UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Raman spectroscopy for optical diagnosis in head and neck tissue Lau, David Pang Cheng 2006

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
831-ubc_2006-0538.pdf [ 13.34MB ]
Metadata
JSON: 831-1.0092915.json
JSON-LD: 831-1.0092915-ld.json
RDF/XML (Pretty): 831-1.0092915-rdf.xml
RDF/JSON: 831-1.0092915-rdf.json
Turtle: 831-1.0092915-turtle.txt
N-Triples: 831-1.0092915-rdf-ntriples.txt
Original Record: 831-1.0092915-source.json
Full Text
831-1.0092915-fulltext.txt
Citation
831-1.0092915.ris

Full Text

RAMAN SPECTROSCOPY FOR OPTICAL DIAGNOSIS IN HEAD AND NECK TISSUE by DAVID PANG CHENG LAU B.Med.Sci. (Hons), University of Nottingham, UK, 1987 BMBS, University of Nottingham, UK, 1989 FRCS, Royal College of Surgeons, Edinburgh, UK, 1996 FAMS, Academy of Medicine, Singapore, 2004 A THESIS SUBMITTED IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE in THE FACULTY OF GRADUATE STUDIES (Surgery) THE UNIVERSITY OF BRITISH COLUMBIA April 2006 © David Pang Cheng Lau, 2006 Abstract Background and Aims Raman spectroscopy (RS) uses light to detect vibrational characteristics of molecules and can distinguish different molecular structures. Extrapolating this to a medical setting, it was hypothesized that RS could be used to make a tissue diagnosis if it could detect molecular changes associated with tissue pathology. The main aims of this study were to determine whether: 1. Raman spectra could be obtained rapidly in-vitro from head and neck tissue, and 2. Differences could be detected between benign and malignant tissue in various sites in the head and neck. Methodology A Raman spectrometer with 785nm excitation laser and a charge-coupled device detector was used to acquire spectra in-vitro from nasopharyngeal, laryngeal and thyroid tissue. Spectral acquisition times ranging from 1 to 30 seconds were studied. Spectra from benign and malignant tissue in the different sites were compared using statistical techniques, with histopathology as the "gold-standard" for diagnosis. Results Good quality spectra were acquired within 5 seconds. Paired analysis of nasopharyngeal specimens (n=6) showed significant differences between benign and malignant tissue at 1297-1305, 1377-1381, 1436-1442, 1541-1555, and 1614-1626cm"1. Paired thyroid tissue analysis (n=5) showed differences at 1264-1266 and 1477cm"1. Unpaired analysis using multivariate statistical techniques showed sensitivity and specificity values in differentiating benign and malignant tissue of 83.3% and 72.7% for nasopharynx (n=23), 69.2% and 94.1% for larynx (n=47), and 86.7% and 70.0% for thyroid (n=65). In addition in the larynx, squamous papilloma could be distinguished from normal tissue and carcinoma with sensitivity and specificity of 87.5% and 93.5% respectively. Conclusions The system could acquire spectra rapidly in-vitro and has potential for in-vivo application in the head and neck, although specialized probes will need to be developed for this purpose. Spectral differences were detected between benign and malignant tissue in the nasopharynx, larynx and n thyroid. Although diagnostic sensitivity and specificity are currently lower than histopathology, they are sufficiently high to warrant further study of this technique as a means of achieving non-invasive tissue diagnosis. Other areas requiring further development include evaluation of wider spectral ranges and characterization of spectra at cellular and biochemical levels. i i i Table of Contents Abstract " Table of Contents i v List of Tables vii List of Figures i x Preface x i i Acknowledgment and dedication xiii PART I INTRODUCTION Chapter 1 Overview of spectroscopy 1 1.1 What is spectroscopy? 1 1.2 Basic properties of electromagnetic radiation 1 1.3 Basic properties of atoms in relation to spectroscopy 3 1.4 Basic properties of molecules in relation to spectroscopy 4 1.5 Origin of Raman spectra 7 1.6 Vibrations in polyatomic molecules 10 Chapter 2 Instrumentation in Raman spectroscopy 12 2.1 Components of a Raman system 12 2.2 Excitation sources 14 2.3 Wavelength selectors 22 2.4 Light detectors 35 2.5 Optical sampling system 42 Chapter 3 Spectral processing 47 3.1 Overview of spectral processing 47 3.2 Pre-processing 47 3.3 Spectral analysis 49 iv Chapter 4 Applications of Raman spectroscopy in medical science 57 4.1 Blood vessels 57 4.2 Brain 5 7 4.3 Breast tissue 5 8 4.4 Gastrointestinal tract 5 8 4.5 Gynaecological tract 59 4.6 Ophthalmology 6 0 4.7 Respiratory tract 60 4.8 Skin, hair and nails 6 1 PART 2 T H E S T U D Y Chapter 5 Background to the study 6 3 5.1 Rationale 6 3 5.2 Hypothesis 6 4 5.3 Objectives 6 4 Chapter 6 Materials and methods 6 5 6.1 Tissue preparation 65 6.2 Spectrometer system 66 6.3 Spectral analysis 6 8 Chapter 7 Results 7 3 7.1 Sampling variability 7 3 7.2 Paired analysis 7 9 7.3 Unpaired analysis 8 9 7.4 Functional group assignments 1 0 6 Chapter 8 Discussion 1 1 0 8.1 Advantages and disadvantages of Raman spectroscopy 110 8.2 Interpretation of results and clinical applicability 112 8.3 Biological correlates 1 1 4 v 8.4 Proposals for future studies 118 8.5 Conclusions 125 PART 3 R E F E R E N C E S 126 PART 4 A P P E N D I C E S 140 Appendix 1 1 4 ° Appendix 2 147 v i List of Tables Table Title Page 1 -1 Energy units for wavelength (X), frequency (v) and wavenumber (v') for different regions of the electromagnetic spectrum 2 1- 2 Relationship of energy level transition to electromagnetic wavelength 6 2- 1 Wavelength characteristics of different types of lasers 19 7-1 Integration areas for normal tissue (n) and nasopharyngeal carcinoma (ca) at wavebands 1297-1305, 1377-1381, 1436-1442, 1541-1555, 1614-1626 cm"1 81 7-2 Integration areas for normal tissue (n) and papillary carcinoma (ca) at waveband 1264-1266 cm" 1 86 7-3 Diagnostic significance and M A N O V A coefficients of the first five principal components for nasopharyngeal tissue 91 7-4 Cross-validation classification results for normal tissue and carcinoma of the nasopharynx 93 7-5 Diagnostic significance and M A N O V A coefficients of the first five principal components for laryngeal tissue 96 7-6 Cross-validation classification results for normal tissue, squamous cell carcinoma and squamous papilloma of the vocal folds 98 7-7 Sensitivity and specificity of RS as a diagnostic test to classify pathology in the larynx 100 7-8 Diagnostic significance and MANOVA coefficients of the first five principal components for thyroid tissue 103 vii 7-9 Cross-validation classification results for benign tissue and papillary carcinoma of the thyroid gland 105 7-10 Functional group assignments and possible biochemical correlates for nasopharyngeal tissue 107 7-11 Functional group assignments and possible biochemical correlates for laryngeal tissue 108 7- 12 Functional group assignments and possible biochemical correlates for thyroid tissue 109 8- 1 Comparison of wavebands and functional group assignments in different tissues where differences were identified between neoplastic and non-neoplastic tissue 117 v i i i List of Figures Figure Title Page 1-1 Energy levels of a diatomic molecule 5 1 -2 Schematic representation of Raman scattering 8 1-3 Energy transitions in normal Raman, resonance Raman and fluorescence spectra .. 9 1- 4 Different vibrational modes of the molecule C H 2 11 2- 1 Main components of a Raman spectroscopy system 13 2-2 Energy diagram illustrating spontaneous emission and stimulated emission in a two-level system 16 2-3 Schematic of a gas laser resonant cavity 17 2-4 Structure of a diode laser 21 2-5 Principle of refraction of light 23 2-6 Dispersion of light into constituent wavelengths using a prism 24 2-7 Diffraction of light at two parallel slits 28 2-8 Geometrical diagram illustrating conditions for constructive or destructive interference when light is diffracted through parallel slits 29 2-9 Czerny-Turner spectrometer arrangement 30 2-10 Axial transmissive spectrograph 31 2-11 Etendue of a spectrometer 32 ix 2-12 The Michelson interferometer 34 2-13 Diagram of a photomultiplier tube 37 2-14 Quantum efficiency of a typical photo-emitter 38 2-15 Schematic representation of the charge-transfer and readout pattern, of a two-dimensional C C D 41 2-16 Basic design of an'W-around 1" probe 45 2- 17 Modifications of the "/V-around-1" probe design 46 3- 1 Schematic representation of a single neuron in an ANN 55 3-2 Schematic representation of a multi-layer perceptron 56 6-1 Optical layout of the Raman system 67 6-2 Example of raw Raman spectra 69 6- 3 Examples of Raman spectra after subtraction of fifth-order polynomial (upper plot) and following five-point smoothing (lower plot) 70 7- 1 Effect of varying signal acquisition time on Raman signal intensity for nasopharyngeal tissue 74 7-2 Effect of specimen size on Raman signal intensity 76 7-3 Effect of varying specimen orientation on the Raman signal 78 7-4 Paired spectra from 6 subjects comparing normal and carcinoma tissue from the nasopharynx 82 x 7-5 Line graphs (a to e) indicating direction and magnitude of change in overall spectral intensity (measured by integration area) at the given wavebands, between normal nasopharyngeal tissue and nasopharyngeal carcinoma 84 7-6 Paired spectra from 5 subjects comparing normal tissue and papillary carcinoma from the thyroid gland 87 7-7 Line graph indicating direction and magnitude of change in overall spectral intensity (measured by integration area) at waveband 1264-1266cm"1, between normal thyroid tissue and papillary carcinoma 88 7-8 Comparison of mean spectra from normal nasopharyngeal tissue and undifferentiated nasopharyngeal carcinoma 90 7-9 Scatter plot allowing visualization of separation of groups for normal tissue and carcinoma of the nasopharynx 92 7-10 Comparison of mean spectra from normal tissue, squamous cell carcinoma and squamous papilloma of the vocal fold 95 7-11 Scatter plot allowing visualization of separation of groups for normal tissue, squamous papilloma and squamous carcinoma of the vocal fold 97 7-12 Comparison of mean spectra from normal tissue, adenoma and papillary carcinoma of the thyroid gland 101 7-13 Comparison of mean spectra from benign thyroid tissue (comprising normal tissue, multi-nodular goitre and adenoma), and papillary carcinoma 102 7- 14 Scatter plot allowing visualization of separation of groups for benign tissue and papillary carcinoma of the thyroid gland 104 8- 1 Artist's impression of a Raman guided biopsy forceps 121 8-2 Optical features of proposed laryngeal imaging Raman probe (LIRAP) 123 xi Preface "...the study of light scattering might carry one into the deepest problems of physics and chemistry..." Sir Chandrasekhara Venkata Raman during his Nobel lecture, December 11, 1930 This thesis is a study of spectra in more than one sense of the word. While Raman spectra form the basis of research in this study, it became increasingly clear to me during my research that I was also studying a spectrum of a different kind. Not only did I have to focus on a problem in clinical medicine related to the health of a human being, but in order to do so also had to try and understand how things function at the most fundamental level, that of the atom (and even a sub-atomic level). One of the problems then was how best to marry these distant and yet intimately related fields of study. Until the early 19th century, a scientist was often also a physician, mathematician, philosopher, chemist, biologist, engineer, or even a politician or artist. Indeed, embarking on this study emphasized to me the importance of each of these fields and often left me wanting. Today science (and art) has grown to such an extent that with few exceptions modern scientists have to limit their attention to one or two specialized areas. Are we not in danger of knowing more and more about less and less, to the extent that we may end up knowing nothing about anything! The deeper I read into the subject, the more I realized that it may take a lifetime, if not longer, to fully grasp all its concepts. It simply may not be possible to be an expert in all aspects of the related fields involved in this type of research, but nonetheless my mentors and colleagues have inspired me to try, and they have also shown me the importance of learning to be a gatekeeper in a multidisciplinary team. The research in this thesis can be described as preliminary, and yet there are areas within it, which I believe break some new ground. I hope that the thesis will serve to an extent as a reference base to further develop this work. Most of all, it has been an interesting journey from which I have made some good friendships. xii Acknowledgment and Dedication This thesis would not have been possible without the help of the following people: My mentors, colleagues and friends, Professors Murray Morrison, Harvey Lui, Haishan Zeng and Zhiwei Huang. Ms Michelle Zeng for compiling the data so accurately and systematically. Dr Ken Berean for making the histological diagnoses. Dr's Chris Man, Don Anderson and Scott Durham for allowing me to use tissue from their patients. Dr Michael Schulzer and Ms Shen Liang for putting in all those hours on the complicated statistics. My dear friend lltfat Hamzavi, whose enthusiasm hatched this project! My dearest children, Matthew and Lucinda, without whose help this thesis would have been completed in a fraction of the time. But I would not have traded that time for anything else in the world. My dearest wife, Rachelle, whose enduring love, patience and encouragement enabled this project to be completed. My father for his patience, ideas and suggestions. Thank you to you all! xiii 1. Overview of Spectroscopy 1.1 What is spectroscopy? Spectroscopy is the study of the absorption and emission of light (and other forms of radiation) by matter (Encyclopedia 2003). These processes depend on the wavelength of the radiation as well as the physical and chemical properties of matter. Spectroscopic techniques have been applied in virtually all fields of science. Examples include radio-frequency spectroscopy to study nuclear moments in magnetic resonance imaging, x-ray spectroscopy to study constituents of distant stars, and optical spectroscopy to determine the physical structure and chemical composition of matter. This thesis is mainly concerned with optical spectroscopy. To understand spectroscopy it is necessary to know more about the basic properties of electromagnetic radiation, and atoms and molecules, which are the fundamental particles of matter. 1.2 Basic properties of electromagnetic radiation Electromagnetic radiation consists of oscillating electric and magnetic fields, which enable energy to be transferred through space. The energy propagates as a wave (with electric and magnetic fields in planes at right angles), with a speed of approximately 3 x 10 8 ms"1 in a vacuum (c). The distance between successive crests in a wave is referred to as its wavelength (X). Different types of electromagnetic radiation have different wavelengths, with visible light forming only a small part of the electromagnetic spectrum between 400 and 800 nanometers (nm). Within the visible part of the spectrum, red light has a longer wavelength than green light, which in turn has a longer wavelength than blue light. Likewise, radio waves have considerably longer (10 1 2 nm) and gamma rays considerably shorter (10"r nm) wavelengths than visible radiation. The ability to decompose electromagnetic radiation into component wavelengths is fundamental to spectroscopy and will be discussed in more detail later. An electromagnetic wave can also be characterized by its frequency (v), which is expressed in hertz (Hz) and given as vA = c , v Another parameter common to vibrational spectroscopy is the "wavenumber", v', defined by v = —, c which has the dimension (1/s)/(cm/s) = 1/cm or cm" 1. Table 1-1 lists the energy units for different regions of the electromagnetic spectrum. 1 Spectral region X (cm) v (Hz) . v' (cm"1) y-ray 10"8 - 10 1 0 3 x 1 0 1 8 - 3 x 1 0 2 0 1 08 - 10 1 0 x-ray 1 0 " 6 - 10"8 3 x 1 0 1 6 - 3 x 1 0 1 8 1 0 6 - 1 0 8 Ultraviolet-visible 1 0 " 4 - 10"6 3 x 1 0 1 4 - 3 x 1 0 1 6 10 4 - 10 6 Raman and infrared 1 0 " 2 - 1 0 - 4 3 x 1 0 1 2 - 3 x 1 0 1 4 1 0 2 - 1 0 4 Microwave 1 - 1 0 " 2 3 x 1 0 1 0 - 3 x 1 0 1 2 1 - 1 0 2 Electron spin resonance 1 0 2 - 1 3 x 1 0 8 - 3 x 1 0 1 0 10"2 - 1 Nuclear magnetic resonance 1 0 4 - 1 0 2 3 x 1 0 6 - 3 x 1 0 8 1 0 " 4 - 1 0 - 2 Table 1-1 Energy units for wavelength (X), frequency (v) and wavenumber (v') for different regions of the electromagnetic spectrum. 2 1.3 Basic properties of atoms in relation to spectroscopy An atom consists of a number of negatively charged electrons bound to a nucleus containing an equal number of positively charged protons, and a generally different number of neutrons. An atom can possess different energy levels. These levels are discreet states termed quantum states. They are largely determined by electronic structure, i.e. the number and arrangement of electrons around the nucleus (electrons possess electrostatic, kinetic and magnetic energy). Energy transitions can also occur in the nucleus although the energy exchange required is greater, in the order of thousands to millions of electron volts (eV) compared to less than thousands of eV for electronic transitions. An atom can make the transition from lower to higher energy-state by absorbing a "quantum" of electromagnetic radiation. A quantum of light energy is termed a "photon". The energy transition, E that can occur between quantum states is dependent on the frequency, v of the electromagnetic radiation such that E = hv where h is Planck's constant (6.626 x 10~3 4 Js). When the atom makes the transition from higher to lower energy-state (the lowest possible energy-state being termed the ground state), it can likewise emit a quantum of light (photon) of the same frequency. Analysis of absorbed and emitted wavelengths can be used to determine the energy states of the atom and thus its electronic structure and properties. 3 1.4 Basic properties of molecules in relation to spectroscopy A molecule is a collection of positively charged nuclei surrounded by a cloud of negatively charged electrons. Like atoms, molecules also possess discreet allowable energy levels, but as their structure is more complex, so are the possible energy transitions. In addition to interactions between electrons and nuclei, forces also result from interactions between the various nuclei. In decreasing order of magnitude, the different energy transitions can be broadly grouped as electronic, vibrational and rotational. The electronic energy of a molecule is the sum of kinetic energy due to electron motion, and electrostatic energy due to attraction between electrons and nuclei, and repulsion between individual electrons and individual nuclei. Rotational energy results from free rotation of the molecule, which can occur in the gas phase. Mathematically, the rotational energy of a diatomic molecule (the simplest form of a molecule, consisting of two atoms) can be calculated by assuming the molecule to behave like a "rigid rotor". However real molecules are not rigid structures, and motion of nuclei within the within the molecular framework gives rise to vibrational energy levels. Using the diatomic molecule as a model, the vibrational energy level can be represented mathematically by assuming the molecule to behave as a "harmonic oscillator". The combined rotational-vibrational energy level, Evj, calculated using the rigid-rotor, harmonic oscillator model is given by EvJ=(v + ^)hv0+BJ(J + \) where v is the vibrational quantum number (v = 0 , 1, 2 . . . ) , h is Planck's constant, v0 is a function of the force constant of the bond characteristic of that particular molecule, B is the rotational constant of the molecule, and J is the rotational quantum number (J = 0 , 1, 2 . . . ) . The energy transition between quantum levels differs considerably in magnitude for rotational, vibrational and electronic transitions (Figure 1-1). It follows from E = hv, that the frequency of electromagnetic radiation absorbed or emitted to make the transition also varies (Table 1-2) . Raman spectroscopy is mainly concerned with changes in the vibrational state of a molecule. Vibrational transitions occur in the 10~ 2 - 10" 4 cm range of the electromagnetic spectrum, also termed the Raman and infrared region of the spectrum. Rotational transitions also occur in this region of the spectrum. However as they are concerned mainly with gases they are not discussed further. 4 1 V=l Electronic ' 1 excited state 3 2-J=0! I Pure rotational transition 2. J=0 : Pure vibrational transition V=0 Pure electronic transition Electronic ground state Figure 1-1 Energy levels of a diatomic molecule. The figure illustrates differences in the magnitude of rotational, vibrational and electronic transitions (although not to scale). J and v represent allowable rotational and vibrational quantum states respectively. It is worth noting that the diatomic molecule does not behave like a true harmonic oscillator but rather as an anharmonic oscillator. As a result the separation between energy levels is not equidistant but decreases with increasing v. This is sometimes illustrated using a potential energy curve (not shown here). 5 Type of transition Range (v\ cm-1) * Spectral region Rearrangement of elementary particles within the nucleus 10 8 -10 1 0 y-ray Between energy levels of inner electrons 10 6-10 8 x-ray Between energy levels of valence electrons 1 0 4 - 106 UV to visible Between vibrational levels of nuclei constituting the molecule ' 10 2 -10 4 Raman to infrared Between rotational levels (change in orientation) 1 - 1 0 2 Microwave Between electron spin levels in magnetic field 10"2 -1 Electron spin resonance (ESR) Between nuclear spin levels in magnetic fields 10"2-10-4 Nuclear magnetic resonance (NMR) Table 1-2 Relationship of energy level transition to electromagnetic wavelength. * v' represents wavenumber where v'=v/c. It is thus closely related to frequency v 6 1.5 Origin of Raman spectra In Raman spectroscopy a sample is irradiated by an intense laser-beam, and scattered light is usually observed in a direction perpendicular to the incident beam (Figure 1-2). Most of the scattered light is at the same frequency as the incident beam (v0). A small amount of light (about 10"5 of the incident beam) is scattered at a new frequency v0 + vm, where v m i s a vibrational frequency of the molecule from which light is scattered. The v0 + vm and v0 - vm lines are called the Stokes and anti-Stokes line respectively. Thus in Raman spectroscopy, the vibrational frequency (vm) is measured as a shift from the incident beam frequency (v0). A Raman spectrum can be produced by plotting the intensity of the scattered light against the Raman shift, usually represented by wavenumber. Figure 1-3 illustrates Raman scattering in terms of a simple diatomic energy level. In normal Raman spectroscopy the exciting line (v0) is chosen so that its energy is well below the first electronic excited state. The dotted line indicates a "virtual state" to distinguish it from the real excited state. The population of molecules at v - 0 is much larger than that at v = 1, as the lower energy state is always preferred (Maxwell Boltzmann distribution law). Hence the Stokes lines (S) are stronger than the anti-Stokes lines (A) under normal conditions. As both provide the same information, it is customary to measure only the Stokes side of the spectrum, the Rayleigh line of course having a Raman shift of zero (Ferraro, Nakamoto et al. 2003). Other types of Raman scattering can also occur. Resonance-Raman scattering occurs when the exciting line is such that its energy intercepts the manifold of the electronic excited state. The resulting resonance-Raman bands demonstrate extremely strong enhancement. Fluorescence spectra can also result when the excited state molecule decays to its lowest vibrational level via radiation-less transitions (these occur via molecular collisions), and then emits radiation. The lifetime of the excited state in resonance-Raman is very short (~10"14s), while those in fluorescence are much longer (~10"8 to. 10~5). 7 Incident light Scattered light Rayleigh scatter Raman scatter Figure 1-2 Schematic representation of Raman scattering. Rayleigh scattering occurs at the same frequency as the incident beam (v0). Raman scattering occurs at a different frequency (v0± vm), where vm is a vibrational frequency of the sample. 8 3 Jk 2 J k 1 i v = 0 I k 3 1 k i k 2 1 A 1 vm 1 1 u v = 0 R S A Normal Resonance Fluorescence Raman Raman Figure 1-3 Energy transitions in normal Raman, resonance Raman and fluorescence spectra. R, S and A represent transitions occurring for Rayleigh, Stokes and anti-Stokes scattering. These also occur in resonance Raman although only the Stokes transition is illustrated. In fluorescence, the dotted line indicates radiation-less transition. The difference between excitation and scattered frequencies, vm is the Raman shift. 