAGING WITH FPA DETECTORS

Inter- and intrasample comparison Chemical distribution images

Statistical analysis Intratablet

Intertablet

Sample heterogeneity Blend uniformity

Morphological analysis

Abundance estimation

Intratablet

Intratablet

Intertablet

Domain statistics Domain size uniformity

Intertablet

Abundance estimation Content uniformity

Figure 2.6 The figure shows the different types of analyses that can be performed on chemical imaging data. The types of analyses that are performed can be grouped into three categories: component abundance estimation, statistical analysis of component distribution, and morphological analysis of discrete particles. All three analyses are used to make inter- and intrasample comparisons, generating abundance and content uniformity estimates, sample heterogeneity and blend uniformity characterization, as well as domain statistics and domain size uniformity data.

dataset, highlighting the ability to characterize both inter- and intrasample distributions of individual sample components. The types of analyses that are performed can be grouped into three categories: component abundance estimation, statistical analysis of component distribution, and morphological analysis of discrete particles. All three analyses are used to make inter- and intrasample comparisons, generating abundance and content uniformity estimates, sample heterogeneity and blend uniformity characterization, as well as domain statistics and domain size uniformity data. Figure 2.6 shows the interrelationship between the analyses and the type of information that is generated. Although the samples used in this example are pharmaceutical tablets, the paradigm will hold across any type of heterogeneous solid state sample. Twenty representative samples from two manufacturers were placed in a matrix, along with three pure component samples, resulting in a sample area of ∼9 × 7 cm2 . The optics were configured so that each of the 81 920 pixels measures an area of ∼275 × 275 μm2 , and data from the entire sample matrix can be collected simultaneously. Data were collected from 1100 to 2450 nm at 10 nm increments with 16 frames coadded at each wavelength and take ∼5 min to acquire. The two sample groups labeled A and B both contain three active pharmaceutical ingredients (APIs): acetaminophen, aspirin and caffeine. The two groups are manufactured by different firms and vary in their reported label concentrations: tablet A group has reported label concentrations of 37%, 37% and 10% for the three components, respectively and tablet B group reports 39%, 39% and 10%. In addition to groups A and B originating from different manufacturers, one group of samples is

46

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

A

B

C

B

A

Aspirin

A

B

C

B

A

Acetaminophen

A

B

C

B

A

Caffeine

A

B

C

B

A

Blank

Figure 2.7 The figure shows the layout of a tablet matrix comprised of tablet types A, B and C, in addition to pure components: acetaminophen, aspirin and caffeine. Each tablet is ∼10 mm in diameter, and the overall sample and pure component matrix covers an area 9 × 7 cm2 .

produced by the innovator, and the other is a generic formulation. Tablet C samples contain only acetaminophen as the active ingredient with a reported label concentration of 79%, and are also included in the array of tablets. The remaining mass of all three tablet-types represents the excipient (binder, disintegrant and lubricant) materials. Pure acetaminophen, aspirin and caffeine samples are also obtained in either tablet form or powder compact and included within the same FOV as embedded reference materials, and measurements on the entire tablet matrix are performed simultaneously. A visible image of the layout of the tablets and pure components is shown in Fig. 2.7. The raw images were corrected for instrument dark response by subtracting a dark cube from both sample and background data cubes, and converting to log10 (1/R). Background data cubes were collected using Spectralon™ (Labsphere, Inc., North Sutton, NH) as the reference. A Savistisky–Golay second derivative was performed as a preprocessing step to minimize contrast caused by baseline offset and slope variations, and then the spectra were mean centered and normalized to unit variance. Data collected from the pure component materials in the same FOV were used to create a reference library of pure component spectra. Using the reference library, a partial least square 2 (PLS2) model was constructed and PLS scores images for each of the library components were generated. Figure 2.8 shows the individual PLS scores images for each component, acetaminophen (Fig. 2.8(a)), aspirin (Fig. 2.8(b)) and caffeine (Fig. 2.8(c)). The channels are scaled independently, and only represent qualitative information about the component distributions. Qualitatively, it is apparent that tablet groups A and B differ. Group A tablets appear to be more homogeneous, with less distinct domains than for group B tablets. Group C samples with only one API and the pure component samples appear extremely homogeneous, which is also to be expected. Although many qualitative conclusions can be drawn from looking at the image alone, it is also possible and desirable to glean quantitative information using a variety of statistical a morphological data analysis tools.

47

NIR SPECTRAL IMAGING WITH FPA DETECTORS

1

1

2

2

3

3

(cm)

(b) 0

(cm)

(a) 0

4

4

5

5

6

6

7

7 0

1

2

3

4 5 (cm)

6

7

8

0

1

2

3

4 5 (cm)

6

7

8

4 5 (cm)

6

7

8

(c) 0 1

(cm)

2 3 4 5 6 7 0

1

2

3

Figure 2.8 (a), (b) and (c) The figures shows the PLS score images for acetaminophen, aspirin, and caffeine, respectively. Each image is scaled independently, and only represents qualitative information about the component distributions.

The result of the PLS calculation is a quantitative, pixel-by-pixel ‘score’ for each of the three components, and statistical information is readily available from these scores. As mentioned above, for sample group A, the sum of the reported weight percents for the three active components is 84%, and for sample group B it is 88%. At each pixel, the three scores, acetaminophen, aspirin and caffeine are summed, and then each component is divided by that sum. By multiplying this result by 0.84 or 0.88 respectively, the normalized mean PLS score for each component in sample groups A and B is obtained. This demonstrates a high-throughput implementation of the abundance example that was previously presented in Section 2.3.1, in which 16 times more samples and additionally pure components are examined in the same amount of time. The results for each sample group are summarized, and presented in Table 2.2. The results show that group A abundance estimates more closely track the values reported on the label, whereas the group B estimations are further from reported values. Because the individual tablets comprise a single dataset, this difference is not due to experimental or data processing anomalies, as the data were

48

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS Table 2.2 Percent abundance estimations for group A and B samples

Acetaminophen Aspirin Caffeine

Sample group A

Sample group B

38 ± 2.7 36 ± 2.3 11 ± 1.0

41 ± 4.8 30 ± 6.0 16 ± 1.6

collected simultaneously and processed identically. It is likely that a significantly more ‘granular’ product (as shown for group B in Figure 2.8) has variations in the surface stoichiometry statistics for any given unit dose may result in poor statistical sampling of the tablet as a whole and an apparent wider variation in potency. A wider range in potency may be real, and not just a statistical sampling effect since it is not possible for a component to be accurately represented within a particular concentration range, in a fixed volume, if the smallest unit (particle) of that component is a sizable fraction of the desired concentration. In other words the more granular the product the more likely it is to have real variations in the mixture, and therefore potency variations from tablet to tablet. Looking at the distribution of individual components in Table 2.2, sample group A tablets appear to be significantly more uniform for aspirin and acetaminophen. The group B tablets have significantly broader distribution of abundance values for these two components, even when normalized to the higher abundance values. The caffeine component though, when normalized, has approximately the same distribution of abundance values for groups A and B. As was demonstrated in the sample statistics example presented in Section 2.3.1 the distribution of pixel intensities in an image can be graphically represented as a histogram, a plot of the number of pixels versus intensity, to provide quantitative statistical information. The normalized standard deviation (the standard deviation divided by mean) for the PLS score images for the three components may be used as a measure of sample heterogeneity. A large normalized standard deviation in a component score image is caused by pixel brightness values that vary significantly across the image, a heterogeneous distribution. If, however, a component is homogeneously distributed, the score values would not vary much on a pixel-to-pixel basis, that is, each pixel would have a similar brightness because each pixel would contain a similar amount of the component. There is enough data density within single tablets (∼1100 individual NIR spectra per tablet) to treat each as a separate sample, and generate sample heterogeneity information for each of the A and B group tablets. These results are presented by component for each sample in Table 2.3, showing the between sample variations for each component. Summarizing the information across tablet groups enables the overall blend uniformity between samples A and B to be compared, and these results are presented in Table 2.4. The results show that the acetaminophen and aspirin components for group A are relatively homogeneous when compared with the caffeine values for these tablets.

49

NIR SPECTRAL IMAGING WITH FPA DETECTORS

Table 2.3 Individual normalized standard deviation values of histogram distributions from PLS score images: a measure of sample heterogeneity for individual group A and B samples

Sample A

Sample B

Sample B

Sample A

Acet.

Aspr.

Caff.

Acet.

Aspr.

Caff.

Acet.

Aspr.

Caff.

Acet.

Aspr.

Caff.

0.30 0.33 0.31 0.33

0.25 0.23 0.25 0.23

0.77 0.72 0.70 0.75

0.25 0.39 0.40 0.37

0.72 0.53 0.51 0.58

0.56 0.70 0.70 0.66

0.45 0.31 0.41 0.38

0.63 0.69 0.60 0.59

0.75 0.60 0.70 0.68

0.36 0.40 0.39 0.29

0.29 0.36 0.31 0.30

0.97 0.82 0.76 0.76

Table 2.4 Summarized normalized standard deviation values of histogram distributions of PLS score images: a measure of sample heterogeneity across group A and B samples

Acetaminophen Aspirin Caffeine

Sample group A

Sample group B

0.34 ± 0.041 0.28 ± 0.046 0.78 ± 0.084

0.37 ± 0.063 0.61 ± 0.072 0.67 ± 0.063

For group B samples, the acetaminophen component is relatively homogeneously distributed, but the aspirin and caffeine components are much more heterogeneously distributed. This correlates with what can be visualized in the image of the samples in Fig. 2.8, but provides quantitative and objective criteria for determining sample heterogeneity. Finally, in addition to abundance estimates and statistical inter- and intratablet comparisons, it is also possible to glean morphological and quantitative particle statistics information for the tablets. Before particle statistic information can be obtained, it is first necessary to determine the extent of the particles for the separate components. This is done by using the PLS score images for each component, and connecting pixels that have the same PLS score or brightness. This is performed in an automated fashion in the data analysis software and appears as a contour in the image, much the same way that height is represented on geological survey maps. The extent of the particles is changed by changing the PLS score value that is chosen for drawing the contour, and this can be automated by determining an appropriate threshold. Not all sample components can be analyzed in this manner. If a component is very well mixed with particle sizes at or below the spatial resolution of the data that component will not appear as discrete particles. For group A samples, referring to Table 2.4 both the aspirin and acetaminophen components are extremely homogeneous, and will be excluded from the particle statistics calculation. Additional sample B tablets are also excluded from one or both of these component calculations because the domains appear to be heavily agglomerated, and difficult to classify as discrete particles. As an example of the process Fig. 2.9

50

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

50

Pixels

100 150 200

250 50

100

150

200

250

300

Pixels Figure 2.9 The figure is a binary image in which the caffeine domains created using particle statistics are visualized. Other sample components are set to zero, so only the caffeine domains are apparent.

Table 2.5 Number of domains highlighted for individual group A and B samples

Sample A Acet.

Aspr.

Sample B

Sample B

Sample A

Caff.

Acet.

Aspr.

Caff.

Acet.

Aspr.

Caff.

13 18 16 18

13 18 17 17

10

16 12 18 23

16

10 15 10 11

21 19 21 18

14

14 21

Acet.

Aspr.

Caff. 28 27 24 17

Table 2.6 Number of domains highlighted averaged across group A and B samples

Acetaminophen Aspirin Caffeine

Sample group A

Sample group B

20.1 ± 5.5

16.6 ± 2.6 11.7 ± 2.3 18.5 ± 3.4

shows the caffeine domains. This is repeated for the acetaminophen and aspirin component where relevant, and the results for each tablet are presented in Table 2.5. Table 2.6 shows the overall differences between tablet groups A and B for the included samples. The mean domain sizes are tabulated in the data analysis software, and are presented in Tables 2.7 and 2.8 for the individual tablets and across sample groups, respectively for the included samples.

51

NIR SPECTRAL IMAGING WITH FPA DETECTORS Table 2.7 Mean domain size for individual group A and B samples

Sample A (mm2 ) Acet. Aspr.

Sample B (mm2 ) Aspr.

Caff.

Sample B (mm2 )

Caff.

Acet.

Acet.

Aspr.

Caff.

0.295 0.379 0.225 0.243

2.627 0.422 0.902 1.218 2.739 0.656 0.838 0.829 0.736 0.605 0.904 3.015 1.002 1.622 0.610 0.786 2.490

0.585 0.678 0.738 0.605

Sample A (mm2 ) Acet. Aspr.

Caff. 0.312 0.270 0.253 0.181

Table 2.8 Mean domain size averaged across group A and B samples

Acetaminophen Aspirin Caffeine

Sample group A (mm2 )

Sample group B (mm2 )

0.27 ± 0.060

1.13 ± 0.825 1.85 ± 1.067 0.70 ± 0.120

The particles statistics show clearly that the caffeine domain sizes for group A and B are very different, for group B the mean domain is more than twice the mean domain size for group A samples. However, the variance in this value is equivalent for groups A and B when normalized to the mean particle size. For aspirin and acetaminophen, no cross-sample group comparisons are possible, as the group A samples did not have clearly identifiable discrete particles of these components. For sample only within group B, these two components show significant variation in size. From these data it is clear that there are significant differences between the products created by these two manufacturers. As stated previously this is an over-the-counter pain medication and used here simply as an example of the type of analyses that are possible using NIR chemical imaging. More than likely, the overall differences between these two products matters little in terms of product effectiveness or safety. It is conceivable, however, that for another type of sample this same information – abundance, heterogeneity and particle statistics, could correlate with important characteristics of the product, such as potency or manufacturability.

2.4 Conclusions Near-infrared chemical imaging using multichannel detectors inherits many of the attributes of conventional NIR spectroscopy using a single-channel detector. In addition to its well-documented capabilities as a spectroscopic technique, in comparison to other vibrational imaging approaches, it has unparalleled flexibility in terms of managing widely varying sample size, placement, shape and color. As has been shown with the three examples presented in this chapter, the technique easily

52

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

spans FOVs ranging from 9.7 × 9.7 μm per pixel to 275 × 275 μm per pixel. In addition to the availability of refractive optics, the ready availability of high power optical sources enables NIR chemical imaging to be employed over widely variable FOVs. While some might consider NIR spectroscopy a rather ‘blunt’ tool when compared with either Raman or MIR spectroscopy, it is relatively straightforward to employ and can be used to tackle a wide variety of applications. Its inherent rugged nature makes it an ideal candidate for automation and gives it the capability to move easily into industrial process environments from a base in the research laboratory. In this new setting, the number of potential uses and value of the technology are multiplied many-fold. We have attempted to show how the attributes of large FOV, high-pixel density and rapid data collection times can be employed in a variety of novel ways that transcends the notion of simple chemical images. Spectrochemical analyses in this imaging modality encapsulates numerical concepts of high-throughput spectroscopy, enhanced sensitivity, blending and particle statistics, embedded spectral libraries and calibration standards as well as simultaneous inter- and intrasample comparisons. While these concepts are not unique to NIR chemical imaging they serve well to highlight the wealth of information that can be collected in a single NIR imaging dataset. The utility of spectrochemical analyses is especially enhanced if the analytical method is designed to take advantage of the extraordinary capability of being able to very rapidly generate tens or hundreds of thousands of NIR spectra using multichannel detection.

References [1] Tkachuk, R. (1981) Protein analysis of whole wheat kernels by near infrared reflectance. Cereal Foods World 26, 584–7. [2] Norris, K. H. (1983) Multivariate analysis of raw materials. In Chemistry and World Food Supplies: The New Frontiers (Chemrawn II. L.W. Shemilt ed.), Pergamon, Oxford, pp. 527–35. [3] Sowa, M. G., Mansfield, J. R., Scarth, G. B. and Mantsch, H. H. (1997) Noninvasive assessment of regional and temporal variations in tissue oxygenation by near-infrared spectroscopy and imaging. Appl. Spectrosc. 51, 143–52. [4] Ellis, J. W. (1929) Molecular absorption spectra of liquids below 3 micron. Trans. Faraday Soc. 25, 888–98. [5] Osborne, B. G., Fearn, T. and Hindle, P. H. (1993) Practical NIR Spectroscopy with Applications in Food and Beverage Analysis, Longman Scientific and Technical with Wiley and Sons, Harlow, UK. [6] Whetsel, K. (1968) Near-infrared spectrophotometry. Appl. Spectrosc. Rev. 2, 1–67. [7] Massie, D. R. and Norris, K. H. (1965) Spectral reflectance and transmittance properties of grain in the visible and near infrared. Trans. Amer. Soc. Agric. Eng. 8, 598–600. [8] Norris, K. H. (1962) Instrumentation of infrared radiation. Trans. Amer. Soc. Agric. Eng. 5, 17–20. [9] Norris, K. H. (1964) Reports on the design and development of a new moisture meter. Agric. Eng. 45, 370–2. [10] Norris, K. H. (1965) Direct spectrophotometric determination of moisture content of grain and seeds. Princ. Methods Meas. Moisture Content Liquids Solids 4, 19–25.

NIR SPECTRAL IMAGING WITH FPA DETECTORS

53

[11] Williams, P. C. (1975) Application of near infrared reflectance spectroscopy to the analysis of cereal grains and oilseeds. Cereal Chem. 52, 561–76. [12] Ciurczak, E. W. and Drennen, J. K. (2002) Pharmaceutical and Medical Applications of Near-Infrared Spectroscopy, Marcel Dekker, New York. [13] Workman, J. J. (1999) Review of process and non-invasive near-infrared and infrared spectroscopy: 1993–1999. Appl. Spectrosc. Rev. 34, 1–89. [14] Kidder, L. H., Kalasinsky, V. F., Luke, J. L., Levin, I. W. and Lewis, E. N. (1997) Visualization of silicone gel in human breast tissue using new infrared imaging spectroscopy. Nature Med. 3, 235–7. [15] Snively, C. M. and Koenig, J. L. (1998) Application of real time mid-infrared FTIR imaging to polymeric systems, diffusion of liquid crystals into polymers. Macromolecules 31, 3753–5. [16] Marcott, C., Reeder, R. C., Paschalis, E. P., Talakis, D. N., Boskey, A. L. and Mendelsohn, R. (1998) Infrared microspectroscopic imaging of biomineralized tissues using a mercury– cadmium–telluride focal-plane array detector. Cell. Mol. Biol. 44, 109–115. [17] Chalmers, J. M., Everall, N. J., Schaerberle, M. D., et al. (2002) FT-IR imaging of polymers: an industrial appraisal. Vib. Spectrosc. 30, 43. [18] Lyon, R. C., Lester, D. S., Lewis, E. N. et al. (2002) Near infrared spectral imaging for quality assurance of pharmaceutical products: analysis of tablets to assess powder blend homogeneity. AAPS PharmSciTech. 3(3), 17. [19] Clarke, F. (2004) Extracting process related information from pharmaceutical dosage forms using near infrared microscopy. Vib. Spectrosc. 34, 25–35. [20] Barer, R., Cole, A. R. H. and Thompson, H. W. (1945) Infra-Red spectroscopy with the reflecting microscope in physics, chemistry and biology. Nature 163, 193. [21] Harthcock, M. A. and Atkin, S. C. (1988) Appl. Spectrosc. 42, 3. [22] Taylor, S. K. and McClure, W. F. (1990) NIR imaging spectroscopy: measuring the distribution of chemical components. In Near Infrared Spectroscopy (M. Iwamoto and S. Kawano, eds), Korin Publishing Co. Ltd., Japan, pp. 393–404. [23] Treado, P. J., Levin, I. W. and Lewis, E. N. (1992) Near-infrared acousto-optic filtered spectroscopy: a solid state approach to chemical imaging, Appl. Spectrosc. 46, 553–9. [24] Robert, P., Bertrand, D., Devaux, M. F. and Sire, A. (1992) Identification of chemical constituents by multivariate near-infrared spectral imaging. Anal. Chem. 64 664–7. [25] Treado, P. J., Levin, I. W. and Lewis, E. N. (1994) Indium antimonide (InSb) focal plane array (FPA) detection for near-infrared imaging microscopy. Appl. Spectrosc. 48, 607–15. [26] Lewis, E. N. and Levin, I. W. (1995) Real-time, mid-infrared spectroscopic imaging microscopy using indium antimonide focal-plane array detection. Appl. Spectrosc. 49, 672–8. [27] Lewis, E. N., Levin, I. W. and Treado, P. J. (1994) Spectroscopic Imaging Device Using Imaging Quality Spectral Filters, US Patent Application 5 377 003, December 27; (1996) US Patent Application 5 528 368, June 18; (2000) US Patent Application RE 036 529. [28] Lewis, E. N., Treado, P. J., Reeder, R. C. et al. (1995) Anal. Chem. 67, 3377–81. [29] Sowa, M. G., Mansfield, J. R., Scarth, G. B. and Mantsch, H. H. (1997) Noninvasive assessment of regional and temporal variations in tissue oxygenation by near-infrared spectroscopy and imaging. Appl. Spectrosc. 51, 143–52. [30] Tran, C. D., Cui, Y., and Smirnov, S. (1998) Simultaneous multispectral imaging in the visible and near infrared region: applications in document authentication and determination of chemical inhomogeneity of copolymers. Anal. Chem. 70, 4701–08. [31] Cassis, L. A., Yates, J., Symons, W. C. and Lodder, R. A. (1998) Cardiovascular near-infrared imaging. J. NIR Spectrosc. 6, 21–5. [32] Treado, P. J., Levin, I. W. and Lewis, E. N. (1992) Near-infrared acousto-optic spectroscopic microscopy: a solid-state approach to chemical imaging. Appl. Spectrosc. 46, 553. [33] Goldstein, S. R., Kidder, L. H., Herne, T. M., Levin, I. W. and Lewis, E. N. (1996) The design and implementation of a high-fidelity Raman imaging microscope. J. Microsc. 184, 33–45.

54

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

[34] Wiencke, D., van den Broek, W. and Buydens, L. (1996) Identification of plastics among nonplastics in mixed waste by remote sensing near-infrared imaging spectroscopy 2. Multivariate image rank analysis for rapid classification. Anal. Chem. 67, 3760–6. [35] http://www.lpi.usra.edu/expmoon/clementine/clementine.html. [36] Lewis, E. N. and Haber, K. H. (2001) Hybrid Imaging Spectrometer, US Patent Application 20 020 135 769. [37] Lewis, E. N., Strachan, D. and Kidder, L. H. (2004) High Volume On Line Spectroscopic Composition Testing of Manufactured Pharmaceutical Dosage Units, US Patent Application 6 690 464. [38] Reay, N. K. and Pietraszewski, K. A. R. B. (1992) Liquid nitrogen-cooled servo-stabilized fabry-perot interferometer for infrared. Opt. Eng. 31, 1667–70. [39] Lewis, E. N., Carroll, J. E. and Clarke, F. M. (2001) A near-infrared view of pharmaceutical formulation analysis. NIR News 12, 16–18. [40] Kraft, M. (2003) Spectrosorting – Industrial On-line Material Classification by NearInfrared Spectral Imaging, Presented at the European Symposium on NIR Spectroscopy. DK September. [41] Lawrence, K. C., Windham, W. R., Park, B. and Buhr, R. J. (2003) A hyperspectral imaging system for identification of feacal and ingesta contamination of poultry carcasses. J. NIR Spectrosc. 11, 269–81. [42] Bellamy, M. K., Norman Mortensen, A., Hammaker, R. M. and Fately, W. G. (1995) NIR imaging by FT-NIR-HT spectroscopy. NIR News 6, 10–12. [43] Bellamy, M. K., Mortensen, A. N., Hammaker, R. M. and Fateley, W. G. (1997) Chemical mapping in the mid- and near-IR spectral regions by Hadamard transform/FT-IR spectrometry. Appl. Spectrosc. 51, 477–86. [44] Fately, W. G., Hammaker, R. M., DeVerse, R. A., Coifman, R. R. and Geshwind, F. B. (2002) The other spectroscopy demonstration of a new de-dispersion imaging spectrograph. Vib. Spec. 29, 163–70. [45] Lewis, E. N. (2004) Things that are new in imaging. Presented at the Eastern Analytical Society Annual Meeting. [46] Spragg, R., Hoult, R. and Sellers, J. (2002) Comparing near-IR and mid-IR microscopic reflectance FT-IR imaging. Presented at the Federation of Analytical Chemists and Spectroscopy Societies Annual Meeting. [47] Carr, G. L. (2001) Resolution limits for infrared microspectroscopy and explored with synchrotron radiation. Rev. Sci. Instrum. 72, 1–7. [48] Sommer, A. J., Tsinger, L. G., Marcott, C. and Story, G. M. (2001) Attenuated total internal reflection infrared mapping microspectroscopy using an imaging microscope. Appl. Spectrosc. 55, 252–6. [49] Everall, N. H. (2000) Modeling and measuring the effect of refraction on the depth resolution of confocal Raman microscopy. Appl. Spectrosc. 54, 773–82. [50] Frenández Pierna, J. A., Baeten, V., Renier, A. Michotte, Cogdill, R. P. and Dardenne, P. (2005) Combination of support vector machines (SVM) and near infrared (NIR) imaging spectroscopy for the detection of meat and bonemeat (MBM) in compound feeds. J. Chemomet. 18, 341–9. [51] Neil Lewis, E. (2002) High-Throughput Infrared Spectroscopy, US Patent Application 6 483 112. [52] Clarke, F. C., Jamieson, M. J., Clark, D. A., Hammond, S. V., Jee, R. D. and Moffat, A. C. (2001) Chemical image fusion. The synergy of FT-NIR and Raman mapping microscopy to enable a more complete visualization of pharmaceutical formulations. Anal. Chem. 73, 2213–20. [53] Clarke, F. C., Hammond, S. V., Jee, R. D. and Moffat, A. C. (2002) Determination of the information depth and sample size for the analysis of pharmaceutical materials using reflectance near-infrared microscopy. Appl. Spectrosc. 56, 1475–83. [54] Clarke, F. C. (2004) Extracting process-related information from pharmaceutical dosage forms using near infrared microscopy. Vib. Spectrosc. 34, 25–35. [55] Lewis, E. N., Carroll, J. E. and Clarke, F. M. (2001) A near-infrared view of pharmaceutical formulation analysis. NIR News 12, 16–18. [56] Lewis, E. N., Schoppelrei, J. and Lee, E. (2004) Near-infrared chemical imaging and the PAT initiative. Spectroscopy Magazine 04, 26–36.

NIR SPECTRAL IMAGING WITH FPA DETECTORS

55

[57] Lewis, E. N., Kidder, L. H., Haber, K. S., Lee, E. and Schoppelrei, J. W. (2005) The role and technology of near infrared chemical imaging in pharmaceutical manufacturing. Presented at the International Forum for Process Analytical Chemistry. [58] Lewis, E. N., Schoppelrei, J. W., Lee, E. and Kidder, L. H. (2005) Near-infrared chemical imaging as a process analytical tool for the pharmaceutical industry. In Process Analytical Technology (Katherine Bakeev, ed.), Blackwell Publishing, Oxford. [59] Lewis, E. N., DuBois, J. and Kidder, L. H. NIR imaging, instrumentation and its applications to agricultural and food engineering. In Near Infrared Spectroscopy in Food Science and Technology (Yukihiro Ozaki, W. Fred McClure, and Alfred Christy, eds), John Wiley & Sons, in press.

3

Multichannel detection with a synchrotron light source: design and potential G. Larry Carr, Oleg Chubar and Paul Dumas

3.1 Introduction For the past decade, synchrotron radiation (SR) has impacted the field of infrared (IR) microspectroscopy and the fields of biological, chemical and physical imaging. Prior studies demonstrated the brightness advantage (also called brilliance or spectral radiance) of the synchrotron IR source and significant effort was devoted to understanding and optimizing the source qualities in the IR region, comprising both the so-called mid-IR (∼2 to ∼20 μm) and far-IR (∼20 to ∼1000 μm) regions. The earliest efforts to develop the synchrotron IR source were motivated by the need for a bright far-IR source in low-throughput experimental methods, such as grazing incidence spectroscopy, from surfaces.1 Although it was understood that the brightness advantage of the synchrotron source extended into mid-IR, an actual demonstration of its use for microspectroscopy did not occur until Hemley and coworkers2 developed a custom instrument to focus light through a high pressure diamond anvil cell at the National Synchrotron Light Source (NSLS) at Brookhaven National Laboratory. The first microspectroscopy experiments that allowed a direct comparison with a thermal source took place in 1993, when a commercial Spectra Tech Irμs IR microscope was installed at U2B, a dedicated mid-IR beamline at the NSLS.3,4 Since then, mid-IR microspectroscopy has expanded to synchrotron facilities throughout the world, with a number of important milestones realized in a variety of scientific disciplines: soft matter,3,5,6 geology,7 biology.7–13 The principal motivation for synchrotron-based IR microspectroscopy is to achieve significantly greater lateral resolution (typically at the diffraction limit) while recording data of superior signal-to-noise characteristics without resorting to prohibitively long acquisition times. Although the brightness advantage of the synchrotron source is its most important quality (especially for microscopy), it possesses other unique and useful features. For example, it is a continuum source with spectral coverage from the very far-IR through X-rays, it is pulsed (on the ps and even fs timescale) and it possesses well-defined polarization states. The source can be used in pump-probe studies of dynamics14,15 and is particularly well suited to experiments where multiple and widely varying photon energies or wavelengths are required. As an example, combined X-ray (fluorescence or absorption) and IR microscopic analysis on the same sample12,13 represents an activity of growing interest in the SR community.

DETECTION WITH SYNCHROTRON LIGHT SOURCE

57

As noted earlier, facilities for performing IR microspectroscopy have expanded throughout the world in response to an increasing demand for beamtime from the scientific community. The NSLS (Brookhaven National Laboratory) presently operates six IR beamlines, with four IR microscopes (Beamlines U2A, U2B, U4IR and U10B), making it a leading synchrotron facility for IR microspectroscopy investigations. Active IR microspectroscopy beamlines can also be found at ALS, Berkeley and SRC, Stoughton (USA); UVSOR, Okasaki and SPring8, Nishi-Harima (Japan); NSRRC, Hsinchu (Taiwan). In Europe, IR activities continue at the SRS, Daresbury (UK); ESRF, Grenoble (France); MAXLAB, Lund (Sweden); Daϕne, Frascati (Italy); ANKA, Karlsruhe (Germany) and BESSY II, Berlin (Germany), while LURE (France), which possessed two IR beamlines, closed in December 2003. Other facilities that are planning IR microspectroscopy programs include Diamond, Rutherford Lab (UK); SOLEIL, Paris (France); DELTA, Dortmund (Germany); ELETTRA, Trieste (Italy); SLS, Villigen (Switzerland); Duke-FEL, Durham (USA); CLS, Saskatoon (Canada); CAMD, Baton Rouge, (USA); SURF-3, Gaithersburg (USA); Australian Synchrotron, Melbourne (Australia) and NSRL, Hefei (China). Though most IR microscopy imaging makes use of the rich and unique absorption features found in the mid-IR for chemical identification, there has been increasing interest in extending the spectral range to lower frequencies, motivated in part by not only the developments in coherent THz spectroscopy and imaging,16 but also in response to the needs of the space science community for the identification of complex minerals found in interplanetary dust particles.17 The broad spectral coverage and high brightness of SR reaches well into the far IR, to below 1 THz.18 With few exceptions,∗ synchrotron-based IR microscopy studies have been performed using a single detector element. The synchrotron serves as a high brightness source with diffraction-limited dimensions capable of illuminating a small region, when focused by an IR microspectrometer. Though the spot is small (typically a few tens of microns or less), the intensity in this region is about 100 times greater than for the thermal globar source. Therefore, confocal aperturing, for example, with a 3×3 μm2 effective aperture size delivers good signal-to-noise ratio data and is routinely used with the synchrotron source. At the time of our writing, no complete study has been reported using both the synchrotron source and a multichannel focal plane array (FPA) detector. It is clear that the introduction of the IR FPA detector has brought Fourier transform infrared (FTIR) microscopy with a thermal source to a new and exciting stage of development. This is illustrated in the other chapters of this volume. Our purpose in this chapter is to address how IR FPA technology could be combined with the synchrotron source to advance IR spectroscopic imaging in ways that would prove quite difficult with a conventional thermal source. To address this question, we will need to understand the detailed nature of the synchrotron IR source, the optical ∗ The ANKA group has recently tested a Bruker Hyperion FPA microscope system at

their IR beamline. Higher signal per pixel has been recorded with the synchrotron source, Y. L. Laurent-Mathis, private communication.

58

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

performance of the microspectroscopy instrument and the issues that presently limit the quality of IR images. We will begin our review with a description of source power, source brightness, intensity distribution at the detector position and dependency with wavelength for both the synchrotron and blackbody (thermal) sources. We will then discuss diffraction and how it affects the image resolution and contrast for a typical IR microspectrometry system. This will lead us to propose a method for enhancing the image resolution and contrast using SR with a high magnification Schwarzschild and FPA detector to achieve spatial oversampling and allow for mathematical recovery of an improved image over regions of modest size. As such, we expect that the synchrotron and thermal sources are likely to play complementary roles – much as they have for single-element detector microspectrometers. The thermal source is capable of illuminating large regions and is well suited to surveying large areas. The synchrotron has its intensity concentrated in a small area, and is typically used for microsampling (i.e. collecting spectra from small objects with minimal contamination from neighboring regions) or to produce high spatial resolution images of small areas – typically smaller than 100 μm on one side. We anticipate a similar role for the IR FPA, with the thermal source surveying areas many millimeters on a side and offering excellent performance down to about 10 μm spatial resolution. With the synchrotron, we anticipate that the resolution limit may be extended down to around 1 μm, but over a much more limited area (perhaps no more than 300 μm on a side).

3.2 Comparisons of thermal and SR sources Let us assume that, in all cases, the optical arrangement is properly optimized, for bringing IR flux both from the source to the sample and from the sample to the detector, with all necessary matching. To estimate the spectral flux on the detector, one needs, first of all, to know the flux emitted by the source. If the spectral flux of the source Fsrc is known, then the average spectral flux reaching one pixel of a multichannel detector can be estimated as Fpix ≈

Fsrc ka Spix Fsrc ka ≈ Sd nx n y

(3.1)

where ka is ‘amortization’ coefficient (0 < ka < 1) that takes into account any losses in the optical scheme and absorption by sample, Spix is the pixel area, Sd is the area of the detector surface which should be illuminated (this area includes all pixels and spaces between them), nx and ny are numbers of pixels in horizontal and vertical direction, respectively. Let us postpone consideration of the questions about illuminated area (spot size) on the sample, and the quality of information that can be brought to the detector by the radiation, until next sections, and estimate now the spectral flux provided by three different types of sources: SR from constant field of bending magnet, edge radiation (ER) from extremities of bending magnets (where the magnetic field changes quickly from 0 to a constant level), and the blackbody (BB) source. For each

DETECTION WITH SYNCHROTRON LIGHT SOURCE

59

of these three cases, the spectral flux will be given below as the emitted power per unit wavenumber, which is related to the number of photons per time unit per unit relative spectral bandwidth as: Fsrc ≡

dN dW = hc d(1/λ) dt (dλ/λ)

(3.2)

where h is the Planck’s constant and c is the speed of light. In practical units:     dW dN W Photons −20 ≈ 2 · 10 (3.3) d(1/λ) cm−1 dt (dλ/λ) s(0.1% bw)

3.2.1 Blackbody radiation Planck’s radiation law determines the power emitted by a small aperture in a cavity, which is at a given equilibrium temperature. The spectral flux emitted by an isotropic blackbody source into a solid angle = 2π sin θr (where θr is the angular radius of the first optical element of the spectrometer) is:    −1   hc dW 2πhc2 Ssrc sin θr − 1 ≈ (3.4) exp d(1/λ) BB λkB T λ3 where h is the Planck constant, c is the speed of light, λ is radiation wavelength, kB is Boltzman constant, T is the temperature of the blackbody and Ssrc is the source area. To estimate the flux in more familiar units (e.g. Watts per wavenumbers), Equation (3.4) becomes:     dW W d(1/λ) BB cm−1  −1   Ssrc [mm2 ] sin θr 1.44 · 104 ≈ 3.74 · 10−2 − 1 (3.5) exp λ[μm]T [K] λ3 [μm] For Ssrc = 10 mm2 , θr = π/8 rad, T = 1000 K, λ = 10 μm, Equation (3.2) gives:   dW W ≈ 3.94 · 10−5 −1 at 1000 K d(1/λ) BB cm and



dW d(1/λ)



≈ 1.24 · 10−4 BB

W at 2000 K cm−1

Figure 3.1 gives the power density of IR radiation emitted in a 1 cm−1 bandwidth by a blackbody at 1000 and 2000 K.

3.2.2 SR as an IR source Electron synchrotrons or storage rings use magnetic fields to bend the electrons onto a closed orbit. SR is produced at each of these ‘bending’ magnets. The emitted

60

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Flux (Watts/cm–1)

1000 K 2000 K 1 × 10–4

1 × 10–5

10–6 100

1000 Wavenumbers (cm–1)

Figure 3.1 Emitted power (expressed in Watt cm−1 ) for a blackbody heated at 1000 and 2000 K versus wavelength (in μm), using Equation (3.5).

radiation spans an extremely broad spectral domain, extending from the X-ray to the very far-IR region. IR radiation is generated by electrons traveling at relativistic velocities, either inside a curved path through a constant magnetic field (bending magnet radiation)19 or on their trajectories inside variable magnetic field, for example, at the edges of bending magnets (edge radiation). The latter type of emission has a lot in common with the transition radiation.20,21 IR photons are emitted in narrow angles. Flux and brightness for the two types of IR emission are almost equivalent, but the opening angle of the edge radiation is narrower than that of the SR from constant field of bending magnet, and the intensity profiles of the two types of emission at various wavelengths are different. As the source will be imaged, after passing through all optical components, onto the detector plane array, it is quite relevant to detail the different beam characteristics, including flux, brightness and observed intensity.

3.2.2.1 (Noncoherent) SR from constant field of bending magnet For bending magnet radiation, the ‘natural opening angle’ (the total angle required to transmit 90% of the emitted light) is given by a simple formula:  1/3 λ θnatural [radians] ≈ 1.66 (3.6) ρ where λ is the wavelength and ρ the radius of the bending magnet (both quantities have same units). In the case of the new French third generation synchrotron facility, SOLEIL (E = 2.75 GeV), this radius is 5.38 m, giving a natural angle of ∼20 mrad for 10 μm wavelength (1000 cm−1 ) and ∼44 mrad for 100 μm wavelength

DETECTION WITH SYNCHROTRON LIGHT SOURCE

61

(100 cm−1 ), while at the NSLS (Second Generation Light Source) in Brookhaven National Laboratories (E = 0.80 GeV), bending magnet radius (1.91 m) gives a natural opening angle of 28 mrad at 10 μm, and ∼62 mrad at 100 μm. In order to efficiently collect the IR photons in the broader energy range possible (extending to the far-IR range), the most important technical details involve a modified dipole chamber and large vertical aperture. The instrumentation has been described in depth in a previous paper.22 Assuming that vertical angular aperture of an IR beamline θy is larger than the natural opening angle of SR, that is, θy (λ/ρ)1/3 , where λ is the radiation wavelength and ρ the magnet radius, the spectral flux of SR can be estimated as23 √    +∞ dW 3 eγ I ≈ θx G(λc /λ) with G(x) ≡ x K5/3 (x  )dx  (3.7) d(1/λ) SR 4π ε0 x where ε0 is the permittivity of free space; e is the charge of electron; γ = E/m0 c2 , γ 1 is the electron relativistic mass enhancement factor; I is the electron beam current; θx the horizontal angular aperture; λc = 4πρ/(3γ 3 ) is the critical SR wavelength for the bending magnet and K5/3 is the modified Bessel function. In practical units, Equation (3.7) becomes     dW W ≈ 4.88 · 10−7 E[GeV]I [A]θx [mrad]G(λc /λ) (3.8) d(1/λ) SR cm−1 where E is electron energy (in GeV). For a storage ring with parameters E = 2.75 GeV, I = 0.5 A, λc = 1.43 Å, a beamline with horizontal angular aperture θx = 40 mrad, at the wavelength λ = 10 μm, Equation (3.8) gives:   dW W ≈ 1.40 · 10−6 −1 d(1/λ) SR cm Total flux emitted by a synchrotron source is plotted in Fig. 3.2 for two synchrotron storage rings: a low energy one (0.8 GeV – VUV ring at the NSLS) and a medium one (2.75 GeV – SOLEIL), for the same collection angles (the maximum stored current is 1000 and 500 mA, respectively). They are of about the same magnitude, with the primary difference due to the maximum stored current. On the same figure, we show the flux emitted by a 2000 K blackbody source. Note that the blackbody source (assuming an area of ∼10 mm2 ) produces about two orders of magnitude more of photons, say at 10 μm, but in a much broader angular range. It is important to note the fall-off of the flux emitted by the blackbody source in the far-IR region, explaining the initial strong impetus to the development of IR synchrotron spectroscopy in this frequency range. The emitted beam has both a specific intensity distribution profile and specific polarization properties. High-accuracy methods of computation of SR allow the calculation of the spectral flux and intensity distributions of emitted radiation in the near- and in the far-field regions.24 In Fig. 3.3(a) three distribution profiles

62

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

10–3

Flux (Watts/cm–1 )

Blackbody at T = 2000 K 1 × 10

–4

1 × 10–5

Synchrotron E = 2.75 GeV

10–6 Synchrotron E = 0.8 GeV 10–7 10

100

1000

Wavenumbers (cm–1 ) Figure 3.2 Comparison of the emitted power from two different synchrotron storage ring: NSLS at 0.8 GeV, and SOLEIL at 2.75 GeV, with that of a blackbody at 2000 K. For SR, bending magnet radiation is considered, with collection angles of 40 mrad × 40 mrad (horizontal × vertical).

have been calculated for low energy storage ring (0.8 GeV – NSLS) and medium energy storage ring (Fig. 3.3(b)) (2.75 GeV – SOLEIL), while retaining the same collection parameters (40 mrad horizontal and 40 mrad vertical). Also displayed are the vertical profiles at the middle of the horizontal aperture. The emitted beam has polarization properties that can be exploited for specific experiments. In Fig. 3.4(a), we plot the distribution of the linear polarization parallel and perpendicular to the orbit plane, while Fig. 3.4(b) and (c) show the intensities at circular left and circular right polarizations.

3.2.2.2 (Noncoherent) Edge radiation from extremities of bending magnets In the near-field observation region, the spectral flux per unit surface of the radiation emitted at two edges of bending magnet limiting one straight section of a storage ring is approximately    2 πLr dW 2eI ⊥ ≈ sin2 (3.9) 2 dS d(1/λ) ER 2λz(z + L) π 2 ε0 r ⊥ 2 = where r⊥ is the distance from observation point to the straight section axis (r⊥ 2 2 x + y ), z is the distance from downstream bending magnet edge to observation plane, which is assumed to be r⊥  z  λγ 2 and L is the distance between bending magnet edges (i.e. the straight section length). This intensity distribution is the result of interference of the emission from two bending magnet edges and the near-field effect.25 Equation (3.9) allows the estimation of the angular size of the

(i)

(ii) 3 μm

(iii) 10 μm

20 μm

40 mrad

(a)

40 mrad

Photons/s/0.1% bw

70 × 109 60

(i)

50 40 30

(ii)

20 (iii)

10 0 –30

–10

(i)

0 10 (mm)

20

30

(ii) 3 μm

(iii) 10 μm

20 μm

40 mrad

(b)

–20

40 mrad

Photons/s/0.1% bw

100 × 109

(i)

80 60 (ii)

40 20

(iii) 0 –20

–10

0 (mm)

10

20

Figure 3.3 Wavefront profile and vertical cuts at the first aperture (40 mrad H × 40 mrad V) for three wavelengths: (i) 3 μm, (ii) 10 μm and (iii) 20 μm. (a) For a storage ring of 0.8 GeV (ρ = 2m), and (b) for a storage ring of 2.75 GeV (ρ = 6m). Note that, for a common wavelength, the optimum vertical angle for collection, which is smaller for higher energy storage ring, increases with wavelength.

64

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a) 40 × 10

Photons/s/0.1% bw/mm2

9

(b)

Parallel to orbit plane

30

20

Perpendicular to orbit plane (c)

10

0 –20

–10

0

10

20

Vertical position (mm)

Figure 3.4 (a) Vertical angular distribution above and below the orbit plane for SR, at a wavelength of 10 μm and for a storage ring of 2.75 GeV. (b) Intensity profile for the right circular polarization. (c) Intensity profile for the left polarization.

first interference ring of the edge radiation intensity distribution as observed at a distance z: r⊥1 /z ≈ 2[2λ(z + L)/(zL)]1/2 . At L = 10 m, z = 3 m and λ = 10 μm, this gives r⊥1 /z ≈ 6 mrad. This is much smaller than the natural opening angle of the SR from constant field of bending magnet. One should note, however, that the first interference ring does not contain all the flux of the edge radiation. After integration of Equation (3.9) over r⊥ within a circular aperture zθr centered on the axis of the straight section (θr is the angular radius of the aperture, assuming the origin at the edge of the downstream magnet), we get:     πzLθr2 eI dW (3.10) ≈ H d(1/λ) ER πε0 λ(z + L) where H (x) ≡ ln(x) − ci(x) + C, and  ci(x) ≡ −

+∞

cos(t)t −1 dt

x

is the cosine integral function, C ≈ 0.577216 is the Euler constant. In practical units, Equation (3.10) can be re-written as       W π · θr2 [mrad] zL dW −7 [m] (3.11) ≈ 5.76 · 10 I [A]H d(1/λ) ER cm−1 λ[μm] z + L

DETECTION WITH SYNCHROTRON LIGHT SOURCE

65

40 mrad

(a)

Horizontal position

Vertical position

Horizontal position

Vertical position

40 mrad

40 mrad

(b)

40 mrad Figure 3.5 Intensity profile at 10 μm wavelength, for edge radiation emission, together with the horizontal and vertical cuts at the center of the aperture of 40 mrad H × 40 mrad V. (a) For E = 2.75 Gev (SOLEIL) and (b) for E = 0.8 GeV (NSLS). Note the larger size of the first ring for the lowest electron energy storage ring.

Taking the following realistic parameters: I = 0.5 A, L = 10 m, z = 5 m, θr = 10 mrad, at λ = 10 μm, we obtain from Equation (3.11): 

dW d(1/λ)



≈ 1.5 · 10−6 ER

W cm−1

We note that one typically needs smaller apertures for edge radiation compared to SR from constant magnetic field. In actual practice, a beamline that is optimized to accommodate maximum flux of edge radiation still receives a certain portion of radiation from constant field regions of bending magnets. To illustrate this, we assume that the aperture size that has been defined in the previous sections (namely, 40 mrad horizontal ×40 mrad vertical), is positioned with its center in the axis of the straight section. Figure 3.5 shows the distribution profile at 10 μm wavelength for (a) SOLEIL (E = 2.75 GeV) and (b) NSLS (E = 0.8 GeV), respectively. It can be noticed that the distribution profile is different from that of bending magnet radiation, and the flux is emitted in a narrower angle.

66

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Table 3.1 Electron source size, electron energy in the storage ring and maximum operating current for some synchrotron facilities where IR microspectroscopy beamlines are under operation or construction (∗ ) of low β values

Synchrotron center

ESRF (France)* Spring-8 (Japan)* Elettra (Italy) MaxII (Sweden) SOLEIL (France) NSLS-Brookhaven (USA)

Energy (GeV)

Maximum operating current (mA)

Horizontal electron source size (μm)

Vertical electron source size (μm)

6.0 8.0 2.0 1.5 2.75 0.80

200 100 300 200 500 1000

∼44 ∼83 ∼239 ∼350 ∼180 ∼550

∼9 ∼19.5 ∼13.5 ∼14.5 ∼8 ∼70

3.2.2.3 Effective source size The source size is in the range of microns, and depends on storage ring parameters, electron beam and energy spread.23 Table 3.1 gives the electron source sizes for various synchrotron facilities. It is important to note that, as IR photons are generated by the electron source, the size of the source is much smaller than any thermal source – this property is the primary reason of the brightness advantage of the SR source. Very near this source, the oscillating E-field strength is high and one could argue that the source is extremely bright. The light, however, cannot be focused back to the original source dimensions. Since we are not performing microspectroscopy with samples placed inside the ring vacuum chamber and directly against the electron beam, the brightness that is available to us is controlled by diffraction. In other words, the brightness at the specimen, and not at the source, is the property to consider. Thus, the diffraction-limited spot size should be taken as apparent source size of an IR synchrotron radiation source. To obtain a rough estimate of the diffractionlimited SR source size, one can divide the radiation wavelength λ by the SR natural opening angle, for example, for the SR from a constant field of bending magnet, the diffraction-limited source size is ∼(λ2 ρ)1/3 , where ρ is the bending radius (see Equation (3.6)). In the general case, the effective SR source size is defined by the quadratic sum of the electron beam size and the diffraction-limited spot size of the single-electron emission. For a somewhat more accurate estimate of the apparent source size in particular emission and observation conditions, numerical methods of Fourier optics can be used. In this framework, the effective source size can be obtained either by backpropagation of the wavefront (at a specific wavenumber) to the source position, or by simulating the radiation focusing at optical magnification equal to 1. To illustrate this, we have considered two cases: an IR beamline at the NSLS (0.8 GeV storage ring) normalized at 1000 mA (electron source size = 550 μm horizontal, 70 μm vertical) and an IR beamline at SOLEIL (2.75 GeV – electron source

67

DETECTION WITH SYNCHROTRON LIGHT SOURCE E = 2.75 GeV Vertical position (mm)

2 1 0 –1 –2 –2

1 2 –1 0 Horizontal position (mm) E = 0.8 GeV

Vertical position (mm)

2 1 –1.0 0

–0.5

0.0

0.5

1.0

Vertical size of image of source (mm)

–1 –2 –2

–1

0

1

2

Horizontal position (mm)

Figure 3.6 Image of the source obtained by backpropagating the wavefront to the first aperture (40 × 40 mrad2 ) and back to the source point. Two cases have been considered: an initial electron source of 70 mm vertical (E = 0.8 GeV – NSLS) and 8 mm vertical (E = 2.75 GeV – SOLEIL).

size = 180 μm horizontal, 8 μm vertical) in the case of bending magnet emission. The images of both sources are displayed on Fig. 3.6. The vertical profiles of the images for both sources, obtained at 10 μm wavelength, shows that they are of equivalent size.

3.2.2.4 Comparing blackbody and SR brightness for IR microscopy Brightness, or spectral radiance, also called brilliance, is defined as B(λ) ≈

Fsrc σx · σy · σx · σy

(3.12)

where Fsrc is the flux, as defined in Equation (3.2), σx and σy are the RMS dimensions of the photon beam in x and y directions, respectively and σx and σy are the emission angles. The values of σx , σy , σx and σy in Equation (3.12) are taken at the photon beam waist. Blackbody globar sources provide higher flux in the near- and mid-IR spectral range compared with non-coherent spontaneous SR. However, the source size and angular divergence of a globar source are much larger than effective source size and divergence of IR SR (note that in third generation SR sources, the

68

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

effective source size of IR emission is given by the diffraction limit – see previous section). Following Equation (3.12), one can calculate the source brightness ratio between the synchrotron source and the blackbody source. (σx · σy · σx · σy )Blackbody (Fsrc )Synchrotron BSynchrotron (λ) = · BBlackbody (λ) (Fsrc )Blackbody (σx · σy · σx · σy )Synchrotron = 10−2 ·

3.3 · (π/8) · (π/8) ≈ 2 × 102 4 × 10−2 · 8 × 10−2

Thus, the apparent brightness of SR is higher than that of a globar source that imparts advantages for experimental spectroscopic measurements, including microspectroscopy. One should mention significant recent progress made in accelerator physics toward stable operation of storage rings with small momentum compaction factor, which results in short (sub-millimeter) longitudinal size of electron beam, and as a consequence, high flux of coherent synchrotron emission in the far-IR spectral range.26 The coherent far-IR SR flux can be much higher than the flux from a blackbody source. One can expect that if this progress will continue, generation of coherent SR may be possible also in the mid-IR range. If this happens, the synchrotron sources will be advantageous over blackbody sources not only in terms of brightness, but also in terms of flux.

3.3 The IR microspectrometer: instrumentation and optical analysis In this section, we discuss the imaging performance of IR microspectrometer systems and attempt to compare various designs and measurement methods. To accomplish this, we consider how images are produced and define metrics for the quality of these images. We review previous work demonstrating how SR-based microspectroscopy delivers images at the theoretical limits of performance, and that this limit is at least a factor of two better than for a typical FPA system. In addition, we describe ways that FPAs can be combined with SR to advance the technique further. We also note that the synchrotron source will not be competitive with labbased thermal sources when large areas are to be imaged and the highest spatial resolution is not required. First sub-section discusses the case of wide-field imaging and compares point microscopy to microscopy using multichannel detectors.

3.3.1 Microspectrometer system components In its typical configuration, the IR microspectrometer is an FTIR spectrometer system combined with a microscope and IR detector. Figures 3.7 and 3.8 show schematics for the two most common configurations. Figure 3.7 illustrates a conventional scanning microspectrometer system where a small area (a ‘point’) is spectroscopically sampled by the instrument, and an image is built-up by scanning the specimen through the focused beam in a raster-style fashion. Since only

DETECTION WITH SYNCHROTRON LIGHT SOURCE

69

Upper focusing mirror

Upper aperture

Schwarzschild ‘objective’

Infrared

Sample stage Interferometer Schwarzschild ‘condenser’

Lower aperture Detector Lower focusing mirror Figure 3.7 Schematic for a scanning IR microspectrometer system using a single-element detector and the possibility for confocal operation where aperturing is used both before and after the sample.

a single point is sampled at a time, these instruments use a single-element detector. The microscope itself uses reflecting Schwarzschild-type objectives to avoid chromatic aberrations. One objective serves to focus the light onto the specimen, while the other collects the light and relays it onto the detector. An aperture can be placed at either or both of the conjugate foci for the objectives, thus constraining the illuminated or detected area on the specimen. On some microspectrometer models, the aperture on the illumination beam is called the ‘upper’ aperture, while that for the detection optics is called the ‘lower’ aperture. We will use these terms for convenience. The Schwarzschild objectives are of catoptric design based on two spherical mirrors centered on a common optic axis. Typical magnifications range from 6X up to 50X, with 15X and 32X (36X) among the most common. Numerical apertures (NA’s) vary from 0.3 up to 0.7, with a value of ∼0.6 commonly available. In the mid-IR, these objectives deliver diffraction-limited performance over an area extending 100–200 μm from the optic axis. The design causes the central portion of the objective’s aperture to be obscured, losing up to 25% of the aperture area. As we will show later, this obscuration leads to significant diffraction effects when compared with a conventional microscope objective. As shown in Fig. 3.7, the two

70

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Infrared Schwarzschild objective Interferometer

Sample stage Schwarzschild condenser

Array detector at focal plane

Figure 3.8 Schematic for an imaging IR microspectrometer system using an FPA detection system. Note that the ‘upper’ aperture must be left ‘open’ or removed to allow light to fall onto the entire area to be imaged.

objectives share a common focus at the specimen, that is, they are confocal. However, we add an additional constraint – that the two apertures (upper and lower) are both present and matched to define the same area (size and shape) – when using the term confocal. As noted above, microspectroscopy imaging with a single-element detector is performed by moving the specimen through a microfocused IR beam in a raster-like fashion. In particular, a complete spectrum is collected individually at each point of the specimen to be imaged, with the spot size controlled by apertures and the step size usually set to have the sampled spots overlap somewhat. The material content of each location is then determined by spectral analysis and the chemical content (or other piece of spectral information) is used to define the color or intensity of a given pixel yielding an image. Since the time spent at each spot (a pixel in the final image) is usually at least 15 s, the time required to build up a 64 by 64 pixel image is close to 18 h, in this case. Desirable characteristics of the resulting image are: (1) high signal to noise at each pixel, (2) good lateral spatial resolution and (3) good contrast fidelity (i.e. the observed change in a particular spectral intensity from point to point accurately matches the specimen’s actual composition). The process of collecting a chemical image using an FPA is somewhat different. By replacing the single-channel detector by a multichannel FPA detector, the spectra in a wide field of view are collected in parallel.

71

DETECTION WITH SYNCHROTRON LIGHT SOURCE

Absorbance

10–1

3500

3000

2500 Wavenumbers

2000 (cm–1

1500

1000

)

Figure 3.9 Synchrotron IR spectrum of a brain cut (5 μm thickness), recorded in reflection mode, using a dual aperture of 3 × 3 μm2 , 64 scans at 8 cm−1 resolution.

High spatial resolution and image contrast are enhanced by reducing the effective aperture size for a point-mapping-type IR microspectrometer system. When the aperture defines a region comparable to or smaller than the wavelength of interest, the resolution will be diffraction limited. This assumes ample signal to noise to render an acceptable image, and the thermal source typically does not offer sufficient performance. This is particularly true when using confocal aperturing, and illustrates the benefit of the synchrotron source’s high brightness. Figure 3.9 shows a spectrum recorded at the MIRAGE beamline at LURE (France) for a 5 μm thick section of a brain tissue, deposited onto a gold-coated substrate (reflection mode), recorded with an aperture of 3×3 μm2 .27 The spectrum was collected as 64 coadded interferometer scans, each at 8 cm−1 resolution, for a total collection time of 30 s. One should note that the ‘extra noise’ below ∼1200 cm−1 is a consequence of severe signal loss due to diffraction and the long wavelength cutoff behavior of such a small aperture (< 21 the wavelength). This confocal configuration allows one to reach a resolution of λ/2.28 Figure 3.10 shows another image collected using the 3 × 3 μm2 confocal aperturing and SR. The sample consists of gold wires deposited onto a silicon wafer with a protective polymer overlayer. The high contrast image shows the absorption strength due to the CH2 stretch modes in the 3 μm wavelength range. The specimen was spatially oversampled by raster scanning through the beam in 1 μm steps. The spectrum for each pixel is an average of 64 FTIR scans at a resolution of 4 cm−1 . The total measurement time was 157 min. The schematic in Fig. 3.8 shows a simplified layout for an IR microspectrometer based on a staring-type FPA detection system. The system also uses Schwarzschild objectives, but the apertures for constraining the microscope’s sensitive location

72 (a)

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(b)

45 40 35 30 25 20 40

15 30

10

45

50

35

25 15

5 0

5

20

10

0

Figure 3.10 (a) Optical image of gold wires which have been coated on a silicon wafer. The gold layers are themselves covered with a thin polymer film. (b) Chemical image of the polymer overlayer (peak height of the νas CH2 , at ∼2920 cm−1 ), recorded with a dual aperture of 3 × 3 μm2 , 64 scans at 4 cm−1 resolution.

are left open, that is, they do not provide any spatial discrimination. Thus, the microscope’s first objective illuminates a rather large area, and this illuminated region is then imaged onto the FPA detector by the second Schwarzschild objective. Spatial discrimination is provided by the individual pixels of the detector, each one serving as its own ‘aperture’. Because there is no matching aperture for the illumination objective, this system does not meet our definition of confocal. Though the images shown in the figures and from FPA systems are of high quality, these are not exact representations of the specimen’s properties. The most significant limitation for high spatial resolution imaging, however, is the diffraction of light, as discussed in the next section. Also, the images are degraded due to limitations of the optical system and other factors (detector fidelity, phase correction, apodization etc.).

3.3.2 Performance: imaging at the diffraction limit In general, the goal for synchrotron-based FTIR microspectroscopy is to deliver images at the diffraction limit. This usually means setting one (or both) of the microscope’s apertures to define a region somewhat smaller than the diffraction limit for the respective objective. For convenience, we use d = λ/NA to define this diffraction-limited dimension. This is consistent with both the calculated FWHM of the Schwarzschild diffraction pattern and is also confirmed by experimental resolution studies on test specimens (to be shown later in this section). In general, diffraction and other degrading effects, such as optical distortion, are quantitatively described through the point spread function (PSF), which describes how the intensity from each point on the object is distributed in the image plane. A detailed discussion of the PSF and how it affects image formation can be found in most text books on optics, so we present only a brief review here. Although the PSF can vary, in general, with position in the image plane (e.g. when optical aberrations are present or from vignetting), we will assume here that it is

DETECTION WITH SYNCHROTRON LIGHT SOURCE

73

Figure 3.11 Calculated intensity profiles for a simple, full aperture objective (Airy pattern, left) and a Schwarzschild objective with central obscuration (right). The same NA and wavelength were used for both calculations. Note the large first-order diffraction ring for the Schwarzschild objective.

a function only of the wavelength of light λ. This assumption will be valid for a properly designed and aligned Schwarzschild objective at locations within about 100 μm of the central (optical) axis at the specimen location. This is certainly the case for an image produced by raster scanning of the specimen through the focused beam. With this, we can write an expression relating the observable image with the true image as Oλ (x, y) = Tλ (x, y) ⊗ Pλ (x, y)

(3.13)

where O is the observable image intensity, T is the true image intensity, P is the optical PSF and the ⊗ symbol represents the convolution operation. Note that each is a function of position x, y in the image for a particular wavelength λ, so the convolution is two-dimensional (over both x and y coordinates). The actual image could be degraded even further depending on the spatial sampling, noise and optical misalignment. We will assume that the system is properly aligned and the pixel sampling density is sufficiently high to not limit the image resolution (although this is not necessarily the case for some FPA systems). The PSF (or diffraction pattern) for a model Schwarzschild objective is shown in Fig. 3.11 along with a standard Airy pattern for comparison. Though the Schwarzschild has a somewhat narrower central maximum, the first-order diffraction maximum (ring surrounding the central peak) is much larger than for the Airy pattern. The significance of this can be understood by integrating the intensity (imaging sensitivity) as a function of radial distance from the optic axis, shown in Fig. 3.12. Only half of the intensity is located in the central peak, with the balance in the firstand higher-order diffraction rings. The effect is apparent when scanning across a sharp test specimen, as shown in the absorption profiles of Fig. 3.13. Rather than a steep transition at the edge, the absorption signal increases gradually over a distance of more than 10 μm. Using a 10–90% criteria, the transition width is about twice the wavelength (13 μm wide for λ ∼ 6 μm). The expected profile can be

74

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Sensitivity –encircled fraction

1.00

0.75

0.50 Schwarzschild Conventional Airy 0.25

0.00

0

5

10

15

20

25

Radial distance (μm) Figure 3.12 Calculated sensitivity as a function of radial distance from the center of an objective’s diffraction pattern, comparing the nonconfocal case for a Schwarzschild and a conventional objective (no obscuration). The ‘plateau’ near 5 μm distance for Schwarzschild corresponds to the first diffraction minima. Thus, only 21 of the sensitivity is located in the diffraction pattern’s central peak. Contrast this to the Airy pattern where more than 80% is contained in the central maximum.

1.0 90% Original data Medium deconvolution Strong deconvolution

Absorption signal (rel.)

0.8

0.6

0.4

0.2 10% 0.0 –15

–10

–5

0

5

10

15

Position (μm)

Figure 3.13 Solid line: Absorption signal at a wavelength of 6 μm, for a linear scan across the edge of a thin PMMA sample. Dashed line: Deconvolution of the original data with the calculated PSF, with high spatial frequencies heavily filtered (suppressed). Dotted line: Deconvolution of the original data, but with less filtering of high spatial frequencies.

75

DETECTION WITH SYNCHROTRON LIGHT SOURCE

Sensitivity –encircled fraction

1.00

0.75

0.50 Schwarzschild Confocal Schwarzschild

0.25

0.00

0

5

10

15

20

25

Radial distance (μm)

Figure 3.14 Calculated sensitivity as a function of radial distance from the center of an objective diffraction pattern, comparing Schwarzschild objective used in a nonconfocal (solid line) and confocal (dash-dot line) configuration. The ‘plateau’ still occurs near 5 μm distance for both cases, but the enclosed sensitivity has been increased to more than 80%. Also note how the confocal case approaches 100% more quickly as a function of radial distance.

predicted, based on Equation (3.13), by convolving the Schwarzschild diffraction pattern with an abrupt edge. This is also shown in Fig. 3.13 as a solid line and clearly demonstrates that the spatial resolution is controlled by diffraction specific to the Schwarzschild objective. One method for improving the measured edge sharpness and contrast is to operate the microscope using a confocal optical arrangement. According to our definition, the confocal IR microscope uses two apertures: one to spatially constrain the illumination and the other to constrain the region being detected. If the two objectives are matched, and apertures are set to the same effective size, the same PSF applies for both illumination and detection and the combined (or total) sensitivity function is the (PSF)2 . This leads to significant improvements in the imaging quality. For example, the first-order diffraction ring (plus higher orders) is immediately reduced, and the central peak narrows substantially. Figure 3.14 shows that more than 80% of the sensitivity lies in the central maximum, and essentially all of the sensitivity is captured within 2λ of the optical axis. Therefore, we expect that a confocal microscope will deliver images having much higher fidelity (both improved resolution and contrast). This is illustrated in Fig. 3.15, showing the absorption profile across a sharp edge for both the nonconfocal and confocal cases. The edge sharpness improves by nearly a factor of three, although additional noise is present due to the lower throughput of the confocal system. Since the measured profile is a convolution of the true profile with the optical PSF (which can be calculated), we have an opportunity to improve the fidelity of the profile by deconvolution. If Pλ (x, y) is known or can be determined, then the

76

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

1.0 Absorption signal (rel.)

Absorption signal (rel.)

1.0

0.5

0.0

0.5

0.0 –10

–5

0

Position (μm)

5

10

–5

0

5

Position (μm)

Figure 3.15 Absorption profiles for a line scan across the sharp edge of a thin PMMA film, providing a measure of spatial resolution and image contrast. Open circles are experimental results and solid lines are the calculated profiles based on the diffraction PSF. Left: nonconfocal absorption profile. Right: confocal profile.

operation can be inverted, a process known as deconvolution. This is accomplished with help from the convolution theorem FT{Oλ (x, y)} = FT{Tλ (x, y)} × FT{Pλ (x, y)}

(3.14)

where FT{ } denotes a Fourier transform (two-dimensional in our case) and ‘×’ is simple multiplication. The true image is then Tλ (x, y) = FT−1 [FT{Oλ (x, y)}/FT{Pλ (x, y)}]

(3.15)

with FT−1 [ ] representing the inverse Fourier transform. In practice, the observable image, that is, the image intensity distribution following the optical system, cannot be precisely measured. For example, the image can never be sampled as a truly continuous function, so we have the usual issues associated with discrete Fourier transforms and sampling theorems. Since the sampling density of the deconvolved image is the same as the ‘input’ (observed) image, the observable image must be sampled at a rather high spatial density. There are also errors due to noise introduced by fluctuations in the illumination intensity and noise in the detector. This problem is compounded by the deconvolution process since FT{Pλ (x, y)} approaches zero for short length scales (high spatial frequencies), resulting in large fluctuations in the calculated image. The remedy is similar to that employed in FTIR spectroscopy, where an apodization function is applied prior to Fourier transforming in order to reduce the spectral resolution and eliminate ‘ringing’ around narrow spectral features. In our case, however, the apodization process is applied to the spatial data and serves to limit the degree of resolution improvement. There is one more analogy with FTIR spectroscopy – higher S/N is needed in the raw data to achieve better

DETECTION WITH SYNCHROTRON LIGHT SOURCE

77

spectral resolution while maintaining the same S/N in the resulting spectra. Thus, the degree by which the spatial resolution of an image can be improved is strongly dependent on the S/N of the raw image. We can illustrate the potential for PSF deconvolution using the measurement data shown in Fig. 3.13 (left panel). This is an absorption profile for a line scan across the sharp edge of a photoresist pattern, so the deconvolution is only one-dimensional for this illustration. The required PSF is calculated for a model 32X Schwarzschild objective and a simple ‘boxcar’ apodization serves to truncate the higher spatial frequencies and limit the resolution improvement, in one case set to 5 μm (medium) and the other to 1.6 μm (strong). The results are shown in Fig. 3.15, along with the original (uncorrected) absorption profile for comparison. The edge sharpness improves by nearly a factor of 3 to a value of 4.4 μm and features associated with the first-order diffraction ring are reduced for the medium case. Preserving more high frequencies in the deconvolution reduces the edge width down to nearly 3 μm, but oscillations in the absorption strength with 5% amplitude have now appeared. Therefore, we conclude that a factor of 4 resolution improvement could be achieved by PSF deconvolution, assuming the S/N of Fig. 3.13 is available. Finally, we point out that PSF deconvolution of a two-dimensional image with a large number of pixels will be less dependent on detailed prior knowledge of the optical system PSF. So-called ‘maximum likelihood’ algorithms can be applied to refine the PSF and converge on one that results in the best resolution improvement across the entire image. Images with at least 64 by 64 pixels are typically needed for this process to be successful.

3.3.3 The FPA microscope system We now have the tools to consider the potential resolution for an IR microspectrometer using an FPA detection system. As illustrated in Fig. 3.10, the FPA microscope has no upper or lower aperture to define the instrument’s sensitivity. Instead, each detector pixel serves as an aperture for this purpose. With only one aperture, the system is not confocal according to our definition, and the PSF will have about half of its intensity in the first- and high-order diffraction rings. Edge sharpness (10–90% transition) will be ∼2λ at best. But the effect of the PSF on other specimen geometries varies with specimen geometry, with some circular shapes resulting in imaging artifacts that could be misinterpreted as actual specimen features. We illustrate two such cases. In the first we consider the transmission through a circular hole in an absorbing material. The hole diameter is 12 μm and the relevant wavelength of light is 6 μm. The absorption profile is shown as a gray-scale in Fig. 3.16. In the left panel we show the true absorption image, while in the right panel we show the calculated absorption image. Note the dark region at the center of the hole, suggesting the presence of absorbing material. This apparent absorption is due to the first-order diffraction ring of the PSF falling onto the region just outside the hole. That such an artefact can actually occur can be demonstrated experimentally.

