JOURNAL OF RAMAN SPECTROSCOPY

J. Raman Spectrosc. 2002; 33: 564–573

Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/jrs.882

Near-infrared Raman spectroscopy for the

classification of epithelial pre-cancers and cancers
Nicholas Stone,1∗ Catherine Kendall,1 Neil Shepherd,2 Paul Crow1,3 and Hugh Barr3
1
Optical Diagnostics Group, Cranfield Postgraduate Medical School, Gloucestershire Royal Hospital, Great Western Road, Gloucester GL1 3NN, UK
2
Department of Cellular Pathology, Gloucestershire Royal Hospital, Great Western Road, Gloucester GL1 3NN, UK
3
Department of Surgery, Gloucestershire Royal Hospital, Great Western Road, Gloucester GL1 3NN, UK

Received 30 October 2001; Accepted 13 March 2002

The use of near-infrared Raman spectroscopy to interrogate epithelial tissue biochemistry and hence
distinguish between normal and abnormal tissues was investigated. Six different epithelial tissues from
the larynx, tonsil, oesophagus, stomach, bladder and prostate were measured. Spectral diagnostic models
were constructed using multivariate statistical analysis of the spectra to classify samples of epithelial
cancers and pre-cancers. Tissues were selected for clinical significance and to include those which develop
into carcinoma from squamous, transitional or columnar epithelial cells. Rigorous histopathological
protocols were followed and mixed pathology tissue samples were discarded from the study. Principal
component fed linear discriminant models demonstrated excellent group separation, when tested by cross-
validation. Larynx samples, with squamous epithelial tissue, were separated into three distinct groups
with sensitivities ranging from 86 to 90% and specificities from 87 to 95%. Bladder specimens, containing
transitional epithelial tissue, were separated into five distinct groups with sensitivities of between 78
and 98% and specificities between 96 and 99%. Oesophagus tissue can contain both squamous and
columnar cell carcinomas. A three group model discriminated the columnar cell pathological groups with
sensitivities of 84–97% and specificities of 93–99%, and an eight group model combining both columnar
and squamous tissues in the oesophagus was able to discriminate pathologies with sensitivities of 73–100%
and specificities of 92–100%. It is likely that any overlap between pathology group predictions will have
been due to a combination of the difficulty in histologically distinguishing between pre-cancerous states
and the fact that there is no biochemical boundary from one pathological group to the next, i.e. there is
believed to be a continuum of progression from the normal to the diseased state. Copyright  2002 John
Wiley & Sons, Ltd.

INTRODUCTION thus making the surveillance or screening of at-risk popu-

The primary requirement for successful treatment of cancer is
early detection. Current methods of detecting malignancies
rely upon surveillance of at risk populations or on diag-
nostic investigations following presentation with suspicious
symptoms. By the time symptoms are present, tumours are
usually of a significant size, and it is often too late to facilitate
a full cure. The majority of malignancies originate in epithe-
lial tissue, usually found lining the surfaces of organs. These
epithelial cancers are called carcinomas. They are especially
of interest because they develop over relatively long time-
scales on the surface or lining of an organ prior to invasion
into deeper tissues. Endoscopic access can often be achieved,
Ł Correspondence to: Nicholas Stone, Optical Diagnostics Group,
Cranfield Postgraduate Medical School, Gloucestershire Royal
Hospital, Great Western Road, Gloucester GL1 3NN, UK.
E-mail: n.stone@medical-research-centre.com
Contract/grant sponsor: MedLINK UK.
lations possible. Furthermore, the process of carcinogenesis
in many epithelial tissues, although not fully understood, is
known to frequently include a pre-malignant state. If this
microscopic cellular change can be detected then early treat-
ments, likely to be less damaging and more successful, can
be performed.
Cancers of the larynx, oesophagus and bladder account
for 9% of all cancers in adults in England and Wales, with
approximately 2000, 5000 and 11 000 new cases per year for
larynx, oesophagus and bladder malignancies, respectively.1
The incidence of all three cancers has risen over the past
decades,1,2 although the rate of incidence of oesophageal
adenocarcinomas occurring in Barrett’s oesophagus has risen
more rapidly than for any other cancer.3,4 The transformation
to carcinoma in all three organs involves an intermediate or
pre-malignant step, whereby the cells in the epithelial layer
begin to grow and mature abnormally. This state is called
dysplasia, although each organ exhibits visually distinct
dysplastic cells depending on the original epithelial cells

Copyright  2002 John Wiley & Sons, Ltd.