9 1.6 Vibrations in polyatomic molecules In diatomic molecules, vibration occurs along the chemical bond connecting the nuclei. In polyatomic molecules, the situation is complicated because each nucleus performs its own oscillations. This can be pictured by imagining a mechanical model of a molecule struck by a hammer. It would vibrate in an extremely complicated manner. However these complicated vibrations can be expressed as a superposition of a number of "normal vibrations" that are completely independent of each other. Figure 1-4 illustrates the various vibrational modes that can occur in a molecule, using a three-atom molecule as an example. The stretching mode can be understood by imagining the C-H bond to be a spring. If the two springs are stretched apart simultaneously and released, "symmetrical stretching" vibrations occur. If the springs are released after stretching one spring and shrinking the other, then "asymmetrical stretching" is said to occur. The third type of normal vibration involves angular changes or bending the C-H bond in what is termed "scissoring". Other bending modes that can occur include "rocking", "twisting" and "wagging", which are also essentially rotational motions of the whole molecule about the three principle axes of rotation that go through its center of gravity. Notations such as ^ for stretching modes and <5for deformation or bending modes are commonly used. Generally, stretching modes have higher vibrational frequencies than bending modes, as the energy required to stretch a bond is higher than that required to bend it. The vibrational frequencies of a number of functional groups (a certain aggregate of atoms) is known, and available in tables and charts (Dollish, Fateley et al. 1974; Socrates 2001). This data has been derived empirically over many years using combined techniques such as gas chromatography - mass spectroscopy, and liquid chromatography - mass spectroscopy to separate and characterize individual components of a sample. Mathematical methods can also be used to predict vibrational frequencies. A functional group may exhibit the same "group frequencies" regardless of the make-up of the rest of the molecule. However neighboring groups and atoms, and the spatial geometry of the molecule can result in perturbations and a shift in the characteristic bands. Biological macromolecules pose a particular challenge, having large and complicated molecular structures. The Raman spectra of biological macromolecules are broad and complicated and it may be difficult to accurately identify the origins of individual bands. Nonetheless, group frequencies for a number of biological macromolecules have been described (Socrates 2001). In biological systems, it can sometimes be easier to consider the spectrum as a whole, representing a "molecular signature", rather than focusing on individual group frequencies. Another problem that can make interpretation of Raman spectra difficult in biological systems is the presence of fluorescence, which is usually stronger than the Raman signal and may obscure it. This is dealt with in more detail in subsequent sections. 10 a) Stretching modes H H H H \ / \ / C c Symmetrical stretching Asymmetrical stretching b) In-plane deformation or bending modes FT* "~H FP FP \ / \ / C C Scissoring Rocking c) Out-of-plane deformation or bending modes © e e e H H H H \ / \ / C C Twisting Wagging Figure 1-4 Different vibrational modes of the molecule C H 2 . Symmetrical and asymmetrical stretching, and scissoring are referred to as "normal" modes of vibration. Rocking, twisting and wagging involve rotational motions of the whole molecule about the three principle axes of rotation. 11 2. Instrumentation in Raman spectroscopy 2.1 Components of a Raman system The following are the major components that make up a Raman spectrometer (Figure 2-1) (Ferraro, Nakamoto et al. 2003): 1. Excitation source. 2. Wavelength selector. 3. Light detector (and computer processing system). 4. Sample illumination and collection system (optical sampling system). Each of these components will be discussed in further detail in the following sections. 12 Illumination Excitation system Sample source Collection system Wavelength selector detector I Computer Figure 2-1 Main components of a Raman spectroscopy system. The illumination and collection systems may include focusing lenses and filters to exclude extraneous wavelengths. 1 3 2.2 Excitation sources Historical aspects In 1928, when Sir Chandrasekhara Venkata Raman discovered the optical phenomenon that bears his name, only crude instrumentation was available. The sunlight was his light source, a telescope the collection system, and his eyes the detector. Much early research focused on development of better excitation sources. Lamps using helium, bismuth, lead and zinc elements proved unsatisfactory because of low light intensities. Subsequently in the 1930's mercury lamps suitable for Raman use were designed. The Toronto mercury arc eventually became the standard source for Raman spectroscopy in the 1950's. This lamp was an unwieldy device consisting of four feet of pyrex tubing coiled about a six-inch radius, connecting two pools of mercury covering high voltage electrodes. A great deal of time was required to warm and ignite the lamp, only to have to shut it down periodically to prevent overheating. A major advance in excitation sources was the advent of lasers in 1962. Lasers are ideal excitation sources for Raman spectroscopy for a number of reasons: (1) Laser beams are monochromatic and extraneous wavelengths are much weaker (notch filters can further be used to eliminate these extraneous lines). (2) Laser beams can be produced over a wide range of wavelengths. (3) A small diameter beam can be produced using a laser. This can range from several millimeters down to 2um, enabling the study of small sample areas. A variety of lasers are now available and are considered the standard excitation sources. These will be discussed in greater detail in the following sections. Principles of laser action The word laser is an acronym for light amplification by stimulated emission of radiation. The principle of laser action is based on the concept of "population inversion". According to the Maxwell-Boltzmann distribution law given by the ratio of the population of molecules in the excited state Py=i, to that in the ground state P^ 0 . reduces exponentially as the transition energy A E (energy required to raise the molecule from ground to excited state) increases, k is Boltzmann's constant 1.3807 x 10~16 erg/degree, and 7" is the absolute temperature. To maintain this equilibrium, excited molecules revert to the ground state by "spontaneous emission" of photons with energy A E , and the population in the ground state is always larger than that in the excited state. In this situation, photons of energy A E are more likely to be absorbed by molecules in the ground state than to cause "stimulated emission" of further photons 14 from excited molecules. Stimulated emission occurs when the population in the excited state is larger than in the ground state. This condition known as "population inversion" can be brought about using energy in the form of an electromagnetic wave (termed the "pump") to raise molecules to an excited state. Excited molecules can revert to a lower energy state by spontaneous emission of photons of energy A E . These photons can trigger emission of further photons from excited molecules, a process termed "stimulated emission" (Figure 2-2). Because each incident photon results in emission of two photons (one by stimulated emission and the other emitted spontaneously), a cascade of emissions results, with emitted photons having the same energy and phase as the original photon. The laser medium can consist of a gas, liquid or solid. The pump may take the form of a flash lamp, electrical discharge or strong laser beam. The laser beam thus obtained is amplified by trapping it in a resonant cavity. Figure 2-3 illustrates an example of a resonant cavity for a typical gas laser. Mirrors are used to reflect emitted photons into the plasma tube thus stimulating emission of further photons, creating a cascade effect and amplification of light. Examples of gaseous lasers include argon, krypton, helium-neon and carbon dioxide. Dye lasers utilize solutions of organic dye molecules as the laser medium. Solid-state lasers utilize materials such as Neodymium-YAG and ruby. Semiconductor diode lasers are another form of solid-state laser that produces optical emissions at semiconductor p-n junctions. The choice of laser as an excitation source may depend on several factors, including the laser wavelength, output power required, laser efficiency, size and cost. 15 2 2 1 Spontaneous emission Stimulated emission Figure 2-2 Energy diagram illustrating spontaneous emission and stimulated emission in a two-level system. Dotted arrow indicates transition from higher to lower energy state. Open arrows represent photons. In spontaneous emission a single photon is emitted during the transition. In stimulated emission an incident photon of the same energy stimulates emission of two photons. 16 Output mirror High reflectance mirror Power Figure 2-3 Schematic of a gas laser resonant cavity. The power source or "pump" is used to raise molecules within the plasma tube to an excited state. As molecules revert to the ground state, spontaneously emitted photons result in stimulated emission from excited molecules. Photons thus emitted are reflected back into the resonant cavity by mirrors eventually resulting in a cascade of stimulated emission of photons. These photons have the same energy and phase and can be released via the output mirror as a laser beam. 17 Choice of laser wavelength Some of the wavelength characteristics of different lasers are listed in Table 2-1. A major consideration when selecting an appropriate excitation wavelength for Raman spectroscopy for tissue diagnosis is that tissue exhibits strong broadband fluorescence. This is most prominent when visible excitation is used, particularly in the blue and green regions of the spectrum (400 - 520 nm) (Hanlon, Manoharan et al. 2000; Carter and Edwards 2001). The Raman signal may be swamped by the fluorescence emission, which may be as much as 10 4 times greater in intensity. Fluorescence can be reduced by continuous exposure of the specimen to the laser to produce "fluorescence burnout" (Clarke, Hanlon etal. 1987; Clarke, Wang et al. 1988; Carter and Edwards 2001). However this technique is not ideal as it can cause heating and drying of tissue specimens. Fluorescence decreases rapidly at longer excitation wavelengths, especially towards the near-infrared (1064nm) and infrared end of the spectrum. This is because longer wavelengths possess lower energy given that E = hv = he IA , and are less able to raise the molecules to their excited electronic states from which fluorescence originates (Figure 1.3). Fluorescence can also be reduced using deep ultraviolet light at wavelengths below 270nm. In biological materials the high lying electronic states excited by ultraviolet light often rapidly relax via non-radiative processes to lower levels, causing fluorescence to occur at much longer wavelengths. As a result the wavelength of the largest shifted Raman band will still be shorter than the start of the fluorescence emission if the excitation wavelength is short enough (Nelson, Manoharan et al. 1992; Asher 1993; Asher 1993). Ultraviolet excitation can further enhance the Raman signal by exciting the electronic states of the molecule thereby causing resonant enhancement (Figure 1-3). However one of the key disadvantages of using ultraviolet wavelengths in biological systems is the potential for mutagenesis, which limits the use of these wavelengths in the clinical setting. In addition, tissue penetration depth of ultraviolet light is in the order of microns, which is significantly less than for longer wavelengths, making ultraviolet light suitable only for studying superficial characteristics of biological tissue (Hanlon, Manoharan et al. 2000). 18 Laser Wavelength Gas Lasers Carbon dioxide 10600 nm (IR) Argon 350-515 nm (UV to green) Helium-Neon 632.8 nm (red) Excimer (e.g. XeCI) 308 nm (UV) Liquid Lasers Dye lasers 4 0 0 - 1 0 0 0 nm (tunable) Solid State Lasers Ruby 694.3 nm (red) Neodymium-YAG 1064 nm (near IR) Diode 3 8 0 - 3 5 0 0 nm (blue to IR) Table 2-1 Wavelength characteristics of different types of lasers 1 9 Other factors determining choice of laser Size, cost and efficiency of a laser are other factors that can determine its choice. In the setting where Raman spectroscopy is used for clinical diagnosis, portability is an important consideration. Diode lasers have several valuable properties that make them potentially useful in a clinical setting. They have high efficiencies (defined as laser power output / electrical power input) of typically 30-50%, which is higher compared to most other lasers. They are small in size and can have dimensions less than 1 mm. Their cost is relatively low due to increasingly widespread use in commercial applications over the past decade, such as the compact disc player, laser printer, barcode scanner and optical fiber communication systems. Diode lasers are available over a range of specific wavelengths from the blue to the infrared region. They do however have a tendency to drift in wavelength in a temperature sensitive fashion. This problem can be improved by careful control of temperatures to within ±0.01 °C, and incorporation of a grating into the laser cavity to maximize one particular laser mode and consequently one particular wavelength (Graham Smith and King 2000; Slater, JM et al. 2001; Ferraro, Nakamoto et al. 2003). Figure 2-4 illustrates the basic functioning of a diode laser. 20 © Laser output p-type AlGaAs n-type AlGaAs p-n junction Polished face O Figure 2-4 Structure of a diode laser. A semiconductor diode is a junction between two doped semiconductors (doping being the introduction of an impurity which alters charge structure). One semiconductor is an n-type with extra negative charge or electrons. The other is a p-type with extra positive charge or "holes". The active region of the diode is the p-n junction. Passing a current through the diode results in electrons from the n-side and holes from the p-side forced into the junction. As electrons and holes recombine, the junction becomes charge depleted, and electrons undergo transition from the conduction band (high energy) to the valence band (low energy) with emission of a photon. Doping provides the necessary population inversion. The parallel faces of the junction layer are polished to produce a resonant cavity. Up to a threshold current light is emitted by spontaneous emission and the diode acts as an LED (light emitting diode). Above the threshold stimulated emission dominates over spontaneous emission and laser action starts. The most common laser diodes use GaAs sandwiched between AlGaAs. 21 2.3 Wavelength selectors The wavelength selector or spectrograph is the heart of the Raman system. It allows light to be dispersed, or separated into component wavelengths. In combination with a photo-detector, it is used to determine intensity of light at different wavelengths. Three basic types of spectrographs are prism and grating spectrographs, and interferometers, which are described in turn. These can be configured into a number of different designs, and the advantages and disadvantages of some of these are also discussed. To understand how these devices work, necessitates a discussion of basic optical physics. Prism spectrographs The earliest Raman spectra were dispersed using a glass prism (Adar 2001). Prisms depend on the optical principle of refraction. When light passes between two different transparent media (e.g. air and glass), different wavelengths are refracted (bent) through different angles, the degree to which is determined by Snell's law, given by nx sin i = n2 sin r where ;' and r are the angles made by incident and refracted light rays with the normal to the plane boundary separating two media of refractive indices n-, and n2 (Figure 2-5). The refractive index n of a medium is in turn defined by n = speed of light c in a vacuum / speed of light v in the medium Because different wavelengths have different speeds in the medium, the angle of refraction is different for each wavelength. This enables light (and other forms of electromagnetic radiation) to be dispersed into constituent wavelengths (Figure 2-6). However because the refractive index of glass varies by only a few percent across the visible spectrum, different wavelengths are separated by small angles and prisms are generally used only when low spectral resolution is required (Graham Smith and King 2000). 22 X . i Medium 1 (refractive index nj) Medium 2 r \ (refractive index n2) Figure 2-5 Principle of refraction of light. Light passing at an angle between two different transparent media is refracted (bent) according to Snell's law. Snell's law is given by m sin ;' = n2 sin r, where /'and rare the angles made by incident and refracted light rays with the normal to the plane boundary, and m and A?2 are the refractive indices of the two media. 23 Indigo Figure 2-6 Dispersion of light into constituent wavelengths using a prism. This example shows the primary spectrum for white light, although dispersion occurs for all types of electromagnetic radiation. Refractive index n is not directly proportional to wavelength X, but rather to X~nas described by the Cauchy formula (n = L + M/Xz+ N/X* +...). 24 Grating spectrographs This second type of spectrograph utilizes a device called a diffraction grating to disperse light (Graham Smith and King 2000; 2003). Diffraction is another property of light that enables it to be separated into component wavelengths. When light passes through an infinitesimally small aperture it spreads outs in a spherical fashion by a process known as diffraction (Figure 2-7). When this occurs through adjacent apertures or slits, the radiating wave fronts interfere with each other. Interference is constructive if the crests of two or more waves meet (in which case they reinforce each other and a bright fringe is seen), or destructive when waves meet out of phase (canceling each other out and producing a dark fringe). In terms of geometrical optics and as illustrated in Figure 2-8, interference is constructive when dsm0 = NA and destructive when d sin0 = 1/2 NA where d is the distance between adjacent slits (also known as the grating element or grating interval, which is fixed for a given device), #is the transmitted angle, N is an integer (also known as the order of interference), and X is the wavelength of the radiation. It can thus be seen that for a given order, different wavelengths are transmitted at different angles, i.e. sinr? oc A There are two basic types of grating: reflection and transmission. The reflection grating comprises a series of small stepped or grooved elements arranged with appropriate spacing and depth such that for a given incidence angle, the output reflections constructively interfere at a different angle for each wavelength, with the net result being that each wavelength is reflected at a different angle. In a transmission grating light is passed through a series of closely spaced slits with different wavelengths transmitted at different angles as described above. The earliest gratings were developed in 1819 by Fraunhofer, and made by winding fine wire between two screws. Later, ruling machines were used to etch lines onto glass. Current gratings are manufactured using a holographic process, by exposing photosensitive film to the interference pattern of two plane waves of monochromatic light. Holographic gratings are used in most modern day spectroscopic applications and can be made with up to 6000 lines per centimeter. Gratings are usually blazed (engraved) so that a particular order will be the most intense. 25 Gratings can be configured in a number of different arrangements within a spectrometer. A typical spectrometer setup using a reflection grating is the Czerny-Turner arrangement (Figure 2-9). This uses collimating and focusing mirrors (rather than lenses) to direct light onto the grating and subsequently the detector. The use of reflecting elements has the advantage of avoiding light attenuation due to passage through gratings or lenses. The grating is rotated mechanically to select different wavelengths within the spectrum for detection. A modification of the basic design is the Rowland circle, which uses a concave grating to diffract and focus light. This reduces losses and aberrations by avoiding the need for focusing lenses and mirrors. One problem with the Czerny-Turner arrangement is that stray reflected light can interfere with the weak Raman scattered light. To overcome this, two, or even three spectrometers can be arranged in tandem to filter out stray light. Another drawback of the Czerny-Turner spectrograph is that the micron scale features of the reflection grating makes it susceptible to dust and fingerprints, and yet the fragile nature of its surface makes it virtually impossible to clean. In addition, its bulky size (it can measure up to 60cm long) and the presence of moving parts make it unsuited for clinical use. Moving parts introduce mechanical instability and slow down the operating speed. A newer spectrometer design is the axial transmissive spectrograph (Slater, JM et al. 2001). This design uses lenses to collimate and focus light, and volume-phase holographic transmission gratings as shown in Figure 2-10. The use of short focal length lenses and absence of moving parts provides a more compact and robust system, with a 5 to 1 calibration stability advantage over the Czerny-Turner arrangement. Transmission gratings are also completely sealed in glass, making them free from contamination and easy to clean. The main drawback of the axial transmissive spectrograph is the limited wavelength range compared to that of reflective designs. However modern holographic transmission gratings have bandwidths sufficiently wide for most Raman applications (0 to 4400cm"1 at 532nm excitation). In addition, avoiding the need for scanning allows a large reduction in time when collecting a complete spectrum. The throughput of light is also claimed to be at least as good as the Czerny-Turner arrangement. When selecting a spectrometer, physical characteristics such as size, speed of operation and ruggedness need to be taken into consideration. A number of optical characteristics also need to be considered including: a. Throughput or etendue. This refers to the light gathering capacity of the spectrometer and is illustrated in Figure 2-11. Throughput is usually expressed as an F-number, defined by F = f/D Where f is the focal length of the collimating lens or mirror (i.e. that which provides parallel rays of light to the grating), and D is the lens diameter, which is related to the side length of a square 26 grating. The F-number is related to the solid angle, which increases with smaller f and larger D. If f is too small resolution decreases. To correct for this D must be large, however this requires a larger more expensive grating. It should also be noted that etendue is a global value and that the smallest value anywhere in the optical path is the etendue of the overall system. The etendue of the spectrograph should be equal to or greater than that of illumination system coupled to it (Ferraro, Nakamoto et al. 2003). b. Bandpass. This refers to the wavelength range that can be measured using a particular grating. It is affected primarily by the linear dispersion of the grating. This refers to the amount of wavelength dispersion per unit length at the detector. The dispersion of a grating in turn refers to the change in angle of diffraction per unit change in wavelength. Angular dispersion increases as the order of diffraction increases and the grating interval decreases. Dispersion is also directly related to the focal length of the focusing system. Simultaneous spectral range is defined as the wavelength range that can just fit on opposite edges of a detector array. Thus the overall detector size is another factor affecting the wavelength range. A "scanning" spectrograph can be used to increase wavelength range. However this design introduces mechanical complexities, which can be disadvantageous as discussed earlier. c. Resolution. Resolution is the closest separation between two monochromatic wavelengths that can be distinguished. Linear dispersion again affects resolution, which is also influenced by the size of the input slit, image quality of the optics, and size of the individual detector elements. • Tradeoffs are usually necessary. For example, the larger the angle of dispersion, the easier it is to separate or resolve two wavelengths that are close together. However this improvement in resolution may decrease the range unless a larger overall detector size is used. Similarly, decreasing the size of the input slit increases resolution but decreases etendue. d. Stray light. Stray light refers to light that is scattered from its intended path. This can create background interference, which can interfere with the weak Raman signal. Attenuation of stray light is usually expressed in terms of optical density (OD). This is the ratio of stray light to incident light in factors of 10. So, for example OD 4 is an attenuation of 10~4. Triple-stage Czerny-Turner spectrographs are considered to have the ultimate in stray-light performance, achieving attenuation of up to OD 12. Much of the stray light arises from the laser line and laser-line filters are often used to attenuate the laser line by 6 to 8 orders of magnitude. 2 7 Advancing wavefront Diffracted wavefronts Figure 2-7 Diffraction of light at two parallel slits. The light can be considered in terms of advancing wavefronts in accordance with Huygens' theory. Waves from adjacent slits, S i and S 2 interfere either constructively or destructively, depending on the phase difference between the two waves. 