78

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Figure 3.16 Example of an imaging artifact. Left: object consisting of an opaque specimen except for a 12 μm diameter circular aperture. Right: calculated transmission image for a single (nonconfocal) Schwarzschild objective with NA = 0.65 and for λ = 6 μm. Note the dark patch in the center, suggesting the presence of absorbing material inside the hole.

In Fig. 3.17 we show the calculated and measured absorption profile for a scan across a 12 μm hole in a photoresist specimen by using the amide I spectral feature at ∼λ = 6 μm. The central ‘peak’ is clearly evident in the experimental measurement. Also note that the absorption intensity never falls below 40% inside the hole. Perhaps the most severe example involves a ring-shaped specimen that happens to match the first-order diffraction ring in size. This is not necessarily an unlikely occurrence, considering the shape and size of biological cells. Figure 3.18 shows such a calculation, starting with the actual object (left) and the expected image for a nonconfocal microscope, such as an FPA system (center). We also include the expected image for a scanning confocal microspectrometer (right).

3.3.3.1 Prospects for PSF deconvolution with FPA microspectrometers The FPA imaging microspectrometer system offers some opportunity to enhance image quality through PSF deconvolution. The method is commonly employed in visible light microscope images. The method may not be directly transferable to the IR system due to the much broader spectral range of interest for molecular spectroscopy, the Schwarzschild diffraction pattern (as opposed to an Airy pattern) and the limited image field for which diffraction-limited performance is achieved. The distortion that occurs is only significant for image locations rather far from the instrument’s optic axis, and for the shorter wavelengths. This is illustrated in Fig. 3.19, showing the calculated image for a circular disk when imaged at λ = 3 μm. The object’s appearance is dominated by diffraction when centered on the optic axis (left) and also when it lies 100 μm from the axis (middle). But when moved to 283 μm (equivalent to the corner position of a 400 by 400 μm image), distortion

79

DETECTION WITH SYNCHROTRON LIGHT SOURCE

Absorption intensity (rel.)

1.00

0.75

0.50

 = 6 μm Measured Calculated

0.25

0.00 0

5

10

15

20

25

30

Position (μm) Figure 3.17 Imaging artifact for a specimen consisting of a thin uniform layer except for a 12 μm diameter hole (aperture), using a nonconfocal 32× Schwarzschild objective (NA = 0.65) and λ = 6 μm. Solid line: calculated absorption profile through the diameter of the hole. Solid circles: measured absorption profile for an actual specimen. Note the poor absorption contrast (∼50%) and ‘bump’ at the holes’ center (15 μm location).

Figure 3.18 Calculated images for a 14 μm diameter ring-shaped object at λ = 6 μm. Left: actual object; center: image for a nonconfocal 32× Schwarzschild objective with NA = 0.65; right: image for a confocal 32× Schwarzschild objective with NA = 0.65.

becomes significant. Such variations in the PSF would be difficult to deconvolve without some prior knowledge of the distortion. Including the Schwarzschild PSF into a deconvolution algorithm presents no particular obstacles, especially if the algorithm is of the ‘maximum likelihood’ type that automatically converges to a suitable PSF given appropriate starting conditions. The real challenge lies in the

80

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Figure 3.19 Calculated image for a 3 μm diameter circular object at a wavelength of 3 μm and at three different locations within the spectrometer’s potential viewing area, illustrating diffraction blurring and optical distortion. Left: centered on instrument’s optic axis; middle: object located 100 μm from optic axis; right: object located 283 μm from optic axis.

broad spectral coverage – more than a factor of 4 in wavelength – sometimes needed for molecular identification.

3.4 Combining SR with an FPA microspectrometer Our final section considers the question: ‘Can the synchrotron source and FPA technology be matched to provide a significant enhancement over either one alone?’. We believe the answer to this question is ‘yes’. But defocusing the synchrotron’s high brightness beam to fill a large area FPA is not the proper approach. If the SR source is only ∼100–1000 times brighter than a globar, then expanding the beam 10–30-fold along each side immediately eliminates the brightness advantage. Instead, we propose on approach based on a modest. In one, a modest size array (e.g. 32 × 32) is used to spatially oversample a region comparable to the illuminated region, which for the synchrotron source is on the order of the diffraction limit. This is about twice the wavelength for a typical Schwarzschild IR objective with NA = 0.5. The goal for the illumination system is to provide fairly homogeneous illumination that matches the viewing objective’s NA for all wavelengths of interest. Therefore, defocusing slightly to cover, say, a 16×16 μm area could be a reasonable solution.

3.4.1 FPA microspectrometer for PSF image deconvolution We can now envision a particular FPA microspectrometer design for very high spatial resolution imaging using SR. The schematic is shown in Fig. 3.20. A low power (e.g. 10–15X), but high NA objective serves as a condenser to illuminate the specimen. The low magnification allows for a large working distance on one side of the specimen for mounting windows. If the sample rests on a transparent

81

DETECTION WITH SYNCHROTRON LIGHT SOURCE Long working distance, ~0.6 NA condenser (~6 ×–10×)

Infrared from synchrotron

~0.6 NA, 74 × objective

128 × 128 pixel 40 mm pitch MCT FPA

Specimen on motorized x –y stage

Figure 3.20 A schematic idea for an imaging IR microspectrometer system using the synchrotron source and FPA detector. The FPA detector serves to spatially oversample a diffraction-limited region of the specimen for subsequent PSF deconvolution and resolution enhancement.

substrate, the substrate side faces this objective (to avoid chromatic aberrations). The beam transmitted through the specimen is collected by a higher power (e.g. 30–75X, NA = 0.65) objective and imaged onto the FPA detector. Since a typical IR FPA has a pixel spacing (‘pitch’) of 40 μm, the geometrical effective dimension of a given pixel is between 1.3 and 0.53 μm. In the latter case (75X), a 32 × 32 FPA would view an area just over 16 μm in size, which is easily illuminated by high-brightness SR without significant defocusing. Also note that, when compared to an IR FPA system with each pixel representing an area of 6 μm, geometric considerations alone imply that the flux per pixel will be 100 times smaller. However, this small effective pixel size is compensated by the very high source brightness so that good signal to noise should be achieved in a short measurement period. The read out of such a modest size array (1024 elements) is now compatible with rapid-scan FTIR methods and a.c. coupling (and even larger format arrays are now available in microscope systems). The high degree of spatial oversampling enables PSF deconvolution. Since an entire 16 × 16 μm area is collected at one time, registration errors are minimized over the critical dimension for PSF deconvolution (less than or on the order of the wavelength). Additionally, only the central portion of the Schwarzschild optic’s field is used, where the imaging quality is highest. The system becomes a combination of FPA imaging system with a raster scan to image larger areas. Potentially, larger areas may be sampled by the FPA before stepping to the next location.

3.4.2 SR as an extended IR source At this point, we return to the intrinsic characteristics of the dipole-type SR source. Recall that the radiation from a single electron is emitted into an opening angle of θν (λ) ≈ (λ/ρ)1/3 , where ρ is the bending radius of the electron beam. This result is basically a statement that the physical dimensions of the source (the transverse size

82

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

of the electron orbit) and the opening angle meet the conditions for a diffractionlimited source. For a storage ring with a 2 m bend radius and a wavelength of 7 μm (1400 cm−1 ), the full opening angle is 20 mrad or about 1◦ . This applies to both the horizontal as well as vertical angle. Of course, the source orbit spans a full 360◦ , although broken into segments by the individual dipole magnets. Therefore, from the perspective of an observer walking around the ring’s perimeter, each time the observer’s angle changes by 1◦ , a new diffraction-limited portion of the orbit, distinguishable from the previous segment, is viewed. The span of a dipole varies with the storage ring design, but most provide at least 10◦ . The NSLS VUV/IR ring uses 45◦ dipoles, although only about 30◦ is physically available due to obstructions from downstream accelerator components. Even if only 16◦ were collected, each pixel of a 16 element linear array could be illuminated with SR having ideal brightness. Other opportunities may present themselves for combining FPAs with the synchrotron source. For example, the multiplex advantage of FTIR spectrometers is negated when compared with a dispersive spectrometer and linear array detector. This suggests that one could design a microspectrometer using an FPA detection system, with one dimension of the array being used for spatial information and the other for spectral information. The system would produce two-dimensional spectral images by the ‘push broom’ method. An extended line source would match such an instrument particularly well. However, a useful comparison of this technique with other spectroscopic imaging methods will require a detailed design and performance analysis and is beyond the scope of this chapter.

3.5 Summary In summary, we have discussed the use of SR with FPA detection systems for IR spectroscopic imaging. Both edge and dipole SR sources possess characteristics that make them attractive for IR microspectroscopy. Though their characteristics approach that of a diffraction-limited point source, there are instances when it can be considered as an extended source, such as when the electron beam cross-section is large or when a broad horizontal angle of collection is available. Microspectroscopy with this high brightness source has allowed the diffraction-limited spatial resolution capabilities for various instrument configurations to be studied. Agreement between theoretical predictions and test measurements indicates that diffraction effects control the available image resolution and contrast. The secondary obscuration of the Schwarzschild objective leads to stronger diffraction ‘rings’ surrounding the pattern’s central peak, which in turn causes a reduction in resolution and contrast. Image artifacts associated with this pattern are also noted and some are confirmed by actual measurements of test specimens. The confocal optical arrangement effectively eliminates these effects, but is inherently incompatible with FPA detection systems having contiguous pixels. An accurate knowledge of the PSF implies that significant correction is possible by deconvolution methods. But for acceptable results, these methods require both high spatial sampling and excellent signal to noise. It is possible that

DETECTION WITH SYNCHROTRON LIGHT SOURCE

83

these requirements may not be achievable with a thermal source. For example, high spatial sampling can be accomplished using an FPA and a high magnification objective, but the flux incident on each pixel may be more than 100× smaller such that the signal-to-noise is not sufficient for a successful deconvolution. The SR source has the potential to solve this issue, and future synchrotron beamline designs may expand the capability for imaging larger areas with similar performance.

Acknowledgements The authors are grateful to F. Polack (SOLEIL), L. M. Miller, R. Smith and G. S. Smith (NSLS), G. P. Williams (Jefferson Lab.), B. Beccard (Thermo Nicolet-France) for interesting discussions and careful reading and comments. We would like to acknowledge Per Uvdal (Lund-Sweden) for providing us with the gold-wires-deposited Si subtrate. We would like to express out thanks to John Reffner for his constant support. The work performed at the National Synchrotron Light Source was supported by the U.S. Department of Energy under contract DE-AC02-98CH10886 at Brookhaven National Laboratory.

References [1] Dumas, P. and Williams, G. P. (2002) Chemical applications of synchrotron radiation. In Advanced Series in Physical Chemistry (T. K. Sham, ed.), World Scientific, Singapore, Vol. 12A, pp. 356–87. [2] Hemley, R. J., Mao, H. K., Goncharov, A. F., Hanfland, M. and Struzhkin, V. (1996) Phys. Rev. Lett. 76, 1667. [3] Carr, G. L., Hanfland, M. and Williams, G. P. (1995) Rev. Sci. Instrum. 66, 1643–5. [4] Carr, G. L., Reffner, J. A. and Williams, G. P. (1995) Rev. Sci. Instrum. 66, 1490–2. [5] Ellis, G., Marco, C., Gómez, M. A., Collar, E. P., García-Martínez, J. M. (2004) J. Macromol. Sci. Part B – Phys. 43, 253–66. [6] Ellis, G., Gómez, M. A. and Marco, C. (2004) J. Macromol. Sci. Part B – Phys. 43, 191–206. [7] Guilhaumou, N., Dumas, P., Carr, G. L. and Williams, G. P. (1998) Appl. Spectrosc. 52, 1029–34. [8] Miller, L. M., Huang, R., Chance, M. R. and Carlson, C. S. (1999) Synchrotron Radiat. News 12, 21–7. [9] Miller, L. M., Carr, G. L., Jackson, M., Dumas, P. and Williams, G. P. (2000) Synchrotron Radiat. News 13, 31–7. [10] Miller, L. M., Dumas, P., Jamin, N., Teillaud, J. L., Bantignies, J. L. and Carr, G. L. (2000) Microbeam Anal. 2000, Proc. 165, 75–6. [11] Miller, L. M., Smith, G. D. and Carr, G. L. (2003) J. Biol. Phys. 29, 219–30. [12] Dumas, P. and Miller, L. (2003) J. Biol. Phys. 29, 201–18. [13] Dumas, P. and Miller, L. (2003) Vib. Spectrosc. 32, 3–21. [14] Carr, G. L. (1999) Vib. Spectrosc. 19, 53–60. [15] Carr, G. L., Lobo, R., LaVeigne, J. and Reitze, D. (2000) Phys. Rev. Lett. 85, 3001–4. [16] Mittleman, D. M., Jacobsen, R. H. and Nuss, M. C. (1996) IEEE J. Sel. Top. Quantum Electron. 2, 679–92. [17] Keller, L. P., Hony, S., Bradley, J. P. et al. (2002) Nature 417, 148–50. [18] Miller, L. M., Smith, G. D. and Carr, G. L. (2003) J. Biol. Phys. 29, 219–230. [19] Duncan, W. and Williams, G. P. (1983) Appl. Opt. 22, 2914. [20] Bosch, R. A. (1997) Nucl. Instrum. Methods Phys. Res. A 386, 525. [21] Matthis, Y. L., Roy, P., Tremblay, B., et al. (1998) Phys. Rev. Lett. 80, 1220. [22] Carr, G. L., Dumas, P., Hirschmugl, C. J. and Williams, G. P. (1998) Nuovo Cimento 20D, 375. [23] Kinsky, S., Perlman, M. L. and Watson, R. E. (1983) Handbook of Synchrotron Radiation, North Holland, Amsterdam.

84

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

[24] Chubar, O. and Elleaume, P. (1998) Accurate and efficient computation of synchrotron radiation in the near field region, Proceedings of the EPAC98 Conference, 22–26 June, pp. 1177–9. [25] Bosch, R. A. and Chubar, O. (1997) Amer. Inst. Phys. Conf. Proc. 417, 35–41. [26] Abo-Bakr, M., Feikes, K., Holldack, P. et al. (2003) Phys. Rev. Lett. 90, 094801, and references therein. [27] Polack, F., Mercier, R., Nahon, L. et al. (1999) SPIE (G. L. Carr and P. Dumas, eds), Vol. 3575, p. 13. [28] Carr, G. L. (2001) Rev. Sci. Instrum. 72, 1–7.

4

Multivariate analysis of infrared spectroscopic image data Scott W. Huffman and Chris W. Brown

4.1 Introduction Hyperspectral images contain spectral features arising from mixtures of chemical components. Over the last decade, a number of self-modeling algorithms have been developed for creating chemical maps from hyperspectral images by extracting pure component spectra from mixture spectra.1–19 These algorithms are very effective when a large number of spectra are available for processing. In the case of hyperspectral images measured with a 64×64 pixel camera, the 4096 spectra are sufficient for using self-modeling algorithms to obtain three or more pure component spectra and for generating chemical maps of the components. Hyperspectral images present a number of difficult challenges for the chemometrician. First of all, there is an extremely large amount of data for each image. This can be both good and bad. It is good since there are plenty of data for extracting chemical pictures of the components. On the other hand, it is bad since many of data points are not independent and do not present new information for processing. Moreover, handling all of the data can be very time consuming and confusing. Finally, some of the data may not be good and can lead to confusion when trying to extract useful information for particular components. This chapter will address the chemometric challenges and demonstrate the methodologies on a number of applications. A certain amount of preprocessing to clean and compress the spectral data can be helpful. Techniques for reducing high frequency noise and random baseline variations are necessary prior to processing the data. Various methods for viewing the data, such as derivatives and transforms, can also aid in processing. Finally, the various methodologies for modeling the useful information will be addressed.

4.2 Preprocessing hyperspectral images Hyperspectral images consist of spectra measured at each spatial location of the object. If each of the spectra measured with a 4K pixel camera consists of absorbance values at 1000 wavenumbers, the total number of data points would be 4 096 000. At a data resolution of 20 bits, this would round out to 80 M bits or 10 M bytes. As the number of camera pixels increases and the spatial resolution improves, the

86

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Absorbance

1657 cm–1

1151 cm–1

0.8

0.4 0 1800

1600

1400

1200

1000

wavelength (cm–1)

Figure 4.1 Data cube showing spectrum of a protein and image slices at selected wavenumbers.

number of bytes becomes astronomical. Storing and transferring all of this data can be a monumental task. However, viewing the data and using all of the data in a meaningful way can be an even more horrendous job. This section will address methods for compressing the data, reducing noise, maximizing information and displaying the final results in a meaningful manner. Spectral imaging datasets are often referred to as data cubes. The reasoning behind this designation can be inferred from Fig. 4.1. The images are considered to be a data cube, since a complete spectrum is obtained at each x − y spatial location. As demonstrated in the figure, an image can be generated at each wavelength; this is called an image slice. Often, the three-dimensional data cube is unraveled to form a set of two-dimensional spectra (absorbance versus wavenumber) for the sake of processing the data. Some algorithms work more effectively on individual spectra or sets of individual spectra, whereas other algorithms are just as effective on the three-dimensional representation.

4.2.1 Data compression Typically, compression and filtering of spectroscopic data go hand-in-hand. Often, the process of compressing the data leads to a certain amount of noise filtering.

MULTIVARIATE ANALYSIS OF IR IMAGE DATA

87

In general, there are two types of compression: (1) individual spectra can be compressed and filtered; and (2) the entire dataset can be compressed and filtered by representing each of the individual spectra as a linear combination of some smaller set of data, which is referred to as a basis set. In this section, we will address the processing of individual spectra by applying the fast fourier transform (FFT) algorithm and followed this discussion with one on processing sets of spectra with principal component analysis (PCA).

4.2.1.1 Compression with FFTs Fourier transforms have been commonly used for several decades by electrical engineers for signal processing. The FFT algorithm20 was introduced in the 1960s and greatly facilitated applications of the technique. A number of analytical instruments, such as interferometers and nuclear magnetic resonance spectrometers, employ Fourier transforms to convert data to a more practical domain for viewing, but the use of FFTs in processing spectroscopic data did not begin until the 1980s and has been rather limited.21−24 Part of the problem with using FFTs in data processing is the difficulty in relating chemical information to peaks and valleys of the transforms. Nevertheless, it is a powerful means of compressing and filtering individual spectral data. Let us consider a very simple example of a noisy spectrum of a protein as shown in Fig. 4.2. The original spectra covering the range of 4000–900 cm−1 were measured at 8 cm−1 resolution; each of the digitized spectra consists of absorbance values for 805 wavenumber data points. Each spectrum was zero-filled to 1024 data points and Fourier transformed. The original spectrum is shown on the top left in Fig. 4.2 and its corresponding FFT of 1024 terms on the top right. The transform was truncated to 512 data points and the inverse transformed to produce the second spectrum on the left. The transform was then truncated to 256 and finally to 128 terms to produce the bottom two spectra. Truncating to128 terms removes some of the higher frequency spectral information, but the 256th-term transform appears to contain most of the important information with a reduction in noise. Thus, the original 805 data point spectrum could be compressed to 256 data points without significant loss in information. Moreover, the effect of this compression on smoothing the spectrum can be noted by comparing the 256 data point spectrum with the original.

4.2.1.2 Compression of sets of spectra with PCA The idea behind PCA is that a spectral dataset of mixtures of the same components can be expressed as a linear combination of a small set of spectral representations. This is most easily understood if we consider a set of mixture spectra for three components. Assuming that there is no noise and that the mixture spectra are simply the sum of absorptivity spectra for the pure components, each of the mixture spectra can

88

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

IFFT

1024 terms

512 terms

256 terms

128 terms

4000

3000

2000

1000

0

Wavenuy mbers (cm–1 )

500

1000

FFT terms

Figure 4.2 Left – spectra – top, original noisy protein spectrum of 805 data points followed by spectrum produced for IFFT with reduced number of terms. Right – FFTs – top, FFT of original spectrum followed by truncated FFTs.

be expressed as a linear combination of the three pure component spectra such that A = c1 k1 + c2 k2 + c3 k3

(4.1)

where A represents the spectrum of the mixture, ci is the concentration of the ith component and ki is the absorptivity spectrum of the pure ith component. For multiple mixture spectra of samples containing the same three components, the linear relation can be written in matrix notation as ⎡ ⎤ ⎡ ⎤ ⎤ A1 c11 c12 c13 ⎡ k1 ⎢ ⎥ ⎢ ⎥ ⎥ ⎢A2 ⎥ ⎢c21 c22 c23 ⎥ ⎢ (4.2) ⎢ ⎥=⎢ ⎥ ⎣ k2 ⎦ ⎣· · ··⎦ ⎣· · · · · · · · ·· ⎦ k3 Am cm1 cm2 cm3 or, A = CK

(4.3)

where A contains the mixture spectra in rows, C contains the concentrations for each mixture in rows and K contains the absorptivity spectra for the pure components in rows. (It should be noted that vectors are normally written as column matrices rather than row matrices as shown here.)

MULTIVARIATE ANALYSIS OF IR IMAGE DATA

89

Principal component analysis makes it possible to find a set of representations for mixture spectra in which noise and interactions are taken into account without knowing anything about the spectra of the pure components or their concentrations. The basic idea is to find a set of representations that can be linearly combined to reproduce the original mixture spectra. In PCA, Equation (4.3) is rewritten as A = SV T

(4.4)

As before, the mixture spectra are in rows of A. However, both the S and V matrices are unknowns; the S matrix is referred to as the score-matrix and V as the loadingmatrix. The first loading vector in the first column of V (or row of V T ) is the best fit of all the spectra in the A-matrix, that is, the first loading accounts for the most variance in the spectra.25 The scores in the first column of S are the projections of the spectra in A onto the first loading vector, that is, they are the coefficients used to reproduce A from the loading vector. The second loading vector in the second column of V accounts for the next most variance in the spectral data and the corresponding column of scores are the coefficients that account for the contributions the second loading vector makes to each of the spectra in the A matrix. Each successive column in V is another loading vector and the corresponding score column is its contributions in fitting the original spectra. An example of PCA on a set of synthetic spectra for three components is given in Fig. 4.3. The three components are represented by single Gaussian peaks at 825, 1250 and 1700 cm−1 . The intensities of these peaks were randomly generated. Ten spectra for the mixtures are shown in the figure; a total of twenty were used in the analysis. Four loadings produced by PCA are shown on the right side of the figure; only three of these are significant and the fourth represents noise. Each of the original spectra and each of the loadings consisted of 1400 data points, so that 28 000 data points in the original data were reduced to 3 × 1400 + 3 × 20 or 4260 data points (3 columns loading vectors and 3 columns of 20 scores). This amounts to approximately sevenfold compression of the data. The loading and scores for PCA can be generated by singular value decomposition (SVD). Instead of expressing the matrix containing the mixture spectra, A, as a product of two matrices as in Equation (4.4), SVD expresses it as a product of three matrices (4.5) A = U V T The V T matrix contains the loading vectors as before, U contains the scores in columns and  contains the singular values. The difference here is that the columns of U are orthonormal to each other, and the original scores matrix, S, is the product U . (In Equation (4.4), the columns of S are orthogonal, but not normalized.) SVD processing finds the eigenvalues and eigenvectors by diagonalizing the matrix product AT A, (4.6) AT A = V  2 V T The singular values contained in  are the square roots of the eigenvalues in  2 . The scores in U are obtained by projecting the mixture spectra onto the loading

90

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Loading 1

Absorbance

Loading 2

Loading 3

Loading 4

2000

1600

1200

800

2000

Wavenumber (cm–1 )

1600

1200

800

Wavenumber (cm–1 )

Figure 4.3 Left – 10 three-component spectra having a single Gaussian band for each component. Right – First four loading spectra produced from 20 two-component spectra.

vectors in VT U = AV

(4.7)

The SVD algorithm is available in a number of software programs.

4.2.2 Smoothing spectra As shown in Section 4.2.1.2, the FFT can be an effective means of removing high frequency noise from spectra, but it is not always clear how much of the spectral information is removed during truncation. The Savitsky–Golay smoothing routine26−28 has been widely accepted in analytical chemical instrumentation for almost four decades. This is a very effective method that is readily available in most software packages. Very simply, the Savitsky–Golay algorithm is a nonlinear weighted smoothing function. In the simplest smoothing procedure, we might use a 3-point routine to smooth a spectrum by calculating a running average of a data point with the previous point and the following point. In this way, all three points are weighted equally. We might improve upon this procedure by using a weighted running average of three points with the point in the center weighted twice that on either side. We could take it a step further and use 5 points, with the center point

MULTIVARIATE ANALYSIS OF IR IMAGE DATA

91

Original

Quadratic-7

Quadratic-11

Cubic-11

Cubic-19

4000

3000

2000

1000

Wavenumber (cm–1 ) Figure 4.4 Top – original spectrum of a protein followed by the same spectrum smoothed by various Savitsky–Golay models.

weighted at 3.0, the next two at 2.0 and the outside two by 1.0. The 3- and 5-point functions are referred to as triangular smoothing functions. The Savitsky–Golay procedure is a modification to the triangular smoothing method. It replaces the triangular weighting function with a (selectable) polynomial curve, such as a quadratic, cubic, quartic or quintic relation. The neat thing about the Savitsky–Golay algorithm is that the coefficients in the polynomial are calculable ahead of time and are usually present in lookup tables. An example of Savitsky– Golay smoothing on the same noisy protein spectrum shown in Fig. 4.2 is given in Fig. 4.4. Examples of smoothing are given with both quadratic and cubic functions and a variety of data points. In the 7-point quadratic, a moving window consisting of the central point and 3 points on either side weighted with a quadratic function. It can be seen that even with a 19-point cubic fit the relative intensities and peak shapes of the important bands are not affected seriously by the smoothing processing. One of the attributes of the Savitsky–Golay method is that high frequency noise is removed before the spectrum information. However, there is a limit to the number of points

92

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

that can be used and, as shown later (Section 4.2.3.1), if too many data points are included in the smooth, important information will be lost.

4.2.3 Noise in hyperspectral images Spectroscopic imaging data is often fraught with noise that prevents clean, easy information extraction. This noise arises from various places such as the detector, the interferometer and others components, found in conventional single detector element equipped Fourier transform infrared (FTIR) spectrometers coupled to infrared microscopes.29 The noise induced by these components can be either random or systematic, and can therefore be minimized in the data by signal averaging in the case of random noise or subtraction of systematic errors from the data. An additional source of noise arises from the heterogeneous nature of samples. Ironically, it is the identification and quantification of these same spatially distributed chemical imperfections of a sample that is often the aim of many investigations utilizing FTIR spectroscopic imaging. This ‘new’ type of noise manifests itself as broad baseline fluctuations in the spectra. To further hinder information extraction from the hyperspectral data, these broad baseline fluctuations are not consistent across the sample plane; thus, causing pixel-to-pixel variations in the data that is not chemical or molecular in nature in the traditional sense of vibrational spectroscopy. As a consequence of this broad, low frequency noise being sample induced and not systematic, the baseline variations cannot be subtracted. These broad spectral features are often a function of one or more optically active features in the sample such as boundaries and interfaces between different materials. As the goal of chemometrics is to extract chemical information from spectroscopic imaging data, we must first eliminate or minimize the extraneous information or noise from the data before we are able to reliably produce chemical maps. The manner in which we approach the elimination of noise in spectroscopic imaging data results in the categorization of two types of random and semirandom noise: high frequency and low frequency. We loosely define high frequency noise as features in the data that are narrower than the narrowest infrared spectroscopic band, and low frequency noise as features that are broader than the broadest vibrational spectroscopic band. Ideally, these sources of noise should be minimized or eliminated by utilizing appropriate measurement methods or processing techniques, such as coherently adding sequential interferograms to minimize random noise. Unfortunately, solutions to these sources of noise are not often available or even practical, since the sample is not under the spectroscopist control. As a consequence, digital processing techniques must be found to overcome spectral irregularities before information can be extracted from the datasets. For clarity, we will utilize synthetic data to illustrate the problems and potential solutions. Images from a synthetic dataset with dimensions of 25 × 25 in the spatial domain and containing 100 channels in the spectral domain are shown in Fig. 4.5. Spectra at each pixel consist of a Gaussian peak at spectral channel 50. The area under the Gaussian change from 0 to 30 across the spatial domain in the patterns

MULTIVARIATE ANALYSIS OF IR IMAGE DATA

(a)

(b)

(c)

(d)

93

Figure 4.5 (a) 25 × 25 image generated from spectra with a single Gaussian peak having intensities 0–30. (b) Low frequency noise superimposed on (a). (c) High frequency noise superimposed on (a). (d) High and low frequency noise superimposed on (a).

depicted in Fig. 4.5(a). (In these figures white is high intensity and black is low.) These images are of the slice at the spectral channel 50. The same dataset with a nonhorizontal baseline is shown in Fig. 4.5(b); in this case a random parabola is superimposed onto the baselines. The original synthetic dataset with synthetic high frequency noise superimposed is shown in Fig. 4.5(c) and the original synthetic dataset when both the high and low frequency fluctuations shown in Fig. 4.5(b) and (c) are superimposed is shown in Fig. 4.5(d). The synthetic data used in constructing Fig. 4.5(d) represents the most realistic of the four synthetic datasets.

4.2.3.1 Derivative spectra Perhaps, the simplest method for minimizing sample-induced spectral baseline fluctuations is by calculating derivative (first, second, etc.) spectra. In derivative spectra, broad baseline fluctuations have much smaller slopes and curvatures than vibrational absorption bands in the original spectra, and as a consequence, the resulting derivative spectra have diminished broader features. This effect is seen in the Fig. 4.6,

94

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS Un derivatized spectra

First-derivative spectra

(a)

60

Second-derivative spectra

(b)

(c)

10

3 2

5

40

1 0 0

20 –5

–1

0 0

50

100

–10

–2 0

50

100

0

50

100

Figure 4.6 (a) Image from Fig. 4.5(b). (b) Image for first-derivative spectra of (a). (c) Image for second-derivative spectra of (a).

where an original image containing broad baseline fluctuations in the spectra is plotted in Fig. 4.6(a), the first-derivative spectra in Fig. 4.6(b), and second-derivative spectra in Fig. 4.6(c). In the case of the second derivative plotted in Fig. 4.6(c), the derivative spectra has been multiplied by −1 to keep the color scheme of the images consistent. In Fig. 4.6(c), the spectra consist mostly of Gaussian features. In the underivatized image (Fig. 4.6(a)), the effect of baseline fluctuations is observed as the reduction in relative intensities in the middle row and middle column of the image. The intensity distribution improves in the image of the first-derivative spectra and is very uniform in the image of the second-derivative spectra. Unfortunately, with more realistic hyperspectral data, such as those with high frequency noise, derivative calculations can accentuate higher frequency spectral features, which detract from their usefulness. For example, the same images shown in Fig. 4.6 are shown in Fig. 4.7 after superimposing the high frequency noise. The decrease in clarity of the image is a result of high frequency noise present in the original data. This high frequency noise amplification can be minimized by smoothing the data while calculating the derivative, as is done in the Savitsky–Golay algorithm discussed in the previous section. As mentioned earlier, caution must be applied while applying the Savitsky–Golay method for smoothing datasets to ‘eliminate’ the high frequency noise because, if too wide of a smoothing function is employed, it will also smooth or ‘eliminate’ desirable information. As a general rule, the larger the smoothing function the more information will be lost. As an example, we have altered our original synthetic data to

95

MULTIVARIATE ANALYSIS OF IR IMAGE DATA Un derivatized spectra

First-derivative spectra

Second-derivative spectra

(a)

(b)

(c)

60

10

4

5

2

0

0

–5

–2

40

20

0 0

50

100

–10

–4 0

50

100

0

50

100

Figure 4.7 (a) Image from Fig. 4.5(d). (b) Image for first-derivative spectra of (a). (c) Image for second-derivative spectra of (a).

include a shoulder on the main peak at spectral channel 50 as shown in Fig. 4.8(a); the image of the second-derivative spectra is shown in Fig. 4.8(b). An important aspect of this modification of the original synthetic data is that it does not alter the image. Additionally, it is important to note the utility of using second-derivative spectra in resolving the peaks. Next, three levels of high frequency noise were superimposed on the modified spectra. Examples of second-derivative spectra obtained after the superimposed noise are shown in Fig. 4.9. The noise levels are designated at n10, n5 and n1, where the 10, 5 and 1 refer to the relative noise levels. At noise levels of n10 and n5, the spectral information is not discernable. In the first row, the spectra were smoothed with a 5-point cubic function using the Savitsky–Golay method. In the second row, the spectra were smooth with a 9-point cubic and in the third row with a 15-point cubic function. The 9-point smoothing makes it possible to clearly identify two peaks in the spectra, but two peaks are more difficult to pick out with a 15-point smooth. The effect of these same noise levels and the same Savitsky– Golay smoothing on the corresponding images are shown in Fig. 4.10. The benefits of smoothing with 9- and 15-point functions are clear from this figure. The 15-point smooth of the third column produced an image as good as the original shown in Fig. 4.8(a). Thus, the 15-point smoothing routine improved the image quality at the expense of the spectral information. In Figs 4.9 and 4.10, the columns represent the same noise level and the rows are the same sized smoothing function. Trends can be seen in both dimensions;

96

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a)

(b)

50

2

40 1 30 0

20 10

–1

0 0

50

100

–2

0

50

100

Figure 4.8 (a) Image from center of overlapping bands. (b) Image from center of overlapping bands after taking second-derivative.

as the size of the smoothing function increases the image is clearer but spectral information is reduced. Additionally, with less noise present a smaller smoothing function can achieve acceptable clear baseline corrected images and maintain the spectral information. An important lessen is also illustrated in this example. If for the purpose of creating a chemical map from noisy data, only one prominent spectral parameter, such as wavenumber position of an isolated or large vibrational band, is required, then second-derivative spectra with a large smoothing function can be employed. In contrast, a large smoothing function will most likely obscure some of the spectral information. An application of derivative spectra for minimizing baseline fluctuations in real data is shown in Fig. 4.11. The sample measured for this image was butter contaminated with nonpathogenic bacteria. This was a microscopic image of a 400 × 400 μm sample. The butter covers the entire sample, and the bacteria contamination was placed on the butter in the lower right corner. An image of the band height of the amide I band positioned at ∼1656 cm−1 is shown on the left. An example spectrum of a pixel containing spectral features of both the bacteria and the butter is shown on the bottom left. In constructing a chemical map of the whereabouts of the bacteria on the slab of butter, the first and simplest identification of bacteria is the amide I vibrational modes of the protein in the bacteria, because butter contains

97

MULTIVARIATE ANALYSIS OF IR IMAGE DATA 5

5

g5n 10

g5n 1 0

0

0

–5

2

g5n 5

0

50

100

4

–5 0

50

100

–2

0

2

2 g9n 10

100

50

100

50

100

g9n 1

g9n 5 2

50

0

0

0 –2

0

50

100

1

–2

0

100

1

0 g15n 1

g15n 5

0

0

0

–2 1

g15n 10

–1

50

50

100

–1

0

0

50

100

–1

0

Figure 4.9 Columns – spectrum in Fig. 4.8(a) with high frequency noise levels of 10, 5 and 1. Rows – spectra after Savitsky–Golay smoothing with cubic having 5, 10 and 15 points.

little or no protein. Unfortunately, because of strong, random baseline interferences, the left image has positive intensities for the amide I band at a number of pixels, particularly in the upper right corner. The image formed from the second-derivative spectra on the right in Fig. 4.12 shows very clearly the location of the bacteria. Later, we will show how the spectra of the bacteria can be separated from that of butter for the same image. The difficulty of obtaining a pure chemical image at this point is due to the strong, variable baselines in the spectra and we will address this problem in Section 4.2.3.2. The most obvious solution to problems where both high and low frequency noises are embedded in the data are better instrumental practices. For example, by sufficient coaddition, the high frequency noise component of spectroscopic imaging data can be averaged out. However, when better instrumental practices are unavailable or impractical, other means must be found to enhance the spectroscopic imaging data so that information may be extracted. The method of using second-derivative spectra, shown in this section can be used with success if the high frequency noise in the data is limited.

4.2.3.2 Spectral and spatial filtering The effects of pixel-to-pixel baseline variation can be greatly reduced by taking the first or second derivatives of the spectra as shown in the section above; however,

98

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

g5n10

g5n5

g5n1

g9n10

g9n5

g9n1

g15n10

g15n5

g15n1

Figure 4.10 Images from spectra shown in Fig. 5.9.

spectra of pure components and chemical image maps cannot be obtained from these manipulations. We really need to reduce the pixel-to-pixel baseline variations in the original hyperspectral images in order to obtain spectra of the components. Recently, it was found that Fourier transform filtering could be used effectively in reducing these baseline interferences.30 Typically, baseline interferences are very broad and they can slope upwards, downwards, be high in the middle or high at both ends. Their very broad nature suggests that they contribute to the very low terms in the Fourier transform. Thus, by taking the Fourier transform of a spectrum and reducing the size the of the low frequency terms, some of the broad baseline will be eliminated from the reverse transform. Instead of using a boxcar filter as we did with the truncation method in Section 4.2.1.1, a curved filter as shown in Fig. 4.12 was used. This low/mid-pass filter consisted of a Gaussian curve centered at 18 th of the number of terms in the transform and having a standard deviation (width) of 18 th the number of terms in the transform. This mild-filter removes the high frequency noise and removes a small amount of the very broad baseline interference. It is important not to try to remove

99

MULTIVARIATE ANALYSIS OF IR IMAGE DATA Second dervature amide I image

Amide I image 06 10

10

0.04

04 20

20 02

30

0

40

–0.2

50

60

–0.4

60

40

0.02

40

50

20

0.03

30

0.01 0

50

20

40

60

Bacteria pixel (4, 54)

Bacteria pixel (4, 54) 0.15 0.4 Absorbance

0.1 0.2 0

0.05 0

–0.2 –0.05 1000

1000

2000 3000 Wavenumber (cm–1 )

2000

3000

Wavenumber (cm–1 )

Figure 4.11 Images and spectra of bacteria contamination of a microscopic butter sample. Butter cover the sample, whereas bacteria is confined to the lower right corner.

FFT

3000

2000

FT terms

1000

Wavenumber (cm–1 ) Low/Mid-Pass Filter

IFFT

3000 2000 Wavenumber cm–1 )

1000

Figure 4.12 Filtering of spectrum at pixel 38 in row 63.

FT terms

100

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

FFT 20

40 Pixel number

60

20

40 FT terms

60

Low-pass filter

IFFT 20

40

60

20

Pixel number

40 FT terms

60

Figure 4.13 Spatial filtering of absorbances at 950 cm−1 across row 63.

too much of the baseline with this method, since it will tend to distort the spectral features. The more significant filtering is applied in the spatial domain. After spectral filtering of each of the 4096 spectra, spatial filtering was applied to the hyperspectral dataset. If we look closely at the spatial variations of the absorption intensity at a particular wavenumber (wavenumber slice), we find a rather high frequency variation across a row or down a column of absorbances. To reduce these rather rapid variations, a two-dimensional FFT was applied to the data at each spectral wavenumber. For demonstration purposes, we show the effect in one-dimensional space by considering only row 63 data for the butter/bacteria example at the spectral channel corresponding to 950 cm−1 as shown in Fig. 4.13. A lot of spatial variation would not be expected for the absorbance at 950 cm−1 ; however, there is relatively high frequency variation across this row. The 64 pixel spatial plot is Fourier transformed and multiplied by a half-Gaussian low-pass filter centered at the first FT term and having a standard deviation of 18 th the number of FT terms (a standard deviation of 8 was used for rows containing 64 pixels). An inverse Fourier transform of the filtered row is then calculated producing a new spatial distribution. The effects of spectral and spatial filtering on nine spectra from row 63 at the butter/bacteria boundary (row 63, pixels 34–42) are shown in Fig. 4.14. The more gross baseline changes are evident in the full spectral range of 4000–900 cm−1 , whereas finer, detailed improvements are more obvious in the fingerprint region of 2000–900 cm−1 . Starting with pixel 38 in row 63 and moving to higher pixel numbers, the field of view changes from all butter to a bacteria/butter mixture; thus, the carbonyl band at 1742 cm−1 from the butter should decrease while the amide I and amide II bands at 1651 and 1547 cm−1 , respectively, for bacteria should increase. These predicted spectral changes are much clearer in the spectra after filtering, that is,

MULTIVARIATE ANALYSIS OF IR IMAGE DATA

Before

3000

2000 Wavenumbers, cn-1

101

After

1000

3000

2000 Wavenumbers, cn-1

1000

Figure 4.14 Row 63 spectra from pixels 34–42 before and after filtering.

the 1742 cm−1 band decreases in intensity, there is an isobestic point at ∼1720 cm−1 and the amide I and amide II bands increase with increasing pixel number.

4.3 Processing hyperspectral images Once spectroscopic imaging data have been acquired and preprocessed, information can be extracted. There are two goals in extracting data from images: one is to divide the data into groups, and the other is to determine the absolute or relative concentrations of the different materials. In most instances, these two goals are accomplished simultaneously, particularly in determining the concentrations. In some cases, it is either not necessary to know the concentrations of individual components represented in spectroscopic imaging data, but it is important to recognize differences and similarities of the individual components. In this section, we will first describe a series of qualitative comparison methods for identifying and classifying heterogeneous materials into groups, and then consider methods for estimating relative concentrations.

4.3.1 Feature extraction The most straightforward method of utilizing spectroscopic imaging data to segment material into different groups utilizes ‘pure’ raw spectral features, such as wavenumbers, bandwidths, band areas or centers of gravity of a band. An example of this can be seen in Fig. 4.15, where synthetic data images are plotted in Fig. 4.15(a) and (b) at spectral channels for band maxima of 68 and 75 corresponding to the pure

102

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a)

(b) 5

5

10

10

15

15

20

20

25

25 5

10

15

20

25

5

(c)

10

15

20

25

(d) 1

1

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2 0

0 20

40

60

80

100

20

40

60

80

100

Figure 4.15 Single wavenumber images of synthetic data: (a) wavenumber image at spectral channel 75, (b) wavenumber image at spectral channel 68, (c) spectrum of the pixel at position (20, 20) and (d) spectrum of the pixel at position (10, 10).

component synthetic spectra shown in Fig. 4.15(c) and (d), respectively. If pure components in a sample exhibit spectroscopically resolved peaks for each component, classification of pixels in the spectroscopic imaging data is easy. Additionally, if all samples were so designed or even possible, numerical analysis and chemometrics would not be needed. In more realistic imaging data, spectral bands are often overlapped, and in such cases, raw spectral parameters do not provide sufficient separation of the groups. For example, two images are shown in Fig. 4.16 for synthetic data using spectral channels 68 for Fig. 4.16(a) and 75 for Fig. 4.16(b) corresponding to spectra shown in Fig. 4.16(c) and Fig. 4.16(d), respectively. The difference between data in this figure and those in the previous one is that the bandwidth of the peak in spectrum Fig. 4.16(d) is twice as large as before and overlaps a band in spectrum Fig. 4.16(c). In this figure, the two images do not provide distinct isolation between the two groups because of the spectral overlap. A slightly more complex utilization of raw spectral features to segment spectroscopically different regions in imaging data is to use combinations of spectral parameters. A common parameter combination is the band ratio. Since these two spectral groups of the synthetic data contain different numbers of bands and peak heights, the ratio of intensities at channels 75/25 and 75/60 should provide different values for pixels containing different amounts of each spectral group. These two band ratio images are shown in Fig. 4.17. Although these images are noisy because

103

MULTIVARIATE ANALYSIS OF IR IMAGE DATA (a)

(b) 5

5

10

10

15

15

20

20

25

25 5

(c)

10

15

20

25

5 (d)

1

0.8

0.6

0.6

0.4

0.4

0.2

0.2

0

0 40

60

80

15

20

25

1

0.8

20

10

20

100

40

60

80

100

Figure 4.16 Single wavenumber images of synthetic data with wider bandwidth of one of the spectroscopically unique groups: (a) wavenumber image at spectral channel 75, (b) wavenumber image at spectral channel 68, (c) is spectrum of the pixel at position (20, 20) and (d) spectrum of the pixel at position (10, 10).

(a)

(b) 5

5

10

10

15

15

20

20 25

25 5

10

20

15

40

20

25

60

5

10

20

15

40

20

25

60

Figure 4.17 Band ratio images of synthetic data: (a) is the band ratio image spectral channels 75/25, (b) is the band ratio image at spectral channels 75/68.

104

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a)

2.5 5

5

2

10

1

(b)

1.5

0.5

10

0

15

1

15

–0.5

20

0.5

20

–1

25

5

10

15

20

25

25

5

(c)

10

15

20

25

(d) 5

10 15

5

2

10

0

15

–4

25 5

10

15

20

25

1

4

–2

20

–1.5

0.5 0 –0.5

20 –1 25 5

10

15

20

25

Figure 4.18 The first four scores of a PCA of the synthetic data with only high frequency noise: (a) first score image, (b) second score image, (c) third score image and (d) fourth score image.

some of the ratios have zero values either in the numerator or denominator, it is very clear that the sample represented by spectrum (a) is in the upper left and the spectrum for (b) is in the lower right of the image. Often there is so much overlap of various bands in image spectra that it is nearly impossible to use raw data or even combinations of spectral bands to classify pixels in the dataset. At such times, it is appropriate to transform the data into alternate spaces. Perhaps the most common of these transformations occurs during PCA, which converts the spectroscopic data into scores and loading vectors as described previously. Scores and loadings are calculated in terms of diminishing variance or, in other words, the first score represents the original data with the largest variance and the second contains the second most variance and so on. The first four PCA score plots for the synthetic data in Fig. 4.16 are shown in Fig. 4.18. The first score image shown in Fig. 4.18(a) demonstrates a feature commonly found in most first score images, that is, they contain spatial features corresponding to pixels with the most prominent spectral features. In the case of the image in Fig. 4.18(a), the synthetic data contains only two spectral groups and a small amount of high frequency noise, and therefore, two prominent features are indicated in the image. The second score image in Fig. 4.18(b) shows promise for separating the two groups. In this image, the group centered at (10, 10) has positive intensity and the group centered at (20, 20) has negative intensity. This is perhaps the best example of segmenting these artificial data into separate groups. The third score image in Fig. 4.18(c) is

105

MULTIVARIATE ANALYSIS OF IR IMAGE DATA

(b)

(a) 6.5

5

1 5 0

10

6

10

15

5.5

15

5

20 25

4.5 5

10

15

20

20

5 1

5

0.5 0

10

–0.5

15

–2

25

25

(c)

–1

10

15

20

25

(d) 5

1

10

0.5

15

0

20

–0.5

–1 20

–1.5

25

25 5

10

15

20

25

5

10

15

20

25

Figure 4.19 The first four scores of a PCA of the synthetic data with high and low frequency noise. (a) The first score image, (b) the second score image, (c) the third score image and (d) the fourth score image.

similar to the one shown in Fig. 4.18(a) and contains the last residual traces of the groups. High frequency noise is observed in the fourth score image (Fig. 4.18(d)). The first four score images of the synthetic data in Fig. 4.16 after the addition of low frequency noise are shown in Fig. 4.19. The major difference between Figs 4.18 and 4.19 is that the presence of the low frequency noise degrades the quality of the images. The reason for this is that the low frequency noise is significant enough for the PCA to try to fit it. More loadings and scores are needed to fit the variable baselines, but more loadings and scores tend to also fit other noise as well. This is another example of why low frequency noise must be dealt with during preprocessing. A unique result of PCA methods over previously described methods is that it often produces score images that are descriptive of the spatial and spectral features with little or no a priori knowledge of the data; this feature is true for the images in both Figs 4.18 and 4.19. PCA is often one of the first chemometric techniques applied to data to obtain some sense of the information available and difficulties in obtaining reasonable results. We have demonstrated the methods in this section using synthetic data. Now the processing will be applied to the bacteria/butter image data described previously. In this experiment, bacteria are obvious contaminants of the butter, and as such, the location on the sample needs to be identified. Four images are shown in Fig. 4.20 illustrating the different group separation methods described earlier in

106

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a)

0.6

(b)

10 20

0

40

50

–0.2

50

60

–0.4

60

40

5

20

20 0

40

50

50 –5 40

60

40

60

2 1

30

40

20

–0.2

3 10

60

0

(d)

10

30

0.2

20

60

(c)

0.4

30

40

20

0.6

20

0.2

30

0.8

10

0.4

0 –1 –2

60 20

40

60

Figure 4.20 FTIR spectroscopic image data of butter with a bacterial contamination in the lower righthand corner: (a) wavenumber slice at amide I (1656 cm−1 ), (b) wavenumber slice of the butter carbonyl band (1720 cm−1 ), (c) amide I carbonyl band area image and (d) butter carbonyl band area image.

this section. Figure 4.20(a) shows an image of a single wavenumber slice at the maximum absorbance of the amide I band (1656 cm−1 ). Since butter contains little or no protein and bacteria is mainly protein, the image in Fig. 4.20(a) indicates, correctly, the lower right-hand corner as the position of the bacterial contamination. Furthermore, Fig. 4.20(a) shows elevated levels near the center of the sample, but on close inspection of the spectra in this region, bacterial spectral features were not found. The butter carbonyl peak height in Fig. 4.20(b) also shows an absorbance increase in the lower right-hand corner indicating that some of this corner region contains larger amounts of butter as well as bacteria. For example, the area near coordinates (40, 55) contains little or no spectral evidence of bacteria, but the peak height images indicate elevated levels for both butter and bacteria. Inspection of the spectra from this region revealed only butter with an elevated spectral baseline in the region of 2500–1400 cm−1 . The two remaining images, that is, Fig. 4.20(c) and (d) are of the band areas of the bacteria’s amide I band and the butter’s carbonyl band, respectively. Unfortunately, these images provide the same confusing information as the single wavenumber images of Fig. 4.20(a) and (b). As discussed earlier, second-derivative spectra can be used to minimize effects of baseline spectral fluctuations. Images prepared in the same manner as in Fig. 4.20 except using the negative of the second derivative of the butter–bacteria data are shown in Fig. 4.21. In this figure, the bacterial contamination region is correctly

107

MULTIVARIATE ANALYSIS OF IR IMAGE DATA

(a) 10

0.03

20

(b) 10

0.04 0.03

20 0.02

30 40

0.01

50 0

60 20

40

40

0.01

50

0

60

60

–0.01 20

(c ) 10

0.02

30

0.1

20

40

60

(d) 10

0

20

30

0.05

40 50

0

60 20

40

60

–0.02

30 40

–0.04

50

–0.06

60

–0.08 20

40

60

Figure 4.21 FTIR second-derivative spectroscopic imaging data of butter with a bacterial contamination in the lower right-hand corner: (a) wavenumber slice at amide I (1656 cm−1 ), (b) wavenumber slice of the butter carbonyl band (1720 cm−1 ), (c) amide I carbonyl band area image and (d) butter carbonyl band area image.

identified. The areas near coordinate (40, 55) in Fig. 4.21(a)–(d) indicate a relatively uniform quantity of butter and show little indication of bacteria. Four PCA score images of the unmodified butter–bacteria data are shown in Fig. 4.22. As with the artificial data, PCA score images can provide intriguing insight into the topology of the data with little or no knowledge of the spectral and spatial features of the data. Both images, Fig. 4.22(a) and (c), highlight regions where the bacteria are located. Unfortunately, as we have seen in previous examples, the raw unmodified data is dominated by pixel-to-pixel baseline fluctuations. As a consequence, these two score images, Fig. 4.22(a) and (c), group together some of the pixels containing only butter with the bacteria in the area near coordinate (40, 55). Therefore, this spectroscopic image data must be preprocessed to minimize the effects of the noise. The same four PCA score images are shown in Fig. 4.23 for the second-derivative spectra. A visual comparison shows an improvement in these preprocessed PCA score images. In Fig. 4.23(b), the troublesome region at coordinate (40, 55) contains only butter. This example illustrates the power of PCA for quickly surveying the data. One can imagine that by setting a threshold at 0.02 (the interface between blue and green in Fig. 4.23(b)), the second score image shown in Fig. 4.23(b) can be used clearly to show the location of the contaminant. We have only shown examples of selecting either raw, derivatized or transformed data parameters for defining different groups of a dataset. This selection process is

108

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a) 10

4

(b) 10

20

2

20

1

30

0.5

40

0

30

0

40 –2

50 60

–4 20

40

50

–0.5

60 20

60

(c)

1.5

10

0.5

20

0

40

60

(d) 10 0.5

20 30

30 –0.5

40 50

–1

0

40 50 60

60 20

40

60

–0.5 20

40

60

Figure 4.22 PCA scores of FTIR spectroscopic imaging data of butter with a bacterial contamination in the lower right-hand corner: (a) is the first PCA score image, (b) is the second PCA score image, (c) is the third PCA score image and (d) is the fourth PCA score image.

only the first part of the classification of pixels or spectra into groups. The next step involves defining a threshold that specifies the boundaries between the groups. An example of these boundaries is best illustrated by the positive and negative values for the pixels in Fig. 4.18(b). In that figure, the two groups are easily distinguishable because one group is positive and the other group is negative. Therefore, a convenient boundary or threshold would be zero, and in such a sample all pixels that contain positive values in this image are in one group and all pixels that contain negative pixels in this image are in the other group. Another type of classification is outlier selection or contamination identification. As an example, in Fig. 4.23(b), the butter is the desired material and bacteria the contamination. An arbitrary threshold for this image would be 0.02, in which all pixels >0.02 are considered suspect, and hopefully, because this is a food product, decontamination procedures are pursued. In these two examples of classification, only arbitrary thresholds have been defined and, as such, confidence in these classifications is lacking. This confidence can be achieved through statistical methods. Although this chapter is not the appropriate place for an involved discussion of application of statistics toward data analysis, we will give one example often used in chemometric classification. In the determination of bacterial contamination of butter, a comparison of all spectra in the dataset with the mean butter spectrum should provide generalized criteria for grouping each pixel into either the butter category or contaminant (bacteria)

109

MULTIVARIATE ANALYSIS OF IR IMAGE DATA

(a) 10

(b) 0.3

10

0.08

20

0.25

20

0.06

30

0.2

30

0.04

40

0.15

40

0.02

50

0.1

50

0

0.05

60 20

40

60

60

–0.02 20

(c)

40

60

(d)

0.02

10

0.03

10

20

0.02

20

0.01

30

0.01

30

0

40

0

40

–0.01

50

–0.01

50

–0.02

60

–0.02

60

20

40

60

–0.03 20

40

60

Figure 4.23 PCA scores of FTIR second-derivative spectroscopic imaging data of butter with a bacterial contamination in the lower right-hand corner: (a) is the first PCA score image, (b) is the second PCA score image, (c) is the third PCA score image and (d) is the fourth PCA score image.

category. Hotelling’s T2 test31,32 is perhaps the simplest and most straightforward of these kinds of comparisons. It provides a level of confidence with its comparison with the mean. The resulting statistical image of the values used in generating the PCA score image shown in Fig. 4.23(b) is shown in Fig. 4.24. The color bar next to the image indicates the correlation between the color and the value for each pixel. The arrows and values on the right are confidence limits. A user of this figure might assign a threshold statistic of 1.9, for example. Pixel values >1.9 would indicate contamination with a probability of error of 0.025 (red number). This statistical test is one of many and provides an example of how spectroscopic imaging data can be analyzed and used to group image pixels with a degree of confidence.

4.3.2 Concentration image maps The final goal of most image processing is to produce concentration maps of each component of the heterogeneous mixture. This is very difficult without at least a limited amount of preprocessing, which was discussed in Section 4.2. The idea of producing concentration maps is to extract pure component spectra from the multitude of spectra obtained with the image and then to find the concentration of each component across the image. These methods are generally referred to as self-modeling.1–19 Most of the currently accepted methods try to find good initial guesses for the spectra of the individual components and then refine these initial

110

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

6

10

5

20

4 30 3 <0.001 <0.005 <0.01 2 <0.025 <0.05 <0.1 1

40 50 60 10

20

30

40

50

60

Figure 4.24 Hotelling’s T image for the second score image of the negative of the bacteria/butter data. The color bar indicates the values of the colors in the image. The numbers adjacent to the arrows on the color bar are the confidence intervals for the statistical test.

guesses using the method of alternating least squares. We will first examine the methods for obtaining good initial guesses and then show how these are used in the alternating least squares model.14 The method referred to as SIMPLISMA (SIMPLe-to-use Interactive Selfmodeling Mixture Analysis) is probably the most widely accepted of the selfmodeling methods.10–13 For this method, absorbances at the wavenumber of the spectral feature exhibiting the greatest relative change in mixture spectra are assigned as the relative concentrations of one component. The spectral band at this frequency is referred to as first purity peak and its intensity is relative to the concentrations of the first pure component. The center of this peak is determined as that frequency from all of the frequencies in the spectra that has the maximum relative standard deviation in the mixture spectra. The relative standard deviation at each frequency is the standard deviation of the absorbances at each frequency divided by the average absorbance at that frequency. To avoid obtaining large relative standard deviations due to excessive noise at low absorbance regions (regions without bands), a small constant value indicative of the noise level is added to the average value in the denominator; we use 3% of the maximum average absorbance in the spectrum. After selecting the first purity peak, the spectra are multiplied by a weighting function, which removes or reduces absorbances at all frequencies that correlate with the first purity peak from all of the mixture spectra. Basically, this weighting function is determined from the correlation matrix; the higher the correlation between two wavenumbers, the lower the weight. The second purity peak is found again by finding the maximum relative standard deviation in the residual spectra. This processing is continued for all potential components. The procedure is interactive since the operator can override the purity peak selection.

Figure 4.25

1543

1000

1000

Bacteria

3000

3000

Butter

1543

1736

------------------After----------------------------------------------------

1000

1000

Infrared images and self-modeling spectra of components of butter–bacteria before (on the left) and after (on the right) FT filtering.

3000

Bacteria

3000

Butter

1736

--------------------Before----------------------------------------------------

112

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

The SIMPLISMA method was recently modified so that principal components (loading spectra) could be used instead of the original spectra.16,17 The modified method referred to as interactive principal component analysis (IPCA) consolidates the spectral information into few loadings and reduces the overall noise. It makes it somewhat easier to deal with noise in regions that lack absorptions. Otherwise, SIMPLISMA and IPCA produce very similar results. The relative absorbance values obtained by these self-modeling procedures are proportional to concentrations of the components in the mixtures and are used as the first estimates for concentrations. The method of alternating least squares14 is then applied to the data. In this method, the mixture spectra in the absorbance matrix, A, are written in terms of Beer’s law as A = CK

(4.8)

where C contains the estimated concentrations and K is the absorptivity matrix. The mixture spectra in the A matrix are regressed onto the concentrations to obtain the K matrix as K = (C T C)−1 C T A (4.9) Then A is regressed onto the new estimate of K to obtain a new estimate for C as C = AK T (KK T )−1

(4.10)

This alternating least squares processing is continued until some convergence criteria is met. We continue until the square root of the sum of all the residuals squared in the relation (Residuals = A − CK) changes by <0.0001. The effect of applying self-modeling to the bacteria–butter image is demonstrated in Fig. 4.25. Prior to Fourier transform filtering, we could not obtain a good spectrum of the bacteria. As can be seen from the figure there were a number of negative bands in the self-modeled spectrum of the bacteria due to the overremoval of the butter spectrum. In addition, there was a sloping baseline in this spectrum. As a consequence, the image of the bacteria was not clear. After filtering,30 the selfmodeling method provided a very good spectrum of bacteria with little or no baseline effects and no interference from butter. The image of bacteria after filtering clearly shows that the bacteria were located in the lower right corner of the field. Moreover, the image of butter following the processing shows that butter was equally distributed over the image field except for a thicker film at the top of the image. The problem of baseline interferences in self-modeling mixture analysis has been addressed recently by using a combination of conventional and second-derivative data in the SIMPLISMA method.13 In that approach, purity peaks could be obtained from either conventional or second-derivative spectra depending upon the spectral bandwidths, and the baselines were extracted as a separate component from the SIMPLISMA analysis. As mentioned earlier in the present report, the pixel-to-pixel variations produce many different shaped baselines, which cannot be accounted for by a single extracted baseline. It seems reasonable that second-derivative spectra could be used effectively to characterize the distribution of chemical components,

MULTIVARIATE ANALYSIS OF IR IMAGE DATA

113

but not for removing baselines to produce adequate pure component spectra from infrared hyperspectral images.

4.4 Conclusions In this chapter, we have tried to demonstrate some of the capabilities and pitfalls of chemometric processing. By no means have we covered all of the potential procedures for processing hyperspectral images, since this would require an entire book. The reader should be aware that there are numerous chemometric methods, which are potentially useful in processing spectral images. These include but are not limited to partial least squares (PLS),33 Mahalanobis distances,34 artificial neural networks,35 minimum entropy19 and fuzzy logic.36 Herein, we have tried to focus on those methods that find the most use at the present time. Other methods have been used and certainly many more will be used in the future. The reader may also want to consult image processing texts from areas such as computer graphics, printing and communication to find methods that have not been included under the umbrella of chemometrics; all of these areas deal with the same type of problems encountered in spectroscopic imaging data and, in many cases, commercialization has led to more rapid solutions than those put forth by spectroscopists and chemometricians. One final piece of advice: the product in data processing is only as good as the informational content of the input data. The field of study we now call chemometrics began in earnest about three decades ago with application of statistical principles to chemical data. Hyperspectral images provide lots and lots of data, and chemometrics can be used to extract important information from these data; however, it is still limited by the quality of the input data.

References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14]

Budevska, B. O., Sum, S. T. and Jones, T. J. (2003) Appl. Spectrosc. 57, 124–31. Lawton, W. H. and Sylvestre, E. A. (1971) Technometrics 13, 617. Knorr, F. J. and Futrell, J. H. (1979) Anal. Chem. 51, 1236. Malinowski, E. R. (1982) Anal. Chim. Acta 134, 129. Schostack, K. J. and Malinowski, E. R. (1989) Chemometr. Intell. Lab. Syst. 6, 21. Osten, D. W. and Kowalski, B. R. (1984) Anal. Chem. 56, 991. Tauler, R., Kowalski, B. R. and Fleming, S. (1993) Anal. Chem. 65, 2040–7. Vandeginste, B., Essers, R., Bosman, T., Reijnen, J. and Kateman, G. (1985) Anal. Chem. 57, 971. Gemperline, P. J. (1989) J. Chemometr. 3, 549. Windig, W., Lippert, J. L., Robbins, M. J., Kresinske, K. R., Twist, J. P. and Snyder, A. P. (1990) Chemometr. Intell. Lab. Syst. 9, 7. Windig, W. and Guilment, J. (1991) Anal. Chem. 62, 1425. Guilment, J., Markel, S. and Windig, W. (1994) Appl. Spectrosc. 48, 320. Windig, W., Antalek, B., Lippert, J. L., Batonneau, Y. and Bremard, C. (2002) Anal. Chem. 74, 1371–9. Andrew, J. J. and Hancewicz, T. M. (1998) Appl. Spectrosc. 52, 797.

114

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

[15] Vacque, V., Dupuy, N., Sombret, S., Huvenne, J. P. and Legrand, P. (1997) Appl. Spectrosc. 51, 407. [16] Brown, C. W., Bu, D., Camacho, N. P. and Mendelsohn, R. (2000) Spectral imaging: instrumentation, applications, and analysis. SPIE Technical Publication 3920 (ISSN 1017-2661), 118–28. [17] Bu, D. and Brown, C. W. (2000) Appl. Spectrosc. 54, 1214–21. [18] Schoonover, J. R., Marx, R. and Zhang, S. L. (2003) Appl. Spectrosc. 57, 154A–70A. [19] Chen, L., Chew, W. and Garland, W. (2003) Appl. Spectrosc. 57, 491–8. [20] Cooley, J. W. and Tukey, J. W. (1965) An algorithm for the machine calculation of complex Fourier series. Math. Comput. 19, 297–301. [21] Brown, C. W., Obremski, R. J. and Anderson, P. (1986) Appl. Spectrosc. 40, 734–42. [22] Brown, C. W., Bump, E. A. and Obremski, R. J. (1986) Appl. Spectrosc. 40, 1023–32. [23] Donahue, S. M., Brown, C. W. and Obremski, R. J. (1988) Appl. Spectrosc. 42, 353–6. [24] Donahue, S. M., Brown, C. W., Caputo, B. and Modell, M. D. (1988) Anal. Chem. 60, 1873–7. [25] Donahue, S. M. and Brown, C. W. (1991) Anal. Chem. 63, 980–5. [26] Savitzky, A. and Golay, M. J. E. (1964) Anal. Chem. 36, 1627. [27] Steinier, J., Termonia, Y. and Deltour, J. (1972) Anal. Chem. 44, 1906–9. [28] Madden, H. H. (1978) Anal. Chem. 50, 1383. [29] Griffiths, P. R. and deHaseth, J. A. (1986) Fourier Transform Infrared Spectrometry, WileyInterScience, New York. [30] Bu, D. Huffman, S. W., Seelenbinder, J. A. and Brown, C. W. (2005) Appl. Spectrosc. 9(5), 575. [31] De Maesschalck, R., Jouan-Rimbaud, D. and Massart, D. L. (2000) Chemometr. Intell. Lab. Syst. 50, 1–18. [32] Ott, R. L. (1992) An Introduction to Statistical Methods and Data Analysis, 4th edn, Duxbury Press, Belmont, CA. [33] Beebe, K. R. and Kowalski, B. R. (1987) Anal. Chem. 59, 1007A–10A. [34] Mark, H. L. and Tunnell, D. (1985) Anal. Chem. 57, 1449–56. [35] Sanchez, M. S., Swierenga, H., Sarabia, L. A., Derks, E. and Buydens, L. (1996) Chemometr. Intell. Lab. Syst. 33, 101–19. [36] Mansfield, J. R., Sowa, M. G., Scarth, G. B., Somorjai, R. L. and Mantsch, H. H. (1997) Anal. Chem. 69, 3370–4.

5

FTIR imaging of multicomponent polymers Jack L. Koenig

5.1 Introduction Properties of polymeric materials depend strongly on their structural organization. The performance of many industrial polymers, in particular, is determined by the microscopic morphology of the polymers. The object of most current chemical imaging research is to identify and characterize heterogeneity of polymers and polymer systems, so as to provide a better understanding of the mechanisms and controlling factors affecting the performance and durability of polymers. Furthermore, bulk morphology often controls mechanical properties, such as toughness, strength, wear and tear resistance. In order to optimize polymer performance, quick reliable methods of determining surface and bulk morphology are essential. In the past, electron microscopy – in particular TEM – has been the primary method of determining polymer morphology. However, the usefulness of electron microscopy has been limited by the destructive nature of the electron beam, the naturally poor contrast between polymer types and the difficulty of preparing (staining, etching, cryogenic microtoming, etc.) high quality specimens. Most structural materials are susceptible to a wide range of defects. Any flaw alters the behavior of a structure, even if only minutely. The larger the flaw the more it reduces the useful properties of the material. One of the challenges in modern materials engineering is defect reduction. Defect reduction involves defect detection, defect source determination and mechanisms and defect elimination. There is no single method of detect review that can fully characterize every defect; each defect classification method has its own strengths.

5.2 Imaging requirements for polymer characterization Most biomedical, forensic, materials and industrial applications of chemical imaging require: • • •

an imaging device with high chemical selectivity and specificity; an imaging device with easy to use image processing and manipulation software; a reliable, easy to use and easy to understand imaging device interface (in the case of imaging, a microscope);

116 • •

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

ability to improve the fidelity of the image by signal averaging; rapid image data acquisition that requires a high frame rate.

5.3 Polymer sampling for FTIR imaging Microsamples for Fourier transform infrared (FTIR) imaging are prepared for analysis by one of the following three methods: • • •

Transmission, where the IR beam passes through the sample; Reflection, where the IR beam reflects from the sample surface; and Attenuated total reflection (ATR), where the IR beam typically penetrates from 0.5 to 2.0 μm into the sample.

Each of these sampling analysis methods has advantages and disadvantages for FTIR imaging of polymeric systems as elucidated next.

5.3.1 Transmission measurements The most desirable sample for FTIR imaging is a film of the sample of thickness sufficiently thin to allow transmission measurements to be made through the film. In the case of polymers, this means that the samples can only be a few microns in thickness. The ideal thickness of samples depends on the molar absorptivity of the material. That is, a strong molecular absorptivity (like a carbonyl mode) requires a thinner sample than for one with a lesser absorptivity. An additional consideration for imaging, compared with single channel spectroscopy, dictates that the thickness of the film should be smaller than the average size of domains to give a meaningful transmission image of the distribution of these domains in the sample. Any film that is to be analyzed is in fact a three-dimensional construction with an internal component and an external component, the surface morphology. The internal structure is related to the chemical composition, inclusions and chemical morphology, and very often the sample is not homogeneous. This can be the nature of the sample (a thin film construction) or a defect or impurity that was generated when the material was made. Thin films for FTIR imaging can be prepared by a large number of methods including spin casting, chemical vapor deposition (CVD), physical vapor deposition (PVD), plasma etching, photolithographic patterning, electroplating, ion implantation, chemical–mechanical polishing, spin-on glass and rapid thermal processing.