present. For example, long periods of exposure to gastro-
oesophageal reflux can lead to a change in the lining of the
oesophagus, whereby the normal squamous epithelial lining
is replaced by protective columnar epithelial cells. This is
often termed Barrett’s oesophagus and it is associated with an
increased risk of developing oesophageal adenocarcinoma.5,6
Barrett’s dysplasia is the most frequently used marker of
increased cancer risk in Barrett’s oesophagus. It is more
commonly seen in areas of intestinal metaplasia (IM), one
of three types of Barrett’s oesophagus, and it is unlikely
that oesophageal adenocarcinoma occurs except in patients
with IM.7,8 The potential benefits of removing an oesophagus
exhibiting dysplasia must be weighed against the relatively
high mortality associated with oesophagectomy (5–15%)9
and the poor outcome in patients who present with invasive
adenocarcinoma of the oesophagus (23% survival at 1 year
and 7% survival at 5 years in England and Wales between
1986 and 1990).1 Although survival rates for laryngeal and
bladder cancers are better (55–65% at 5 years),1 it is delayed
diagnosis leading to systemic metastatic invasion that is the
main reason for poor survival from epithelial cancers.10
The ‘gold standard’ for the detection of malignan-
cies and pre-malignancies is excisional biopsy followed by
histopathological analysis. This technique relies upon sec-
tioning tissue less than one cell thick (<10 µm), staining with
haematoxylin and eosin (H&E) and viewing under a con-
ventional light microscope. The analysis depends mainly on
the subjective recognition of tissue morphology and architec-
tural patterns. There are significant difficulties in obtaining
an accurate diagnosis using the ‘gold standard.’ First, the
selection of biopsies for the detection of invisible microscopic
surface lesions must rely upon a blind targeting protocol. This
can lead to a high probability of missing abnormal tissue and
large numbers of normal samples will be generated. A study
in the oesophagus following 16 patients undergoing surveil-
lance for Barrett’s oesophagus demonstrated that even a
rigorous sampling protocol will be likely to miss abnormal
lesions previously found in the organ.11 Second, in the anal-
ysis of pre-cancerous lesions there are often high levels of
discrepancy between pathologists, owing to the subjective
nature of histopathological analysis. Studies grading dys-
plasia in ulcerative colitis and colonic adenomas produced
an overall agreement between any two pathologists ranging
from 42 to 65%.12,13 Similar studies demonstrated agreement
of between 58 and 61% for oesophageal dysplasia14 and 54%
for laryngeal dysplasia.15 Furthermore, the removal of tis-
sue samples can have a damaging effect on the function of
the organ, especially in sensitive organs such as the larynx,
where permanent loss of voice may result.
Histopathological examination of biopsy samples relies
upon the subjective assessment of tissue architecture, which
is likely to exhibit abnormal changes at a later stage than sub-
cellular biochemical changes. Evidently, the development
of a technique, permitting the objective, non-invasive
qualitative biochemical analysis of tissue would be of great
Copyright  2002 John Wiley & Sons, Ltd.
Classification of epithelial pre-cancers and cancers 565
value. The potential of Raman spectroscopy to distinguish
between normal and malignant tissue has been demonstrated
by a number of workers.16 – 20
The objective of this study was to evaluate the potential
for near-infrared (NIR) Raman spectroscopy, a highly specific
optical analysis technique, to classify or discriminate between
the pathological state of different epithelial tissues and thus
demonstrate a future capability for in vitro classification
and in vivo detection of early epithelial malignancies. The
larynx, oesophagus and bladder are the main subjects of
this work. Carcinomas in each organ develop from three
distinctly different types of epithelial cells. This has allowed
the evaluation of Raman spectral analysis for detection
and classification of disease in squamous, transitional and
columnar epithelial cells. To date only the oesophagus has
been examined in any detail with Raman spectroscopy,
although only selected pathologies were compared.17,21
This study followed extremely rigorous sampling pro-
cedures to reduce the errors caused by discrepancies in
histopathology, sample misorientation and by measuring
samples with mixed pathology. In addition, the full spec-
trum of malignant disease found in each organ was studied,
in order to remove any bias likely to be found in studies
comparing normal with cancerous tissue, i.e. the extreme
pathologies, but not the stages in between.
EXPERIMENTAL
Tissue samples
Tissue specimens were collected for this study during routine
endoscopic or surgical procedures. Informed written consent
was obtained from each patient and the study was approved
by the relevant Local Research Ethics Committees (UK). The
tissue samples were removed from the biopsy forceps and
orientated on acetate paper to enable a depth cross-section
through the tissue to be seen. The mounted sample was then
placed in a 2 ml cryovial (BDH) and dropped into liquid
nitrogen. Histological tissue cross-sections were cut from
each of the samples using a freezing-microtome and stained
with H&E. The use of the acetate backing paper made it
possible to adhere the sample to the microtome without
the use of a cutting agent and thus minimize chemical
contamination to the tissue sample. Stained sections were
analysed by between one and three registry histopathologists
according to organ-specific architectural and morphological
grading systems.22 Only homogeneous samples with clearly
defined pathologies were used in this investigation and those
with mixed or indeterminate pathologies were discarded.
In addition only oesophagus samples with a unanimous
agreement from all three pathologists were used for this
study. Both of these precautions were implemented to reduce