28 Figure 2-8 Geometrical diagram illustrating the conditions for constructive or destructive interference when light is diffracted through parallel slits. Constructive interference occurs when d sin 9 is an integer number of wavelengths. This integer usually represented by N, is known as the order of interference. Lower orders produce brighter fringes. When d sin 9 is VzX a dark fringe is produced due to destructive interference. 29 Illumination slit Reflection grating Detection slit Collimator mirror Focusing mirror Figure 2-9 Czerny-Turner spectrometer arrangement. 30 C C D detector Focusing lens (ft 1.4) Collimating lens (f/1.8) Holographic grating Entrance slit Figure 2 -10 Axial transmissive spectrograph. This design is produced by Kaiser Optical System Inc (Slater, JM et al. 2001). 31 Light source 7\ u Solid angle Collimating lens Ti il Slit area / \ / y Grating Figure 2-11 Throughput or etendue of a spectrometer. The amount light entering the spectrometer and reaching the grating is proportional to the solid angle. The solid angle is related to the focal length, f and diameter, D of the lens, and etendue is usually expressed as an F-number, where F = f/ D. Throughput is also proportional to slit area. 32 Interferometers The third type of device for isolating frequencies or wavelengths in spectra is the interferometer. This instrument divides light using semi-transparent mirrors, into two or more beams that travel different paths and then recombine either constructively or destructively in wavelength dependent fashion. The principal interferometers used in spectroscopy are those developed by the American physicist Michelson (1881) in an attempt to find "luminiferous ether", a hypothetical medium thought at the time to pervade all space, and by the French physicists Fabry and Perot (1896). The Michelson interferometer is used for high-resolution Fourier transform Raman spectroscopy, and is described further in Figure 2-12. The interferometer has two main advantages: Firstly, the frequency resolution of the interferometer increases with increasing length of travel of the movable mirror, and is not a function of slit width (the throughput advantage). Secondly, the entire spectrum is recorded simultaneously with one detector (the multiplex advantage). The interferometer offers high-resolution spectroscopy at longer excitation wavelengths (moving closer to the infrared reduces unwanted fluorescence signal). However movement of the mirror to scan the spectral range of interest necessitates longer signal acquisition times, making the device unsuited for in-vivo measurements. 33 D Photo-detector Figure 2-12 The Michelson interferometer. An incident beam, A from light source, S is divided into reflected and transmitted beams (B and C respectively) by a partially reflecting mirror, P also called a beam splitter. These beams are reflected from their respective mirrors, M-i and M 2 back to the partially reflecting mirror. If the total number of oscillations of the two waves during their separate paths adds up to an integral number just after combining on the partially reflecting mirror, the light from the two beams will add constructively and be directed towards the photo-detector. Certain wavelengths are thus preferentially transmitted, resulting in a visible interference pattern. Typically, one mirror is mounted on a movable carriage that can vary the length of the light path. By photo-electrically recording the intensity of the interference pattern as the mirror is moved, a spectrum can be obtained. The resulting signals contain information about many wavelengths simultaneously and a mathematical operation, called a Fourier transform, converts the recorded modulation in the light intensity at the detector into the usual frequency domain. 34 2.4 Light detectors As Raman signals are inherently weak, detection and amplification of the signal is a key challenge in Raman spectroscopy. Most early work utilized photographic detection but this involved long exposure times, and developing and examining photographic plates was cumbersome. Modern day photonic (light) detectors are of two main types: photo-emissive detectors and semiconductor detectors. Both convert light energy to electrical energy, although by different mechanisms. These will be discussed together with advantages and disadvantages. Photo-emissive Detectors In a photo-emissive detector, an incident photon strikes a solid surface, and frees an electron from it by a process known as the "photoelectric effect". The solid surface emitting electrons is termed the photo-cathode, and electrons are collected within a vacuum tube by an anode, creating an output signal. The resultant current is proportional to the rate of incidence of photons. At low light intensities detection of individual photons can be amplified using a photo-multiplier tube (PMT) (Figure 2-13). The PMT was the first modern electronic detector available for Raman spectroscopy (Ferraro 1967). Coupling the PMT with a scanning monochromator allowed a wide Raman spectral range to be measured, albeit one resolution element at a time. In modern-day Raman spectroscopy, the PMT is infrequently used, having been superceded by semiconductor detectors, in particular the charge-coupled device (CCD) detector. This is because being a single-element detector, measuring the full spectrum requires scanning across the frequency range, which is time consuming. In addition the device requires high voltages and is fragile and susceptible to damage. Finally, the performance characteristics are generally below those of modern semiconductor detectors. Two important performance characteristics of photonic detectors are "noise" and "quantum efficiency". Noise refers to unwanted signals that may obscure the already weak Raman signal. Most significant is "dark noise" or "dark current". This arises from electrons emitted due to thermal excitation and can create a signal even when no light is falling on the detector. Cooling is used to reduce dark noise but can cause unwanted condensation, necessitating periodic disassembly and drying. Other sources of detector noise include "read-out noise" due to inaccuracy of the read-out electronics measuring the number of electrons arriving at the detector; and "quantization noise", which arises during analogue to digital conversion of the read-out signal. In an ideal detector, these sources of noise should be well below the "shot noise" level, which is the square root of the number of detected photons. 3 5 The quantum efficiency of a detector is the percentage of incoming photons that are converted to detectable electrons. Quantum efficiency depends on the material used in the detector. It also varies according to the wavelength of light detected. It can be seen from the quantum efficiency curve that there is a maximum wavelength above which electrons cannot be released (Figure 2-14). Metals typically have quantum efficiencies around 10%, whereas semiconductor materials such as gallium arsenide or caesium oxide can provide quantum efficiencies of up to 30%. This can be increased further using various design modifications. 36 Figure 2-13 Diagram of a photomultiplier tube. Electrons emitted by the photocathode are accelerated to the first dynode. At each successive dynode each electron stimulates emission of several secondary electrons, resulting in amplification of the current eventually collected at the anode. 37 o CD • i - H o CD a 3 CS a 200nm lOOOnm Wavelength (nm) Figure 2-14 Quantum efficiency of a typical photo-emitter. There is a maximum wavelength beyond which photons of light cannot release electrons. Typically this is below 1000nm. 38 Semiconductor detectors In semiconductor detectors incident photons excite and mobilize electrons from the valence band (immobile, bound state) to the conduction band (mobile state) within the semiconductor. The excitation of an electron to the conduction band leaves a "hole" in the valence band. Both electron and hole are made to move (in opposite directions) through the solid by applying an external electric field, and are detected as a photo-induced current. The amount of energy required to mobilize electrons is less that that required for the electron to escape completely, as is necessary in photo-emissive detectors. Semiconductor detectors are therefore more sensitive to lower light levels. Semiconductor detectors are usually arranged in linear or two-dimensional arrays such that each detector within the array measures a different wavelength. This creates a "multi-channel" system that can detect the entire spectrum simultaneously. Multi-channel detection enables recording of the entire spectrum in the same time taken to record one wavelength with a scanning system. Array based spectrometers also offer better stability and reproducibility, as they do not have moving parts. Two main types of arrays are photodiode arrays (PDAs) and charge-coupled device (CCD) arrays. A PDA is a linear array of photodiodes manufactured on an integrated circuit board. There may be as many as 1024 elements in the array, and as it would be impractical to have a separate wired connection to each one, electronic switching is used to allow each diode to be read in turn. Typically P D A s are not sensitive enough to detect weak Raman signals. An alternative arrangement currently in wide use is the C C D . The C C D was invented by George Smith and Willard Boyle at Bell Labs in 1969. They are also used in imaging devices such as digital cameras and astronomical telescopes. C C D s are usually two-dimensional arrays of photosensors that, like a PDA, come as an integrated circuit board or semiconductor "chip". However the readout mechanism is greatly different. The three principles of a C C D are (a) photodetection, (b) charge storage and (c) charge transfer. A typical C C D is the metal-oxide semiconductor device. This device consists of three layers: a metal electrode layer, a semiconductor layer (e.g. silicon), and a non-conducting oxide layer (e.g. silicon-oxide), which separates the first two layers. Photons of light striking the semiconductor layer form mobile electron-hole pairs as described earlier. A positively biased charge in the electrode layer repulses each hole and attracts each electron. However the silicon-oxide layer prevents electrons from moving on to the electrode. Differential voltages applied through the electrode layer also ensure that electrons in the semiconductor layer are trapped within a well (or pixel), the amount of charge within each pixel being proportional to the number of photons striking that pixel. Shifting the voltage allows electrons to be transferred from one well to another. This enables the charge in each pixel to 39 be read sequentially using an output amplifier, beginning with the first row (Figure 2-15). An analogy for this is a "bucket brigade" in which buckets are passed from person to person along a line to fill a tank. Once all the pixels in the first row have been read, the entire next row is shifted to the first row and the process repeated. Current C C D s have formats of more than 1024 x 1024 pixels, with a total size of about 25 mm 2 . As with PMTs, C C D s are also subject to the limitations of noise and quantum efficiency. Cooling is required to reduce noise and is carried out using liquid nitrogen, or solid-state thermoelectric (Peltier) coolers, which can reduce the temperature of the C C D to as much as 70 to 9 0 ° C below ambient. Design modifications such as "multi-pin phasing" to prevent thermally generated electrons being drawn into the potential wells can further reduce dark noise by a factor of 50 to 200 (Sims 1994). The quantum efficiency of silicon based CCDs, reaches a maximum of around 40% within the operational wavelength range. Design modifications such as "back illumination" and "deep depletion" have been reported to increase quantum efficiency to as much as 80%, however this is usually a trade off against increased noise (Slater, JM et al. 2001). A major weakness of silicon C C D s is the hard upper limit on usable excitation wavelengths (around 1000nm), beyond which quantum efficiency drops dramatically. Although excitation wavelengths closer to the infrared (e.g. 1064nm) would reduce fluorescence background in the detected signal, these wavelengths cannot be used with most current C C D s . Germanium, Indium Galium Arsenide and Platinum silicide are alternative materials to silicon that can detect light closer to the infrared range (1000 to 2000nm). Such materials have been used in single channel detectors for FT Raman spectroscopy. However their application in C C D technology is still in the early phase of evolution (Slater, JM et al. 2001). 40 Figure 2-15 Schematic representation of the charge-transfer and readout pattern, of a two-dimensional C C D . 41 2.5 Optical sampling system The optical sampling system directs light from the excitation source to the sample, and collects scattered light from the sample for input to the spectrometer (Slater, JM et al. 2001). The sample can be coupled with the laser and spectrometer using lenses and mirrors, or by means of optical fibres. Lenses and mirrors are usually used in Raman microscopes and FT-Raman systems (and may be referred to as "direct coupling"). Optical fiber coupling is useful when remote sensing is required, for example in in-vivo Raman spectroscopy. Most of the discussion that follows will focus on fiber optic coupled sampling. The function of optical fibers is based on the optical principle of total internal reflection. An optical fiber consists of a core material having relatively high refractive index, surrounded by a cladding of lower refractive index. The radiation is totally reflected at the core-cladding interface, which allows it to be directed along the fiber. Most commercial optical fibers use silica core / silica cladding cables. Generally, fiber optic probes used in Raman spectroscopy operate in backscatter mode. This means that light collected for analysis is scattered backwards from the sample by 180° in relation to the exciting beam. Key variations include filtering techniques, and probe design for directing light from the fiber to the sample and back. Filtering Optical fibers can produce significant Raman and fluorescence signals, which may be stronger than the Raman signature of the sample itself. In order to reject these fiber emissions a "band-pass filter" is placed in the excitation arm, which allows only the excitation wavelength and not the fiber emissions to pass to the sample. The ideal band-pass filter would allow 100% throughput of a narrow bandwidth, while completely attenuating all other wavelengths. To be effective it must be placed as close to the sample as possible. Even if this "clean up" filtering is achieved perfectly, Rayleigh scatter and directly reflected laser light are still eight orders of magnitude stronger than Raman scatter, and must be rejected from the collection arm so that they do not interfere with the Raman signal of the sample. This is achieved using a "notch filter". The notch filter is placed in the collection arm close to the sample, because if the Rayleigh scattered or reflected laser light were to enter the collection fiber, it would generate further fiber emissions in the return path. The Raman scattered light can also cause this to happen, but the effect is weaker. An ideal notch filter would attenuate a narrow bandwidth (containing only the laser wavelength) by at least eight orders of magnitude, while allowing transmission of 100% of all other wavelengths. 42 A number of different filter technologies have been used including absorption filters, triple monochromators, and reflective interference filters. Colored-glass absorption filters date back to Raman's original experiments (Raman and Krishnan 1928), however most of these filters have absorption bands too broad for use in modern Raman spectrometer systems. The triple monochromator, introduced in the early 1980's can attenuate the laser wavelength by a factor of 10 1 0 and allow signals as close as a 5cm"1 Raman shift to be detected. However the devices are physically large, mechanically complex and have multiple potential sources of failure and error. Reflectance interference filters, introduced in the late 1980's are currently the preferred choice in most Raman applications (Slater, JM et al. 2001). These filters essentially reflect selected wavelengths while transmitting others. The two basic designs are dielectric and holographic. Holographic notch filters are considered to have higher rejection capability and higher Raman transmission. Commercially available notch filters can reduce the laser signal by four to eight orders of magnitude, while allowing transmission of up to 90% of the out of band signal, with Raman shifts as low as 50cm"1 (Clarke, Hanlon et al. 1987; Schoen, Sharmma et al. 1993). Fluorescence can be generated within the filter, producing an unwanted albeit usually weak background signal. Generally though, holographic filters produce fewer artifacts than dielectric filters. Probe design There are two basic types of probe design, which are termed non-imaging and imaging. The difference is that in non-imaging probes, separate excitation and collection fibers interface directly with the sample. Whereas in imaging probes, the images of the fibers are optically combined between the sample and the probe, which operates in non-contact mode. The "/V-around-1" is the simplest non-imaging probe design. Laser light is directed through a single excitation fiber, surrounded by a variable number of collection fibers, which return the weaker Raman scattered light to the analyzer. The number of collection fibers that can be used around the excitation fiber, are limited by fiber size, overall diameter of the probe, and the size of the C C D detector. Figure 2-16 illustrates a 6-around-1 design. In order to collect light efficiently, the excitation and collection cones must overlap as much as possible. A number of modifications to the basic design have been proposed to improve its performance, and some are illustrated in Figure 2-17 (Schwab and McCreery 1984; Cooney, Skinner et al. 1996; Cooney, Skinner et al. 1996; de Lima, Sathaiah et al. 2000). Angled or beveled fiber tips can skew collection cones and increase overlap and collection efficiency. Spacing devices such as sapphire or air windows can protect the delicate fiber tips and provide a constant distance between probe tip and sample. Lenses can be introduced to relay images to a remote point. This allows the probe to be confocal, enabling it to look through a transparent medium. 43 This property may be of value if a thick spacing window is required, to avoid picking up Raman spectra of the window instead of the sample. Filters can also be introduced into the excitation and collection bundles. These are typically thin dielectric filters. Although holographic filters have better performance than their dielectric counterparts, they cannot be used, being much thicker and resulting in too much loss of light. The main advantage of the "A/-around-1" probe design is that it is small, rugged, and can be used to reach remote areas. Such probes have been used to study gynecological, gastrointestinal and arterial tissue in-vivo (Mahadevan-Jansen, Mitchell et al. 1998; Bakker Schut, Witjes et al. 2000; Buschman, Marple et al. 2000; Shim, Song et al. 2000). Imaging probes tend to be larger and more complex than "/V-around-1" designs. However they generally offer better signal detection and fiber background rejection performance. In the basic design of an imaging probe, light from the excitation fiber is collimated with a lens and then passed through a band-pass filter to remove laser-induced fiber emissions. A further lens focuses the beam onto the sample. The backscattered Raman signal is collected and collimated through the same lens, and is directed via a combining element to the collection fiber or fibers. Between the combining element and the collection fiber, a notch filter is used to remove the laser wavelength. As filtering is not required for individual fibers, larger holographic filters can be used, which allow sufficient transmission of light while providing higher filtering performance. The images of the excitation and collection fibers are optically combined and therefore considered confocal, enabling the probe to "look through" a transparent medium. Smaller spot sizes can also be achieved compared to "A/-around-1" probes. A variation in the design separates the excitation and collection pathways. The excitation and collection arms are positioned at a 40° angle to each other. An additional collimating lens in the collection arm directs scattered light from the sample to the notch filter, from where it is refocused onto the collection fiber bundle. This helps to reduce collection of specularly reflected laser light (Huang, Zeng et al. 2001). 44 Protective sheath Excitation fiber Collection fiber Epoxy filler Excitation cone Collection cones Figure 2-16 Basic design of an "A/-around 1" probe. Diagram shows (a) axial cross-section of a 6-around-1 arrangement, and (b) longitudinal section indicating overlap of excitation and collection cones. 45 Secondary collection cone Sapphire window Mirrored fiber face Excitation fiber Collection fiber Dielectric notch filter WWW////////////WWW I ^^m//////////// • " • • • ^ Dielectric bandpass filter Figure 2-17 Modifications of the "A/-around-1" probe design. These include mirrored fiber faces to provide a secondary collection cone and wider overall collection field; sapphire window spacer; and filters. As described in (Slater, JM et al. 2001). 46 3 . Spectral Processing 3.1 Overview of spectral processing Raman spectra typically contain multiple peaks with a wealth of information about the physical and chemical properties of molecules. Interpretation of this information can be challenging, especially for spectra from complicated biological molecules. The past two decades has seen increasing use of mathematical and statistical methods to extract information from chemical data. These methods have developed into the field known as chemometrics. A number of processing techniques are available, which enable analysis of Raman spectra. Pre-processing techniques remove noise and standardize spectra so that comparisons can be made. A number of different spectral analysis techniques can then be used to compare spectra from different samples. The mathematics behind some of these methods forms the basis of entire textbooks and the following sections deal with the topics more on a conceptual level. 3.2 Pre-processing Pre-processing techniques may be used to achieve smoothing, background reduction and normalization (Shaver 2001; Ferraro, Nakamoto et al. 2003). Smoothing Smoothing is applied to reduce sharp peaks due to background noise, with the aim of improving signal to noise (SIN) ratio of the Raman spectrum. Poor S/N ratio results from low scattering intensities, inherent in Raman spectra of tissue. It can also result from inefficiencies of the spectrometer grating and detector. A number of methods have been applied to achieve smoothing. Examples include adjacent averaging, Savitsky-Golay smoothing and fast Fourier transform (FFT). A description of adjacent averaging provides the simplest explanation of smoothing. In this method, each point of the smoothed spectrum is an average of adjacent points in the original spectrum. An equal number of data points on each side of the target data point are included in the average. This creates a problem at the beginning and end of the spectrum. For example, in a 5-point smoothing function with two points on either side of the target point, the first two and last two points in the spectrum should be discarded. Weighted averaging can be used. For example in a 5-point weighting function, the center point may be weighted by 1.0, the two points adjacent to this by 0.75, and the next set of points on either side by 0.50. 47 Savitsky-Golay smoothing applies a polynomial curve to a given number of points on either side of the center weight. FFT applies a mathematical function that removes high frequencies, leaving somewhat broader bands. Whatever the technique applied, the main drawback of smoothing is a loss of information. It results in broadening of spectral peaks, which effectively lowers spectral resolution. The number of points to which smoothing is applied is also somewhat subjective, often decided by visualization of the spectral plot. In an ideal situation, the need for smoothing could be reduced by improving efficiency of the spectrometer/detector system. Background Reduction Raman spectra often appear with sloped or curved backgrounds. This is due to fluorescence, Rayleigh "wings" or other aberrations. It may be necessary to reduce the background to emphasize Raman peaks and enable comparisons between spectra. Several mathematical methods can be used, such as derivatives and baseline flattening. A derivative is calculated by plotting the difference between intensities of adjacent points (or between any selected spectral gap). The method reduces broad background contributions but creates positive and negative spikes, resulting in loss of the original spectral pattern. It also tends to create sharper peaks and thus emphasize spectral noise. This effect increases with the order of the derivative but can be reduced by applying a smoothing function. In baseline flattening, the baseline curve or slope is fitted with a function, which is then subtracted from the spectrum. Selecting the appropriate function is a subjective process requiring a degree of trial and error. Linear and polynomial functions with a varying number of points are often used. Other curve fitting techniques are available e.g. splines and la-grange, and are the subject of entire mathematical texts. A major drawback of baseline flattening is that Raman bands can be distorted, especially if the background slope is steep. Another drawback is that a function, which fits well to one slope, may not fit as well to another. This can make standardization difficult. As with smoothing, mathematical manipulation to reduce the background can distort spectral information. Background reduction using optical techniques may reduce this problem. Excitation wavelengths closer to the infrared produce less fluorescence (this is limited by C C D detector technology, which currently does not perform well in the infrared range). Other methods of combating 48 fluorescence have involved use of pulsed laser excitation, dc chopping, and synchronization of scattered radiation to electronically filter out background emissions (Ferraro, Nakamoto et al. 2003). Normalization Normalization converts the intensity range of different spectra to the same or similar scale, which allows spectra to be combined and compared. Spectra can be normalized to a given frequency or to the total area under the spectrum. Normalization to a given frequency involves dividing the intensity of each frequency by the intensity of the given frequency. The intensity of the given frequency becomes 1.0 in all normalized spectra. This method depends on features of a single band and may distort the normalized spectrum if that band is not consistent between spectra. One way to overcome this problem is to normalize the spectra to the total area under the spectrum, such that the area under the spectrum becomes 1.0. This method is independent of the features of any single band. A disadvantage is that background contributions can distort the normalized spectrum, and extraneous background information should be removed first. 