5.3.1.1 Solvent casting A thin film for FTIR imaging can be prepared by solvent casting. Solvent casting requires a solvent that dissolves the sample preferably at room temperature, which can be easily removed (high vapor pressure) without formation of bubbles in the sample film. Solvents such as chloroform (boiling point (bp.) 61.2◦ C), acetone

FTIR IMAGING OF MULTICOMPONENT POLYMERS

117

(bp. 56.2◦ C), trichloroethanol (bp. 151◦ C) o-dichlorobenzene (bp. 180.5◦ C) and water (bp. 100◦ C) can be used. Several drops of the solution are placed on an inert infrared (IR)-transmitting substrate (such as KBr) and the solvent is evaporated to leave a film of the sample on the substrate. If the film is too thin, an additional deposition on top of the original can be made. In some cases, the film can be peeled off the substrate, and placed free-standing in a film holder in the spectrometer. The limitations of the solvent casting technique are residual solvent interfering with the spectrum of the sample and bubbles or nonuniform thickness of the cast film.

5.3.1.2 Spin coating Spin coating is a widely used technique for the preparation of thin polymer films on top of solid substrates for IR analysis. Depending on the used polymer–solvent combination, films with a very uniform thickness and small surface roughness can be obtained. Controlled by the preparation parameters, film thicknesses in a wide range from the monolayer regime up to several microns may be prepared. A solution containing dissolved polymer is sprayed onto a rotating substrate to ensure a complete and even coating. By spinning the substrate at high speed, the excess solution is ejected almost instantly leaving a thin film, which during the next few seconds continues to flow radially owing to the action of the centrifugal force. As the film thins down, the solvent evaporates and the viscosity increases to a point at which the film motion stops. The process is completed by evaporation of the remaining solvent. To prepare films of different thickness, the concentration of the polymer solution and the spinning rate are varied.1 The preparation of thin films by spin coating is used to obtain flat and homogeneous films with thicknesses down to the radius of gyration of the unperturbed molecules. Film thickness, which is determined mainly by the viscosity of the polymer solution and spinning speed, can range from tens of microns to well below a micron; however, it is difficult to deposit high quality films of submicron thickness by spin coating. The lower limit of film thickness for continuous spin-coated films of polymers ranges from 2 to 30 nm. Efforts to produce thinner films lead to discontinuous coatings. Controlling the surface roughness of the film can be critical for the final optical properties of the film. Solvent-rich films may have insufficient time to level and heal surface roughness due to Marangoni (surface-tension-gradient driven flow) instabilities, when the solvent is rapidly evaporating.2 Evaporation induces temperature gradients or leads to a local enrichment of polymer molecules leading to a higher surface tension in the surface layer. This causes an upward flow of solution from areas of lower surface tension. The upward flow induces a concentration gradient across the polymer film, which reduces its free energy by creating Benard cells (Benard instability), in which vertical circulation takes place. With slowly evaporating solvents, there is no Marangoni flow, internal leveling occurs and the films surfaces are smooth. The surface roughness increases with spin speed for any given solvent.

118

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

5.3.1.3 Microtoming Thin samples can be sliced from an appropriately mounted bulk sample. Smaller samples can be imbedded in a wax block. Automated sliding microtomes can be used for slicing. Thin samples are hard to handle and may curl. A microcompression cell can be used to hold the samples. Unfortunately, microtoming can introduce arifacts into the samples by tearing and orientation. 5.3.2 Reflection FTIR imaging measurements Of course, most samples for microspectroscopy are not accessible for transmission measurements, particularly in those circumstances where modification of the sample occurs by preparation of a thin film.

5.3.2.1 Reflection from ancient art works One particularly interesting example is the study of ancient art works where FTIR reflection is required to obtain information.3 First, removal of samples from the surface of the painting is not only undesirable but difficult as aged samples of paint are very brittle and crumbling occurs easily. A painting consists of several superimposed paint layers with a finishing varnish layer. The pigment particles in the binding medium can be identified in the reflection image. This method was applied to a paint cross-section of Rembrandt’s Portrait of a Standing Man (1639). FTIR imaging of this cross-section identified and localized different compounds present in the layers of this sample. Identification of these compounds based on their IR spectra was confirmed by results from art historical and conservation literature. Special attention was given to a discoloration that was observed in large parts of the described painting. This discoloration was clearly visible in the paint crosssection. The limitation of the use of reflection methods is that only the surface layer is measured. It is difficult to acquire spectra from lower paint layers. The manual separation of different layers is difficult because of the thinness of the individual paint layers which normally range from 1 to 50 μm. Fourier transform infrared microscopes are equipped with a reflection capability that can be used under these circumstances. External reflection spectroscopy (ERS) requires a flat, reflective surface, and the results are sensitive to the polarization of the incident beam as well as the angle of incidence. Additionally, the orientations of the electric dipoles in the films are important to the selection rules and the intensities of the reflected beam. In reflectance measurements, the spectra are a function of the dispersion in the refractive index and the spectra obtained are completely different from that obtained through a transmission measurement that is strongly influenced by the absorption index, k. However, a complex refractive index, n + ik can be determined through a well-known mathematical route, namely, the Kramers–Kronig analysis. Infrared reflection techniques are used for surface characterization because they provide highly detailed information about the molecular structure, orientation on

FTIR IMAGING OF MULTICOMPONENT POLYMERS

119

the surface and intermolecular interactions, they are nondestructive and are relatively easy to conduct as they do not require high vacuum systems. The surface-induced IR selection rule states that only vibrating dipoles with a nonzero component perpendicular to the substrate surface will be excited by p-polarized IR radiation. This provides a means for determining the average orientation of surface-confined molecules.4

5.3.2.2 Reflection measurements from metal surfaces Reflection FTIR imaging spectroscopy has been applied to the oil stains on sandblasted metals in order to establish micrometer thick oil stains of finite areas and sub-milligram masses. The integration of some reduced spectral feature (peak height, peak area, etc.) from such an image yields a calibration factor that becomes a basis for cleaning verification. Beyond obtaining useful calibration data, these measurements revealed stain migration dynamics that operated on time scales from seconds to years and that were dependant on the oil and the surface cleaning process. FTIR images were obtained in a time frame of minutes based on near normal reflectance measurements. The interaction of oil droplets with various substrates can be determined by FTIR imaging.5

5.3.3 ATR FTIR imaging Attenuated total reflection FTIR is a well-established technique for obtaining absorbance spectra of opaque samples. The mode of interaction is unique because the probing radiation is propagated in a high index-of-refraction internal reflection element (IRE). The radiation interacts with the material of interest, which is in close contact with the IRE, forming an interface across which a nonpropagating evanescent field penetrates the surface of the material of interest to a depth in the order of one wavelength of the radiation. The electric field at the interface penetrates the rarer medium in the form of an evanescent field whose amplitude decays exponentially with distance into the rarer medium. In the ATR experiment, total reflection of a light beam occurs at the interface between a medium with high refractive index (ATR crystal, n1 ) and one of lower refractive index (sample, n2 ). With the ATR technique, a crystal material of high refractive index is used as an IRE. The probe beam enters the IRE under an angle that exceeds the critical angle for total internal reflection. At the interface of the highly refractive IRE and the low refractive medium, the penetration of the electromagnetic field causes the formation of an evanescent wave propagating in all directions, decaying exponentially with distance from the surface into the bulk rarer medium. The evanescent wave decay can be represented by the following: E = E0 exp(−γ z)

(5.1)

120

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

where E0 is the electrical field strength at the surface and    2 1/2  n2 1 2 γ = 2n1 π sin θ − λ n1

(5.2)

where θ is the angle of incidence of the IR radiation. The penetration depth is in the order of the radiation wavelength. ATR sampling in FTIR imaging allows a wider variety of materials to be examined. For example, opaque or very thick samples that are not easily examined in transmission mode can be simply analyzed in the ATR mode. The ATR mode requires very little sample preparation: the sample must be flat in order to have good contact when pressing the crystal against the sample at a reasonable pressure. Attenuated total reflection has a controlled optical pathlength. The effective absorbance path length in the sample using the ATR method depends largely on the refractive index of the IRE crystal. With germanium, which has a refractive index of 4 (and assuming that the refractive index of the samples is ∼1.5 and that the incident angle is 45◦ ), the effective path length is ∼0.6 μm at 1100 cm−1 and ∼0.4 μm at 1600 cm−1 . Therefore, ATR images obtained reflect only the compositions on the surface layer of the ∼0.5 μm thick sample.6 In principle, it may be possible to construct three-dimensional images using the pathlength variability for different absorbance modes and different IREs. In addition, ATR microspectroscopy yields higher spatial resolution than transmission. The magnification factor is equal to the refractive index of the IRE. This capability arises from the fact that the sample is immersed in a medium of high refractive index. Consequently, the diffraction-limited spot size at the microscope’s focus is reduced, thereby increasing spatial resolution. The magnification factor associated with the refractive index (n) of the IRE allows one to work with aperture sizes that are n times greater, thereby eliminating diffraction effects introduced by small aperture dimensions.7,8 This corresponds to a 4× magnification factor associated with the germanium IRE. Experimentally, this translates into an increased spatial and volumetric resolution. The expected value for the spatial resolution, d, of the instrument: d = 1.22λ/n1 sin θ

(5.3)

Using a value of NA = 0.3, a wavelength of 6.25 μm (1650 cm−1 ) and a refractive index of 4.0, the resolution is obtained as 6.26 μm. Experimentally, the spatial resolution was determined to be 8 μm.9 The amount of sample being examined by ATR imaging in a single pixel can be calculated. The signal that one detector element is sensing arises from a sample volume element, which is ∼11 fl. This volume element is based on a spatial resolution of 8 μm, a penetration of 1 μm and a conical geometry for the evanescent beam present at the sample–IRE interface. Assuming a density of 1.2 g cm−3 , the volume elements translates into a mass of 13 fg. One disadvantage of ATR imaging is that the imaging area is relatively small,

FTIR IMAGING OF MULTICOMPONENT POLYMERS

121

usually of the order of 50 μm2 . A macro-ATR element can be used when larger samples areas are being studied. However, there is a distortion (elongation) of the image due to the optics of the illumination of the macro-ATR crystal. Mapping surfaces with the ATR objective is more complicated than transmission or external reflection IR because the stage must be moved in the z-direction downward away from the IRE, then translated in the x or y-direction to a new position to before being raised again to bring the sample back in contact with the IRE. An electronic pressure measuring device can be used to achieve the same pressure for each measurement. The contact between the sample and the ATR crystal is very important in ATR/FTIR spectroscopy.10 This problem becomes a critical issue for imaging applications. The absorbance variation shown in an ATR image is strongly influenced by the quality of the contact made between the crystal and the sample. If the sample has poor contact with the crystal, the absorbance will be much weaker than when there is good contact. This can be the case for hard or non-easily deformed samples or materials with rough surfaces. It may be possible to obtain uniform contact between all areas of the sample and the crystal by applying a relatively high but known pressure using a torque wrench. For a sample with a rough surface, if some areas make good contact while others are poor, the generated image could be very misleading, because one may think that an image was showing a heterogeneous distribution of a substance, but actually it was an image of ‘contact quality’.11 Because the diamond ATR allows one to apply a high contact pressure without damaging the crystal, it can reduce image artifacts due to possible poor contact of a sample with large contact area. Attenuated total reflection microspectroscopy has been substantially improved through the use of a shaped germanium hemispherical IRE. This cell design permits only one reflection to take place as the light traverses through the crystal. Half of a microscope objective is employed to introduce light into the IRE and the other half is employed to collect the exiting light. ATR images based on different absorption bands at different wavelengths can exhibit different results due to differences in penetration. The effective pathlength of the ATR method depends largely on the refractive index of the crystal. With germanium, which has a refractive index of 4 (assuming that the refractive index of the samples is ∼1.5 and the incident angle is 45◦ ), the effective pathlength is ∼0.6 μm at wavenumber 1100 cm−1 and ∼0.4 μm at wavenumber 1600 cm−1 . Therefore, the image obtained will reflect only the compositions of the surface layer of the sample. The good news is that by using ATR materials with different refractive indexes resulting in different depths of penetration holds the possibility of constructing three-dimensional FTIR images using the images of the different layers.

5.4 FTIR image analysis When a focal plane array (FPA) with a moderate size of 64 × 64 elements is used, the resulting spectral cube consists of 4096 spectra. This massive amount of data is

122

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

acquired in a very short time. Converting the data into chemically and physically significant information is the task of the analysis. A discussion of some of the complex data processing strategies is discussed in Chapter 4; here, we provide an overview of the common procedures employed in the analysis of polymeric systems. Once the FTIR image has been collected, preprocessed, and processed, there are a series of steps that are followed to interpret the data. Initially, the acquired image should be inspected visually for the presence of artifacts. An artifact is any feature in the image that does not correspond to any property of the sample. Any recognizable artifacts should be removed before image analysis.

5.4.1 Selection of characteristic spectral stains for each component For determination of concentration profiles in images, one of the most important issues is the selection of IR frequencies for the individual components at which the absorbance is measured and analyzed. Therefore, it is necessary to select a component optimized spectral region based on the knowledge of the spectra of the components being measured. Images may be thresholded for maximum contrast between components or between values for comparison across a set of images. Thresholding is affected by nonuniform thickness effects, which may lead to erroneous conclusions about a species’ distribution. Hence, peak height (or area) ratios are used to normalize thickness effects and relate the observed absorbance to species’ concentration. With the FTIR method, we want to utilize all the 4096 data points. One approach is to generate a two-dimensional plot of the species absorbances12 (Fig. 5.1(a)). This type of representation plots the absorbance of each peak against the other for the same pixel, generating a point on the scatter plot. In this manner, a large number of points (equal to the number of pixels in the image) are plotted on the graph. On the basis of the distribution of chemical species seen on the two-dimensional plot (Fig. 5.1(c)), a color composite may be prepared to determine to which phase the scatter points correspond. This composite (Fig. 5.1(b)) indicates that the red points, in this case, correspond to the matrix phase (termed phase I) and the blue points correspond to the droplet phase (termed phase II). Then, the average absorbance of each phase can be calculated on the basis of the average of all points in the color-coded areas delineated by the ellipses, as shown.

5.4.2 Construction of contour plots Once the characteristic absorbance frequencies for the components have been determined, one can view the image data with contour plots (intensity versus spatial position). A contour plot for a given frequency is a representation of the data in three dimensions: spectral intensity (shown by colors), and plotted in the x- and ydimensions. From the contour map, one can determine initially whether the acquired image is homogeneous or heterogeneous. If the image is homogenous and is expected

123

FTIR IMAGING OF MULTICOMPONENT POLYMERS

(a)

(c)

(b)

Absorbance (CN)

1.0 0.8 0.6 0.4 0.2 0.0 0.0

0.1

0.2

0.3

0.4

0.5

Absorbance (C:CH) Figure 5.1 (a) FTIR image of a PB–E7 blend. (b) False color composite of absorbance distribution for the blend image. (c) Two-dimensional scatter plot of the absorbance of a PB specific peak (X-axis) against an E7 specific peak. Reproduced from figure 1 of Ref. 12, with permission.

to be heterogeneous, it is desirable to examine the image acquisition process to establish greater spectral contrast in the image reflecting the expected heterogeneity. The next step in the image analysis process is to search for unique chemical or morphological features in the images. This is accomplished by retrieving representative spectra from distinct spatial regions and evaluating these spectra. Then one should create a chemically specific image profile using the information obtained from the spectral interpretation of the chemistry or structural components. Finally, one interprets the chemically-specific image profiles and relates them to the spatial structure of the sample.

5.4.3 Histograms A histogram is a bar chart that shows the possible values for a pixel and the corresponding number of pixel that has that value in an image. Histograms can be used to show characteristics about the image. Taking these characteristics, changes can be made to each pixel in order to have an image conform to a desired histogram. Histograms can represent the distribution of absorbances in images where each histogram bin represents an absorbance value in the image. The histogram of an image is a graph of the number of pixels at each of the intensities represented in the digitized image. A histogram analysis of the image will display the complete pixel distribution across the gray scale. The histogram represents a high-dimensional distribution (atleast 32 or 64 gray scale bins). A histogram of the absorbances contains no spatial information; it only shows the distribution of the absorbances in terms of the total number for each value.

124

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Histograms have many uses. One of the more common is to decide what value of threshold to use when converting a gray scale image to a binary one by thresholding. If the image is suitable for thresholding then the histogram will be bimodal – that is, the pixel intensities will be clustered around two well-separated values. A suitable threshold for separating these two groups will be found somewhere in between the two peaks in the histogram. If the distribution is unimodal it is unlikely that a good segmentation can be produced by thresholding. Histograms can be used for thresholding images, which assumes that images are composed of regions with two different gray level ranges; one set of grey level values is associated with the background, and the other higher set of values with the absorbance of the chemical components in the image. When histogram equalization is applied to the image, valuable detail separates from the background and becomes visible. However, under circumstances where the histogram of the image is a single broad band, other approaches must be used. Histograms are used in dealing with statistical data and distributions. Histograms can be compared using statistical tests as well as using vector norms.13 The combination of digital imaging and spectroscopy provides a chemical basis for a fast, easy and noninvasive method for analyzing the homogeneity of a dispersion of one component, for example, a drug in a polymer matrix. The distribution of the drug can be evaluated visually, as shown in Fig. 5.2. Quantitative data is then produced using the image analysis software, where the absorbance value of the drug from each detector element is extracted. The data is graphed as a histogram showing the number of pixels that register each absorbance value. Various attributes, such as the shape of the histogram, can be used to further analyze the homogeneity of the dispersion. The system used here is a two-component system consisting of testosterone in PEO. Testosterone images collected C=O stretch at 1657 cm−1 were used to examine the distribution and cluster size of the testosterone as a function of initial drug concentration.14 A qualitative evaluation of the drug images shows cluster size variations as well as dispersion heterogeneities. For each image, a histogram 10% THP

20% THP

30% THP

40% THP

50% THP

20x

THP

Figure 5.2 Optical (upper set) and characteristic IR images (lower set) of testosterone in poly(ethylene oxide) (PEO).

125

FTIR IMAGING OF MULTICOMPONENT POLYMERS (a) 300

(b)

10%

20% 200

Number of pixels

Number of pixels

250 200 150 100

150

100

50

50 0

0 0.0

0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.18 Absorbance (a.u.) (c) 180

0.4

0.6

0.8

Absorbance (a.u.) (d)

30%

250

160 140

40%

200 Number of pixels

Number of pixels

0.2

120 100 80 60 40

150 100 50

20 0

0 0.0

0.2 0.4 0.6 Absorbance (a.u.)

0.8

0.0

0.2

0.4

0.6

0.8

Absorbance (a.u.)

Figure 5.3 Histograms of testosterone distribution in samples containing (a) 10%, (b) 20%, (c) 30% and (d) 40% testosterone. The width as well as the symmetry of the histograms indicate the homogeneity of the drug distribution. Reproduced from figure 3 of Ref. 12, with permission.

was generated from absorbance values of the drug band registered at each detector element, as shown in Fig. 5.3. An ideal distribution would be a narrow curve, indicating a large amount of detector elements recording the same absorbance (concentration) value. Conversely, a curve with a large breadth would indicate a wide range of concentration values, leading to a high degree of heterogeneity. Tailing at values either above or below the main curve indicate nonuniform distribution. Significant tailing is seen at all drug concentrations, a cluster of lower than average values at 10% and 40%, and higher than average at 20% and 30%. Although a visual inspection of the histogram accompanied by the drug image may sufficiently characterize the distribution, qualitative data can also be extracted. The range in absorbance units of the histogram indicates the breadth of the distribution, with narrow and wide ranges indicating small and large distributions, respectively. The breadth of the distribution increased with an increase in drug concentration. A comparison of the mean and median absorbance values demonstrated the symmetry of the histogram. If the values are equal, the curve would be completely symmetrical, while differences in the two values would indicate a lopsided distribution. With the exception of 40% testosterone, the difference between the average and mean increased with concentration, with

126

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

the maximum difference being 0.0230 a.u. at 30% testosterone. Also noted was the absorbance value registered over the highest number of detector elements, which increased with increasing drug loading. The most populated absorbance value increased as the number of detector elements at that population decreased as a function of drug loading, with the exception of 40% testosterone. These results show that an increase in drug concentration tends to increase the heterogeneity of the matrix. The data could also be compared to pre-established uniformity limits that could accept or reject the tablet based on the amount of deviation.

5.5 Applications of FTIR imaging to complex polymer systems 5.5.1 FTIR imaging of polymer laminate films Multilayer polymer films can be formed with coextrusion. Most commercial films contain seven or fewer layers, although layer-multiplication technology can be used to create films with thousands of layers. Each layer provides a desired characteristic (e.g. tensile strength, permeability control, adhesion), but in many cases an additive is blended into one or more layers to confer other properties to the films. The first reported application of FTIR microspectroscopic imaging to polymeric materials was the imaging of a cross-section of a laminated polymer film. The adhesive layer between two polymers was identified.15 A three-layered film composed of poly (ethylene terephthalate), poly(ethylene-co-vinyl alcohol) and low density polyethylene was studied using this technique. By plotting a specific frequency corresponding to each layer, chemical images were generated which demonstrated the degree of segregation of the three components in the film. In another report, an FTIR of a cross-section of a laminated polymer film showed that the adhesive layer (7.5 μm thick) could be isolated.16 The use of FTIR imaging with an FPA detector has been used for the characterization of multilayer laminates with layers of polyolefin, polyamide, ethylvinyl alcohol and ethylvinyl acetate.17 High contrast was observed by using IR bands associated specifically with each polymer layer. A three-layer film laminate comprising an ethylene–vinyl acetate (EVA) copolymer layer sandwiched between layers of a polyethylene (PE) and a polyester was studied. In the following Fig. 5.4, a schematic diagram of the laminate is shown. An FTIR absorbance intensity gray-scale image for each component based on specific bands at 2850, 2980 and 1730 cm−1 was used. FTIR absorbance spectra from a single pixel within each of the three layers are also shown.18

5.5.2 Chemical morphology of multi-component polymeric materials Paints, adhesives and lubricants are typically multicomponent polymer systems. The behavior of phase-separated blends in the bulk after quenching into the unstable region of the phase diagram is variable. In the bulk, the concentration fluctuations

FTIR IMAGING OF MULTICOMPONENT POLYMERS (a)

PE

EVA

127

Polyester

(b)

2850 cm–1

1730 cm–1 Polyester

EVA

1000 1200 1400 1600 1800 2000 Wavenumber/cm–1

Absorbance

Absorbance

PE

Absorbance

(c)

2980 cm–1

1000 1200 1400 1600 1800 2000 Wavenumber/cm–1

1000 1200 1400 1600 1800 2000 Wavenumber/cm–1

Figure 5.4 (a) Schematic of a three-layer polymer laminate. (b) FTIR absorbance intensity gray scale at ∼2850, 2980 and 1730 cm−1 at spectral resolution of 16 cm−1 . (c) FTIR absorbance spectra from a single pixel within each of the three layers. Reproduced from figure 1 of Ref. 18, with permission.

that govern the phase-separation process are random. As a result, the final morphology consists of mutually interconnected domain structures rich in a given blend component that coarsen slowly with time. Multicomponent polymer systems consist of morphologically distinguishable objects, such as particles dispersed in a continuous matrix. Discrete objects can be described by their morphometry, for example, size, shape, center of gravity, surface area, volume, orientation, etc. Polymers are industrial material systems with a variety of uses and it is important to understand the evolution of their microstructure and morphology. Conventional optical microscopy can be used to determine the positions of objects in such samples when those objects are separated by distances greater than several hundred nanometers, as restricted by the diffraction limit of light. However, optical microscopy is limited for these systems by the differences in the refractive indexes. In contrast (no pun intended), polymer blends are examples where the IR images yield valuable new morphological information. The spectroscopic images allow the measurement of the size, shape and distribution of the morphological units in the continuous matrix. Additionally, it is possible to determine chemical composition within the morphological objects using the FTIR spectra from pixels in the object and the surrounding matrix. The accuracy of using imaging to measure morphological size distributions in multicomponent samples depends on several factors. First and for most, the effective

128

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

spatial resolution of the instrument relative to the size of the morphological objects is important. For FTIR imaging, the primary limitation is the effective spatial resolution of ∼10 μm for morphological objects. The ability to accurately recognize the location of the boundaries between adjacent objects can be a source of measurement error. With FTIR imaging, changes in the refractive index results in a measurable discontinuity which can be detected.19 The determination of the spatial variation of composition of a sample is accomplished by obtaining spectra at successive adjacent positions in the sample and measuring the intensities of the unique IR lines of each component of the mixture. The images can be used to determine the number of morphological objects, the area of the objects as well as the shape and orientation within the objects. Histograms generated from different sections of the same sample can be used to yield information about the distribution of the various parameters. The determination of the temporal changes in composition can be made by monitoring the spectral changes for a given spatial position in the sample. A tedious process is currently used to determine the phase diagram of binary polymeric systems. A single concentration is cycled through a temperature range and the onset of a refractive index difference (scattering or microscopy) is observed. Hence, a number of measurements are required at small temperature differences and accuracy depends on time spent at the experiment. This procedure is then repeated for many different concentrations to determine the phase diagram. Small refractive index (which may be temperature dependent!) differences between the components may complicate analysis. Moreover, the phase boundaries for many systems are accessible only at elevated temperatures leading to chances of degradation in some systems. Finally, the phase diagram may correspond to transitions in one (or both) components leading to a very complicated situation. The basic issue in such systems is the determination of phase composition as a function of temperature. It must also be realized that each starting composition in the single-phase region is capable of yielding information about two phases in the two-phase region. The goal of determining the composition of each phase is readily achieved using FTIR imaging coupled to thermal control of the sample.20,21 A single composition is held in the two-phase region and quenched to different temperatures in the two-phase region. A single step is required. Once hydrodynamic and coarsening equilibrium is attained for a few minutes, the sample may be imaged. Images are acquired continuously over time and the data is used from an image only when two consecutive images were found to not change. The agglomerates smaller than the spatial resolution are not resolved in the images, but can be detected in the extracted FTIR spectra from pixels in the image. Subsequent quantitative analysis using the FTIR spectra yields the phase composition of the two phases at the temperature of quench. Thus, the two points of the coexistence curve at the temperature of quench are simultaneously determined. Now, the sample can be subjected to higher temperatures and allowed to equilibrate at the higher temperature and the phase diagram determined by sequentially stepping back to the single-phase temperature. Deeper thermal quenches are not used to avoid secondary phase separation. Alternately, multiple

FTIR IMAGING OF MULTICOMPONENT POLYMERS

129

quenches to different temperatures in the two-phase region may be carried out from the single phase. A calibration curve was independently determined by starting from high temperatures. The average spectrum at each temperature was taken to indicate phase composition. The phase diagram obtained corresponds well with the phase diagram obtained using optical microscopy following the classical methods of obtaining phase diagrams. It was demonstrated that, using this approach, phase diagrams are readily obtained in a fast, straightforward manner, that is, decoupled from the kinetics of the phase-separation process.

5.5.2.1 Polymer/liquid crystalline dispersions Polymer dispersed liquid crystals (PDLCs) are composed of low molecular weight liquid crystal (LC) microdomains in a polymeric matrix. Devices incorporating PDLCs have a wide variety of applications in the electrooptical industry and as light modulators in windows and displays, mainly due to potential difference modulated, reversible light scattering by dispersed LC domains. PDLC films are prepared by inducing phase separation in an initially homogeneous mixture of an LC and a polymer or polymer precursor. Generally, they are fabricated by phase-separation methods (thermally induced phase separation, solvent evaporation-induced phase separation and polymerization-induced phase separation (PIPS)). The morphology of the LC domains in the polymer matrix depends largely on the conditions during the phase-separation process and affects the electrooptical properties significantly. Among the various methods of inducing phase separation, PIPS appears to be the most attractive route to prepare PDLCs. Thermally curable systems are expected to have a high LC solubility in the matrix due to curing at high temperatures. The polymerization process, solubility and size of the dispersed phase are inseparably linked in such systems. The LC droplets in PDLCs are formed by the diffusion and coalescence of LCs during the solidification of the polymer matrix. In fast polymerization, small LC domains are observed owing to early freezing of the LC domains that are formed at the beginning of the polymerization. In slow polymerization, large LC domains are formed, owing to the sufficient time for LC diffusion. A typical UCST curve for a curing oligomer–LC mixture is shown in Fig. 5.5. The solid lines represent the state prior to curing, and the dashed lines represent the postcuring state. The coexistence curves in such phase diagrams define onephase and two-phase regions. Upon polymerization, the coexistence curve shifts toward higher temperatures in the direction of the arrow to the curve close to matrix gelation shown as a dashed curve. Clearly, there are different regions of interest in temperature–composition space that may influence formation and the ultimate properties of the PDLC. A mixture in region A will not phase separate even after extensive polymerization. Starting from a homogeneous mixture in region B (between the coexistence curves), one obtains a phase-separated PDLC system after polymerization due to the coexistence curve now being the dashed one for the polymerized matrix–LC system. The very appearance of two phase-separated dispersions implies

130

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS T

A

Tp1

B

Tp2 C

Tg,m fM2

fM1

f fD1 fD2

Figure 5.5 Generic phase diagram showing the coexistence curve and solubility for uncured oligomers– LC (solid line) and curing PDLC (dashed line). Taken form Ref. 20, with permission.

that the timescale for phase separation is faster than that for polymerization to gelation. The composition of the phase-separated domains at any temperature can be given by tie lines to the coexistence curve at that temperature (see Fig. 5.5). For polymerizing at temperature Tp1 and a lower temperature, Tp2 , compositions of the two formed phases can be seen. A large difference results in compositions of the two phases formed after curing simply by lowering the temperature. Significantly, lower amounts of LC will remain dissolved in the matrix after polymerization at the lower temperature and subsequent use at higher temperatures. Residual solubility of matrix material is also expected to be lower. On the basis of these observations, a new route to produce PDLCs resulting in lower matrix has been suggested and tested using FTIR imaging.22 The new process, termed TICS (thermally induced curing stabilization) phase separation, involves cooling the nonpolymerized matrix–LC mixture to the two-phase region (e.g. to temperature Tp2 in Fig. 5.5), allow sufficient time for phase separation, and subsequently polymerize the dispersion rapidly at the lowered temperature. The polymerization freezes in the formed structures, thereby maintaining solubility at levels below that obtained from carrying out the process at higher temperatures. A well-studied system (photocurable, thermosetting matrix, NOA65 and LC E7) which has high LC solubility was chosen as a model system to examine the route for PDLC formation by TICS phase separation. Infrared absorption bands that correspond to the reacting species can be readily identified and tracked in NOA65. The sulfur–hydrogen stretching (∼2570 cm−1 ) was used to track the curing of NOA65. IR images of representative samples, obtained at different temperatures for increasing times of rest prior to polymerization, are seen in Figs 5.6 and 5.7. Figure 5.6 displays a set of images for samples cured after a prepolymerization rest time from 21 to 8 h at 270 K. Figure 5.7 displays the set for same time intervals at a rest temperature of 255 K. The average droplet diameter is seen to monotonically increase and the droplet density to decrease with rest time in both cases. This is possibly the result of increased coalescence. Droplet

131

FTIR IMAGING OF MULTICOMPONENT POLYMERS (a)

(b)

100 μm (d)

0.06 (e)

(c)

0.45

100 μm

(f)

Figure 5.6 IR images of PDLCs formed by curing at 270 K after a prepolymerization cooling time of (a) 30 min (b) 1 h (c) 2 h (d) 4 h (e) 6 h and (f ) 8 h. Taken from Ref. 20, with permission.

(a)

(b)

100 μm (d)

0.02 (e)

(c)

0.70

100 μm

(f)

Figure 5.7 IR images of PDLCs formed by curing at 255 K after a prepolymerization cooling time of (a) 30 min (b) 1 h (c) 2 h, (d) 4 h (e) 6 h and (f ) 8 h. Taken from Ref. 20, with permission.

132

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

density and average droplet size are both greater for samples at 255 K. This points to reduced LC miscibility. Domain size distribution is also narrower for samples at 255 K. Qualitatively, the longer the standing time, the larger was the domain size. Polydispersity in sizes is smaller for shorter cooling times. The FTIR images are plotted as a ratio of the absorbance of a matrix-specific band (OH or SH) to the absorbance of the LC-specific nitrile group stretching peak. Using mixtures of the two components in known amounts, composition ratio curves were determined for the amount of matrix in the LC phase and the LC dissolved in the matrix phase. This allowed a determination of the composition of the LC and matrix separated phases.

5.5.3 Immiscible polymer blends An application of FTIR to imaging has been the study of domain sizes in chemically separated polymer blends.23 Optical microscopy of these systems is particularly challenging as the resultant phases have similar refractive indices. Good image contrast was achieved by FTIR imaging due to the inherent spectral differences arising from the characteristic absorbance bands of the components. Polymer blends provide an important route to property combinations not generally available in a single polymeric material.24 The attractiveness of blending lies in the property tailoring possibilities. However, most binary mixtures of polymers are not miscible on the molecular level because the entropy of mixing is not favorable for high molecular weight polymers. As most polymer pairs are immiscible, they form multiphase systems with weak physical and chemical interactions across the phase boundaries. The kinetics of phase segregation in immiscible polymer blends is very complex and involves various mechanisms that include diffusion, nucleation, growing and variation of size and form of the dispersed domains. This is the case, for instance, in binary mixtures in the neighborhood of their mixing–demixing temperature. Immiscibility of polymers in the melt is a common phenomenon, typically leading to a two-phase random morphology. If the phase separation occurs by a spinodal decomposition process, it is possible to control the kinetics in a manner that leads to multiphase polymeric materials with a variety of co-continuous structures. Common morphologies of polymer blends include droplet, fiber, lamellar (layered) and co-continuous microstructures. The distinguishing feature of co-continuous morphologies is the mutual interpenetration of the two phases and an image analysis technique using TEM has been described for co-continuous evaluation.25 Poly(vinyl chloride) (PVC) is a useful polymer that is becoming less popular due to its degradation behavior. However, blends of PVC with some other polymers have the capability to yield composites with greater stability. In one such case, images of phase-separated PVC–poly(methyl methacrylate) (PMMA) blends are reported.26 FTIR images representative of PVC or PMMA can be readily obtained. The PVC and PMMA FTIR images in Fig. 5.8 show exact phase inversion, indicating phase separation of polymers in the bulk for all PMMA molecular weights.

133

FTIR IMAGING OF MULTICOMPONENT POLYMERS

100 mm

0.15

0.0

0.20

PMMA, 1330 cm–1

PVC, 1366 cm–1

0.0

Figure 5.8 Images from a PMMA–PVC blend. These images show that both components are highly segregated. Taken from Ref. 26, with permission.

The behavior of chemical phase-separated blends in the bulk after thermal quenching into the unstable region of the phase diagram is variable. In the bulk, the concentration fluctuations that govern the phase-separation process are random. As a result, the final morphology consists of mutually interconnected domain structures rich in a given blend component that coarsen slowly with time. Phase-separated mixtures of uncured poly(butadiene) (PBD) and diallyl phthalate (DAP) were studied to characterize morphology differences before and after the curing process. These blends exhibit UCST behavior, and the composition of the domains reflects the phase-separation behavior. The changes in domain sizes and composition with time (prior to curing) are driven by minimization of the interfacial area. Upon heating, the phase-separated blend above the UCST for purposes of curing, a homogeneous cured sample results. The differences in absorbance intensities in the C−H stretching (2950 cm−1 ) and C=O stretching (1730 cm−1 ) from PBD and C−O ester stretching modes (1280 cm−1 ) from DAP were used as probes for FTIR imaging contrasts. Comparison of the blend morphologies can be made by comparing the chemical images shown as Figs 5.9 (immediately after preparation) and 5.10 (24 h after preparation). The image analysis results are shown in Fig. 5.11 The average diameters of DAP domains were calculated from the measured area (A). An increase in particle sizes and a decrease in the number of particles were observed on aging. The driving force of the growth in particle sizes is the minimization of interfacial area. When the system is sufficiently mobile, the particle sizes grow to reduce the free energy of the system. Calibration curves were constructed and the chemical composition of the domains were measured. The measured value is 24.9 wt% ± 3.5 in the PBD matrix and 91.3 wt% ± 4.3 in the DAP domain before aging. After aging of the sample at room temperature, the DAP concentration dropped to a value of 17.8 wt% ± 4.1 in PBD matrix and increased to 88.3 wt% ± 4.9 in the DAP domain. The high concentration of DAP in the PBD phase before aging was thought to be due to submicrometer particles of DAP in PBD. The decrease in DAP concentration in the DAP domain on aging arises from the inclusion of submicrometer PBD particles into the DAP domain during the aging process.

134 (a)

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

250 mm

0.2 0.7 C9H stretching image 2950 cm–1

(b)

0.7 1.8 C:O stretching image 1730 cm–1

(c)

0.7 1.8 C9O ester bending image 1280 cm–1

Figure 5.9 IR spectroscopic images of phase separation of PBD–DAP blends (50 wt.% DAP) taken immediately after sample preparation. Reproduced from figure 2 of Ref. 27, with permission.

(a)

250 mm

0.1 0.8 C9H stretching image 2950 cm–1

(b)

0.6 1.1 C:O stretching image 1730 cm–1

(c)

0.6 1.1 C9O ester stretching image 1280 cm–1

Figure 5.10 IR spectroscopic images of phase separation of PBD–DAP blends (50 wt.% DAP) taken 24 h after sample preparation. The images are reported (a) at 2950, (b) 1730 and (c) 1280 cm−1 . DAP domains in the PED matrix and interfacial regions are observed. Reproduced from figure 2 of Ref. 27, with permission.

The morphological changes were characterized over a period of time and the postcure sample exhibited homogeneity at the resolution of the instrument. By constructing Beer’s law calibration plots, the amount of each component that is present in the two phases was accurately determined. Additionally, this chapter demonstrated the application of standard image processing routines to the analysis of chemical images. The number and size of domains in the blend were monitored over time, and it was found that the coalescence and ripening of the domains could be quantified using these routines. Similar FTIR imaging studies have been reported for PBD–zinc diacrylate blends27 and polyisoprene–DAP blends.28

FTIR IMAGING OF MULTICOMPONENT POLYMERS Average diameters of DAP particles (0 h after sample preperation) 150

(b)

Average diameter (mm)

Average diameter (mm)

(a)

100 50 0 1 8 15 22 29 36 43 50 57 64 71 78 Particle number

135

Average diameters of DAP particles (24 h after sample preperation) 150 100 50 0 1

8 15 22 29 36 43 50 57 64 71 78 Particle number

Figure 5.11 Particle size distribution by image analysis: (a) immediately after sample preparation (b) 24 h after sample preparation. The increase in particle size and the decrease in particle numbers are observed. Volume fraction of DAP increased from (a) 15.6 to (b) 43.5 vol.% DAP. The increase in volume fraction results from the coalescence of submicron particles. Reproduced from Ref. 27, with permission.

5.5.3.1 Polymer liquid crystalline blends Polymer dispersed LCs are composites of polymers with low molar mass LCs. They consist of small (1–20 μm) LC-rich domains embedded in a polymeric matrix. The polymer is chosen to have a refractive index as close to one of the refractive indices of the anisotropic LC. It is well known that the electrooptical properties of LCs (or liquid crystalline domains) may be controlled by the application of an electric potential across the film. When an electric field is applied, the LC orients with respect to (parallel or antiparallel) the direction of the field. Thus, in the oriented state, the refractive index of the LC and that of the polymer are matched, and little scattering occurs at the phase boundaries. This feature leads to the film becoming almost optically transparent. On the other hand, when the field is removed, the LC domains relax back to having randomly oriented directors, resulting in a mismatch of refractive indices at the phase boundaries. This condition results in the film becoming highly scattered or optically opaque. The switching behavior of each domain depends on its size and shape. The electrooptical properties of the liquid crystalline domain depend on the type of LC and the amount of matrix material dissolved in the LC. Both these aspects of formation may be determined by using a single technique – FTIR imaging. The evolution of domains for the PB/E7 is observed by using FTIR imaging (Fig. 5.12).29 The indicated times for each figure denote the time elapsed since the onset of phase separation was visually observed and image collection was started (t = 0). The imaging technique is able to capture the evolution of domains and their growth by coalescence.

5.5.4 Crosslinking-induced phase separation of elastomers The competition between the phase separation and crosslinking processes opens up the possibility of controlling pattern formation and the structure of the resulting material by varying the reaction temperature and the blend composition. There is considerable effort to modify the useful properties of elastomers by introducing a coagent that is copolymerized with the elastomer and contributes positively

136

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

t =1 min

t = 5 min

t =12 min

t = 20 min

002

035 100 mm

t = 30 min

t = 50 min

t =100 min

Figure 5.12 FTIR images of the evolution of morphology in the PB–E7 blend. Reproduced from figure 1 of Ref. 29, with permission.

to the performance properties of the elastomer. The macroscopic properties are determined by microscopic and molecular factors, such as the degree of phase separation, the domain sizes, interfacial interactions and the composition of the different phases. The FTIR imaging technique was used to study the spatial analysis of the domains in the blends of PBD, zinc diacrylate (ZDA), cured with dicumyl peroxide (DCP).30 ZDA is a metallic salt of acrylic acid which is a crosslinking coagent in the peroxide curing of elastomers. It has two terminal double bonds that are reactive in the presence of free radicals and readily grafts to elastomer chains to form a complex crosslinked network. ZDA exists as phase-separated forms in the networks because it is a solid at curing temperatures. The chemical crosslinking occurs both within ZDA agglomerates and across the ZDA–PBD interface. The distribution of the ZDA coagent in the PBD matrix was monitored before and after curing. ZDA exists as agglomerates of different sizes because of its polar property. The inhomogeneous distribution of ZDA in the blend was observed with FTIR imaging using the CH rocking band (1270 cm−1 ) of ZDA (Fig. 5.13(a)) or integrating the CH stretching region of 3115 cm−1 to 2810 cm−1 (Fig. 5.13(b)). The PBD-rich matrix has higher intensity in CH stretching region and lower intensity at the 1270 cm−1 band than the ZDA-rich domains. The agglomerates smaller than the spatial resolution are not resolved in the images, but are detected in the extracted FTIR spectra from the image. The spectra from each phase show the distinct peaks of asymmetric CO2 stretching mode (1555 cm−1 ) and symmetric CO2 stretching mode (1435 cm−1 ) from ZDA (Fig. 5.13(c)), with different intensities. When the samples are cured, the inhomogeneous distribution of ZDA agglomerates is maintained, but the chemical structures in each phase changes. The spectra change in the CH stretching region. In the PBD-rich matrix, the aliphatic CH stretching mode

FTIR IMAGING OF MULTICOMPONENT POLYMERS

(a)

3492 cm–1 image of ZDA

137

PBD-rich matrix ZDA-rich domain

(c) 4 PBD-rich matrix 0.20

3115 – 2810 cm–1 image of PBD

3 Abs

250 mm

0

ZDA-rich domain

2

(b) 1 0 3000 10.2

2500 2000 1500 Wavenumber (cm–1 )

1000

140.1

Figure 5.13 FTIR images of uncured PBD–ZDA blends: (a) 1270 cm−1 image of ZDA; (b) the image of PBD from integration of 3115–2810 cm−1 ; (c) the individual spectra extracted from the image. Reproduced from figure 1 of Ref. 28, with permission.

(2916 cm−1 ) increases and unsaturated CH stretching mode (3075 cm−1 ) decreases upon curing because single bond crosslinks are formed at the expense of double bonds. The same trend is also observed in the CH stretching mode in the ZDA-rich domain. The absorbance differences in the CH stretching mode between PBD-rich and ZDA-rich domains exist even after the curing reaction. Figure 5.14 shows the toluene distribution of the solvent-diffused sample (5 min cured). The image contrast is based on the difference in swelling capability throughout the sample. More toluene is imbibed into the PBD-rich matrix and less solvent is imbibed into the ZDA-rich domains. Therefore, the ZDA-rich domain is shown as low intensity (blue) and PBD-rich region is shown as high intensity (red, yellow or green). The difference in toluene distribution results from the differences in solubility and crosslink density.

5.5.5 Semicrystalline polymer systems Fourier transform infrared imaging allows for the examination of semicrystalline polymers via their dichroism.31 Typical images of the center of a PEG (Mw 35k, Tc ∼ 35◦ C) spherulite are shown in Fig. 5.15. These images were generated by plotting the baseline corrected absorbance intensity of the 1343 cm−1 CH2 wagging

138

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

250 μm

ZDA domain PBD matrix

0.6

1.4

Figure 5.14 Toluene distribution of the solvent-diffused sample (5 min cured). Reproduced from figure 2 of Ref. 28, with permission.

1.0

0.0 100 mm 2.0

0.5 Figure 5.15 Top, images of the center of a PEG spherulite obtained by plotting the 1343 cm−1 band from data obtained using either parallel (right) or perpendicular (left) radiation. Bottom, a dichroic ratio image obtained by ratioing the top images. Reproduced form figure 2 of Ref. 31, with permission.

band. This band was chosen because it has the strongest dichroic behavior of all of the spectral bands (vide infra), therefore, generating the best image contrast. The spherulitic structure, as visible in polarized optical microscopy, can be reproduced based on the orientation of transition dipole moments of functional groups in the sample. With an IR polarizer in the beam path, the degree of orientation for each vibrational mode was determined by the generation of spatially resolved dichroic

FTIR IMAGING OF MULTICOMPONENT POLYMERS

139

0.75

0.0 100 mm 2.0

0.5

Figure 5.16 Images of a polymer spherulite plotted as in Fig. 5.15 except the material experienced faster quenching to below the crystallization temperature. Reproduced from figure 2 of Ref. 31, with permission.

ratio images. Images from the same system, except quickly quenched are shown in Fig. 5.16. The images at the top were generated by placing the polarizer in the path of the IR beam such that the radiation was polarized in one of two orthogonal directions. The parallel direction was defined as being oriented top to bottom in the sample plane (the plane of the paper in the images), while the perpendicular direction was defined as being oriented left to right. Additionally, the image at the bottom of the figure is a dichroic ratio image, formed by a pixel-by-pixel division of the absorbance values from the parallel image by the absolute value of the perpendicular image. This dichroic ratio image shows the spatial distribution of orientation within the system, and additionally allows the orientation to be quantified. Using standard (nonmicroscopic) polarized FTIR spectroscopy, these systems demonstrate little or no dichroic behavior because a large number of spherulites are included in the IR beam, and because the spherulites possess cylindrical symmetry.

5.5.6 Semicrystalline polymer blends While FTIR images are similar to those that can be generated with polarized optical microscopy, the utility of fast FTIR imaging to the study of semicrystalline systems becomes apparent in the examination of blend systems. When a melt-miscible polymer system is analyzed at temperatures lower than the crystallization temperature of one constituent, a phase-separated or single crystalline phase structure might result. When the Tg of the noncrystalline component is lower than the crystallization temperature, the component phase separates as it is rejected from the crystal structure. The extent of phase separation may be easily visualized

140

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a)

(b)

(d)

(e)

(c)

100 mm

Figure 5.17 Images of a PEG–PVAc blend system showing the typical behavior for (a) parallel polarized, (b) perpendicular polarized and (c) dichroic ratio images. The bottom images generated by plotting the 1721 cm−1 carbonyl stretching band of PVAc for both (d) parallel and (e) perpendicular polarization show that PVAc is located mainly in the interspherulite regions. Reproduced from figure 3 of Ref. 31, with permission.

using IR imaging. PVAc–PEO and high molecular weight–low molecular weight PEO blends have been imaged to examine the degree of segregation.32 From the images in Fig. 5.17 it is apparent that the PVAc is mostly excluded from the crystalline regions, and is located mainly within the inter-spherulitic regions.

5.6 Summary and conclusions Fourier transform infrared images have been utilized in a large number of applications for complex polymer systems. The future is bright as new instruments will have faster data collection times and higher fidelity.

References [1] Mellbring, O., Olseth, S. K., Krozer, A., Lausman, J. and Hjertberg, T. (2001) Spin coating and characterization of thin high-density polyethylene films. Macromoleucles 34, 7496. [2] Stawhecker, K. E., Kumar, S. K., Douglas, J. F. and Karim, A. (2001) The critical role of solvent evaporation on the roughness of spin-cast polymer films. Macromoleucles 34, 4669. [3] Mansfield, J. R., Attas, M., Majzels, C., Cloutis, E., Collins, C. and Mantsch, H. H. (2002) Near infrared spectroscopic reflectance imaging: a new tool in art conservation. Vib. Spectrosc. 28(1), 59–66. [4] Dubois, L. H. and Nuzzo, R. G. (1992) Ann. Rev. Phys. Chem. 43, 437.

FTIR IMAGING OF MULTICOMPONENT POLYMERS

141

[5] Powell, G. L., Beecroft, M., Dombrowski, M., Mattison, P. and Szczesniak, M. M. (2003) FTIR imaging of the interaction of oil droplets with various substrates. PITTCON Paper 1703–4. [6] Gupper, A., Wilhelm, P., Schmied, M., Kazarian, S. G., Chan, K. L. A. and Reuû ner, J. (2002) Application of imaging methods for the characterization of a polymer blend. Appl. Spectrosc. 56, 1515. [7] Otts, D. B., Zhang, P. and Urban, M. W. (2002) High fidelity surface chemical imaging at 1000 nm levels: internal reflection IR imaging (IRIRI) approach. Langmuir. 18, 6473. [8] Zhang, P., Otts, D. B. and Urban, M. W. (2002) Recent advances in imaging of polymers; toward nano-level 3D spatial resolution using infrared spectroscopy. Polym. Mater.: Sci. Eng. 87, 170. [9] Sommer, A. J., Tisinger, L. G., Marcott, C. M. and Story, G. M. (2001) ATR infrared mapping microspectroscopy using an imaging microscope. Appl. Spectrosc. 55(3), 252–6. [10] Ekgasit, S. and Padermshoke, A. (2001) Appl. Spectrosc. 55, 1352. [11] Chan, K. L. A. and Kazarian, S. G. (2003) New opportunities in micro- and macro-attenuated total reflection infrared spectroscopic imaging: spatial resolution and sampling versatility. Appl. Spectrosc. 57, 381. [12] Bhargava, R., Wall, B. G. and Koenig, J. L. (2000) Comparison of the FTIR mapping and imaging techniques applied to polymeric systems. Appl. Spectrosc. 54(4), 470–9. [13] Brunelli, R. and Mich, O. (2001) Histograms analysis for image retrieval. Pattern Recog. 43, 625. [14] Coutts Carrie, A., Wright, Norman A., Mieso, Ellen V. and Koenig, Jack L. (2003) The use of FT-IR imaging as an analytical tool for the characterization of drug delivery systems. J. Control. Release 93(3), 223–48. [15] Marcott, C., Story, G. M., Dowrey, A. E., Reeder, R. C. and Noda, I. (1997) Mikrochim. Acta Suppl. 14, 157–63. [16] Marcott, Curtis and Reeder, C. Robert (1998) Industrial applications of FTIR microspectroscopic imaging using a mercury-cadmium-telluride focal-plane array detector. Proceedings of the SPIE – Infrared Technology and Applications XXIV, Vol. 3436, 285–9. [17] Miseo, E. (2002) Determination of inhomogeneities in spectroscopic samples. Amer. Lab. News 34(14), 20. [18] Chalmers, J. M., Everall, N. J., Schaeberle, M. D., et al. (2002) FTIR imaging of polymers – an industrial appraisal. Vib. Spectrsc. 30, 42. [19] Bhargava, R., Wang, S. Q. and Koenig, J. L. (1998) FTIR imaging of the interface in multicomponent spatially-separated systems using optical effects induced by differences in refractive indices. Appl. Spectrosc. 52(3), 323–8. [20] Bhargava, R., Wang, S. Q. and Koenig, J. L. (1999) Studying polymer dispersed liquid crystal formation by FTIR spectroscopy: 1. monitoring curing reactions. Macromolecules, 32(26), 8982–88. [21] Bhargava, R., Wang, S. Q. and Koenig, J. L. (1999) Studying polymer dispersed liquid crystal formation by FTIR spectroscopy: 2. phase separation and ordering. Macromolecules, 32(26), 8989–95. [22] Bhargava, R., Wang, S.-Q. and Koenig, J. L. (1999) FTIR imaging studies of a new two-step process to produce polymer dispersed liquid crystals. Macromolecules, 32, 2748–60. [23] Oh, S.-J. and Koenig, J. L. (1998) Phase and curing behavior of polybutadiene/diallyl phthalate blends monitored by FTIR imaging using facal array detection. Anal. Chem. 70, 1768–72. [24] Coleman, M. M., Graf, J. F. and Painter, P. C. (1991) Specific Interactions and the Miscibility of Polymer Blends, Technomic Publishing, Inc., Basel. [25] Galloway, J. A., Montminy, M. D. and Macosko, C. W. (2002) Image analysis for interfacial area and cocontinuity detection in polymer blends. Polymer, 43, 4715–22. [26] Artyushkova, K., Wall, B., Koenig, J. and Fulghum, J. E. (2000) Correlative spectroscopic imaging: XPS and FTIR studies of PVC/PMMA polymer blends. Appl. Spectrosc. 54(11), 1549–58. [27] Oh, S. J. and Koenig, J. L. (2000) Studies of peroxide curing of polybutadiene/zinc diacrylate blends by fast FTIR imaging. Rubber Chem. Technol. 73(1), 74–9. [28] Oh, S. J. and Koenig, J. L. (1999) Studies of peroxide curing of cis-polyisoprene/diallyl phthalate blends by spectroscopic techniques. Rubber Chem. Technol. 72(2), 334–42.

142

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

[29] Bhargava, R., Wall, B. G. and Koenig, J. L. (2000) Comparison of the FTIR mapping and imaging techniques applied to polymeric systems. Appl. Spectrosc. 54(4), 470–9. [30] Oh, S. J. and Koenig, J. L. (2000) Studies of peroxide curing of polybutadiene/zinc diacrylate blends by fast FTIR imaging. Rubber Chem. Technol. 73(1), 74–9. [31] Snively, C. M and Koenig, J. L. (1999) J. Polym. Sci. Part B: Polym. Phys. 37(17), 2353–9. [32] Bhargava, R., Ribar, T. and Koenig, J. L. (1999) Towards faster FTIR imaging by reducing noise. Appl. Spectrosc. 53, 1313.

6

Combinatorial approaches to catalyst development with multichannel detectors Christopher M. Snively and Jochen Lauterbach

6.1 Introduction – combinatorial materials development The basic idea behind the combinatorial approach to the discovery process is to synthesize a large, highly diverse population of potential candidates, and then rapidly screen them in order to determine which of them possess properties that are potentially useful. This approach has been employed in the field of organic synthesis for decades, and has been exploited especially for the discovery of novel pharmaceuticals.1,2 The main experimental innovation driving this particular application area is the ability to attach molecules to solid supports. These solid supports are then exposed to the desired reaction conditions, and the products can be isolated simply by isolating the support material, usually in the form of insoluble polymeric objects with micron to millimeter dimensions. This is readily applicable because synthetic methods are available for generation, modification and linking of a wide range of organic functional groups. Owing to the added complexity brought about by the synthesis of hard materials, such as heterogeneous catalysts, superconductors or phosphors, the approaches applied in the pharmaceutical area are not directly transferable to the realm of materials development and discovery. Thus, the growth of combinatorial materials science as a discipline is highly dependent on the development of both novel high-throughput processing techniques to manufacture a large number of structurally diverse samples and high-throughput analytical techniques that are capable of rapidly providing the information required for the determination of the material structure and performance. High-throughput methodologies for the design and evaluation of inorganic materials were first described in the early 1970s.3 The parallel testing concept was first employed in the field of heterogeneous catalysis as early as 1986.4 There was a research explosion in the field of high-throughput materials discovery in the late 1990s. Many advances in experimental philosophy pertaining to this research area have been made in a relatively short time. There is currently a debate in the catalysis community over whether the high-throughput approach is the way in which the field should advance. The argument that is often heard on the side of opposition of high-throughput experimentation (HTE) is that this approach goes against the very foundation of the scientific method and the Edisonian approach to experimental research.

144

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Most of this opposition stems from the fact that the initial studies in the HTE area were concerned primarily with speed. Libraries of samples were generated rapidly, however, very little characterization was performed on these libraries to determine if the desired compositions were achieved or if these model materials accurately mimicked their real-world counterparts. While it is possible to employ such a ‘shotgun’ approach, this is ultimately self-defeating, and most researchers in the HTE field have come to this realization. Much effort is currently being placed not only on determination of final material properties, but also on rapid characterization methods that can be applied to the myriad formulations generated in these studies. As with many new developments in science, it is the manner in which it is employed that determines the validity of the approach. While some early research in the high-throughput materials area was focused merely on deriving ‘yes or no’ answers to questions, the field has matured to the point where researchers are beginning to incorporate computer modeling methods in order to intelligently design materials. The high-throughput approach is seen to simply be a more efficient way of performing the same experiments that have always been performed. The development of approaches that will incorporate HTE into a rational design approach that is designed to complete the experimental–theoretical loop is also a topic of much interest in the field.5 The first applications of analytical techniques to the study of combinatorial catalyst libraries involved the use of traditional analytical techniques that were modified in order to increase their throughput. Examples of these include fast gas chromatography (GC), resonance-enhanced multiphoton ionization, and scanning mass spectrometry.6 While this was a necessary first step in the field, these approaches are all based upon sequential techniques. Since only one sample can be studied at a time, the total analysis time is directly proportional to the total number of samples. In order to maximize the throughput of an analytical technique, it is desirable to have a way of collecting information from multiple samples simultaneously, in a parallel fashion. There are several ways in which this can be approached. In the limit of infinite resources, one could apply the brute force method, in which a multitude of analytical instruments would be applied to the problem. For example, in order to collect GC data from 50 reactors in parallel, 50 gas chromatographs would be purchased and connected to the reactors. While this approach will work for a relatively small number of samples, it is obvious that the extension to larger sample sets is infeasible for even the deepest of pockets. A more intelligent way of solving this problem is to ‘parallelize’ as many parts of the setup as possible. The crux of this concept lies in the ability to do more with less, allowing the study of a multitude of samples while requiring a minimum of additional resources per additional sample to be studied. In order to take maximum advantage of available resources, the parallel approach should be applied to all aspects of the analytical approach. If a highly parallel analytical technique is added to a sequential reactor system, or vice versa, the total throughput of the system may not increase at all. Just as a chain is only as strong as its weakest link, the high-throughput setup is only as fast as its slowest component.

CATALYST DEVELOPMENT

145

A large part of optimizing analytical setups involves removing the main bottleneck that is slowing down the entire process. Often, several iterations are required in which the bottleneck is shifted back and forth between the reactor and analytical technique before an optimum is found. This optimization can also be performed on the synthetic side of the problem, with the utilization of automated liquid dispensing and sample handling systems. Unfortunately, the end point of this endeavor is often determined by purely monetary considerations. In the area of heterogeneous catalysis, the design of dedicated reactor systems is crucial to the successful implementation of an HTE strategy. Most of the initial work on the development of reactors for high-throughput testing of heterogeneous catalysts consisted of the development of reactor systems amenable to a specific analytical technique.7–10 Catalysts tested using such reactors were usually not analyzed under well-controlled conditions, nor under conditions that were directly comparable with traditional single reactor studies. Many groups began to recognize the value in developing high-throughput reactors that tested catalysts under nearly conventional conditions.11–16

6.2 Array detection schemes for high-throughput analysis The ideal high-throughput analytical technique would be efficient in terms of required resources and would be scalable to accommodate an arbitrarily large number of samples. In addition, this scalability would be such that the dependence of the cost of the equipment to perform the experiments would scale in a less than linear manner as a function of the number of samples that could be studied. The only way to accomplish this is to have one or more aspects of the experimental setup utilize an array-based approach. Array detectors are massively multiplexed versions of single-element detectors composed of a rectangular grid of small detectors. The most commonly encountered examples are CCD cameras, which are used to acquire ultraviolet, visible and near-infrared (IR) photons in a parallel manner. Other examples include IR focal plane arrays (FPAs) for the collection of IR photons and channel electron multipliers for the collection of electrons. Array detectors are typically employed to provide a spatial aspect to the collected data. Thus, a dataset from a single experiment is composed of data from a variety of spatial locations of a single sample. This is commonly used to map out chemical or physical heterogeneity in multicomponent systems.17,18 The approach taken in high-throughput studies is a bit different, in that many samples are simultaneously placed in the field of view of the instrument. Thus, a single experimental dataset is composed of information from multiple samples, as illustrated in Fig. 6.1. This leads to a multisample multiplex advantage, since many samples can be analyzed with the collection of a single dataset. By designing the experimental setup and associated sampling accessories appropriately, it becomes possible to analyze an arbitrarily large number of samples up to, and including, the total number of pixels in the array detector. Since the arrays employed to date have been used to detect

146

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

A

A

Radiation from

B B

spectrometer

C ne

m

ple

Sa

r

cto

pla

A FP

te de

C

Figure 6.1 Basic concept of the application of FTIR imaging as a high-throughput technique. Placing multiple samples in the field of view of the imaging spectrometer allows the collection of a single dataset that contains information from all samples. Reprinted with permission from ACS.

either photons or electrons, developing a particular experimental setup is a matter of optical design. The first array-based technique was designed specifically to study reactions on solid phase catalysts as IR thermography.9,19 This approach utilizes IR sensitive FPA detectors to measure the temperature of catalysts under reaction conditions. This approach has the advantages of a theoretical high thermal sensitivity, typically several tens of millikelvin, and the ability to study both endothermic and exothermic reactions. The main disadvantage of this approach, however, is the lack of chemical information. It must be assumed that the temperature change is associated entirely with the desired reaction pathway. The presence of unexpected side reactions will not be detected in this approach, as long as they have similar thermal behavior as the reaction under study. It is also possible to partially alleviate the problem of chemical insensitivity by incorporating narrow bandpass filters into the optical setup.20 Thus, by choosing an appropriate frequency region, it becomes possible to detect the presence of a particular reactant or product species. While this adds some measure of chemical sensitivity to the thermography approach, it is only capable of monitoring one species at a time. Additionally, the success of this approach relies upon the fact that the spectral bands of the desired species do not overlap with any other species and that unexpected reaction products that have spectral contributions in the region of interest are not present.

6.3 FTIR imaging as a high-throughput technique To remove some of the limitations imposed by the approaches mentioned above, Fourier transform infrared (FTIR) imaging has been developed as a high-throughput technique for the study of heterogeneous catalysts.21,22 By combining an array

147

CATALYST DEVELOPMENT

Sampling accessory FPA

AB C

D

Rapid scanning FTIR spectrometer E

F

Figure 6.2 General layout of the FTIR imaging spectrometer. Optical elements F and G illuminate the sampling accessory with light from the spectrometer; optical elements B and E focus the light from the sampling accessory onto the FPA detector; aperture C is used to control the overall light level reaching the detector, and filter B is used to block light from outside the desired spectral region.

detection scheme with spectral multiplexing, it becomes possible to acquire high fidelity spectral information from multiple samples simultaneously. True chemical sensitivity is therefore achieved, as well as the ability to monitor several chemical species simultaneously. This is not possible using the above techniques or other array-based techniques, including laser-induced fluorescence imaging23,24 and colorimetric assays.25 A diagram of the experimental setup employed in our lab is shown in Fig. 6.2. A Bruker Equinox 55 spectrometer provides the modulated IR radiation, which is directed into the sampling accessory via refractive optical elements. The light emerging from the sampling accessory is subsequently focused onto a 64 × 64 pixel mercury cadmium telluride FPA detector via another set of refractive optical elements. An optical filter is also included in the setup to minimize Fourier foldover noise. When the interferometer is operated in step-scan mode, this setup is capable of acquiring a complete imaging dataset in a matter of several minutes. This is adequate for static or slowly varying systems, but does not allow the study of faster dynamic processes. This was improved upon by operating the spectrometer in rapid-scanning mode,26 which greatly reduced the acquisition time for an imaging dataset. For example, an imaging dataset with 8 cm−1 spectral resolution over a 1300 cm−1 spectral range can be acquired in <2 s, with a resulting signal-to-noise ratio better than 100 : 1. This temporal resolution and data quality is adequate for the study of many dynamic processes in real time. The successful application of FTIR imaging to high-throughput studies requires the use of dedicated sampling accessories due to the unique sample geometries encountered in these systems.27 These sampling accessories are parallel counterparts to the single-sample accessories that are familiar sights in an IR spectroscopy laboratory. The diameter of optics that can be readily obtained commercially is typically limited to 3 in, which imposes an upper limit to the overall size of the sampling accessory. There, therefore, exists a tradeoff between the number of samples that can be simultaneously analyzed and the area per sample. This requires one to establish a balance between throughput (number of samples) and data quality (number of pixels per sample). By incorporating larger format FPA detectors, it will become possible to increase the throughput while maintaining the data quality due to the increased pixel density. An example of a dedicated sampling accessory is the parallel gas phase array (GPA) that was utilized in the gas phase effluent studies mentioned below. For

148

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Figure 6.3 Photograph of the GPA sampling accessory.

the study of gas phase reactor effluents, 16 was chosen as a reasonable number of catalysts to study. This provides an adequate limit of detection, which is typically between 10 and 100 ppm v/v for acquisition times in the range of several seconds, depending on the gas to be analyzed. It also provides a balance between the speed at which catalysts can be synthesized and analyzed. The GPA is composed of 16 tightly packed stainless steel tubes, each with a dedicated gas inlet and outlet to eliminate crosstalk between samples. The ends of the tubes are sealed via silicone sealant to IR transparent windows, which are chosen such that they will be inert to the reaction gases under study (see Fig. 6.3). Each stainless steel tube is 20 cm in length with an outer diameter of 0.375 in. The residence time of the gas in the tube is ∼4 s when the flowrate is set to 125 ccm per channel. The IR light is transmitted through all tubes simultaneously, allowing the products from all reactors to be analyzed in parallel. Figure 6.4 shows a typical set of time-resolved IR spectra obtained from a single channel. The IR bands of different components in the gas phase and their temporal evolution can clearly be distinguished. Another approach to gas phase sampling, utilizing a reflection–absorption sampling geometry, has been demonstrated recently in which 49 catalysts can be analyzed simultaneously.28

6.4 Applications 6.4.1 Application I: resin-supported ligands Although our main research interest has been in the field of heterogeneous catalysis, we have performed several proof of concept studies showing that FTIR imaging is flexible enough to be applied to the study of resin-supported combinatorial libraries, of the type commonly employed in split-and-pool syntheses. It was shown that the

149

CATALYST DEVELOPMENT

.) Absorbance (A.U

0.6

N2O

0.4

NO2 0.2

NO

CO

0.0 75

2200 2000

50

Tim

e ( 25 s)

0

1600

–1 )

cm

1800

( ber

m

u ven Wa

Figure 6.4 A series of typical spectra acquired from a single sample, demonstrating the spectral and temporal resolution of rapid-scan FTIR imaging. Reprinted with permission from Wiley.

chemical identity of several different amino acid ligands supported on polystyrene resin beads could be identified simultaneously in a single experiment using only their IR signatures.21 Additionally, it was shown that the reaction kinetics of supported ligands could be followed in a quantitative manner. Previous studies have shown that the reaction kinetics could be followed in parallel using near IR spectral imaging.29 However, the lack of chemical specificity that is associated with this spectral region, as compared to the mid-IR, limits this approach to the study of select chemical systems. The oxidation of a primary alcohol to an aldehyde was chosen as a model reaction in our studies. Aliquots of beads were removed from the reaction mixture several times during the course of the reaction. These beads were quenched and placed in small, separate groups onto an IR transparent substrate. A single imaging dataset containing spectral information from all of these groups was collected. After analyzing the dataset, the kinetics of the reaction was quantified.21 It was also shown that this analysis could be performed in situ by placing the beads inside of a flow cell and acquiring imaging data from them during the reaction.30 This approach could be applied, for example, to the parallel study of the reaction kinetics of a variety of different ligands. The major advantage to all of these approaches is their nondestructive nature. After analysis, the beads are able to be recovered and submitted for further analysis or additional reaction steps.

6.4.2 Application II: adsorbates on catalyst surfaces Over the past several years, FTIR imaging has been applied to the study of a variety of catalytic systems in several different sampling geometries. The first examples in

150

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

this area demonstrated that the chemical species adsorbed on the surface of multiple catalyst samples could be analyzed. It was shown that the adsorption of CO onto both zeolites and transition metal catalysts supported on oxide materials could be differentiated based solely upon their IR absorption signatures.26 In later studies, it was shown that the kinetics of this adsorption process could be followed in real time.22 Thus, the difference in adsorption kinetics of molecules from the gas phase onto the catalysts of interest was determined quantitatively. This approach has also been used to characterize the types of sites present on catalytic materials by detecting various vibrational bands of adsorbed pyridine.31 These types of experiments allow not only the parallel, in situ characterization of available adsorption sites on catalytic surfaces, but also the study of adsorbed reactants, products, reaction intermediates and spectator species. This type of experimental approach can lead to an accelerated understanding of structure–activity relationships, the key information necessary for rational design of catalytic materials.