errors from using misclassified and mixed pathology tissue
samples. The retained samples (biopsy blocks) were stored
at
85 ° C until spectroscopic studies were carried out. Prior
J. Raman Spectrosc. 2002; 33: 564–573
566 N. Stone et al.
to performing spectral measurements, the specimens were
passively warmed to room temperature.
A total of 41 laryngeal biopsy specimens were collected
from 35 patients undergoing routine microlaryngoscopy.
The larynx sections were classified by a histopathologist as
either normal squamous epithelium; squamous dysplasia
or squamous cell carcinoma. Twenty-five samples from 20
patients were retained for analysis with Raman spectroscopy,
including 14 histologically normal samples, five exhibiting
dysplasia and six with squamous cell carcinoma.
Over 150 oesophageal jumbo biopsy samples were har-
vested from patients undergoing endoscopic surveillance
for dysplasia in Barrett’s oesophagus. Following section-
ing and histopathological analysis, only 89 homogeneous
samples from 44 patients remained. A blind consensus of
histopathological opinion was achieved on 50 of these. The
samples with consensus, retained for study with Raman
spectroscopy, exhibited eight different pathological states
[normal squamous epithelium, cardiac Barrett’s (CB), fundic
Barrett’s (FB), intestinal metaplasia (IM), high-grade dys-
plasia (HGD), adenocarcinoma, squamous dysplasia and
squamous cell carcinoma (SCC)]. Of all the samples in this
study there was no consensus of agreement on samples
exhibiting low-grade dysplasia (LGD) and therefore this
group was omitted from the consensus prediction model
described in this paper.
Bladder samples exhibiting normal, carcinoma in situ
(CIS), low-grade carcinoma (LGC), moderate-grade carci-
noma (MGC) and high-grade carcinoma (HGC) were har-
vested during routine cystoscopy. Twelve homogeneous
samples from 12 patients were retained for analysis.
Other samples
Homogeneous normal samples of stomach, prostate and
tonsil were also studied to demonstrate the facility of
NIR-Raman spectroscopy to provide repeatable analysis of
tissue biochemistry in other epithelial tissues. Additionally,
confounding factors or substances likely to contaminate
epithelial tissue spectra were collected and studied. These
included mucus, bile, blood, a coffee and spittle mixture and
endoscope lubricant.
Raman spectroscopy
Raman scattering measurements were performed in vitro
with a customized Renishaw System 1000 Raman micro-
spectrometer. A Spectra-Physics argon ion (2017) pumped
Ti:sapphire laser (3900S) was tuned to provide an excita-
tion wavelength of 830 nm and a custom set of edge filters
(Renishaw) were used to reject elastic scattered light of the
laser wavelength. A single dispersion grating was used to
separate the collected light into its spectral components:
1200 lines mm
1 for laryngeal tissue measurements and
300 lines mm
1 for oesophagus and bladder tissue mea-
surements. To allow analysis of the full fingerprint region
(400–1800 cm
1 ) of the tissue spectra it was necessary, with
Copyright  2002 John Wiley & Sons, Ltd.
the higher line density grating, to scan the spectrum across
the CCD using the ‘extended scan mode.’ The lower density
grating allowed measurement of the full fingerprint region
of the tissue spectrum without the need to move the grating.
However, a compromise in spectral resolution was made,
but this has been shown not to affect the quality of the data.
High-quality NIRPLAN optics were selected to minimize
their spectral contribution in the fingerprint region, reducing
the requirement for optics signal subtraction.
Samples were orientated so that the incident laser beam
illuminated the epithelial surface, with a similar geometry to
that which would be used in vivo. Tissue samples were held
in position on the microscope stage between two calcium
fluoride slides. A ð 80 ultra-long working distance lens
was used to focus the laser beam (power at the sample
of 32 š 1.1 mW) to a spot size of 2–3 µm on the tissue surface
and collect the scattered photons in non-confocal mode. At
least five spectra were acquired from each larynx sample
(mean number D 7.96, range 5–12) and a minimum of 10
spectra were acquired from each oesophagus sample (mean
number D 14, range 10–20). The scattered Raman signal
was integrated for 30 s and measured over a spectral range
of 400–1800 cm
1 with respect to the excitation frequency.
Bladder spectral measurements were performed in the same
manner, except that a ð20 objective was used with an
illumination power of ¾85 mW. At least 15 spectra were
acquired from each sample and spectra were integrated for
10 s. The changes in spectrometer configuration between
laryngeal to oesophageal to bladder spectral datasets were
caused by ongoing optimization processes with the aim of
maximizing the signal and minimizing integration times.
The Raman spectrometer was calibrated for wavenumber
shift using a neon lamp standard and Renishaw WiRE
software. Tissue spectra were corrected for the energy-
dependent response of the system by multiplying the
measured spectrum by the energy transfer function, which
was measured using a tungsten-filament lamp with a
spectral output calibrated by the UK National Physical
Laboratory. No further manipulation of the tissue spectra
was performed prior to analysis. Intra-sample repeatability
was evaluated and the mean spectra for each tissue pathology
were obtained. Any correlation between Raman spectral
variations and histopathology was assessed using empirical
and multivariate data analysis techniques.
RESULTS AND DISCUSSION
The use of 830 nm excitation radiation facilitated the
acquisition of high-quality, low-fluorescence spectra in short
time-scales. Little background fluorescence is present in the
spectra and the background slope is almost exclusively
caused by the diminishing CCD sensitivity at longer
wavelength. Signal-to-noise ratios between 10 and 20 were
achieved (depending on the signal strength of the mode)
J. Raman Spectrosc. 2002; 33: 564–573
 Classification of epithelial pre-cancers and cancers