3.3 Spectral analysis Spectral analysis techniques aim to identify unique spectral features that distinguish one compound from another, or in the case of biomedical applications, one tissue type from another. Early methods for analyzing spectra involved comparison of a few representative frequencies. Sometimes peak ratios were used. Selection of peaks was subjective, and consequently so also were the results. Some spectroscopists became proficient at recognizing spectral patterns and could readily match unknown spectra with known library spectra. These techniques were restricted because only limited spectral information was used. Current methods utilize multivariate analysis techniques to take advantage of as much spectral data as possible. The advent of the personal computer greatly facilitated application of these techniques. Multivariate analysis techniques for analyzing spectral data aims to address two problems: (1) How to reduce the dimensionality of the data with minimal loss of information, and (2) How to use the data to separate two or more groups of individuals. A number of multivariate techniques are available. However, review of existing studies applying multivariate analysis to Raman spectra, and discussion with statistical experts, led to principal components analysis being chosen as the main multivariate technique to determine the diagnostic potential of Raman spectroscopy in this study (Wolthuis, Bakker Schut et al. 1999; Stone, Stavroulaki et al. 2000; Shaver 2001; Ferraro, Nakamoto et al. 2003). The basis and value of this technique is well described (Manly 1994; Everitt and Dunn 2001) and is discussed below. A newer multivariate technique known as artificial neural networks, has also 49 been used in some Raman spectroscopy studies, and is discussed in contrast to the more conventional technique of principal components analysis. Principal components analysis (PCA) Raman spectra from different samples create a set of multivariate data. This can be represented by the matrix X, where X = \ v21 "22 Mp "2p Sil An2 -Sip J In this matrix, n is the number of individual spectra (or independent observations), and p is the number separate wavelengths at which the Raman signal intensity is measured (known as variables). In spectral data the variables are (a) typically many, relative to the number of observations, and (b) highly correlated (i.e. an increase in intensity of one peak, due to an increase in concentration of the substance, typically results in corresponding increases in intensity of other peaks). Both situations may lead to problems when applying standard regression techniques. These problems can be overcome by the multivariate technique, P C A . P C A was originally introduced by Pearson (1901) and independently by Hotelling (1933). The basic aim is to reduce the number of explanatory variables with minimal loss of information, while removing correlation between the variables. This is achieved by creating a series of new variables, or principal components, each of which is a linear combination of the original variables. The first principal component accounts for as much variation in the original data as possible, while the second principal component accounts for as much as possible of the remaining variation subject to being uncorrelated with the first component, and so on. If the first few components account for most of the variation in the original data, they can be used to summarize the data with little loss of information. A reduction in dimensionality of the data is thus achieved, which can simplify later analysis. In algebraic terms, the first principal component of the observations, yx, is the linear combination v, =anxx +aux2 +--- + a]pxp whose sample variance is greatest among all such linear combinations. In order that the variance of v, cannot be increased indefinitely by simply increasing the value of any of the coefficients au ,au • ••alp , the equation is subject to the constraint 50 an +an +... + alp = 1 In other words, the sum of the squares of the coefficients must equal 1. Multiplying each variable by its given coefficient results in each observation being represented by a single figure, which is the first principal component of that observation. The second principal component, and that v, and v 2 are uncorrelated. The third and subsequent components are derived in a similar manner such that each component is not correlated to the preceding components. If there are p variables, there can be up to p principal components, however principal components that only account for a small proportion of the variation in the data are usually discarded. Classification Analysis As discussed above, P C A reduces dimensionality of the multivariate data and removes collinearity or correlation between the variables. The problem remains how to use the principal components to discriminate between two or more groups of spectra, which in Raman spectroscopy may represent normal and pathological tissue types. A number of discrimination function analysis techniques can be used to address this problem. One approach is to use the Mahalanobis distance metric (Manly 1994; Ferraro, Nakamoto et al. 2003). The Mahalanobis distance metric is a measure of the Euclidean (geometrical) distance of an individual observation from the group mean. When several groups are present, the individual can be allocated to the group that it is closest to. This may or may not be the group that the individual actually came from. The percentage of correct allocations is an indicator of how well the groups can be discriminated using the available variables. In its simplest form, the Mahalanobis distance can be represented by the equation where DeucM is the Euclidean distance between the target spectrum and the average spectrum of the training set, runkJ is the intensity of the ith frequency of the unknown or target spectrum, raveii is the intensity of the / h frequency of the average spectrum of the training set, and the sum is over all / frequencies. If this equation is applied to the full spectrum so that the number of coordinates is equal to the number of spectral frequencies, so called over-fitting occurs, with introduction of noise as well is such that the variance of y2 is as large as possible subject to the constraint that a22 +a222 +... + a2p2 =1 51 as real information, because many frequencies are dependant or correlated. The use of P C A scores instead of individual frequencies helps to eliminate this problem. If there are n principal components then the Mahalanobis distances can be measured in n dimensions. Separation of the groups can be visualized by plotting P C A scores in two or three dimensions to create a scatter-plot. If more than three principal components are used in the analysis, then visualization becomes difficult because a scatter-plot cannot be created in more than three dimensions in space. An alternative method to enable visualization of group separation is to plot derivatives that take into account all the principal components being used. One possibility is to plot linear discriminants of the principal components. In general however, the number of linear functions that can be derived is the smaller of (m - 1) and p (where m is the number of groups being separated, and p is the number of variables). Thus for three groups a two-dimensional plot could be created, whereas for two groups, a linear plot would need to be used. Cross validation The performance of the discriminant function can be determined by applying the function to the data from which it was derived. A table can be created to calculate the percentage of correctly classified individuals Correct group Allocated group 1 2 1 A b 2 C d The percentage of correctly classified individuals can be calculated as ( a + d ) xlOOo/o (a + b + c + d) Sensitivity and specificity values can also be calculated from a two by two table. Because each individual contributes to the group mean, this method of classification tends to be biased in favor of allocating individuals to the correct group. This bias may be reduced by the so-called "cross-validation", "jack-knife" or "leave-one-out" classification methods. These involve deriving the discriminant function from n - 1 members of the sample, and then using the function to classify the member not included. The process is carried out n times, leaving out each sample member in turn. In this way each individual is allocated to its closest group without that individual being used to determine the group center. This method of classification usually gives a slightly smaller number of correct allocations than the straightforward classification. 52 Other methods for deriving classification rules The use of P C A and Mahalanobis distances has been described. The "k nearest neighbors" technique is a variation on this and has also been suggested for analysis of Raman spectra (following PCA). This technique measures the distance of an individual from its k nearest neighbors. The individual is then assigned to that group to which the majority of its neighbors belong, as determined from a training set. Another technique that could be used to classify spectra is Fisher's discriminant function, which creates linear functions of variables to maximize the between group to within group ratio of variances. Logistical discrimination is another technique that may be more robust if multivariate distributions of the populations are not normal. These techniques may however be affected by collinearity between a large number of variables. Tree-based models are an alternative technique, but are usually considered "unsupervised" in that groups to which individuals are allocated are not known in advance. Finally, newer and more sophisticated pattern recognition techniques such as artificial neural networks have been proposed. This technique, and its possible advantages and disadvantages are examined further below. Artificial Neural Networks (ANN) ANN provide an alternative means of classification analysis. ANN are highly interconnected networks of relatively simple computing elements called "neurons", that have the ability to respond to input stimuli and learn to adapt to the environment (Schmitt and Udelhoven 2001; Shaver 2001). They have their basis in artificial intelligence and pattern recognition. McCulloch a neuro-physiologist and Pitts a mathematician, are credited with pioneering the field of ANN in 1943. Neural networks have three essential components. These are: a) the basic computing elements, usually referred to as neurons, b) the network architecture, which describes interconnections between the neurons, and c) the training algorithm, which allows network parameters to be modified so that the desired and actual response of the network to a given input signal are as close as possible. In a single neuron (Figure 3-1), an input signal x,is multiplied by a weight w-,. Input signals are combined by a summation function, and an activation function is applied to limit the output of the neuron to 1 or - 1 . Thus the neuron "fires" or not, depending on whether the argument is positive or negative. The output response y can be represented by the equation 53 Individual neurons are linked to form a network. Typically the network is arranged in layers, with the output of each neuron feeding into neurons at the next level. This is termed a layered feed-forward network or multi-layer perceptron (Figure 3-2). Layers in between the input and output layers are termed hidden layers, as they are not directly observable. As an analogy, the neurons in each layer can be considered as individual specialists, each trained to recognize specific features in the data. Specialists in the input layer feed data to each specialist in the hidden layer. The specialist in the hidden layer then advises each output specialist, who in turn makes their own individual assessment on whether or not a given sample belongs to a specific class. The number of layers, the number of neurons in each layer, and the functions used in each individual neuron among other parameters, can be varied to create the best performing A N N (Shaver 2001). It is worth noting that the predictive ability of an ANN decreases proportionally with the number of layers, as errors are accumulated through the network layers. Therefore in spectroscopic applications, feed forward networks with only one or even no hidden layers are preferred (Schmitt and Udelhoven 2001). When ANN are used for classification analysis, weightings and other function parameters are determined by a process known as training. In the first step, a forward pass is performed through the ANN with a set of training samples, using random weightings. The magnitude of error between desired and actual output responses is computed and used to adjust the weightings. This procedure is called a training cycle or epoch. A new cycle is then initiated with the adjusted parameters and repeated until optimal performance and error levels are achieved (Schmitt and Udelhoven 2001). Neural networks have received a lot of attention in recent years as a promising method for discrimination and classification analysis. Part of the allure of this technique is its apparent sophistication, and an ability to learn and adapt, and therefore solve complex large-scale problems. However several comparative studies suggest that claims made for neural networks have often been exaggerated, and the performance of ANN may not be significantly better than conventional multivariate techniques (Everitt and Dunn 2001). 54 Output Neuron i Figure 3-1 Schematic representation of a single neuron in an ANN. Inputs are weighted, summed, processed by an activation function, and output to the next layer of neurons 55 Figure 3-2 Schematic representation of a multi-layer perceptron. The circles represent individual neurons. Those marked x are the input layer. Those marked z are the hidden layer. Those marked y are the output layer, which determine the class to which the input data belong. 56 4 . Applications of Raman Spectroscopy in Medical Science Raman spectroscopy has been used in numerous industrial and biological applications. In medical science its potential lies mainly in its ability to enhance medical diagnosis, although to date it remains largely a research tool. It has been applied in a number of different areas of the human body, both in vitro and in vivo. Broadly speaking there have been two types of applications. The first has been to examine biochemical characteristics of pathological tissue, for example atherosclerotic plaques, cataracts, and gallstones. The second type of application has been its use to distinguish neoplastic from non-neoplastic tissue. The different applications will be discussed according to anatomical site, in alphabetical order. 4.1 Blood vessels Raman spectroscopy has been used to study atherosclerotic plaques in blood vessels since the late 1980's (Clarke, Isner et al. 1988). The main aim of this field of research has been to identify plaques more likely to be associated with thrombo-embolic complications (Hanlon, Manoharan et al. 2000; Stefanadis, Vavuranakis et al. 2003). Early in vitro studies on human artery tissue suggest that concentrations of collagen, elastin, and cholesterol compounds can be quantified using Raman spectroscopy (Manoharan, Baraga et al. 1992; Romer, Brennan et al. 1998; Salenius, Brennan et al. 1998; Weinmann, Jouan et al. 1998). Raman spectroscopy has been used to distinguish normal arterial tissue from non-calcified and calcified plaques, with sensitivities and specificities between 80 to 100% using P C A and other statistical methods (Silveira, Sathaiah et al. 2002; Silveira, Sathaiah et al. 2003; van de Poll, Kastelijn et al. 2003). An important advance in this field of research has been the development of miniature fiber optic probes for in vivo use. 2.5 mm diameter forward viewing and 1 mm diameter side viewing probes have been described (Buschman, Marple et al. 2000). These probes use a 7-around-1 arrangement of optical fibers with dielectric filters placed close to the tip, in an arrangement similar to that described in Figure 2-17. The side-viewing probe used an additional 90° reflector, comprising a fused silica prism and gold-plated mirror, in the probe tip. This probe was small enough to pass through a 1 mm diameter stainless steel needle, and Raman signals were obtained using the probes in vivo from sheep aorta. 4.2 Brain In the brain, Raman spectroscopy has been applied to distinguish between tumors and normal or necrotic tissue (Hanlon, Manoharan et al. 2000; Koljenovic, Choo-Smith et al. 2002). Mizuno suggested that relative changes in phospholipid to protein content, and accumulation of mineral and caroteinoid pigments could be responsible for differences observed (Mizuno, Hayashi et al. 1992). 57 More recently Raman spectroscopy has been used to study Alzheimer's disease in cadaveric brain tissue (Hanlon, Manoharan et al. 2000). Differences were noted between normal brain and Alzheimer's disease, which it was thought, could relate to amyloid proteins in senile plaques found in patients with dementia. The authors proposed that in vivo application of this technique could involve using a fiber optic probe inserted through the nostril and positioned at the olfactory epithelium. 4.3 Breast Breast tissue has been studied using Raman spectroscopy since the early 1990's. Alfano et al (Alfano, Liu et al. 1991) used Fourier transform Raman spectroscopy with 1064 nm excitation. Differences were reported between normal and malignant tissue in relative intensities of the bands at 1445 and 1651 cm" 1. Frank et al (Frank, McCreery et al. 1995) studied normal and malignant breast tissue using 784 nm excitation. A shift was noted in the 1439 cm" 1 band in normal tissue to 1450 cm"1 in infiltrating duct carcinoma. Differences in the intensity ratio between 1439 and 1654 cm" 1 were also reported for normal tissue and infiltrating duct carcinoma. However differences between infiltrating duct carcinoma and fibrocystic change were less pronounced. Using an excitation wavelength of 830 nm, Hanlon et al (Hanlon, Manoharan et al. 2000) reported the use of P C A to study differences between benign and malignant breast tissue. With this method, 14 of 15 normal, 13 of 15 benign, and 31 of 31 malignant tissue specimens were correctly diagnosed. When the criteria of Alfano et al and Frank et al were applied, normal and malignant tissue could be separated easily but benign tumors could not be well separated from either group. Haka et al (Shafer-Peltier, Haka et al. 2002) used Raman spectroscopy to specifically study the chemical composition of micro-calcifications occurring in benign and malignant breast lesions. Using PCA, micro-calcifications from benign and malignant tissue could be distinguished with a sensitivity of 88% and specificity of 93%, a significant improvement over current x-ray mammography techniques. Differences were related to different concentrations of calcium carbonate and protein in micro-calcifications in benign and malignant disease. Raman spectroscopy has also been used to study metastatic axillary lymph nodes in breast cancer (Smith, Kendall et al. 2003), and the nature of foreign body inclusions in breast implant capsules (Luke, Kalasinsky et al. 1997; Centeno, Mullick et al. 1999; Pasteris, Wopenka et al. 1999). 4.4 Gastrointestinal tract The upper and lower gastrointestinal tracts have been studied using Raman spectroscopy. Manoharan studied colorectal carcinomas using UV resonance Raman spectroscopy in vitro (Manoharan, Wang et al. 1995). Spectral differences were identified corresponding to lower amino acid to nucleotide ratios, and low adenyl levels in dysplastic and carcinoma tissue. The same group used near-infrared Raman spectroscopy to study normal tissue and adenocarcinoma of the colon 58 (Feld, Manoharan et al. 1995). Increased signal intensity was reported in carcinoma at 1340, 1458, 1576, and 1662 cm* 1 , and attributed to higher nucleic acid content. More recently, the diagnostic potential of near-infrared Raman spectroscopy was assessed, both in vitro and in vivo in colonic tissue by evaluating its ability to distinguish adenomatous and hyperplastic polyps. A custom made 2mm diameter fiber optic probe, passed through the instrument channel of a standard gastrointestinal endoscope, enabled in vivo measurements to be made. Multivariate statistical techniques, including P C A and linear discriminant analyses, produced diagnostic sensitivities and specificities of 91% and 95% in vitro, and 100% and 89% in vivo respectively (Molckovsky, Song et al. 2003). In the esophagus, Wolthius (Wolthuis, Bakker Schut et al. 1999) demonstrated in vitro differences in Raman spectra from normal tissue, Barretts epithelium (metaplasia), and adenocarcinoma. (Shim, Song et al. 2000) used ANN to determine how well Raman spectroscopy could distinguish dysplasia from metaplasia, and reported a sensitivity of 73% and specificity of 93%. This group also reported the first in vivo Raman spectra of human esophageal tissue, using a similar probe as described above for the colon. Spectra were obtained with good signal to noise ratio in 5 seconds, and varying pressure on the probe tip, or the probe-tissue angle did not alter measurements significantly. In a limited set of spectra from normal and diseased tissues, only subtle differences were revealed, and the need for powerful spectra sorting algorithms was emphasized. Raman spectroscopy has also been studied in oral cavity tissues. Our group studied Raman spectroscopy of normal oral cavity tissues (unpublished data). Raman spectra of teeth showed an overwhelmingly intense peak at 955cm"1 compared to soft tissue sites. This peak corresponded to calcium compounds in bone and teeth, and illustrated how weak the soft tissue Raman signal is compared to mineralized substances. Other soft tissue sites in the oral cavity showed similar overall patterns with trends toward specific characteristics for individual sites. For example, the ratio of peaks at 1299 and 1445 cm" 1 was lower in the tongue than at other soft tissue sites. Bakker Schut (Bakker Schut, Witjes et al. 2000) studied chemically induced palatal carcinoma in a rat model. Using P C A and linear discriminant analysis, high-grade dysplasia and carcinoma-in-situ could be distinguished from low-grade dysplasia and normal tissue with a sensitivity and specificity of 1. Finally, Raman spectroscopy has also been used to analyze the chemical composition of gallstones (Paluszkiewicz, Kwiatek et al. 1998). 4.5 Gynecological tract Liu et al (Liu, Das et al. 1992) used near-infrared Fourier transform Raman spectroscopy to study benign and cancerous tissue of the cervix, uterus and ovary. Empirical analysis of the spectra 59 demonstrated differences at 1262, 1445 and 1659 cm" 1. Mahadevan-Jansen et al (Mahadevan-Jansen, Mitchell et al. 1998) used near-infrared Raman spectroscopy to study cervical tissue in vitro. Multivariate statistical methods were used to analyze spectra from 36 tissue specimens. Pre-cancers could be differentiated from normal tissues with a sensitivity of 82% and specificity of 92%. The same group described a probe that could be used to acquire spectra from cervical tissue in vivo (Mahadevan-Jansen, Mitchell et al. 1998). More recently, raft cultures have been used as a tissue model to study the Raman spectra of cervical tissue in vitro (Viehoever, Anderson et al. 2003). Such techniques will enable more detailed understanding in the future, of the cellular and biochemical bases for spectra differences between normal tissue and diseased states. 