6.4.3 Application III: reactor effluent quantification Most often, the goal of high-throughput catalytic studies is the rapid quantification of chemical species in effluent gases from parallel reactors. This gives a direct measure of the rate and selectivity performance of the catalysts, which are the parameters that are most indicative of their real-world applicability. All of the studies in our laboratory have employed a 16-well reactor22,32–34 for the study of several different reactions. Here, we report briefly on a study of NOx storage and reduction (NSR) catalysts for lean burn engine applications. The inability of current three-way catalytic converters (TWC) to reduce nitrogen oxides (NOx ) under net-oxidizing conditions is impeding the widespread implementation of lean burn engines. To address the apparent conflict of high fuel efficiency of lean burn engines and low NOx emissions, NSR catalysts were designed to store NOx during a fuel lean cycle and reduce the stored NOx during a subsequent fuel rich cycle.35–38 We performed a systematic study of the catalytic performance over a wide range of operating conditions for catalysts of varying compositions. The catalysts were studied with respect to three performance criteria: saturation NOx storage capacity, total N2 O production, and the steady-state lean NOx reduction to both partially reduced nitrous oxide and fully reduced molecular nitrogen. The effluent of each reactor contains a mixture of NO, N2 O, NO2 , CO and C2 H4 in He. Therefore, multivariate techniques had to be employed to quantify these mixtures (vide infra). A typical result of effluent concentration as a function of time is shown in Fig. 6.5. It becomes immediately clear that only a truly parallel, chemically sensitive technique could be used to study such a system in a high-throughput fashion. FTIR imaging allows us to rank catalytic materials according to the above-defined performance criteria. The current experimental setup, with its excellent temporal resolution afforded by rapid-scan data collection, allows us to perform kinetic measurements of catalysts in parallel. An example of this can be seen in the measurements of reaction order and apparent activation energies during CO oxidation over supported rhodium

151

CATALYST DEVELOPMENT

1500 NOx

Concentration (ppm)

N2O CO 1000

500

0 0

50

100

150

Time (s) Figure 6.5 Concentrations of three effluent species during several rich-to-lean switches from an NSR catalyst.

catalysts. Rhodium is one of the metals used in automotive catalytic converters, and, although its main role is to promote reduction of NO, it is also known to be a good catalyst for CO oxidation. Steady-state reaction experiments were performed at four different catalyst temperatures with various Rh and noble metal catalysts. Figure 6.6 shows an example of the measured reactivity versus O2 partial pressure for one temperature. The calculated reaction orders agree well with literature data,39 which report a positive order close to unity. To determine the overall activation energy for the CO oxidation over rhodium, rate measurements were performed at fixed-feed concentrations of CO and O2 while the reaction temperature was varied. The results for a typical Arrhenius plot with the calculated apparent activation energy values are shown in Fig. 6.7. The results again compare very favorably with the literature, where values between 83 and 142 kJ mol−1 have been reported.39 This example shows that it is possible to extract quantitative kinetic data in parallel using arraybased detection techniques, thereby considerably speeding up traditional catalysis research.

6.5 Data management One of the greatest challenges that accompanies the implementation of array detectors is the handling of the massive amounts of data that are generated. Since the detector is essentially a collection of many small detectors, each dataset is composed of data orders of magnitude more than from traditional experimentation. To make matters worse, this massive amount of data is collected in a time period that is

152

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

0.8

–5.8 –6.0

0.6 –6.2 –6.4

1.2

log (r)

–6.6 –6.8 –7.0 –7.2 –7.4 –7.6 –7.8 –1.8

–1.6

–1.4

–1.2

–1.0 logPO2

–0.8

–0.6

–0.4

Figure 6.6 Plot used for the determination of O2 reaction order for various catalysts undergoing CO oxidation. A reaction order around unity was obtained for all experiments.

comparable to the time for a single experiment on conventional instrumentation. As a practical example, the collection of a single dataset using a 64 × 64 pixel FPA with 8 cm−1 spectral resolution over the 2600–1300 cm−1 spectral range occupies ∼5 MB of disk space. During a transient reaction kinetics experiment, one of these data sets can be collected approximately every 2 s. Thus, during a typical day of experimentation, several tens of gigabytes of data are collected. This figure gets even larger when one considers that the data must be Fourier transformed to produce single-beam spectra and finally processed into absorbance spectra. From this example, it is clear that a highly optimized strategy for data processing and reduction is necessary to eliminate any potential bottleneck in this area. This leads to the necessity of implementing intelligent strategies to (1) more effectively manage the data, such that data that is unusable or of questionable quality is eliminated and (2) collect data more efficiently by reducing the total number of experiments required to solve a particular problem. General approaches to data processing in high-dimensional spectrochemical data are discussed in Chapter 4; here, we summarize the unique issues and data handling in typical HTE. Due to the unique properties of the optical setup and the resulting datasets generated in our laboratory, custom data processing software was developed. Common operations, such as fast Fourier transformation, absorbance calculation, and species quantification were implemented with efficient algorithms.40 The FFT of a single

CATALYST DEVELOPMENT

153

–11

–12

–13 176 (kJ/mol)

ln[r]

–14

–15

–16

103 (kJ/mol)

–17

–18 0.270 0.275 0.280 0.285 0.290 0.295 0.300 0.305 0.310 0.315 1000/RT Figure 6.7 Arrhenius plot for a set of supported Rh catalysts, showing a range of apparent activation energies.

dataset, which is the most time consuming operation, requires <1 s per image on a modern desktop PC (3.0 GHz processor, 1 GB memory, Windows OS, RAID0 hard disk configuration). One method that has been implemented to reduce the computing time for each dataset is to realize that, although there are thousands of pixels in the array, there are only 16 samples being studied. An example of a typical brightfield image collected through the GPA is shown in Fig. 6.8. It is obvious that a large number of pixels in the array are not used to collect any meaningful data. This stems from the inability to realize a completely close-packed arrangement of the tubes in the GPA and also from residual distortion in the optical setup. These factors lead to the result that, for the image in Fig. 6.8, <20% of the pixels contain useful data. The data from the pixels corresponding to each sample are averaged together to form 16 separate spectra. This results in a reduction from 5 MB to 20 kB for each dataset. Univariate analysis methods have been applied for the quantification of gas mixtures in several systems that we have studied, however, it becomes clear that more advanced techniques must be applied in more complicated systems. An example of this is shown in Fig. 6.9, which shows spectra of a gas mixture of C2 H4 , NO and H2 O. Owing to the significant band overlap present between these components, it is clear that the implementation of a univariate approach will not allow the quantification of

154

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Figure 6.8 Typical brightfield image, as viewed through the GPA sampling accessory.

0.12 C2H4

Absorbance (a.u.)

0.10 0.08

NO

0.06 0.04

NO + C2H4 + H2O

0.02 0.00 2000

1900

1800

7000

1600

Wavenumbers (cm–1 ) Figure 6.9 Spectra showing the complete overlap of the spectral bands of NO, C2 H4 and H2 O.

all of these species. Multivariate processing methods must be used to extract reliable concentration information from the data.41,42 Multivariate methods, in principle, allow any number of components to be analyzed and extend the quantitative usefulness of IR analysis to complex mixtures. We typically employ the factor-based methods of principal component regression (PCR) and partial least squares (PLS-1 and PLS-2). The basic premise of these methods is that the entire IR spectrum can be utilized to quantify concentrations instead of relying upon a single spectral frequency. A calibration dataset was collected with the same spectral parameters as the sample data. This calibration set spanned the entire expected concentration range,

CATALYST DEVELOPMENT

155

and was designed such that the variables of interest were varied independently. Separate calibrations were performed for each channel in the GPA, in order to minimize any effect that the optical aberrations may have on the results. The use of a validation dataset revealed43 that the average difference between the expected experimental concentration and the predicted concentration is typically below 100 ppm v/v for NO and C2 H4 , which demonstrates the excellent performance obtained using the multivariate factor-based calibration. The effect of increasing the total number of coadded scans for each spectrum was also investigated. An increase in the number of scans from 1 to 8 improved the calibration model substantially, but further increases in the number of scans, and therefore the collection time, did not significantly increase the calibration effectiveness. Typical concentration information extracted from a dataset using the multivariate approach can be seen in Fig. 6.5. The vast amount of quantitative data generated by array-based high-throughput techniques makes it necessary to approach catalysis research from a slightly different angle. It becomes almost impossible for the researcher to follow all results obtained and to keep the ‘big picture’ in focus. It therefore becomes mandatory to make use of the statistical methods of design of experiments (DoE), which helps to guide experiments and extract the maximum amount of information in a systematic fashion. We have employed both screening designs and response surface designs to the NSR problem.44 This organized approach has led to a considerable reduction in experiments, while preserving the quantity and quality of information. This level of understanding allows the derivation of mathematical models of catalyst performance as a function of catalyst composition and operating conditions. The ultimate goal of this vein of research is to gain a complete understanding of this system, such that the intelligent design of novel catalytic materials with improved performance will be possible.5

6.6 Summary Fourier transform infrared imaging is a versatile analytical technique that has been adapted as a high-throughput tool for the study of heterogeneous catalysts. High fidelity, chemically sensitive information can be obtained from a variety of sample geometries and chemical systems with a temporal resolution sufficient to follow many dynamic processes in real time. Future developments in the field of high-throughput IR spectroscopy are currently limited by the development of FPA detectors. As this technology matures, faster and more sensitive detectors with larger numbers of pixels will become available. This will improve the performance of highthroughput IR spectroscopic instrumentation by allowing the detection of lower concentrations of species from an increasing number of samples at a higher time resolution. Similar approaches are being applied to other analytical techniques, including the development of array detectors in mass spectrometry45,46 and the development of parallel data collection schemes in NMR.47,48 It is expected that all of these developments will have a major impact on the field of heterogeneous

156

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

catalysis and will open up new avenues of research, eventually allowing for the implementation of a data-driven approach to the design of new and novel catalytic materials.

References [1] Czarnik, A. W. and DeWitt, S. H. (eds) (1997) A Practical Guide to Combinatorial Chemistry, American Chemical Society, Washington, DC. [2] Terrett, N. K. (1998) Combinatorial Chemistry, Oxford University Press, New York. [3] Hanak, J. J. (1970) J. Mater. Sci. 5, 964. [4] Creer, J. G., Jackson, P., Pandy, G., Percival, G. G. and Seddon, D. (1986) Appl. Catal. 22, 85. [5] Caruthers, J. M., Lauterbach, J. A., Thomson, K. T. et al. (2003) J. Catal. 216, 98. [6] Senkan, S. (2001) Angew. Chem. Int. Edn. 40, 312. [7] Senkan, S., Krantz, K., Ozturk, S., Zengin, V. and Onal, I. (1999) Angew. Chem. Int. Edn. 38, 2794. [8] Cong, P., Doolen, R., Fan, Q. et al. (1999) Angew. Chem. Int. Edn. 38, 484. [9] Moates, F. C., Somani, M., Annamalai, J., Richardson, J. T., Luss, D. and Willson, R. C. (1996) Indust. Eng. Chem. Res. 35, 4801. [10] Orschel, M., Klein, J., Schmidt, H. W. and Maier, W. F. (1999) Angew. Chem. Int. Edn. 38, 2791. [11] Hagemeyer, A., Jandeleit, B., Liu, Y. M. et al. (2001) Appl. Catal. A-Gen. 221, 23. [12] Hoffmann, C., Wolf, A. and Schuth, F. (1999) Angew. Chem. Int. Edn. 38, 2800. [13] Desrosiers, P., Guram, A., Hagemeyer, A. et al. (2001) Catal. Today 67, 397. [14] Thomson, S., Hoffmann, C., Ruthe, S., Schmidt, H. W. and Schuth, F. (2001) Appl. Catal. A-Gen. 220, 253. [15] Perez-Ramirez, J., Berger, R. J., Mul, G., Kapteijn, F. and Moulijn, J. A. (2000) Catal. Today 60, 93. [16] Hahndorf, I., Buyevskaya, O. V., Langpape, M. et al. (2002) Chem. Eng. J. 89, 119. [17] Paschalis, E. P., Recker, R., DiCarlo, E., Doty, S., Atti, E. and Boskey, A. L. (2003) J. Bone Min. Res. 18, 1942. [18] Chalmers, J. M., Everall, N. J., Schaeberle, M. D. et al. (2002) Vib. Spectrosc. 30, 43. [19] Taylor, S. J. and Morken, J. P. (1998) Science 280, 267. [20] Qin, F. and Wolf, E. E. (1996) Catal. Lett. 39, 19. [21] Snively, C. M., Oskarsdottir, G. and Lauterbach, J. (2000) J. Combin. Chem. 2, 243. [22] Snively, C. M., Oskarsdottir, G. and Lauterbach, J. (2001) Catal. Today 67, 357. [23] Su, H. and Yeung, E. S. (2000) J. Amer. Chem. Soc. 122, 7422. [24] Su, H. and Yeung, E. S. (2002) Appl. Spectrosc. 56, 1044. [25] Busch, O. M., Hoffmann, C., Johann, T. R. F., Schmidt, H. W., Strehlau, W. and Schuth, F. (2002) J. Amer. Chem. Soc. 124, 13 527. [26] Snively, C. M., Katzenberger, S. and Oskarsdottir, G. and Lauterbach, J. (1999) Opt. Lett. 24, 1841. [27] Snively, C. M. and Lauterbach, J. (2002) Spectroscopy 17, 26. [28] Kubanek, P., Busch, O., Thomson, S., Schmidt, H. W. and Schuth, F. (2004) J. Combin. Chem. 6, 420. [29] Fischer, M. and Tran, C. D. (1999) Anal. Chem. 71, 2255. [30] Lauterbach, J., Snively, C. M. and Oskarsdottir, G. (2002) Chemically sensitive highthroughput parallel analysis of resin bead supported combinatorial libraries. In Combinatorial Materials Development (R. Malhotra, ed.), American Chemical Society, Washington, DC. [31] Busch, O. M., Brijoux, W., Thomson, S. and Schuth, F. (2004) J. Catal. 222, 174. [32] Hendershot, R. J., Fanson, P. T., Snively, C. M. and Lauterbach, J. (2003) Angew. Chem. Int. Edn. 42, 1152. [33] Hendershot, R. J., Lasko, S. S., Fellmann, M. F. et al. (2003) Appl. Catal. A-Gen. 254, 107. [34] Snively, C. M., Oskarsdottir, G. and Lauterbach, J. (2001) Angew. Chem. Int. Edn. 40, 3028.

CATALYST DEVELOPMENT

[35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48]

157

Hirata, H., Hachisuka, I., Ikeda, Y., Tsuji, S. and Matsumoto, S. (2001) Top. Catal. 16/17, 145. Bögner, W., Krämer, M., Krutzsch, B. et al. (1995) Appl. Catal. B: Environ. 7, 153. Takahashi, N., Shinjoh, H., IIjima, T. et al (1996) Catal. Today 27, 63. Matsumoto, S. (1996) Catal. Today 29, 43. Campuzano, J. C. (1990) In The Chemical Physics of Solid Surfaces and Heterogeneous Catalysis (D. A. Kind and D. P. Woodruff, eds), Elsevier, Amsterdam, Vol. 3, p. 389. Frigo, M. and Johnson, S. G. (1998) ICASSP, IEEE, 35, 1381. Kramer, R. (1998) Chemometric Techniques for Quantitative Analysis, Marcel Dekker, New York. Otto, M. (1999) Chemometrics, Wiley-VCH, New York. Hendershot, R. J., Vijay, R., Feist, B. J., Snively, C. M. and Lauterbach, J. (2005) Meas. Sci. Technol. 16(1), 302–5. Hendershot, R. J., Rogers, W. B., Snively, C. M., Ogunnaike, B. A. and Lauterbach, J. (2004) Catal. Today, 98(3), 375–85. Winograd, N. and Braun, R. M. (2001) Spectroscopy 16, 14. Barnes, J. H., Schilling, G. D., Sperline, R. et al (2004) Anal. Chem. 76, 2531. Macnaughtan, M. A., Hou, T., Xu, J. and Raftery, D. (2003) Anal. Chem. 75, 5116. Reetz, M. T., Tielmann, P., Eipper, A., Ross, A. and Schlotterbeck, G. (2004) Chem. Commun. 12, 1366.

7

Materials analysis systems based on real-time near-IR spectroscopic imaging Martin Kraft, Raimund Leitner and Herwig Mairer

7.1 Introduction Near-infrared (NIR) spectroscopy is a tool widely used in material analysis, since a majority of substances have characteristic absorption features in this spectral range. Applications range from simple qualitative analysis and material classification to in-line process monitoring.1,2 A range of NIR chemical sensor probes is commercially available. Probably the most critical restriction in the applicability of single-point probes is the confinement to samples where it can be guaranteed that the investigated objects are sufficiently homogeneous and the investigated point(s) are representative for the entity. While this is the case for the majority of fluid samples, many solid materials will not fulfil this requirement. For example, modern high-tech products usually contain a wide range of different materials, while in food industry an item that is analysed as edible at one spot may well be rotten at a different one. While for precisely defined objects these problems can be overcome (e.g. by measuring at pre-selected, representative points), hardly any industrially applicable systems are available for reliably analysing and classifying samples of unknown or variable shape or material distribution. This creates a significant commercial demand for contact-free methods capable of spatially resolved, qualitative and/or quantitative analysis in real time at a possibly high throughput. Further key requirements are industrial suitability, high reliability, reasonable investment and low operation costs and the capability of delivering data on size and shape of the sample and/or inhomogeneous areas. These requirements clearly indicate the application of spectroscopic imaging methods based on multichannel detection, also known as ‘Spectral Imaging’ (SI).3–5 While until recently such technologies have been restricted mostly to simple filter solutions or to the lab, the commercialisation of new components now makes SI systems from the UV to the NIR available for industrial applications at competitive prices.

7.2 Data acquisition 7.2.1 Image acquisition The classical scheme of sample scanning with a single-point probe is flexible in applicable methodology and both spectral and spatial resolution. In practice,

MATERIALS ANALYSIS SYSTEMS

159

the instrumentation required for the scanning movement and, especially, the prohibitively high time consumption for the acquisition of one single image renders this approach useless for industrial imaging applications. This is even more true when the aim is to devise real-time material analyses protocols based on spectroscopy. Hence, for spectroscopic imaging single-point probes are restricted to off-line laboratory applications. Currently, the most widely used method is wavelength scanning SI, which is better known as ‘staring imaging’ in remote sensing. This technique is based on acquiring a number of single two-dimensional images of a sample at different wavelengths. For industrial applications, the wavelength selection can be achieved by employing either (1) a number of discrete filters,6 (2) tuneable filters,7,8 for example, acousto-optical tuneable filters (AOTFs) or liquid crystal tuneable filters (LCTFs) or (3) by illumination of the sample at selected, discrete wavelengths.9 Under laboratory conditions, other principles, such as FT-imaging, are also available. In wavelength scanning spectroscopic imaging, both spatial dimensions are acquired simultaneously, while the spectral information is acquired sequentially. After completion of the data acquisition, the single, spectrally encoded images are combined in the computer, allowing calculation of the spectra for each pixel. This method is highly useful for a range of applications, in particular when only a few images at characteristic wavelengths have to be recorded, for example, for humidity measurements. Filter solutions can easily and quickly be implemented, but are restricted to target analysis and are generally less flexible than fully spectroscopic instruments. The more different wavelengths that have to be recorded, the more stringent becomes the restriction that the sample must be kept stationary under the SI unit during image acquisition for all wavelengths of interest. As the acquisition of a single, spectrally highly resolved spectroscopic image may last several tens of seconds, this may present a technical and process-logistical problem in an industrial process. The samples would have to move in a step-wise manner, which limits the application to processes where the method of movement may be adapted to the requirements of the sensor and to off-line laboratory applications, for example, distribution and dissolution studies of active ingredients in pharmaceuticals. A further limitation when aiming at real-time systems is that the full spectral information is available only after having acquired images for all wavelengths of interest. Hence, chemometric evaluation algorithms can only be applied after having recorded the full spectral hypercube. This necessarily involves handling of large amounts of data, which makes demands on the storage and processing capabilities of the processing system and limits the real-time capability. More suitable for real-time industrial material analysis using spectroscopic imaging is a third method, spatial scanning SI, more commonly known as push-broom scanning SI. The term ‘push-broom scanning’ originates from remote sensing and implies the line-wise acquisition of the image data, making use of a constant, relative movement between sample and imager. Instead of recording a two-dimensional image, a line across the sample, perpendicular to the direction of the relative movement, is projected into an imaging spectrograph. The radiation originating

160

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

along this observation line is spectrally analysed and the spectral information for each pixel along the investigating line projected along the second axis of the two-dimensional detector chip. The spectral encoding can be provided either by dispersive optics forming an imaging spectrograph10,11 or by linearly variable filters.12 Also some developments involving micro-opto-electro-mechanical systems (MOEMS) have been reported.13 Since the spatial information along the line is retained, the computerised images contain the spatial information along the first axis and the full spectral wavelength information along the second axis, as depicted in Fig. 7.1. The spectral and the first spatial dimension are simultaneously acquired, while the second spatial dimension is recorded sequentially due to the movement of the sample relative to the SI sensor. By combining the slices, the second spatial axis can be derived, resulting in a full image. In contrast to the stop-motion requirement of wavelength scanning SI, spatial scanning SI actively requires a continuous relative movement between imager and sample. This property makes the principle of scanning SI preferable for many industrial applications, as it allows analysing moving samples, for example, on a conveyor belt, without loss of time and without having to disrupt existing, continuous processes. In contrast to the accumulation of single frames to a full spectral hypercube and the subsequent evaluation of the information contained therein, classification and/or quantification algorithms can be applied to the pixels’ spectra directly and in real time. Instead of full spectra, only the analysis results have to be stored for the single pixels. This significantly reduces the amount of data to be handled during image reconstruction and processing, and makes realtime evaluation of spectrally fully resolved spectroscopic images possible in the first place.

c

d

e

Intensity

b

Wavelength

a

W

av

ele

ng

th

l atia

n

itio

pos

Sp

Spatial position Figure 7.1 Spectral image for an observation line across five different polymer samples with different sizes (left to right: polystyrene, PS; polyoxymethylene, POM; high density polyethylene, PE–HD; polypropylene, PP; acrylonitrile–butadiene–styrene, ABS). The left-side image shows the twodimensional grey-scale image as delivered from the spectral imaging system, the right-hand image shows it as a three-dimensional profile plot to illustrate the spectral content.

MATERIALS ANALYSIS SYSTEMS

161

7.2.2 Sample–radiation interaction Diffuse reflection (DR) is the dominant method for the acquisition of spectroscopic images in material analysis. For most samples, in addition to specular (mirror) reflection off the surface, a part of the radiation penetrates into the bulk of a sample upon illumination. The penetrating radiation undergoes multiple sample interactions – reflection, refraction and diffraction – before parts of the light will eventually make it back to the surface. Owing to the multitude of sample interactions, this diffusely reflected radiation has a significantly higher information content about the sample than specularly reflected light with only a single radiation – sample interaction. The information depth is very much dependent on the wavelength and the sample, in particular its (optical) density, and can range from a few microns to tens of millimetres, thus allowing even bulk analysis for many materials. It follows that diffuse reflectance set-ups are highly suitable for qualitative material identification sensors, but less so for quantitative analysis. Absolute quantification would require a constant optical density of the investigated materials, a condition that is rarely met in real-world applications. A possible workaround is to quantify two or more materials in a sample relatively to each other. As the reflected radiation is emitted from the sample in a random direction, diffusely reflected radiation can be separated from, potentially sensor-blinding, specular reflections. Common techniques are off-angle positioning of the sensor with respect to the position(s) of the illumination source(s) and the use of polarisation filters. Application restrictions apply to optically clear samples with little to no scattering centres, thin samples on an absorbing background and dark samples. In either of these cases, the intensity of radiation diffusely reflected off such samples is frequently insufficient for spectral analysis. While dark objectives remain a problem, thin and/or transparent samples can be measured in transmission or in transflectance. Transmission arrangements for SI applications follow the classical transmission layout, with a radiation source on one side of a sample and the SI analyser on the other side. The information depth is either constant or can easily be measured, which is an advantage in particular for (semi-)quantitative analysis. Although occasionally applied, such transmission systems suffer from the restriction that the sample must either be freely suspended or rest on a support that is transparent in the investigated spectral range. Transflection is essentially a cross between transmission and reflection. When light is shined onto a reflective surface covered by an optically clear sample, two processes occur – reflection off the top surface of the sample and transmission and reflection off the mirror, followed by a second transmission. As for most materials the surface reflection is low in comparison to the reflection off the mirror, the total collected radiation corresponds to a transmission measurement with double pathlength. Instead of a transparent support, transflection systems require only a spectrally neutral, or at least constant, broadband reflector under the sample. The optical layout depends on the reflector. With diffusely reflecting reflectors, for example dense polytetrafluoroethylene (PTFE) or compacted titanium dioxide,

162

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

the layout is essentially identical to a DR set-up. DR transflectance arrangements are particularly useful for the measurement of transparent and/or thin samples, for example, plastic films. A setback is the significantly higher effort for system implementation and maintenance, since the transport system, or at least a part of it in the sensing zone, has to be broadband-reflecting and considerable care must be taken to avoid soiling. For the analysis of clear organic coatings, in particular on metal, placing the SI analyser in a position where it records the specularly reflected radiation – attenuated by two transmissions through the material to be analysed – is preferable. As the metal under the coating acts as mirror reflector, such systems have high reliability and low maintenance requirements, provided the reflection properties of the coated metal remain reasonably constant.

7.3 Instrumentation The currently available instrumentation allows building affordable, industrially applicable NIR SI systems up to around 1800 nm. While research is going on to extend the wavelength range towards longer wavelengths, the limiting factor at present is the non-availability of suitable NIR cameras at competitive prices. A further complication is that high-resolution NIR detector devices are also used in military surveillance, tracking and weapon guidance systems. Therefore, export restrictions apply frequently. The system hardware of a real-time NIR SI system is very much reliant on methodology, wavelength range and application requirements. Apart from the materials that have to be detected, sample size, required spatial resolution and the necessary throughput influence the design of an SI sensor unit. The following concentrates on the description of the most common case in NIR real-time material analysis and classification, a spatial-scanning SI system in DR mode. A typical spatial-scanning SI system for material analysis and classification is shown in Fig. 7.2. The instrument consists of several components: • • • • •

A camera (A) for the chosen spectral range, with application-matched resolution and frame rate; attached to it, a spectrally matched imaging spectrograph (B) and suitable optics; optical elements in the beam path, for example, a secondary line aperture (C) or beam stops; a spectrally matched, homogenous illumination (D) of the samples; and a sample transport system, for example, a conveyor belt (E) moving the samples (F) perpendicular to the observation line (G).

In the implemented system (Fig. 7.2, right image) additional elements are the control and data evaluation unit (c) and an actuator controlled by the SI sensor, in the shown case a pneumatic sorter (f).

163

MATERIALS ANALYSIS SYSTEMS

H

A H

a

B b f C

D D

d d

E

F

G

c

F e

Figure 7.2 Layout and practical example of an industrial SI system for qualitative material analysis and sorting. Left image: basic design; camera (A), imaging spectrograph (B), light path with a secondary line aperture (C), illumination system (D), conveyor belt (E) moving the samples (F) perpendicular to the observation line (G) and an (optional) second imaging system (H). Right image: industrial unit for a 1800 mm conveyor belt moving at 2.2 m s−1 ; SI unit (a), beam path compartment (b), control unit (c), near-IR radiation sources (d), conveyor belt (e) and pneumatic sorter (f) controlled by the shown SI sensor. A

B C

e

D b

c

d

c

a

Figure 7.3 Layout of an SI unit based on an ImSpector® (SpecIm, SF) imaging spectrograph; A: nearIR camera with two-dimensional-InGaAs chip (e), B: ImSpector® with line aperture (b), transfer optics (c) and PGP spectral analyser component (d); C: view optics; D: sample plane with observation line (a). Reproduced with permission from SpecIm after an optical rendering.

The core of an industrially applicable SI sensor is the imaging spectrograph coupled to a spectrally matched NIR camera. For real-time SI applications using a spatial-scanning arrangement, the most suitable imaging spectrograph on the market is based on a direct vision diffraction prism-grating-prism (PGP) component, featuring a transmission volume grating cemented, together with a shortpass and a longpass filter, between two optically matched prisms (ImSpector® , SpecIm, SF). The working principle is shown in Fig. 7.3. The radiation enters the imaging spectrograph (B) through a line slit aperture (b) and is projected by optics (c) onto the dispersive PGP element (d). The spectrally

164

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

resolved radiation is subsequently projected by another set of optics onto the camera chip (e) of the attached camera (A). The resulting two-dimensional image (Fig. 7.1) represents the radiation intensities for a defined spatial (spatial position axis) and spectral (wavelength axis) coordinate. The choice of imaging spectrograph, together with the camera resolution, determines the spectral range and the spectral resolution of the SI sensor. Attached to the front of the spectrograph are spectrally optimised optics (C), that determine the field of view (a) of the first spatial axis and, together with the camera resolution, influence the spatial resolution in both spatial axis. The optics are selected to the specific needs of an application and can range from telescope optics for remote sensing to an NIR-optimised microscope for material classification of microsamples. Imaging spectrograph, two-dimensional detector (camera) and optics form the actual ‘spectral imaging unit’. An SI system can feature one or several of these units (Fig. 7.2, item H), to cover a wider spectral range or to increase the achievable spatial resolution. Supplementary optical elements, such as beam stops or additional beam apertures, can be introduced to increase the system performance and reduce stray light and other interferences. The second building block is the illumination unit. Depending on the application, broadband light sources ranging from LEDs to specially modified kW-floodlights may be used. The light sources must provide an even, stable illumination of the sample, as shadows or other inhomogeneities in illumination may lead, for example, to erroneous classification results. The third module in a spatial-scanning SI system is a sample transport unit providing a relative movement of the samples perpendicular to the first spatial axis (observation line) at a constant speed. In industrial applications, mostly microprocessor-controlled conveyor belts are used (Fig. 7.2, item e), but, in case of large, heavy or otherwise immobile items, it is also possible to mount the spectral imaging unit(s) and the illumination onto a movement system and move them relative to the stationary, investigated object. The movement velocity together with the frame rate of the camera determine the width of the observation line and hence the resolution along the second spatial axis. By varying the two parameters, the resolution can be adapted to meet specific application requirements. The fourth module is the control unit, consisting primarily of a state-of-the-art industry PC equipped with a suitable frame-grabber card for image frame acquisition, spectral and spatial real-time data evaluation and control of subsequent actuators or communication with process control systems. The technical data of a typical industrial SI system for macroscopic samples, as shown in Fig. 7.2, are detailed in Table 7.1.

7.4 Real-time data analysis In contrast to environmental or military remote sensing applications, where huge amounts of data are stored and processed off-line using highly evolved chemometric

MATERIALS ANALYSIS SYSTEMS

165

Table 7.1 Technical data of a typical industrial real-time spectroscopic imaging system for material analysis and sorting (SI system – technical data)

Wavelength Spectral resolution Spatial resolution 1. axis Spatial resolution 2. axis Band speed VIS unit

1000–1700 nm <10 nm 1/200–1/1000 of the inspected line (absolute: cm–μm, depending on optics) Customisable Up to 2.5 m s−1 Add-on-option

algorithms, in industrial material analysis systems the data evaluation must be performed on-line and in real time. Hence, suitable data pre-processing and an application-oriented spectral evaluation with optimised parameters are crucial factors. On the average, the total time for acquisition, processing, data evaluation and post-processing of one sampling line is less than 15 ms. To complete the entire data processing in such a short time using a standard PC, all operations must be carefully time-optimised. In particular, suitable pre-processing and fast but yet robust and reliable evaluation algorithms are essential. In addition, the overall system must be easy to use and provide the necessary interfaces to subsequent processing units dependent on the sensor data or an integrated process control system.

7.4.1 Pre-processing The raw images recorded by the camera have to be pre-processed to (1) correct for inevitable measurement deficiencies of the acquisition system, (2) reduce the amount of data that has to be processed to the necessary minimum and (3) prepare the data for the chosen evaluation algorithm. The first step is a removal of ‘dead’ pixels from the data raw images, followed by (1) dark-current correction based on a dark-current image B(x, λ) and (2) reference scaling against a broadband NIR white diffuse reflectance standard (reference image W (x, λ)). The actual reflectance images R(x, λ) are calculated from the recorded sample raw image X(x, λ) and the two reference images according to Equation (7.1). R(x, λ) =

X(x, λ) − B(x, λ) W (x, λ) − B(x, λ)

(7.1)

The second step is data reduction, by spectral range selection or selective binning of redundant data. The algorithms involved herein are optimised to reduce the data that has to be processed to the required minimum without losing relevant information. The proper execution of data reduction directly influences the real-time capabilities of the system. The final pre-processing step is data preparation for the evaluation. The particular operations are reliant on the chosen evaluation algorithm. Generally, a baseline

166

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

correction and/or the calculation of the first or second derivatives using a standard algorithm (e.g. Savitzky Golay) are involved. For applications involving only qualitative analysis or relative quantification, a spectral normalisation, based, for example, on the Euclidean distance (L2) norm algorithm, may be useful.

7.4.2 Spectral data evaluation The selection of an optimal spectral data evaluation algorithm is essential for satisfactory system performance, but is usually not easily predictable. Apart from the chemometrical performance, the execution time of the algorithm is crucial for realtime systems. As the execution time depends mainly on the number of mathematical operations of the algorithm, expressed by the run-time complexity, a mathematically simpler method involving fewer operations is often preferable to a (potentially) more powerful method that takes longer to calculate.

7.4.2.1 Qualitative analysis/classification Spatially resolved material identification and classification is currently the prevalent application for SI systems. Of the many powerful spectral classifiers available, only two types, each with a number of different algorithms,14 could successfully be applied for real-time SI applications: discriminant classifiers and dissimilarity-based classifiers. In addition, occasionally dedicated algorithms, such as fuzzy-classifiers, may be useful for special applications, for example, when there is no ab inito knowledge about the number and properties of the classification classes. 7.4.2.1.1 Discriminant classifiers. The two most important discriminant classifiers for material analysis using spectroscopic imaging systems are the Fisher linear discriminant classifier (FLDC) and the quadratic discriminant classifier (QDC). Other classfiers, such as the classical linear disriminant classifier (LDC), have frequently exhibited an inferior performance. Prior to the actual classification, the FLDC performs a linear mapping to a lower dimensional subspace optimised for class separability, based on the betweenclass scatter and the within-class scatter of the training set. In classification, each sample is assigned to the class giving the highest log-likelihood using a linear classifier. The QDC calculates a mean and the covariance for each class, allowing the classifier to find more accurate discriminative functions. Each sample is assigned to the class with the highest log-likelihood calculated following Equation (7.2), with the class mean μ and the class covariance . (x − μ)T  −1 (x − μ) 1 − ln(det ) + const (7.2) 2 2 Assignments with likelihoods below a preset threshold value are rejected and classified into a common rejection class. By adjusting the rejection threshold, the tolerance ln(p(x)) = −

MATERIALS ANALYSIS SYSTEMS

167

of the classification against sample deviations and modifications can be optimised. This algorithm is statistically very powerful, but since usually a PCA of the input data is required to reduce the dimensionality, the algorithm is restricted to applications with lower amounts of data, that is, slower processes or imaging at comparatively low spatial resolutions.

7.4.2.1.2 Dissimilarity-based classifiers. Dissimilarity-based classifiers (DBCs) use a dissimilarity measure MD to transform the input data into a dissimilarity space, where each trained class represents a separate dimension. Any dissimilarity measure with the following properties can be used:   0 x=y (7.3) MD (x, y) = >0 x  = y By designing the dissimilarity measure, it is possible to use expert knowledge or application-dependent properties to construct an efficient dissimilarity space, where classification is often easier than in the original feature space. Suitable discriminative dissimilarity measures for the real-time classification of the spectra are, for example, the Euclidian distance, the cumulative sums of spectral differences or selected geometric measures. The algorithm chooses the class with the smallest distance between the sample spectrum and the class reference spectrum. Individually adjustable class thresholds allow optimising the generalisation properties of the DBC method, thus determining how tolerant the classifier is against sample deviations and chance variations of the spectra. Spectra not falling into any of the match zones are assigned to a rejection class. In the practical application, DBCs work best for a limited number of possible classes, typically up to around 20. If this requirement is met, stable real-time classification can also be achieved for both spectrally and spatially highly resolved spectra.

7.4.2.2 (Semi-)Quantitative analysis Quantitative analysis applied on SI data in real time is a considerable challenge. Most algorithms that are capable of more than just integrating the intensity under a peak, and perhaps relating that to another peak, involve complex mathematical operations. While able to evaluate one spectrum in real time, this is usually not true when having to extract quantitative data from around a hundred spectra in parallel. One possible workaround is to reduce quantification tasks to classification problems by defining a number of class representatives for certain concentrations of the analytes. For process control it may be sufficient to define nominal conditions plus a tolerance field as class ‘Acceptable’, a wider range as class ‘Warning’ and everything else as ‘Out of Tolerance’. Although this will be only a semi-quantitative analysis with preset readings instead of a classical, continuous calibration, such a solution is sufficient for many process control applications. The reduction into a

168

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

classification problem is particularly successful for relative quantification, meaning that the required information can be extracted from the relative intensities and/or positions of the spectral features; it can also be used for absolute quantification of major components. For analysis of minor components down to trace amounts or when continuous calibration curves are necessary, it is possible in certain cases to use specially adapted, mathematically simplified and hence speeded-up algorithms. Anyhow, here, even more than with classification tasks, the performance depends very much on the selection of a suitable algorithm and the careful adaptation towards the specific problem. In any case, such algorithms frequently are not overly stable against unforeseen spectral interferences.

7.5 Integrated image processing Spectroscopic images contain a wealth of information, not only in the spectral dimension but also in the spatial dimensions. By applying smart image processing algorithms to extract and enhance the spatial information contained in the acquired data, the practical importance of SI material analysis systems can be significantly improved. Parameters that can be extracted and classified by computer programs encompass sample size, shape, homogeneity, structure, etc. A typical application is the determination of position, size and orientation of an object together with the material information as control parameters for an automatic sorter. For clean, non-overlapping samples it is sufficient to determine the centre of each object and the approximate size to control the sorting unit. However, frequently these ideal conditions cannot be guaranteed, as samples overlap or surface adhesions may interfere with the classification, as shown in Fig. 7.4. Such artefacts will lead to (partially) false information about size and position of an object and hence to incomplete or erroneous sorting. As an example, a bottle (A, B, C, D) with a (paper) label should be detected as one single object, not as two separate entities with a piece of paper between (A, B) or on top (C) and a different polymer part next to it (PE cap in object A, PP caps in objects B, C, D). Otherwise, based on the classification result, a subsequent sorting or turnout stage would try to handle all these as separate objects (as indicated by the blob centre markers in Fig. 7.4(c)) instead of one object. This requires an image processing step to process the classification results and deliver the correct size and shape of the different objects, regardless of the presence of, for example, interfering paper labels that partially obscure the objects’ surfaces. To achieve that, an object reconstruction algorithm detects connected components (blobs) in the binary images of each material class. In a first step, blobs that do not meet pre-defined size restrictions are considered as classification errors and filtered out (e.g. the black line due to a faulty camera pixel in object D). The binary image of each material class (material and overlay) is then morphologically dilated

MATERIALS ANALYSIS SYSTEMS

(a)

(b)

(c)

(d)

169

Figure 7.4 Classification and image processing results of a typical situation in polymer waste recycling; (a): digital image; (b): initial classification result; (c): calculation of separation data based on the initial classification result; (d): classification result after real-time image processing. A, B: Polyethylene terephthalate (PET) bottles with paper labels, C: PE bottle with paper label, D: PE bottle with PE film label, E: PP cup, F: PS cup. Classification colour code: red: high-density PE; green: PS; dark blue: PET; yellow: PP; light blue: paper.

with a box mask and the dilated paper blobs are overlaid with the dilated blobs of each polymer class to detect overlaps.19 These overlaps identify regions in the classification result where object material and overlay substance regions touch each other. Adjacent regions will be merged to a single blob, if (1) the material blob surrounds the overlay blob and the centre of the overlay blob and the material blob lie inside a circle with a certain diameter or (2) the centre of the overlay blob lies between two material blobs. If one of these conditions is met, the whole object is defined as one material blob (Fig. 7.4c/d), which significantly increases the sorting efficiency and overall process stability. Further applications can be found, for example, in process control, where an image processing algorithm classifies flaws in a coating according to size and distribution and makes the data available to the process management system. This data allows close monitoring of the process and optimised pre-failure maintenance, for example, in nozzle cleaning, thus significantly reducing outage times.

7.6 Material analysis applications 7.6.1 Industrial waste classification and sorting One application for an automated real-time classification system is sorting of waste into pure fractions for material recycling. Concurrently with boosting production and use, the amount of waste increases. Although efforts are being made for material recycling rather than landfill deposition and thermal treatment, the main stumbling block for practical large-scale applications is the purity of fractions. Currently in only

170

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

about 30% of the total consumer plastic waste can be recycled and even then mostly low-quality products are being produced. Standard methods for the separation of polymers in waste involve a lot of manual work, making recycled material frequently more expensive than fresh polymers and hence rendering the recycling pointless from the economical point of view. In waste sorting for recycling, a reflectance-type NIR-SI system as sensor15 allows the identification of practically any important polymer (Fig. 7.4) both in municipal and industrial polymer waste and allows to sort them accordingly in real time into pure material fractions, for example with a pneumatic line sorter. During sorting, only positively identified items are sorted into the individual containers, while non-polymer waste, unknown materials and (heavily) contaminated items are ejected from the system and can be recycled thermally. The real-time classification is achieved using a specially adapted dissimilarity-based classifier (DBC-NN). When applying an image-processing algorithm on the classification results, it is even possible to distinguish, for example, between PET bottles without a cap, those with a PE cap and those with a PP cap (Fig. 7.4). Based on the same principle, it is also possible to distinguish between highquality deinking paper and low-quality non-deinking fractions, like cardboard, even if they are laminated with a high-quality top layer finish. These SI sorting systems operate over conveyor belts with a width between 1.5 and 2.2 m, moving at up to 2.5 m s−1 . A different waste grading application is shown in Fig. 7.5. Here, shredded mixed waste (size around 2×2 cm) from a recycling plant for electric and electronic equipment is analysed and graded according to its calorific value. This can be achieved with sufficient accuracy by determining the relative abundance of some key polymers with respect to the entity. This is of importance since especially with highly diverse, mixed waste and overlaps of already small objects, which are encountered in a stream of shredded material, a full identification is almost impossible, in particular in real time. In addition to the prediction of the calorific value, a screening for certain key materials can also be performed. Currently, polyvinyl chloride (PVC), in particular, is regarded as problematic, as it will emit hydrogen chloride fumes during incineration and is one source for the formation of highly toxic dioxins. Depending on the classification of the incinerator, only a certain amount of PVC is acceptable, while above that level either the whole batch of the material has to be incinerated elsewhere or deposited.

7.6.2 Surface coating inspection A typical example of semi-quantitative analysis of an organic coating on a metal surface is shown in Fig. 7.6. A matted metal sheet is industrially spray-coated with a colourless two-component coating. The task was to determine not only the coverage, something that could also be achieved by using reflectance sensors, but also whether the mixing ratio is within the specification.

MATERIALS ANALYSIS SYSTEMS

171

Figure 7.5 Classification result of shredded mixed electronics waste, superimposed onto a greyscale image of the waste fraction. Only objects that could be identified with a likelihood >95% have been classified. Classification colour code: yellow: metal, mostly aluminium; red: polymethyl metacrylate (PMMA); orange: polyolefines (PE, PP); pink: styrene polymers (PS, PS–E, etc.); violet: polyamides (PA 6, PA 6.6, etc.); green: acrylonitrile–butadiene–styrene (ABS); blue: polyvinyl chloride (PVC).

By using NIR imaging spectroscopy in a diffuse transflectance set-up, using the matted metal sheet as diffuse reflector, it is possible to detect coverage, mixing ratio of the coating and damages in the metal surface. By applying an imaging algorithm, it is additionally possible to automatically classify the detected abnormalities according to size as well as position and feed that data into the process control system.

7.6.3 Food control Near-IR sensors are also a topic of increased interest for the analysis of biological materials, from food quality control to bioactivity measurements. First applications

172

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Figure 7.6 Semi-quantitative analysis of an organic two-component coating on a metal surface. The classifier grades the results into ‘within specification’ (light grey), ‘warning’ (white), mixing ratio/coverage outside specification (dark grey) and surface defects on the metal carrier (black). A: significant isolated problem, B: localised constant problem, C: acceptable localised random flaws, E: start of massive failure, F: localised application of primarily one component only due to spray nozzle obstruction.

encompass (1) ripeness measurements of fruits and vegetables, based both on VIS measurement of the surface and NIR measurements correlating to internal properties, for example, firmness, total sugars or acidity; (2) foreign objects in or among food and (3) defect and edibility control, checking, for example, for ruptures, rotten or mildewed areas on the surface or hidden in the object, as well as surface contaminations, for example, on poultry after depluming.16,17.

7.6.4 Mineralogical material analysis Gemstones are frequently subject to alterations or plain forgeries to achieve higher prices. For example, turquoises (CuAl6 [(OH)2 /PO4 ]4 · 4H2 O) are popular gems that are increasingly rare and at the same time easy to ‘polish up’ or counterfeit. Various metal ions can give the stones shades from green to blue, making pure colour measurement insufficient for a reliable identification, while standard analytical methods are destructive and/or time-consuming and expensive. Hence, a simple, quick and reliable method to detect fraudulent jewellery would be of high interest.

173

MATERIALS ANALYSIS SYSTEMS

a

b

c

d

e

f

g

h

i

Figure 7.7 Classification results for nine turquoise samples. Left: digital image; right: combined VIS/NIR material classification result. Samples a, e and f are true, massive turquoises, sample b is turquoise set in silicate bedrock, h is a true turquoise with different ionic dopants and hence a different colour, sample d is true turquoise powder pressed with a polymeric binder, g is an artificially produced turquoise and c and i are other minerals that look superficially similar to turquoise and are frequently used in counterfeits.

With a combined VIS + NIR SI system and a fuzzy-C-means classification algorithm, it was possible to set up a fast, compact, contact-less and non-destructive analytical instrument. As shown in Fig. 7.7, the two-range spectroscopic imager is capable of distinguishing among true turquoises, even if they are set into bedrock (b), treated turquoises (pressed turquoise powder, etc.), synthetic turquoises and nonturquoise materials without sample pre-treatment with high reliability.18

References [1] Chalmers, J. M. (ed.) (2002) Handbook of Vibrational Spectroscopy, John Wiley & Sons, New York. [2] Meyers, R. A. (ed.) (2000) Encyclopedia of Analytical Chemistry, Vols. 12 and 15, John Wiley & Sons, New York. [3] van der Meer, F. and De Jong, S. M. (eds) (2002) Imaging Spectrometry, Kluwer Academic Publishers, Dordrecht. [4] Bearman, G. H., Levenson, R. M. and Cabib, D. (eds) (2000) Spectral imaging: instrumentation, applications, and analysis. Proc. SPIE 3920. [5] Smith, R. D., Nelson, M. P. and Treado, P. J. (2000) Proc. SPIE 3920, 14. [6] Hopkins, M. F. (1995) Proc. SPIE 2599, 294. [7] Eisenreich, N. and Rohe, T. (2000) Encyclopedia of Analytical Chemistry, John Wiley & Sons, New York, pp. 7623–44. [8] Gat, N. (2000) Proc. SPIE 4056, 50. [9] Carstensen, J. M. (1999) European Patent Application EP 1 051 660 B1. [10] Fately, W. G., Hammaker, R. M., DeVerse, R. A., Coifman, R. R. and Geshwind, F. B. (2002) Vib. Spectrosc. 29, 163. [11] Aikio, M. (1993) European Patent Application EP 0 635 138 B1. [12] Gat, N. (1992) US Patent Application US 5 166 755. [13] Faeley, W., Coifman, R. R., Geshwind, F. and DeVerse, R. A. (2003) PCT Patent Application WO 03 023 692 A1. [14] Duda, R. O., Hart, P. E. and Stork, D. G. (2000) Pattern Classification, 2nd edn, John Wiley & Sons, New York.

174

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

[15] Kulcke, A., Gurschler, C., Spö ck, G., Leitner, R. and Kraft, M. (2003) J. Near Infrared Spectrosc. 11, 71. [16] Lawrence, K. C., Windham, W. R., Park, B. and Buhr, R. J. NIR News 12, 3. [17] Windham, R. W., Lawrence, K. C., Park, B. et al. (2002) PCT Patent Application WO 02 063 939 A2. [18] Gurschler, C., Serafino, G., Spö ck, G., Del Bianco, A. and Kraft, M. (2002) Proc. OPTO, p. 10. [19] Leitner, R., Mairer, H. and Kercek, A. (2003) Real-time imaging 8, 245.

8

Industrial applications of near-IR imaging Anthony E. Dowrey, Gloria M. Story and Curtis Marcott

8.1 Introduction The near-infrared (NIR) region of the electromagnetic spectrum has seen increasing research activity in recent years.1–6 This region of the spectrum, between 14 285 and 4000 cm−1 (700–2500 nm), is between the visible portion of the spectrum, where electronic transitions occur, and the mid-infrared (mid-IR) region, where fundamental vibrations absorb. The NIR spectral region is generally where overtone and combination bands of the fundamental vibrations absorb, but some low-lying electronic transitions also absorb in this region. Overtone and combination bands of vibrational fundamentals are due to anharmonic transitions, that is, they result from higher-order perturbations to the harmonic oscillator model of molecular vibrations. Since they are second-order or higher-order transitions, they show less absorbance intensity, typically 1–2 orders of magnitude lower for first overtone and combination bands. Higher-order transitions are even weaker. Since only vibrational fundamentals with frequencies above 2000 cm−1 will have first overtone and combination bands lying above 4000 cm−1 in the NIR, virtually all the vibrational absorptions of interest involve the motion of hydrogen atoms (overtones and combinations of CH-, OH- and NH-fundamental vibrations). The reduced intensity of absorptions occurring in the NIR spectral region leads to certain advantages, when compared with analyses in the mid-IR spectral region. NIR spectroscopy can be performed successfully with much thicker samples (than for mid-IR), using either transmission or reflectance modes of spectral data collection.6 Also, low-cost materials are available for working in the NIR region and many of these materials are suitable for use in a manufacturing setting for on-line or process monitoring. As a result, NIR instrumentation is generally less expensive than analogous mid-IR instrumentation and tends to be more robust and less chemically sensitive than mid-IR spectrometers. Lastly, optical fiber technology, spurred by advances in optical materials for communications applications, lends itself well to NIR spectroscopy, allowing remote detection.7 Until recently, nearly all forms of analytical measurement in the mid-IR and NIR spectroscopic regions involved detection at a single detector element where all of the transmitted or reflected source energy was focused and measured. This

176

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

‘single point’ of measurement sometimes involved sample areas as small as a few microns using microscope optics to deliver the energy, or sample areas as large as several square centimeters using macroscopic optical arrangements. As long as the single-element detector does not saturate due to too much energy reaching it, higher sensitivity can usually be achieved when larger areas of sample are illuminated and all of the available energy is focused onto a single-point detector. However, collecting data from large sample areas can become a disadvantage if a sample is not homogeneous. Microscopic sampling techniques may be necessary to evaluate small areas of specific interest or nonhomogeneous samples. But microscopic sample analysis can become difficult and tedious if the material of interest does not lend itself well to microscopic sampling, or if many different areas need to be analyzed. Focalplane array (FPA) detectors are devices containing a grid of individual detector elements arranged in an ‘array’, similar to those currently used for digital cameras.8 In fact, an FPA detector can be thought of, and is often referred to, as a ‘camera’. What differentiates FPAs used for spectroscopic imaging from those used for visible imaging (digital cameras) is that each individual FPA element is effectively a miniature version of a standard spectroscopic detector element. In other words, it has the capacity to sense changes in the amount of radiation impinging at that small point, just like a single-element detector. The result is that it is possible to collect a full spectrum (depending on the spectral range of the FPA and the optics used) at each of the FPA detector elements. FPA detectors used in the NIR spectral imaging studies described here include a 1×16 mercury–cadmium–telluride (MCT) linear array, 320 × 256 and 320 × 240 indium–gallium–arsenide (InGaAs) FPAs and a 320 × 256 indium–antimonide (InSb) FPA. The resulting data collected from an FPA detector designed for spectroscopic imaging is a data ‘cube’, where spatial information is collected in the x–y plane and spectral information is collected along the z-axis. One can make images of the absorption of a sample at a particular wavelength to show distribution of a chemical species exhibiting absorption at the wavelength chosen. It is also possible to extract spectral data at either a particular image point, or from several points combined together to show an average spectrum representing the x–y region chosen. In addition, advanced mathematical manipulations, such as integrated peak areas, first and second derivatives of each spectrum and principal component analysis (PCA) can be applied to spectral imaging datasets in order to enhance the differences among the data. A detailed account of the mathematical processing of spectral imaging datasets is provided in Chapter 4. In summary, spectroscopic imaging in the NIR spectral region can make good use of the advantages associated with NIR analysis. Thicker samples can be imaged. Reflection or transmission can be used to collect data. Less expensive optics and detectors/cameras are available to accomplish these tasks in the NIR spectral region. And most of these are more mechanically and chemically robust than that available for the mid-IR spectral region, and often do not usually require special cooling apparatus.

INDUSTRIAL APPLICATIONS OF NIR IMAGING

177

8.2 Experimental Four different NIR spectral imaging instruments were used in the applications presented in this chapter. An NIR spectral imaging system designed for the specific task of imaging water was constructed from a 320 × 240 InGaAs FPA (Sensors Unlimited) which is sensitive between 1100 and 1700 nm. The first overtone of the OH-stretching fundamental vibration of water, which occurs near 1400 nm, is one of the strongest absorbers in this spectral range, making this particular camera a good one for imaging water. In addition to the InGaAs FPA, our ‘water camera’ has dual eight-position filter wheels mounted between the FPA and the lens, and a tungsten–halogen fiber-optic ring lamp that surrounds the f/1 camera lens. One filter wheel contains various 1 inch optical filters that can be switched into the beam path under computer control. The second filter wheel contains various neutral density filters that are used to compensate for the fact that the camera response varies depending on the peak wavelength and bandwidth of each optical filter. In order to make quantitative measurements, it is important not to readjust the gain and offset on the camera itself between collection of the sample and background wavelengths. For qualitative studies, like the one described in this chapter, images are simply collected through a 1.4 μm optical bandpass filter at 10 frames s−1 . Analysis time is determined by the length of time needed to capture the event of interest. The second NIR imaging system used for applications discussed in this chapter is a commercial MatrixNIR® system manufactured by Spectral Dimensions, Inc. It consists of a 320 × 256 InGaAs FPA camera coupled to a liquid crystal tunable filter (LCTF) capable of tuning a 10 nm bandpass through a spectral range from 1100 to 1800 nm. By using different focusing lenses, the spatial resolution of this system can be varied from 5.7 μm × 5.7 μm per pixel (field of view – FOV = 1.82 mm × 1.46 mm) to 181 μm × 181 μm per pixel (FOV = 5.8 cm × 4.6 cm). Energy is supplied to the sample by four tungsten–halogen lamps. Analysis time using this system is typically 5–10 min. Measurements were also made on a Sapphire® NIR imaging system at Spectral Dimensions, Inc. (Olney, MD). This system is similar to the MatrixNIR® , but has a spectral range from 1600 to 2400 nm which is achieved by using an extended wavelength range LCTF and a 320 × 256 InSb FPA. Finally, a commercial Perkin-Elmer Spotlight® 300 Fourier transform infrared (FTIR) imaging system equipped with a 1×16 MCT linear array detector was found to have sufficient sensitivity at 4323 cm−1 (2313 nm) where a long-chain CH2 bendstretch combination band absorbs to do quantitative studies. This system was used to generate NIR images over a 5 mm × 5 mm area of the sample with an individual pixel size of 25 μm × 25 μm. This system uses a motorized ‘mapping’ sample stage to move the sample into position for each 1 × 16 point data collection as it collects the multiple measurements needed to completely cover the area of interest. Analysis time using this system depends on the size of the area to be analyzed; typically about 50 min for a 5 × 5 mm2 area. Once the spectral data cubes have been collected, they are often converted to another file format where the data can be examined in more detail. Two software

178

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

packages used in this chapter to further analyze NIR spectral imaging data are ISYS® (Spectral Dimensions, Inc., Olney, MD) and ENVI® (Research Systems, Inc., Boulder, CO).

8.3 Application using NIR spectroscopic imaging We will now discuss some examples where NIR spectroscopic imaging has been used to evaluate some materials of commercial interest. This will include examples from a variety of home/commercial product areas including fabric wetting, spray deposition, food composition and surfactant deposition on nonwoven materials. These examples will demonstrate both qualitative and quantitative uses of NIR spectral imaging for determining chemical distributions in nonhomogeneous materials.

8.3.1 Water migration on fabrics One of the easiest substances to image in the NIR spectral region is water. It is also an important substance to be able to monitor in industries concerned with absorbent materials, such as toilet tissue, paper towels, diapers, catamenial products and fabrics. The NIR images shown in Fig. 8.1 were collected with a 320 × 240 element InGaAs FPA through a 1.4 μm optical bandpass filter. The four images shown in the top row of Fig. 8.1 were captured just as a drop of water from a pipette hit the fabric. All four fabric samples were from the same new fabric pretreated to provide stain resistance. Fabrics treated for stain resistance also tend to be water resistant. The fabric sample on the far left had never been washed. The second fabric from the left had been washed once in a detergent, the third 5 times and the fabric on the far right had been washed 10 times prior to the experiment. The four images shown in the second row of Fig. 8.1 were collected after the drop of water had been allowed to spread for 0.5 s (five frames later). It can be clearly seen that the successive detergent washings are affecting the surface energy of the fabric and allowing the water droplet to spread more quickly.

t=0 s

Unwashed

1 cycle

5 cycles

10 cycles

t = 0.5 s

Figure 8.1 Images collected at 0 and 0.5 s of a water droplet applied to unwashed, pretreated fabric and fabric washed 1, 5 and 10 cycles.

INDUSTRIAL APPLICATIONS OF NIR IMAGING

179

This likely is the result of the washing removing some of the pretreatment material that is often applied to new fabrics. It would not be possible to monitor this type of water migration with a visible camera, as water does not absorb light in the visible region of the spectrum. But by using an NIR imaging experiment, specifically designed to visualize water, this becomes relatively easy.

8.3.2 Spray nozzle patterns Figure 8.2 shows an example where NIR spectral imaging was used to perform the seemingly simple task of measuring the distribution of material deposited from a spray nozzle. The purpose of this test was to aid in the development of more efficient spray delivery systems. Imaging was accomplished using a macro-lens attached to the MatrixNIR® system described earlier. The resulting FOV was 5.8 cm × 4.6 cm and the spectra were collected from 1100 to 1800 nm. Images were collected from a pattern deposited on a paper substrate, sprayed from a distance of about 3 in. Evaluation of the spectral data showed that a component of the material being sprayed exhibited a significant absorption near 1480 nm (CH-bend fundamental combined with CH-stretch first overtone). This absorption was evident even after the spray had dried. Once this feature of the sprayed material was identified, it was a simple matter to apply peak area integration for the identified absorption across the entire image. The images shown on the left in the figure are maps of the peak area of the 1480 nm NIR absorption. These images, obtained in <10 min, led to a method for rapidly assessing the performance of spray delivery nozzles. Spatial distribution of spray, as well as some quantitative information about distribution uniformity, was quickly obtained. This test replaced a previous method where spray was deposited onto a 96-well plate (shown on the right half of the figure). The previous method required individual measurements at each of the 96 wells, took several hours to complete and resulted in much poorer spatial resolution. The NIR spectral imaging method provided a much quicker test.

8.3.3 Surfactant deposition on a nonwoven substrate Figure 8.3 shows the distribution of multiple materials deposited on a multiple-ply nonwoven substrate. This substrate included a polymer mesh material embedded between two nonwoven layers to enhance strength. The 5.8 cm × 4.6 cm image in Fig. 8.3 shows three separate areas on the substrate where different surface composition is observed. The image on the left of Fig. 8.3 was constructed by taking the first derivative of all NIR spectral data in the data cube. First-derivative spectra tend to negate the effect of baseline fluctuation in the raw data and can also aid in resolving overlapping absorptions. PCA score plots were used to identify five major component groups from within this dataset. These unique groups of similar spectra were identified using the ENVI® software package to visualize the principal component scores

Pixels

250

200

150

100

50

250

200

150

100

50

100 150 200 250 300 Pixels

0.06 0.05 0.04 0.03 0.02 0.01 0 –0.01 –0.02

–0.02

0

0.02

0.04

0.06

0.08

100 150 200 250 300 Pixels Four-hole nozzle deposition 1480-nm images

50

Single-hole nozzle deposition

50

NIR images of spray deposition on filter paper

0 0 0 0 0 0 0 0 0 9

0 0 0 0 0 0 0 0 0 4.5

14

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0

= 8–10% = 10–12%

= 0.1–2% = 2–4% = 4–6% = 6–8%

0 0 0 0 0 0 0

0 0 0 0 0 0 0

Percentage of total

18

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0

23

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0

27

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0

36

0 0.2 0.3 0.6 1.1 1.1 0.9 0.2 0

0 0 0 0 0 0 0

45

0.1 0.4 1.5 3.3 4.9 4.4 2.9 1.1 0.2

0 0 0 0 0 0 0

50

0.5 1.5 3.1 4.5 6.1 6.8 3.3 1.3 0.2

0 0 0 0 0 0 0

54

0.4 1.5 2.4 2.8 2.9 2.5 1.7 0.6 0.2

0 0 0 0 0 0 0

Distance in mm

41

0 0.4 0.6 1.5 1.5 2.5 1.3 0.4 0

0 0 0 0 0 0 0

59

0.3 0.7 1.2 2.4 2.5 2.5 2 0.7 0.1

0 0 0 0 0 0 0

63

0.2 0.4 0.7 2.5 2.3 1.7 1 0.5 0.1

0 0 0 0 0 0 0

0 0 0 0 0 0 0

68

0 0.2 0.3 1.2 1.1 0.3 0.2 0.1 0

96-well plate spray deposition chart

32

0 0 0 0.3 0.1 0.3 0 0 0

0 0 0 0 0 0 0

72

0 0 0 0.1 0 0 0 0 0

0 0 0 0 0 0 0

77

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0

81

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0

86

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0

90

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0

95

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0

36 41 45 50 54 59 63 68 72

4.5 9 14 18 23 27 32

Figure 8.2 NIR 1480 nm peak area images (5.8 cm × 4.6 cm) of a material deposited on a paper substrate from a spray nozzle, and a chart of measurements from spray material deposited on a 96-well plate.

Pixels

Absorbance Absorbance

INDUSTRIAL APPLICATIONS OF NIR IMAGING

181

Log(1/R)

0.005 0 –0.005

1100 1200 1300 1400 1500 1600 1700 Wavelength (nm) Figure 8.3 A 5.8 cm × 4.6 cm image colored according to component groups identified from a PCA score plot of the first derivative spectra of each pixel in the image. The average first derivative spectra for each group is shown on the right. The colors of the spectra correlate with the colors shown on the image.

from a rotatable n-dimensional scatter plot, where n is the total number of scores determined from the PCA calculation. The spectra on the right is the average first-derivative spectra of each of the five grouped sets of pixels determined from the score plots. The actual image shown was constructed by converting each pixel identified as being in one of the groups identified from the PCA three-dimensional score plot to a color assigned for that group. The colors used are the same for both the image and the average first-derivative spectra. The resulting image shows where these components reside spatially in the sample area analyzed.

8.3.4 Flavored chips An extended spectral range (1600–2400 nm) NIR imaging system (Sapphire® , Spectral Dimensions, Inc., Olney, MD) was used to examine component distributions in flavored, salted snack chips. In this case, PCA was performed on the spectral image data cubes using the ISYS® software. NIR spectral images of two different flavored chips from the same manufacturer were concatenated before the PCA analysis was performed on normalized spectra in the chip-containing regions of the linked images. Figure 8.4 shows three principal component loading spectra, which bear a striking similarity to the NIR spectra of starch, oil and water. The score images of these three principal components are shown on the right of Fig. 8.4 for the two different flavored chips. It is readily apparent that there are differences related to the components imaged, in particular, the chip on the left (flavor 1) contains more oil that the chip on the right (flavor 2). In addition, a comparison of the images of principal component #3 suggests that the water distribution is more uniform for the chip on the right (flavor 2). This example illustrates how NIR spectral imaging using a wavelength range extended to 2400 nm can be useful for understanding the distribution of components in manufactured food products.

182

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Flavor 1

Flavor 2

PC # 1 Score image

0 10 20 30 40

0.2

PC # 2 Score image

0 10

0 mm

PCA loading

mm

0.4

30 40

1800 2000 Wavelength (nm)

2200

Red: PC # 1 Loading (starch) Green: PC # 2 Loading (oil) Blue: PC # 3 Loading (water)

PC # 3 Score image

0 10 mm

–0.2 1600

20

20 30 40 0

20

40

60 mm

80

100

Figure 8.4 Principal component loading spectra and corresponding score images representative of starch, oil and water components on two different flavored chips.

8.3.5 Lotion distribution on nonwoven paper In this example application, the local basis weight of a long-chain hydrocarboncontaining lotion applied to a nonwoven paper sample is quantitatively determined by transmission spectral imaging in the NIR region. The top ply of a lotioned tissue, called Tempo® , was examined by FTIR spectral imaging in transmission mode (absorption spectroscopy) using a Perkin-Elmer Spectrum Spotlight® 300 instrument. Because of the sample thickness, the CH2 -stretching fundamental IR absorbances at 2920 and 2850 cm−1 that are typically used to visualize materials con−(CH2 )− − units are too intense taining linear hydrocarbon components of repeated − for quantitative determinations (absorbances > 1.0). As a result, it was necessary to use the spectral region between 4000 and 4500 cm−1 , in particular, the integrated area of the band centered at 4323 cm−1 , for analysis. This vibrational absorption, due to a bend-stretch combination band of a long-chain CH2 group, is about 50 times less intense than the corresponding CH2 -stretching fundamental vibration at 2920 cm−1 . Figure 8.5 shows a spectral image of a 5 mm × 5 mm area of the top ply of a Tempo® sheet scanned at a spatial resolution of 25 μm × 25 μm per pixel. The spectral resolution was 8 cm−1 . One of the two CH2 bend-stretch combination bands (4323 cm−1 ) was used to generate the image. We chose the end points for the band integration to be 4296 and 4368 cm−1 . Red areas represent pools, or puddles

183

Absorbance

CH-combination band region

4323

INDUSTRIAL APPLICATIONS OF NIR IMAGING Average of 4271 spectra in petrolatum pools

Average of all 39 800 spectra image

Average of 29 983 spectra with little petrolatum

4500

4400

4300 4200 Wavenumber

4100

Figure 8.5 NIR image of Tempo® lotioned nonwoven tissue generated from the integrated area under the band at 4323 cm−1 . Red areas represent puddles of petrolatum (lotion) with integrated area under the CH bend-stretch combination band at 4323 cm−1 > 0.18. Black areas represent thin areas or holes in the paper. Blue areas represent thick areas of the paper.

of long-chain hydrocarbon-containing lotion (e.g. petrolatum) where the integrated area under the CH2 bend-stretch combination band at 4323 cm−1 > 0.18. Black areas represent thin areas or holes in the paper. Blue areas represent thick areas of the paper. The red spectrum on the right of Fig. 8.5 represents the average spectrum of all the red pixels (those with the 4323 cm−1 band area >0.18). The black spectrum is the average of all 39 800 pixels in the entire 5 mm × 5 mm FOV. The green spectrum is the average of 29 983 spectra in the rest of the image where there are not locally high concentrations of lotion. Clearly, these NIR spectral imaging data demonstrate that the majority of the lotion is concentrated in about 11% of the area of the sheet. Figure 8.6 shows that the integrated area of the NIR combination band centered at 4323 cm−1 produced linear Beer’s law curves for a series of calibration samples with lotion loadings ranging from 0 to 72 g m−2 . Once the lotion loading level reached about 15 g m−2 , all of the surface area of the nonwoven paper sample is coated with lotion (above an arbitrary detection threshold of 0.2 g m−2 ). Evidence of lotion pools of higher concentration is apparent on all samples. Using this calibration plot, it is possible to generate spectral images where an equivalent lotion loading in units of g m−2 (or g (sqm)−1 ) can be determined for each pixel in the image. Figure 8.7 shows an image of the local lotion loading at each pixel (in g (sqm)−1 ) for a sample of Tempo® tissue (top) and another commercial tissue (bottom). The corresponding local loading of lotion for each pixel along the line drawn on the images is shown on the right. These linescans clearly illustrate that Tempo® has almost no lotion in most area, but a high concentration of lotion (∼20 g (sqm)−1 ) in a few local pools, while the other commercial tissue has a nearly uniform distribution of lotion at ∼5 g (sqm)−1 . This NIR imaging method is clearly able to differentiate these two types of lotioned tissues. There are other NIR bands besides the CH2 bend-stretch combination band at 4323 cm−1 that could have been used to measure the lotion distribution on Tempo®

184

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

4.50

Integrated peak absorption

4.00 3.50 3.00

y = 0.0509x R 2 = 0.9763

2.50 2.00 1.50 1.00 0.50 0.00

0

20 40 60 Lotion basis weight of calibration sample

80

Figure 8.6 Beer’s law plot constructed from calibration samples of Tempo® with known amounts of lotion plotted versus the integrated area under the NIR absorption band at 4323 cm−1 averaged over all pixels in the image.

sheets. Similar experiments to those done on the Perkin Elmer Spotlight® imaging system were done using a Spectral Dimensions MatrixNIR® instrument on a single top ply of commercial Tempo® product (lotion loading unknown). In this case, it was necessary to use second overtone CH-stretching absorption band areas near 1200 nm to create chemical maps of lotion level/distribution. Since the earlier experiments on the Perkin-Elmer Spotlight® instrument involved using the CH-stretch/bend combination absorption band areas near 2300 nm, we would expect that method to be more sensitive to the lotion concentration because these combination band absorptions are about 10 times more intense than second overtone absorptions near 1200 nm. On the other hand, the MatrixNIR® instrument can collect data over much larger sampling areas more quickly because it has a much larger format array detector (81 920 pixels versus 16 pixels for the Spotlight). The MatrixNIR® data shown in Figs 8.8–8.10 collected in reflectance using macro, 1× and 10× objectives illustrate the lotion level/distribution on three different size scales (181, 38, and 5.7 μm per pixel, respectively). These data suggest the MatrixNIR® instrument could also be used to quantitate lotion levels and distribution at loading levels of typical product.