0.65 5
Normal Squamous Mucosa
Squamous Dysplasia
4
0.6
Squamous Cell Carcinoma

3
0.55
Intensity / arbitr. units

2
0.5
1

LD 2
0.45
0

0.4
-1

0.35
-2

0.3 -3
Normal Larynx
Squamous Dysplasia
Squamous Cell Carcinoma
0.25 -4
400 600 800 1000 1200 1400 1600 -6 -4 -2 0 2
-1
Wavenumber / cm LD 1

Plate 1. Mean Raman spectra from each pathological group Plate 2. Plot of linear discriminant function weights for each
in the laryngeal tissue. Spectra have been normalized to the larynx spectrum, when tested against the optimized model using
intensity of the 1446 cm
1 peak. a cross-validation process.

0.6

0.55 Normal Squamous Mucosa

4 Barrett's Oesophagus

0.5 Dysplasia & Cancer

0.45
Intensity / arbitr. units

0.4

0.35 0
LD 2

0.3
-2
0.25 Normal Oesophagus
Cardiac Barrett's
Fundic Barrett's
0.2 Intestinal Metaplasia
Low Grade Dysplasia -4
High Grade Dysplasia
0.15 Adenocarcinoma
Squamous Dysplasia
Squamous Cell Carcinoma
0.1 -6
400 600 800 1000 1200 1400 1600 1800
-5 -3 -1 1 3 5 7
-1
Wavenumber / cm LD 1

Plate 3. Mean Raman spectra from all pathologies present in Plate 4. Plot of linear discriminant function weights of the
the oesophageal tissue specimens studied. Spectra have been three-group oesophagus model for each spectrum, when tested
normalized to the intensity of the 1446 cm
1 peak. against the model using a cross-validation process.

6
Normal Bladder
0.9
Carcinoma in situ

4 Low Grade Carcinoma

Intensity / arbitr. units

Moderate Grade Carcinoma

0.8
High Grade Carcinoma
2

0.7
LD 2

0.6
-2

0.5 Normal Bladder

Carcinoma in situ -4
Low Grade Carcinoma
Moderate Grade Carcinoma
High Grade Carcinoma
0.4
400 600 800 1000 1200 1400 1600 -6
-5 -3 -1 1 3 5
-1
Wavenumber / cm LD 1

Plate 5. Mean Raman spectra from all bladder specimens Plate 6. Plot of linear discriminant function weights for
studied, grouped by pathology. Spectra have been normalized each spectrum, when tested against the model using a
to the intensity of the 1446 cm
1 peak. cross-validation process.