4.6 Ophthalmology Most Raman spectroscopy studies in ophthalmology have been in relation to cataract formation (Erckens, Jongsma et al. 2001). The lens is particularly suited for investigation with Raman spectroscopy because of its high protein content (Schachar and Solin 1975). Changes in protein content, sulphur bonds and hydration in relation to ageing and cataract formation have been studied (Iriyama, Mizuno et al. 1982; Ozaki, Mizuno et al. 1987). The use of Raman spectroscopy for in vivo studies in humans has been limited by the need to use only very low light levels. Nonetheless, it has been used to study the nature of intra-ocular lenses (Erckens, March et al. 2001). Raman spectroscopy has also been used to study other ocular structures. It has been used to study biochemical changes in the corpus vitreous and retina in relation to diabetic retinopathy (Sebag, Nie et al. 1994; Katz, Kruger et al. 2003) and macular degeneration (Ermakov, Ermakova et al. 2004). In the cornea, measurement of hydration and protein content using Raman spectroscopy has been suggested as a means of quality control for organ cultured corneas (Siew, Clover et al. 1995). Raman spectroscopy has also been used to quantify transport of glaucoma medication through the cornea (Bauer, Hendrikse et al. 1999). More recently the technique has been applied to study corneal refractive surgery in animals, to examine effects of corneal hydration (Fisher, Masiello et al. 2003), and to demonstrate that plastic particles persist in the cornea following micro-keratome surgery (Ivarsen, Thogersen et al. 2004). 4.7 Respiratory tract Preliminary studies suggest that Raman spectroscopy could be used to distinguish benign and malignant lesions in the upper and lower respiratory tracts. Stone (Stone, Stavroulaki et al. 2000) studied Raman spectra from laryngeal tissue using 830nm excitation, and Raman microscopy to acquire spectra. Using a combination of P C A and linear discriminant analysis, sensitivities of 83%, 6 0 76%, and 92%, and specificities of 94%, 91 %, and 90% were reported for classifying normal tissue, dysplasia and squamous cell carcinoma of the larynx respectively. Kaminaka (Kaminaka 2001) reported an exploratory study of human lung tissues by near infrared Raman spectroscopy at 1064 nm excitation. The use of this longer wavelength efficiently eliminated background fluorescence enabling detector-noise limited Raman measurements. However the photo-multiplier detector used (Hamamatsu R5509-42) required a measurement time of four hours to acquire the full Raman spectrum. More recently systems enabling signal acquisition in the order of seconds have been applied in lung tissue. Huang used 785 nm excitation and acquisition times of 5 seconds (Huang, McWilliams et al. 2003). Spectral shape differences were observed between normal and tumor tissue in the ranges 1000-1100, 1200-1400 and 1500-1700 cm" 1, which contain signals related to protein and lipid conformations, and nucleic acid C H stretching modes. The ratio of Raman intensities at 1445 to 1655 cm"1 provided good differentiation between normal and malignant bronchial tissue. Using a system similar to the one described by Kaminaka, but with multi-channel detectors, Yamazaki was able to acquire fluorescence-free Raman spectra of lung tissues within one second (Yamazaki, Kaminaka et al. 2003). 210 Raman spectra from cancerous and non-cancerous lung tissues were analyzed by a least-squares fitting procedure, which produced a sensitivity and specificity of 91% and 97% respectively for cancer prediction. Huang also examined the effects of acquiring spectra from formalin fixed tissue samples, and showed that spectra differed from fresh tissue specimens. Irrigation with phosphate buffered saline (PBS) prior to acquiring spectra could reduce the effects of formalin fixation (Huang, McWilliams et al. 2003). 4.8 Skin, Hair, Nails Skin is an ideal organ for study using Raman spectroscopy, being an external structure and easily accessible. One of the earliest reported applications of Raman spectroscopy in human skin was a study of the "Iceman', a late Neolithic man found in a glacier in 1991. Raman spectroscopy of the "Iceman's" 5200 year-old skin demonstrated degradation of the protein moiety although a largely intact lipoidal component (Williams, Edwards et al. 1995). Raman characteristics of normal skin, as well as those of hair and nail have been described (Caspers, Lucassen et al. 1998; Gniadecka, Faurskov Nielsen et al. 1998; Caspers, Lucassen et al. 2001). Variations with body site, age, hydration, and characteristics of drug absorption have also been studied (Lawson, Edwards et al. 1998; Puppels, Bakker Schut et al. 2001; Caspers, Williams et al. 2002; Caspers, Lucassen et al. 2003). Raman spectroscopy has been used to study a number of dermatological diseases in vitro and in vivo. Initial reports suggest the technique may have a role in diagnosis of benign skin conditions such as cutaneous tophi, atopic dermatitis and psoriasis (Gniadecka, Wulf et al. 2001; Wohlrab, Vollmann et al. 2001), as well as diagnosis of malignant skin conditions (Gniadecka, Wulf et 61 al. 1997; Hata, Scholz et al. 2000; Nijssen, Bakker Schut et al. 2002; Gniadecka, Philipsen et al. 2004). Recently ANN was used to determine if Raman spectroscopy could distinguish malignant melanoma from pigmented nevi, basal cell carcinoma, seborrheic keratoses, and normal skin. The sensitivity and specificity for diagnosis of melanoma achieved by neural network analysis of Raman spectra were promising, being 85% and 99% respectively. Raman spectroscopy may also aid understanding of biochemical changes associated with conditions such as vitilgo (Schallreuter, Zschiesche et al. 1998; Schallreuter, Kothari et al. 2003). Surface probes for in vivo use have been described (Huang, Zeng et al. 2001; Puppels, Bakker Schut et al. 2001). The system described by Huang et al was able to acquire a good quality in vivo Raman signal in less than one second. 62 5. Background To This Study 5.1 Rationale Raman spectroscopy can provide detailed information about molecular structure. A potentially valuable application in clinical medicine is an ability to make a tissue diagnosis, based on detection of molecular changes associated with tissue pathology. This could be applied in a number of different head and neck sites, and could reduce the need for tissue biopsy. In the nasopharynx, Raman spectroscopy would be a valuable diagnostic tool if it could distinguish cancer from normal tissue more accurately than white light endoscopy. White light endoscopy provides an illuminated, magnified view of the nasopharynx, and is the standard means of examining the nasopharynx clinically. Yet with white light endoscopy, early nasopharyngeal carcinoma can be difficult to distinguish from normal tissue, and recurrent carcinoma following radiotherapy can also be hard to detect. Multiple biopsies are often required with associated patient discomfort and morbidity. Raman spectroscopy could be used to direct biopsies in patients with suspected nasopharyngeal carcinoma. Raman spectroscopy has not been described previously in the nasopharynx. In the larynx, non-invasive real-time optical diagnosis techniques could provide an adjunct to in-clinic videolaryngostroboscopy. Examples where non-invasive identification of pathology may be of particular value include surveillance of conditions such as recurrent laryngeal papillomatosis and dysplasia. The ability to make a non-invasive tissue diagnosis may help to determine the need for examination under anesthesia and may save the patient unnecessary biopsy. Even intra-operatively, non-invasive real-time optical diagnosis may be useful to distinguish tumor from normal tissue, and to aid determination of tumor margins. . Preliminary data from another group suggests Raman spectroscopy can distinguish normal laryngeal tissue and dysplasia from squamous cell carcinoma at a microscopic level, with signals acquired over 30 seconds (Stone, Stavroulaki et al. 2000). Although the results are promising with sensitivities and specificities of 90% and above, for detection of carcinoma, shorter signal acquisition times are required for the technique to be useful in clinical practice. Till now there is also no data assessing Raman spectroscopy on benign laryngeal neoplasms. In the thyroid gland, pre-operative diagnosis of a thyroid nodule can assist with surgical planning. For example, a benign nodule can be treated expectantly or with hemithyroidectomy, whereas a malignant one will often require total thyroidectomy. Fine needle aspiration cytology (FNAC) is frequently used for preoperative evaluation but requires a skilled cyto-pathologist and cannot be 63 conducted in real time. A tool able to provide spectroscopic diagnosis in real time may be a useful adjunct to FNAC. 5.2 Hypothesis As tissue pathology involve changes in molecular composition, Raman spectroscopy could be used to detect these molecular changes and make a tissue diagnosis. 5.3 Objectives The objectives of this study were: 1. To determine whether Raman spectra could be obtained from a number of head and neck tissue sites, including the nasopharynx, larynx and thyroid gland. 2. To determine whether rapid signal acquisition was possible in these tissues, using an existing Raman probe in an in vitro setting. 3. To determine whether spectral differences could be demonstrated to distinguish normal and pathological tissue in these head and neck sites. 64 6. Materia ls a n d M e t h o d s 6.1 Tissue preparation Nasopharyngeal, thyroid and laryngeal tissue was studied. Informed consent was obtained prior to obtaining tissue specimens, using a protocol approved by the University of British Columbia Clinical Research Ethics Board. From the nasopharynx, a total of 37 tissue specimens were obtained from 17 separate patients. All were Chinese adults (12 male, 5 female), with ages raging from 27 to 64 years (mean 45 years). Specimens were obtained during biopsy for suspected primary nasopharyngeal carcinoma. Tissue samples were obtained trans-orally using 4 mm punch forceps, 2% topical lidocaine, and endoscopic guidance. After biopsy, each specimen was placed in a container with saline and surrounded with ice. Raman analysis was carried out on fresh tissue within two hours of biopsy. Although freezing allows more time before analysis, it was decided to use fresh specimens to avoid tissue artifacts, which can make histological diagnosis less accurate. From the larynx, 47 tissue specimens were obtained from 20 separate patients. There were 13 males and 7 females. Patients' ages ranged from 19 to 91 years (mean 54 years). All specimens were obtained from the membranous vocal folds or lesions involving this region. Tissue was obtained during microlaryngeal surgery (26 specimens) or immediately following laryngectomy (21 specimens). Because of the logistical difficulties of analyzing large numbers of fresh specimens, laryngeal specimens were frozen for storage before analysis. Immediately following biopsy, tissue specimens were washed with normal saline and embedded in OCT ("Optimal cutting temperature compound", Tissue-Tek, Torrance, CA), after which they were stored in a -80°C freezer until analysis. Prior to Raman analysis, specimens were thawed at room temperature and rinsed with phosphate-buffered isotonic saline solution (pH 7.4). From the thyroid, 89 tissue specimens were obtained from 16 separate patients. There were 5 male and 11 female patients, ranging from 18 to 62 years in age (mean 39 years). Tissue was obtained during partial or total thyroidectomy for nodular lesions of the thyroid gland. Tissue was frozen and prepared prior to Raman analysis using the same steps as for laryngeal tissue. Following acquisition of Raman spectra specimens were marked with toluidine blue for orientation and fixed in formalin for histological processing. Routine histology was the standard against which Raman spectra were compared. 65 6.2 Spectrometer system The excitation source was a 785 nm diode laser (Model 8530, SDL, Inc., San Jose, CA) with a maximum output power of 300 mW. The spectrograph was an axial-transmissive imaging design (HoloSpec-f/2.2-NIR, Kaiser Optical Systems, Inc., Ann Arbor, Ml). The detector was a back-illuminated, deep depletion C C D detector (LN/CCD 1024 E H R B , Princeton Instruments, Trenton, NJ). The C C D was NIR optimized with a quantum efficiency of >75% at 900 nm. It was liquid nitrogen cooled, and its S/N ratio was read-out noise limited when weak Raman signals were acquired. The laser was coupled by a 200 jum core-diameter silica fiber to a Raman probe, originally designed for in vivo studies in skin. The probe had separate illumination and collection arms. The half-inch diameter illumination arm incorporated a collimating lens, a band-pass filter (785±2.5 nm) to block spectral contributions from silica in the excitation fiber, and a focusing lens. Laser light was focused onto the sample surface with a spot size of 3.5 mm. The one-inch diameter collection arm contained a collimating lens, a holographic notch filter of OD>6.0 at 785 nm (HSG-785-LF, Kaiser Optical Systems, Inc., Ann Arbor, Ml) to attenuate the laser line while allowing passage of Raman scattered light, and a refocusing lens. The refocusing lens focused the filtered beam onto a fiber bundle consisting of fifty-eight 100 /jm fibers arranged in a circular shape of 1.6 mm diameter. The fiber bundle directed the collected light back to the spectograph and C C D detector, from where data was processed and displayed using a personal computer. The output end of the fiber bundle was arranged in a linear array. A separate 50 /jm fiber was split into the bundle and directed into the center of the linear array for wavelength calibration. The number of fibers used in the bundle was the maximum allowable by the height of the C C D detector (6.9 mm). The illumination arm of the probe was angled at 40° to the collection arm, which was angled perpendicularly to the sample surface. This was to avoid collection of specularly reflected laser light. All the optical components of the probe were mounted in a rigid case to keep the in the desired position. A schematic representation of the system is given in Figure 6-1. A detailed description of the spectrometer system is also given in the reference by Huang et al (Huang, Zeng et al. 2001). Biopsy samples were placed on aluminum foil, which produced a negligible Raman signal in our system. The epithelial surface was directed upwards and the specimen position was adjusted in relation to the excitation beam to achieve the strongest attainable signal. Raman spectra were recorded between 950 and 1,650 cm" 1, with signal acquisition times ranging between 1, 5, 10, 15, and 30 seconds. 66 C C D detector Spectro-meter Single fiber Excitation arm (l/2in diameter) Collimator Fiber bundle Desk-top computer Band-pass filter Focusing lens Collection arm (lin diameter) 4 Re-focusing lens 1 Notch filter ^ Collimator y Specimen Figure 6-1 Optical layout of the Raman system. 6 7 6.3 Spectral analysis Preprocessing A fifth order polynomial was fitted to the background tissue autofluorescence signal for each set of raw data, between 840 and 900 nm. The polynomial was then subtracted from the measured signal to obtain the Raman signal, which was plotted as intensity (arbitrary units) against wavenumber (cm"1) (Figures 6-2 & 3). The same order polynomial was fitted over the same waveband individually for each data set to correct for variations in autofluorescence between specimens. This technique has been used effectively by other groups (Mobley, Kowalski et al. 1996; Mahadevan-Jansen, Mitchell et al. 1998; Wolthuis, Bakker Schut et al. 1999), and as with other studies we found that a fifth order polynomial gave the best fit to the background autofluorescence. Five-point smoothing by adjacent averaging was then carried out to reduce noise effects (Figure 6-3). Each spectrum was then normalized to the integration area under the curve to correct for variations in absolute spectral intensity and enable comparison of relative intensities at different wavenumbers. 6 8 15000 1 1 1 1 1 1 1 1 800 820 840 860 880 900 920 940 Wavelength (nm) Figure 6-2 Example of raw Raman spectra. Note the large contribution from tissue autofluorescence, making the Raman peaks almost invisible prior to subtraction of autofluorescence. 69 1600 - i 1200 1300 1400 1500 1600 1700 Wavenumber (cm"1) T T 1400 1500 Wavenumber (cm"1) Figure 6-3 Examples of Raman spectra after subtraction of fifth-order polynomial (upper plot) and following five-point smoothing (lower plot). Subsequent normalization retains the shape of the spectrum but makes different spectra comparable. 70 Sampling variability Variations related to signal acquisition time, specimen size, and specimen orientation were studied qualitatively in a number of specimens. The effect of different spectral acquisition times on the quality of spectra obtained was studied by superimposing spectra obtained from the same patient with different signal acquisition times, and visually examining the variability between spectra. The effect of specimen size on signal quality was studied using a similar method. A single specimen was divided into large and small fragments, and spectra obtained from each fragment were superimposed, so that spectral variability could be examined visually. The effect of specimen orientation was studied by obtaining spectra from the mucosal and deep surfaces of a specimen. The spectra were again superimposed to examine signal variability. This was applied to tissue from the nasopharynx and larynx, which have mucosal and deep surfaces. Paired data Paired analysis was carried out in nasopharyngeal and thyroid gland specimens in which separate normal and cancer tissue biopsies were obtained from the same subject. Spectra from these subjects were analyzed as paired data. This was achieved by performing a paired t-test at each wavenumber in the spectrum, comparing the logi0-transformed signal intensity of normal and cancer tissue. The aim was to identify consistent trends in the differences between normal and cancer tissue at individual wavenumbers. The log10-transformed data was used to account for the possibility of a skewed distribution (Altman 1991). Wavenumbers with highly significant differences (p < 0.01) between normal and cancer tissue were identified. If these wavenumbers made up a certain range or waveband, then the log-transformed integration area of those wavebands were compared separately, using the paired t-test. The integration area was an indication of the overall signal intensity for a given waveband. Unpaired data The ability of RS to classify tissue pathology was evaluated using multivariate analysis. Several steps were applied: Step 1: Screening. One-way analysis of variance (ANOVA) was used to select wavenumbers where the difference between normal and carcinoma tissue (normal, squamous carcinoma and squamous papilloma in the larynx) reached a significance level of 0.00001. Step 2: Principal Components Analysis (PCA). PCA was applied to these selected variables, to remove correlation between them and reduce the dimensionality of the data. This created an orthogonal set of linear combinations of the wave-numbers reaching significance on ANOVA. Multiple analysis of variance (MANOVA) was used to identify the most diagnostically significant principal components i.e. 71 those showing a significant difference in value for normal and carcinoma (normal, squamous carcinoma and squamous papilloma in the larynx). Step 3: Classification Analysis. A small set of principal components, describing more than 99% of the total variance, was used to construct a classification model. The classification model was tested by the cross-validation method, with its accuracy being compared to histopathological diagnosis as the gold standard (Rencher 2002). To visualize separation of the groups, two linear discriminant functions were constructed from the selected principal components and used to create a scatter plot. The standardized coefficients of the first linear discriminant function were also compared, to give an indication of the relative importance of each wavenumber in predicting tissue diagnosis. 72 7. Resu l t s 7.1 Sampling variability Signal acquisition time Signal acquisition times of 1, 5, 10, 15 and 30 seconds were studied in specimens from nasopharyngeal, laryngeal and thyroid tissue. Similar results were found and a representative example from normal nasopharyngeal tissue is shown in Figure 7-1. Increasing the signal acquisition time resulted in stronger overall signal intensity (Figure 7-1 a). However, when the spectra were normalized and combined in a single plot, they fell within a similar intensity range (Figure 7-1 b). The greatest variability in the normalized signal intensity resulted from the spectrum with a 1 second acquisition time. This may be due to increased noise or a less accurate signal. When this spectrum was eliminated from the combined plot, the remaining spectra were almost identical (Figure 7-1 c) and increasing the signal acquisition time from 5 to 10, 15 or 30 seconds did not appear to improve the quality of the spectra significantly. 73 Figure 7-1 Effect of varying signal acquisition time on Raman signal intensity for nasopharyngeal tissue, (a) Before normalization. Longer signal acquisition times display stronger intensity, (b) After normalization. The spectrum with 1-second acquisition time has the greatest signal variability, (c) After elimination of spectrum with 1-second signal acquisition time, spectra with 5, 10, 15 and 30 second acquisition times appear almost identical. The 30-second spectrum is not illustrated to improve clarity. 74 Specimen size The effect of biopsy size on the quality of the spectra is illustrated in one patient whose tissue specimen was divided into a large fragment of 4 mm diameter and small fragment of 2 mm diameter. Tissue in this patient was obtained from the nasopharynx. Both fragments were diagnosed as benign lymphoid tissue on histology. Figure 7-2 demonstrates the differences in the spectra. The larger specimen had a higher intensity signal than the smaller one (Figure 7-2a). After normalization, both spectra displayed similar patterns. However, the smaller specimen demonstrated a more variable signal (Figure 7-2b). The spot diameter of the laser used in this study was 3.5 mm. For nasopharyngeal tissue, specimens ranged from 2 to 7 mm diameter, with a mean of 4 mm. For laryngeal tissue, specimens ranged from 2 to 6 mm diameter, with a mean of 3.5 mm. For thyroid tissue, specimens ranged from 3 to 15 mm diameter, with a mean of 8 mm. 75 CO -t—' ' c CO -Q CD CO c CD 0 > -I—' _ c o CD a: 600 400 £ 200 3.0i sma l l 1000 1200 1000 1200 large 1400 1600 1400 1600 W a v e n u m b e r (cm" 1 ) Figure 7-2 Effect of specimen size on Raman signal intensity, (a) Before normalization. The larger specimen has stronger signal intensity, (b) After normalization. Both signals display similar patterns but the smaller specimen has a more variable signal. Spectra illustrated are from nasopharyngeal tissue. Similar results were seen in laryngeal and thyroid tissue. 76 Specimen orientation In thin specimens of about 1 mm thickness, spectra from mucosal and deep surfaces tended to be very similar. Figure 7-3 illustrates spectra from the mucosal and deep surfaces of a representative patient with normal nasopharyngeal tissue. When the spectra were super-imposed they appeared almost identical. The thickness of this specimen was about 1 mm and the two surfaces were distinguishable macroscopically. The effect of varying specimen thickness was not specifically evaluated. 77 M u c o s a l sur face D e e p sur face Figure 7-3 Effect of varying specimen orientation on the Raman signal. Normalized spectra almost identical. This patient had a normal nasopharyngeal tissue biopsy, measuring 1 mm in thickness. Similar results were observed from laryngeal tissue specimens. 78 7.2 Paired analysis Paired data was available for nasopharyngeal tissue in six patients and for thyroid tissue in five patients. Nasopharyngeal tissue 33 wavenumbers were identified where the difference in signal intensity between normal tissue and carcinoma reached highly significant levels, defined by p<0.01 when comparing the log10-transformed signal intensity using a two-tailed, paired student's t-test. These wavenumbers were found to fall into five separate ranges or wavebands, namely 1297-1305, 1377-1381, 1436-1442, 1541-1555, and 1614-1626 cm"1. When the overall signal intensity at each waveband (represented by login-transformed integration area) was compared between normal tissue and carcinoma, the differences remained highly significant (i.e. p<0.01) for all five wavebands (Table 7-1). The location of the wavebands where these differences existed is indicated in Figure 7-4. The direction and magnitude of change between normal tissue and carcinoma can be appreciated from the line graphs in Figure 7-5. 