8.4 Conclusions In this chapter, we talked about the advantages of using spectral imaging in the NIR spectral region compared with spectral imaging in the mid-IR spectral region (the former can more easily analyze thicker samples and use less expensive, more

0.00

5.00

10.00

15.00

20.00

25.00

30.00

35.00

–5.00

0.00

5.00

10.00

15.00

20.00

25.00

30.00

35.00

0

0

3000

Distance (µm)

2000

4000

1000

3000 Distance (µm)

2000

4000

Line scan across a commercial lotioned tissue

1000

Line scan across Tempo® tissue

5000

5000

Figure 8.7 NIR spectral images representative of the integrated area under the band at 4323 cm−1 for Tempo® (top) and another commercial lotioned tissue (bottom). The corresponding line scans on the right represent the amount of lotion determined for each pixel along the line shown on the image.

1000 μm

10 000 μm

Local lotion basis weight (g (sqm)–1) Local lotion basis weight (g (sqm)–1)

186

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

0 5

0.2

10

(mm)

15

0.18

20 0.16

25 30

0.14

35 40

0.12

45 0

10

20

30 mm

40

50

Figure 8.8 A 58 mm × 46 mm NIR image of a Tempo® top ply taken through a macroobjective on the MatrixNIR® instrument (Spectral Dimensions). This image is a band integration map of the lotion absorption (1160–1260 nm) from a 4 scan, 24 frame data collection smoothed 13 points and offset corrected. The white box in the upper left shows the 5 mm × 5 mm size of images collected by the Perkin-Elmer Spotlight® instrument (see Figs 8.6 and 8.7) for comparison.

0

0.2

1

0.18

2

0.16

3

0.14

mm

4

0.12

5 0.1 6 0.08

7

0.06

8

0.04

9 0

2

4

6 mm

8

10

12

Figure 8.9 A 12.1 mm × 9.8 mm NIR image of a Tempo® top ply taken through a 1× objective on the MatrixNIR® instrument. The white box in the upper left shows the 5 mm × 5 mm size of images collected using the Perkin-Elmer Spotlight® instrument for comparison.

robust optics and instrumentation). We implemented several types of FPA detectors and instrumentation for the applications described in this chapter. We also discussed the data ‘cube’ one gets from an analysis using an FPA and some of the mathematical manipulations (first derivative, PCA) one can apply to such datasets.

187

INDUSTRIAL APPLICATIONS OF NIR IMAGING

0 0.18 0.17

0.2

0.16

0.4

0.15 mm

0.6 0.14 0.8

0.13

1

0.12

1.2

0.11 0.1

1.4 0

0.5

1

1.5

mm Figure 8.10 A 1.4 mm × 1.8 mm NIR image of a Tempo® top ply taken through a 10× microscope objective. This image is about three times smaller than the 5 mm × 5 mm images collected on the Perkin-Elmer Spotlight® instrument.

The examples given demonstrated several applications where we have used NIR spectroscopic imaging. It was shown how water migration can be monitored using a filter-type imaging NIR system. Spray nozzle patterns were quickly measured looking at spectral images of deposited spray on a paper substrate. Absorption peak area (integration), first-derivative spectra, and three-dimensional PCA score plots were used to highlight different components deposited on paper substrates, including lotion distribution and local amount. PCA, combined with an extended spectral region, was used to evaluate the distribution of starch, oil and water in flavored chips. There are many potential applications in the commercial arena where NIR imaging can provide useful and unique information. The authors hope this overview of applications may help the reader determine whether or not NIR spectral imaging could be a viable approach to their particular problem.

Acknowledgements The authors wish to acknowledge Brian Johnston, Raffaele Scoccianti, Giovanni Cataldo, Luigi DiGirolamo and Domenico Quirino (P&G, Pescara, Italy); Joerg Kleinwaechter (P&G, Schwalbach, Germany); David Henry (P&G, Cincinnati, OH), Tom Cambron (P&G Pharmaceuticals, Norwich, NY); Robert Hoult (Perkin Elmer); Rohit Bhargava (NIH) and Neil Lewis, Eunah Lee, Ken Haber and Joe Schoppelrei (Spectral Dimensions, Inc.).

188

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

References [1] Williams, P. and Norris, K. (1987) Near-Infrared Technology in the Agricultural and Food Industries, American Association of Cereal Chemists, St. Paul, MN. [2] Osborne, B. G., Fearn, T. and Hindle, P. H. (1993) Practical NIR Spectroscopy with Applications in Food and Beverage Analysis, 2nd edn, Longman Scientific & Technical, Harlow, UK. [3] Burns, D. A. and Ciurczak, E. W. (eds) (1992) Handbook of Near-Infrared Analysis, Marcel Dekker, New York. [4] Hassell, D. C. and Bowman, E. M. (1998) Appl. Spectrosc. 52, 18A. [5] Workman, J. J. Jr. (1999) Appl. Spectrosc. Rev. 34, 1. [6] Ozaki, Y. and Berry, R. J. (2001) In Handbook of Vibrational Spectroscopy, Vol. 2 (J. M. Chalmers and P. R. Griffiths, eds), John Wiley & Sons, Chichester, UK, pp. 953–9. [7] Todd, T. R. (2002) In Handbook of Vibrational Spectroscopy Vol. 2 (J. M. Chalmers and P. R. Griffiths, eds), John Wiley & Sons, Chichester, UK, pp. 1574–86. [8] Kidder, L. H., Haka, A. S. and Lewis, E. N. (2001) In Handbook of Vibrational Spectroscopy Vol. 2 (J. M. Chalmers and P. R. Griffiths, eds) John Wiley & Sons, Chichester, UK. p. 1386.

9

IR spectroscopic imaging From cells to tissue Max Diem, Melissa J. Romeo, Susie Boydston-White and Christian Matthäus

9.1 Introduction: definition and goals of spectral mapping The advent of infrared (IR) microscopes, coupled to interferometric optical benches, permits the collection of IR spectra of sample volumes of the order of 20 × 20 × 5 μm3 . Here, 5 μm describes the approximate thickness of the sample at the focal point of the microscope, and 20 × 20 μm2 the area, selected by knife edges or a fixed aperture.1 Second- and third-generation IR microspectrometers have further decreased the required sample volume. These sampling dimensions are of the same order of magnitude as those of a human cell. Thus, IR microspectroscopy (IR-MSP) was deemed a useful method to investigate processes in cellular biology, histopathology and cytology. However, particularly for tissue samples, point-bypoint methods of investigating properties such as state of health are hampered by the difficulty in correlating the observed spectral properties with the histopathology at exactly the same spot of tissue from which the spectrum was collected. This is particularly so since IR-MSP is carried out on unstained tissue sections, for which visual identification of different tissue types and stages of disease is virtually impossible. The original efforts were further complicated by the fact that different tissue types have significantly different spectral properties. For example, fatty tissue and connective tissue differ enormously in their chemical composition, and thus present quite different IR spectra. Thus, if a spectrum of breast connective tissue was contaminated by fatty tissue and this contamination was not recognized, the observed spectral changes were not interpreted correctly.2 Spectral mapping techniques circumvent these problems, since spectra are collected for thousands of pixels and the spectral and spatial information are analyzed together. Early IR maps were created by scanning the sample in a raster pattern through the focal point of the IR microscope in steps of the same size as the x- and y-dimensions of the pixel element.3 Although time consuming, spectral mapping, coupled to uni- or multivariate methods of data analysis, did demonstrate unambiguously that IR-MSP distinguished different tissue types and states of disease in tissue with a sensitivity comparable to those of standard histopathology.4 For the analysis of exfoliated cells (such as an automated ‘Pap’ test for cervical cancer screening), the ability to collect the spectrum of each individual cell separately by mapping methods is even more important, since statistical analysis

190

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

of all the observed spectra appears necessary for arriving at a correct diagnosis. However, as in the case of tissues discussed above, such an analysis would be very time consuming if carried out by mapping methods where the sample is moved into the focal point of a single detector. About a decade ago, the first results based on focal plane array (FPA) detectors were published.5 In these experiments, the sample was left at a fixed position, but the spatial information was collected via a detector array, typically consisting of 64 × 64 individual detectors. Details of FPA-based technology, often referred to as ‘spectral imaging’, are discussed elsewhere6 in this chapter (Chapter 1, this volume); however, these detectors, originally developed for military applications, had some serious shortcomings for spectroscopy applications. In 2001, Perkin Elmer, Inc. introduced an imaging spectrometer based on a high sensitivity linear array detector, coupled with rapid stage motion. This instrument allows IR spectral maps (images) to be collected independent of the sample size and rivals FPA-based instruments in data acquisition time and operating convenience at a significantly lower cost. In this contribution, we summarize our work with individual (cultured) cancer cells, ensembles of cancer cells and tissues containing cancer metastases in order to demonstrate the power of the new IR imaging techniques, coupled with multivariate methods of analysis. Particularly, we shall attempt to correlate the results obtained from studying individual cells with those collected from tissue sections.

9.2 Experimental In other sections of this volume,6 detailed discussions on the principles of data acquisition for IR spectral imaging are presented (Chapter 1, this volume). Here, we shall restrict ourselves to a description of the instrumentation used for the research described in the following review.

9.2.1 Instrumental aspects: PE Spotlight 300 All data presented below were collected using a Perkin Elmer imaging microspectrophotometer, consisting of a Spectrum One FTIR bench coupled to a Spectrum Spotlight 300 IR microscope (henceforth referred to as the PE 300). This totally integrated instrument incorporates a 16 × 1 element (400 × 25 μm2 ) HgCdTe (MCT) array detector and a single point, 100 × 100 μm2 MCT detector in the same Dewar. Both detectors operate in photoconductive mode at liquid nitrogen temperature. The D ∗ of each element in the array detector exceeds 4.5 × 1010 (cm Hz1/2 W−1 ). The detectors were designed for use with 1300 K sources typically used in IR spectroscopy, and cover the spectral range down to 720 cm−1 . (The single-point MCT detector even allows data collection down to 650 cm−1 .) The symmetrically arranged objective and condenser provide an image magnification of 6 × (at 1 : 1 imaging, see below), and have a numerical aperture of 0.58. Specifically designed optics permit 1 : 1 or 4 : 1 imaging of the sampled area on

IR SPECTROSCOPIC IMAGING

191

the detector elements, resulting in 25 × 25 or ∼6.25 × 6.25 μm2 nominal spatial resolution (the actual spatial resolution is wavelength dependent, and is determined by the diffraction limit). Spectra are collected in rapid-scan mode at a maximum rate of 80 pixels s−1 . For the single-cell spectra reported below, between 64 and 128 interferograms were coadded for each pixel at ∼6 × 6 μm2 spatial resolution. For tissue studies, 1–4 interferograms were coadded at 25 × 25 μm2 spatial resolution, which results in a data acquisition rate of 80 or 20 pixels s−1 , respectively. Thus, very large datasets of about 200×200 pixels (corresponding to tissue areas of 5×5 mm2 ) can be collected in about 30–120 min. Visual image collection via a CCD camera is completely integrated with the microscope stage motion and IR spectra data acquisition. The visible images are collected under white light LED illumination, and are ‘quilted’ together to give pictures of arbitrary size and aspect ratio. The desired regions for the IR maps are selected on the visual images and are restricted in size only by available memory.

9.2.2 Samples All tissue and cell ‘smear’ data were collected in reflection mode from samples on low-e glass slides (Kevley Technologies, Chesterfield, OH) for analysis. These slides are made of glass coated with a thin Ag/SnO2 layer. They are chemically inert and nearly transparent to visible light. However, they reflect mid-IR radiation almost completely, and thus are ideal and inexpensive substrates for reflection IR-MSP, as they allow both visual and IR images to be collected from the same sample. Some of the single-cell data were collected in transmission mode with the cells grown on CaF2 windows.

9.2.3 Spectral maps of individual cells Cervical cancer (HeLa) cells (cell line CCL-2, ATCC, Manassas, VA) were grown in 75 cm2 cell culture bottles or 100 mm2 square culture dishes (Sigma Aldrich, Milwaukee, WI) at 37◦ C in a 5% CO2 atmosphere in Dulbecco’s Minimum Essential Eagle Medium supplemented with 10% (by volume) fetal calf serum (both from ATCC). Samples for the study of individual cells were prepared by depositing carefully cleaned low-e slides, or IR transmitting windows, directly into the cell culture dishes. Cells were allowed to attach to and grow on these slides as sparse monolayers. The cell density is determined mostly by the incubation time of the slide in the cell culture. Generally, a few days of incubation was required to produce adequate specimens. Cells grown onto the slides assumed a relatively large size (up to 80 μm) with pseudopods resembling those commonly observed in fibroblasts. This cell shape is adequate for imaging the nucleus, but the cytoplasm is often too thin and diffuse to give good spectra. Individual cells were imaged at ∼6 μm spatial resolution and 4 cm−1 spectral resolution. To achieve acceptable signal-to-noise ratios, 64 or 128 interferograms

192

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

were coadded. Cell nuclei were found to exhibit excellent spectra with as much as 0.2 OD units in the amide I band under these collection conditions, due to the larger sample thickness in the nucleus and the high concentration of proteins and nucleic acids. However, the cytoplasm, at the edges of the cell, produces spectra of much lower amplitude (0.02 OD units or less) and, consequently, much higher noise levels.

9.2.4 Spectral maps of ‘smears’ Cytology, that is, the visual inspection of exfoliated cells, is most commonly carried out for smears of cells, which are deposited (‘smeared’) directly from brushes, spatulas or other exfoliation devices onto microscope slides. Such smears are unsuitable for spectral analysis, since they contain clumps of cells, cellular debris, erythrocytes and other contamination. However, better methods of cell slide preparations have been introduced into cytology, among them the ThinPrepTM methods developed by Cytyc, Inc. (see ref. 7), and spin centrifugation deposition techniques. These methods are very good for real exfoliated cell samples,7 since they permit the purification of the cell exfoliate, enrichment in the cells desired for analysis and produce good monolayers for visual cell inspection. Our initial experiments on cell ensembles were also carried out for cells grown onto the slides, as described above, but to somewhat higher cell density. However, in order to reproduce the conditions found in real cytology, we imaged large areas of the slides (in the mm2 range) at high spatial resolution (6 μm), even though the slides were sparsely populated (∼10% of the slide surface contained cells). Although this is a fairly time-consuming task, even with fast imaging spectrometers such as the PE 300, we could mimic the conditions encountered for real cytological samples, and analyze datasets containing hundreds of individual cells from one slide.

9.2.5 Spectral maps of tissues Samples were cut from paraffin-embedded tissue blocks, and de-paraffinized using standard procedures.8 Adjacent tissue sections (5 μm thick) were mounted on low-e slides. One tissue section was stained immediately using H&E stain, and used as a reference for data acquisition. An adjacent section was used for IR mapping, and stained subsequently for detailed histopathological analysis. It is imperative that the IR spectral maps are compared with visual images obtained from the same tissue section to avoid artifacts due to tissue changes between adjacent sections. Tissue sections were imaged in reflectance mode at 25 μm spatial resolution, and 4 cm−1 spectral resolution. These conditions produce spectra with very good signalto-noise ratios (>1000 : 1), but the use of microspectroscopy in reflection mode may produce dispersion shaped artifacts at the edge of tissue section, or for very discontinuous tissues, such as glandular structures. These will be discussed in more detail below.

IR SPECTROSCOPIC IMAGING

193

9.2.6 Mathematical analysis Preliminary data analysis carried out for the spectral datasets were functional group mapping, and/or hierarchical cluster analysis (HCA). This latter method, which is well described in the literature,4,9 is an unsupervised approach that does not require any reference datasets. Like most of the multivariate methods, HCA is based on the correlation matrix CLM for all spectra in the dataset. This matrix, defined by Equation (9.1), N L M L M i=1 (Si − S )(Si − S ) CLM =   N N L M L 2 M 2 i=1 (Si − S ) i=1 (Si − S )

(9.1)

expresses the similarity, or ‘distance’, between each spectrum and all other spectra of the data. Each element of the covariance matrix (also known as the correlation coefficient between the spectra) is obtained by forming the inner product of two normalized spectral vectors L and M, which are represented by one-dimensional columns of N absorbance (or derivative) values. The symbols S L and S M represent the mean values for each spectral vector. Since all spectra are normalized, identical spectra exhibit a correlation coefficient of unity. The resulting covariance matrix CLM contains P 2 entries, where P is the total number of spectra within the dataset. However, since the matrix is symmetric, only p(p − 1)/2 spectral distance elements CLM need to be computed. Nevertheless, for large datasets, the size of the covariance matrix may exceed the address space of Pentium IV-class processors, and the computational problems described next are rather unyielding. Subsequently, the two most similar spectra in the dataset, that is, the spectra whose correlation factors are closest to unity, are merged into a new object. A new covariance matrix is calculated for the new object and all existing spectra. The process of merging spectra or clusters into new clusters is repeated, and the CLM matrix is recalculated, until all spectra have been combined into a few clusters. This process combines the most similar spectra into the same cluster, while keeping track of which spectra have been incorporated into each cluster (cluster membership). The merging process is depicted in a dendrogram, which indicates at what level of similarity spectra or clusters are merged. The dendrogram suggests the number of spectral classes at which the clustering process should be terminated. We have used Ward’s algorithm10 to carry out the merging process of spectra and clusters, although other methods exist to perform this task (straight average, weighted average, etc.). Next, pseudocolor maps based on cluster membership are constructed by assigning a color to each spectral cluster, and displaying this color at the coordinates at which each spectrum was collected. Thus, all pixels in a pseudocolor map that have the same color are from closely related spectra. The number of clusters is adjusted such that good correspondence with the pathological images is obtained.

194

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Reasonable noise in the spectral data does not affect the clustering process. In this respect, cluster analysis is much more stable than other methods of multivariate analysis, such as principal component analysis (PCA), in which an increasing amount of noise is accumulated in the less relevant clusters. The mean cluster spectra can be extracted and used for the interpretation of the chemical or biochemical differences between clusters. HCA, per se, is ill-suited for a diagnostic algorithm. We have used the spectra from clusters to train artificial neural networks (ANNs), which may serve as supervised methods for final analysis. This process, which requires hundreds or thousands of spectra from each spectral class, is presently ongoing, and validated and blinded analyses, based on these efforts, will be reported. Cluster analysis was performed by importing datasets, in native Perkin-Elmer data format, into the CytoSpecTM FTIR imaging software package.11 The maps were processed on a personal computer equipped with a 2.2 GHz, 64-bit Athlon processor and 8 GByte of RAM. Data pretreatment included the removal of pixels with too high or too low absorbance values, or with poor signal-to-noise ratio, from the dataset. The remaining spectra were expanded between 1800 and 800 cm−1 , derivatized using a Savitsky–Golay algorithm,10 and vector normalized before cluster analysis.

9.3 Results and discussion Although one logical presentation of the results would be from cells, through cell ensembles, to tissue sections, we present the discussion of tissue sections first, because tissue results demonstrate the power of the methodology in a more intuitive way. We refer to these studies as ‘spectral histopathology’, since results analogous to histopathology, based strictly upon spectral measurements, are obtained.

9.3.1 Spectral histopathology of lymph nodes The tissues to be discussed here are sections of lymph nodes excised during surgery. Lymph nodes form an important part of the body’s immune system. Lymph nodes filter bacteria and viruses from the lymphatic fluid and act as important barriers to the spread of cancerous cells throughout the body. In this process, they may become the site of metastatic tumors, which spread to the lymph nodes via primary lymphatic drainage from adjacent tumors. The analysis of lymph nodes is an area of great importance, since many cancers (e.g. malignant breast lesions) are often detected only after metastases have formed. Thus, axillary lymph nodes are routinely excised during lumpectomies and mastectomies, and analyzed for the presence of cancer cells. Present methods of detecting cells that have metastasized from the primary tumor into adjacent lymph nodes are generally adequate, although there are cases where the identification of the primary tumor or distinction between, for example, hyperplasia and neoplasia may be difficult. The major shortcoming of standard histopathology, however, is the time delay (typically several hours up to days) between the excision of a lymph node and the availability of the diagnosis.

IR SPECTROSCOPIC IMAGING

195

This delay presents an enormous emotional stress to the patient and postpones the onset of treatment such as chemo- or radiation therapy. Spectroscopic analysis of lymph nodes via IR imaging may offer a powerful, objective and rapid alternative to traditional histopathological analysis and diagnosis. IR imaging has already been demonstrated by many groups to be a powerful tool both for accurately reproducing tissue architecture and morphology, and also for the identification of malignant cells in tissue. We now turn to an anatomical description of lymph nodes. The lymph node is surrounded by a thick, fibrous capsule and is subdivided into compartments by trabeculae. Inside the capsule is the subcapsular or marginal sinus, which forms the entry point of lymphatic fluid into the node, via the afferent vessel. The lymph node cortex, which lies beneath the subcapsular sinus, is the location of the primary and secondary lymphoid follicles. The primary follicles are comprised of B-lymphocytes. An immune response stimulates B-cells to replicate and differentiate, converting the primary follicle into a secondary follicle or germinal center, surrounded by a zone of small lymphocytes. The paracortex surrounds the germinal centers and primary follicles and contains mostly T-lymphocytes. The medulla is composed of medullary cords, consisting of macrophages and plasma cells, and medullary sinuses. The medullary vessels include the arteries and veins, and the afferent and efferent lymphatic vessels, respectively, deliver the lymphatic fluid into and out of the lymph node. In Fig. 9.1(a) and (b) we wish to demonstrate the ability of IR spectral histopathology to differentiate the different tissue types found in lymph nodes. The section shown does not contain any abnormal cells. Figure 9.1(a) shows the hematoxylin/eosin (H&E) stained section of the lymph node, typically analyzed by a histopathologist. This image was collected using the PE 300 as well, and is composed of over 50 individual exposures which are quilted together. Figure 9.1(b) represents the pseudocolor map constructed from about 12 000 pixels. This map is generated from HCA with five clusters. The unsupervised cluster analysis and resulting pseudocolor map is able to accurately reproduce the morphological architecture of the lymph node, seen as the paracortex containing T-lymphocytes (blue), follicles containing B-lymphocytes (brown), medullary cords (orange), node capsule (dark blue) and fatty tissue (green). Each pixel of the pseudocolor map represents an individual spectrum. The mean spectra from each of the five clusters (not shown) can be interpreted easily in terms of the biochemical composition: the capsule, for example, exhibits spectral features typical of connective tissue (collagen), whereas the spectra of the fatty tissue (green regions) exhibit high phospholipid components. Figure 9.1(c) and (e) show the H&E stained sections from lymph node containing colon adenocarcinoma metastases. The corresponding pseudocolor maps are presented in Fig. 9.1(d) and (f ). The color schemes of the two pseudocolor maps are not correlated, and are assigned strictly by cluster size. Thus, the cancerous regions in Fig. 9.1(d) correspond to the dark and light blue areas, whereas in Fig. 9.1(f ), the cancerous glandular region is shown as dark brown areas. As discussed before for the results of Fig. 9.1(b), these pseudocolor maps are obtained in an unsupervised

196 (a)

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(b)

(c)

(d)

(e)

(f)

Figure 9.1 (a) Section of normal area of a lymph node, stained with H&E after IR data acquisition. (b) Pseudocolor spectral map of lymph node section shown in (a), constructed from about 12 000 individual spectra (five clusters). Blue: paracortex containing T-lymphocytes; brown: follicles containing B-lymphocytes; orange: medullary cords; dark blue: node capsule; green: fatty tissue. (c) and (e), H&E stained sections from lymph node containing colon adenocarcinoma metastases. (d) and (f ) Pseudocolor maps of sections depicted in (d) and (e), respectively. The color schemes of the two pseudocolor maps are arbitrary, and are assigned according to cluster size.

fashion; that is, no spectral databases are used in the analysis, which is strictly based on spectral similarity and dissimilarity. The only input into the cluster analysis is the number of clusters at the end point. This number may vary between about four and eight clusters, and is indicated by the dendrogram, as discussed in the experimental section above. We generally opt for a clustering end point such that the pseudocolor map reproduces the visually discernible features, since, at this point, the goal of these studies is to obtain spectral classes that correspond to the histopathological description of the sample. Furthermore, the pseudocolor maps do not change much by adding or subtracting one cluster. It appears likely that more clusters will be utilized once spectral maps are correlated against more sophisticated techniques, such as immunohistochemical methods. At present, two problems persist for the clinical application of IR spectral imaging methods. One of them is the enormous size of the datasets. In order to image an entire lymph node section, about 5 × 5 mm2 in size, on the order of 200 × 200, or 40 000, individual spectra are collected. Although the data acquisition presents little problem and could be performed with existing array detector-based spectrometers within a few minutes, the data reduction by HCA takes over a day, even when using powerful personal computers such as the 64 bit AMD processor in our laboratory. The situation will undoubtedly improve tremendously when true 64 bit, WINDOWS-based operating system will be available to address the enormous

IR SPECTROSCOPIC IMAGING

197

matrix size required for these computations. Furthermore, analysis of a spectral dataset, even consisting of 40 000 spectra, can be performed in less than one minute by a diagnostic algorithm, based, for example, on ANNs. The other problem, which has been observed for many years, is the distortion of the absorptive band shapes by reflective contributions. The problem is due to the fact that in any microscopic measurements, light rays very far from normal incidence are employed. Owing to the high reflective index of many of the substrate materials, significant amount of reflected light is collected. The situation becomes worse for the reflection microspectroscopy routinely carried out in our laboratory for cost reasons (a low-e slide costs about 1% of a similarly shaped and sized CaF2 microscope slide). The reflective component has a distinct dispersive line shape that, when added to the absorptive line shape, results in strongly distorted spectra with peak shifts of up to 30 cm−1 in strongly affected peaks, such as the amide I band. This artifact occurs predominantly at the edges of tissue, and is therefore observed frequently in very loose, glandular tissue structures. We have proposed a method of removing this artifact based on a phase correction procedure in Fourier space.13 Details of this method will be reported elsewhere. We now turn to the discussion of the cluster spectra (cf. Fig. 9.2) that were extracted from lymph node tissue maps. Due to the dispersion-shape artifact discussed above that occurs in some of the spectra of glandular tissue, we have cut the spectral range for the analyses to exclude the amide I region, which is most affected by this problem. Thus, the second-derivative spectra shown in Fig. 9.2 are displayed between 1580 and 950 cm−1 . Trace A is for the connective tissue of the lymph node capsule. The bands marked by asterisks are typical for collagen, and have been described before.12 Trace B is a typical spectrum observed for mixed tissues: it contains the collagen features, and, in addition, shows peaks due to phospholipids found in fatty tissue. Traces C and D are for T- and B-lymphocytes, respectively. As expected, these traces are quite similar; nevertheless, we can distinguish these lymphocytes (or, possibly, their state of activation) based on the location within the lymph node from where the spectra were collected. Two different spectra were observed for the tumor areas. Trace E corresponds to the lighter blue area in Fig. 9.1(c), whereas trace F corresponds to the darker, and more dense, areas of the tumor. The differences in these spectra could be due to more glandular tissue type in the former region, or to different stages of the tumor. Detailed histopathological analysis will be required to answer this question after complete reduction of the dispersion artifact.

9.3.2 Spectral maps of individual cells Next, we report an example of a spectral image of an individual cell, which was grown onto a low-e slide as described above. We reported results from similar experiments before,14 which used a synchrotron light source and spectral mapping methodology. In these earlier experiments, we were able to achieve higher spatial resolution, due to the collimated nature of the IR beam from a synchrotron light

198

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

*

*

A *

* B

C

D

E

F

1500

1300 1100 Wavenumber (cm–1)

Figure 9.2 Second derivative IR Spectra of lymph node tissue types. Trace A: connective tissue of the lymph node capsule. The bands marked by asterisks are typical for collagen. Trace B: spectrum observed for mixed tissues, containing collagen features peaks due to phospholipids found in fatty tissue. Traces C and D: spectra of T- and B-lymphocytes, respectively. Traces E and F: Spectra observed for the tumor areas. Trace E corresponds to the lighter blue area in Figure 9.1(c), whereas trace F corresponds to the darker, and more dense, areas of the tumor.

source. We were, furthermore, able to treat the cell enzymatically to selectively remove RNA and DNA from the dried cell. These efforts allowed us to unambiguously assign the spectral signature of several of the cellular components. We also established that the spectra of actively growing and dividing cells were quite similar, regardless of whether the cells were normal or cancerous. In addition, we found that inactive cells, such as terminally differentiated epithelial cells exhibit quite different spectra devoid of DNA and RNA features. The absence of RNA features in the cytoplasm of cells can be readily interpreted in terms of the differences in ribosome concentration, which may vary by as much as two orders of magnitude between cells that actively synthesize proteins, and cells that do not.15 The differences in DNA signals in the nuclei of active and inactive cells is more difficult to interpret. In particular, we found that in cells with pyknotic nuclei, DNA

199

IR SPECTROSCOPIC IMAGING (a)

(c)

(b) 1800

1600

1400 1200 Wavenumber (cm–1 )

1000

Figure 9.3 (a) Visual image of a HeLa cell grown directly onto a CaF2 window. (b) IR spectral image based on amide I (1655 cm−1 ) band intensity (darker hues indicate stronger signals). (c) Raw spectra from the nucleus (top), the cytoplasm close to the nucleus (middle) and outer cytoplasm (bottom). The intensity of the amide I band for the nuclear spectrum is 0.16 OD units.

signals were absent. We interpreted this observation in terms of a hypothesis stating that completely condensed chromatin is so optically dense and so compact that its signal may be unobservable in IR-MSP.2 Synchrotron-based maps of individual cells are difficult to acquire, due to the inherent limitation of synchrotron beam time, the travel associated with visiting a synchrotron, and the fact that the measurements are basically point-by-point mapping experiments. Thus, we were eager to collect spectral maps, at a spatial resolution approaching that of the synchrotron-based experiments, using the PE 300 microspectrometer at Hunter College. Figure 9.3(a) shows a visual image of a HeLa cell grown directly onto a CaF2 window. When cells are grown onto such windows, rather than spin-deposited from suspension, they adhere strongly to the surface, and can be washed or stained subsequently. However, cells assume different shapes when they are growing onto a slide and develop spindlelike appendages, known as pseudopods, similar to those of fibroblasts. In Fig. 9.3(a), the nucleus and cytoplasm can be clearly distinguished. The IR spectral map (Fig. 9.3(b)) is based on amide I (1655 cm−1 ) band intensities; with darker hues indicating stronger signals. This figure demonstrates clearly that the protein signal is strongest for the nucleus, and nearly an order of magnitude less in the cytoplasm. The large protein signals of the nucleus may be due to two factors: in the cells grown onto a substrate, the nucleus is significantly thicker than the surrounding cytoplasm, thus offering the IR beam a longer absorption path. Second, the nucleus exhibits high protein concentrations (histones and proteins associated

200

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

with transcription and DNA replication). Thus, the large protein signals observed in the nucleus are not unexpected. Figure 9.3(c) shows unscaled spectra from the nucleus (top), the cytoplasmic area close to the nucleus (middle) and outer cytoplasm (bottom). The intensities of the protein signals can be estimated from these spectra. The intensity of the amide I band for the nuclear spectrum is 0.16 OD units. The nuclear and the cytoplasmic regions exhibit quite pronounced phosphodiester vibrations (∼1235 and 1185 cm−1 ). This observation confirms our earlier findings that actively dividing cells (such as HeLa cells in the exponential growth phase) show pronounced DNA/RNA signals for the nucleus, and RNA/phospholipid signals in the cytoplasm.13,15 The spectral quality of the traces shown in Fig. 9.3(c) is excellent, and exceeds those collected using the synchrotron source (however, the synchrotron data were collected at somewhat higher spatial resolution). These observations suggest that excellent IR spectra of nuclei can be collected using modern laboratory IR-MSP instruments.16 In fact, we are collecting, at the time of this writing, hundreds of spectra and visual images from the nucleus of cells in an effort to correlate spectral changes in the nuclei with cell biological effect. These studies have revealed that collecting spectra from entire cells is less advantageous than from the nuclear regions, since the spectral changes are averaged over larger areas in the former case. Unfortunately, the diffraction limit does not permit spectral images to be collected from cells unless the near-field advantage is exploited. However, we have mapped mytotic cells using Raman imaging microspectroscopy, and have reported spectral images of mytotic cells in the telophase and metaphase.17

9.3.3 Spectral maps of ‘cell smears’ Finally, we wish to report the results for simulated cell smears. These experiments are being carried out to determine the best parameters for data acquisition, and to establish the variations of the cellular spectra from a homogeneous cell culture. Similar efforts have been undertaken before, where the cells were spin-deposited onto the microscope slides. However, for cultured cervical cancer (HeLa) cells, the resulting samples contained many cells that maintained their morphology in suspension quite well, and produced dried cells that were nearly spherical. Once dried, these cells could not be stained easily, and thus, their divisional activity could not be established after IR data acquisition. Furthermore, large spherical cells often gave spectral artifacts that made interpretation impossible.16 Mature exfoliated cervical cells present quite a different appearance when spindeposited or smeared onto a slide. This appearance is more similar to that we obtain when we grow cells directly onto a substrate. Thus, we carried out these initial methodology tests for cells grown as sparse monolayers onto low-e slides. Figure 9.4(a) shows a visual image of such a cell sample. Figure 9.4(b) shows a raw intensity map, based on the amide I integrated intensity of the same region. Although these cells are from a homogeneous, exponential cell culture, there are quite large

201

IR SPECTROSCOPIC IMAGING

(a)

(c)

(b)

(d) 0.16 0.12 0.08 0.04

1700

1200 1450 Wavenumber (cm–1 )

950

Figure 9.4 (a) Visual image of HeLa cells grown as a sparse monolayer on a low-e slide. (b) Raw intensity map, based on the amide I integrated intensity of the same region. (c) Spectral map based on cluster analysis, revealing large differences in spectral intensities associated with the cells. (d) Mean spectra of clusters shown in (c) (same color scheme). Notice the frequency shifts in the amide I and II peaks that cause the different cluster memberships.

differences in spectral intensities associated with the cells. However, these differences are not just related to the overall intensity: a pseudocolor map, based on cluster analysis of the normalized spectra, shows distinct differences between the distribution of spectra. The cells marked within the black circles have a similar distribution of spectra in the pseudo-color map whereas other cells, notably the cluster northeast of the circle, have different spectral distributions, although they appear visually quite similar to each other. The mean cluster spectra shown in Fig. 9.4(d) not only differ in intensity quite strongly, as pointed out before, but they show small frequency shifts in the amide I and II peaks which are, most likely, the cause for different cluster memberships. In order to correlate the spectral differences with cellular features, we have used biochemical stains (incorporation of bromodeoxyuracil and an immuno-histochemical stain for cycline E) to establish the exact stage of the cells within the cell cycle.

202

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

These efforts will be used to interpret the differences in the spectral distribution in the observed maps.

9.4 Conclusions In this review, we demonstrate that excellent IR spectra from microscopic regions of cells and tissue can be collected. These spectra are extremely sensitive to variations in the biochemical composition of the pixels from which the spectra were acquired. Multivariate analyses of the spectra datasets of cells, cell smears and tissue sections produce pseudocolor maps in a totally unsupervised fashion that reproduce the histopathology of tissue sections and cytological features of cells and cell smears. In the case of tissue, we demonstrated that many features of the tissue that can be visualized microscopically after appropriate staining can be seen in the spectral pseudocolor maps without staining. Thus, IR spectral imaging detects features, which are closely related to the tissue histopathology. For individual cells, the spectral maps reproduce the demarcations between cytoplasm and nucleus, and reveal information on the chemical composition, and variations therein, of cells depending on their activity level. For cell ensembles, finally, we are able to determine the heterogeneity of the observed spectra for homogeneous cell cultures. The understanding of the variation of spectral features of the same cell type will be required for the application of IR-MSP in cytology.

Acknowledgements Partial support of this research from NIH grants GM 60654 and CA 090346 (both to MD), and RR-03037, which supports the infrastructure of the Chemistry Department at Hunter College, is gratefully acknowledged.

References [1] For review of the early work in this field, see, for example: Jackson, M. and Mantsch, H. H. (2002) Pathology by infrared and Raman spectroscopy. In Handbook of Vibrational Spectroscopy, Vol. 5 (J. M. Chalmers and P. R. Griffiths, eds), John Wiley & Sons Ltd, Chichester, UK, pp. 3227–45. [2] Diem, M., Boydston-White, S. and Chiriboga, L. (1999) Infrared spectroscopy of cells and tissues. Shining light onto an unsettled subject. Appl. Spectrosc. 53(4), 148–61A. [3] Lasch, P. and Naumann, D. (1998) FT-IR microspectroscopic imaging of human carcinoma in thin sections based on pattern recognition techniques. Cell. Mol. Biol. 44(1), 189–202. [4] Wood, B. R., Chiriboga, L., Yee, H., Quinn, M. A., McNaughton, D. and Diem, M. (2004) FTIR mapping of the cervical transformation zone, squamous and glandular epithelium. Gynecol. Oncol. 93(4), 59–68. [5] Lewis, E. N., Treado, P. J., Reeder, R. C., et al. (1995) Fourier transform step-scan imaging interferometry: high definition chemical imaging in the infrared spectral region. Anal. Chem. 67, 337. [6] Bhargava, R. and Levi, I. W. (2005) Fourier transform mid-infrared spectroscopic imaging.

IR SPECTROSCOPIC IMAGING

203

[7] van Driel-Kulker, A. M. J., Ploem-Zaaijer, J. J., van der Zwan, M. and Tanke, H. J. (1980) A preparation technique for exfoliated and aspired cells allowing different staining procedures. Anal. Quant. Cytol. 2(4), 243–6. [8] Chiriboga, L., Yee, H. and Diem, M. (2000) Infrared spectroscopy of human cells and tissue. VI. A comparative study of histopathology and infrared microspectroscopy of liver tissue. Appl. Spectrosc. 54(1), 1–8. [9] Helm, D., Labischinski, H., Schallen, G. and Naumann, D. (1991) Classification and identification of bacteria by FT-IR spectroscopy. J. Gen. Microbiol. 137, 69–79. [10] Ward, J. H. (1963) Hierarchical grouping to optimize an objective function. J. Amer. Stat. Assoc. 58, 236–44. [11] See http://www.Cytospec.com. [12] Camacho, N. P., West, P., Torzilli, P. A. and Mendelsohn, R. (2001) FT-IR microscopic imaging of collagen and proteoglycan in bovine cartilage. Biopolymers (Biospectrosc.) 62, 1–8. [13] Griffith, P. R. and De Haseth, J. A. (1986) Fourier Transform Infrared Spectrometry, John Wiley & Sons, New York, chap. 3. [14] Lasch, P., Pacifico, A. and Diem, M. (2002) Spatially resolved IR micro-spectroscopy of single cells. Biopolymers: Biospectrosc. 67, 335–8. [15] Lasch, P., Boese, M., Pacifico, A. and Diem, M. (2002) FT-IR spectroscopic investigations of single cells on the subcellular level. Vib. Spectrosc. 28(1), 147–57. [16] Diem, M., Romeo, M., Matthäus, C., Miljkovic, M., Miller, L. and Lasch, P. (2004) Comparison of Fourier transform infrared (FTIR) spectra of individual cells acquired using synchrotron and conventional sources. Infrared Phys. Technol. 45, 118–331. [17] Diem, M., Romeo, M., Boydston-White, S., Miljkovic, M. and Matthäus, C. (2004) A decade of vibrational micro-spectroscopy of human cells and tissue. The Analyst, 129(10), 880–5.

10 FPA imaging and spectroscopy for monitoring chemical changes in tissue Bayden R. Wood and Don McNaughton

10.1 Introduction Fourier transform infrared (FTIR) spectroscopy is becoming an increasingly powerful tool in the study of the composition of cells and tissues. The technique relies upon the principles of interferometry and Fourier transformation for its speed and sensitivity. When the resolution is sufficient, the infrared (IR) spectrum of any compound is in effect a unique molecular fingerprint. In medicine, the diagnostic and monitoring capabilities of FTIR are based on the fundamental premise that, in any pathologic process, a chemical change must accompany any morphological or symptomatic manifestation. FTIR spectroscopy is sensitive to the macromolecular differences between diseased and nondiseased cells. The observed spectroscopic changes relate to changes in the concentration and conformational orientation of functional groups associated primarily with proteins, lipids, nucleic acids and carbohydrates. Early studies undertaken on diseased versus nondiseased cells indicated that the FTIR technique showed potential as a diagnostic tool in the clinical environment. However, it was difficult to compare the spectroscopic results directly with routine clinical tests such as pathology results, because in many cases it was virtually impossible to know the type and number of cells that were actually being targeted with the FTIR technique. The coupling of multichannel detectors with FTIR spectrometers in the mid-1990s by Lewis, Levin and coworkers1 propelled IR chemical imaging analysis into the twenty-first century. The new generation of FTIR spectrometers equipped with multichannel array detectors, along with developments in FTIR substrates and multivariate imaging with ultrafast computers, now make it possible to gain insight into the type, number and distribution of cells within the tissue matrix. In this chapter the application of multichannel IR detector arrays to mammalian tissue is discussed from a practical perspective. The major emphasis will be on the application of the technology to cervical cancer diagnosis focusing on sample preparation, instrumental parameters, data processing techniques and correlation with histology. The next part of the chapter focuses on the application of the technique to arthritis research by investigating the effect of specific antibodies that induce an arthritic type of response in bovine cartilage explants. In the final part of the chapter

MONITORING CHEMICAL CHANGES IN TISSUE

205

the concept of FTIR three-dimensional (3D) imaging in combination with unsupervised hierarchical clustering is introduced through the investigation of FTIR images recorded from multiple adjacent sections from the gut sample of a monkey.

10.2 Applications of FTIR tissue imaging to cervical cancer 10.2.1 History of FTIR spectroscopy applied to cervical cancer diagnosis The application of FTIR imaging in combination with unsupervised hierarchical clustering shows potential as a diagnostic tool for cervical neoplasia. In the early 1990s a series of studies by Wong et al.2,3 on exfoliated cervical cells reported spectral differences between samples from patients diagnosed by cytology as being normal and those from patients diagnosed as having dysplasia or cancer. The authors correlated a decrease in the intensity of glycogen bands and an increase in the − νs (PO− 2 ) and νas (PO2 ) bands associated with nucleic acid moieties with dysplasia and cancer. Factors such as hypomethylation, the influences of hydrogen bonding on proteins, nucleic acids and carbohydrates and the degree of disorder of the methylene chains in lipid membranes were used to explain the differences in the spectra recorded with a pressure-tuning FTIR diamond anvil cell.4 Studies undertaken by Diem and coworkers5–15 and independently by McNaughton and coworkers16–24 indicated that the spectral changes observed between normal and diseased samples may not be related to the number and molecular composition of dysplastic cells per se but other factors such as localized inflammation effects,8,18 the number of dividing versus nondividing cells5 and the overall divisional activity of the cells.13 Other biological products, such as mucin, erythrocytes, leukocytes and other debris, can also obscure diagnostic regions of the spectra.8,18 Because of the intrinsic variation encountered when analyzing exfoliated cells with FTIR-based techniques, multivariate statistical approaches and artificial neural networks (ANNs) have been applied in the analysis.17,21,23,25,26 Some of these techniques provide information on the important variables that may distinguish normal from diseased samples, but they are limited in that they do not provide information on the cervical cell types and their stages of differentiation and maturation within the cervix. This is emphasized by the fact that dysplastic lesions are often surrounded by leukocytes and, in particular, activated lymphocytes. Products such as ThinPrepTM improve the diagnostic capability by removing mucins and erythrocytes.23 White cell lysis buffer can be added to remove leukocytes but care must be taken not to damage cells that may be diagnostically useful. A study by Romeo et al.22 demonstrated that samples recorded −O) at different stages of the menstrual cycle revealed dramatic changes in the ν(C− glycogen region (1200–1000 cm−1 ) due to variations in glycogen concentration over the cycle. A series of studies by Diem and coworkers9,11,12,27 revealed that the size of the nucleus resulting from changes in cell physiology and morphology is − a critical factor in determining the intensity of the νs (PO− 2 ) and νas (PO2 ) bands of DNA. Moreover, the nuclei size may provide an explanation regarding why the

206

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

divisional activity of the cells is important in the FTIR detection of cancer in cells and tissues.5,9,10 We compared spectra of dehydrated chicken erythrocytes, which contain nuclei, with nonnucleated adult human cells. In both cases 10 μl of cells was deposited onto a KRS-5 substrate and rapidly desiccated. The spectra are essentially the same, especially in the phosphodiester region, and had similar intensity νsym (PO− 2) − and νasym (PO2 ) modes associated with nucleic acid vibrations. This was indeed unexpected given that the chicken erythrocytes contain a nucleus packed full of DNA and the human cells do not. We also compared spectra of isolated nuclei extracted from the chicken erythrocytes with intact chicken erythrocytes. The extracted nuclei swelled in size to about 20 μm during the process and were still visible at this size − after desiccation. The spectra show the intensity of the νsym (PO− 2 ) and νasym (PO2 ) modes to be much greater in the extracted nuclei compared with intact cells. It is thought that the DNA within the nuclei of an intact cell is compacted to a degree, which precludes interaction with the majority of the incident radiation, and hence contributions from DNA do not appear in the spectra. In summary, these studies demonstrated that a detailed understanding of the spectral features of the cell types and spectral variations resulting from differentiation, maturation and cell cycle stages is a prerequisite before interpreting the spectral differences between normal and dysplastic cytological diagnosed samples.5

10.2.2 FTIR point-to-point mapping of cervical tissue In order to gain more insight into FTIR cervical pathology, a detailed investigation into the spectral cytology of the various cell types that comprise the cervical transformation zone, which surrounds squamous epithelium and glandular endothelium, was deemed necessary so that the origin of the major spectral types observed in cervical exfoliates and the spectral characteristics of abnormality could be understood. In collaboration with the Diem group we applied FTIR point-to-point IR microscopic mapping to the analysis of cervical tissue sections deposited on Ag–SnO2 coated slides and incorporated unsupervised hierarchical cluster analysis (UHCA) in CytospecTM 28 to construct IR images. The resulting cluster maps (or images) highlighted the different tissue regions, lymphocyte exudes and regions of abnormality within the tissue matrix.16 By analyzing different spectral windows with the UHCA approach and comparing the results with histology, we found that the changes most important in identifying anatomical and histopathological features are differences in the amide I and II region (1740–1470 cm−1 ). Such differences reflect changes in the protein secondary structure, which are characteristic of the individual cell types. The results infer that protein secondary structure is consistent for cells of the same type but varies significantly between cell types. Other spectral regions, including the glycogen–phosphodiester region (1200–1000 cm−1 ), also exhibit major changes for the different cell types, but the intercell variability for cells of the same type was found to be greater in the cluster maps using this spectral range. While these maps show excellent correlation with the anatomical and

MONITORING CHEMICAL CHANGES IN TISSUE

207

histopathological features in cervical tissue, they are not applicable to the clinical environment mainly because of the time required to collect the maps, which is around 8–10 h. In addition, the spatial resolution using conventional instrumentation is limited to 20–30 μm, precluding the spectral differentiation of cells and objects below this limit. The spatial resolution can be increased to the diffraction limit using a synchrotron source but the same or an even greater time is required to produce a map with sufficient S/N spectra to generate meaningful images.

10.2.3 FTIR focal plane array imaging of cervical tissue The new generation of multichannel array detectors has changed both the timescale and spatial resolution with which data is collected; instruments capable of collecting large, essentially diffraction limited, arrays of spectra in a short period of time are now available. Fourier transform infrared imaging spectrometers in the midIR can be equipped with focal plane arrays (FPAs) of mercury–cadmium–telluride (MCT) or InSb detectors with varying sizes. The DigiLab FPA ‘Stingray’ pixel has a theoretical spatial resolution of 5.5 μm, when using a 36× objective. Instruments such as the DigiLab ‘Stingray’ FTIR imaging spectrometer operate in rapid-scan mode and can be used to generate a mosaic of FTIR images that are subsequently stitched together to produce a multitiled image that enables the encapsulation of larger objects such as tissue sections. The unit can also be fitted with a macrosampling compartment, enabling a spatial resolution of 40 μm2 per pixel to be achieved over larger samples. Linear array detectors, such as those used in the Perkin Elmer ‘Spotlight’ instrument usually consist of 16 or 32 MCT detectors arranged in linear fashion. The individual detectors can achieve a theoretical spatial resolution of 6.25 or 25 μm depending on the magnification and numerical aperture of the Cassegrain lenses. In our investigations we have primarily utilized a DigiLab ‘Stingray’ FTIR imaging spectrometer to analyze mammalian tissue. We have found the linear arrays to be extremely useful in targeting small areas of tissue, where sufficient information is already available to make decisions on target areas. The following sections detail a practical approach to experimental procedures, instrumental parameters, data preprocessing steps and multivariate analysis for FTIR FPA imaging of mammalian tissue, when primarily using cervical tissue as a model example.

10.2.3.1 Preparation of cervical samples Cervical samples were obtained by cone biopsy from Royal Women’s Hospital (Melbourne) from patients diagnosed with high-grade cervical dysplasia by cytology. The adjacent sections were separated into two groups. One group was mounted on a glass slide and stained with hematoxylin and eosin (H&E) for light microscope examination. The other group was de-paraffinized as delineated below and mounted on Ag–SnO2 coated absorption–reflection slides (Keveley TechnologiesTM ) for FTIR FPA imaging with a Digilab Stingray FTIR microscope system equipped with a (MCT) 64 × 64 FPA (4000–950 cm−1 ). The stained slides were examined and

208

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

reported by a histopathologist. The Ag–SnO2 coated absorption–reflection slides are used in preference to plain reflective surfaces and standard transmission plates for a number of reasons. First, they are cost-effective; second, they allow the transmission of visible light and hence allow direct correlation with normal microscope slides; and third, slides can be stained for direct pathology/cytology after IR images have been taken. They also provide a substrate with properties similar to that used in a cytology/pathology laboratory and having the correct size to match the equipment in such a laboratory.

10.2.3.2 Fixation and preparation of cervical tissue for FTIR imaging Critical to obtaining high-quality spectral maps of tissue sections is a rigorous fixation and sectioning protocol designed to maintain cell morphology and minimize chemicals that may obscure diagnostically important bands. Our work has primarily utilized paraffin-embedded tissue because of its availability and relevance in the pathology lab. The importance of a rapid fixation approach for the preservation of cell morphology and for the removal of water, which would otherwise obscure the conformationally sensitive amide I mode, cannot be overemphasized. The various steps in tissue preparation for FTIR are divided into four stages, namely, fixation, paraffin processing, sectioning and de-waxing. 10.2.3.2.1 Stage 1 – fixation. Immediately after the tissue is removed from the cervix it is immersed into a solution comprising one part phosphate buffered saline (PBS) and three parts 16% paraformaldehyde and fixed for 72 h. This is mainly because the collection is done on the day of the operation, which occurs on Friday, and further tissue processing cannot commence until Monday. The PBS provides an isotonic environment and acts as a buffer for the paraformaldehyde, which when dissolved is acidic. Formaldehyde, as 4% buffered formaldehyde (10% buffered formalin), is the most widely employed universal fixative for paraffin embedded sections. Aqueous formaldehyde exists principally in the form of its monohydrate – methylene glycol, CH2 (OH)2 – and as low molecular weight polyoxymethylene glycols or polymeric hydrates. The reactive component is thought to be the hydrated form, namely methylene glycol, but this is disputed in Ref. 29. The main advantages of using formaldehyde are that it arrests autolysis and bacterial decomposition and also stabilizes the cellular and tissue constituents so that they can withstand the subsequent stages of tissue processing. It also crosslinks both the proteins and lipids thus preserving the basic macromolecular chemistry of these important molecules. Aldehydes are thought to form crosslinks between proteins, essentially creating a gel, which enables retention of cellular constituents. Soluble proteins are fixed to structural proteins and rendered insoluble, giving some mechanical strength to the entire structure, which enables it to withstand subsequent processing. Aldehydes form crosslinks between protein molecules, the reaction occurring with the basic amino acid lysine, although other groups including thioxyl, hydroxyl, imino, peptide, amido, carboxyl, guanidyl and aromatic rings could be involved.29

209

MONITORING CHEMICAL CHANGES IN TISSUE Table 10.1 Tissue processing protocol

Solution 10% formalin 80% ethanol 95% ethanol 100% ethanol 50/50% ethanol/xylene 100% ethanol Xylene Wax

Time in solution (h)

Repeats

2 1 1 1 1 1 1 1

0 1 2 2 0 2 2

Only lysine residues on the exterior of the protein molecule are thought to react and these usually account for 40–60% of the total lysyl residues.29 Formalin does not precipitate proteins and only slightly precipitates other components of the cell; it is also a good fixative for complex lipids but has no effect on neutral fats and neither preserves nor destroys adipose tissue. Formalin also traps carbohydrates because it preserves proteins so that they hold glycogen, which is otherwise readily leached from the cell. Another advantage of the compound is that it does not produce large bands that obscure the major bands associated with the macromolecules in tissue samples. Furthermore formalin is inexpensive and available in most laboratories. 10.2.3.2.2 Stage 2 – paraffin processing. The next stage is to prepare the tissue for wax embedding. This entails a series of alcohol immersions that are aimed at softening the tissue for wax impregnation. The process, which is detailed in Table 10.1, is carried out overnight in a Shandon Citadel 1000 rotary processor. The major disadvantage of this process is that the lipids are dissolved and consequently removed from the dehydrated cells. 10.2.3.2.3 Stage 3 – sectioning. The advent of silver-doped tin oxide coated slides developed by Keveley TechnologiesTM offers the advantages of affordability (∼US$2 per slide) and sample visualization with a conventional microscope facility. One can record the IR image of the sample and then stain the section with an appropriate stain for a direct comparison with the IR image. The slides are nonhygroscopic unlike many conventional IR substrates and work essentially in double transmission mode, that is, the IR beam passes through the tissue and is reflected off the slide back through the tissue. This effectively doubles the sample pathlength and hence the signal to noise. It is, therefore, necessary to section tissue samples on the order of 4–5 μm to avoid detector saturation. The major draw back with Ag–SnO2 slides is light dispersion from areas where there is only thin sample coverage or near tissue edges where there is a component of the slide background and sample. A broad baseline feature in the 1700–1000 cm−1 region, which results in a dramatic

210

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

A

B

C

3500

3000

2500

2000

1500

1000

Wavenumber (cm–1) Figure 10.1 Mean extracted spectra from a cluster analysis performed on cervical tissue. The spectra were extracted from the three outermost clusters surrounding a gland that was devoid of tissue. The clusters represent, approximate distances of (A) 5 μm, (B) 10 μm and (C) 20 μm penetration into the surrounding stroma. Spectra (A) and (B) show varying degrees of the dispersion artifact while spectrum (C) is devoid of the artifact.

increase in the amide II mode intensity relative to the amide I mode, characterizes this effect. The dispersion can also result in an apparent shift of the amide I mode to a lower wavenumber value. Figure 10.1 compares mean extracted spectra from a cluster analysis performed on cervical tissue. The spectra were extracted from the three outer-most clusters surrounding a gland that was devoid of tissue. The clusters represent distances of approximately 5, 10 and 20 μm penetration into the surrounding stroma. The mean extracted spectrum at 5 μm (labeled A) clearly shows the dispersion artifact that is characterized by a sharp dip in the baseline at ∼1720 cm−1 and an apparent increase in the amide II mode (1544 cm−1 ) relative to the amide I mode (1650 cm−1 ). The spectrum at 10 μm (labeled B) exhibits only a minor contribution from the dispersion artifact showing a slight dip at ∼1720 cm−1 and a minor increase in the amide II/amide I ratio, while the spectrum recorded at 20 μm (labeled C) is devoid of the artifact. These artifacts can be removed through modifications in the data collection and fast Fourier transformation. Diem has developed an algorithm, which will soon be incorporated into CytospecTM ,28 that involves removing the reflective components by performing a phase correction on the spectra.30 This is achieved by carrying out a reverse complex Fourier transform resulting in two interferograms due to the reflective and absorptive components of the complex refractive index. A subsequent, real forward Fourier transform of the individual interferograms produces the pure absorptive and reflective spectral components, R(ν) and I (ν), respectively.30 The phase-corrected IR spectrum S(ν) is

MONITORING CHEMICAL CHANGES IN TISSUE

211

obtained from the real and imaginary (absorptive and reflective) parts of the spectra according to S(ν) = R(ν) cos θ + I (ν) sin θ

(10.1)

where the phase angle θ is given by θ (ν) = arctan(I (ν)/R(ν))

(10.2)

While transmission measurements negate this artifact, the current substrates are very expensive and therefore require cleaning after each sample. Transmission measurements on substrates, such as ZnSe, KRS5, CaF2 , require the sample to be sectioned between 8 and 10 μm to obtain optimal signal to noise without saturating the detector. Another disadvantage of the transmission substrates is that they are not suited for optical microscopy because of their poor opacity and color. Hence, it is difficult to record a useful visual image of the tissue under investigation to compare with the IR image. The recent development of an FTIR microscope with infinity corrected objectives by SensIR will improve the visible imaging capability and will be particularly relevant to tissue analysis, especially when coupled with linear array detectors. It is also important to mount the sections for FTIR imaging in the same direction as the H&E stained templates to facilitate the location of anatomical and histopathological features. Particular care must be taken to avoid orientation artifacts resulting from uneven sectioning that can influence the results of the cluster analysis. Figure 10.2 depicts a schematic of a possible orientation effect in a hypothetical piece of tissue. At the right end of the tissue the nuclei and most of the other organelles are present in the tissue layer, while at the other end of the tissue the cells are halved due to uneven sectioning. This effect can influence the results of UHCA, introducing another level of heterogeneity within the sample. An orientation artifact like the one presented in Fig. 10.2 can be identified by a consistent change in total absorbance across the sample, indicating a consistent change in the sample pathlength. However, this is not the only type of orientation artifact and more work is required to identify and develop methods to compensate for these artifacts. 10.2.3.2.4 Stage 4 – dewaxing. The final process after the tissue sections have been deposited onto the FTIR substrate is the removal of paraffin from tissue that would

Figure 10.2 Schematic of a possible orientation effect in a hypothetical piece of tissue.

212

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

−H stretching (3000–2800 cm−1 ) and deformation region otherwise obscure the C− (∼1450 cm−1 ). This can be achieved by washing the tissue section three times in clean xylene. It is imperative that the xylene is fresh for each wash otherwise the paraffin permeates back into the tissue. This is not a step that is critical for the preparation of normal stained images and its importance for IR imaging should be stressed.

10.2.3.3 Instrumental parameters for FTIR imaging There are a number of factors to consider when choosing instrumental parameters, especially for large sections of tissue (≈5 mm) when time is a constraint. In general, for large tissue sections we use 6 cm−1 resolution and coadd eight scans for each pixel. At 4 cm−1 resolution, water vapor becomes a factor, because although the FTIR bench is fully purged the microscope is not and the background is usually recorded some hours before the sample image is processed. The resulting spectra can show large contributions from atmospheric water bands; the absorbance of which varies quite dramatically across the image. By using a higher resolution the time taken to record an image is much longer because the moving mirror has to travel further to generate the interferograms. On the other hand, by selecting an 8 cm−1 resolution, subtle details such as inflection points from underlying bands could be missed, although it would take significantly less time to record the images. In view of these considerations, we select the 6 cm−1 resolution. The number of scans to coadd also becomes a consideration. While it is desirable to maximize the signal to noise ratio this is usually not practical for large samples. We generally use 16 scans and improve our signal to noise by mathematical processing as will be discussed. Another parameter that is critical when recording large images is the level of pixel aggregation. We are restricted by the Windows 1 Gbit application limit and can only process a maximum of a 128 × 128 distance matrix in CytospecTM . Consequently, all of our images are reduced to these dimensions and our spatial resolution compromised. We have found that even with large numbers of aggregated pixels (e.g. 64 pixels) most anatomical and histopathological features can be identified. One disadvantage of the DigiLab FPA mosaic imaging is that the data matrix must be constructed as a square, which is not appropriate when one has elongated sections of tissues. Linear array instruments, such as the Perkin Elmer Spotlight, do not have this limitation, however, they are limited by the fact that the detector elements cannot be aggregated and hence large sections of tissue require enormous data matrices which cannot be processed with CytospecTM .28 10.2.3.4 Multivariate image reconstruction of FTIR array data The data processing can be divided into three phases. Phase 1 is the removal of poor quality spectra with an automated routine. Phase 2 is the data preprocessing of the spectra, which passed the quality test. This usually entails some type of baseline correction and normalization process. Phase 3 is multivariate image reconstruction where the spectra are classified and reproduced as color points

MONITORING CHEMICAL CHANGES IN TISSUE

213

in a two-dimensional (2D) color image. Phase 4 is comparison of the FTIR constructed images with the histological stained sections. This final step requires the input of one or more expert histologists. In the following examples, we demonstrate the effects of modifying various instrumental and data processing parameters on the resulting cluster maps. This type of procedure is vitally important in order to establish a consistent set of parameters that maximize information and minimize diagnostic time. These parameters vary depending on the type of tissue and detail of information required. 10.2.3.4.1 Phase 1 – quality testing. After reading the data into the CytospecTM program the first step is the ‘quality test’. The aim of this routine CytospecTM test is to remove spectra that have weak absorbance, poor signal to noise and/or the dispersion artifact or spectra that contain large contributions from water vapor. In general, we only use the ‘sample thickness’ criterion to minimize errors associated with thin samples with low absorbance and poor S/N , or alternatively to remove spectra with too high absorbance that result in a nonlinear detector response. The integrated absorbance over a specified spectral range is used to assess the sample thickness. A spectrum fails the sample thickness test if the integration value is higher or lower than the defined thresholds. To remove spectra of poor absorbance the ‘sample thickness’ criterion for cervical tissue sectioned at 4 μm is generally set to a minimum of 50 and maximum of 1000 in the 1800–950 cm−1 region. To remove spectra affected by the dispersion artifact requires a trial and error approach together with some interpretation of the spectra, since the magnitude of the artifact appears to vary depending on the amount of background versus sample encapsulated by the aggregated pixels. In general the ‘additional parameter’ is set at 0.6 on the amide II mode at ∼1540 cm−1 . It may not be imperative to remove the dispersion artifact if the information of interest is not around the edges of tissue or on the periphery of holes within the tissue. The dispersion artifact will be classified into its own individual cluster/s so that it can be easily identified in the resultant cluster maps. Once the spectra have passed the quality test they are ready for the second stage of processing. 10.2.3.4.2 Phase 2 – data preprocessing. There are many ways to process spectral data prior to multivariate image reconstruction and there is no ideal method that can be generally applied to all types of tissue. It is usual practice to correct the baseline to account for nonspecific matrix absorptions and scattering induced by the physical or bulk properties of the dehydrated tissue. One possible procedure is to fit a polynomial function to a preselected set of minima points and zero the baseline to these minima points. However, this type of fit can introduce artifacts because baseline variation can be so extreme that one set of baseline points may not account for all types of baseline variation. A more acceptable way to ‘correct’ spectral baselines is to use the derivatives of the spectra. This can only be achieved if the S/N of the individual spectra is high and if an appropriate smoothing factor is introduced to reduce noise in the derivatized spectra. Derivatives serve two purposes: they minimize broad

214

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

gentle slopping baseline features and maximize inflection points on the spectral profiles for ease of visualization. Spectra that passed a quality test were converted to second-derivative spectra using a Savitsky–Golay algorithm. The process of converting to second-derivative spectra essentially eliminates the need for baseline correction of normal spectra. One very important parameter that has a major influence on the resulting cluster maps is the number of smoothing points chosen for the smoothing function that is convolved to the derivative function. Figure 10.3 shows a series of UHCA maps for the same section using identical multivariate and preprocessing parameters except for the number of smoothing points. When only 5 smoothing points are used in the calculation as in Fig. 10.3(b) no histological feature can be identified in the cluster maps. The clusters essentially represent the four tiles that constitute the mosaic. This indicates that the major difference between the clusters is essentially related to noise and water vapor bands and not to any chemical information from the tissue. As the number of smoothing points increases so too does the ability to recognize tissue structures and layers. The tiling effect is also greatly reduced when more smoothing points are applied. The tiling effect is exacerbated by the fact that pixels toward the periphery of the FPA detector receive fewer IR photons than those in the center. Figure 10.4(a) depicts the IR illumination over the 64 × 64 FPA detector in reflection mode from a clean KeveleyTM Ag–SnO2 coated microscope slide. The dark areas around the periphery of the FPA image are not as strongly illuminated as those in the center, indicating that only a small number of IR photons are reaching these pixels. Figure 10.4(b) shows representative spectra from the four pixels closest to the periphery of the detector for a slide with deposited tissue that exhibit extremely poor signal to noise. By smoothing out the noise and normalizing the spectra, the tiling effect observed in the mosaic images after cluster analysis is removed. Care must also be taken not to over-smooth the data and remove inflection points that could potentially be diagnostic. To avoid over-smoothing, reduce tiling and resolve anatomical and histopathological features, we use the second-order derivative convolved to a smoothing function with 13 points selected as a compromise. Combining a synchrotron source to an FPA system may overcome the FPA illumination problem. Recently, Moss and coworkers (Laurent Mathis, Y. and Moss, D. ANKA synchrotron, Karlsruhe – personal communication) have demonstrated that significant improvement in terms of signal to noise and spatial resolution can be obtained by coupling an FPA microscope with a synchrotron source. After performing the second derivative or carrying out an alternate form of baseline correction, the data is normalized to account for differences in sample thickness that cause changes in the optical pathlength and hence the absorbance values in the spectrum. We tend to use a vector normalization approach, because unlike min/max normalization it does not rely on a single band being invariant. In the past it has been common practice to normalize against the amide I mode. This, however, is not necessarily the best approach because first, this band overlaps bands from other macromolecules and second, the amide I region is diagnostically important as we demonstrated in our FTIR mapping study on cervical tissue.16 For vector

MONITORING CHEMICAL CHANGES IN TISSUE

(a)

215

(b)

(c)

(d)

(e)

(f )

(g)

(h)

Figure 10.3 (a)H&E stained section of cervical tissue together with a series of UHCA maps using identical multivariate and preprocessing parameters except for the number of smoothing points. (b) 5 points, (c) 7 points, (d) 9 points, (e) 11 points, (f ) 13 points, (g) 15 points and (h) 17 points.

normalization the average value of the absorbance is calculated over the spectral region specified. This value is then subtracted from the spectrum so that the new average value equals zero. Finally, the spectra are scaled such that the sum squared deviation over the indicated wavelengths equals 1.  i

(xi )2 = 1

(10.3)

216

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a)

(b)

4000

3500

3000

2500

2000

1500

1000

Wavenumber (cm–1 )

Figure 10.4 (a) IR illumination over the 64 × 64 FPA detector in reflection mode from a clean KeveleyTM Ag–SnO2 coated microscope slide. (b) The figure shows representative spectra of a proteinbased sample from four pixels along the same pixel column closest to the periphery of the detector and designated by the arrow.