7 9 (a) WavebancM 2 9 7 - 1 3 0 5 cm"1 Patient n ca ig n ig ca 1 0.01491 0 .01615 -1 .82652 -1 .79183 2 0 .01292 0 .01395 -1 .88874 -1 .85543 3 0 .01220 0 .01303 -1 .91364 -1 .88506 4 0 .01702 0 .01795 -1 .76904 -1 .74594 5 0 .01865 0 .02033 -1 .72932 -1 .69186 6 0 .01762 0 .02028 -1 .75399 -1 .69293 p=0.00105 (b) Waveband 1377-1381 cm"1 Patient n ca Ig n Ig ca 1 0 .00563 0 .00480 -2 .24949 -2 .31876 2 0 .00512 0 .00473 -2 .29073 -2 .32514 3 0 .00539 0 .00525 -2.26841 -2 .27984 4 0 .00628 0 .00556 -2 .20204 -2 .25493 5 0 .00590 0 .00507 -2 .22915 -2 .29499 6 0 .00565 0 .00488 -2 .24795 -2 .31158 (c) Waveband 1436 -1442 cm"1 Patient n ca Ig n Ig ca 1 0 .00703 0 .00859 -2 .15304 -2.06601 2 0 .00969 0 .01079 -2 .01368 -1 .96698 3 0 .00897 0 .01029 -2.04721 -1 .98758 4 0 .01080 0 .01219 -1 .96658 -1 .91400 5 0 .00799 0 .01133 -2 .09745 -1 .94577 6 0 .00872 0 .01122 -2 .05948 -1.95001 p=0.00374 80 (d) Waveband 1541-1555 cm"1 Patient n ca Ig n Ig ca 1 0.02565 0.02063 -1.59091 -1.68550 2 0.02736 0.02384 -1.56288 -1.62269 3 0.02677 0.02539 -1.57235 -1.59534 4 0.02308 0.01872 -1.63676 -1.72769 5 0.02838 0.02524 -1.54699 -1.59791 6 0.02822 0.02435 -1.54944 -1.61350 p=0.00201 (e) Waveband 1614-1626 cm-1 Patient n ca Ig n ig ca 1 0.01557 0.02089 -1.80771 -1.68006 2 0.01198 0.01394 -1.92154 -1.85574 3 0.01020 0.01223 -1.99140 -1.91257 4 0.01352 0.01873 -1.86902 -1.72746 5 0.01648 0.01966 -1.78304 -1.70642 6 0.01804 0.01981 -1.74376 -1.70312 Table 7-1 (a-e) Integration areas for normal tissue (n) and nasopharyngeal carcinoma (ca) at wavebands 1297-1305, 1377-1381, 1436-1442, 1541-1555, 1614-1626 cm"1. Ig n and Ig ca are the corresponding log10 transformed data. The p-value is obtained by performing a two-tailed, paired student's t-test comparing Ig n and Ig ca. It represents the significance level of the difference in overall signal intensity (measured by integration area), between normal tissue and carcinoma at these wavebands. 81 1200 1300 1400 1 500 1600 1 ' T— 1200 130 1 1400 1500 1600 A 1 1 T" 1200 130 ) 1400 1500 1600 1200 130 1 1400 1500 1600 j 1 -1 1 • 1 r — 1200 1300 1 1 1 ' — 1200 1300 1400 1400 1500 —' 1 1500 1600 -* 1— 1800 Wavenumber (cm-1) Normal Carcinoma Figure 7-4 Paired spectra from 6 subjects comparing normal and carcinoma tissue from the nasopharynx. Shaded columns indicate the five wavebands, 1297-1305, 1377-1381, 1436-1442, 1541-1555 and 1614-1626 cm" 1, where p-values reached greatest significance on paired t-test. 82 Logio integration area (arbitrary units) Logio integration area (arbitrary units) Logio integration area (arbitrary units) Figure 7-5 Line graphs (a to e) indicating direction and magnitude of change in overall spectral intensity (measured by integration area) at the given wavebands, between normal nasopharyngeal tissue and nasopharyngeal carcinoma. P-values were 0.001, 0.003, 0.004, 0.002, and 0.002 respectively, for wavebands 1297-1305, 1377-1381, 1436-1442, 1541-1555, and 1614-1626 cm"1. 84 Thyroid tissue In paired data from thyroid tissue, only three wavenumbers were identified where differences in signal intensity between normal tissue and carcinoma reached highly significant levels, using the same criteria as when comparing paired nasopharyngeal tissue specimens. These included a single wavenumber at 1477cm"1, and a narrow waveband between 1264-1266cm"1. When the overall signal intensity at the 1264-1266cm"1 waveband (represented by logi0-transformed integration area) for normal tissue and carcinoma was compared, the p value was no longer <0.01, although the difference remained significant (p=0.01059) (Table 7-2). The location of the wavenumbers where these differences were found is indicated in Figure 7-6. The direction and magnitude of change between normal tissue and carcinoma at the 1264-1266cm"1 waveband can be appreciated from the line graph in Figure 7-7. 85 Waveband 1264-1266 cm"1 Patient n ca ig n Ig ca 1 0.00797 0.01270 -2.09854 -1.89620 2 0.00749 0.00884 -2.12552 -2.05355 3 0.00834 0.00974 -2.07883 -2.01144 4 0.00848 0.01215 -2.07160 -1.91542 5 0.00742 0.00925 -2.12960 -2.03386 p=0.01059 Table 7-2 Integration areas for normal tissue (n) and papillary carcinoma (ca) at the waveband 1264-1266 cm"1. Ig n and Ig ca are the corresponding log 1 0 transformed data. The p-value is obtained by performing a two-tailed, paired student's t-test comparing Ig n and Ig ca. It represents the significance level of the difference in overall signal intensity (measured by integration area), between normal tissue and carcinoma at this waveband. 86 ^ 1 I ' 1 1 1 1 r— 1200 1300 1400 1500 1600 Wavenumber (cm'1) — — Normal Carcinoma Figure 7-6 Paired spectra from 5 subjects comparing normal tissue and papillary carcinoma from the thyroid gland. The shaded column indicates waveband 1264-1266cm"1, where individual wavenumbers reached p-values of <0.01 on the paired t-test. The vertical line at 1477cm"1 indicates the only other wavenumber where the p-value was O.01. 8 7 1264-1266 cm-1 o.o •1= -0.5 ro xi ro, -1.0 ro CD i ro c o ro l_ O) 0 -2.0 -2.5 J Normal Tumor D=0.011 Figure 7-7 Line graph indicating direction and magnitude of change in overall spectral intensity (measured by integration area) at waveband 1264-1266cm" 1, between normal thyroid tissue and papillary carcinoma. 88 7.3 Unpaired analysis Nasopharyngeal tissue In the nasopharynx, 11 histologically normal specimens (from 6 subjects) and 12 with undifferentiated carcinoma (from 5 subjects) were available for unpaired analysis. All these specimens were collected after 18 September 2001. Specimens collected prior to this date were not included in the unpaired analysis, as the machine had undergone several configuration changes up to this date. No further machine configuration changes were made after 18 September till the end of the study. The mean spectra from normal tissue and carcinoma are illustrated by the graph in Figure 7-8. Data screening with ANOVA identified 69 points with a significant difference between normal tissue and carcinoma (significance level at 0.05). PCA showed that the first 5 principal components described more than 99.9% of the total variance. MANOVA demonstrated a significant difference between normal tissue and carcinoma (Fapprox= 3.50, p = 0.023), and identified principal components 1, 3 and 5 as the most diagnostically significant (p = 0.006, p = 0.028 and p = 0.034 respectively) (Table 7-3). A scatter plot was created to allow visualization of separation of groups (Figure 7-9). The classification results of the model obtained by cross-validation are displayed in Table 7-4. In total, 78.3% of the cross-validated grouped cases were correctly classified using this method. The sensitivity and specificity of this technique to differentiate normal tissue and undifferentiated carcinoma in the nasopharynx was 83.3% and 72.7% respectively (Prevalence 47.8% and 52.2% respectively for normal tissue and undifferentiated carcinoma). The data used to derive these results was based on a signal acquisition time of 10 seconds. The sensitivity and specificity values were the highest achieved when compared to results derived using signal acquisition times of 5 and 15 seconds. 89 1200 1300 1400 Wavenumber (cm-1) 1500 1600 Figure 7-8 Comparison of mean spectra from normal nasopharyngeal tissue and undifferentiated nasopharyngeal carcinoma. 90 Principal Component 1 2 3 4 5 % of total variance 89.27% 9.95% 3.44% 0.21% 0.03% F value 9.43 0.48 5.57 0.39 5.11 P value 0.006 0.498 0.028 0.541 0.034 Table 7-3 Diagnostic significance and M A N O V A coefficients of the first five principal components for nasopharyngeal tissue. These account for over 99.9% of the total variance in the data for normal nasopharyngeal tissue and undifferentiated nasopharyngeal carcinoma. 91 0.002 0.003 0.004 0.005 0.006 0.007 LD1 Figure 7-9 Scatter plot allowing visualization of separation of groups for normal tissue and carcinoma of the nasopharynx. LD1 and LD2 are the first two linear discriminant functions derived from the first five principal components. 92 Raman Predicted Group Membership Pathology Normal Tumor Total Number Normal 8 (72.7%) 3 (27.3%) 11 (100%) Tumor 2 (16.7%) 10 (83.3%) 12 (100%) Table 7-4 Cross-validation classification results for normal tissue and carcinoma of the nasopharynx. 93 Laryngeal tissue Tissue specimens were grouped into 3 histological types: normal tissue, squamous cell carcinoma and squamous papilloma. Owing to difficulties in obtaining normal tissue from the membranous vocal folds, tissue with inflammatory change on histology was included in the normal group. 18 tissue specimens obtained from 5 patients were histologically normal or showed inflammatory change (and will be referred to as normal tissue). 13 specimens from 6 patients showed squamous cell carcinoma. 16 specimens from 10 patients showed squamous papilloma. Of these, 7 biopsies showed carcinoma-in-situ and 6 showed invasive carcinoma. The 13 specimens were grouped together in the statistical analysis as visual inspection and A N O V A of the spectra did not show a clear cut-off between in-situ and invasive carcinoma, based on the number of biopsies obtained. It was also felt that the presence of malignant change, rather than the degree of invasion was a more important diagnostic feature. The mean spectra from normal tissue, carcinoma and papilloma are illustrated by the graph in Figure 7-10. Data screening using A N O V A identified 86 points with a significant difference between the three tissue types (significance level at 0.00001). P C A showed that the first 5 principal components described more than 99% of the total variance. M A N O V A demonstrated a significant difference between normal tissue, carcinoma and papilloma (F a p p r 0 ) (=9.14, p<0.0001), and identified principal components 1, 2 and 5 as the most diagnostically significant. The results are listed in Table 7-5. Two linear discriminant functions constructed from the first five principal components are illustrated in the scatter plot in Figure 7-11 for visualization of separation of groups. Data points from each tissue category tend to cluster at different regions in the plot. The classification results of the model obtained by cross-validation are displayed in Table 7-6. In total, 83.0% of the cross-validated grouped cases are correctly classified using this method. A calculation of the sensitivity and specificity of this technique to differentiate normal tissue, carcinoma and papilloma is shown in Table 7-7. These results are based on collapsing the data into a 2 by 2 table when calculating the sensitivity and specificity results for each category. 94 Figure 7-10 Comparison of mean spectra from normal vocal fold tissue, squamous cell carcinoma and squamous papilloma. 95 Principal Component 1 2 3 4 % of total variance 72.3% 24.8% 1.5% 0.7% 0.3% F value 38.74 16.28 1.93 0.97 6.89 P value < 0.0001 < 0.0001 0.1569 0.3889 0.0025 Table 7-5 Diagnostic significance and MANOVA coefficients of the first five principal components for laryngeal tissue. These account for over 99% of the total variance in the data. 96 o O ' d CO o O ' d 0 Normal A Carcinoma + Papilloma , 0 0 A (SD 0 o o r o + * 0 o Q o o o A + A O d 0.0 I 0.005 I 0.010 LD1 —I 0.015 0.020 Figure 7-11 Scatter plot allowing visualization of separation of groups for normal tissue, squamous papilloma and squamous carcinoma of the vocal fold. LD1 and LD2 are the first two linear discriminant functions derived from the first five principal components. 9 7 Raman Predicted Group Membership Pathology Normal Carcinoma Papilloma Total Number Normal 16 (88.8%) 1 (5.6%) 1 (5.6%) 18 (100%) Carcinoma 3 (23.1%) 9 (69.2%) 1 (7.7%) 13 (100%) Papilloma 1 (6.25%) 1 (6.25%) 14 (87.5%) 16 (100%) Figure 7-6 Cross-validation classification results for normal tissue, squamous cell carcinoma and squamous papilloma of the vocal folds. 98 Normal Carcinoma Papilloma Sensitivity 88.8% 69.2% 87.5% Specificity 86.2% 94.1% 93.5% Prevalence 38.3% 27.7% 34.0% Table 7-7 Sensitivity and specificity of RS as a diagnostic test to classify pathology in the larynx. 99 Thyroid tissue Thyroid tissue available for unpaired analysis comprised normal tissue or multi-nodular goitre (38 specimens from 12 subjects), adenoma (7 specimens from 2 subjects), and papillary carcinoma (20 specimens from 5 subjects). Normal tissue and multi-nodular goitre were grouped together as the tissue can appear the same on histology, and the distinction is usually made on gross examination of the thyroid gland. This group will be referred to as normal tissue. As with laryngeal tissue, the thyroid specimens were kept frozen and spectroscopic readings were carried out as a batch. Readings were therefore not affected by changes in configuration of the machine. The mean spectra from normal tissue, adenoma and papillary carcinoma are displayed in the graph in Figure 7-12. At most wave-numbers the mean spectra from normal tissue and adenoma were very close. The unpaired student's t-test at each wave-number showed significant differences at only 29 points among 221 (p<0.05). M A N O V A on the first 10 principal components did not demonstrate a significant difference between mean spectra of normal tissue and adenoma (p=0.446). Individually, none of the 10 principal components showed a significant difference between the two groups either. This lack of difference may be a result of a truly small difference between the two groups, or may be because of the small sample size of the adenoma group. As it was considered more important to try and distinguish carcinoma from benign tissue, it was decided to combine spectra from normal tissue and adenoma into one group, which will henceforth be referred to as benign tissue. The mean spectra from benign tissue and papillary carcinoma are illustrated by the graph in Figure 7-13. Data screening with A N O V A identified 112 points with significant differences between benign tissue and papillary carcinoma (significance level at 0.00001). P C A demonstrated that the first 5 principal components described more than 99% of the total variance. M A N O V A showed a significant difference between benign tissue and papillary carcinoma (F a p p r o x=11.23, P<0.0001). M A N O V A on first 5 principle components demonstrated diagnostic significance only for the first principal component (PO.0001) (Table 7-8). A linear discriminant function constructed from the first 5 principle components was used to visualize separation between groups. The scatter plot of the linear discriminant function is shown in Figure 7-14. The first 5 principle components were used to construct a classification model. The classification results obtained by cross-validation of the prediction model are displayed in Table 7-9. In total, 81.5% of the cross-validated grouped cases were correctly classified. The sensitivity and specificity of this technique in differentiating benign thyroid tissue from papillary carcinoma was 86.7% and 70.0% respectively (prevalence 69.2% and 30.8% respectively for benign tissue and carcinoma). The data used to derive these results was based on a signal acquisition time of 5 seconds. 100 Figure 7-12 Comparison of mean spectra from normal tissue, adenoma and papillary carcinoma of the thyroid gland (Note: Normal tissue and multi-nodular goitre are grouped together as the distinction can be difficult to make on histology. This group is labeled as "Normalormng"). 101 Wavenumber (cm-1) Figure 7-13 Comparison of mean spectra from benign thyroid tissue (comprising normal tissue, multi-nodular goitre and adenoma), and papillary carcinoma. 102 Principal Component 1 2 3 4 5 % of total variance 97.35% 1.67% 0.45% 0.10% 0.06% F value 55.00 3.62 0.13 1.23 0.04 P value < 0.0001 0.062 0.722 0.271 0.848 Table 7-8 Diagnostic significance and MANOVA coefficients of the first five principal components for thyroid tissue. These account for over 99% of the total variance in the data. 103 Q o -- J o 4> A ° A A A A O A A A A 0 papillary benign -0.008 1 -0.006 -0.004 1 -0.002 0.0 0.002 1 0.004 I 0.006 LD1 Figure 7-14 Scatter plot allowing visualization of separation of groups for benign tissue and papillary carcinoma of the thyroid gland. LD1 and LD2 are the first two linear discriminant functions derived from the first five principal components. 104 Pathology Raman Predicted Group Membership Benign Papillary Total Number Papillary 6(30.0%) 14(70.0%) 20 (100%) Benign 39(86.7%) 6(13.3%) 45(100%) Table 7-9 Cross-validation classification results for benign tissue and papillary carcinoma of the thyroid gland. 105 7.4 Functional group assignments The Raman frequencies with the 20 highest standardized coefficients of the first linear discriminant function (LD1) are listed in Tables 7-10 to 7-12 for nasopharynx, larynx and thyroid respectively. Rankings, functional group assignments and possible biochemical correlates are also given. LD1 was used in preference to the second linear discriminant function LD2 for this analysis as it has a much higher standard deviation (2.4 for LD1 compared to 0.5 for LD2 for laryngeal tissue, 1.0 compared to <0.00001 for nasopharyngeal tissue, and 0.95 compared to <0.00001 for thyroid tissue), and can separate the groups better. 106 Wavenumber (cm-1) Rank Assignment Reference 1436 20 1438 8 1440 5 1442 3 1444 1446 1 2 C H 2 scissoring and C H 3 bending in lipids (Socrates 2001) 1448 4 1452 6 1454 9 1456 11 1499 18 1501 19 1505 17 1516 14 1518 1520 10 7 CN stretching & bending in cytosine monophosphate (dCMP), a pyramidine nucleotide (Fodor, Rava et al. 1985) 1522 12 1524 16 1526 13 1528 15 Table 7-10 Functional group assignments and possible biochemical correlates for nasopharynx tissue. Raman frequencies selected are those with the highest standardized coefficients of linear discriminant function LD1. Relevant Raman frequencies are listed in ascending order but ranked according to the standardized coefficient value. 107 Wavenumber (cm"1) Rank Assignment Reference 1210 18 CC6H5 stretching mode present in tyrosine and phenylalanine (Carter and Edwards 2001) 1324 17 1326 19 1328 14 1330 13 CH3CH2 wagging mode present in (Perno, Grygon et collagen and purine bases of DNA al. 1989) 1332 12 1334 11 1336 15 1441 16 1443 10 1445 6 1447 4 1448 2 1450 1452 5 1 CH2 scissoring and CH3 bending present in lipids (Socrates 2001) 1454 3 1456 7 1458 9 1460 8 1462 20 Table 7-11 Functional group assignments and possible biochemical correlates for laryngeal tissue. Raman frequencies selected are those with the highest standardized coefficients of linear discriminant function LD1. Relevant Raman frequencies are listed in ascending order but ranked according to the standardized coefficient value. 108 Wavenumber (cm"1) Rank Assignment Reference 1315 20 1319 8 CH3CH2 wagging mode present in collagen and purine bases of DNA (Perno, Grygon et al. 1989) 1321 5 1399 3 1401 1 1403 2 1405 4 NH & CO stretching & bending in nucleic acid bases (Perno, Grygon et al. 1989) 1407 6 1409 9 1411 11 1446 18 1448 1452 19 17 CH2 scissoring and CH3 bending present in lipids (Socrates 2001) 1463 14 1479 10 1480 7 1482 1484 12 16 CH stretching and bending in deoxyguanosine monophosphate (dGMP), a purine nucleotide (Fodor, Rava et al. 1985) 1486 13 1488 15 Table 7-12 Functional group assignments and possible biochemical correlates for thyroid tissue. Raman frequencies selected are those with the highest standardized coefficients of linear discriminant function LD1. Relevant Raman frequencies are listed in ascending order but ranked according to the standardized coefficient value. 109 8. D i s c u s s i o n 8.1 Advantages and disadvantages of Raman spectroscopy A number of techniques utilize light-tissue interactions with the aim of achieving a tissue diagnosis non-invasively, in vivo and in real time. To understand the advantages and disadvantages of Raman spectroscopy it is useful to appreciate the basis of some of these techniques, which include contact endoscopy, fluorescence spectroscopy and imaging, reflectance spectroscopy, and optical coherence tomography. In contact endoscopy, a magnifying endoscope makes contact with the tissue surface to allow direct examination of cell morphology. This technique has been used to facilitate diagnosis of laryngeal malignancies (Andrea, Dias et al. 1995; Arens, Glanz et al. 2003). However interpretation is qualitative and the operator must be trained to make a cytological diagnosis. Another drawback is that only superficial abnormalities can be detected. Also, because stable contact is needed between endoscope and tissue surface, the procedure requires general anesthesia. As the endoscope is of relatively large diameter, this technique would be difficult to apply in the nasopharynx and thyroid gland. Fluorescence spectroscopy utilizes monochromatic light to excite exogenous or endogenous fluorophores in tissue. As fluorescent light is of longer wavelength than the excitation light (dictated by Stoke's law), fluorescence can be detected by filtering out the excitation wavelength. Application of an exogenous fluorophore that localizes to tumor cells can enable tumor localization by detection of fluorescence. Exogenous fluorophores can be administered intravenously or topically. Intravenous fluorophores can result in generalized skin photosensitization, which limits this technique (Profio and Balchum 1985). Topical application of fluorophores, such as 5-aminolevulinic acid or 5-ALA can reduce photosensitization and have been studied in the detection of malignancy in the oral mucosa (Kennedy, Pottier et al. 1990; Leunig, Mehlmann et al. 2001; Betz, Stepp et al. 2002). Auto-fluorescence spectroscopy is another type of fluorescence spectroscopy that detects endogenous fluorophores. This technique avoids use of drugs and their associated cost and side effects (Schantz and Alfano 1993; Harries, Lam et al. 1995; Dhingra, Perrault et al. 1996; Gillenwater, Jacob et al. 1998; Betz, Mehlmann et al. 1999; Delank, Khanavkar et al. 2000). In the oral cavity, sensitivity and specificity of as high as 88 and 100% respectively have been reported for distinguishing neoplasia from normal tissue (Gillenwater, Jacob et al. 1998). In the larynx, sensitivity and specificity values around 90% have been reported for distinguishing normal from abnormal tissue (Zargi, Fajdiga et al. 2000; Eker, Rydell et al. 2001; Malzahn, Dreyer et al. 2002; Arens, Dreyer et al. 110 2004), however the specificity may be as low as 50% when using the technique to distinguish malignant from benign laryngeal lesions (Delank, Khanavkar et al. 2000). Nasopharyngeal tissue has also been studied by autofluorescence, using the Lung Imaging Fluorescence Endoscopy (LIFE) system originally developed to detect bronchial carcinoma (Hung, Lam et al. 1991; MacAulay, Lam et al. 1995). The LIFE system allows autofluorescence to be displayed as a video image as well as spectroscopically. Differences between normal tissue and tumor in the nasopharynx were not very marked (MacAulay, Lam et al. 1995). There are other controversies in autofluorescence spectroscopy. For example, fluorescence in the red part of the spectrum is associated with oral neoplasia, but could relate to mucosal ulceration or bacterial porphyrins, rather than tumor cells per se. Reduced fluorescence in the blue and green part of the spectrum has also been associated with tumors but may in fact reflect attenuation of the signal from submucosal fluorophores due to mucosal thickening in tumors (Koenig, McGovern et al. 1996), or increased perfusion and light absorption by hemoglobin (Betz, Mehlmann et al. 1999). In reflectance spectroscopy, broadband visible light is directed onto the tissue surface and the intensity of backscattered light is measured over a spectrum of wavelengths. Backscattering is dependent on the excitation wavelength and physical characteristics of the scattering surface such as refractive index and particle size. This enables information to be obtained about morphological features such as tissue thickness or even nuclear size (the nucleus can be considered a scattering particle). In the diagnosis of epithelial malignancies, information must be obtained from the superficial epithelial layers. However most backscattered light arises from beneath the epithelium, to a depth of 700um (Sokolov, Drezek et al. 1999), with only a small component from the epithelial layers. A number of techniques have been used to eliminate the sub-epithelial tissue signal. These include analysis of wavelengths such as 420, 540 and 580nm, which are absorbed by hemoglobin in deeper tissues and thus are more representative of superficial structures (Nordstrom, Burke et al. 2001). Modeling techniques have also been used to eliminate predicted contributions from deeper layers. More recently, polarized light has been used, which alters its polarity after multiple scattering in the deeper layers; light scattered from more superficial epithelial layers undergoes significantly fewer scattering events and can be detected using polarizing filters (Sokolov, Drezek et al. 1999). Preliminary studies using reflectance to distinguish normal, dysplastic and cancer cells in the oral cavity suggest promising results but the diagnostic ability of reflectance spectroscopy for this purpose remains to be evaluated fully (Backman, Wallace et al. 2000). Optical coherence tomography (OCT) is a more recent extension of the principles of reflectance spectroscopy. In O C T , infrared light is focused onto tissue, and the time delay of light reflected from the internal microstructure at different depths is measured by an interferometer (Fujimoto, Brezinski et al. 1995; Brezinski, Tearney et al. 1998; Fujimoto, Bouma et al. 1998; Fujimoto, Pitris et al. 2000). By 111 scanning the beam across the tissue, a two-dimensional map of the reflectance from internal architecture and cellular structures is built up. The technique is analogous to ultrasound imaging except that infrared light instead of acoustic waves is used. Imaging resolution of up to 10|im can be achieved, which is equivalent to low power light microscopy and ten times higher than currently available radiological diagnostic imaging. Probes can have diameters as small as 1mm and images can be acquired rapidly, close to real time, to depths of 2 to 3mm (Fujimoto 2003; Herz, Chen et al. 2004). While the technique may be useful in detecting micro-structural abnormalities, its ability to distinguish different types of pathology remains to be well established (Pantanowitz, Hsiung et al. 2004). The main advantage of Raman spectroscopy over other techniques is that it can provide information about molecular structure rather than physical or morphological characteristics alone. It therefore has the potential ability to detect tissue pathology at a more fundamental level, and does not depend on operator skill or subjectivity. Its main disadvantage is that the Raman signal from soft tissue is weak and requires sophisticated technology for detection. In addition, autofluorescence, which is also induced by molecular excitation, can obscure the weak Raman signal. A major challenge in biomedical applications of Raman spectroscopy is how to eliminate background autofluorescence without distorting Raman spectral information. While mathematical subtraction is a commonly used technique, this inevitably introduces error. Future work needs to focus on optical techniques to reduce autofluorescence. Being wavelength dependent, autofluorescence is reduced at longer excitation wavelengths, which is why near-infrared (NIR) light has superceded visible light for Raman excitation in biomedical applications. However autofluroescence is not completely eliminated at these wavelengths. Longer wavelengths may further reduce background autofluroescence but create detection difficulties with current CCD detector technology. Research is needed to advance detector technology and this issue is addressed further in section 8.4 on "Proposals for future studies". NIR light has an added advantage in Raman spectroscopy in that absorption is minimal resulting in tissue penetration in the order of millimeters. This creates a "diagnostic window", which is of possible value in diagnosing submucosal lesions (Hanlon, Manoharan et al. 2000). Our finding of equivalent spectra at opposite sides of a 1mm thick specimen is consistent with this. 8.2 Interpretation of results and clinical applicability For Raman spectroscopy to be applicable in a clinical setting, several criteria must be met: First, the signal must be acquired rapidly with good signal to noise (SIN) ratio. Second, the technique must distinguish pathological and non-pathological tissue with adequate sensitivity and specificity. Third, the device must be able to access the tissue concerned. 