MONITORING CHEMICAL CHANGES IN TISSUE

217

10.2.3.4.3 Phase 3 – multivariate image reconstruction. Built into CytospecTM 28 are four central types of algorithm to analyze spectroscopic data. These include UHCA, principal components analysis (PCA), K-means clustering (K-MC) and fuzzy C-means clustering (F-CMC). Each has certain advantages and disadvantages in terms of computational time and for the degree of supervision required. In developing any automated diagnostic technique the degree of supervision must be as limited as possible to avoid the introduction of systematic errors and data biasing. The computation time must also be as minimal as possible to be clinically useful and for the resultant data output to be easily interpreted. For the purpose of cervical tissue FTIR image data we have found UHCA to satisfy most of these requirements, although the time taken to generate the distance matrix and perform the UHCA is still a concern. The matrix size is also restricted to approximately 128 × 128 spectra due to the memory allocation per application of 1 GBit for WINDOWS. This no doubt will be overcome in the future with continued development in computers and the computational efficiency of CytospecTM . While it is possible to work with larger matrices for the K-MC method, this, however, requires much more time to generate similar types of images because one must initially specify the number of clusters to be found and then perform a separate cluster analysis for each specified number. In UHCA, all clusters from N = 1 to N = N − 1, where N is the number of spectra, are calculated simultaneously and the clusters maps can be displayed almost instantly. On the other hand, the fuzzy C-means method does enable the user to generate all clusters simultaneously, but the clusters in the resultant maps appeared blurred and not as definitive compared to the UHCA method. In its present form the UHCA also requires some supervision in terms of deciding how many clusters adequately describe the anatomical and histopathological features within the tissue matrix. This part of the analysis can only be done with an experienced histologist. The alternative is to construct a routine to determine the number of clusters required to account for the majority of the variance between spectra based on experience and/or statistical analysis. The first part of the analysis requires a distance matrix to be calculated. This can be achieved using a number of different algorithms including D-values using Pearson’s correlation coefficient, Euclidean distances, normalized Euclidean distances, Euclidean squared distances and City Block. All of these algorithms are part of the CytospecTM software package and appear to produce similar cluster maps although the time taken for each method can vary. We used the D-values method because this is a well-established linear regression method that is suited to relative concentration data. One disadvantage of this algorithm is that it is computationally more demanding than the others, therefore more time is required for the distance matrix calculation. Equation (10.3) is the equation for the D-values of Pearson’s correlation co-efficient based on CytospecTM notation. n  ¯ ¯ i=1 xji · xki − n · Xj · Xk (10.4) rjk =    n −2 2 2 − n · x −2 · ni=1 xki i=1 xj i − nxj k

218

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

The result of the calculation is a distance or dissimilarity matrix, which is a symmetric matrix with zero diagonal elements, such that the ijth element represents how far apart or how dissimilar the ith and j th objects are. The matrix is now in a suitable form for hierarchical clustering. In cluster analysis a measure of similarity is established for each class of related spectra and a mean characteristic spectrum can be extracted for each class. In the final step, all spectra in a cluster are assigned the same color. In the false color maps, the assigned color for each spectral cluster is displayed at the coordinates at which each data cube was collected. The mean spectrum of a cluster represents all spectra in a cluster and can be used for the interpretation of the chemical or biochemical differences between clusters. There are also a variety of algorithms to select from, including Ward’s algorithm, which we employ because it minimizes the heterogeneity of the clusters. The central equation for Ward’s algorithm is as follows.   1 di,j k = (10.5) · [(nj + ni )dji + (nk + ni )dki − ni · dji ] ni + n j + n k

10.2.3.5 Normal epithelium Figure 10.5(a)–(i) shows a H&E stained section, univariate maps and cluster maps from a normal section of tissue showing a small stromal inclusion, diagnosed by histology, in the basal layer. The univariate maps based on intensity of the 1651 cm−1 band (Fig. 10.5(b)) and the integrated area underneath the 1300–1200 cm−1 spectral range (Fig. 10.5(c)) are also presented. The absorbance at 1651 cm−1 gives an indication of relative protein concentration, while the 1300–1200 cm−1 area gives an indication of relative collagen concentration in the stromal layer and the concentration of nucleic acids and other phosphate moieties in the superficial and intermediate layers. These univariate maps give some information on the distribution of the functional groups in the tissue; however, there is little anatomical detail one can correlate with the H&E stained section. This is unlike the UHCA maps, where an excellent correlation exists between tissue layers observed in the cluster map and those observed in the H&E stained section. Essentially five clusters account for the majority of cell types observed using the spectral range (1800–950 cm−1 ) in the stained tissue section. These include the (1) superficial, (2) intermediate, (3) parabasal, (4) basal and (5) connective tissue. The five cluster map (Fig. 10.5(f )) shows two small blue clusters in the parabasal–intermediate cluster (orange). These two clusters are the same color as the cluster that accounts for the intermediate–superficial cluster. Analysis of the raw spectra revealed moderate concentrations of glycogen appearing in both features, similar to the concentration observed in the superficial–intermediate mean extracted spectrum from the UHCA. This would infer that cells in these positions are similarly differentiated in terms of their biochemistry. The fact that these glycogen inclusions cannot be picked up in the H&E section provide an example of the power and potential of IR FPA imaging in histology.

MONITORING CHEMICAL CHANGES IN TISSUE

219

(b)

(a)

4 2

3 5

(c)

1

(d)

(e)

(f)

(g)

(h)

(i)

Figure 10.5 Stained section, univariate maps and cluster maps (1800–800 cm−1 ) from a section of tissue that is cytologically diagnosed as normal with a small stromal inclusion in the basal layer. (a) H&E section, (b) absorbance map of 1651 cm−1 band, (c) integrated intensity between 1300–1200 cm−1 , (d) two clusters, (e) three clusters, (f ) five clusters, (g) six clusters, (h) seven clusters and (i) eight clusters; ( j) mean spectra from (f ) and (k) mean spectra from a five cluster analysis of 1800–1700 cm−1 region only. Originally published in the Australian Journal of Chemistry.31

The IR diagnosis report in this case would state: (1) (2) (3)

Small inclusions identified in the intermediate layer (orange cluster). Analysis of raw spectra indicates moderate to high concentrations of glycogen compared with surrounding intermediate tissue. The moderate to high concentration of glycogen within the inclusion indicates that these cells are not very metabolically active and therefore not cancerous.

220

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(j)

1750 1650 1550 1450 1350 1250 1150 1050 950 Wavenumber (cm–1 ) (k)

1750 1650 1550 1450 1350 1250 1150 1050 950 Wavenumber (cm–1 ) Figure 10.5 continued.

The stromal inclusion observed in the H&E section and located on the basal layer appears in the sixth cluster in blue and surrounded by a green cluster that defines the basal layer. At eight clusters (Fig. 10.5(i)) the basal layer is clearly delineated (mid-blue), the small stromal inclusion is apparent, and a region between intermediate and parabasal appears as a separate cluster. The connective tissue (green) toward the upper region of Fig. 10.5(i) is now resolved into two clusters. The additional mid-blue cluster correlates well with red blood cells observed in the H&E stained section. We get a similar result when only the amide I spectral region is used in the UHCA calculations, although the small inclusion appears at a lower cluster number

MONITORING CHEMICAL CHANGES IN TISSUE

221

and the basal and parabasal layers are not quite as defined. Thus by minimizing the spectral region for analysis and using fewer clusters no real useful information has been lost. The inclusion of more clusters in the analysis resulted in further differentiation based mainly on baseline variation and artifacts introduced by the mosaic nature of the images. Investigation of clusters based on only the nucleic acid spectral region in the same fashion was not as useful with differentiation of all but the three major cell types lost even at cluster numbers >3. The mean preprocessed spectra for the five cluster analyses over the 1800–800 cm−1 are presented in Fig. 10.5( j) while the corresponding spectra for the 1700–1600 cm−1 are shown in Fig. 10.5(k). The extracted spectra for both spectral windows are very similar. The spectra show a clear reduction in glycogen, proceeding from the superficial layer to the basal layer. The connective tissue shows the characteristic collagen triplet indicative of the triple helix contributions. Major differences in terms of band intensity, position and width are observed in the amide I region for the different clusters. Given these differences it is, therefore, not surprising that we obtain similar cluster maps when only using the amide I region in the cluster maps.

10.2.3.6 Squamous metaplasia Squamous metaplasia is defined as the replacement of the endocervical epithelium with undifferentiated subcolumnar reserve cells, which differentiate into squamous epithelium. It is characterized by the appearance of stratified undifferentiated cells known as reserve cells, which are cuboidal, have round to oval nuclei that are not enlarged and have scant cytoplasm, resulting in prominent cytoplasmic vacuolization. The metaplastic squamous epithelium lacks intracytoplasmic glycogen and usually overlies endocervical glands. Figure 10.6 shows an H&E stained section (a), along with a univariate map based on the intensity of the amide I mode (b) and a UHCA map derived from the 1300–1000 cm−1 region (c). The mean extracted spectra of these UHCA maps are presented in Fig. 10.6(d). The UHCA maps correlate well with the H&E stained section and clearly differentiate the main cell layers including the reserve cells (orange), cuboidal squamous metaplastic cells (brown), superficial layer (dark green), mucus secreting columnar cells (blue) connective tissue (light blue). The mean extracted spectrum that is representative of the superficial layer (dark green) shows a higher concentration of glycogen than the other cell types, however, this is a lot lower than that observed in the normal epithelium, which is consistent with squamous metaplasia. The mean extracted spectrum for the connective tissue (light blue) clearly shows the collagen triplet between 1300 and 1200 cm−1 . The spectrum also shows that the ratio of the 1450 cm−1 band (assigned to the methyl/methylene deformation mode from amino acid side chains) to the 1400 cm−1 band is greater in the connective tissue than in the other regions of tissue. The mean spectrum of the columnar cells around the periphery of the duct shows characteristic mucin bands in the 1100–1000 cm−1 region. The spectra of the reserve cells and the basal/parabasal are quite similar although differences in the amide modes and the 1300–1000 cm−1 region can be observed.

222

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a)

(b)

(c)

(d)

1800

1600

1400

1200

1000

Wavenumber (cm–1 )

Figure 10.6 (a) H&E stained section showing characteristic features indicative of squamous metaplasia along with (b) a univariate map based on the intensity of the amide I mode and (c) a UHCA map, six clusters, derived from the 1300–1000 cm−1 region. (d) The mean extracted spectra of these UHCA maps are presented.

10.2.3.7 Villoglandular adenocarcinoma The cervical section presented in Fig. 10.7 has been diagnosed as a rare form of neoplasma known as villoglandular adenocarcinoma. Characteristic cytologic features of this condition include the presence of long villous fronds and papillae lined by columnar cells with intact cytoplasmic borders and minimal atypia.32 Threedimensional ball-like clusters of cells with smooth intact communal cytoplasmic rings are also associated with this condition.32 The section under examination is interesting in terms of the various anatomical and histopathological features, which include inflammatory exudate, red blood cells, connective tissue and glandular cells. A four tile (2 × 2) image mosaic was constructed from 64 × 64 image arrays collected as four pixel aggregates. Therefore, the area of tissue was 704 μm2 and represented a subsection of the tissue under investigation. The resultant cluster image in Fig. 10.7(b) thus has ∼11 μm spatial resolution, which matches the diffraction-limited spatial resolution at 1000 cm−1 .

223

MONITORING CHEMICAL CHANGES IN TISSUE

(a)

(b)

(c)

1800

1700

1600

1500 1400 1300 1200 Wavenumber (cm–1 )

1100

1000

(d)

1 2 3 4 Figure 10.7 (a) H&E section from a section of tissue with anatomical features consistent with villoglandular adenocarcinoma. (b) UHCA map of five clusters calculated over the 1800–950 cm−1 region on second-derivative vector normalized spectra. (c) Mean extracted spectra from the raw data representing each cluster. (d) Stack plot showing cluster maps performed on five adjacent sections from the same tissue.

The wavenumber resolution was set at 6 cm−1 and 16 scans were coadded for each spectrum. UHCA was performed on the 1800–950 cm−1 region on secondderivative vector normalized spectra. The brown cluster and resultant mean extracted brown spectrum in Fig. 10.7(c) are representative of the dispersion artifact from

224

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

the outer edges of the gland and also the periphery of the tissue. The light blue cluster represents a combination of mainly red blood cells embedded in the stromal matrix. The dark blue cluster is predominately stroma, while the orange is mainly lymphocyte exudates. The light green cluster accounts for a mixture of glandular cells and stroma. The mean extracted spectra exhibit dramatic changes in the amide I mode in terms of both width and band position. The band center varies from ∼1643 to 1659 cm−1 . The spectra also show intensity differences in the band center at 1237 cm−1 , which probably represents a combination of the νasym (PO− 2 ) from nucleic acids and contributions from the collagen amide III mode. The stack plot in presented Fig. 10.7(d) shows cluster maps performed on four adjacent sections from the same tissue. The maps and corresponding spectra (data not shown) are very similar for each section indicating that the biochemistry between the adjacent sections is also consistent. This figure also demonstrates that minor differences in sectioning will not affect the overall spectra to any significant degree. The time involved in analysis of such large full resolution images is prohibitive with our current computing system (1.6 GHz Athlon workstation with 2 Gb RAM) and will remain so in the short term. The images presented in Fig. 10.5 were collected and processed with UHCA in ∼18 min, some 20 times faster than using a conventional FTIR microscope, which is rapid enough for the combination of IR imaging and UHCA analysis to be useful as a pathological tool.

10.3 FPA imaging and spectroscopy for monitoring chemical changes associated with collagen-induced arthritis Infrared spectroscopy is particularly useful in determining changes in macromolecular composition in samples and hence FPA imaging provides an excellent route into monitoring changes that occur upon the progression of disease. Often, the major changes on disease progression are changes in protein structure and the example above on cervical tissue is a characteristic example of this. We have also used the technique to follow the changes that occur in diseases such as hepatitis and muscle degeneration through aging (manuscripts in preparation). The images derived from liver biopsies and cadaver muscle can be compared with traditional histology and pathology in the same manner as described above using cluster analysis or similar techniques. FPA imaging also finds a role in more fundamental disease research, where experimental animal models are often used to ‘mimic’ a disease. One such study is outlined below. Fluorescence imaging is widely used to follow processes such as antibody binding, usually by attaching a fluorophore to the antibody. This allows for monitoring processes such as antibody penetration, which when combined with traditional dye stains can be used to correlate the immunofluorescence with gross morphological changes. This combination, however, gives no insight into the antibody-specific chemical changes that provide a better understanding of the various processes. IR imaging of tissue samples upon antibody penetration, however, provides a unique

MONITORING CHEMICAL CHANGES IN TISSUE

225

opportunity to follow both the physical penetration and the consequential changes in macromolecular composition. An experimental model of the human autoimmune disease, rheumatoid arthritis, is collagen-induced arthritis (CIA), which is induced in animals by injection with type II collagen (CII), a major component of articular cartilage. In the following autoimmune response, antibodies to CII are formed and the resultant disease has expression and histopathological appearance similar to rheumatoid arthritis. A similar form of arthritis develops when antibodies to CII are passively transferred from mice with CIA to naïve mice. Antibodies to CII are known to be present in the sera and synovial fluid of patients with rheumatoid arthritis and some of these react with the same epitopes of CII as arthritogenic monoclonal antibodies from CIA (mAbs).33 However, the question remains whether autoantibodies to CII actively participate in the pathogenesis of rheumatoid arthritis, or whether they occur merely as a secondary effect of cartilage degradation due to an inflammatory immune response of other provenance. It has been shown previously34 that arthritogenic mAbs to CII can inhibit collagen fibrillogenesis in vitro and affect chondrocyte morphology and matrix formation in chondrocyte cultures. However, it is unclear whether mAbs would penetrate and disrupt a preexisting cartilage matrix, so in recent work we have examined the effects of mAbs to CII on bovine cartilage explants using a combination of immunofluorescence, IR imaging and toluidine blue. Toluidine blue stains proteoglycan (PG) and so intense color correlates with high PG content. Cartilage explants from adult bovine metacarpal phalangeal joints were cultured with the mAbs CII-C1, M2.139, F4, GAD6 or medium alone for periods up to 21 days. CII-C1 and M2.139 are arthritogenic mAbs that bind to conformational epitopes on CII, F4 is a nonarthritogenic control mAb that binds to a conformational epitope and GAD6 is a control mAb that binds to an irrelevant antigen glutamic acid decarboxylase. To determine whether any effects were the result of crystallizable fragment (Fc) binding of the mAbs to chondrocytes, the explants were cultured with a fragment of arthritogenic antibody with two antigen binding sites F(Ab)2 from CII-C1. After incubation, the cartilage was sectioned and adjacent sections used for IR imaging, fluorescence imaging and toluidine blue staining for PG. Tissue sections were mounted on Ag–SnO2 coated slides in an essentially identical fashion to that described above for the cervical tissue. The IR FPA images have been used in a number of ways to help understand the experimental outcomes. First, hierarchical cluster analysis has been used to determine the protein/collagen types and their distribution within the tissue; second, univariate images have been used to show the distribution of collagen and PG; third, average spectra from the cluster types or from regions of similar content have been used to determine the macrochemical changes and finally, the images have been correlated with both immunofluorescent and stained images.37 In Fig. 10.8 a cluster map of an explant control is shown together with the average cluster spectra. The hierarchical cluster map in Fig. 10.8(a) shows a distinct layering of protein types through the cartilage used as a control, which can be further understood through examining the average spectra for each cluster. From the amide I band positions in

226

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS (a)

(b)

Absorbance

0.20 Amide II (total protein)

0.15 0.10

Collagen triplet

0.05 0.00 1800 1700 1600 1500 1400 1300 1200 1100 1000 Wavenumber (cm–1 )

1700

1650

1640

1660 1656

(c)

1600

1550

Wavenumber (cm–1 )

Figure 10.8 (a) 704 × 704 mm explant section on day 14, normal (control). Cluster analysis performed on amide I band (1715–1600 cm−1 ). (b) Average spectra for three clusters. (c) Average second-derivative spectra for three clusters.

the second-derivative spectra in Fig. 10.8(c), it can be seen that there are three distinct protein types with band positions that correspond with α-helical (1656 cm−1 ), β-sheet (1640 cm−1 ) and random coil (1660 cm−1 ) proteins. A toluidine blue stained section of adjacent tissue shows consistent dark staining over the whole section with no breakdown of collagen or PG, while IR univariate images based on total protein content show consistent protein concentration across the sample.

227

MONITORING CHEMICAL CHANGES IN TISSUE (b)

(d)

(f)

0.20

Arbitrary units

Absorbance

(e)

(c)

0.15 0.10 0.05

1800

1400

1000

1666 1655 1641

(a)

1750

1650

1550

Figure 10.9 (a) 704 × 704 mm explant section on day 7 cultured with arthritogenic Mab. (b) Image constructed from collagen band (1300–1190 cm−1 ). (c) Image constructed from ratio of collagen/ protein (1300–1190 cm−1 )/(1600–1500 cm−1 ). (d) Cluster analysis performed on amide I band (1715–1600 cm−1 ). (e) Average spectra for three clusters. (f ) Average second-derivative spectra for three clusters.

The images in Fig. 10.9 highlight the changes in protein and collagen distribution that occur upon incubation with an arthritogenic mAb. The multivariate three cluster map of Fig. 10.9(d) based on the protein region correlates particularly well with the univariate map of Fig. 10.9(b), which shows the areas of high collagen/protein ratios. Higher-order cluster maps show even better correlation. The staining of the section incubated with GAD6, the irrelevant mAb (Fig. 10.10(a)), is consistent across the whole section, while for the arthritogenic mAb (Fig. 10.10(b)) the outer portions are essentially unstained. For the section incubated with F(Ab)2 the stain shows that all the PG has been destroyed. The IR univariate maps based on the intensity of bands assigned to PG in Fig. 10.3 correlate very well with the stained sections and give the same information. Examination of the average spectra extracted from the high and low PG regions show that there is indeed a reduction in PG and changes in the protein structure. The work described above shows that the FPA IR imaging can provide the information required for such a study, essentially replacing the immunofluorescence and traditional staining techniques, while also providing additional information on the chemical changes occurring. One of the problems encountered while using FPA imaging to monitor changes occurring in treated tissue specimens is the ability to make direct comparisons between two sections treated in different fashions. Univariate chemical maps allocate

228

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS 6

(a)

5 4 3 2 1 0 1750

1550

1350

1150

950

1750

1550

1350

1150

950

1750

1550

1350

1150

950

1750

1550

1350

1150

950

1750

1550 1350 Wavenumber

1150

950

6

(b)

5 4 3 2 1 0 6

(c) Absorbance

5 4 3 2 1 0 (d)

6 5 4 3 2 1 0

(e)

6 5 4 3 2 1 0

Figure 10.10 Explant sections on day 14 cultured with (a) GAD6, (b) FA and (c) CII-C1 (d) M2.139 and (e) the F(Ab)2 fragment of CII-C1. The left image is toluidine blue stained and the right is an IR image based on the PG content (integrated bands in the 1100–1000 cm−1 region). The spectra are ten averaged spectra from the PG high and low areas and the error bars have a standard deviation of 1. Reprinted with permission from Arthritis Research & Therapy, biomedical central open access publishing.37

MONITORING CHEMICAL CHANGES IN TISSUE

229

colors to all intensities between the highest and lowest intensities across the imaged section and so the colors cannot be directly compared between sections. In techniques such as cluster analysis there is no simple way to ensure that the color allocations across clusters in different samples are the same and the same problem occurs. In order to enable direct comparison between sections, we have developed our software to allow for adjacent sections to be tiled together prior to image analysis. The images constructed after this data pretreatment ensures direct comparison between sections. This stitching can also be used to compare adjacent sections of a tissue block as shown in Fig. 10.7(d).

10.4 Application of FTIR 3D imaging to histology The ability to generate and manipulate 3D images of body parts or tissue sections is extremely useful in determining the extent of disease or tissue degeneration. Once the data are collected, there are numerous ways of representing them. The most useful ways of generating such 3D images is through X-ray-based techniques, such as computerized tomography (CT), positron emission tomography (PET), magnetic resonance imaging (MRI) and 3D ultrasound. When compared with IR images they have the advantage of being nondestructive. On coupling X-ray-based techniques with synchrotron radiation an increase in contrast35 is achieved. Recently a hospital at Spring8 in Japan has benefitted from the advantages of synchrotron sources. Unlike IR imaging, the contrast in these techniques does not supply chemical information so 3D IR imaging would provide a useful and novel alternative. A newly emerging imaging technique is based on Terahertz (THz) or far-IR radiation,36 and although this may display chemical information the THz region is an extremely narrow frequency region and the spectroscopy is almost unexplored. Although 3D THz images of quite large objects have been generated, the chemical information provided from this technique in the recent past, and highly likely in the future, is extremely limited. The lack of penetration of mid-IR radiation into tissue precludes real-time imaging, however, it is possible to build images from adjacent sections of tissue. This has the disadvantage of destroying the tissue but if information is required on the extent of penetration of the disease, then it may find some use. We have developed software that allows us to construct and visualize 3D images of tissue by stacking univariate images or multivariate images such as cluster maps. Once a 3D block is constructed then the tissue image may be examined by slicing through the tissue to view sections of any width, with a choice of the number of slices and the angle of each slice. Figure 10.11 shows two ways of presenting such data for a block of a monkey gut sample. In Fig.10.11(a) and (c) the whole block is presented with Fig. 10.11(a) being a reconstruction of seven 2D images of amide I intensity and Fig. 10.11(c) the same from seven 2D hierarchical cluster maps. The finger-like projections into the tissue, which are apparent in all the representations,

230

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a)

(b)

(c)

(d)

Figure 10.11 Section of monkey gut reconstructed from FPA images of seven 6 μm sections. Images from 1650–1500 cm−1 region. (a) and (b) from univariate maps; (c) and (d) from hierarchical clustering using three clusters and showing two perpendicular sections.

are the villi of the gut. Figure 10.11(b) and (d) show one method of slicing the data to show the internal structure of the tissue block.

10.5 Conclusions Having shown that IR imaging and UHCA analysis is capable of extracting mean spectra for the major cell types in cervical tissue samples, there are a number of remaining goals for the cervical cancer work. First, we wish to develop a methodology to determine the optimum number of clusters in any analysis that

MONITORING CHEMICAL CHANGES IN TISSUE

231

accounts for the true spectral variation. In this way, any subjective choice of cluster number will be eliminated from the analysis. We will then develop a databank of spectra for all the normal and abnormal cell types present in tissue showing dysplasia (including human papilloma virus, HPV) and carcinoma in situ. This databank could provide the basis to develop and test a diagnostic tool for cervical tissue based not on cluster analysis but on a more rapid system such as ANNs. The rate determining step in the analysis will then be the recording of the hyperspectral cube and not the multivariate statistical analysis. The current study has shown that these goals are well within reach. Our other works on bovine cartilage explants, facial muscle tissue and liver biopsies have also shown that FPA imaging by itself or in combination with cluster analysis provide a useful tool at the research level where it provides information far surpassing techniques such as traditional staining or fluorescence imaging. In the research environment, the timescale involved when cluster analysis is necessary is of little consequence. The question remains as to whether FPA imaging will find a role as a tool in pathology or cytology. Although we have shown that FPA images provide useful morphological information together with chemical information, which is available from no other technique, the time involved in gathering and displaying the information on tissue samples of the size generally subjected to pathology is prohibitive at present. FPA imaging, however, is in its infancy and given the rapid advances in computer technology and the expected advances in FPA technology, it is conceivable that FPA imaging instruments could one day become a common tool in the pathology laboratory.

Acknowledgements The cervical cancer work is supported by a Commonwealth National Health and Medical Research Council grant (Appl. No. 236812). Wood is supported by an Australian Synchrotron Research Program Fellowship. We would also like to thank Michael Quinn (Royal Women’s Hospital, RWH) for input on all gynecological aspects of the project; Merrill Rowley (Department of Biochemistry, Monash University) for direction and advice in all aspects of the arthritis project; Keith Bambery (Centre for Biospectroscopy) for assistance in image acquisition and data processing for cervical cancer project; Virginia Billson (Royal Women’s Hospital, Melbourne) for histological diagnosis of cervical sections; Corey Evans (Department of Chemistry, Leicester University) for generation of 3D univariate and multivariate maps of the monkey gut sample; Duncan Abercrombie (Department of Biochemistry, Monash University) for image acquisition and data processing of cartilage explants in arthritis project; Clyde Riley (RWH) for cervical tissue sample preparation; Juleen Hallo (RWH) for organizing collection of cervical samples and Finlay Shanks for instrumental support.

References [1] Lewis, E. N., Treado, P. J., Reeder, R. C. et al. (1995) Fourier transform spectroscopic imaging using an infrared focal-plane array detector. Anal. Chem. 67, 3377–81. [2] Wong, P. T. T., Wong, R. K., Caputo, T. A., Godwin, T. A. and Rigas, B. (1991) Infrared spectroscopy of exfoliated human cervical cells: evidence of extensive structural changes during carcinogenesis. Proc. Natl. Acad. Sci. USA 88, 10988–92.

232

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

[3] Wong, P. T. T., Wong, R. K. and Fung, M. F. K. (1993) Pressure tuning FT-IR study of human cervical tissues. Appl. Spectrosc. 47, 1058–63. [4] Fung, M. F. K., Senterman, M., Eid, P., Faught, W., Mikhael, N. Z. and Wong, P. T. T. (1997) Comparison of Fourier-transform infrared spectroscopic screening of exfoliated cervical cells with standard Papanicolou screening. Gynecol. Oncol. 66, 19–15. [5] Diem, M., Boydston-White, S. and Chiriboga, L. (1999) Infrared spectroscopy of cells and tissues: shining light onto a novel subject. Appl. Spectrosc. 53, 148–61A. [6] Mohlenoff, B., Romeo, M., Diem, M. and Wood, B. R. (2005) Mie-type scattering and nonBeer-Lambert absorption behavior of human cells in infrared microspectroscopy. Biophys. J. 88, 3635–3640. [7] Diem, M., Chiriboga, L. and Yee, H. (2000) Infrared spectroscopy of human cells and tissue. VII. Strategies for analysis of infrared tissue mapping data and applications to liver tissue. Biopolymers (Biospectroscopy) 57, 282–90. [8] Chiriboga, L., Xie, P., Vigorita, V., Zarou, D., Zakim, D. and Diem, M. (1997) Infrared spectroscopy of human tissue. II. A comparative study of spectra of biopsies of cervical squamous epithelium and of exfoliated cervical cells. Biospectroscopy 4, 55–9. [9] Chiriboga, L., Xie, P., Yee, H. et al. (1998) Infrared spectroscopy of human tissue. I. Differentiation and maturation of epithelial cells in the human cervix. Biospectroscopy 4, 47–53. [10] Chiriboga, L., Xie, P., Yee, H., Zarou, D., Zakim, D. and Diem, M. (1998) Infrared spectroscopy of human cells and tissues. IV. Detection of dysplastic and neoplastic changes in human cervical tissue via infrared microscopy. Cell Mol. Biol. 44, 219–29. [11] Chiriboga, L., Yee, H. and Diem, M. (2000) Infrared spectroscopy of human cells and tissue. Part VI: a comparitive study of histopathology and infrared microspectroscopy of normal, cirrhotic, and cancerous liver tissue. Appl. Spectrosc. 54, 1–8. [12] Chiriboga, L., Yee, H. and Diem, M. (2000) Infrared spectroscopy of human cells and tissue. Part VII: FT-IR microscopy of DNAase- and RNAase-treated normal, cirrhotic, and neoplastic liver tissue. Appl. Spectrosc. 54, 480–5. [13] Boydston-White, S., Gopen, T., Houser, S., Bargonetti, J. and Diem, M. (1998) Infrared spectroscopy of human tissue. V. Infrared spectroscopic studies of myeloid leukemia (ML-1) cells at different phases of the cell cycle. Biospectroscopy 5, 219–27. [14] Boysden-White, S., Gopen, T., Houser, S., Bargonetti, J. and Diem, M. (1999) Infrared spectroscopy of human tissue. V. Infrared spectroscopic studies of myeloid leukemia (ML-1) cells at different phases of the cell cycle. Biospectroscopy 5, 219–27. [15] Chiriboga, L., Diem, M. and Wood, B. R. (2003) Letter in response to ‘infrared spectral features of exfoliated cervical cells, cervical adenocarcinoma tissue and adenocarcinoma cell line’. Gynecol. Oncol. 91, 275–6. [16] Wood, B. R., McNaughton, D., Chiriboga, L., Yee, H. and Diem, M. (2003) Fourier transform infrared mapping of the cervical transformation zone, and dysplastic squamous epithelium. Gynecol. Oncol. 93, 59–68. [17] Wood, B. R., Quinn, M. A., Burden, F. R. and McNaughton, D. (1996) An investigation into FTIR spectroscopy as a biodiagnostic tool for cervical cancer. Biospectroscopy 2, 143–53. [18] Wood, B. R., Quinn, M. A., Tait, B., Hislop, T. and Romeo, M. (1998) FTIR microspectroscopic study of cell types and potential confounding cells in screening for cervical malignancies. Biospectroscopy 4, 75–91. [19] Wood, B. R., Quinn, M. Q. and McNaughton, D. (1997) Detection of potential confounding variables and cell types in Pap smears using Fourier transform infrared spectroscopy. In Spectroscopy of Biological Molecules (P. Carmona, R. Navarro and A. Hernanz, eds), Kluwer Academic Publishers, Dordrecht, pp. 445–6. [20] Wood, B. R., Quinn, M. Q., Tait, B., Romeo, M. and Mantsch, H. H. (1998) A FTIR Spectroscopic study to identify potential confounding variables and cell types in screening for cervical malignancies. Biospectroscopy 4, 75–91. [21] Romeo, M., Burden, F. R., Wood, B. R., Quinn, M. A., Tait, B. and McNaughton, D. (1998) Infrared microspectroscopy and artificial neural networks in the diagnosis of cervical cancer. Cell Mol. Biol. 44, 179–87. [22] Romeo, M., Wood, B. R. and McNaughton, D. (2002) Observing the cyclical changes in cervical epithelium using infrared microspectroscopy. Vib. Spectrosc. 28, 167–75.

MONITORING CHEMICAL CHANGES IN TISSUE

233

[23] Romeo, M., Wood, B. R., Quinn, M. A. and McNaughton, D. (2003) The removal of blood components from cervical smears: implications for cancer diagnosis using FTIR spectroscopy. Vib. Spectrosc. 72, 69–76. [24] Romeo, M. J., Quinn, M. A., Wood, B. R. and McNaughton, D. (1999) An investigation into hormonal influences on cervical cells using FTIR spectroscopy. In Spectroscopy of Biological Molecules: New Directions (J. Greve, G. J. Puppels and C. Otto, eds), Kluwer Academic Publishers, Dordrecht, pp. 447–8. [25] Cohenford, M. A., Godwin, T. A., Cahn, F., Bhandare, P., Caputo, T. A. and Rigas, B. (1997) Infrared spectroscopy of normal and abnormal cervical smears: evaluation by principal component analysis. Gynecol. Oncol. 66, 59–65. [26] Cohenford, M. A. and Rigas, B. (1998) Cytologically normal cells from neoplastic cervical samples display extensive structural abnormalities on IR spectroscopy: implications for tumor biology. Proc. Natl. Acad. Sci. USA 95, 15327–32. [27] Lasch, P., Boese, M., Pacifico, A. and Diem, M. (2002) FT-IR spectroscopic investigations of single cells on the subcellular level. Vib. Spectrosc. 28, 147. [28] Lasch, P. CytospecTM , a Matlab based application for infrared imaging. See http://www.cytospec.com for details. [29] Pearse, A. G. E. (1980) Histochemistry, Theoretical and Applied, 4th edn, Churchill Livingstone, Edinburgh. [30] Diem, M. (2004) Correction of dispersion artifacts in IR microspectral data. SPEC2004 Shedding New Light on Disease: Optical Diagnosis for the New Millenium, Third International Conference, 19–24 June, Rutgers University, Newark, NJ. [31] Bambery, K., Wood, B. R., Quinn, M. A. and McNaughton, D. (2004) Fourier Transform Infrared Imaging and Unsupervised Hierarchical Clustering applied to cervical Biopsies. Aust J. Chem. 57, 1139–43. [32] Novotny, D. B. and Ferlisi, P. (1997) Villoglandular adenocarcinoma of the cervix: cytologic presentation. Diag. Cytopath. 17, 383–7. [33] Burkhardt, H., Koller, T., Engstrom, A., Nandakumar, K. S., Turnay, J. and Kraetsch, H. G. (2002) Arthritis Rheum. 46, 2339–48. [34] Gray, R. E., Seng, N., Mackay, I. R. and Rowley, M. J. (2004) J. Immunol. Methods 285, 55–61. [35] Lewis, R. (1997) Medical applications of synchrotron radiation X-rays. Phys. Med. Biol. 42, 1213–43. [36] Zhang, X.-C. (2002) Terahertz wave imaging: horizons and hurdles. Phys. Med. Biol. 47, 3667–77. [37] Crombie, D. E., Turer, M., Biurrun Zuasta, B., Wood, B., McNaughton, D., Nandakumar, K. S., Holmdhal, R., Van Damme, M. P. and Rowly, M. J. (2005) Destructive effects of murine arthritogenic antibodies to type II collagen on cartilage explants in vitro. Arthritis Research & Therapy 7, R927–37.

11 Infrared microscopy and imaging of hard and soft tissues Applications to bone, skin and cartilage Richard Mendelsohn, Adele L. Boskey and Nancy P. Camacho

11.1 Introduction Infrared (IR) spectroscopy has been used for over half a century for structural analysis in biophysics and biochemistry. The approach has provided important information about tissue composition, protein secondary structure and interactions, DNA conformation and structural transitions, and lipid conformational order and phase behavior. Traditional applications of the method require that samples be homogenized. This restriction places a limit on applications involving biological tissue, since the spatial distribution of molecular components in tissue provides an important determinant for their function, and sample homogenization evidently destroys this information. A second limitation of traditional sampling approaches is that relatively large sample quantities are required, so that the measured spectrum is an average over the entire sampled region. These limitations of IR spectroscopy were recognized long ago. The coupling of a microscope to an IR spectrometer was described in the late 1940s.1 The device was used for microanalytical applications. For the study of heterogeneous materials, the problem of slow data accumulation in point-by-point IR microscopy has limited the utility of this technique. For example, the examination of tissues with the same number of data points as in modern 64 × 64 array detectors would require about 68 h (at 1 min/spectrum) with single-point detectors. This fact evidently precludes the investigation of the large number of samples needed for statistically significant biomedical results. The commercial availability in the mid to late 1990s of midIR array detectors has overcome the above limitation of traditional IR microscopy, resulting in the introduction of IR imaging spectrometers and the opening up of a new field of biomedical research into normal and pathological states of tissues. This chapter illustrates the power of this new technology by considering selected applications of IR microscopy and imaging to hard and soft tissues. Some of the applications described here began as a collaboration between two of the current authors (RM, ALB) in the late 1980s, which produced the first systematic study of the technique to investigate molecular structure changes between normal and pathological states of bone. The chapter will also demonstrate the advantages of

IR IMAGING OF HARD AND SOFT TISSUES

235

this approach to pathological states of cartilage and illustrate the potential of the method to the study of permeation of exogenous substances in skin.

11.2 IR imaging protocols Two instruments were used for these experiments. The first generation of data were collected on a Bio-Rad (‘Sting Ray’) instrument equipped with an IR microscope, a step-scanning interferometer and an IR detector consisting of an array of 64 × 64 MCT elements (total size ∼4 mm × 4 mm) imaged to a sample area of ∼400 × 400 μm. Interferograms were collected at 8 cm−1 spectral resolution and spectra were generated by Fourier transformation of the 4096 interferograms. Each detector element is sampled 80 times at each stopped position of the moving mirror. Thus the final datasets consist of 4096 IR spectra spanning the approximate frequency range of ∼4000–900 cm−1 . The second generation of IR microscopic images was acquired with the Perkin-Elmer ‘Spotlight’ system, consisting essentially of a 16 × 1 linear array of detector elements along with a computer controlled XY sample stage automated to move the sample relative to the array. Sample sizes were typically 850 × 700 μm, acquired with a spatial resolution of 6.25 μm, so that ∼15 000 complete spectra at high signal-to-noise ratios (∼1000/1 for strong bands) may be acquired in about 1 h. Spectral resolution is typically 8 cm−1 . A somewhat wider spectral range is available with the ‘Spotlight’ system (4000–700 cm−1 ).

11.3 Applications of FTIR microscopy and imaging to tissues 11.3.1 Bone 11.3.1.1 The tissue Bone is a composite tissue essential to vertebrates because it provides strength and rigidity; thus enabling mobility. Bone is also the storehouse for calcium and phosphate ions required for the metabolism of the organism. Bone consists of an organic matrix (principally type I collagen with smaller amounts of noncollagenous proteins, lipids and water) and inorganic crystals. The collagen gives bone its elasticity and provides a template for the deposition of mineral. Noncollagenous proteins are involved in the metabolism of the tissue; they control cell–cell and cell–matrix interactions and regulate the deposition of mineral crystals. The lipids are predominately components of the multiple cell types found within bone: osteoblasts (bone-forming cells), osteocytes (osteoblasts surrounded by mineral and connected to each other by long extensions known as caniculae) and osteoclasts (cells that remodel bone). The mineral is a nanocrystalline analog of the geologic mineral hydroxyapatite (Ca10 (PO4 )6 (OH)2 ) with numerous substituents including magnesium, carbonate,

236

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Resting zone

Hypertrophic zone Provisional calcification Growth plate

Blood vessel

Trabecular bone

Osteon

Marrow cavity Periosteum

Cross-section cortical bone

Endosteum

Cortical bone

Figure 11.1 Schematic showing long bone with higher power views of the epiphyseal growth plate and cortical bone.

acid phosphate and lattice vacancies. It is the mineral that gives bone structural rigidity and enhances the mechanical properties of the organic matrix.2 The bones in the body can be categorized by shape (long and flat), structural arrangement (cortical and trabecular/lamellar, woven, compact) or their mechanism of development (endochondral and intramembranous ossification). Intramembranous ossification refers to direct bone formation such as is seen in the skull. Endochondral ossification describes the replacement of a cartilage model with bone. This is the way one grows in length with the growth zones (growth plate) at the ends of the bones turning in to bone. It is also the way in which fractures heal. Figure 11.1 shows a schematic drawing of a long bone in a developing animal with the different categories of bone indicated by the inserts.

11.3.1.2 Questions solved by IR microscopy Traditional IR spectroscopy was initially applied to homogenized bone3 to identify the presence of hydroxyapatite in bone and to quantify the presence of bone substituents, such as carbonate and acid phosphate. Because carbonate can substitute for both hydroxyl ions and phosphate ions, IR techniques were developed to distinguish between these two sites for substitution. These spectra also provided

IR IMAGING OF HARD AND SOFT TISSUES

237

information on the crystal size and perfection within the homogenized tissues. In fact, the first study to show differences between normal bone and bones from individuals with osteoporotic fractures was based on IR analysis.4 Similar analyses were applied to the dentin and enamel of teeth demonstrating the differences in crystal size and composition of these two components.3 The limitation in these studies was that these hard tissues had to be homogenized and thus the spatial heterogeneity characteristic of bone could not be documented. The availability of IR microspectroscopy and IR microspectroscopic imaging permits questions related to spatial variations in bone to be addressed. The spatial resolution of these techniques are of the order of 10 μm, thus spectral differences varying from the surface to the interior of different bone types can be characterized. While some of these spatial variations were observable using histochemical stains and microradiography, IR provided quantitative descriptors. One of the key questions addressed by IR microspectroscopy concerned the changes in mineral content, composition and properties during development, and the impact of diet, environment, species and specific proteins (whose genes were mutated, ablated or overexpressed). The compositional changes in bone as a function of developmental site,5 age, diet or disease had been addressed by density fractionation of pulverized bone6 but there was no information on spatial variation. Since bone is such a heterogeneous tissue, spatially resolved studies of changes during growth and development, and with age7 opened a new area of investigation and suggested new therapeutic modalities. Additionally, these studies provided paradigms for analysis of bone formation in cell culture.8,9 The second major area to be addressed once the normal variations in spectral properties of bone were established, was the spatial and temporal alterations in bone disease. The ability of IR microspectroscopy to nondestructively characterize the composition in thin sections of bone that could then be evaluated by histologic techniques allowed investigators to address questions concerning mineral and matrix changes during fracture healing,10 and in prevalent diseases, such as osteoporosis,11 or rarer diseases, such as osteopetrosis12 and osteogenesis imperfecta.13 Osteoporosis, a widespread disease in both women and men is characterized by loss of bone density with an increased risk of fracture. More significantly, once an individual sustains an osteoporotic fracture, the probability that they will sustain an additional fracture greatly increases. The first IR studies of osteoporotic bone were performed in the early 1970s4 based on analyses of homogenized biopsies. A series of IR and X-ray diffraction studies reviewed in Ref. 10 indicated that while the mineral content was consistently decreased in an osteoporotic patient when compared with normal age-matched healthy bone, crystal size and perfection was either increased, decreased or unchanged. To address this variation, studies that considered the ultrastructure of the bone, and excluded callus at the fracture site or microcracks in the tissue, could only be performed by IR microspectroscopy and imaging. These studies as described below demonstrated decreased mineral content, increased crystallinity and increased matrix maturity in the osteoporotic bone.11 As disease changes are becoming better characterized, with knowledge of

238

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

what age-matched normal healthy bone spectra are like, the IR analysis is facilitating unbiased quantitative evaluation of the efficacy of therapeutic manipulation on the quality of bone.14–17 There are some limitations to IR analyses of bone. First, because bone is a hard tissue, thin tissue sections are difficult to prepare. Generally, to circumvent this problem, bone is fixed and embedded in a hard substance so that it can be cut with a microtome. This is time consuming, and the embedding materials commonly in use have spectral contributions that overlap the mineral and matrix components of bone.18 Additionally, for studies on humans, biopsies are needed and these are not often obtained, limiting the number of samples available for analysis and thus the generality of the study results. For this reason, animal models of these disease states are often analyzed, with the caveat that the findings in growing animals and adult humans may not be in total agreement.

11.3.1.3 Representative studies The vibrations that provide the most information about bone structure are the phosphate ν1 , ν3 (∼1200–900 cm−1 ), phosphate ν4 (640–560 cm−1 ), carbonate ν2 (890–850 cm−1 ), and amide I and II (1720–1580 cm−1 ). These are shown in Fig. 11.2(a), which is a typical spectrum of adult vertebrate cortical bone. This spectrum is different in appearance from most of the serial spectra shown in Fig. 11.2(b) demonstrating the progression of spectra through the growth plate of an immature mouse. The mature bone spectra in Fig. 11.2(a) and (b) are similar. Maps of variation in growth plate mineral spectra enabled us to challenge the statement that the mineral content of calcified cartilage is greater than that of the trabecular bone that lies just beneath it. The conflict was suggested by variable literature reports based on ash weight determination that appeared controversial because of difficulties in dissecting out different regions for bulk analysis. By mapping the ratio of the integrated area of the phosphate ν1 , ν3 band to the amide I band a mineral : matrix ratio (related to the mineral content of the tissue) was calculated through the growth plate showing a gradient of values (Fig. 11.2(a)). There are spectral differences moving across trabecular bone (Fig. 11.2(c)). Fourier transform infrared (FTIR) microscopic imaging and microspectroscopy have enabled hypotheses about the functions of bone matrix proteins and bonerelated cytokines to be evaluated and sometimes challenged through detailed analysis of knockout and transgenic animals. For example, osteocalcin is one of the most abundant noncollagenous proteins found in all bones. It contains several gammacarboxy glutamate residues, whose synthesis requires vitamin K. This anionic structure enables it to bind to bone mineral. Initial examination of the appearance of the bones of mice in which the gene for osteocalcin was ablated (knockout animals) showed wider bones, which was interpreted as demonstrating a role for osteocalcin in bone formation.19 IR analyses of the bones of the knockout and control animals (Fig. 11.3(b)) revealed that the osteocalcin-deficient animals had a significantly greater mineral : matrix ratio but a lower crystallinity (crystal size and formation)

239

IR IMAGING OF HARD AND SOFT TISSUES PO4(n1,n3)

Absorbance

(a)

PO4(n4) Amide I Amide II

CO3(n2)

600

800

1000

1200

1400

1600

1800

Wavenumber (cm–1) (c) 60 Scaled absorbance

Scaled absorbance

(b) 15 10 5 0

800

1000 1200 1400 1600 1800 Wavenumber (cm–1)

50 40 30 20 10 0 –10 800 1000 1200 1400 1600 1800 Wavenumber (cm–1)

Figure 11.2 IR spectra of bone. (a) Single spectrum of adult bovine long bone showing relevant peaks. (b) Series of spectra with the least intense originating from the resting zone of the growth plate and intensities increasing as the metaphyseal bone under the calcifying cartilage is approached. (c) Serial spectra moving across a trabecular bone in the mouse, from outside to center. The spectrum from the periosteal surface is the least intense, and that in the center of the trabecular bone most intense. Note the change in the shape of the phosphate band as the bone becomes more mineralized.

than the age-matched wildtype controls examined at all sites.20 This and the greater effects seen in bones that are known to make more osteocalcin supported the idea that this protein was more important as a regulator of mineral growth and formation. From other data sources, it was known that osteocalcin was also involved in the recruitment of cells (osteoclasts) important for bone remodeling. The IR data could not separate these two mechanisms, although further studies in calcium-deficient animals indicated that both mechanisms were important for the behavior of osteocalcin.21 There is a second gamma-carboxy glutamic acid containing protein in bone and cartilage – matrix gla protein (MGP). Ablation of this protein causes an increase in mineral content in the cartilage of these mice, and excessive growth plate calcification is illustrated in Fig. 11.3(a). In the MGP-deficient animals (MGP −/−) the gradient normally seen in the control (wildtype) animals’ growth plate is not apparent, demonstrating that matrix gla protein is a mineralization inhibitor.

300

200

100

0

0

8

16

24

32

40 *

WT

Periosteum

MGP –/–

100 200 300

mm 0

Center

300

200

100

0

KO

Endosteum

*

MGP +/+

100 200 300

0

5

10

0

3

6

9

12

15

18

21

0.0

3.0

6.0

9.0

12.0

(b) 15.0

WT

r Pe

t en C

er En

KO

Cortical bone

i

um te os

d

t os

eu

m

Figure 11.3 Studies in mutant mice. (a) Image of mineral to matrix in the growth plate of a young mouse lacking matrix gla protein (MGP−/−) and its wildtype control (MGP+/+). Note the gradient of mineral : matrix ratios in the wildtype is not apparent in the knockout. (b) Mineral : matrix increases at three sites in the cortices of the osteocalcin knockout (KO) mouse (6 month data), while crystallinity is decreased relative to wildtype (WT) controls. (c) In the osteonectin knockout mouse (4 months old) the collagen maturity assessed in terms of the 1660 : 1690 peak area ratio is increased on the periosteal and endosteal surface.

(c)

mm

0

1660/1690 (% area)

(a) Crystallinity (% area) Mineral : matrix ratio

IR IMAGING OF HARD AND SOFT TISSUES

241

Osteonectin is another predominant matrix protein whose function in bone was not known before the generation of osteonectin-deficient animals.22 Osteonectin is a calcium-binding protein; animals deficient in osteonectin were first reported to have cataracts and show impaired wound healing, but the effects of this predominant bone protein were not known. The osteonectin-deficient mice had less strong bones, and the trabeculae within the marrow cavities seemed to vanish with age.22 The FTIR analyses demonstrated that the trabecular bone had lower mineral content and lower crystallinity, but most significant (Fig. 11.3(c)) was the increase in the collagen maturity seen on the sides of the bone where new formation (periosteal edge) and remodeling (endosteal edge) were occurring.23 This alteration in collagen maturity was most likely a major contributor to the mechanical properties in the bones of these animals. The other major area in which IR imaging and microspectroscopy have been useful is in the analysis of mineral quality in disease states. The mineral and matrix properties in a mouse model of the rare genetically inherited brittle bone disease – osteogenesis imperfecta – showed smaller crystals with an abnormal composition.23 Patients with this disease have a high risk of fracture, depending on the site of the mutation in the collagen gene. IR offers the opportunity to address the impact of different mutations in the collagen molecule on mineral deposition. Osteopetrosis is another rare disease in humans that has been analyzed by IR imaging. Patients with this disease have bones that are like rock, impeding their mobility and increasing their pain. In many patients, the cause of the disease is the inability of the bone resorbing cells, osteoclasts, to remodel the bone. For this reason there is a persistence of calcified cartilage. The IR data demonstrated increased mineral content and decreased crystal size, consistent with the properties seen in the bones of animal models of this disease.13 In contrast to these rarer bone diseases, osteoporosis is a very common disease affecting men and women as they age. It is associated with a loss of bone (mainly trabeculae) and an increased risk of fracture. IR has enabled significant differences between normal and osteoporotic tissues to be identified. Osteoporotic animals and humans have fewer trabeculae, but in the trabeculae that are present the mineral content is reduced (Fig. 11.4(a)) while the crystallinity and collagen maturity are increased (Fig. 11.4(b) and (c)). Moreover, there is a distinct difference in the distribution of these parameters as a function of distance from the surface of the trabeculae. This is illustrated for four cases (three osteoporotic and one control) in which a line was drawn across the trabecular width to detect these parameters (Fig. 11.4(d)); but this is seen throughout every biopsy and in all biopsies examined for osteoporotic or normal control patients. The osteoporotic studies are now being extended to analysis of effects of a variety of widely used therapies compared with the properties existing in age-matched controls that did not have osteoporosis. Analyses of multiple sites in biopsies from larger numbers of women and men with osteoporosis showed a decrease in mineral : matrix ratio, an increase in crystallinity, and an increase in collagen maturity.12–15 Hormone replacement therapy15 and parathyroid hormone16 shift the osteoporotic bone closer

–200

0

100

200

mm from center of trabeculae

–100

Normal OP

Normal OP

300

(c)

0

1

2

3

(d) 4

0

0 0

1

2

3

240

360

480

60

90

120 mm from trabecular edge

30

mm from trabecular edge

120

150

600

Normal OP

Figure 11.4 IR analyses of osteoporotic and control bone (data presented here come from post-menopausal women with reduced bone mineral density as contrasted with age-matched normals who show no evidence of bone disease). (a) Average mineral : matrix ratio going across a trabecular bone in a representative individual. Mean and SD represent reproducibility of measurement across similar lines. (b) Collagen maturity expressed as 1660 : 1690 intensity ratios going across the same trabeculae as indicated in (a). (c) Variation in crystallinity (1030 : 1020 peak area ratios) across the same trabeculae. (d) Individual data from four patients (three with osteoporosis) and one control (large solid diamonds). The line connecting the normal data indicates the typical variation seen in the normal. This is not noted in any of the three other female patients.

–300

0.5

1.0

1.5

2.0

2.5

3.0

(b) 3.5

0.5

1.0

1.5

2.0

2.5

3.0

3.5

(a) 4.0

Mineral : matrix

1660 : 1690

Crystallinity Mineral : matrix

IR IMAGING OF HARD AND SOFT TISSUES

243

to normal. We are currently evaluating a selective estrogen modulator and a bisphosphonate to determine their effects on bone mineral and matrix properties, while trying to establish the association between the parameters measured in the IR and bone mechanical strength.

11.3.2 Skin 11.3.2.1 The tissue Skin provides a barrier for permeability at points at which body surfaces come into contact with the environment. A visible micrograph of a piece of pig skin (frozen section, unstained) of dimensions 400×400 μm is shown in Fig. 11.5(a). Three main anatomical regions of the sample as shown are the stratum corneum, the epidermis and the dermis. The stratum corneum, the outermost layer of the epidermis, is a layer typically 15 μm thick (although values up to 60 μm have been reported for areas such as the wrist), which provides the major barrier for permeability, in particular the retardation of water loss from the interior. The epidermis is the overall protective barrier and is usually about 50 μm thick although on palms and soles this value is increased. Below the stratum corneum are three successive layers named (from the surface down), the stratum granulosum, containing polygonally shaped cells, the straum spinosum, containing polyhedron-shaped cells possessing a spiny appearance, and the stratum basale, the area of the epidermis where stem cells provide daughter cells that migrate to the surface layers. Underlying the epidermis is the dermis, a tough supportive connective tissue matrix containing numerous specialized structures. There is substantial history regarding the application of conventional vibrational spectroscopy methods to study the intact surface of skin, the extracted stratum corneum and the ceramide–cholesterol–fatty acid mixtures that constitute the primary lipid components of the barrier. The complexity of the barrier and the multiple phases formed by the interactions of the barrier components have begun to reveal the role of each of these substances in barrier structure and stability. The use of bulk phase IR to monitor lipid phase behavior and protein secondary structures in the epidermis, as well as in stratum corneum models, is also well established.24–28 In addition, in vivo and ex vivo attenuated total reflectance (ATR) techniques have examined the outer layers of skin to probe hydration levels, drug delivery and percutaneous absorption at a macroscopic level.29–32 Both mid-IR and near-IR spectroscopy have been used to differentiate pathological skin samples.33,34 The above studies, and many others too numerous to mention, lend confidence to the fact that the extension to IR imaging will produce useful results. 11.3.2.2 Questions solved by IR microscopy In this section, we outline the applications of IR microscopic imaging to the molecular level determination of dermal and transdermal percutaneous absorption. The rationale for these experiments is that drugs often exhibit low penetration rates

Wavenumber (cm )

–1

3600 3200 2800 1600 1200

C=O

3100 3000 2900 2800 Wavenumber (cm–1 )

2856

nsCH2

Lipid-rich area Epidermis

400 mm

Dermis (c)

(e)

1600 Wavenumber (cm–1 )

3600 3200 2800

1200

3100 3000 2900 2800 Wavenumber (cm–1 )

2875

nsCH3

Protein-rich area

Figure 11.5 (a) An optical microscopic view (frozen section 5 μm) showing a 400 × 400 μm section of pig skin. (b) An IR spectrum from a lipid-rich region of the section. The location at which the spectrum was taken is indicated with an arrow. Note the relatively intense C=O band near 1720 cm−1 . (c) An IR spectrum from a protein-rich region of the section. The location at which the spectrum was taken is indicated with an arrow. Note the weakness of the C=O band near 1720 cm−1 . (d) An expanded version of the C–H stretching region from the lipid-rich area of the sample. The strong features near 2856 cm−1 (labeled) and 2920 cm−1 arise from the symmetric and asymmetric CH2 stretching vibrations of the lipid chains. (e) An expanded version of the C–H stretching region from the protein-rich area of the sample. The feature near 2875 cm−1 (labeled) arises from the symmetric CH3 stretching vibrations primarily from protein.

(b)

(d)

(a) Stratum corneum

IR IMAGING OF HARD AND SOFT TISSUES

245

through the hydrophobic layers of the stratum corneum. It is known that the stratum corneum provides the rate-limiting step to both solute transport and water loss. Control of transport rates is of interest in several areas of dermatological and drug delivery research. For example, optimal therapeutic intervention may require delivery of drugs at particular concentrations to well-defined locations within the skin. In addition, knowledge of permeation pathways of penetration-enhancing molecules would provide a quantitative basis for understanding the thermodynamic and kinetic mechanisms of action of these substances. Two strategies for improving and controlling transport rates are in vogue. First, penetration enhancers, such as dimethylsulfoxide (DMSO), oleic acid, and terpenes, have been recruited for this purpose. The second strategy involves the use of encapsulation agents, such as liposomes, for delivery. To evaluate the permeation of exogenous substances, a variety of physical techniques of varying degrees of sophistication have been employed. Traditional methods for evaluation of skin permeation involve measuring the delivery of radioactive materials through skin, or alternatively analyzing sequential ‘tape-strips’, successively removed from the outermost epidermal layers. These approaches predate modern biophysics, lack spatial resolution and provide neither molecular structure information about physical factors that control enhancer performance, nor insight into permeation pathways. However, in the past decade, a variety of modern microscopic and spectroscopic approaches have provided some insight into these issues. These include electron microscopy and small-angle X-ray scattering techniques to track adsorption and permeation of liposomes, and fluorescencemicroscopy-based approaches, such as confocal laser scanning microscopy35 and two-photon fluorescence microscopy to provide optical sectioning of samples without the need for physically altering the specimen.36,37 Infrared microscopic imaging provides the significant advantages of direct spatially resolved concentration and molecular structure information for sample constituents. Raman microscopy (not further discussed in this chapter) possesses the additional benefit of confocal acquisition of this information and a 10-fold increase in spatial resolution at the expense of reduced signal-to-noise ratios compared with IR. The interested reader is urged to check the seminal studies of the Puppels group in Rotterdam,38–40 as well as our own initial efforts in this direction.41 The current section describes the initial applications of IR microspectroscopic imaging to monitor the permeation and tissue distribution of the dermal penetration enhancer, DMSO, in porcine skin as well as to track the extent of permeation of phospholipid vesicles.

11.3.2.3 Representative results Infrared spectra from two regions of a microscopic section skin, namely, a lipidrich region and a protein-rich region are shown in Fig. 11.5(b) and (c), respectively. Arrows point to the respective regions. Numerous differences are evident between the two spectra. Most obvious is the presence of a strong lipid C=O stretching

246

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

band near 1720 cm−1 in Fig. 11.5(b), which has only a very weak counterpart in Fig. 11.5(c). Striking differences are also evident in the C–H stretching region of each sample that are presented on an expanded frequency scale in Fig. 11.5(d) (lipid-rich) and Fig. 11.5(e) (protein-rich). The lipid-rich spectrum reveals contributions mainly from CH2 groups, in particular, the symmetric stretching mode near 2856 cm−1 , while the protein-rich C–H stretching region presents bands primarily from methyl groups, as typified by the symmetric CH3 stretching vibration near 2875 cm−1 . 11.3.2.3.1 Permeation of DMSO. To illustrate the power of IR imaging for studies of penetration enhancers through skin, the spatial distribution of a sample of DMSO-d6 that had been topically applied is depicted in Fig. 11.6. For these studies, solutions of DMSO at ∼15 μl cm−2 were applied to the stratum corneum surface of porcine skin and left for 3 h, after which the skin surface was wiped clean to

Depth DMSO-d6 (microns) 0

Lipid CH2

Protein

Amino acids

Micrograph air

SC

400 350

750 700

1100 1000 400 mm

1400 400 mm Figure 11.6 Permeation of DMSO-d6 through skin: visible micrograph and various FTIR parameters. The IR images are color coded red>yellow>green>blue for each spectral parameter. The value corresponding to each color varies from image to image. The following vibrational modes are used to construct the IR images: DMSO-d6 distribution is estimated from the area under the CD3 stretching vibrations; lipid distribution is measured from the CH2 symmetric stretching mode intensity. Protein distribution is measured in two ways, from the integrated intensity of the amide II mode, and from amino acid side-chain modes near 1340 cm−1 . (SC = stratum corneum.)

IR IMAGING OF HARD AND SOFT TISSUES

247

remove all remaining exogenous material prior to microtoming. Frozen sections were placed on BaF2 IR windows and used directly for IR imaging. A series of visible and IR images, the latter corresponding to the spatial distributions of the endogenous lipid or protein, or the exogenous DMSO, were generated from spectra of a section cut perpendicular to the surface layer. The spatial distribution of the protein was monitored in two ways, from the amide II intensity (∼1550 cm−1 ) and from the intensity of amino acid side-chain modes near 1340 cm−1 . Lipid localization was derived from the CH2 symmetric stretching intensity. The DMSO distribution was determined from the peak areas under the CD3 stretching vibrations. Although the DMSO intensity apparently parallels that of the bands derived from protein, it is not necessarily correct to conclude that DMSO is excluded from lipid containing regions, since the area around the corneocytes is surrounded by lipidic structures that are of a spatial dimension much smaller than the spatial resolution of the current IR experiment. However, the ability of IR imaging to track the permeation of exogenous materials is clearly demonstrated. 11.3.2.3.2 Permeation of liposomes. Figure 11.7 reports on IR determination of the permeation of liposomes through the stratum corneum. An optical micrograph of a skin section to which acyl chain 1-palmitoyl-d31 , 2-oleoylphosphatidylcholine (P-d31 OPC) liposomes had been topically applied is shown in Fig. 11.7(a). The stratum corneum is at the bottom of the figure, the liposomes on the surface are barely detectable. Spectra of the CH stretching region acquired along the dashed arrow shown in Fig. 11.7(a) are plotted in Fig. 11.7(b). The stratum corneum location is deduced from the presence of a signal near ∼2849 cm−1 arising from ordered chains (arrow in Fig. 11.7(b)). Further into the epidermis, methyl CH stretching vibrations from protein constituents appear near 2880 cm−1 . An image of the CH2 spatial distribution constructed from the intensity near 2849 cm−1 is shown in Fig. 11.7(c). This frequency is characteristic of highly ordered chains known to exist in the stratum corneum, and thus provides a direct identification of this tissue region. The color coding ranges from white (highest levels) to black (lowest levels). The stratum corneum is clearly delineated in the white regions near the surface. Spectra of the CD2 stretching region (2250–2000 cm−1 ) acquired along the dashed arrow shown in Fig. 11.7(a) are plotted in Fig. 7(d). The location of the exogenous lipid is clearly revealed from this spectral region. An image of the CD2 spatial distribution constructed from the intensity near 2090 cm−1 is shown in Fig. 11.7(e). The image reveals the extent of liposome permeation, which in this instance is through the stratum corneum and into the epidermis. A variety of important issues may be addressed through experiments of this type, for example, kinetics of permeation, spatial localization of liposomes, etc. The second issue upon which the current results depend, namely the physical state of the exogenous lipids upon permeation, was addressed through the wellknown sensitivity of the methylene stretching modes to chain conformational order. Varying views pertaining to this issue have been propounded and have been reviewed recently by Bouwstra et al.42

(d)

2000

500 mm

2050

2100 2150 Wavenumber

–0.05 2800

0.00

0.05

0.10

0.15

0.20

0.25

0.30

(b) 0.35

2200

2850

2250

(e)

500

400

300

200

100

0

0

2900 2950 Wavenmunber

100

3000

200

0

0

300

500

400

300

200

100

mm

(c)

400

100

500

200 300 mm

400

500

Figure 11.7 Permeation of liposomes through a skin section. (a) Visible micrograph of a skin section (stratum corneum at the bottom of the figure) to which P-d31 OPC liposomes had been applied topically. The liposomes are barely detectable. (b) Spectra of the CH stretching region acquired along the dashed arrow shown in (a). The stratum corneum location is deduced from the presence of a signal near ∼2849 cm−1 arising from ordered chains. (c) An image of the CH2 spatial distribution constructed from the intensity near 2850 cm−1 . The color coding ranges from white (highest levels) to black lowest levels. The stratum corneum is clearly delineated in the white regions near the surface. (d) Spectra of the CD2 stretching region (2250–2000 cm−1 ) acquired along the dashed arrow shown in (a). (e) An image of the CD2 spatial distribution constructed from the intensity near 2090 cm−1 . The image reveals the extent of liposome permeation.

(a)

IR absorbance

IR IMAGING OF HARD AND SOFT TISSUES

249

The CD2 stretching mode near 2090 cm−1 is very sensitive to chain conformational order in a manner similar to the CH2 stretching modes, that is, a lower frequency indicates more of all-trans conformers. We have examined the CD2 stretching frequencies of P-d31 OPC liposomes after they were allowed to permeate into the stratum corneum. This molecular species was chosen since at room temperature, these vesicles are conformationally highly disordered. Thus, if upon permeation of P-d31 OPC into the SC, the liposomes were to be disrupted, a decrease in this frequency would be expected, since if the highly disordered chains were to mix with the highly ordered ceramides of the SC, the former would tend to become conformationally ordered. If, on the other hand, the liposomes were to remain intact on permeation, the frequency should remain approximately constant. The experimental results are shown in Fig. 11.8. Factor analysis was applied to analyze the data. An optical micrograph of the section is shown in Fig. 11.8(a), and the exogenous liposomes are observable visibly (albeit with difficulty) above the stratum corneum. Two relevant factors (labeled 1 and 2) are displayed in Fig. 11.8(b). The factors, although similar, exhibit a slight (∼1 cm−1 ) difference in the position of the maximum of the CD2 stretching vibration, factor 1 showing a slightly lower frequency. An image of the scores derived from each factor are shown in Fig. 11.8(c) (factor 1) and Fig. 11.8(d) (factor 2). In regions where the liposomes have penetrated into the stratum corneum, factor 1 dominates. The ∼1 cm−1 frequency decrease is consistent with conformational ordering of the perdeuterated palmitate chain of those liposomes that have permeated the skin. While small changes in the appearance of the factors must still be observed many times for gaining confidence in the above interpretation, these data are consistent with disruption of the vesicle structure and mixing of the P-d31 OPC with the stratum corneum lipids. In contrast, those vesicles that remain on the skin surface are represented by factor 2, that is, with a slightly increased CD2 stretching frequency. Three advantages of IR microscopic imaging for the studies of permeation into skin are clear from the above data. First, the location of exogenous material is directly monitored. Second, conformation-sensitive spectral features provide useful indicators of changes in molecular structure. Third, if the exogenous material perturbs the molecular structure of the native skin components, the nature and location of the disruption can be imaged directly from the spectra of the endogenous material.