112 To maximize detection of the weak tissue Raman signal, most experimental studies have used signal acquisition times ranging from 30 seconds to several minutes (Mahadevan-Jansen, Mitchell et al. 1998; Bakker Schut, Stone et al. 2000; Hanlon, Manoharan et al. 2000; Stone, Stavroulaki et al. 2000). Shorter times are required to make the technique applicable clinically. One study described the acquisition of spectra within 5 seconds from esophageal tissue but did not discuss the effect of varying the acquisition time (Shim, Song et al. 2000). The system described in our study was able to acquire a signal with good S/N ratio within 5 seconds. Although the signal intensity was higher at 10, 15 and 30 seconds, spectral quality did not appear significantly better after normalization. It is necessary to note that these findings were based on qualitative descriptions and that quantitative methods of evaluating S/N ratio would be preferable in future studies. With a signal acquisition time of several seconds, steady contact between probe and tissue needs to be maintained. This would be feasible in the nasopharynx even in an awake subject. However the larynx is considerably more sensitive and general anesthesia would be necessary to maintain contact between probe and tissue in almost all human subjects. Possible applications in the larynx in this situation include intra-operative tissue diagnosis and determination of tumor margins during laryngeal surgery. Further reduction to sub-second signal acquisition times may enable RS to be applied using "non-contact" mode in the larynx in awake subjects. This potentially could be done in conjunction with fiber-optic or rigid videolaryngostroboscopy. In the thyroid gland contact between probe and tissue could potentially be maintained for several seconds even in an awake patient if trans-cutaneous access to the gland could be achieved via a probe small enough (ideally less than a millimeter) to avoid excessive tissue trauma. The issue of probes is discussed further in section 8.4. Prediction sensitivity and specificity values for distinguishing pathological and non-pathological tissue were 83.3% and 72.7% respectively in the nasopharynx, 69.2% and 94.1% in the larynx, and 86.7% and 70.0% in the thyroid. In addition in the larynx, squamous papilloma could be distinguished from normal tissue and carcinoma with sensitivity and specificity of 87.5% and 93.5% respectively. While these figures do not equal that of histopathology, they are sufficiently high to warrant further study of the technique. Several possible sources of error could be addressed to improve diagnostic ability: The first is "system noise", which could be reduced by using more advanced components for example more sensitive detector systems in the future. The second source of error is mathematical data manipulation. Background reduction, smoothing, normalization and multivariate processing all introduce possible error into the analysis. Of these, background reduction to eliminate autofluorescence is likely to be the greatest culprit. Ideally, optical rather than mathematical techniques should be employed. Some of the challenges surrounding this are alluded to in section 8.1 and discussed further in section 8.4. A third source of error relates to the nature of the tissue sample and sampling technique. Tissue in-homogeneity can introduce inconsistencies to the signal. For 113 example, if a tissue specimen is reported as normal on histology, there have to be no malignant cells seen. However if a tissue specimen is reported as carcinoma, islands of normal cells may still exist in varying degrees. When the two-dimensional scatter plot in Figure 7-11 is examined it is seen that three carcinoma specimens were misdiagnosed as normal using Raman spectroscopy. Of these, two comprised carcinoma-in-situ and one had evidence of invasion. Similar information is not available from nasopharyngeal and thyroid tissue. Finer spot size to sample smaller tissue area or measuring Raman spectra at a microscopic level could reduce inaccuracies due to in-homogeneity. This study also demonstrates that different biopsy sizes can introduce variability in the signal. This effect is reduced but not eliminated by normalizing the spectrum. Biopsies are inevitably of different sizes in vitro but more uniformity could be achieved with a smaller excitation spot size to ensure that the spot is smaller than the biopsy specimen. In vivo studies could also help overcome signal variability due to different biopsy sizes. Genetic changes not yet reflected phenotypically may also account for signal differences. Tissue appearing benign on histology may be in the process of carcinogenesis having undergone changes in gene expression. These more fundamental changes may be reflected in the Raman spectrum (Peticolas, Patapoff et al. 1996). The study by Stone (Stone, Stavroulaki et al. 2000) using NIR Raman spectroscopy and multivariate techniques to detect laryngeal malignancy, demonstrated similar specificity as our study (90% v 94%) but higher sensitivity values (92 v 69%). Background reduction was not performed in Stone's study. This may have reduced errors introduced by mathematical manipulation, but may also mean that the autofluorescence signal was a major contributor to the results achieved. An additional difference is that microscopic Raman spectroscopy was used, which may have reduced effects of tissue in-homogeneity as discussed earlier. There are no comparable studies on nasopharyngeal or thyroid tissue. Addressing these sources of error is likely to improve diagnostic accuracy and clinical relevance of Raman spectroscopy. In addition larger sample sizes may also improve sensitivity and specificity values. However, ability to access the tissue concerned is a further challenge in the head and neck region, and specially designed probes will be required. This is addressed in detail in section 8.4. 8.3 Biological correlates Assignment of a Raman frequency or frequencies to a functional group of atoms is achieved using techniques such as chromatographic separation and mass spectroscopy. For biological macromolecules this process can be difficult as these molecules are large and complex. Nonetheless, functional group assignments for certain biological macromolecules are known. 1 1 4 Possible functional group assignments and biochemical correlates for wavenumbers exhibiting differences between pathological and non pathological tissue on multivariate analysis are listed in tables 7-10 to 7-12 of the results section. Many of the wavenumbers exhibiting differences are in close proximity to each other, and can be summarized into a series of wavenumber bands. These are listed in Table 8-1 for the three tissue sites. Possible functional group assignments and biochemical correlates are determined from tables and other previously reported data. The selected functional group assignments are known to be associated with biological molecules However the process remains empirical as functional groups are often represented by a band rather than a single wavenumber, and bands from different functional groups overlap considerably, with a given wavenumber representing up to 20 known functional groups (Socrates 2001). The process of assigning functional groups is complicated further because even simple molecules are usually identified by a number of peaks or an overall spectral pattern rather than a single wavenumber or wavenumber band. This process becomes even more complex in a large biological macromolecule. Although peak assignments can be made in regions of the spectrum where differences exist between normal and pathological tissue, the exact relationship between these spectral regions and the cellular or molecular changes occurring remains incompletely understood. More detailed studies at a cellular or molecular level will be necessary to increase our understanding in this area. At present it probably more accurate to consider the spectrum as a whole, in the form of a "molecular signature". Despite drawbacks in analyzing individual wavenumbers, it is worth noting that other authors have associated certain Raman peaks and bands with malignancy (Liu, Das et al. 1992; Mahadevan-Jansen, Mitchell et al. 1998; Stone, Stavroulaki et al. 2000). Some of these peaks and bands have been associated with different DNA content in normal and pathological tissues, although changes in composition of other biochemical compounds such as lipids has also been implicated. Relevant studies identifying Raman peaks or bands corresponding to those observed in this study and are listed in Table 8-1. Other studies have not attempted to assign relative importance to the different peaks in terms of predicting pathology. In this study wavenumbers were ranked by identifying the highest 20 standardized coefficients of LD1 of the first five principal components. Using the technique in the larynx, three spectral regions appeared more important in predicting laryngeal pathology. These included the 1210 cm"1 peak and wavebands 1324-1336 cm"1 and 1441-1462 cm"1 (Tables 7-11 and 8-1). In the only other RS study of the larynx, Stone et al identified peaks in each of these three spectral regions but did not attempt to assign relative importance to them. Similar comparative data is not available for nasopharyngeal or thyroid tissue. When wavebands are compared for the three tissue types (nasopharynx, larynx and thyroid), it is interesting to see that only the band at 1436-1465cm"1 demonstrates differences between normal and pathological tissue for all three tissues. This band was also implicated by Liu (Liu, Das et al. 1992) in 115 gynecological tract cancers and may represent C H 2 scissoring & C H 3 bending in lipids (Socrates 2001). The band at 1315-1336cm"1 was found in both larynx and thyroid tissue and has also been implicated by other authors (Liu, Das et al. 1992). This band may relate to C H 3 C H 2 wagging in collagen & purine bases of DNA (Perno, Grygon et al. 1989). The four remaining bands (1210, 1399-1411, 1479-1488 and 1499-1528cm"1 were each seen in only one tissue type. This may be a reflection of differences in cellular composition and structural morphology of the three tissue types. Results of both paired and unpaired analysis were available for nasopharyngeal and thyroid tissue. Wavenumber bands demonstrating differences between normal and pathological tissue could be expected to be similar in paired and unpaired analysis. However comparison showed only partial overlap. In paired nasopharynx data, five bands were identified of which only one overlapped with the two bands from unpaired analysis (i.e. less than 30% overlap). In paired thyroid data, two bands were identified of which one overlapped with the four bands from unpaired analysis (i.e. just over 30% overlap). The small degree of overlap could reflect variability in absolute peak intensities between specimens. It may be that pathological changes are better detected by observing relative differences in peak intensity between normal and abnormal tissue in the same subject. The question could be addressed further by subjecting a larger number of paired specimens to paired t-test as well as multivariate analysis. 116 Waveband (cm"1) Possible functional assignment Other studies implicating this waveband in tumor Nas Lar Thy 1210 C C 6 H 5 stretching in tyrosine & phenylalanine (Carter and Edwards 2001) • 1315-1336 C H 3 C H 2 wagging in collagen & purine bases of DNA (Perno, Grygon etal. 1989) (Liu, Das et al. 1992) (1240-1330cm-1; gynecological tract) (Stone, Stavroulaki et al. 2000) (1336cm'1; larynx) • • 1399-1411 NH & CO stretching & bending nucleic acid bases (Perno, Grygon et al. 1989) • 1436-1463 C H 2 scissoring & C H 3 bending in lipids (Socrates 2001) (Liu, Das et al. 1992) (1445cm"1; gynecological tract) • + • • 1479-1488 CN stretching & bending in deoxyguanosine monophosphate, a purine nucleotide (Fodor, Rava et al. 1985) (Mahadevan-Jansen, Mitchell et al. 1998) (1480cm"1; gynecological tract) • + 1499-1528 CN stretching & bending in cytosine monophosphate (dCMP), a pyramidine nucleotide (Fodor, Rava etal. 1985) • Table 8-1 Comparison of wavebands and functional group assignments in different tissues where differences were identified between neoplastic and non-neoplastic tissue. Nas = Nasopharynx, Lar = Larynx, and Thy = Thyroid. Black dot indicates that a difference was noted. Plus sign indicates that difference was also observed in paired data. Studies in which wavenumbers or wavenumber bands associated with malignancy corresponded with those observed in our study are also listed. 117 8.4 Proposals for future studies Some areas requiring future study include (1) Improving diagnostic accuracy, (2) Understanding the cellular basis for spectral changes, and (3) Development of probes for in vivo application. Improving diagnostic accuracy A key area that needs to be addressed is how to eliminate the background tissue fluorescence signal. A number of mathematical and optical techniques attempt to achieve this. In this study we subtracted a polynomial function of the background curve. This technique has been used successfully by other authors (Mobley, Kowalski etal. 1996; Mahadevan-Jansen, Mitchell et al. 1998; Wolthuis, Bakker Schut et al. 1999) but nonetheless can create some distortion of spectral information. Background reduction using optical techniques may reduce this problem and methods such as pulsed laser excitation, dc chopping, and synchronization of scattered radiation to electronically filter out background emissions have been suggested (Ferraro, Nakamoto et al. 2003). Excitation wavelengths closer to infrared produce less fluorescence, which is why NIR excitation is typically used in biomedical applications. Even longer wavelength light, for example above 1000nm may reduce this problem further although currently is limited by CCD detector technology, which does not perform well in the infrared part of the spectrum. The development of more versatile detectors is an area requiring further research. A recent interesting development has been the use of 1064nm excitation with a multi-channel detector described by (Yamazaki, Kaminaka et al. 2003). The authors claim that this system is able to obtain a fluorescence free signal in under 1 second with good signal to noise ratio. The detector, being a multi-channel photomulitplier device, was not subject to the constraints of current CCD detectors using higher wavelength excitation. Another consideration in dealing with fluorescence is to include the fluorescence signal in the analysis rather than eliminate it. Combining the two modalities of Raman and autofluorescence could improve diagnostic accuracy. However mathematical separation of the signal would still need to be achieved, with its attendant problems. Aside from the fluorescence issue it may be useful to compare results from Raman, fluorescence and reflectance spectroscopy on the same specimens, and to assess the effects of combining modalities on diagnostic accuracy. Even with significant reduction in background autofluorescence the spectrum is still subject to noise. In our system, signal analysis was limited to a range between 1200 and 1650 cm"1. This was in part related to noise outside of this range. The spectrum measured in our study only corresponds to a 118 small part of the overall emission spectrum for biological tissue. Hardware refinements including improved filtering may help to broaden the analyzable spectrum. Broader waveband detection may avoid elimination of potentially valuable signal peaks, and could further improve diagnostic sensitivity and specificity. Cellular basis for spectral differences Further work is required to determine more accurately the origin of the Raman peaks in pathological and non-pathological specimens. Using a micro-spectrometer, work could be carried out at a cellular level (Ferraro, Nakamoto et al. 2003). Studying spectra from individual cells has advantages over studying gross tissue specimens in that it reduces problems from in-homogeneity of the gross specimen. At a light microscopy level, it may also be possible to study spectra for example specifically from the nucleus. This may help to narrow down the origin of specific Raman peaks. Cell lines of specific tissue types could provide useful material for this type of study. While Raman peaks ultimately arise from specific functional groups, it may be more practical to identify spectral patterns at a cellular or organelle level. Development of probes for in vivo application of Raman spectroscopy in the head and neck Raman spectroscopy has the potential to be used in an in vivo setting, either as a tool for early diagnosis, or as a guidance method during intervention and treatment. In early diagnosis, Raman spectroscopy may detect molecular changes that precede changes in cellular morphology and tissue structure during disease processes. It can do so without the need for dyes, labels, radiation or destruction of tissue and is thus highly suited for in vivo use. In intervention it can be used to guide biopsies and excision margins, or possibly identify different types of tissue during surgical procedures. The three tissues studied (nasopharynx, larynx and thyroid) illustrate different requirements for Raman systems to be applied in vivo. The requirements relate mainly to the mode of access to the region concerned and the signal collection times. a. Nasopharynx The clinical standard for diagnosis in the nasopharynx, is white-light endoscopy and where necessary biopsy. Endoscopy can be carried out trans-nasally under topical anesthesia, using either a flexible or rigid instrument up to 4 mm in diameter. When biopsy is required, appropriate forceps can be passed parallel to the endoscope either through the same or opposite nasal cavity. An alternative approach is trans-oral, using a right-angled viewing rigid endoscope and curved forceps. 119 In the nasopharynx, a potential application of Raman spectroscopy is to guide biopsies of suspicious lesions in nasopharyngeal carcinoma. A Raman-guided biopsy would require three instruments to be passed into the nasopharynx simultaneously: namely an endoscope for visualization, a Raman probe, and biopsy forceps. To facilitate this process and minimize patient discomfort, these instruments should be combined if possible, with the overall diameter being about 4 mm. Two possible methods of achieving this are (a) To design a fine flexible Raman probe that could be passed through a 2 mm instrument channel in a standard fiber-optic endoscope. Biopsy forceps could be passed separately. Adaptations of the "n around 1" design that has been used in the gastrointestinal tract (Shim, Song et al. 2000) could be applied to this situation, (b) Probe and forceps could be combined in one unit, while endoscopic visualization is carried out separately using a standard fiber optic nasal endoscope. Raman guided biopsy forceps have been proposed for gastrointestinal tract application (Bohorfoush 1996; Puppels, Bakker Schut et al. 2001) (Figure 8-1) and the concept could be adapted for use in the nasopharynx. In both these situations, the ability of the probe to make relatively stable contact with the nasopahryngeal mucosa would enable the signal to be acquired over several seconds. The main challenge would be to achieve miniature instrument size while maintaining sufficiently high signal to noise ratio. 120 Figure 8-1 Artist's impression of a Raman guided biopsy forceps (As described in (Puppels, Bakker Schut et al. 2001). 121 b. Larynx The larynx provides a different set of challenges to probe design. In the outpatient setting the larynx can be examined trans-nasally using a flexible fiber optic endoscope (less than 4 mm diameter). It can also be examined trans-orally using a 70 or 90-degree angle-vision magnifying rigid endoscope (usually about 1 cm in diameter). In both situations no contact is made with the larynx, as this often results in violent coughing and can cause laryngospasm. Spectroscopy of the vocal cords in the outpatient setting could be useful to distinguish benign and malignant lesions and determine the need for biopsy under general anaesthesia. However this would need to be carried out in non-contact mode, with the instrument about 2 to 3 cm above the vocal cords. A proposed design showing the optical features that could be incorporated into a rigid endoscope for trans-oral diagnosis, in given in the diagram (Figure 8-2). We are not aware of any other designs for this application. A potential problem is that the signal path may be unstable, as the probe cannot make contact with the vocal cords. To minimize effect of variations in vertical distance between probe and vocal cords, a long focal length focusing lens is necessary (the distance between probe and vocal cords can vary by approximately 1 to 2 cm). As there is no stable contact between probe and sample, signal acquisition times need to be especially short, in the sub-second region. However in comparison to the narrower diameter nasopharyngeal probe, the larger diameter optical components in the proposed laryngeal probe may help to compensate for this. A different situation in which a laryngeal probe could be applied is in contact mode during microlaryngoscopy under general anaesthesia. In this procedure, a rigid tubular endoscope is passed through the patient's mouth, and contact can be made through this with the vocal cords. A contact probe similar to that described in the nasopharynx could be used in this situation, although it is worth noting that the area of the vocal cords is relatively small (they may measure only 2 cm in length) and the contact area would need to be less than 1 to 2 mm to be useful. 122 Figure 8-2 Optical features of proposed laryngeal imaging Raman probe (LIRAP). (a) Collimating lens, (b) Band-pass filter (dichroic mirror), (c) Reflecting prism, (d) Focusing lens (long T'for non-contact imaging; angled at 70 degrees to the vertical), (e) Mirror, (f) Notch filter (g) Lens for focusing light onto collection optic fiber, (h) & (i) Parallel optical system for direct visualization (shown in a transverse section of the probe), (j) cylindrical stainless steel casing. The probe measures approximately 1 cm in diameter and measures about 15 cm long, to allow it to be placed in the subject's oropharynx to enable imaging of the larynx. 