11.3.3 Cartilage 11.3.3.1 The tissue Articular cartilage is a specialized connective tissue covering the articular ends of diarthrodial joints. It provides a near-frictionless load-bearing surface,43 and in humans it ranges from ∼1 to 3 mm thick in the knee and hip joints.44 The framework of cartilage is composed of a network of type II collagen fibrils that are associated with type IX and XI collagens, noncollagenous proteins and proteoglycan (PG) components.45 Aggrecan, the primary cartilage PG, contains covalently attached side chains of the glycosaminoglycans chondroitin sulfate and keratin sulfate linked

400

300

200

100

0

0

100

200

300 mm

400

500

Factor 1 – permeated P-d31OPC

600

–0.04

–0.02

0

0.02

0.04

0.06

(d)

B

400

300

200

100

0

0

0.0 2160

0.1

0.2

0.3

(b) 0.4

2200 Wavenumber

2220

100

200

300 mm

400

500

Factor 2 – surface P-d31OPC

2180

1

Factors: 2

600

2240

–0.05

0

0.05

Figure 11.8 Structural information from IR images. (a) Optical micrograph of a frozen skin section to which P-d31 OPC liposomes had been applied topically. (b) Factors corresponding to permeated P-d31 OPC (factor 1) or surface P-d31 OPC (factor 2). A slight (∼1 cm−1 ) frequency shift is detected between the peak positions. The slightly lowered frequency for factor 1, suggests a more ordered lipid chain. (c) A score image of factor 1, showing permeation. (d) A score image for factor 2, showing this factor residing primarily on the surface.

mm

(c)

(a)

mm

IR IMAGING OF HARD AND SOFT TISSUES

251

Superficial zone

Middle zone

Deep zone

Tidemark Calcified zone Subchondral bone

Figure 11.9 Schematic of articular cartilage structure showing the superficial, middle and deep zones, the tidemark boundary between the noncalcified and calcified cartilage layer, and the subchondral bone that underlies the articular cartilage.

to a protein core. Cartilage has substantial zonal heterogeneity with respect to its molecular composition and cellular morphology (Fig. 11.9). The surface layer of cartilage, the superficial tangential zone (STZ) (∼10% of total thickness), consists of densely packed type II collagen fibrils, a small amount of PGs, a high water content and flattened, ellipsoidal-shaped chondrocytes with their longer sides parallel to the articular surface.46,47 The midzone (MZ) (40–60% of the total thickness) has a lower water content than the STZ, the chondrocytes are randomly distributed and have a rounded, spherical shape, and there is a higher concentration of PG and a lower concentration of collagen fibrils (larger diameter, more widely spaced and randomly oriented compared with the STZ). The cells in the deep zone (DZ) (about 30% of the total thickness) have rounded cells similar to the MZ but are aligned in columns that are perpendicular to the surface of subchondral bone. This zone contains collagen fibrils with the largest diameter, the highest concentration of PGs and a lower percentage of water than the other zones. Here, the collagen fibrils are oriented radially and penetrate through the tidemark, a thin basophilic boundary between the uncalcified and calcified cartilage. Finally, the calcilifed zone is a thin layer of calcified cartilage that separates the hyaline articular cartilage from the subchondral bone. The chondrocytes in the calcified zone are fewer and smaller in size than the cells in the DZ. The collagen fibers from the DZ are embedded into the calcified zone, and anchor themselves to the subchondral bone. Pathological changes in cartilage turnover involve the degradation of collagen and PGs in the cartilage matrix triggered by either cellular or mechanical signals.48–50 Such changes are characteristic of osteoarthritis (OA), which is a

252

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

progressively disabling musculoskeletal disease affecting millions of people in the United States,51 with prevalence directly correlated with age. Advanced stages of the disease are characterized by extensive cartilage damage that can eventually lead to total loss of joint function and joint arthroplasty. A significant complication in the treatment of OA is the inability to diagnose the disease process at an early stage, and the lack of methods to evaluate the tissue response to therapeutic and tissue engineering interventions. While macroscopic, later-stage cartilage damages such as lacerations, ruptures and chondral fractures can be recognized via existing technology,52 early stages of OA that involve disruption of the cartilage matrix but no obvious mechanical damage are much more difficult to identify. In addition, detection and quantitation of compositional and structural changes in articular cartilage repair tissue formed by various tissue engineering approaches is hindered by the difficulty in nondestructively evaluating the matrix. Knowledge of specific ultrastructural changes at early stages of degeneration and in repair tissue would be extremely important during arthroscopic procedures, where crucial decisions are made regarding salvaging or removing cartilage and meniscus. Magnetic resonance (MR) imaging has emerged as a primary modality for such assessment, with anatomic definition of cartilage on a spatial scale of roughly 100 μm now being supplemented by use of MR techniques that are more or less specific for hydration and for collagen and PG content.53,54 However, to truly characterize the microstructure and biochemistry of cartilage, extraction of tissue for in vitro studies is required. FTIR spectroscopy is a convenient tool to study the previously described changes in degenerative cartilage structure. Although there have been many conventional FTIR studies on collagen (including the effects of hydration on secondary structure55,56 and molecular changes during self-assembly57 and some FTIR studies on extracted PGs from various tissues,58–60 until very recently, IR spectroscopic studies on articular cartilage were nonexistent. The first reported IR studies on cartilage were imaging experiments from our group61 followed closely by those of Potter et al.62 These investigations provided a unique opportunity to study the relative amount, molecular nature, distribution and orientation of the components of cartilage at a pixel size of ∼6.3 μm and a spatial resolution of ∼10–12 μm. In our studies, IR images acquired through the superficial, middle, and deep zones of thin sections of bovine articular cartilage showed variation in the intensities of bands arising from the primary nonaqueous components of cartilage, collagen and PG (primarily aggrecan), and thus reflected the differences in quantity of these specific components. In addition, spectra of mixtures of model compounds, which had varying proportions of type II collagen and aggrecan, were analyzed to identify spectral markers that could be used to quantitatively analyze these components in cartilage. It was found that the integrated area of the amide I absorbance (1720–1590 cm−1 ) correlated with collagen concentration, and the ratio of the integrated area of the PG sugar ring C–O absorbance (1140–985 cm−1 ) to the amide I absorbance area correlated to the quantity of PG. Polarization experiments were performed to determine the spatial distribution of the collagen orientation in the

IR IMAGING OF HARD AND SOFT TISSUES

253

different zones of cartilage, and the ratio of the collagen amide I/amide II absorbance was utilized as a spectral marker of collagen orientation. In Potter et al.,62 IR imaging data were analyzed with multivariate techniques to define the spatial distribution of matrix constituents in native and engineered cartilage samples. A more recent study by Rieppo et al.63 advanced the quantitative IR imaging analysis of cartilage by incorporation of a standard into histological sections to account for variability in section thickness. Taken together, these studies provided a framework in which complex pathological changes in cartilage could now be further assessed with this technology. Characterization of molecular changes in native, repair and tissue-engineered cartilage is an essential element in the development of therapeutic approaches to OA. Cartilage is spatially heterogeneous at the microscopic level, and superimposed degradation processes greatly increase heterogeneity. Therefore, quantitation of the molecular composition at high spatial resolution is critical. Toward these goals, IR imaging studies of degenerative, repair and tissue-engineered cartilage, and of model systems of cartilage degradation that mimic OA, have been undertaken to identify molecular changes in cartilage. The studies detailed below provide further insight into the mechanisms underlying the initial stages of matrix disruption that eventually lead to total joint degradation. Furthermore, the data from these studies form the foundation for the use of IR spectroscopic techniques that can be applied clinically in an in situ environment, that is, infrared fiber optic probe methods.64

11.3.3.2 Representative results 11.3.3.2.1 Spectral markers of collagen degradation in OA. Figure 11.10(a) shows a photomicrograph and histology images of cartilage from normal and osteoarthritic regions of a tibial plateau harvested during knee replacement surgery. These tissues are initially visually identified and graded as either grossly normal (no obvious macroscopic damage) or degraded (fibrillations, clefts or fissures present), corresponding to the Collins’ Visual Scale65 grade 1 and grade 3, respectively. IR imaging analysis of thin sections from these tissues revealed subtle, but consistent changes between grade 1 and grade 3 sites (Fig. 11.10(b)).64 Quantitation of spectral parameters associated with the collagen molecule demonstrated area and peak height changes in the amide II peak and in the absorbance centered at 1338 cm−1 . Specifically, an increased amide II/1338 cm−1 area ratio was found for the more degraded tissue regions (Fig. 11.10(c)). Polarized IR imaging studies of the grade 3 tissues also found changes in collagen orientation throughout the tissue zones as reflected by the ratio of the amide I/amide II absorbance. To understand the molecular origin of these spectral changes, normal bovine articular cartilage was treated with collagenase, an enzyme specific for collagen breakdown.66 This enzymatic treatment serves as an in vitro model of the pathological collagen degradation that occurs in OA. The control and treated specimens showed spectral differences that had previously been observed for normal and OA cartilage (Fig. 11.11(a)–(c)). With collagenase

254

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a)

(c) Amide II/1338 cm–1 area ratio

Grade 1

Higher higher

(b)

Absorbance

Grade 1 Grade 3

Lower lower

Grade 3 1338 cm–1

400 mm

1700 1600 1500 1400 1300 1200 1100 1000 Wavenumber (cm–1 )

Figure 11.10 (a) Tibial plateau cartilage harvested from joint replacement surgery (left side, cm scale on bottom). Nearly normal ‘grade 1’ and degraded ‘grade 3’ are shown. On the right are histological sections from grade 1 and grade 3. Note the increased surface fibrillation on the grade 3 cartilage. (b) FTIR imaging spectra collected from grade 1 and grade 3 cartilage. The absorbance centered at 1338 cm−1 is decreased in the spectrum from grade 3 cartilage. (c) IR images created from the ratio of the integrated area of the amide II absorbance to the 1338 cm−1 absorbance. This ratio is increased substantially in grade 3 tissue.

treatment, the amide II/1338 cm−1 area ratio increased in the superficial zone, and polarized IR measurements demonstrated a more random orientation of the collagen fibrils (decreased amide I/amide II ratio in the superficial zone) that correlated spatially with the immunohistochemical determination of broken type II collagen. These studies demonstrated that IR imaging is sensitive to changes related to the degradation of cartilage, specifically to the breakdown of type II collagen. The 1338 cm−1 collagen absorbance arises from side-chain vibrations (possibly from the wagging mode of methylene groups) and is apparently sensitive to the order of the triple helix. This absorbance band had previously been examined in other studies and was shown to decrease in intensity as the collagen denatures,67 and thus appears to be a reasonable spectral marker to monitor collagen degradation in biological tissues. The changes in the polarized amide I/amide II area ratios evidently arise from disrupted orientation of the cleaved collagen molecules, and strongly support the association of the observed spectral changes with the unraveling of the collagen

255

IR IMAGING OF HARD AND SOFT TISSUES

(a)

Control Treated Amide II/1338 cm–1 area ratio

(b)

60

Polarized IR spectra from the superficial zone of an untreated and collagenase-treated specimen Untreated

Amide I 0.6 0.5

Area ratio

50

Collagenasetreated

Amide II

0.4 40

0.3 0.2

30

0.1 20 (c)

1700

1600

1500

1400

1300

1200

IR images of polarization data Highest amide I/II ratio Lowest amide I/II ratio Higher SZ 400 µm

Lower lower Control (untreated)

Collagenasetreated

Figure 11.11 (a) Data from control and collagenase-treated bovine articular cartilage sections. The ratio of the integrated area of the amide II absorbance to the 1338 cm−1 absorbance is increased with collagenase treatment. Polarized IR imaging data (b), (c) show a decreased amide I/amide II area ratio, indicative of reduced collagen orientation in the collagenase-treated sections.

triple helix structure due to enzymatic activity. Thus, two key spectral parameters have been elucidated to monitor cartilage degradation in disease. Although comparisons were made between normal and somewhat degraded tissues, the sensitivity of IR spectroscopy to such changes bodes well for the eventual capability to monitor molecular changes at earlier stages of disease. 11.3.3.2.2 Cartilage repair tissue. Cartilage lesions occur due to various etiologies, such as intraarticular fractures or as a result of ligament injuries. The defects typically do not heal, and as such, are the target of varied therapeutic protocols aimed at improved healing of repair tissue or at improved integration of engineered cartilage into the defect site. Since the morphology and composition of repair and engineered cartilage are likely to be predictive of their ultimate in vivo functionality,68 knowledge of the composition and distribution of molecular components in these tissues, and how they compare to native cartilage is essential. We utilized IR imaging to examine paraffin-embedded biopsies from patients with medial femoral condyle

256

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Higher

Lower

Figure 11.12 (A) Light microscope image of unstained section of native cartilage and repair tissue. Chondrocytes (ch) are also visible. (B) FTIR image calculated from collagen amide I peak area shows distribution of collagen. (C) FTIR image of PG distribution calculated from the ratio of the PG sugar C–O absorbance (1140–985 cm−1 ) to the amide I absorbance. (D) FTIR image from polarized data calculated from the collagen amide I/amide II ratio. In this image, a higher ratio indicates greater collagen fibril orientation vertically.

cartilage defects and compared the repair tissue to the native cartilage. The collagen amide I distribution enabled clear differentiation of native and regenerative cartilage, and of chondrocytes (Fig. 11.12(A) and (B)). The regenerative tissue typically had a lower collagen content compared to native cartilage, and a distinct interface was present between the regenerative and native tissue. The PG content of the regenerative tissue was also significantly lower than that of the native articular cartilage (Fig. 11.12(C)), and the orientation of the repair collagen was more random compared with the collagen in the native cartilage (Fig. 11.12(D)). Thus, overall the results indicated that the composition and structure of the regenerative tissue in these biopsies was significantly different in several ways from the native cartilage. In contrast to this comprehensive set of results obtained by imaging one section by FTIR, typical histological evaluation of repair tissue would involve the use of at least four different special stains to similarly characterize the tissue: H&E for morphology, immunostaining for collagen, safranin O staining for qualitative PG distribution, and picrosirius red with polarized light microscopy for collagen orientation. Clearly, IR microscopic imaging analysis has many advantages over traditional analyses for the characterization of cartilage repair tissue and the monitoring of cartilage degradation. The ability to simultaneously obtain information about collagen and PG

IR IMAGING OF HARD AND SOFT TISSUES

257

contents, distribution and orientation will enhance our understanding of cartilage repair, and aid the design of new therapeutic approaches for OA and other cartilage diseases.

Acknowledgements This work was supported by NIH grants AR041325 (ALB), AR046121(ALB, NPC) GM29864 (RM) and EB00744 (NPC).

References [1] Barer, R., Cole, A. R. H. and Thompson, H. W. (1949) Infra-red spectroscopy with the reflecting microscope in physics, chemistry and biology. Nature 163, 198–201. [2] Boskey, A. L. (2001) Bone mineralization. In Bone Biomechanics, 3rd edn (S. C. Cowin, ed.), CRC Press, Boca Raton, FL, pp 5.1–5.34. [3] Elliott, J. C. (1973) The problems of the composition and structure of the mineral components of the hard tissues. Clin. Orthop. 93, 313–45. [4] Cohen, L. and Kitzes, R. (1981) Infrared spectroscopy and magnesium content of bone mineral in osteoporotic women. Isr. J. Med. Sci. 17, 1123–25. [5] Mendelsohn, R., Hassankhani, A., DiCarlo, E. and Boskey, A. (1989) FT-IR microscopy of endochondral ossification at 20 mu spatial resolution. Calcif. Tissue Int. 44, 20–4. [6] Grynpas, M. D. and Marie, P. J. (1990) Effects of low doses of strontium on bone quality and quantity in rats. Bone 11, 313–19. [7] Hanschin, R. G. and Stern, W. B. (1995) X-ray diffraction studies on the lattice perfection of human bone apatite (Crista iliaca). Bone 16, 355S–63S. [8] Luppen, C. A., Smith, E., Spevak, L., Boskey, A. L. and Frenkel, B. (2003) Bone morphogenetic protein-2 restores mineralization in glucocorticoid-inhibited MC3T3-E1 osteoblast cultures. J. Bone Min. Res. 18, 1186–97. [9] Bonewald, L. F., Harris, S. E., Rosser, J. et al. (2003) Von Kossa staining alone is not sufficient to confirm that mineralization in vitro represents bone formation. Calcif. Tissue Int. 72, 537–47. [10] Ouyang, H, Sherman, P. J., Paschalis, E. P., Boskey, A. L. and Mendelsohn, R. (2004) Fourier transform infrared microscopic imaging: effects of estrogen and estrogen deficiency on fracture healing in rat femurs. Appl. Spectrosc. 58, 1–9. [11] Boskey, A. L. (1990) Bone mineral and matrix. Are they altered in osteoporosis? Orthop. Clin. North Amer. 21, 19–29. [12] Paschalis, E. P., Betts, F., DiCarlo, E., Mendelsohn, R. and Boskey, A. L. (1997) FTIR microspectroscopic analysis of human iliac crest biopsies from untreated osteoporotic bone. Calcif. Tissue Int. 61, 487–92. [13] Boskey, A. (2003) Mineral changes in osteopetrosis. Crit. Rev. Eukaryot. Gene Expr. 13, 109–16. [14] Gadeleta, S. J., Boskey, A. L., Paschalis, E. et al. (2000) A physical, chemical, and mechanical study of lumbar vertebrae from normal, ovariectomized, and nandrolone decanoate-treated cynomolgus monkeys (Macaca fascicularis). Bone 27, 541–50. [15] Paschalis, E. P., Boskey, A. L., Kassem, M. and Eriksen, E.F. (2003) Effect of hormone replacement therapy on bone quality in early postmenopausal women. J. Bone Min. Res. 18, 955–9. [16] Paschalis, E. P., Burr, D. B., Mendelsohn, R., Hock, J. M. and Boskey, A. L. (2003) Bone mineral and collagen quality in humeri of ovariectomized cynomologus monkeys given rhPTH(1-34) for 18 months. J. Bone Min. Res. 18, 769–75. [17] Bohic, S., Rey, C., Legrand, A. et al. (2000) Characterization of the trabecular rat bone mineral: effect of ovariectomy and bisphosphonate treatment. Bone 26, 341–8.

258

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

[18] Aparicio, S., Doty, S. B., Camacho, N. P. et al. (2002) Optimal methods for processing mineralized tissues for Fourier transform infrared microspectroscopy. Calcif. Tissue Int. 70 422–9. [19] Ducy, P., Desbois, C., Boyce, B. et al. (1996) Increased bone formation in osteocalcindeficient mice. Nature 382, 448–52. [20] Boskey, A. L., Gadaleta, S., Gundberg, C., Doty, S. B., Ducy, P. and Karsenty, G. (1998) Fourier transform infrared microspectroscopic analysis of bones of osteocalcin-deficient mice provides insight into the function of osteocalcin. Bone 23, 187–96. [21] Shapses, S. A., Cifuentes, M., Spevak, L. et al. (2003) Osteopontin facilitates bone resorption, decreasing bone mineral crystallinity and content during calcium deficiency. Calcif. Tissue Int. 73, 86–92. [22] Delany, A. M., Amling, M., Priemel, M., Howe, C., Baron, R. and Canalis, E. (2000) Osteopenia and decreased bone formation in osteonectin-deficient mice. J. Clin. Invest. 105, 915–23. [23] Camacho, N. P., Landis, W. J. and Boskey, A. L. (1996) Mineral changes in a mouse model of osteogenesis imperfecta detected by Fourier transform infrared microscopy. Connect. Tissue Res. 35, 259–65. [24] Ongpipattanakul, B., Francoeur, M. L. and Potts, R. O. (1994) Polymorphism in stratum corneum lipids. Biochim. Biophys. Acta 1190, 115–22. [25] Bommannan, D., Potts, R. O. and Guy, R. H. (1990) Examination of stratum corneum barrier function in vivo by infrared spectroscopy. J. Invest. Dermato. 95, 403–8. [26] Golden, G. M., Guzek, D. B., Harris, R. R., McKie, J. E. and Potts, R. O. (1986) Lipid thermotropic transitions in human stratum corneum. J. Invest. Dermatol. 86, 222–59. [27] Rerek, M. E., Chen, H.-C., Markovic, B., Garidel, P., Mendelsohn, R. and Moore, D. J. (2001) Phytosphingosine and sphingosine ceramide headgroup bonding structural insight through thermotropic hydrogen/deuterium exchange. J. Phys. Chem. B 105, 9355–62. [28] Moore, D. J., Rerek, M. E. and Mendelsohn, R. (1997) FTIR spectroscopy of the conformational order and phase behavior of ceramides. J. Phys. Chem. B 101, 8933–40. [29] Potts, R. O., Guzek, D. B., Harris, R. R. and McKie, J. E. (1985) A noninvasive, in vivo technique to quantitatively, measure water concentration of the stratum corneum using attenuated total-reflectance infrared spectroscopy. Arch. Dermatol. Res. 277, 489–95. [30] Mak, V. H. W., Potts, R. O. and Guy, R. H. (1990) Percutaneous penetration enhancement in vivo measured by attenuated total reflectance infrared spectroscopy. Pharm. Res. 7, 835–41. [31] Potts, R. O. and Francoeur, M. L. (1990) Lipid biophysics of water loss through the skin. Proc. Natl. Acad. Sci. USA 87, 3871–3. [32] Pirot, F., Kalia, Y. N., Stinchcomb, A. L., Keating, G., Bunge, A. and Guy, R. H. (1997) Characterization of the permeability barrier of human skin in vivo. Proc. Natl. Acad. Sci. USA 94, 1562–7. [33] McIntosh, L. M., Jackson, M., Mantsch, H. H., Stranc, M. F., Pilavdzic, D. and Crowson, A. N. (1999) Infrared spectra of basal cell carcinomas are distinct from non-tumor-bearing skin components. J. Invest. Dermatol. 112, 951–6. [34] McIntosh, L. M., Summers, R., Jackson, M. et al. (2001) Towards non-invasive screening of skin lesions by near-infrared spectroscopy. J. Invest. Dermatol. 116, 175–81. [35] Veiro, J. A. and Cummins, P. G. (1994) Imaging of skin epidermis from various origins using confocal laser scanning microscopy. Clin. Lab. Invest. 189, 16–22. [36] Hanson, K. M., Behne, M. J., Barry, N. P., Mauro, T. M., Gratton, E. and Clegg, R. M. (2002) Two-photon fluorescence lifetime imaging of the skin stratum corneum pH gradient. Biophys. J. 83, 1682–90. [37] Yu, B., Kim, K. H., So, P. T. C., Blankschtein, D. and Langer, R. (2003) Visualization of oleic acid-induced transdermal diffusion pathways using two-photon fluorescence microscopy. J. Invest. Dermatol. 120, 448–55. [38] Caspers, P. J., Lucassen, G. W., Carter, E. A., Bruining, H. A. and Puppels, G. J. (2001) In vivo confocal Raman microspectroscopy of the skin: noninvasive determination of molecular concentration profiles. J. Invest. Dermatol. 116, 434–41. [39] Caspers, P. J., Lucassen, G. W. and Puppels, G. J. (2003) Combined in vivo confocal Raman spectroscopy and confocal microscopy of human skin. Biophys. J. 85, 572–80.

IR IMAGING OF HARD AND SOFT TISSUES

259

[40] Caspers, P. J., Lucassen, G. W., Wolthuis, R., Carter, E.A., Bruining, H. A. and Puppels, G. J. (1998) In vitro and in vivo Raman spectroscopy of human skin. Biospectroscopy 4, S31–9. [41] Xiao, C., Flach, C. R., Marcott, C. and Mendelsohn, R. (2004) Uncertainties in depth determination and comparison of multivariate with univariate analysis in a laminated polymer and skin. Appl. Spectrosc. 58, 382–9. [42] Bouwstra, J. A., Honeywell-Nguyen, P. L., Gooris, G. S. and Pnec, M. (2003) Structure of the skin barrier and its modulation by vesicular formulations. Prog. Lipid Res. 42, 1–36. [43] Huber, M., Trattnig, S. and Lintner, F. (2000) Anatomy, biochemistry, and physiology of articular cartilage. Invest. Radiol. 35, 573–80. [44] Shepherd, D. E. and Seedhom, B. B. (1999) Thickness of human articular cartilage in joints of the lower limb. Ann. Rheum. Dis. 58, 27–34. [45] Goldring, M. B. (1997) The musculoskeletal system: articular cartilage. In Primer on the Rheumatic Diseases (J. H. Klippel, ed.), The Arthritis Foundation, Atlanta, pp. 14–18. [46] Aydelotte, M. B., Greenhill, R. R. and Kuettner, K. E. (1988) Differences between sub-populations of cultured bovine articular chondrocytes. II. Proteoglycan metabolism. Connect. Tissue. Res. 18, 223–34. [47] Aydelotte, M. B. and Kuettner, K. E. (1988) Differences between sub-populations of cultured bovine articular chondrocytes. I. Morphology and cartilage matrix production. Connect. Tissue Res. 18, 205–22. [48] Iannone, F. and Lapadula, G. (2003) The pathophysiology of osteoarthritis. Aging Clin. Exp. Res. 15, 364–72. [49] Malemud, C. J., Islam, N. and Haqqi, T. M. (2003) Pathophysiological mechanisms in osteoarthritis lead to novel therapeutic strategies. Cells Tissues Organs 174, 34–48. [50] Mort, J. S. and Billington, C. J. (2001) Articular cartilage and changes in arthritis: matrix degradation. Arthritis Res. 3, 337–41. [51] Jackson, D. W., Simon, T. M. and Aberman, H. M. (2001) The articular cartilage repair dilemma: symptomatic articular cartilage degradation (the impact in the new millennium). Clin. Orthop. 391S, S14–25. [52] Buckwalter, J. A. and Mow, V. C. (1994) Injuries to cartilage and meniscus: sports injuries to articular cartilage. In Orthopaedic Sports Medicine Principles and Practice. (J. C. DeLee and D. Drez, Jr., eds), W. B. Saunders Company, Philadelphia, PA, pp. 82–107. [53] Burstein, D., Bashir, A. and Gray, M. L. (2000) MRI techniques in early stages of cartilage disease. Invest. Radiol. 35, 622–38. [54] Burstein, D. and Gray, M. L. (2003) New MRI techniques for imaging cartilage. J. Bone Joint. Surg. 85-A (Suppl. 2), 70–7. [55] Lazarev, Y. A., Grishkovsky, B. A. and Khromova, T. B. (1985) Amide I band of IR spectrum and structure of collagen and related polypeptides. Biopolymers 24, 1449–78. [56] Lazarev, Y. A., Grishkovsky, B. A., Khromova, T. B., Lazareva, A. V. and Grechishko, V. S. (1992) Bound water in collagen-like triple helical structure. Biopolymers 32, 189–95. [57] George, A. and Veis, A. (1991) FTIRS in H2 O demonstrates that collagen monomers undergo a conformational transition prior to thermal self-assembly in vitro. Biochemistry 30, 2372–7. [58] Bychkov, S. M. and Kuz’mina, S. A. (1991) Study of eye proteoglycans by means of infrared spectroscopy. Biull. Eksp. Biol. Med. 111, 475–7. [59] Bychkov, S. M. and Kuz’mina, S. A. (1992) Study of tissue proteoglycans by means of infrared spectroscopy. Biull. Eksp. Biol. Med. 114, 246–9. [60] Cael, J. J., Isaac, D. H., Blackwell, J. and Koenig, J. L. (1976) Polarized infrared spectra of crystalline glycosaminoglycans. Carbohydr. Res. 50, 169–79. [61] Camacho, N. P., West, P., Torzilli, P. A. and Mendelsohn, R. (2001) FTIR microscopic imaging of collagen and proteoglycan in bovine cartilage. Biopolymers 62, 1–8. [62] Potter, K., Kidder, L. H., Levin, I. W., Lewis, E. N. and Spencer, R. G. (2001) Imaging of collagen and proteoglycan in cartilage sections using Fourier transform infrared spectral imaging. Arthritis Rheum. 44, 846–55. [63] Rieppo, J., Hyttinen, M. M., Jurvelin, J. S. and Helminen, H. J. (2004) Reference sample method reduces the error caused by variable cryosection thickness in Fourier transform infrared imaging. Appl. Spectrosc. 58, 137–40. [64] West, P. A., Bostrom, M. P., Torzilli, P. A. and Camacho, N. P. (2004) Fourier transform infrared spectral analysis of degenerative cartilage: an infrared fiber optic probe and imaging study. Appl. Spectrosc. 58, 376–81.

260

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

[65] Collins, D. H. (1949) The Pathology of Articular and Spinal Diseases, Edward Arnold and Co., London, pp. 76–9. [66] West, P. A., Torzilli, P. A., Chen, C., Lin, P. and Camacho, N. P. (2005) Fourier transform infrared imaging spectroscopy analysis of collagenase-induced cartilage degradation. J. Biomed. Opt. 10(1), 14015. [67] Jackson, M., Choo, L. P., Watson, P. H., Halliday, W. C. and Mantsch, H. H. (1995) Beware of connective tissue proteins: assignment and implications of collagen absorptions in infrared spectra of human tissues. Biochim. Biophys. Acta 1270, 1–6. [68] Mainil-Varlet, P., Aigner, T., Brittberg, M. et al. (2003) Histological assessment of cartilage repair: a report by the Histology Endpoint Committee of the International Cartilage Repair Society (ICRS). J. Bone Joint Surg. 85-A (Suppl. 2), 45–57.

12 Mid-infrared imaging applications in agricultural and food sciences Douglas L. Elmore, Carrie A. Lendon, Sean A. Smith and Chad L. Leverette

12.1 Introduction Materials commonly studied in agriculture and food science include natural commodities, such as corn, beef, cotton and soybeans; isolated components, such as starch, soy protein and corn oil; and final food products, such as snack foods, breads, sauces and candy. Nonfood materials such as paper, clothing, biodegradable polymers and animal feed are also of interest. Most agriculture and food materials (agri-food materials) are complex mixtures that are chemically and physically heterogeneous. The spatial distribution of these properties can have a desirable or undesirable effect on the final product. Desirable properties are often referred to as functionalities by agriculture and food scientists; and include such things as physical stability of starting materials; texture, consistency and flavor of processed foods; or strength and appearance of nonfood products. Homogeneous systems, such as cooking oil, exist in thermodynamic equilibrium and the properties of these systems are determined by their chemical composition.1 Heterogeneous systems are not in thermodynamic equilibrium. The properties of these systems are governed by both the chemical composition and the internal framework formed by the spatial arrangement of the individual chemical components present. The formation of steric, electrostatic and covalent forces between these individual components can have a dramatic effect on the properties of the product. Ice cream is a classic example, which is primarily composed of ice cream mix and air. The pure components of ice cream have different structural properties in the mixture than they do in their isolated form. Frozen ice cream has a certain consistency and texture, which is quite different from any of the individual frozen ingredients. Infrared spectroscopy provides a wealth of chemical and physical information for both pure compounds and mixtures,2–4 and the technique has been used extensively in agri-food sciences for both macroscopic5–11 and microscopic12–14 investigations. Infrared spectroscopic imaging (infrared imaging) is a form of spatially resolved vibrational spectroscopy.15–22 The chemical and physical information obtained by infrared imaging is complementary to the morphological information obtained by visible microscopy. The combination of these two methods allows scientists to gain deeper insight into agri-food materials.

Water

Lipids Proteins Water

2925 2870 2855 2930 2890

2150

1740

1650

1630

1460

νa (CH2 ) νs (CH2 ) νs (CH3 ) ν(CH)

(OH)comb

ν(C=O)

ν(C=O)

δ(OH)

νa (CO2− 3 ) Carbonates

Complex carbohydrates and amorphous or fully hydrated sugars

Water,c carbohydrate Protein Unsaturated lipids Lipids, proteins

3300 3250 3010 2960

ν(OH)b ν(NH) ν(=C−H) νa (CH3 )

Compound class

∼Frequencya (cm−1 ) The band has a more rounded shape than ν(NH). This band has a more rounded shape than ν(OH). This band can be used to monitor saturation in fats and oils. Large CH2 /CH3 intensity ratios suggest long alkyl chains and visa versa. CH stretching features for fatty compounds (and proteins) and carbohydrates are significantly different. Frequencies of the νa (CH2 ) and νs (CH2 ) bands can be used to qualitatively monitor conformational order. Peak maxima are observed near 2930 and 2890 cm−1 . CH stretching features for fatty compounds (and proteins) and carbohydrates are significantly different. More bands are observed for crystalline simple sugars, than complex, amorphous, or fully hydrated carbohydrates. This band is a combination of deformation and librational modes, and is a very broad band. The band appears in a region typically free of interference in many agri-food samples. In some cases, the band can be used to monitor water content. Carbonyl bands typically shift to lower frequency with increased hydrogen bonding. This is a great band for tracking fat and oil. Amide I (secondary amide). This band is sensitive to changes in secondary structure and overlaps the δ(OH) band for water. This is the fundamental deformation band. It overlaps the protein amide I band. This band can be used to monitor changes in water content in a variety of materials including meats, protein isolates, and starch. This is a very strong broad band. Carbonates are ubiquitous in agri-food products.

Comments

Useful bands for infrared imaging in agri-food materials2,23,24

Functional group

Table 12.1

700

(OH)libr

This band shifts depending on its complex state. Amide II (secondary amide). Shifts to higher frequency with increasing hydrogen bonding. Splitting occurs for orthorhombic crystalline structures. In some cases, this band can be used to identify crystalline regions of fatty acids, such as palmitic acid. This band can be used to confirm changes observed in the νa (CH3 ) and νs (CH3 ) band intensities. Lecithin contains phospholipids. This band can be used to monitor phospholipids in the presence of triglycerides. This is the peak maxima in the so-called C−O stretch region. This region contains highly coupled contributions from ν(C−O), ν(C−C), ν(C−O−C), δ(C−OH) and δ(CH) vibrations. Crystalline simple sugars often produce numerous sharp bands. Amorphous simple sugars, starch and cellulose produce broader, less resolved bands. This is a great band for monitoring cis-disubstituted alkene groups in unsaturated fat. This is a great band for monitoring trans-disubstituted alkene groups in unsaturated fat. This band is due to librational motions (restricted rotation), and is very broad. In some cases this band is useful for monitoring water content.

most prevalent form of lipids in food.

b Abbreviations: ν, stretch; δ, deformation; s, symmetric; a, asymmetric. c Water refers to condensed phase water. Lipids include triglycerides, fatty acids, glycolipids, phospholipids and steroids. Triglycerides are the

a Common frequency observed in agri-food materials.

Water

trans Fat

965

Carbohydrates

1000

ν(C−O)

δ(=C−H)trans

Phospholipids

1240

νa (PO2 )

cis Fat

Lipids

1375

δ(CH3 )

720

Lipids

1470

δ(CH2 )

δ(=C−H)cis

Fatty acid salts Proteins

1560 1550

νa (CO− 2) δ(NH)

264

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

12.2 Spatially resolved chemical and physical information Infrared imaging can provide both chemical and physical information about heterogeneous samples. For either purpose, intense bands are usually chosen to provide acceptable signal-to-noise ratios (SNRs) in the resulting images. A table of useful bands in agri-food studies is presented in Table 12.1.23 The table only contains medium to very strong bands, since these are most likely to provide acceptable SNRs in an imaging experiment. For a given functional group, the table identifies the approximate band frequency for most agri-food materials, the compound class of interest and relevant comments. The band frequency is provided as a practical convenience to help agriculture and food scientists identify useful bands for infrared imaging. For example, in theory, a carbonyl group may be encountered that absorbs anywhere from ∼1900 to 1550 cm−1 . However, most edible fats and oils contain high levels of triglycerides, which usually have a band maxima near 1740 cm−1 . Clearly, the more specific, commonly observed frequency of ∼1740 cm−1 is worth noting by a food scientist. For readers who are interested in a comprehensive consideration of infrared group frequencies, numerous excellent books are available.2,23,24 Infrared imaging can be used to determine spatial distribution of chemical species, and this is currently most common use. Applications exploiting this capability has been described in a large number of studies.25–30 The technique can also be used to effectively perform nondestructive separations by identifying image pixels corresponding to relatively pure components. If multiple pixels can be identified for this purpose, the extracted spectra can be averaged to improve the SNR. Since infrared spectroscopy also provides information about physical structure, infrared imaging can be used to determine spatial distribution of physical properties as well. Some of the properties include intermolecular and intramolecular order, hydrogen bonding, protein secondary structure, complexation and functional group orientation. Infrared spectroscopy has been used extensively to probe the physical structure of proteins, carbohydrates and fats in biological materials.4,31,32 Since agri-food materials are biological in origin, the infrared methodology applicable to biochemical and biophysical studies is often equally well suited for agri-food. It is important for the agriculture and food scientist to know that an enormous amount of literature exists in which infrared spectroscopy has been used to probe biologically relevant fatty compounds, proteins and carbohydrates. These biological studies can provide important background and experimental information relevant to infrared imaging studies of agri-food materials. Lipids are fats and oils. Fats exist as solids while oils exist as liquids at room temperature. Lipids are mainly composed of triglycerides, but may also contain steroids, phospholipids, fatty acids, monoglycerides, diglycerides and other amphipathic molecules with long alkyl chains. The intramolecular order of these molecules is largely determined by the conformational order of the alkyl chain. The conformational order is dependent on many factors such as chain length, temperature,

AGRICULTURE AND FOOD SCIENCE APPLICATIONS

265

A

B

1800

1600

1400

1200

1000

800

Figure 12.1 FTIR ATR spectrum of (A) anhydrous crystalline glucose and (B) hydrated amorphous glucose.

pressure and other components present in the mixture.33 A large increase in conformational order is associated with a liquid to solid phase transition for a lipid. Smaller changes in conformational order are associated with solid–solid phase transitions that are experienced by these types of compounds. The spatial variation of conformational order in both solid and liquid dispersions often plays an important role in the functionality of food systems. The ν(CH2 ) bands (3000–2800 cm−1 ) are qualitative indicators of conformational order in the alkyl chain.34 Specifically, the frequency of both the νa (CH2 ) and νs (CH2 ) vibrations shift to lower frequency as conformational order increases. These bands, which are very large, are ideal for infrared imaging. It is also worth noting that the splitting of the δ(CH2 ) band (∼1470 cm−1 ) is associated with an orthorhombic crystalline structure.35 Both simple and complex carbohydrates are found in agri-food materials. The simple carbohydrates (sugars) encountered are mainly sucrose, glucose, fructose, maltose and lactose. Sugars provide energy and the sensation of sweetness. A combination of different sugars is often used in processed foods, particularly in situations where the sugar’s sweetness is masked or hidden by other ingredients. Spatial distribution of these sugars and their physical states can affect product functionality. When carbohydrates are examined with infrared spectroscopy, large spectral differences are observed between amorphous and crystalline sugars. The Fourier transform infrared (FTIR) attenuated total reflectance (ATR) spectra of crystalline and amorphous glucose provide a good example. In Fig. 12.1 note the relatively sharp bands observed for crystalline anhydrous glucose (A) compared with the much broader bands of amorphous hydrated glucose (B). In contrast to sugars, only small spectral differences are observed between amorphous and crystalline forms of complex carbohydrates, such as starch.36

266

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

In many cases, the secondary structure of a protein may adopt an α-helix, a β-sheet or a random-coil conformation. The secondary structure of food proteins can have a dramatic effect on product functionality.37 The amide I band has been used extensively to determine protein conformation,31,32 and the methodology is applicable to infrared imaging studies. Specifically, the largest contributions near 1655, 1650, and 1625 cm−1 are usually attributed to α-helix, random coil, and β-sheet, respectively (noting in some cases that random coil, is the largest frequency band). Unfortunately, the contributions are so overlapping that quantification often requires a deconvolution procedure.38 Scientists can validate their infrared methodology with concurrent X-ray diffraction, NMR and/or circular dichroism measurements. Vibrations sensitive to changes in short-range order and degree of hydration are observed in the so-called C–O stretch region of starch.36 Most bands in this region are highly coupled and possess contributions from various ν(C–O), ν(C–C), δ(C–OH) and δ(CH) vibrations in the carbohydrate.2,24 Because of the highly coupled and overlapping nature of the absorbance bands in this region it is very difficult, to assign specific bands unambiguously. Nonetheless, several diagnostic bands do exist. The ∼1047 cm−1 band intensity increases as the starch becomes more ordered, while the ∼1022 cm−1 band intensity increases with amorphous content. The ∼995 cm−1 band has a large amount of δ(C–OH) character and is very sensitive to water content. The frequency of this band shifts ∼12 cm−1 as the water content increases from 0% to 50%. No additional shift is observed for water levels over 50%. All three bands appear as increasing or decreasing shoulders on others bands. For this reason, very high SNR measurements and deconvolution methods are required to extract order and hydration information in an infrared imaging study of starch-containing material. If the molecules in a sample have some degree of preferred orientation (anisotropy), it is possible to probe the orientation of the functional groups with polarized measurements. This can be particularly useful in infrared imaging studies of protein films.

12.3 Chemical infrared imaging of protein, carbohydrates and fat in agri-food mixtures Infrared imaging can be used for the non-destructive, spatially resolved determination of protein, carbohydrate, and fat in a mixture. The corresponding infrared images are often referred to as chemical images. Chemical imaging is extremely useful in agri-food studies for determining the spatial distribution of protein, carbohydrates, fat and water mixtures, and because of its importance, this subject will be addressed in some detail. In Fig. 12.2 ATR infrared spectra of starch (a complex carbohydrate) (A) soy protein (B), canola oil (mostly triglycerides) (C), and a 1 : 1 : 1 mathematical mixture of the three (D) are presented. The regions containing diagnostic bands appropriate for infrared imaging are identified with dashed boxes. Note how easily all three components may be identified in the mixture spectrum (D).

267

Carbohydrates

Proteins

Fats and oils

AGRICULTURE AND FOOD SCIENCE APPLICATIONS

A

B

C

D 1800

1600

1400

1200

1000

800

–1

cm

Figure 12.2 FTIR ATR spectrum of (A) starch, (B) soy protein, (C) canola oil and (D) a mathematical 1 : 1 : 1 mixture.

Since the ν(NH) band (∼3250 cm−1 ) for proteins is complicated by overlapping ν(OH) bands (∼3300 cm−1 ) from both water and carbohydrates, the amide I and II bands near 1650 cm−1 and 1550 cm−1 are usually chosen for functional group mapping. However, it should be noted that the δ(OH) band for water near 1630 cm−1 , overlaps the amide I band. Proteins and carbohydrates can vary significantly in water content depending on their environment. Four bands are attributed to condensed phase water ν(OH) ∼3300 cm−1 , (OH)comb ∼ 2075 cm−1 , δ(OH) ∼1635 cm−1 , (OH)libr ∼ 700 cm−1 . The two fundamental modes, ν(OH) and δ(OH), are typically used to monitor changes in water content. However, the often overlooked (OH)comb and (OH)libr bands can also serve this purpose. While the (OH)comb band is much weaker than the other bands, it is extremely broad. The large bandwidth and the fact that interfering bands are rarely present in the 2300–1900 cm−1 region for many agri-food materials can make the (OH)comb a desirable band for monitoring changes in water content. In some cases the (OH)libr band can be useful when ATR measurements are made due to the relatively high depth of penetration produced at ∼700 cm−1 , that is, the (OH)libr band is a high intensity band in ATR spectra. Since the ν(OH) for carbohydrates is complicated by overlapping ν(NH) bands from protein and the ν(OH) band from water, large bands in the C–O stretch region between ∼1250 cm−1 and ∼900 cm−1 are usually chosen for functional group mapping. The ν(CH2 ) and ν(CH3 ) bands for fats and oils overlap ν(CH2 ) and

268

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

ν(CH3 ) contributions from proteins and the ν(CH) contributions from carbohydrates. For this reason, the ν(C=O) is usually chosen for functional group mapping. In some cases, the strong C–O stretch band near 1080 cm−1 for triglycerides can also be useful.

12.4 Sampling A working knowledge of the currently available sampling techniques is vitally important to scientists who use infrared imaging. The most common technique used for agri-food studies are presented in Table 12.2. The most desirable sample for infrared imaging is an optically smooth, thin film (∼0.5–20 μm in thickness depending on the application), which is appropriate for transmission measurements.39,40 Unprepared agri-food samples are rarely encountered in this form. Two methods are commonly used to prepare samples for transmission measurements, cryogenic slicing (microtoming) or diamond-cell compression. Microtoming does not compromise spatial distribution of chemical species and is currently the most commonly used sample preparation method for infrared imaging. Sections are typically placed on conventional transmission substrates (e.g. BaF2 ). Diamond-cell compression works on any deformable material. However, the original spatial distribution of chemical species may be slightly altered as the sample is compressed between the two diamond windows. Also, note that hard, undeformable material (such as SiO2 ) can damage diamond windows. Researchers should probe the sample for hard material with a mortar and pestle before using the diamond compression cell. The diamond phonon absorptions compromise the 2550–1850 cm−1 region, but fortunately this region is rarely of interest in agri-food studies. ATR sampling is gaining popularity in infrared imaging.41–44 The method is fast, easy and often requires no sample preparation. Standard ATR issues including depth of penetration, the sample–substrate intimate contact requirement, and crystal properties should all be considered when using this technique.45,46 The original spatial distribution of the chemical species may be slightly altered when pressure is applied. Macro-ATR imaging accessories are commercially available, while microsystems have recently appeared in the literature. Chan and Kazarian have recently reported the first FTIR images, which were obtained with a conventional single bounce diamond-anvil ATR accessory.47 FTIR images were produced with 13 μm spatial resolution with no infrared microscope objectives. This diamondanvil ATR accessory is extremely versatile, easy to use and easy to clean. However, the ZnSe focusing lens in accessory must be carefully aligned. The resulting image is elongated and requires a compensation procedure. The authors of this chapter expect to see commercially available diamond anvil imaging accessories in the near future and believe that these accessories will greatly increase the popularity of infrared microscopic imaging in both agri-food research and in general. An embedding and polishing method can be used as an alternative to microtoming.48 In this case, a sample is placed in an uncured polymer resin, cured

AGRICULTURE AND FOOD SCIENCE APPLICATIONS

269

Table 12.2 Common sampling methods for mid-Infrared imaging of agri-food materials

Technique/sample preparation Transmission Thin film sample (little or no preparation required) Cryogenic sectioning (microtome)

Diamond compression cell

Reflection absorption Thin film sample (little or no preparation required)

ATR No sample preparation required

Comments

Samples are optically smooth with ∼0.5–20 μm thickness. This is the most desirable sample type. The physical properties are uncompromised. In some cases, optically smooth protein and starch films can be prepared (without cryogenic slicing). This is the most common sampling technique used for infrared imaging. Chemical integrity is maintained, but the technique can compromise physical properties. Samples are typically prepared on BaF2 or CaF2 substrates. Typically no sample preparation is required. Allows transmission measurements to be performed on samples that would normally produce significant scattering anomalies. Works on any deformable material. Can compromise physical properties. Original spatial distribution of chemical species is slightly altered (sample is compressed between two diamond plates). The diamond phonon absorption compromises the 2550–1850 cm−1 region; fortunately this region is often of little interest in agri-food studies. Hard material, such as SiO2 , can damage diamond windows. SiO2 and CO2− 3 are common in agri-food materials – test deformability of the sample with a mortar and pestle before using diamond compression cell to avoid window damage. Optically smooth samples with ∼0.3–10 μm thicknesses are placed on a gold covered surface or other high reflecting substrate. Similar to the transmission method except that the pathlength is twice as long. The physical properties are uncompromised. In some cases, optically smooth protein and starch films can be prepared (without cryogenic slicing). Typically no sample preparation required. Macro-ATR imaging accessories are commercially available, while microsystems have been described in the literature.42 ATR issues include depth of penetration, intimate contact requirement and crystal properties. The original spatial distribution of chemical species may be slightly altered when pressure is applied.

and polished to the level of optical smoothness. Specular reflectance spectra are then obtained from the polished surface at normal incidence. The reflectance spectra are converted to absorbance spectra using the Kramers–Kroenig transform. The embedding and polishing method enables infrared, Raman and brightfield imaging to be performed on the same sample. The technique can also be used for depth profiling by successively polishing away more sample or microtoming the sample laden resin.

270

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

Table 12.3 Continued

Technique/sample preparation Specular reflection Optically smooth sample (little or no preparation required)

Embedding and polishing

Emission Irregular-shaped material

Comments

This technique involves a reflectivity measurement, therefore, the Kramers–Kroenig transform is required to convert the resulting reflectance spectrum to a pseud-absorbance spectrum. This is a simple procedure for normal incidence measurements. Optically thick samples are desirable. Additional interfaces, such as air pockets, in the sample can induce reflection absorption bands – resulting in a spectrum of convolved reflection and reflection absorption features. In some cases, optically smooth protein and starch films can be prepared (without cryogenic slicing), The sample is embedded in an uncured polymer resin, cured, then polished to the point of optical smoothness. Specular reflectance issues may complicate the images. The technique can be used for depth profiling and offers an alternative for samples that are difficult to microtome. The technique produces samples that can be imaged with multiple techniques. Undeveloped in agri-food studies. May gain popularity in future.

The final sampling technique for consideration is emission. Little has been reported on the feasibility of this technique for infrared imaging of condensed phase materials, such as bulk polymers, biological, pharmaceutical or argri-food materials. As higher sensitivity imaging spectrometers appear, emission measurements may gain popularity.

12.5 Chemometrics The discipline of chemometrics can play an important role in the effective application of infrared imaging to agri-food materials. In a univariate approach, a researcher extracts a specific observable (e.g. frequency or band height) from a set of spectra as a series of scalar quantities. The process is then repeated for additional observables. In many cases, this approach is sufficient to meet the researcher’s needs.49 However, the task of extracting the most useful information possible from an infrared image dataset, which may contain thousands of spectra measured at hundreds or even thousands of individual frequencies, often requires a different strategy. It simply may not be possible for the researcher to manually examine and extract information from each of these spectra. The researcher may, therefore, choose to employ a chemometric method instead. Chemometric methods can reduce a large dataset

AGRICULTURE AND FOOD SCIENCE APPLICATIONS

271

to a convenient mathematical representation that can be used to quickly determine artifact and information content. This reduction proceeds by decomposing the collection of spectra into a much smaller set of spectral trend vectors in an N -dimensional space (where N is the number of spectral frequencies) that describe as much of the variation among the spectra as possible. The process of automating the extraction of these maximum-information vectors from a dataset is often referred to as the multivariate approach for obtaining information from a collection of measurements. Some multivariate methods useful in agri-food studies include cosine correlation analysis (CCA), principal component analysis (PCA), multivariate curve resolution (MCR), partial least squares (PLS) and artificial neural networks (ANNs). Cosine correlation analysis is a quick way to generate infrared images of spectral similarity as long as the dataset has reasonable signal to noise, has good spectral resolution of absorbance features from chemically different constituents and is free from significant baseline disturbances.49 In agri-food infrared imaging applications, CCA is often useful when trying to identify the distribution of constituents with significantly different fingerprint-region spectra, such as proteins and carbohydrates. For maximum (dis)similarity resolution, the CH/OH/NH stretching region (∼3600–2800 cm−1 ) should be excluded from CCA calculations, since the corresponding water, carbohydrate and protein bands overlap considerably in this region and will induce significant correlation. Principal component analysis is the workhorse of chemometrics; in fact, virtually all other multivariate analysis techniques employ PCA as part of their algorithm.50–55 Optimally, PCA will reduce a spectral image dataset to two much smaller matrices that describe the variation in identity and concentration of the individual constituents across the field of view of the imaged sample. Unfortunately, the method can by complicated by nonlinearities. The most pervasive nonlinearity that occurs in agri-food imaging applications is frequency shifting. For example, if the level of crystallinity of a carbohydrate constituent varies across an image, the frequency of certain ν(C–O) bands may shift significantly. As another example, proteins show significant shifts in the positions of their amide I and amide II peaks with changes in secondary structure and hydrogen bonding. Another method that can be used to quickly extract useful chemical information from an infrared image dataset is MCR.50–52,54,56–58 In some cases, this method can be used to obtain the concentration and absorbance spectra for each constituent in the original dataset. However, if the goal is not necessarily to resolve the constituents’ spectra, but rather to empirically classify them, a regression method may be more appropriate.53,59,60 The most prominent multivariate regression methods include PLS and ANNs. Partial least squares proceeds much like PCA, in that it extracts sets of scores and loadings that capture the most variation in the dataset. Unlike PCA, however, PLS relies on a set of constituent values to determine the weights of the image data prior to extracting relevant variation; therefore, PLS usually does a better job than PCA of rejecting artifacts from its model of the data. Unless a set of mixture spectra with known mixture concentrations is available to construct the PLS calibration,

272

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

the set of constituent values will most likely be integer class identifiers that are imposed on the pixel spectra based on a priori information about the constituents (this particular implementation of PLS is usually called discriminant-PLS). Class information can also be obtained using the methods discussed earlier. Finally, since PLS is an empirical calibration, the model is not always interpretable in a classical spectroscopic sense, depending on the influence of artifacts and the complexity of the relationship between the constituent types and their spectra. Artificial neural networks ‘learn’ the correct constituent classification model through iterative trial-and-error calculations to determine which frequencies in the data show the best ability to classify the pixels according to the constituent type. They can explore nonlinear as well as linear relationships among the frequencies by using, for example, squared and cross-product terms of the frequency data. However, because of the extreme nonlinearities that may be found in a neural net model, their results are often uninterpretable in a classical sense, even though they may predict quite well.

12.6 Applications Only a few studies have described agri-food applications, which employed multichannel infrared detection,56,61,62 while a larger number of studies have described applications that used single channel or point detection to obtain infrared maps or images.13,14,63–65 Undoubtedly, more and more multichannel detection will be used in the future as commercially available infrared imaging systems become more common in the agri-food laboratory. Multichannel detection systems are gaining popularity since they provide higher acquisition rates, higher sensitivity and the ability to obtain a line or global image in a single scan, a capability that is essential for time-dependent studies.18,20,66–70 However, since fruitful research has been described using both techniques, both single-point and multichannel infrared imaging will be considered in this section. In this section, we refer to both infrared microscopic imaging and infrared microscopic mapping as simply infrared imaging. If a study employed multichannel detection, it is noted in the text. Marcott et al.61 described the first agri-food application of infrared imaging using a multichannel detector. In this study, a Fourier transform infrared (FTIR) imaging spectrometer with a 64 × 64 MCT FPA was used to image wheat kernels that were cyromicrotomed to 8 μm. Spatial variation of the carbohydrate, protein and lipid components were apparent in the images. The aleurone cells, pericarp layer, endosperm and germ could be distinguished from one another based on changes in the protein/lipid absorbance ratio. Budevska et al.56 used an FTIR imaging spectrometer equipped with a multichannel 64 × 64 MCT to characterize mature corn kernels. A bright-field visible image of a corn kernel section and a corresponding pseudocolor infrared image (2000 cm−1 region) from this study are presented in Fig. 12.3(a) and (b), while univariate infrared images of the total carbohydrate (1026 cm−1 ), total protein

273

AGRICULTURE AND FOOD SCIENCE APPLICATIONS Endosperm (b)

60 Row detector number

(a)

Absorbance

40 30 20 10

0.4 0.2

60

60 50 Absorbance

40 30 20 10 30 40 50

1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0

1656 57.4

30 40 50

1.2 1.0 0.8 0.6

60

60 50 Absorbance

40 30 20 10 10 20 30 40 50 60 Column detector number

0.9 0.8 0.7 0.6 0.5 0.4 0.3

1737 2.9

Row detector number Row detector number

10

1.4

10 20

Row detector number

20

10 20 30 40 50 60 Column detector number

50

10 20

(e)

30

Subaleurone

60

(d)

40

1026 93.7

Aleurone (c)

50

3500 3000 2500 2000 1500 1000 Wavenumber

Figure 12.3 (a) A brightfield visible image of a corn kernel section and (b) a corresponding pseudocolor infrared image from the study by Budevska et al.56 are presented above the corresponding univariate images of (c) the total carbohydrate, (d) total protein and (e) total lipids and cutin (1737 cm−1 ). The spectral region used to produce each image is indicated in a corresponding single pixel spectrum to the right of the infrared images. Reproduced from Ref. 34 with permission, copyright 2003 Applied Spectroscopy.

274

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(1656 cm−1 ) and total lipids and cutin (1737 cm−1 ) are presented in Fig. 12.3(c)–(e), respectively. Some single-pixel spectra are also presented. Multivariate techniques were used to identify and extract spectra of pure components for a library search that returned good matches for zein and starch. The multivariate technique was able to resolve two different types of protein. An infrared image study of corn and oat-flour-based extrudates was reported by Cremer and Kaletunc.71 Spatial distributions of starch, protein and lipid were determined. Starch was ubiquitous over the entire cross-section. The protein distribution possessed the smallest degree of uniformity and was inversely correlated with the starch distribution. The lipid distribution was correlated with neither the starch nor the protein distribution. In 1998, Himmelsbach et al.63 used infrared imaging to correlate chemical composition with the morphological structure in cross-sections of flax stem tissue (Linum usitatissimum L). The group was able to spatially differentiate between waxes, pectate salts, cellulose, aromatics and acetylated polysaccharides. In a more recent study, this group investigated the effect of enzymatic retting of flax stem.64 Microtomed cross-sections were treated with a retting enzyme, and in some cases, exposed to a chelating agent. Separation of the bundles into single fibers was accompanied by chemical changes in the flax components. Spatially resolved chemical information facilitated a correlation between the chemical and physical changes that occurred as a result of the retting. The process of meat denaturation in the connective tissue and single muscle fibers of the bovine longissimus dorsi muscle has been studied by Kirschner et al.62 In this study the meat was aged for 7 days, heat treated at temperatures ranging from 45◦ C to 70◦ C and imaged with an FTIR imaging spectrometer. The spectrometer was equipped with a 64 × 64 MCT. Second-derivative spectra of bovine tissue that was heat treated at four different temperatures (raw, 45◦ C, 60◦ C and 70◦ C) are presented in Fig. 12.4. Band assignments were made and are indicated by black arrows. Each spectra was obtained by averaging 90 single-pixel spectra. The relative increase in both the 1695/1654 cm−1 and 1630/1654 cm−1 ratio indicates an increase in antiparallel β-sheet and a decrease in α-helical structure as the temperature increases. Images of the raw and heat-treated samples were generated using the intensity ratio of the 1630/1654 cm−1 bands, which measures the degree of denaturation. These images are presented in Fig. 12.5 along side the corresponding brightfield images. In the false color composite infrared images, the red and blue colors correspond to the highest and lowest levels of denaturation, respectively. The images show evidence of denaturation at 45◦ C, as the image begins to take on shades of yellow or green. As the temperature increases, not only does the amount of denaturation increase, but the fibers begin to shrink and the intermyofibrillar space increases. In terms of a denaturation gradient, the image at 45◦ C suggests that the denaturation is most prominent in the outer portions of the fibers, while at higher temperatures, the denaturation appears to be more homogeneous. To explain this observation, the authors suggested that denatured protein was expelled from the myofibers.

AGRICULTURE AND FOOD SCIENCE APPLICATIONS

1630 cm–1 b-sheet

1695 cm–1 b-sheet

Raw 45°c 60°c 70°c

1654 cm–1 a-helix

1750

1700

1650

1600

275

1550

1500

Wavenumber (cm–1 ) Figure 12.4 Normalized second-derivative spectra averaged from 90 single spectra of heat-treated bovine tissue presented in a study by Kirschner et al.62 From bottom to top, the temperatures are raw, 45◦ C, 60◦ C and 70◦ C, respectively. Reproduced from Ref. 40, with permission; copyright 2004 American Chemical Society.

Kirkwood et al.72 recently developed the first high-throughput system for rapid identification of pathogenic foodborne microorganisms using a rapid-scan FTIR imaging spectrometer equipped with an MCT FPA. FTIR images of over 80 isolates from various genera including Clostridium, Listeria, Salmonella, Escherichia, Staphylococcus and Shigella were recorded. Dendrograms were generated based on the spectral data in the region between 1350 cm−1 and 970 cm−1 by applying hierarchical cluster analysis using the Ward linkage together with Euclidean distance as the metric. Successful discrimination of the strains was demonstrated. The benefits of this methodology include rapid analyses, elimination of chemical reagents, reduction of analysis cost and potential for automation. Infrared microspectroscopy is used in the pulp and paper industry73 for compositional analysis of pulp, wet end and size press chemical components in paper contaminants, additives and cellulose, and inks. Wilkinson et al.74 studied ink–paper interactions to discern changing ink properties when applied to paper. In this work, the authors employed synchrotron-based reflectance infrared microspectroscopy as a rapid, direct and nondestructive analysis approach for the study of inks on paper. The authors of this chapter expect infrared imaging to gain popularity as a technique for paper analysis. Yu et al.13,65 reported two FTIR imaging studies of plant tissue in grain barley that employed a synchrotron light source. In the first study, this group demonstrated the capability of this technique to spatially resolve the chemical composition in plant tissue. In the second study, spectral bands for protein, lipid, lignin, total

276

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a)

60

12

50

11 1 0.9

40 30 20 10 10 20 30 40 50 60 (b)

60

12

50

11 1

40 30 20 10 10 20 30 40 50 60 (c)

60 50 40 30 20 10 10 20 30 40 50 60

(d)

0.8 0.7 0.6

60 50 40 30 20 10 10 20 30 40 50 60

0.9 0.8 0.7 0.6

12 11 1 0.9 0.8 0.7 0.6

12 11 1 0.9 0.8 0.7 0.6

Figure 12.5 The brightfield (left) and infrared images (right) of bovine tissue at four different temperatures: raw, 45◦ C, 60◦ C and 70◦ C (from top to bottom). The infrared images were generated using the intensity ratio of the 1630/1654 bands to track the changes in denaturation. Intensity bars next to the chemical images indicate the degree of denaturation, with red being high and blue being low. The areas colored black represent spectra that did not pass a quality test. Reproduced from Ref. 40, with permission; copyright 2004 American Chemical Society.

carbohydrates, nonstructural carbohydrates, hemicelluloses and cellulose were identified and used to create functional group images. Infrared and brightfield images were compared. The locations of the pericarp, seed coat, aleurone and endosperm were identified and correlated with the spatially resolved chemical information.

AGRICULTURE AND FOOD SCIENCE APPLICATIONS

277

Huffman et al.75 used multichannel infrared imaging to study the thermal gel to liquid crystalline phase transition, vesicle diffusion rates and bilayer packing of disteroylphosphatidylcholine-derived aqueous dispersions of multilamellar lipid vesicles. The study is biophysical in nature, but nonetheless relevant to agri-food aqueous dispersions. The reader should note that a large number of biological, medicinal, pharmaceutical and materials science studies provide useful information that is relevant to infrared imaging of agri-food materials. Some examples are therefore cited.25,26,43,46,76–84

12.7 Complementary imaging techniques Infrared microscopic imaging spectrometers are typically coupled with a traditional brightfield microscope. Corresponding infrared and visible microscopic images are absolutely essential. Additional imaging techniques can also complement the infrared technique. For example, scanning electron microscopy (SEM) is routinely used in agri-food studies to obtain images at both micrometer and submicrometer spatial resolutions.85 SEM analysis provides images with greater resolution than traditional brightfield microscopes, while also providing a wider range of magnification and an increased depth of field. SEM provides important information about the morphological and topographical properties of food and beverage ingredients, properties that can affect appearance, taste, texture or other functionalities of the final product. Sample preparation for SEM analysis includes vapor deposition of a metal overlayer. For samples not amenable to this procedure, environmental scanning electron microscopy (ESEM) can sometimes be used instead, albeit resulting in lower quality images.86,87 Transmission electron microscopy (TEM)85,88 can be used to probe the internal structure of a sample. Numerous strategies exist for using this technique.89,90 Atomic force microscopy (AFM) provides information about surface roughness and topography on the nanometer scale and is gaining popularity in agri-food studies.91 Infrared imaging can be used to investigate the spatial variation of polyatomic inorganic ions. Unfortunately, many of these ions have very intense, broad absorbances near 1000 cm−1 , such as sulfates, phosphates and silicates. To make matters worse, complex carbohydrates also have large absorbances in this region. However, when infrared imaging is combined with a spatially resolved energy dispersive X-ray spectroscopy (EDS), specific ions can be monitored.92 Energy-dispersive spectrometers can be coupled to scanning electron microscopes. Combining other molecular imaging techniques with infrared can provide additional insight into agri-food materials.93 Complementary methods include Raman, confocal laser scanning microscopy (CLSM), hyperspectral reflectance and nearinfrared imaging. The combination of Raman and infrared imaging provides many benefits,12,94 including complementary gross selection rules. Specifically, Raman activity requires a change in the polarizability of the molecule, while infrared activity requires a change in the dipole moment. For this reason, Raman is more sensitive to nonpolar bonds, while infrared is more sensitive to polar bonds. Highly conjugated

278

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

compounds such as lignin and carotenoids are present in food materials, and the carbon–carbon double bond vibrations present in the compounds produce intense Raman signals. In many cases, Raman is very convenient for imaging lignin and carotenoids in hydrated agri-food mixtures, while these same compounds can be very difficult to image using infrared under similar conditions. In Raman, the amide I protein band experiences little interference from the δ(OH) water band (a weak Raman scatterer). Conversely, in infrared, the amide I band is complicated by the δ(OH) water band (a strong infrared absorber). The amide II band is strong in the infrared spectrum, while almost undetected in the Raman. For lipid studies, C–C stretching modes provide insight into the conformational order of the alkyl chains. These bands are weak in the infrared and are, therefore, unsuitable for functional group imaging. Raman can more easily differentiate between amorphous glucose, starch and cellulose present in agri-food materials. Note that Raman analysis can suffer from relatively poor signal to noise, autofluorescence of natural systems and susceptibility of natural systems to laser damage.12 Confocal laser scanning microscopy is commonly used in food and agriculture science. Through monitoring either autofluorescence from certain biological components or sensitive site-specific biomarkers coupled to particular food ingredients, detailed three-dimensional chemical images can be produced for agri-food materials. CLSM has been utilized in applications ranging from monitoring the interactions between protein and lipids in fat spreads and cheeses to observing the competition of emulsifiers at the interfaces of water and lipid.95 Hyperspectral reflectance imaging, which is based on ultraviolet visible spectroscopy, has been used mostly to investigate food quality and safety issues.96 Polarized light microscopy (PLM) produces images based on the contrast generated by differences in native birefringence, and is therefore very sensitive to changes in crystallinity. The information provided by PLM in the study of starch-based materials is very complementary to infrared imaging, since the latter technique may only detect large changes in crystallinity for complex carbohydrates. Both traditional and automated PLM techniques are used in agri-food studies.97 Magnetic resonance imaging (MRI) is a relatively new imaging technique98–100 that is gaining popularity in the agri-food arena. This technique is capable of studying dynamic processes in food systems, such as fluid flow, processing, and moisture and fat migration.101 The technique is particularly sensitive to water. Near-infrared reflectance imaging is a popular technique for spatially monitoring chemical species in agricultural materials.102–115 The technique has been used to investigate maize kernels,116 sugar content in the flesh of melons117 and soluble solids in kiwifruit.118

12.8 Conclusions The spatial arrangement of chemical, physical and morphological structure in agrifood materials play an important role in product functionality. Infrared imaging

AGRICULTURE AND FOOD SCIENCE APPLICATIONS

279

is a powerful technique that provides spatially resolved chemical and physical structural information. Knowledge and understanding of relevant spectroscopic observables, sampling techniques, chemometric methods and complementary imaging techniques will enable agriculture and food scientists to more effectively apply this technique.

References [1] Walstra, P. (2003) Physical Chemistry of Foods, 1st edn, Marcel Dekker, New York. [2] Colthup, N., Daly, L. and Wiberley, S. (1990) Introduction to Infrared and Raman Spectroscopy, 3rd edn, Academic Press, Boston, MA. [3] Griffiths, P. R. and De Haseth, J. A. (1986) Fourier Transform Infrared Spectrometry. Chemical Analysis, John Wiley & Sons, New York. [4] Chalmers, J. and Griffiths, P. R. (eds) (2002) Handbook of Vibrational Spectroscopy, John Wiley & Sons, New York. [5] Cerna, M., Barros, A. S., Nunes, A. et al. (2003) Carbohydr. Polym. 51, 383–9. [6] Wilson, R. H. and Tapp, H. S. (1999) Trac-Trends Anal. Chem. 18, 85–93. [7] Vandevoort, F. R. (1992) Food Res. Int. 25, 397–403. [8] Hiroaki, I., Toyonori, N. and Eiji, T. (2002) IEEE Trans. Instrum. Meas. 51, 886–890. [9] Dahlberg, D. B., Lee, S. M., Wenger, S. J. and Vargo, J. A. (1997) Appl. Spectrosc. 51, 1118–24. [10] Jaaskelainen, A. S., Nuopponen, M., Axelsson, P., Tenhunen, M., Loija, M. and Vuorinen, T. (2003) J. Pulp Paper Sci. 29, 328–31. [11] Kumosinski, T. F. and Farrell, H. M. (1993) Trends Food Sci. Technol. 4, 169–175. [12] Thygesen, L. G., Lokke, M. M., Micklander, E. and Engelsen, S. B. (2003) Trends Food Sci. Technol. 14, 50–7. [13] Yu, P., McKinnin, J. J., Christensen, C. R. and Christensen, D. A. (2004) Agric. Food Chem. 52, 1484–94. [14] Wetzel, D. L., Eilert, A. J., Pietrzak, L. N., Miller, S. S. and Sweat, J. A. (1998) Cell. Mol. Biol. 44, 145–67. [15] Lewis, E. N., Treado, P. J., Reeder, R. C. et al. (1995) Anal. Chem. 67, 3377–81. [16] Kidder, L. H., Levin, I. W., Lewis, E. N., Kleiman, V. D. and Heilweil, E. J. (1997) Opt. Lett. 22, 742–4. [17] Bhargava, R. and Levin, I. W. (2001) Anal. Chem. 73, 5157–67. [18] Dumas, P., Jamin, N., Teillaud, J.-L., Miller, L. M. and Beccard, B. (2004) Faraday Discuss. 126, 289–302. [19] Wetzel, D. L. and LeVine, S. M. (1999) Science 285, 1224–5. [20] Elmore, D. L., Tsao, M. W., Frisk, S., Chase, D. B. and Rabolt, J. F. (2002) Appl. Spectrosc. 56, 145–9. [21] Pellerin, C., Snively, C. M., Chase, D. B. and Rabolt, J. F. (2004) Appl. Spectrosc. 58, 639–46. [22] Michaels, C. A., Stranick, S. J., Richter, L. J. and Cavanagh, R. R. (2000) J. Appl. Phys. 88, 4832–9. [23] Socrates, G. (2001) Infrared and Raman Characteristic Group Frequencies: Tables and Charts, 3rd edn, John Wiley & Sons, New York. [24] Lin-Vien, D., Colthup, N. B., Fately, W. G. and Grasselli, J. G. (1991) The Handbook of Infrared and Raman Characteristic Frequencies of Organic Molecules, Academic Press Limited, London. [25] Ouyang, H., Sherman, P. J., Paschalis, E. P., Boskey, A. L. and Mendelsohn, R. (2004) Appl. Spectrosc. 58, 1–9. [26] Marcott, C., Reeder, R. C., Paschalis, E. P., Tatakis, D. N., Boskey, A. L. and Mendelsohn, R. (1998) Cell. Mol. Biol. 44, 109–15. [27] Chalmers, J. M., Everall, N. J., Schaeberle, M. D. et al. (2002) Vib. Spectrosc. 30, 43–52.

280

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

[28] Marcott, C., Reeder, R. C., Sweat, J. A., Panzer, D. D. and Wetzel, D. L. (1999) Vib. Spectrosc. 19, 123–9. [29] Bhargava, R., Wang, S. Q. and Koenig, J. L. (2003) Adv. Polym. Sci., 163, 137–91. [30] Kidder, L. H., Colarusso, P., Stewart, S. A. et al. (1999) J. Biomed. Opt. 4, 7–13. [31] Mantsch, H. H. and Chapman, D. (1996) Infrared Spectroscopy of Biomolecules, WileyLiss, New York. [32] Gremlich, H. and Yan, B. (eds) (2000) Infrared and Raman Spectroscopy of Biological Materials, Marcel Dekker, New York. [33] Scheuing, D. (ed.) (1990) Fourier Transform Infrared Spectroscopy in Colloid and Interface Science, American Chemical Society, Washington, DC. [34] Snyder, R. G., Hsu, S. L. and Krimm, S. (1978) Spectrochim. Acta Part A-Mol. Biomol. Spectrosc. 34, 395–406. [35] Snyder, R. G. (1979) J. Chem. Phys. 71, 3229–35. [36] vanSoest, J. J. G., Tournois, H., deWit, D. and Vliegenthart, J. F. G. (1995) Carbohydr. Res. 279, 201–14. [37] Yada, R., Jackman , R. and Smith, J. (eds) (1994) Protein Structure – Function Relationships in Foods, Kluwer Academic Publishers, Boston, MA. [38] Yang, W. J., Griffiths, P. R., Byler, D. M. and Susi, H. (1985) Appl. Spectrosc. 39, 282–7. [39] Messerschmidt, R. and Harthcock, M. (1988) Infrared Microscopy, Marcel Dekker, New York. [40] Katon, J. E. and Sommer, A. J. (1992) Anal. Chem. 64, A931–40. [41] Sommer, A. J., Tisinger, L. G., Marcott, C. and Story, G. M. (2001) Appl. Spectrosc. 55, 252–6. [42] Chan, K. L. A. and Kazarian, S. G. (2003) Appl. Spectrosc. 57, 381–9. [43] Chan, K. L. A., Hammond, S. V. and Kazarian, S. G. (2003) Anal. Chem. 75, 2140–6. [44] Reich, G. (2002) Pharmazeut. Indust. 64, 870–4. [45] Sommer, A. J., Tisinger, L. G., Marcott, C. and Story, G. M. (2001) Appl. Spectrosc. 55, 252–6. [46] Lewis, L. L. and Sommer, A. J. (2000) Appl. Spectrosc. 54, 324–30. [47] Chan, K. L. A. and Kazarian, S. G. (2003) Appl. Spectrosc. 57, 381–9. [48] Elmore, D. L., Leverette, C., Lendon, C., Smith, S., Anderson, B. and Muroski, A. (2004) Mid-infrared spectroscopic imaging: from thin film to food systems, 50th Gordon Research Conference on Vibrational Spectroscopy, Bristol, RI. [49] Harthcock, M. A. and Atkin, S. C. (1988) Appl. Spectrosc. 42, 449–55. [50] Widjaja, E., Crane, N., Chen, T., Morris, M., Ignelzi, M. and McCreadie, B. (2003) Appl. Spectrosc. 57, 1353–62. [51] Lavine, B., Davidson, C., Ritter, J., Westover, D. and Hancewicz, T. M. (2004) Microchem. J. 76, 173–80. [52] de Juan, A., Tauler, R., Dyson, R., Marcolli, C., Rault, M. and Maeder, M. (2004) Trends Anal. Chem. 23, 70–9. [53] Winson, M., Goodacre, R., Timmins, E. et al. (1997) Anal. Chim. Acta 348, 273–82. [54] Budevska, B. O. (2000) Vib. Spectrosc. 24, 37–45. [55] Windig, W., Antalek, B., Lippert, J. L., Batonneau, Y. and Bremard, C. (2002) Anal. Chem. 74, 1371–9. [56] Budevska, B. O., Sum, S. T. and Jones, T. J. (2003) Appl. Spectrosc. 57, 124–31. [57] Pudney, P., Hancewicz, T., Cunningham, D. and Gray, C. (2003) Food Hydrocolloids 17, 345–53. [58] Wan, J., Hopke, P., Hancewicz, T. M. and Zhang, S. (2003) Anal. Chim. Actaf 476, 93–109. [59] van den Broek, W., Derks, E., vande Ven, E., Wienke, D., Geladi, P. and Buydens, L. (1996) Chemometr. Intell. Lab. Syst. 35, 187–97. [60] Goodacre, R. (2003) Vib. Spectrosc. 32, 33–45. [61] Marcott, C., Reeder, R. C., Sweat, J. A., Panzer, D. D. and Wetzel, D. L. (1999) Vib. Spectrosc. 19, 123–9. [62] Kirschner, C., Ofstad, R., Skarpeid, H.-J., Host, V. and Kohler, A. (2004) J. Agric. Food Chem. 52, 3920–9. [63] Himmelsbach, D. S., Khalili, S. and Akin, D. E. (1998) Cell. Mol. Biol. 44, 99–108. [64] Himmelsbach, D. S., Khalili, S. and Akin, D. E. (2002) J. Sci. Food Agric. 82, 685–96.

AGRICULTURE AND FOOD SCIENCE APPLICATIONS

281

[65] Yu, P., McKinnon, J. J., Christensen, C. R., Christensen, D. A., Marinkovic, N. S. and Miller, L. M. (2003) J. Agric. Food Chem. 51, 6062–7. [66] Bhargava, R., Wall, B. G. and Koenig, J. L. (2000) Appl. Spectrosc. 54, 470–9. [67] Snively, C. M., Katzenberger, S., Oskarsdottir, G. and Lauterbach, J. (1999) J. Opt. Lett., 24, 1841–3. [68] Huffman, S. W., Bhargava, R. and Levin, I. W. (2002) Appl. Spectrosc.,56, 965–9. [69] Bhargava, R. and Levin, I. W. (2003) Appl. Spectrosc., 57, 357–66. [70] Snively, C. M., Pellerin, C., Rabolt, J. F. and Chase, D. B. (2004) Anal. Chem., 76, 1811–16. [71] Cremer, D. R. and Kaletunc, G. (2003) Carbohydr. Polym. 52, 53–65. [72] Kirkwood, J., Ismail, A., Gour, L. et al. (2004) Developmaent of the high-throughput system for rapid identification of microorganisms based on rapid-scan focal plane array Fourier transform infrared (FPA-FTIR) spectroscopy. International Food Technologies Annual Meeting, Las Vegas, NV. [73] Conners, T. and Banerjee, S. (eds) (1995) Surface Analysis of Paper, CRC Press, London. [74] Wilkinson, T. J., Perry, D. L., Martin, M. C., McKinney, W. R. and Cantu, A. A. (2002) Appl. Spectrosc. 56, 800–3. [75] Huffman, S. W., Schlucker, S. and Levin, I. W. (2004) Chem. Phys. Lipids, 130, 167–74. [76] Camacho, N. P., Carroll, P. and Raggio, C. L. (2003) Calcif. Tissue Int. 72, 604–9. [77] Potter, K., Kidder, L. H., Levin, I. W., Lewis, E. N. and Spencer, R. G. S. (2001) Arthritis Rheumatism, 44, 846–55. [78] Masaki, T., Goto, K., Inouye, Y. and Kawata, S. (2004) J. Appl. Phys. 95, 334–8. [79] Lasch, P., Pacifico, A. and Diem, M. (2002) Biopolymers (Biospectroscopy), 67, 335–8. [80] Coutts-Lendon, C. A., Wright, N. A., Mieso, E. V. and Koenig, J. L. (2003) J. Control. Release, 93, 223–48. [81] Elmore, D. L., Chase, D. B., Liu, Y. J. and Rabolt, J. F. (2004) Vib. Spectrosc. 34, 37–45. [82] Elmore, D. L., Leverette, C. L., Chase, D. B., Kalambur, A. T., Liu, Y. J. and Rabolt, J. F. (2003) Langmuir, 19, 3519–24. [83] Bhargava, R., Wang, S. Q. and Koenig, J. L. (1998) Appl. Spectrosc. 52, 323–8. [84] Bhargava, R. and Levin, I. W. (2004) Vib. Spectrosc. 34, 13–24. [85] Kalab, M. (1983) Physical Properties of Foods (M. Peleg and E. B. Bagley, eds), AVI Publishing Company, Incorporated, Westport, CT, pp. 43–104. [86] Stokes, D. J. (2003) Phil. Trans. Roy. Soc. London Ser. A-Math. Phys. Eng. Sci. 361, 2771–87. [87] Stokes, D. J. and Donald, A. M. (2000) J. Mater. Sci. 35, 599–607. [88] Yada, R. Y., Harauz, G., Marcone, M. F., Beniac, D. R. and Ottensmeyer, F. P. (1995) Trends Food Sci. Technol. 6, 265–70. [89] Steinbrecht, R. A. and Zierold, K. (1987) Cryotechniques in Biological Electron Microscopy, Springer-Verlag, Berlin, Heidelberg. [90] Revel, J.-P., Barnard, T. and Haggis, G. F. (1984) The Science of Biological Specimen Preparation for Microscopy and Microanalysis, Scanning Electron Microscopy, Inc., AMF O’Hare, IL. [91] Kirby, A. R., Gunning, A. P. and Morris, V. J. (1995) Trends Food Sci. Technol. 6, 359–65. [92] Charbonneau, J. E. (2001) Scanning, 23, 51–7. [93] Scotter, C. N. G. (1997) Trends Food Sci. Technol. 8, 285–92. [94] Pudney, P. D. A., Hancewicz, T. M. and Cunningham, D. G. (2002) Spectrosc. Int. J. 16, 217–25. [95] Blonk, J. C. G. and Vanaalst, H. (1993) Food Res. Int. 26, 297–311. [96] Kim, M. S., Chen, Y. R. and Mehl, P. M. (2001) Trans. ASAE, 44, 721–9. [97] Oldenbourg, R. and Mei, G. (1995) J. Microsc. 180, 140–7. [98] Mariette, F. (2004) Co. R. Chim. 7, 221–32. [99] Schrader, G. W., Litchfield, J. B. and Schmidt, S. J. (1992) Food Technol. 46, 77–83. [100] Schmidt, S. J., Sun, X. Z. and Litchfield, J. B. (1996) Crit. Rev. Food Sci. Nutrition, 36, 357–85. [101] Simoneau, C., McCarthy, M. J. and German, J. B. (1993) Food Res. Int. 26, 387–98. [102] Lewis, E. N. and Levin, I. W. (1993) Biophys. J. 64, A108. [103] Lewis, E. N., Treado, P. J. and Levin, I. W. (1994) Amer. Lab. 26, 16–21. [104] Treado, P. J., Levin, I. W. and Lewis, E. N. (1994) Appl. Spectrosc. 48, 607–15. [105] Mehl, P. M., Chen, Y. R., Kim, M. S. and Chan, D. E. (2004) J. Food Eng. 61, 67–81.

282

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

[106] Borregaard, T., Nielsen, H., Norgaard, L. and Have, H. (2000) J. Agric. Eng. Res. 75, 389–400. [107] Cheng, X., Tao, Y., Chen, Y. R. and Luo, Y. (2003) Trans. ASAE, 46, 551–8. [108] Kim, M. S., Lefcourt, A. M., Chao, K., Chen, Y. R., Kim, I. and Chan, D. E. (2002) Trans. ASAE, 45, 2027–37. [109] Lu, R. F. (2004) Postharvest Biol. Technol. 31, 147–57. [110] Park, B., Lawrence, K. C., Windham, W. R. and Buhr, R. J. (2002) Trans. ASAE, 45, 2017–26. [111] Park, B., Chen, Y. R. and Huffman, R. W. (1996) J. Food Eng. 30, 197–207. [112] Sugiyama, J. and Ogawa, Y. (2001) J. Japan. Soc. Food Sci. Technol.-Nippon Shokuhin Kagaku Kogaku Kaishi, 48, 263–7. [113] Sugiyama, J. (1999) J. Agric. Food Chem. 47, 2715–18. [114] Tran, C. D. and Grishko, V. I. (2004) Microchem. J. 76, 91–4. [115] Wen, Z. and Tao, Y. (2000) Trans. ASAE, 43, 449–52. [116] Cogdill, R. P., Hurburgh, C. R. and Rippke, G. R. (2004) Trans. ASAE, 47, 311–20. [117] Tsuta, M., Sugiyama, J. and Sagara, Y. (2002) J. Agric. Food Chem. 50, 48–52. [118] Martinsen, P. and Schaare, P. (1998) Postharvest Biol. Technol. 14, 271–81.

13 Applications of near-infrared imaging for monitoring agricultural food and feed products Vincent Baeten and Pierre Dardenne

13.1 Introduction The use of near-infrared spectroscopy (NIRS) for the analysis of agro-food products began in the 1970s with Karl Norris’ pioneering work demonstrating the high potential of this approach.1,2 NIRS techniques are now regarded as attractive and promising analytical tools for use in research, control and industrial laboratories. Analysts increasingly consider them as future technologies for food and feed analysis. This trend stems from the extensive use of computers as well as developments in instrumentation, and in appropriate chemometric procedures. New applications of spectroscopic techniques in chemical, pharmaceutical, life science, agro-food and environmental analysis are now emerging with great frequency. The evolution of instrumentation in infrared spectroscopy has been particularly rapid in process analytical chemistry, a crucial aspect of the pharmaceutical, chemical and agrofood industries. Analysis is moving closer to the sampling point or the field by means of fibre optics, allowing real-time analysis and continuous control of processes or embedded instruments. There is also a trend towards developing analytical chemistry systems combining the instrument, the interface between the instrument and the sample (i.e. sample presentation device), and the software integrating data acquisition, data archiving and chemometrics for specific applications. A more recent development is the use of the spectroscopic imaging instruments, based on multichannel infrared detectors, for the control and monitoring of agricultural food and feed products. For analysts and researchers, this is something of a revolution with hundreds or thousands of spectra being collected for each sample, process or experience, instead of one spectrum using classical NIRS instruments. The challenge is to extract and exploit the relevant information contained in the huge amount of data now available.3 While earlier chapters in the book discussed the instrumentation and applications of NIRS imaging, this chapter looks at how this technique can meet challenges in agricultural food and feed product analysis. A wide range of applications for agricultural purposes have been developed, including satellite and aircraft remote-sensing, macroscopic imaging for food quality

284

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

control, microscopic imaging for feed control and the study of plant physiology.4 Taylor and McClure in Ref. 5 demonstrated the great potential of NIRS imaging for measuring the distribution of chemical components. They worked with a CCD video camera with an effective sensitivity of 400–1100 nm and having narrow-band interference filters, and used the combination of spectral image data recorded at the different wavelengths to measure the distribution of chlorophyll and moisture in the plant. More recently, using an AOTF-based near-infrared (NIR) multispectral imaging instrument, Tran and Grishko in Ref. 6 demonstrated the potential of this technique for determining the water content of olive leaves. Two points need to be made here about the following discussion. First, most of the studies featured in this chapter concern the use of a system that is simultaneously active in the visible (400–700 nm) and NIR (700–2500 nm) ranges in relation to various problems. Second, the focus is on a limited number of papers that provide a broad overview of the potential applications of this technique for the agricultural food and feed sectors. The chapter reviews the application of the NIRS imaging for: (1) the remote control and monitoring of agriculture; (2) the analysis of food products as part of food-chain quality and safety control; and (3) the control of compound feeds at the ingredient particles level, in compliance with animal feedstuff regulations. All the NIR images shown in the chapter were collected at Walloon Agricultural Research Centre with a MatrixNIRTM instrument (Spectral Dimensions Inc., Olney, USA). This instrument includes an InGaAs array detector (240 × 320 pixels), which is active in the 900–1700 nm range.

13.2 Use of NIR imaging for remote control and monitoring in agriculture 13.2.1 The problem Traditionally, remote-sensing images, which observe the light from the sun that is reflected by the Earth, have been divided into two main categories. The first category includes the satellite-based images that cover large areas, with relatively fixed time intervals and a low spatial resolution. The second category includes airborne-based images that usually cover small areas, with flexible flight schedules and high spatial resolutions.7 Remote sensing of the earth began with the Landsat 1 satellite – a multispectral spectroscopic instrument – launched by NASA in 1972.8 Other similar and well-known instruments were the SPOT (Satellite Pour l’Observation de la Terre) and AVHRR (Advanced Very High Resolution Radiometer) satellites. Multispectral techniques involve the use of a small number of spectral bands (generally <10) from the visible and NIR ranges. In multispectral instruments the optimal bands are for a specific analysis or monitoring. Multispectral systems are less expensive, produce smaller datasets and have a greater signal-to-noise ratio (S/N). However, they produce data only from a limited number of broadbands.

MONITORING AGRICULTURAL FOOD AND FEED PRODUCTS

285

The first airborne-based system, the Airborne Imaging System (AIS), became operational in 1983. AVIRIS (Airborne Visible and Infrared Imaging Spectrometer), based on the AIS concept and covering the electromagnetic spectrum from 400 to 2500 nm with narrow-bands, became operational in 1989.9 Another example of a hyperspectral airborne sensor is CASI (Compact Airborne Spectrographic Imager). Tools developed for geological science have greatly influenced the increasing interest in hyperspectral information for remote sensing and the rapid development of imaging spectrometry to meet agricultural challenges. Hyperspectral technologies produce contiguous spectra with tens or hundreds of bands with a narrow bandwidth in the same spectral range of the multispectral instruments. The resolution of a hyperspectral instrument is considered to be sufficient for the identification of most biological materials.10 Satellite and airborne remote-sensing data are subject to large distortions that could substantially reduce their usefulness. There are three main distortion factors that could interfere with the collection of spectroscopic imaging data from space or airborne instruments for agricultural applications9,10 : (1) (2)

(3)

the system distortions, including those resulting from sensor calibration, scanner construction and engine (e.g. aircraft) instability; The sun, an important factor to take into account in remote sensing using spectroscopic imaging instruments. The passive optical system and the atmosphere through which the energy passes, both from the sun to the earth’s surface and back to the instrument, interferes with the data collected. Atmospheric distortions include the effect of scattered dry air molecules (haze) and absorption by air molecules. The geometric distortions related to an unstable platform, a low acquisition altitude and a large field of view. Sophisticated algorithms have been developed to reduce the impact of these distortions on the collected image data.

In the 1990s, it was recognised that the reflectance data obtained with remote sensing were highly valuable for modelling biophysical parameters. The spectral reflectance data extracted from hyperspectral instruments, however, are more sensitive than those collected with multispectral instruments for detecting different parameters and have extended the field of application of spectroscopic remotesensing technologies. The scope of the application of remote sensing is broad, and image data are essential inputs in environmental and ecological research (examples of applications include observation of ecosystems, fluvial geomorphic features, plant species mapping, woody debris distribution, vegetation water content, natural resource management, forest mapping, detection of pollution and fisheries oceanography).11–13 Another important area for the application of remote-sensing technologies is agriculture. The following section outlines the use of hyperspectral data for the determination of vegetation indices (VIs), precision agriculture and characterisation of grassland canopy.

286

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

13.2.1.1 Determination of VIs In order to determine agricultural crop characteristics, the VIs are calculated using spectroscopic remote-sensing imaging data.14–16 To minimise the undesirable disturbances of differences in soil background and atmospheric conditions, there are combinations of the reflectance in different wavelength bands. The VIs used to quantify crop variables include the leaf area index (LAI), wet biomass (WBM), plant height (PLNHT) and grain yield (YLD) indices, all of which are affected by climate, soils, cultivars, cultural practices, management and technological inputs, and are regional in nature.17 LAI is known to be one of the first plant responses to stress.18 It is directly related to the exchange of energy, CO2 and mass from plant canopies,19 and is sensitive to leaf cell enlargement due to water deficit.20 LAI is used to compute evapotranspiration21 and to compare terrestrial ecosystems worldwide.22 The WBM and PLNTH indices are excellent indicators of crop growth conditions and potential yield. WBM is also a good indicator of leaf and crop moisture. All these VIs involve two wavelength bands. The hyperspectral imaging data may be crucial for providing additional information and improving the traditional VIs. Another important index is the red-edge index derived from hyperspectral data, which is increasingly being used to determine crop characteristics.15,16 The position of the red-edge is defined as the inflexion point (or maximum slope) of the redNIR slope. Its determination requires a large number of spectral measurements in narrow spectral bands in the 680–780 nm region. The position and shift of the rededge index combined with other parameters has been successfully used to determine crop nitrogen status or nitrogen deficiency and to monitor the growth of some crops, such as potatoes and sugar beet.23

13.2.1.2 Precision agriculture Spectroscopic imaging data acquired by aircraft, could also play an important role in the development of precision agriculture. These technologies have the potential to detect crop stress and diagnose its cause before a farmer is able to spot the problem with the naked eye. The aim of remote sensing in precision agriculture is to provide farmers with detailed information that can be used to tailor the application of water, pesticides or fertilisers to crop needs. Other aims include: verifying the effectiveness of variable-rate fertiliser applications, verifying the effectiveness of fungicide applications, quantifying loss due to accidental spray drift damage and monitoring physical damage due to insects, flooding, wind or hail.24 Spectroscopic remote-sensing imaging technologies used with geographical information systems (GIS) and global positioning systems (GPS) may provide tools that will enable farmers to maximise the economic and environmental benefits of precision agriculture.24 By carefully identifying in-field variability, farmers can find a balance between production maximisation and environmental stress reduction. It is important to mention that while spatial resolution is valuable in monitoring crop appearance, it is the spectral signature that reveals the most information about plant stress and health. Also, the hyperspectral technologies produce more data than multispectral

MONITORING AGRICULTURAL FOOD AND FEED PRODUCTS

287

technologies, allowing a farmer to determine whether the stress observed in a field is caused by water depletion, insect infestation, poor fertilisation or weed invasion.7 One of the most useful applications of spectroscopic remote-sensing imaging is the possibility of the early detection of infestation in crops, enabling eradication using local treatments. Late detection results not only in financial loss, but also requires aerial spraying and the use of chemicals that are harmful to the environment. Agricultural remote-sensing approaches using spectroscopic imaging technology have been proposed, for instance, for characterising and determining the severity of fungal diseases in wheat.25 The aim of this study was to discriminate between healthy and diseased areas in a spring wheat crop suffering from fungal infestation, and to determine the plant-cover damage level in the affected areas. The main advantage of this approach is that it is suitable for real-time use. The hyperspectral crop reflectance data used consisted of 164 bands in the 360–900 nm region. Another important area is the detection of invasive weeds – also called noxious weeds – using the spectroscopic imaging systems. This approach could offer the possibility of generating timely and accurate weed maps.26

13.2.1.3 Characterisation of grassland canopy The spectroscopic imaging data provided by airborne hyperspectral imagers could also be used to characterise grassland canopy.27,28 Grasslands are an important component of agricultural landscape in some European countries, such as Belgium, and under current regulations it is mandatory to monitor them. At regional level, grassland monitoring is closely linked to the knowledge of regional management systems, the inventory of forage production and the quality and control of agrienvironmental measures. The work conducted by Buffet and Oger28 sought to demonstrate that hyperspectral remote-sensing imaging helps crop management by providing a continuous spatial and temporal assessment of the parameters characterising the canopy structure of each grassland area, as well as its biochemical and biophysical properties. This study was based on spectral remote-sensing observations and field observations of representative meadows in Lorraine in southeast Belgium. This region is characterised by important grassland areas under a wide variety of management systems. The remote-sensing data were collected in September 2002 from two spectroradiometric imaging systems mounted on a Dormier 228 aircraft. Hyperspectral data were acquired using a CASI sensor working in the 400– 950 nm region and a SASI (Shortwave Airborne Spectrographic Imager) sensor working in the 850–2500 nm region, with a ground resolution of 2.5 × 2.5 m and 2 × 2 m, respectively. The spectral resolution of the CASI and SASI sensors was 6 and 10 nm, respectively. Figure 13.1 presents the spectral features observed for the grasslands not harvested, grasslands just harvested and bare soil. The study showed that relationships exist between physicochemical parameters and hyperspectral data, enabling the quality of grassland canopy to be assessed and regional inventories of grass production potential to be drawn-up. The study also demonstrated the possibility of discriminating between different types of meadows (e.g. pasture and mowed

288 0.7

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

CASI sensor

SASI sensor

Bare soil

0.6

Not harvested Reflectance

0.5

Just harvested

0.4 0.3 0.2 0.1 0.0 400

600

800

1000

1200 1400 1600 1800 Wavelength (nm)

2000

2200

2400

2600

Figure 13.1 Spectral features of grasslands not harvested, grasslands just harvested and bare soil obtained with the CASI and SASI sensor. Courtesy of Dr Robert Oger and Ir Dominique Buffet, Walloon Agricultural Research Centre, Gembloux, Belgium.

meadows) and the potential of combining information from two sensors operating in different regions of the electromagnetic spectrum. The potential of spectroscopic imaging for mapping grass quality in Africa’s savanna range lands has been also investigated.29

13.3 NIR imaging for food analysis 13.3.1 The problem The search for new methodologies to improve the overall quality and safety of products in the food-chain is ongoing. Rapid and non-invasive methods that can be easily implemented to assess hazardous conditions in food production are required. Vibrational spectroscopy is one such method; it has been used successfully to evaluate the quality and safety of agro-food products. Its suitability for qualitative and quantitative analysis with little or no sample preparation, and its speed and high throughput, make it very attractive for the industrial sector. Traditional NIRS analyses in which one spectrum is obtained for each sample has the disadvantage of the results coming from an analysis (i.e. spectrum) of one small area or of several specimens of the sample. In contrast, the spectroscopic imaging method allows one to take account of the spatial variability of the samples analysed.4,30 The following section outlines the application of NIR imaging for analysing fruits, cereals and meat products.

13.3.1.1 Analysis of fruits The main quality attributes of fruits are appearance (colour, size, shape, lack of defects), texture and flavour (ripeness). Consumer satisfaction, however, relates

MONITORING AGRICULTURAL FOOD AND FEED PRODUCTS

289

mainly to texture and flavour. The technology for sorting fruits according to appearance exists, but the analytical challenge of developing methods to determine quality parameters, in a non-destructive way, has still to be met. It is necessary to develop a powerful tool to identify and detect spectral and spatial anomalies due to defects and/or contamination.31–33 Near-infrared spectroscopy imaging has been proposed as a means of assessing the external and internal quality of apples.34 Its potential has been demonstrated, for instance, in detecting external contamination in apples. Kim et al.30 used an imaging system working in the 430–930 nm region, with a spatial resolution of 1 mm, to detect fungal contamination in apples. This technology has since been applied to detect faecal contamination on the surface of fruit, using just two NIR wavelengths,33,35 and to detect soil contamination.35 These works showed that detecting faecal contamination using only information from the NIR region is less sensitive to variations in apple coloration. The detection is limited by the thickness of the faecal spot. Near-infrared imaging is also promising for detecting defects on the surfaces of apple varieties. Work has been also conducted on developing spectroscopic imaging systems to detect bruise defects and open skin cuts.30,32,35 The detection of bruises on apples is crucial for the industry – and for retailers in particular – as this defect reduces the quality of the fruit considerably, leading to financial losses. Lu32 developed an NIR hyperspectral imaging system covering the 900–1700 nm spectral region to detect bruises. He was able to detect bruises on apples, but he also showed that the spectral information included in the 1000–1340 nm region was not appropriate. Mehl et al.35 using an imaging system in the 430–900 nm range showed the potential of the method for detecting bruises in various cultivars. They concluded that the NIR bands were not subject to colour changes from the various apple cultivars studied and that there was no single waveband image that could be used easily to differentiate normal apples from bruised apples. They used the asymmetric second difference method to extract the relevant information from the spectra. NIRS imaging has also been investigated for analysing internal quality parameters. Studies have shown that the firmness and soluble solid content of apples can be predicted using selected NIR wavelengths (880, 905 and 940 nm).32,36 A method to measure starch distribution and the starch index of apples during maturation has been developed using NIRS imaging technology.37 The use of this technology to detect bitterpit in apples has been also proposed.38 Near-infrared spectroscopy imaging data has been used for assessing the internal characteristics of other fruits. Martinsen and Schaare39 were able to calibrate an imaging instrument for predicting the concentration of soluble solids in ripening kiwifruit using the NIR reflectance spectra in the 650–1100 nm range. The calibration models had a standard error of 1.2◦ Brix over a range of 4.7–14.1◦ Brix and were used to show the spatial distribution of soluble solids in cut sections of the fruit. Sugiyama40 and Tsuta et al.41 used NIRS imaging for evaluating the quality of melons, and studied the distribution of sugar in the fruit. Other applications include distinguishing between dark and berries in green grapes38 and measuring

290

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS Indexed factor scores (factor 3) (b)

i

PC 3

iii

100 150

PCA loading

Pixels

50

ii

PCA loading Raw spectrum

0.3 0.2

0

0.1

–20

0

–40

–0.1

–60

–0.2

–80

200 50

100

150 200 Pixels

250

20

Reflectance (counts)

(a)

–100 900 1000 1100 1200 1300 1400 1500 1600 1700 Wavelength (nm)

300

Figure 13.2 Analysis of raspberries by NIR imaging (third PC image) showing a grading in the maturity (Berries i = low maturity; ii = medium maturity; iii = ripe). (a) The figure shows the third PC image and (b) is a plot of the raw spectrum and the third PCA loading. From Walloon Agricultural Research Centre, Belgium.

(b) 50

50

100

100

Pixels

Pixels

(a)

150 200

150 200

50

100 150 200 250 300 Pixels

50

100 150 200 250 300 Pixels

Figure 13.3 Analysis of white currants by NIR imaging. NIR images at (a) 1340 nm and (b) 1410 nm, respectively. From Walloon Agricultural Research Centre, Belgium.

the ripeness of tomatoes.42 Figure 13.2 shows the PC 3 image obtained from the NIR spectroscopy image of raspberries with different grades of ripeness. Figure 13.3 present the NIR images at (a) 1340 nm and (b) 1410 nm of white currants. The fine external and internal structures of the berries can be observed.

13.3.1.2 Single-kernel analysis of cereals Among the most important ingredients in diets throughout the world are cereals such as wheat, rice and maize. Significant efforts need to be made to ensure that there is no contamination of cereals during production and storage. Such contamination includes adulteration by other cereal species or seeds from other crops, decayed and damaged grains (e.g. mouldy grains), animal faeces (e.g. from birds and rodents),

MONITORING AGRICULTURAL FOOD AND FEED PRODUCTS

291

seed contaminated with mycotoxins and insect infestations. Cereal contamination results in a loss in the quality, which in turn leads to financial losses for both the producers and the retailers if it is not detected at an early stage. Contamination rarely occurs at significant levels; methods that are able to detect defects at the grain level are needed. The target of analytical tools should be to detect one contaminated grain per kilogram. An important factor in developing methods for performing singlekernel analysis is their potential in the breed sector. It has been shown that genetic improvements can be significant when breeding selection is performed on a singlekernel basis.43 As a first step, classical NIRS instruments were adapted for single-seed analysis in order to meet the specific analytical requirements in detecting defect and contamination.44–46 Subsequently, work was carried out to demonstrate the potential of the NIRS imaging technique, which has the advantage of not being influenced by the heterogeneous morphology of products, such as maize, or by the distribution inside the grain of the parameter to be determined, since the entire kernel is analysed using imaging methodology. Ridgway and Chambers47 demonstrated the potential of NIRS imaging for detecting insects inside wheat kernels. They used a Hamamatsu NIR vidicon camera (C2400 series) and collected the reflectance images at wavelengths of 1202 nm and 1300 nm using appropriate filters. These wavelengths were selected on the basis of a previous study in which classical NIRS instruments were used.44 They used the subtracted image (1202–1300 nm) to detect grains internally infested with grain weevil larvae, based on the detection of changes in kernel composition. Later, they demonstrated the potential of NIRS imaging not only for detecting weevil larvae in kernels but also for detecting grains infested by insect pests (adult beetles and larvae in wheat) and ergot (parasitic fungi that form mycotoxins poisonous to both humans and animals) in wheat.48 This methodology can be used to detect infested kernels in other plant species. Figure 13.4 presents the analysis of single kernels by NIR imaging spectroscopy to detect insect infested grains. Figure 13.4(a) presents the image at 1400 nm of three infested wheat kernels. Figure 13.4(b) shows the fifth PC image of the NIR image of intact (1) and infested (2 & 3) coffee beans. Hurburgh and collaborators49–51 studied the potential of hyperspectral NIR techniques for analysing single-kernel maize. They developed an automated kernel analysis and sorting system to single-out maize kernels and place them in the field of view of an NIR spectrometer for analysis and classification. The imaging system used included an NIR camera active in the 700–1100 nm region and a liquid crystal tuneable filter to select individual wavelengths. They showed that it was possible to predict moisture and oil concentration in a single kernel with a standard error of cross-validation of 1.2% and 1.4%, respectively. One of their conclusions was that a major limitation of this technique is the poor quality of single-kernel reference chemistry. Similarly, a project aiming to use an NIR imaging system active in the 900–1700 nm region has been initiated at CRA-W to study the distribution of component in wheat kernel. Figure 13.5 shows the results of wheat grain analysis by NIR imaging. The figure displays the NIR image at 1140 nm spectra of the germ

292

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a)

1400 nm

(b)

1

PC 5

3

2

Figure 13.4 Analysis of single kernels by NIR imaging to detect insect infested grains. (a) The figure presents the image at 1400 nm of three infested wheat kernels. (b) The figure shows the fifth PC image of the NIR image of intact (1) and infested (2 and 3) coffee beans. From Walloon Agricultural Research Centre, Belgium.

1140 nm

PC 6 00 50

Reflectance (Counts)

0 –50 –100 –150 –200 –250 –300 900 1000 1100 1200 1300 1400 1500 1600 1700 Wavelength (nm)

Figure 13.5 Analysis of wheat grains by NIR imaging. NIR image at 1140 nm, spectra of the germ (dotted line) and albumen (continuous line), as well as the sixth PC image bringing to the fore the germ of each kernel. From Walloon Agricultural Research Centre, Belgium.

MONITORING AGRICULTURAL FOOD AND FEED PRODUCTS

0

Vegetal feed

Mammals’ meal

Poultry meal

Fish meal

293

1000

mm

2000

3000

4000

0

1000

2000

3000 4000 mm

5000

6000

Figure 13.6 Analysis of mixtures of vegetal feed, mammals’ meal, poultry meal and fish meal by NIR imaging. NIR image at 1380 nm. From Walloon Agricultural Research Centre, Belgium.

(dotted line) and albumen (continuous line), as well as the sixth PC image bringing to the fore the germ of each kernel.

13.3.1.3 Analysis of meat Spectrometric imaging has been also proposed for analysing meat products, particularly poultry carcasses. Currently, meat control (i.e. the detection of faecal and ingesta contamination) is performed by visual observation. The interest in using spectroscopic imaging is because it will be able to provide physical and chemical information of the products analysed. This information could include physical and geometric observations of size, orientation, shape, colour and texture, as well as chemical and molecular information (e.g. water, fat and protein content). Several papers have proposed using a VIS/NIR imaging system operating in the 400–900 nm region to detect faeces and ingesta contamination in poultry carcasses.52,53 They have shown the usefulness of this method and concluded that the impact of using different feed ingredients in the formulation of compound feed on the VIS/NIR spectra should be investigated. The selection of the most interesting wavelengths has been also investigated.54 A dual-wavelength spectral imaging system has been proposed for classifying poultry carcasses.52,55 Chao et al.55 proposed hyperspectral and multispectral systems for detecting chicken skin tumours. They used a spectroscopic imaging system active in the 420–850 nm region and a fuzzy inference system to distinguish tumours from normal skin tissue. The technique has also been used to characterise heart disease in chickens.56,57

294

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a)

0 A

B

C

D

500 1000 1500 mm

2000 2500 3000 3500 4000 4500 0

(b)

0

1000

2000

3000 mm

4000

A

B

C

D

5000

6000

5000

6000

500 1000 1500 mm

2000 2500 3000 3500 4000 4500 0

1000

2000

3000 4000 mm

Figure 13.7 Analysis of processed animal proteins (A = bovine and pig; B = poultry; C = feather; D = fish) by NIR imaging. (a) and (b) show the NIR images at 1210 and 1360 nm, respectively. From Walloon Agricultural Research Centre, Belgium.

13.4 NIR imaging for feed analysis 13.4.1 The problem With the emergence of the BSE crisis, first in Europe and later in other parts of the world, authorities have taken a lot of legal decisions in order to assure the human safety. One of them is the partial or total ban of the use of animal protein in compound feed. Originally, classical microscopy was the only method available for the fight against fraud or accidental contaminations of feedstuffs with meat

MONITORING AGRICULTURAL FOOD AND FEED PRODUCTS

295

Pixels

50 100 150 200 50

100

150 200 Pixels

250

300

Figure 13.8 Detection of feather meal in compound feed by NIR imaging. NIR image at 1480 nm. From Walloon Agricultural Research Centre, Belgium.

and bone meal (MBM).58 NIRS has also been proposed to solve this analytical challenge.59,60 With this technique, a single spectrum is obtained from the analysis of one sample (e.g. feed ingredient or compound feed). Recent developments have led to hyphenated instrument coupling an NIR spectrometer and a microscope (NIRM instrument). With this instrument, spectra of up to hundreds or thousands of particles can be obtained from the analysis of one feed ingredient or one compound feed. NIRM has proved to be an essential tool in the strategy aiming to tackle the detection of MBM in the frame of the BSE epidemic.61–63 The principal limitation of this technique is the sequential collection of the spectra (particle by particle) that has been solved by the introduction of the NIR imaging technology.3

13.4.1.1 Detection of MBM in feedstuffs Since 2001, a method based on NIR imaging is developed to detect animal ingredient particles in compound feeds.3,58,64 This NIR imaging system allows the analysis of about 400 particles in 5 min. Figure 13.6 presents the NIR image at 1380 nm of four materials from vegetal and animal origin (i.e. compound feed without MBM, mammals meal, poultry meal and fish meal). Each material used is a mixture of eight different individual samples. Figure 13.6 shows that the animal ingredients can be easily discriminated from the vegetal ones, courtsey of the images obtained with the NIR camera. The simultaneous analysis of hundreds or thousands of spectra using an NIR imaging system has the advantages of the speed and sensitivity that requires a screening method. The first results indicated that the NIR imaging method has a limit of detection of about 0.1% (depending on the number of particles analysed), and allows the discrimination between most of the fish and terrestrial particles. Combined with the recent chemometric method, SVM (support vector machines), used as classification algorithm, the NIR imaging method has proven to be promising

296

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

(a)

0 500

A

B

C

D

1000 1500 mm

2000 2500 3000 3500 4000 4500 0

(b)

0

1000

2000

3000 4000 mm

E

F

G

H

5000

6000

1000

mm

2000

3000

4000

0

1000

2000

3000 4000 mm

5000

6000

Figure 13.9 Analysis of processed vegetal meals (A and E = beet; B = rapeseed; C = linseed; D = wheat; F = lucerne; G = soya; H = corn) by NIR imaging. NIR images at 1480 nm. From Walloon Agricultural Research Centre, Belgium.

for the future.65 One of the main advantages of this NIR imaging methodology is that it is a non-destructive method and that it can be used in combination with classical microscopy or other techniques in order to get additional information about the particles analysed. Moreover, this method has been applied with success not only on the raw particles, but also on the particles coming from the sediment fraction. The NIR imaging methods have been tested on a wide diversity of compound feeds (with MBM in a range of 0.1–8% and also free of MBM) showing results

MONITORING AGRICULTURAL FOOD AND FEED PRODUCTS

297

with a level of false positives and false negatives inferior to 5%.66,67 NIR imaging could be also applied to give an indication of the origin of animal ingredients present in contaminated food. Figure 13.7 presents the NIR images at 1210 and 1360 nm of four materials of animal origin (bovine and pig, poultry, feather and fish meal). Figure 13.8 presents another important advantage of the NIR imaging method: in addition to spectroscopic information, the morphologic structure of the analysed particle can also be obtained. The figure is an NIR image at 1480 nm of about 400 particles of compound feed with feather meal particles. The feather meal particles (in white) are clearly visible.68

13.4.1.2 Detection of the vegetal source of feed ingredients Near-infrared imaging spectroscopy has also been applied for the complete screening of feedstuffs in order to detect and quantify all feed ingredients included in a compound feed. Figure 13.9 presents two NIR images at 1480 nm that demonstrates the potential of the method to discriminate different vegetable feed ingredients. Figure 13.9(a) includes beet, rape seed, linseed and wheat; while Fig. 13.9(b) includes beet, lucerne, soya and corn.66 13.5 Conclusion The application of NIR imaging to the remote control and monitoring of agriculture, the analysis of food products as part of the food-chain quality and safety control, as well as the control of compound feeds at the ingredient particles level has been briefly reviewed in this chapter. The benefits of NIR imaging technology for the monitoring of the agro-food products are obvious and will increase in the next few years. During the early stages, the main effort was concentrated on the exploration and development of new applications. NIR imaging techniques are now continuously evolving and have been introduced into newer fields. We assist in the development of fully automated systems for the control of agro-food products. The extraction of useful data, which contribute to the application and rejection of the voluminous data that cause confusion, remains the main challenge.

References [1] Williams, P. and Norris, K. (eds) (2001) Near-Infrared Technology in the Agricultural and Food Industries, 2nd edn, AACC, Minnesota. [2] Bertrand, D. and Dufour, E. (2000) La spectroscopie infrarouge et ses applications analytiques. Editions Tec & Doc, Paris, 566 pp. [3] Baeten, V. and Dardenne, P. (2002) Spectroscopy: developments in instrumentation and analysis. Grasas y Aceites, 53(1), 45–63. [4] Budevska, B. O., Sum, S. T. and Jones, T. J. (2003) Appl. Spectrosc. 57(2), 124–131. [5] Taylor, S. K. and McClure, W. F. (1989) NIR imaging spectroscopy: measuring the distribution of chemical components, Proceedings of the Second International NIRS Conference (M. Iwamoto and S. Kawano, eds), Tsukuba, Japan, pp. 393–404.

298

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

[6] Tran, C. D. and Grishko, V. I. (2004) Determination of water contents in leaves by a near-infrared multispectral imaging technique. Microchem. J. 76, 91–94. [7] Yao, H., Tian, L. and Noguchi, N. (2001) Hyperspectral imaging system optimization and image processing. ASAE Meeting Paper Number 01-1105, 30 July–1 August, Sacramento, CA. [8] Curran, P. J. (1989) Remote sensing of foliar chemistry. Remote Sensing Environ. 29, 271–8. [9] Vane, G. and Goetz, A. F. H. (1993) Terrestrial imaging spectrometry: current status, future trends. Remote Sensing Environm. 44, 117–26. [10] Jacobsen, A. (2000) Analysing airborne optical remote sensing data from a hyperspectral scanner and implications for environmental mapping and monitoring – results from a study of casi data and Danish semi-natural, dry grasslands. PhD Thesis. National Environmental Research Institute, Denmark, 148 pp. [11] Santos, A. M. P. (2000) Fisheries oceanography using satellite and airborne remote sensing methods: a review. Fish. Res. 49(1), 1–20. [12] Curran, P. J. (2001) Imaging spectrometry for ecological applications. JAG 3(4), 305–12. [13] Kerr, J. T. and Ostrovsky, M. (2003) From space to species: ecological applications for remote sensing. Trends Ecol. Evolution 18(6), 299–305. [14] Wessman, C. A., Bateson, C. A. and Benning, T. L. (1997) Detecting fire and grazing patterns in a tallgrass prairie using spectral mixture analysis. Ecol. Appl. 7, 493–511. [15] Clevers, J. G. P. W. (1999) The use of imaging spectrometry for agricultural applications. ISPRS J. Photogrammetry Remote Sensing 54, 299–304. [16] Clevers, J. P. G. W. and Jongschaap, R. (2001) Imaging spectrometry for agricultural applications In Imaging Spectrometry ( F. D. van der Meer and S. M. de Jong, eds), Kluwer Academic Publishers, Dordrecht, pp. 157–99. [17] Thenkabail, P. S., Smith, R. B. and De Pauw, E. (2000) Hyperspectral vegetation indices for determining agricultural crop characteristics. Remote Sensing Environ. 158–82. [18] Barlett, D. S., Hardisky, M. A., Johnson, R. W., Gross, M. F., Klemas, V. and Hartman, J. M. (1988) Continental scale variability in vegetation reflectance and its relationship to canopy morphology. Int. J. Remote Sensing 9, 1223–41. [19] Fassnacht, K. S., Gower, S. T., MacKenzie, M. D., Nordheim, E. V. and Lillesand, T. M. (1997) Estimating the leaf area index of north central Wisconsin forest using the landsat thematic mapper. Remote Sensing Environ. 61, 229–45. [20] Shibayama, M., Takahashi, W., Morinaga, S. and Akiyama, T. (1993) Canopy water deficit detection in paddy rice using high resolution field spectroradiometer. Remote Sensing Environ. 45, 117–26. [21] Clevers, J. G. P. W., Bü ker, C., van Leeuwen, H. J. C. and Bouman, B. A. M. (1994) A framework for monitoring crop growth by combining directional and spectral remote sensing information. Remote Sensing Environ. 50(2), 161–70. [22] Running, S. W. (1989) Estimating terrestrial primary productivity by combining remote sensing and ecosystem. In Remote Sensing of Biosphere Functioning (R. J. Hobbs and H. A. Mooney, eds), Springer-Verlag, New York. pp. 65–86. [23] Broge, N. H., Hvidberg, M., Hansen, B. U., Andersen, H. S. and Nielsen, A. A. (1997) Analysis of biophysical relatioships for a wheat canopy, Proceedings of the Third Airborne Remote Sensing Conference and Exhibition, Vol. 2, pp. 373–9. [24] Seelan, S. K., Laguette, S., Casady, G. M. and Seielstad, G. A. (2003) Remote sensing applications for precision agriculture: a learning community approach. Remote Sensing Environ. 88(1–2), 157–69. [25] Hamid Muhammed, H. and Larsolle, A. (2003) Feature-vector based analysis of hyperspectral crop reflectance data for discrimination quantification of fungal disease severity in wheat. Biosyst. Eng. 86(2), 125–34. [26] Lamb, D. W. and Brown, R. B. (2002) Remote-sensing and mapping of weeds in crops, J. Agric. Eng. Res. 78(2), 117–25. [27] Buffet, D. and Oger, R. (2003) Hyperspectral imagery for environmental mapping and monitoring: ‘Case Study of Grassland in Belgium’, geographical information systems and remote sensing: environmental applications, Proceedings of the International Symposium, 7–9 November, Volos, Greece.

MONITORING AGRICULTURAL FOOD AND FEED PRODUCTS

299

[28] Buffet, D. and Oger, R. (2003) Characterisation of grassland canopy using CASI-SASI hyperspectral imagery, Presented at the CASI-SWIR2002 Workshop, 4 September, Bruges. [29] Mutanga, O. and Skidmore, A. K. (2003) Integrating imaging spectroscopy and neural networks to map grass quality in the Kruger National Park, South Africa. Remote Sensing of Environ. 90, 104–15. [30] Kim, M. S., Chen, Y. R. and Mehl, P. M. (2001) Hyperspectral reflectance and fluorescence imaging system for food quality and safety, Trans. ASAE 44(3), 721–9. [31] Martinsen, P., Schaare, P. and Andrews, M. (1999) A versatile near infrared imaging spectrometer. J. Near Infrared Spectrosc. 7, 17–25. [32] Lu, R. (2003) Detection of bruises on apples using near-infrared hyperspectral imaging. Trans. ASAE 46(2), 523–30. [33] Kim, M. S., Lefcourt, A. M., Chao, K., Chen, Y. R., Kim, I. and Chan, D. E. (2002) Multispectral detection of fecal contamination on apples based on hyperspectral imagery : Part I. Application of visible and near-infrared reflectance imaging. Trans. ASAE 45(6), 2027–37. [34] Lu, R. and Chen, Y. R. (1998) Hyperspectral imaging for safety inspection of food and agricultural products, Proceedings of SPIE: Pathogen Detection and Remediation for Safe Eating, Vol. 3544, pp. 121–33. [35] Mehl, P. M., Chen, Y. R., Kim, M. S. and Chan, D. E. (2003) Development of hyperspectral imaging technique for the detection of apple surface defects and contaminations. J. Food Eng. 61(1), 67–81. [36] Lu, R. (2004) Multispectral imaging for predicting firmness and soluble solids content of apple fruit. Postharvest Biol. Technol. 31(2), 147–57. [37] Peirs, A., Scheerlinck, N., De Baerdemaeker, J. and Nicolaï, B. M. (2003) Starch index determination of apple fruit by means of a hyperspectral near infrared reflectance imaging system. J. Near Infrared Spectrosc. 11, 379–89. [38] Bellon-Maurel, V., Steinmetz, V., Guizard, C. and Flores Porras, L. (1999) Near infrared imaging: new hardware opens new opportunities, NIR-99 Ninth International Conference, Verona, Italy, pp. 159–63. [39] Martinsen, P. and Schaare, P. (1998) Measuring soluble solids distribution in kiwifruit using near-infrared imaging spectroscopy. Postharvest Biol. Technol. 14, 271–81. [40] Sugiyama, J. (1999) Visualization of sugar content in the flesh of a melon by near infrared imaging. J. Agric. Food Chem. 47, 2715–18. [41] Tsuta, M., Sugiyama, J. and Sagara, Y. (2002) Near-infrared imaging spectroscopy based on sugarabsorption band for melons. J. Agric. Food Chem. 50(1), 48–52. [42] Polder, G., van der Heijden, G. W. A. M. and Young, I. T. (2000) Hyperspectral image analysis for measuring ripeness of tomatoes, Presented at the 2000 ASAE International Meeting, Paper Number 003089, 9–12 July, Milwaukee, WI. [43] Silvela, L., Rodgers, R., Barrera, A. and Alexander, D. E. (1989) Effect of selection intensity and population size on percent oil in maize, Zea mays Theor. Appl. Genet. 78, 298–304. [44] Chambers, J. and Ridgway, C. (1996) Rapid detection of contaminants in cereals, Proceedings of the Seventh International Conference on Near-Infrared Spectroscopy (A. M. C. Davies and P. Williams, eds), 6–11 August 1995, Montréal, Canada, pp. 484–9. [45] Ridgway, C., Chambers, J. and Cowe, I. (1999) Detection of grain weevils inside single wheat kernels by a very near infrared two-wavelength model. J. Near Infrared Spectrosc. 7, 213–21. [46] Wang, D., Ram, M. S. and Dowell, F. E. (2002) Classification of damaged soybean seeds using near-infrared spectroscopy. Trans. ASAE 45(6), 1943–8. [47] Ridgway, C. and Chambers, J. (1998) Detection of insects inside wheat kernels by NIR imaging. J. Near Infrared Spectrosc. 6, 115–19. [48] Ridgway, C., Davies, R. and Chambers, J. (2001) Imaging for the high-speed detection of pest insects and other contaminants in cereal grain in transit, Presented at the 2001 ASAE International Meeting, Paper Number 013056, 30 July–1 August, Sacramento, CA. [49] Codgill, R. P., Hurburgh, C. R. Jr, Jensen, T. C. and Jones, R. W. (2002) Single-kernel maize analysis by near-infrared hyperspectral imaging, Proceedings of the 10th International Conference on Near-Infrared Spectroscopy, 2001 (A. M. C. Davies and R. K. Cho, eds), Korea, pp. 243–7.

300

SPECTROCHEMICAL ANALYSIS USING IR DETECTORS

[50] Stevermer, S. W., Steward, B. L., Codgill, R. P. and Hurburgh, C. R. (2003) Automated sorting and single kernel analysis by near-infrared hyperspectral imaging, Presented at the 2003 ASAE International Meeting, Paper Number 036159, 27–30 July, Las Vegas, NV. [51] Codgill, R. P., Hurburgh, C. R. Jr and Rippke, G. R. (2004). Single kernel maize analysis by near-infrared hyperspectral imaging. Trans. ASAE 47(1), 311–20. [52] Park, B., Windham, W. R., Lawrence, K. C., Buhr, R. J. and Smith, D. P. (2002) Imaging spectrometry for detecting feces and ingesta contamination on poultry carcasses, Proceedings of the 10th International Conference on Near-Infrared Spectroscopy (A. M. C. Davies and R. K. Cho, eds), Korea, 2001. pp. 457–61. [53] Lawrence, K. C., Windham, W. R., Park, B. and Buhr, R. J. (2003) A hyperspectral imaging system for identification of faecal and ingesta contamination on poutry carcasses. J. Near Infrared Spectrosc. 11, 269–81. [54] Windham, W. R., Park, B., Lawrence, K. C., Buhr, R. J. and Smith, D. P. (2002) Selection of visible/near infrared wavelengths for characterising fecal and ingesta contamination of poultry carcasses, Proceedings of the 10th International Conference on Near-Infrared Spectroscopy (A. M. C. Davies and R. K. Cho, eds), Korea, 2001. pp. 453–6. [55] Chao, K., Mehl, P. M. and Chen, Y. R. (2002) Use of hyper- and multi-spectral imaging for detection of chiken skin tumors. Appl. Eng. Agric. 18(1), 113–19. [56] Chao, K., Chen, Y. R., Hruschka, W. R. and Park, B. (2001) Chicken heart disease characterization by multi-spectral imaging. Appl. Eng. Agric. 17(1), 99–106. [57] Chao, K., Chen, Y. R., Hruschka, W. R. and Gwozdz, F. B. (2001) On-line inspection of poultry carcasses by a dual-camera system. J. Food Eng. 51(3), 185–92. [58] Gizzi, G., van Raamsdonck, L. W. D., Baeten, V. et al. (2003) An overview of tests for animal tissues in animal feeds used in the public health response against BSE. In Risk Analysis of Prion Diseases in Animals (C. I. Lasmézas and D. B. Adams, eds), Sci. Tech. Rev, 22(1), 311–31. [59] Garrido, A. and Fernández, V. (1998) NIRS technology for the quantiftative prediction of the percentage of meat and bone meal added to a feed compound. A feasibility study, Report of the Workshop Identification of Animal Ingredients in Compound Feeds, CEMA – DG VI – SMT Program, May, Lyngby, Denmark. [60] Murray, I, Aucott, L. S. and Pike, I. H. (2001) Use of discriminant analysis on visible and near infrared reflectance spectra to detect adulteration of fish meal with meat and bone meal. J. Near Infrared Spectrosc. 9, 297–311. [61] Piraux, F. and Dardenne, P. (1999) Feed authentication by near-infrared microscopy, Proceedings of the Ninth International Conference on Near-Infrared Spectroscopy (R. Giangiacomo, et al., eds), June, Verona, Italy. [62] Baeten, V., Michotte Renier, A., Sinnaeve, G. and Dardenne, P. (2001) Analyses of feedingstuffs by near-infrared microscopy (NIRM): detection and quantification of meat and bone meal (MBM), Proceedings of the Sixth International Symposium on Food Authenticity and Safety, 28–30 November 2000, Nantes, pp. 1–11. [63] Baeten, V. and Dardenne, P. (2001) The contribution of near infrared spectroscopy to the fight against the mad cow epidemic. NIRS News 12(6), 12–13. [64] Michotte Renier, A., Baeten, V, Sinnaeve, G., Fernández Pierna, A. and Dardenne, P. (2004) The NIR camera: a new perspective for meat and bone meal detection in feedingstuffs, Proceedings of the 11th International Conference on Near-Infrared Spectroscopy (A. Garrido, et al. eds), April, Cordoba, Spain. [65] Fernández Pierna, J. A., Baeten, V., Michotte Renier, A., Cogdill, R. P. and Dardenne, P. (2004) Combination of support vector machines (SVM) and near infrared (NIR) imaging spectroscopy for the detection of meat and bone meat (MBM) in compound feeds. J. Chemometr. 18(7–8):341–349 [66] Fernández Pierna, J. A., Baeten, V., Michotte Renier, A. and Dardenne, P. (2004) NIR camera and chemometrics: the winner combination for the detection of MBM, Book of Abstracts of the International Symposium on Food and Feed Safety in the Context of Prion Diseases, 6–18 June, Namur, Belgium. pp. 119–20. [67] Fernández Pierna, J. A., Didelez, G., Baeten, V., Michotte Renier, A. and Dardenne, P. (2004) Chemometrics for multispectral and hyperspectral data: application to screening

MONITORING AGRICULTURAL FOOD AND FEED PRODUCTS

301

of compound feeds, Poster Presented at CAC-2004, 20–23 September, Lisboa, Portugal. [68] Baeten, V., Michotte Renier, A., Fernández Pierna, A. et al.(2004) Review of the possibilities offered by the near infrared microscope (NIRM) and near infrared camera (NIR Camera) for the detection of MBM, Oral Presentation, Intenational Symposium Food and Feed Safety in the Context of Prion Diseases, 16–18 June, Namur, Belgium.

Index

abundance estimation, 45–51 AC detectors, see detector coupling, alternating current acousto-optic tunable filter, 28, 159 acquisition ratio, 17 agricultural applications: damage assessment, 286–8 NIR spectroscopy, 25–7, 264–9 rationale, 26, 284–5 airborne imaging systems, 285 airborne visible and infrared imaging spectrometer, 285 airy pattern, 73–7 AIS, see airborne imaging systems amide vibrational modes, spectroscopy of, see protein, spectroscopy of ANNs, see artificial neural networks AOTF, see acousto-optic tunable filter apertures, 9, 12, 57, 63–4, 69–71 artifacts, 12–13, 73–8 confocal, 57, 69, 71–2, 75 imaging of circular, 78–80 array detectors: focal plane, 13–16, 19, 27–9, 57, 70, 78–82, Ch. 5 passim, 145–8, 176, 190 formats, 19–21, 28–9, 176 modes, 14, 28–9, 81 near-infrared, 28–9 parameters, 14, 21, 28–9, 176–7 artificial neural networks, 197, 205, 271 atmospheric distortions, 285 attenuated total reflection (ATR), 22, 25, 119–21, 262, 266–7, 270 sample contact, see sample contact, ATR AVIRIS, see airborne visible and infrared imaging spectrometer background, thermal, 14 bandpass filters, see filters, bandpass barley, 275 baseline variations, 38, 92–4, 98–100

bending magnet radiation, 60–65 blackbody radiation, 58–9, 61 bone: in feed, see meat and bone meal limitations of spectroscopy, 238 microspectroscopy relevance, 236–8 spectroscopy, 238–40 structure, 235–6 bovine spongiform encephalopathy, 294–5 bovine tissue, 273, 276 breeding, 291 brightfield imaging, 14–15 brilliance, see source, brightness BSE, see bovine spongiform encephalopathy calibration, 36 transfer, 37 cartilage: spectroscopy of, 225–9, 250, 252–6 structure, 249, 251–3 CASI, see compact airborne spectrographic imager Cassegrain, see objectives, Schwarzschild catalysis: resin-supported, 148–9 surface effects, 149–50 CCD, see charge coupled device cells: imaging, see imaging applications, cytology spectroscopy, 191–2, 197–202, 205–6 centerburst, 3 cereal analysis, 290–293 cervical cancer, 189–90, 204 spectroscopy of, 205–6 cervical tissue: processing, 208–12 spectroscopy of, 207–8 charge coupled device, 27–8, 145 chemical map, see image, spectroscopic chips, 181–2

304

INDEX

classification, multivariate, 32, 166–8, 217–18, 295–7 classifier: discriminant, 166–7, 217–18 dissimilarity-based, 167, 171, 193–4, 217 cluster analysis, 193–4, 217–24; see also hierarchical clustering analysis combinatorial approach, 143–5 compact airborne spectrographic imager, 285, 287–8 compression, see data compression concentration map, see image, spectroscopic confocal, see apertures, confocal contamination, 108–9, 290 contour plots, 122 convolution theorem, 76–7 corn, 272–4 correlation-based analysis, 26 cosine correlation analysis, 271 crop imaging, 286–8 crosslinking, 129–33, 135–7 cytology, 192

dark current/response, 31–2, 37–8, 165 data: acquisition, efficiency of, 13–17 compression, 86–90 quality, see signal to noise ratio verification of quality, 212–16 DC detectors, see detector coupling, direct current deconvolution, 76–7 derivative, see spectra, derivatives design of experiments, 155 detector coupling: alternating current, 5, 20 direct current, 14, 31–2 detectors: electronic noise, 18, 20–21 linear array, 10, 19–22 specific detectivity, see specific detectivity two-column, 20 see also array detectors dewaxing, 211–12 diamond compression cell, 262 diffraction, 67, 73–5, 200 diffuse reflection, 33, 161–2 disease diagnosis, 190, 194–6, 200, 218–24, 238–43, 253–7, 287

dispersion artifacts, 197, 210–211 distribution homogeneity, 35, 38–41, 124–6 PLS measure, 48 sampling considerations, 42 DNA, spectroscopy, 198–200, 205–6

early detection, infestation, 287 electron source size, 65–7 emission sampling, 263 Euclidian distance, 167, 217, 275

Fabry-Perot filters, 30 fast Fourier transform, see Fourier transform feature extraction, 101–4, 122, 152–3 field of view, 9, 13, 22, 32–4, 42, 52 filtering: spatial, 97–101 spectral, 111 filters: acousto-optic, see acousto-optic tunable filter bandpass, 3, 98–101, 146 Fourier transform, 87, 98–101, 111 liquid crystal tunable, see liquid crystal tunable filter near-infrared imaging, 30–31, 159 fixation, 208–9 flax, 274 fluctuations, baseline, see baseline variations food analysis, 172, 181–2, 266–9, 272–7, 288–97 Fourier transform, 3–5, 76–7, 87, 153 infrared, 1 FOV, see field of view FPA, see array detectors, focal plane frame co-addition, 17–18 fruit analysis, 288–90 FT, see Fourier transform FTIR, see Fourier transform, infrared

gain ranging, 18 gas phase array, 147–8 globar source, see source, blackbody GPA, see gas phase array grain: spectroscopy, 25–6 yield, 286

INDEX

grassland imaging, 287–8 gut, 230

HCA, see hierarchical clustering analysis heterogeneity, see distribution homogeneity hierarchical clustering analysis, 193–4, 196, 205, 217–21, 274–5 high frequency modulation, 20 high frequency noise, see noise, high frequency high throughput: experimentation, 143–4 sampling, 1, 42–3, 143–5, 274–5 histograms, 35, 40, 123–6, 128 histopathology, 189, 194–202, 218–24 homogeneity, see distribution homogeneity Hotelling T2 test, 109 hyperspectral imaging, see spectroscopic imaging

image: segmentation, 104–5 spectroscopic, 96, 109–13 imaging: aperture artifacts, 78–80 data processing and display, 122–6 FTIR, 13–16, 28, 33, 146–8, 189–91, 204, 212, 263 near-infrared, 25–7, 29–31, 177, 278, 283–4 push-broom, 159–60 raster scan, 20–1, 82 3-D, 229–30 visible, 284 imaging applications: abundance measurement, 46–51 agriculture, 272–4, 284–8 art, 118 bacteria, 96–7, 106–12, 274–5 bone, 237–43 catalysis, 145–6, 149–51 cytology, 192, 200–202 disease, 190, 194–6, 200, 218–24, 238–43, 253–7, 287 fabrics, 178–81 food, 172, 181–2, 266–70, 273–7, 294–7 histopathology, 192, 194–202, 218–30, 238–43, 252–7, 293 liquid crystal composites, 22 mineralogical, 172

305

paper, 182–4, 275 pharmaceutical, 37–42, 45, 49–51 plants, 274–7, 284 polymers, 123–6, 128–40, 169–72 poultry, 293 reaction kinetics, 148 skin, 243–50, 293 surfaces, 118–19, 172 waste, 169–72, 290, 293 water migration, 178 imaging instrumentation: mid-infrared, 13–22, 27–9, 33, 78–80, 145–8, 177, 190–191, 212, 235 near-infrared, 29–31, 162–4, 177–8 infestation, 287 infrared image, brightfield, 15 infrared microscope, see microscopes, infrared interference, 2, 63 interferogram, 3–4 multipass acquisition, 16 processing, 4, 151–3 recording, 3–7, 147–8 interferometer, see interferometry, instrumentation interferometry: continuous scan, 3, 5, 16 instrumentation, 2, 146–8, 189–91 rapid scan, 5, 16, 147 step scan, 6, 28, 147

kernel analysis, 291–2 kinetics, 150

laminate film, 126–7 larvae, 291 LCTF, see liquid crystal tunable filter leaf area index, 286 linear discriminant classifier, see classifier, discriminant lipid dynamics, 275–6 liquid crystalline dispersions, 129–32, 135–7, 275–6 liquid crystal tunable filter, 30–34, 159 lotion, 182–7 low frequency noise, see noise, low frequency lymph nodes, spectroscopy, 194–7 Lyot filter, 30

306

INDEX

map, chemical, see image, spectroscopic mapping, 8–13, 69, 158–9, 177, 189, 206 Marangoni instabilities, 117 matrix representation, 88–90 maximum likelihood algorithms, 77, 80 MBM, see meat and bone meal meat, 274, 276, 293 meat and bone meal, 294–7 microscopes, infrared, 8–11, 69–73, 189, 190–191 microscopy: FTIR, see microspectroscopy, FTIR near-infrared, 164 microspectroscopy: FTIR, 1, 8, 27, 56–8, 67–73, 78–82, 92, 177, 190–191, 206–7, 234 polarized, 139–40 microtoming, 118, 209–10 mid-infrared: imaging, data preprocessing, 4, 85–101, 121–6, 152–4, 213–15 spectroscopy, see interferometry morphology, 45, 127–40 multispectral techniques, 284–5 multivariate methods, 33, 35, 101–13, 154–5, 193–4, 212–18, 270–272 near-infrared: region, 25 spectroscopy, applications, 25–7 vibrational modes, 25–6 near-infrared imaging: commercialization, 28 configurations, 29–31, 177–8 data pre-processing, Ch. 2 passim, 165–6 see also imaging, near-infrared near-infrared region, 25–7 noise, 16–22, 28, 92–3, 194 amplifier, 18 focal plane array, 17–19 high frequency, 91–3, 95, 104 low frequency, 17, 92–3 numerical aperture, 9, 13, 34, 39, 69, 73, 81 objectives: diffraction pattern, 73–7 Schwarzschild, 8, 69, 71–5, 78, 81–2 working distance, 81 object reconstruction algorithm, 168–9

optical PSF, see point spread function orientation, molecular, 119, 139 osteoporosis, 237–8 osteoporotic bone, spectroscopy of, 237–8, 241–3, 253–4 outlier selection, 108

paper analysis, 275 ‘Pap’ test, see cervical cancer partial least squares, 46–51, 154, 271 quantification, 46–51 score images, 46–51 particle analysis, 36 pathlength, ATR, 120 pathogens, 274–5 PCA, see principal components analysis PDLCs, see polymer dispersed liquid crystals penetration depth, 25 pests, 291 phase: composition, 128–40 correction, 210–211 diagram, 128–32, 133–7 separation kinetics, 129 PLS, see partial least squares point mapping, see mapping point spread function, 73–82 polarization, 61–2, 65, 139–40 polymer: composites, 124–40 dispersed liquid crystals, 22, 128–32, 135–7 laminate, see laminate film polymer blends, 132–7, 140 semicrystalline, 137–40 polymer characterization: preparation techniques, 116–21 requirements, 115–16 sampling, 116 precision agriculture, 286 principal components analysis, 87–90, 104–8, 154, 166, 176, 179, 186, 194, 271, 290–292 protein, spectroscopy of, 206, 214, 252–4, 264–5, 273–8 PSF, see point spread function push-broom scanning, see imaging, push-broom

INDEX

quadratic discriminant classifier, see classifier, discriminant quality control, 43–51, 288–90 Raman spectroscopy, 33–4, 277–8 raster scanning, see imaging, raster scan real-time data processing, 160, 164–5, 168–9 recycling, 170–171 red-edge, 286 reflection, 118–21, 161, 192, 197, 214, 262–3, 267, 270, 285 artifacts, 197 diffuse, see diffuse reflection reflection-absorption, 119–21, 148, 161, 262, 267 remote sensing, 284–5 distortions, 285 resolution: spatial, 56, 70–71, 73–7, 127–8 spectral, 30, 33–4, 37 RNA, spectroscopy, 198, 200 sample contact, ATR, 121 sampling methods, 262, 267, 270 SASI, see shortwave airborne imaging spectrographic imager satellite imaging, 284–5 Savitsky–Golay, 42, 90–96, 166, 194, 214 scattering: atmospheric, 285 contrast mechanism, 32 microscopic, 14–15, 127–8, 210–211 Schwarzschild, see objectives, Schwarzschild sectioning, see microtoming segmentation, see image, segmentation shortwave airborne imaging spectrographic imager, 287–8 signal to noise ratio, 7, 77, 207, 214, 235, 262 imaging, 16–18, 147, 284 SIMPLISMA, 110–113 singular value decomposition, 89 skin: permeation, 246–9 spectroscopy, 243–9 smoothing, 90–92, 95–6, 214 triangular, 91 solid phase catalyst, 145–6 solvent casting, 116–17

307

source: blackbody, 58–9, 61, 68 brightness, 56–8, 60, 65, 67–8 flux, 58, 60, 68 polarization, 61, 65 size, 65–7 specific detectivity, 16–17, 28–9 spectra, derivatives, 38, 44, 93–7, 214 spectral: flux, see source, flux imaging, see spectroscopic imaging radiance, see source, brightness spectrograph, imaging, 31, 163–4 spectroscopic imaging, 13–16, 27–31, 146–8, 162–3, 176–8, 189–91, 285 comparison, 33–4, 284–5 complementary techniques, 33–4, 224–5, 252, 277–8 data size, 85, 121–2, 152, 196–7 mid-infrared, see imaging, FTIR near-infrared, see imaging, near-infrared specular reflection, 263 spherulite, 138–140 spin coating, 117 spot size, 12, 33–4, 58, 67 spray nozzle, 179 squamous metaplasia, 221 statistics, 43 imaging data, 35, 37–42 stray light, 12–13 structure–activity relationship, 150 support vector machines, 295 SVD, see singular value decomposition SVMs, see support vector machines synchrotron: beam lines, 57, 66 source, 57–68, 81 source polarization, 61, 65

Terahertz spectroscopy and imaging, 57 thermography, 146 time-resolved spectroscopic imaging, 22 time-resolved spectroscopy, 7, 148 tissue: imaging, 189, 192, 194–7, 218–30 processing, 208–12 transflection, see transmission-reflection transmission, 9–11, 116, 161, 262 transmission-reflection, 9–11, 118–20, 161, 262

308 vegetation indices, 285–6 vibrational spectroscopic imaging, see spectroscopic imaging Ward’s algorithm, 193, 218, 275

INDEX

waste: sorting and classification, 169–72 see also imaging applications, waste wet biomass, 286 wheat, 272, 287