1 2 3 Thyroid Spectroscopy of the thyroid gland represents yet a different set of scenarios. In the outpatient setting, fine needle aspiration using an 18-guage needle can be used to provide cytological diagnosis of thyroid lesions. This requires time and the services of a skilled cytopathologist. Raman spectroscopy has the potential to provide a real time online diagnosis and could be helpful to determine the need for further intervention where cytopathology services are not readily available. A narrow diameter (1 to 2 mm) needle based probe would be necessary for this application. The larger the diameter of the needle the greater is the risk of hematoma formation and patient discomfort. Needle based probes have been proposed for use in the diagnosis of intra-arterial atherosclerotic plaques (Buschman, Marple et al. 2000). Spectroscopic diagnosis may also be of value in the intra-operative setting. Diagnosis of a malignant lesion may determine the extent of surgery. Currently, frozen section histology is used for this purpose, although Raman spectroscopy could provide an adjunct to this. Real time on-line diagnosis could also be useful to identify important structures such as the parathyroid glands and the laryngeal nerves, to avoid inadvertent injury during surgery. A rigid contact probe similar to that described in the nasopharynx and larynx sections could be used. 124 8.4 Conclusions The study suggests that Raman spectroscopy can detect differences between neoplastic and non-neoplastic tissue, in the nasopharynx, larynx and thyroid gland. At present the diagnostic sensitivity and specificity does not equal that of histopathology, but nonetheless is sufficiently high to warrant further study of this technique. The system described could acquire spectra in vitro from the tissues studied within several seconds and has potential for in vivo application. Further research into Raman spectroscopy as a means of non-invasive tissue diagnosis is warranted. Areas requiring further development include evaluation of a wider spectral range, studies of Raman spectroscopy at a cellular and biochemical level, and development of probes for in-vivo studies. 125 R e f e r e n c e s (2003). Spectroscopy, Encyclopaedia Britannica. 2003. Adar, F. (2001). Evolution and revolution of Raman instrumentation - Application of available technologies to spectroscopy and microscopy. Handbook of Raman spectroscopy. From the research laboratory to the process line. H. Edwards. New York, Marcel Dekker Inc: 11-40. Alfano, R. R., C. H. Liu, et al. (1991). "Human breast tissues studied by IR Fourier transform Raman spectroscopy." Lasers Life Sci 4: 23-28. Altman, D. (1991). Practical statistics for medical research, Chapman & Hall. Andrea, M., O. Dias, et al. (1995). "Contact endoscopy during microlaryngeal surgery: a new technique for endoscopic examination of the larynx." Ann Otol Rhinol Larynqol 104(5): 333-9. Arens, C , T. Dreyer, et al. (2004). "Indirect autofluorescence laryngoscopy in the diagnosis of laryngeal cancer and its precursor lesions." Eur Arch Otorhinolaryngol 261(2): 71-6. Arens, C , H. Glanz, et al. (2003). "Compact endoscopy of the larynx." Ann Otol Rhinol Larynqol 112(2): 113-9. Asher, S. (1993). "UV resonance Raman spectroscopy for analytical, physical and biophysical chemistry. Part 1." Anal Chem 65: 59A-66A. Asher, S. A. (1993). "UV resonance Raman spectroscopy for analytical, physical, and biophysical chemistry. Part 2." Anal Chem 65(4): 201A-210A. Backman, V., M. Wallace, et al. (2000). "Detection of preinvasive cancer cells." Nature 406(6791): 35-36. Bakker Schut, T. C , N. Stone, et al. (2000). "Progress in the detection of neoplastic progress and cancer by Raman spectroscopy." Proc SPIE 3918: 106-113. 126 Bakker Schut, T. C , M. J . Witjes, et al. (2000). "In vivo detection of dysplastic tissue by Raman spectroscopy." Anal Chem 72(24): 6010-8. Bauer, N. J . , F. Hendrikse, et al. (1999). "In vivo confocal Raman spectroscopy of the human cornea.' Cornea 18(4): 483-8. Betz, C. S., M. Mehlmann, et al. (1999). "Autofluorescence imaging and spectroscopy of normal and malignant mucosa in patients with head and neck cancer." Lasers Surg Med 25(4): 323-34. Betz, C. S., H. Stepp, et al. (2002). "A comparative study of normal inspection, autofluorescence and 5-ALA-induced PPIX fluorescence for oral cancer diagnosis." Int J Cancer 97(2): 245-52. Bohorfoush, A. (1996). "Tissue spectroscopy for gastrointestinal diseases." Endoscopy 28: 372-380. Brezinski, M. E., G . J . Tearney, et al. (1998). "Optical biopsy with optical coherence tomography." Ann N Y Acad Sci 838: 68-74. Buschman, H. P., E. T. Marple, et al. (2000). "In vivo determination of the molecular composition of artery wall by intravascular Raman spectroscopy." Anal Chem 72(16): 3771-5. Carter, E. and H. Edwards (2001). Biological applications of Raman spectroscopy. Infrared and Raman spectroscopy of biological materials. B. Yan. New York, Marcel Dekker Inc: 421-475. Caspers, P. J . , G . W. Lucassen, et al. (2001). "In vivo confocal Raman microspectroscopy of the skin noninvasive determination of molecular concentration profiles." J Invest Dermatol 116(3): 434-42. Caspers, P. J . , G . W. Lucassen, et al. (2003). "Combined in vivo confocal Raman spectroscopy and confocal microscopy of human skin." Biophys J 85(1): 572-80. Caspers, P. J . , G . W. Lucassen, et al. (1998). "In vitro and in vivo Raman spectroscopy of human skin." Biospectroscopy 4(5 Suppl): S31 -9. 127 Caspers, P. J . , A. C. Williams, et al. (2002). "Monitoring the penetration enhancer dimethyl sulfoxide in human stratum corneum in vivo by confocal Raman spectroscopy." Pharm Res 19(10): 1577-80. Centeno, J . A., F. G. Mullick, et al. (1999). "Laser-Raman microprobe identification of inclusions in capsules associated with silicone gel breast implants." Mod Pathol 12(7): 714-21. Clarke, R., E. Hanlon, et al. (1987). "Laser Raman spectroscopy of calcified atherosclerotic lesions in cardiovascular tissue." Applied Optics 26: 3175-3177. Clarke, R., Q. Wang, et al. (1988). "Laser Raman spectroscopy of atherosclerotic lesions in human coraonary artery segments." Applied Optics 27: 4799-4800. Clarke, R. H., J . M. Isner, et al. (1988). "Spectroscopic characterization of cardiovascular tissue." Lasers Surg Med 8(1): 45-59. Cooney, T., H. Skinner, et al. (1996). "Comparative study of some fiber-optic remote Raman probe designs. Part I: Model for liquids and solids." Appl Spectrosc 50: 836-848. Cooney, T., H. Skinner, et al. (1996). "Comparative study of some fiber-optic remote Raman probe designs. Part II: Tests of single-fiber, lensed, flat, and bezel-tip multifiber probes." Appl Spectrosc 50: 849-860. de Lima, C. J . , S. Sathaiah, et al. (2000). "Development of catheters with low fiber background signals for Raman spectroscopic diagnosis applications." Artif Organs 24(3): 231-4. Delank, W., B. Khanavkar, et al. (2000). "A pilot study of autofluorescent endoscopy for the in vivo detection of laryngeal cancer." Laryngoscope 110(3 Pt 1): 368-73. Dhingra, J . K., D. F. Perrault, Jr., et al. (1996). "Early diagnosis of upper aerodigestive tract cancer by autofluorescence." Arch Otolaryngol Head Neck Surg 122(11): 1181-6. Dollish, F., W. Fateley, et al. (1974). Characteristic Raman freguencies of organic compounds. New York, John Wiley & Sons, Inc. 1 2 8 Eker, C , R. Rydell, et al. (2001). "Multivariate analysis of laryngeal fluorescence spectra recorded in vivo." Lasers Surg Med 28(3): 259-66. Encyclopedia, B. (2003). "Spectroscopy." Encyclopaedia Britannica Premium Service 21 May, 2003(http://www.britannica.com/eb/article?eu=119267). Erckens, R. J . , F. H. Jongsma, et al. (2001). "Raman spectroscopy in ophthalmology: from experimental tool to applications in vivo." Lasers Med Sci 16(4): 236-52. Erckens, R. J . , W. F. March, et al. (2001). "Noninvasive Raman spectroscopic identification of intraocular lens material in the living human eye." J Cataract Refract Surg 27(7): 1065-70. Ermakov, I., M. Ermakova, et al. (2004). "Macular pigment Raman detector for clinical applications." J Biomed Opt 9(1): 139-48. Everitt, B. and G. Dunn (2001). Chapter 3. Principal components analysis. Applied Multivariate Data Analysis. London, Arnold: 48-73. Everitt, B. and G . Dunn (2001). Chapter 11. Discrimination, classification and pattern recognition. Applied Multivariate Data Analysis. London, Arnold: 248-270. Feld, M. S., R. Manoharan, et al. (1995). P rocSPJE 2388: 99-105. Ferraro, J . (1967). Advances in Raman instrumentation and sampling techniques. Raman spectroscopy: Theory and practice. H. Szymanski. New York, Plenum Press: 44-81. Ferraro, J . , K. Nakamoto, et al. (2003). Chapter 1. Basic theory. Introductory Raman spectroscopy. San Diego, Academic Press: 1-94. Ferraro, J . , K. Nakamoto, et al. (2003). Chapter 2. Instrumentation and experimental techniques. Introductory Raman spectroscopy. SanDiego, Academic Press: 95-146. 129 Ferraro, J~, K. Nakamoto, et al. (2003). Chapter 5. Analytical chemistry. Introductory Raman Spectroscopy. San Diego, Elsevier Science: 267-293. Fisher, B. T., K. A. Masiello, et al. (2003). "Assessment of transient changes in corneal hydration using confocal Raman spectroscopy." Cornea 22(4): 363-70. Fodor, S., R. Rava, et al. (1985). "Ultraviolet resonance Raman spectroscopy of the nucleotides with 266-, 240, 218, and 200-nm pulsed laser excitation." J Am Chem Soc 107: 1520-1529. Frank, C . J . , R. L. McCreery, et al. (1995). "Raman spectroscopy of normal and diseased human breast tissues." Anal Chem 67(5): 777-83. Fujimoto, J . G . (2003). "Optical coherence tomography for ultrahigh resolution in vivo imaging." Nat Biotechnol 21(11): 1361-7. Fujimoto, J . G. , B. Bouma, et al. (1998). "New technology for high-speed and high-resolution optical coherence tomography." Ann N Y Acad Sci 838: 95-107. Fujimoto, J . G. , M. E. Brezinski, et al. (1995). "Optical biopsy and imaging using optical coherence tomography." Nat Med 1(9): 970-2. Fujimoto, J . G . , C. Pitris, et al. (2000). "Optical coherence tomography: an emerging technology for biomedical imaging and optical biopsy." Neoplasia 2(1-2): 9-25. Gillenwater, A., R. Jacob, et al. (1998). "Noninvasive diagnosis of oral neoplasia based on fluorescence spectroscopy and native tissue autofluorescence." Arch Otolaryngol Head Neck Surg 124(11): 1251-8. Gniadecka, M., O. Faurskov Nielsen, et al. (1998). "Structure of water, proteins, and lipids in intact human skin, hair, and nail." J Invest Dermatol 110(4): 393-8. 130 Gniadecka, M., P. A. Philipsen, et al. (2004). "Melanoma diagnosis by Raman spectroscopy and neural networks: structure alterations in proteins and lipids in intact cancer tissue." J Invest Dermatol 122(2): 443-9. Gniadecka, M., H. C. Wulf, et al. (2001). "Cutaneous tophi and calcinosis diagnosed in vivo by Raman spectroscopy." Br J Dermatol 145(4): 672-4. Gniadecka, M., H. C. Wulf, et al. (1997). "Distinctive molecular abnormalities in benign and malignant skin lesions: studies by Raman spectroscopy." Photochem Photobiol 66(4): 418-23. Graham Smith, F. and T. King (2000). Prism and grating spectrometers. Optics and photonics: An introduction. Chichester, John Wiley & Sons, Ltd: 233-250. Graham Smith, F. and T. King (2000). Semiconductors and semiconductor lasers. Introduction to optics and photonics. Chichester, John Wiley and Sons, Ltd: 301-314. Hanlon, E., R. Manoharan, et al. (2000). "Prospects for in vivo Raman spectroscopy." Phys Med Biol 45(2): R1-59. Hanlon, E. B., R. Manoharan, et al. (2000). "Prospects for in vivo Raman spectroscopy." Phys Med Biol 45(2): R1-59. Harries, M. L., S. Lam, et al. (1995). "Diagnostic imaging of the larynx: autofluorescence of laryngeal tumours using the helium-cadmium laser." J Larynqol Otol 109(2): 108-10. Hata, T. R., T. A. Scholz, et al. (2000). "Non-invasive raman spectroscopic detection of carotenoids in human skin." J Invest Dermatol 115(3): 441-8. Herz, P. R., Y. Chen, et al. (2004). "Micromotor endoscope catheter for in vivo, ultrahigh-resolution optical coherence tomography." Opt Lett 29(19): 2261 -3. Huang, Z., A. McWilliams, et al. (2003). "Effect of formalin fixation on the near-infrared Raman spectroscopy of normal and cancerous human bronchial tissues." Int J Oncol 23(3): 649-55. 131 Huang, Z., A. McWilliams, et al. (2003). "Near-infrared Raman spectroscopy for optical diagnosis of lung cancer." Int J Cancer 107(6): 1047-52. Huang, Z., H. Zeng, et al. (2001). "Rapid near-infrared Raman spectroscopy system for real-time in vivo skin measurements." Optics Letters 26(22): 1-4. Hung, J . , S. Lam, et al. (1991). "Autofluorescence of normal and malignant bronchial tissue." Lasers Surg Med 11(2): 99-105. Iriyama, K., A. Mizuno, et al. (1982). "An application of laser Raman spectroscopy to the study of a hereditary cataractous lens; on the Raman band for a diagnostic marker of cataractous signatures." Curr Eve Res 2(7): 489-92. Ivarsen, A., J . Thogersen, et al. (2004). "Plastic particles at the LASIK interface." Ophthalmology 111(1): 18-23. Katz, A., E. F. Kruger, et al. (2003). "Detection of glutamate in the eye by Raman spectroscopy." J Biomed Opt 8(2): 167-72. Kennedy, J . C , R. H. Pottier, et al. (1990). "Photodynamic therapy with endogenous protoporphyrin IX: basic principles and present clinical experience." J Photochem Photobiol B 6(1-2): 143-8. Koenig, F., F. J . McGovern, et al. (1996). "Laser induced autofluorescence diagnosis of bladder cancer." J Urol 156(5): 1597-601. Koljenovic, S., L. P. Choo-Smith, et al. (2002). "Discriminating vital tumor from necrotic tissue in human glioblastoma tissue samples by Raman spectroscopy." Lab Invest 82(10): 1265-77. Lawson, E. E., H. G . Edwards, et al. (1998). "Interaction of salicylic acid with verrucae assessed by FT-Raman spectroscopy." J Drug Target 5(5): 343-51. 132 Leunig, A., M. Mehlmann, et al. (2001). "Fluorescence staining of oral cancer using a topical application of 5-aminolevulinic acid: fluorescence microscopic studies." J Photochem Photobiol B 60(1): 44-9. Liu, C , B. Das, et al. (1992). "Raman, fluorescence and time-resolved light scattering as optical diagnostic techniques to separate diseased and normal biomedical media." J Photochem Photobiol B 16: 187-209. Liu, C . H., B. B. Das, et al. (1992). "Raman, fluorescence, and time-resolved light scattering as optical diagnostic techniques to separate diseased and normal biomedical media." J Photochem Photobiol B 16(2): 187-209. Luke, J . L., V. F. Kalasinsky, et al. (1997). "Pathological and biophysical findings associated with silicone breast implants: a study of capsular tissues from 86 cases." Plast Reconstr Surg 100(6): 1558-65. MacAulay, C , S. Lam, et al. (1995). "Optical properties of normal and malignant tissues in the nasopharnyx and larynx." Proc SPIE 2 3 9 5 : 153-157. Mahadevan-Jansen, A., M. F. Mitchell, et al. (1998). "Near-infrared Raman spectroscopy for in vitro detection of cervical precancers." Photochem Photobiol 68(1): 123-32. Mahadevan-Jansen, A., M. F. Mitchell, et al. (1998). "Development of a fiber optic probe to measure NIR Raman spectra of cervical tissue in vivo." Photochem Photobiol 68(3): 427-31. Malzahn, K., T. Dreyer, et al. (2002). "Autofluorescence endoscopy in the diagnosis of early laryngeal cancer and its precursor lesions." Laryngoscope 112(3): 488-93. Manly, B. (1994). Chapter 6. Principal components analysis. Multivariate Statistical Methods. A Primer. London, Chapman & Hall: 76-92. 133 Manly, B. (1994). Chapter 8. Discriminant function analysis. Multivariate Statistical Methods. A Primer. London, Chapman Hall: 107-127. Manoharan, R., J . J . Baraga, et al. (1992). "Quantitative histochemical analysis of human artery using Raman spectroscopy." J Photochem Photobiol B 16(2): 211 -33. Manoharan, R., Y. Wang, et al. (1995). "Ultraviolet resonance Raman spectroscopy for detection of colon cancer." Lsers in Life Sci 6: 217-227. Mizuno, A., T. Hayashi, et al. (1992). "Near-infrared FT-Raman spectra of the rat brain tissues." Neurosci Lett 141(1): 47-52. Mobley, P., B. Kowalski, et al. (1996). Applied Spectroscopy Reviews 31: 347-368. Molckovsky, A., L. M. Song, et al. (2003). "Diagnostic potential of near-infrared Raman spectroscopy in the colon: differentiating adenomatous from hyperplastic polyps." Gastrointest Endosc 57(3): 396-402. Nelson, W., R. Manoharan, et al. (1992). "UV resonance Raman activity of bacteria." Applied Spectroscopy Reviews 27: 67-124. Nijssen, A., T. C. Bakker Schut, et al. (2002). "Discriminating basal cell carcinoma from its surrounding tissue by Raman spectroscopy." J Invest Dermatol 119(1): 64-9. Nordstrom, R. J . , L. Burke, et al. (2001). "Identification of cervical intraepithelial neoplasia (CIN) using UV-excited fluorescence and diffuse-reflectance tissue spectroscopy." Lasers Surg Med 29(2): 118-27. Ozaki, Y., A. Mizuno, et al. (1987). "Inter- and intramolecular disulfide bond formation and related structural changes in the lens proteins. A Raman spectroscopic study in vivo of lens aging." J Biol Chem 262(32): 15545-51. 134 Paluszkiewicz, C , W. M. Kwiatek, et al. (1998). "FT-Raman, FT-IR spectroscopy and PIXE analysis applied to gallstones specimens." Cell Mol Biol (Noisv-le-qrand) 44(1): 65-73. Pantanowitz, L , P. L. Hsiung, et al. (2004). "High-resolution imaging of the thyroid gland using optical coherence tomography." Head Neck 26(5): 425-34. Pasteris, J . D., B. Wopenka, et al. (1999). "Analysis of breast implant capsular tissue for crystalline silica and other refractile phases." Plast Reconstr Surg 103(4): 1273-6. Perno, J . , C . Grygon, et al. (1989). "Raman excitation profiles for the nucleotides and for the nucleic acid duplexes poly(rA)-poly(rU) and poly(dG-dC)." J Phys Chem 93: 5672-5678. Peticolas, W., T. Patapoff, et al. (1996). "Laser Raman microscopy of chromosomes in living eukaryotic cells: DNA polymorphism in vivo." J Raman Spectrosc 27: 571-578. Profio, A. E. and O. J . Balchum (1985). "Fluorescence diagnosis of cancer." Adv Exp Med Biol 193: 43-50. Puppels, G. , T. Bakker Schut, et al. (2001). Chapter 13. In vivo Raman spectroscopy. Handbook of Raman spectroscopy. From the research laboratory to the process line. H. Edwards. New York, Marcel Dekker Inc: 549-574. Puppels, G. , T. C . Bakker Schut, et al. (2001). In vivo Raman spectroscopy. Handbook of Raman spectroscopy. From the research laboratory to the process line. H. Edwards. New York, Marcel Dekker Inc.: 549-574. Raman, C . and K. Krishnan (1928). "A new type of seconday radiation." Nature 212: 501. Rencher, A. (2002). Method of multivariate analysis. New York, John Wiley and Sons. Romer, T. J . , J . F. Brennan, 3rd, et al. (1998). "Histopathology of human coronary atherosclerosis by quantifying its chemical composition with Raman spectroscopy." Circulation 97(9): 878-85. 135 Salenius, J . P., J . F. Brennan, 3rd, et al. (1998). "Biochemical composition of human peripheral arteries examined with near-infrared Raman spectroscopy." J Vase Surg 27(4): 710-9. Schachar, R. A. and S. A. Solin (1975). "The microscopic protein structure of the lens with a theory for cataract formation as determined by Raman spectroscopy of intact bovine lenses." Invest Ophthalmol 14(5): 380-96. Schallreuter, K. U., S. Kothari, et al. (2003). "Molecular evidence that halo in Sutton's naevus is not vitiligo." Arch Dermatol Res 295(6): 223-8. Schallreuter, K. U., M. Zschiesche, et al. (1998). "In vivo evidence for compromised phenylalanine metabolism in vitiligo." Biochem Biophys Res Commun 243(2): 395-9. Schantz, S. P. and R. R. Alfano (1993). "Tissue autofluorescence as an intermediate endpoint in cancer chemoprevention trials." J Cell Biochem Suppl 17F: 199-204. Schmitt, J . and T. Udelhoven (2001). Use of artificial neural networks in biomedical diagnosis. Infrared and Raman spectroscopy of biological materials. H. Gremlich. New York, Marcel Dekker Inc: 379-419. Schoen, C , S. Sharmma, et al. (1993). "Performance of a holographic super-notch filter." Applied Spectroscopy 47: 305-308. Schwab, S. and R. McCreery (1984). "Versatile efficient Raman sampling with fiber optics." Anal Chem 56: 2199-2204. Schwartz, M. T. (1976). "Evolving ecosystems: role in dinosaur extinction." Nature 12: 16-17. Sebag, J . , S. Nie, et al. (1994). "Raman spectroscopy of human vitreous in proliferative diabetic retinopathy." Invest Ophthalmol Vis Sci 35(7): 2976-80. Shafer-Peltier, K. E., A. S. Haka, et al. (2002). "Model-based biological Raman spectral imaging." J Cell Biochem Suppl 39: 125-37. 136 Shaver, J . (2001). Chapter 7. Chemometrics for Raman spectroscopy. Handbook of Raman spectroscopy. From the research laboratory to the process line. H. Edwards. New York, Marcel Dekker Inc: 275-306. Shim, M. G. , L. M. Song, et al. (2000). "In vivo near-infrared Raman spectroscopy: demonstration of feasibility during clinical gastrointestinal endoscopy." Photochem Photobiol 72(1): 146-50. Siew, D., G . Clover, et al. (1995). "Micro-Raman spectroscopic study of organ cultured corneas." Journal of Raman spectroscopy 26: 3-8. Silveira, L , Jr., S. Sathaiah, etal. (2002). "Correlation between near-infrared Raman spectroscopy and the histopathological analysis of atherosclerosis in human coronary arteries." Lasers Surg Med 30(4): 290-7. Silveira, L., Jr., S. Sathaiah, et al. (2003). "Near-infrared Raman spectroscopy of human coronary arteries: histopathological classification based on Mahalanobis distance." J Clin Laser Med Surg 21(4): 203-8. Sims, G . (1994). Priciples of charge-transfer devices. Charge-transfer devices in spectroscopy. M. Denton. Weinheim, V C H : 9-58. Slater, J . , T. JM, et al. (2001). Raman spectroscopy and its adaption to the industrial environment. Handbook of Raman spectroscopy. From the research laboratory to the process line. H. Edwards. New York, Marcel Dekker, Inc: 41-144. Smith, J . , C. Kendall, et al. (2003). "Raman spectral mapping in the assessment of axillary lymph nodes in breast cancer." Technol Cancer Res Treat 2(4): 327-32. Socrates, G . (2001). Infrared and Raman characteristic group freguencies: tables and charts. Chichester, UK, John Wiley & Sons Ltd. 137 Sokolov, K., R. Drezek, et al. (1999). "Reflectance spectroscopy with polarized light: is it sensitive to cellular and nuclear morphology." Optics Express 20(5 (13)): 302-317. Stefanadis, C , M. Vavuranakis, et al. (2003). "Vulnerable plaque: the challenge to identify and treat it." J Interv Cardiol 16(3): 273-80. Stone, N., P. Stavroulaki, et al. (2000). "Raman spectroscopy for early detection of laryngeal malignancy: preliminary results." Laryngoscope 110(10 Pt 1): 1756-63. van de Poll, S. W., K. Kastelijn, et al. (2003). "On-line detection of cholesterol and calcification by catheter based Raman spectroscopy in human atherosclerotic plaque ex vivo." Heart 89(9): 1078-82. Viehoever, A. R., D. Anderson, et al. (2003). "Organotypic raft cultures as an effective in vitro tool for understanding Raman spectral analysis of tissue." Photochem Photobiol 78(5): 517-24. Weinmann, P., M. Jouan, et al. (1998). "Quantitative analysis of cholesterol and cholesteryl esters in human atherosclerotic plaques using near-infrared Raman spectroscopy." Atherosclerosis 140(1): 81-8. Williams, A. C , H. G . Edwards, et al. (1995). "The 'Iceman': molecular structure of 5200-year-old skin characterised by Raman spectroscopy and electron microscopy." Biochim Biophys Acta 1246(1): 98-105. Wohlrab, J . , A. Vollmann, et al. (2001). "Noninvasive characterization of human stratum corneum of undiseased skin of patients with atopic dermatitis and psoriasis as studied by Fourier transform Raman spectroscopy." Biopolymers 62(3): 141-6. Wolthuis, R., T. Bakker Schut, et al. (1999). Raman spectroscopic methods for in vitro and in vivo tissue characterization. Fluorescent and luminescent probes for biological activity. Apractical guide to technology for quantitative real-time analysis. W. Mason. Cambridge, Academic Press: 433-456. 138 Yamazaki, H., S. Kaminaka, et al. (2003). "The diagnosis of lung cancer using 1064-nm excited near-infrared multichannel Raman spectroscopy." Radiat Med 21(1): 1-6. Zargi, M., I. Fajdiga, et al. (2000). "Autofluorescence imaging in the diagnosis of laryngeal cancer." Eur Arch Otorhinolarynqol 257(1): 17-23. 139 A p p e n d i c e s Appendix 1 His topatho logy photographs demonst ra t ing the t issue types be ing s tud ied , (a) nasopha rynx , (b) larynx, and (c) thyroid t i ssue. (a) Nasopharynx (a .1) No rma l nasopha ryngea l t i ssue. Lympho id s t roma cove red by respi ratory type epi the l ium (original magni f icat ion 1 0 0 x ) . 140 (a.2) Undifferentiated nasopharyngeal carcinoma. Sheets of malignant epithelial cells with a lympho-plasmacytic infiltrate. Cells have a syncytial appearance with prominent nucleoli (original magnification 100x). 141 (b) Larynx (b.1) Normal vocal fold tissue. Non-dysplastic squamous epithelium overlies fibrous tissue (original magnification 100x). 142 (b.2) Invasive moderately differentiated squamous cell carcinoma. Irregularly shaped nests of squamous epithelial cells are seen within a reactive stroma in the upper half of the picture. The tissue in the lower left half of the picture shows benign minor salivary gland lobules within the vocal fold tissue (original magnification 40x). 143 (b.3) Squamous papilloma showing multiple papillary fronds covered by non-dysplastic squamous epithelium. Note the relative homogeneity of the tissue compared with the in-homogeous squamous cell carcinoma specimen in (b.2). (original magnification 40x). 144 (c) Thyroid (c.1) Normal thyroid tissue. Normal thyroid follicles are lined by simple cuboidal epithelium (original magnification 40x) f fa '*' I fit* f 4 i r - * £' ft • I - * • * \ •* * ,., T / f V "** * I «** , ,» > i t 1 • B i o " f t 145 (c.2) Papillary carcinoma. Papillary structures surround a fibrovascular core (original magnification 40x). 146 Appendix 2 List showing the number of samples collected from each tissue type at each signal acquisition time. The number in brackets is the number of patients from which the samples were collected. * In thyroid tissue, longer signal acquisition times were associated with signal saturation by autofluorescence, making it impossible to derive a Raman spectrum. 147 Tissue Type Signal Acquisition Time Nasopharynx 1s 5s 10s 15s 30s Normal 14(8) 25(15) 19(11) 14(9) -Cancer 14(6) 18 (8) 16(7) 13(6) -After 18 Sept 2001 Normal 9(6) 9(6) 11 (6) 9(6) -Cancer 12(5) 12(5) 12(5) 12(5) -Larynx Normal 17(5) 18(5) 13(4) 4(3) 1 (1) Cancer 12(4) 13(6) 10(6) 2(1) -Papilloma 14(7) 16(10) 16(10) 14(8) 2(2) Thyroid Benign 69(16) 45(14) 6(3) * * Papillary carcinoma 20 (5) 20 (5) 11 (5) * * 148 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0092915/manifest

Comment

Related Items