Copyright  2002 John Wiley & Sons, Ltd. J. Raman Spectrosc. 2002; 33
N. Stone et al.

8 Normal Squamous Mucosa 10

Cardiac Barrett's
Fundic Barrett's 8
6 Intestinal Mataplasia
High Grade Dysplasia
4
Adenocarcinoma 6
Squamous Dysplasia
Squamous Cell Carcinoma
4
2
2

LD 3
LD 2
0
0
-2
-2
-4 -4
-6 -6
-8 -8
-6 -1 4 9 -5 0 5 10 15
LD 1 LD 1

4 5

2 3

0 1

LD 5
LD 4

-2 -1

-4 -3

-6 -5

-8 -7
-5 0 5 10 15 -6 -1 4 9 14 19
LD 1 LD 1

6 6

4 4

2 2
LD 6

LD 7

0 0

-2 -2

-4 -4

-6 -6
-5 0 5 10 15 -5 0 5 10 15
LD 1 LD 1
Plate 7. Plot of linear discriminant function weights of the eight-group oesophagus model for each spectrum, when tested against
the model using a cross-validation process.

Copyright  2002 John Wiley & Sons, Ltd. J. Raman Spectrosc. 2002; 33
 Classification of epithelial pre-cancers and cancers 567

in epithelial spectra when using the ð80 objective, the 300
lines mm
1 grating and 30 s of signal integration.
Plots of mean, energy sensitivity corrected, Raman
spectra from the normal epithelial tissues studied (larynx,
oesophagus, bladder, tonsil, stomach and prostate) are
shown in Fig. 1 to enable similarities and differences to
be visualized. The Raman spectra of the normal epithelial
tissues are very similar, but subtle variations exist. The most
significant of these can be seen in the wavenumber ranges
800–980 and 1200–1400 cm
1 .
The spectral repeatability over 100 spectra from 10
normal oesophageal specimens from different patients was
evaluated. The spectral shape was consistent and the
wavenumber positions of the peaks vary about the mean
positions by š1.12 cm
1 . However, there was a significant
baseline intensity variation from one sample to the next. The
mean spectrum from 1200 oesophagus spectra, covering all
pathologies exhibited in the measured specimens, is shown
in Fig. 2. The most significant vibrational modes have had
their mean wavenumber positions labelled. An attempt to
assign these peaks tentatively has been made in Table 1.
Larynx
Laryngeal spectra were combined into three separate groups,
each defined by the tissue pathology of the measured spec-
imen. The mean spectra for each group, normal squamous
mucosa, squamous dysplasia and squamous cell carci-
noma, were calculated, normalized to the peak intensity
at 1446 cm
1 and plotted in Plate 1. Visual inspection of this
plot shows that the differences observed in the spectra from
specimens at different stages of cancer development are very
subtle. At first glance the spectra appear almost identical.
However, a closer study of the data reveals miniscule changes
in peak heights from one group to the next. Difference spectra
indicate that the magnitude of the spectral variance between
pathology groups is between 1 and 2% of the Raman spec-
tral signal. Therefore, rather than empirically select peaks
or peak ratios for significance of correlation between inten-
sity and pathology group, all major peak intensities were
measured for each spectrum in the data set and their ratios
calculated. These peak intensities and peak intensity ratios
were then tested for significance of group separation using
analysis of variance (ANOVA). The inter-sample variation
of individual peak intensities was large enough to obscure
the inter-group differences. However, the use of peak ratios
compensated for the overall baseline intensity changes from
one sample to the next. The most significant peak ratio (F
value 33.46, p value −0.0001) for separation of the three
groups was shown to be I754 /I780 . Mean intensity ratios of
0.953 š 0.036, 0.942 š 0.035 and 0.906 š 0.041 were observed
for normal, dysplastic and cancerous tissue, respectively.
Although on average this peak ratio does vary significantly
with pathology group, the variation between spectra were

Larynx Oesophagus Bladder

0.8 0.65
0.6
Intensity / arbitr. units

0.7 0.6
0.55 0.55
0.6
0.5 0.5
0.5
0.45
0.45
0.4 0.4
0.4 0.3 0.35
0.35 0.2 0.3
0.25
600 800 100012001400 1600 600 800 100012001400 1600 600 800 100012001400 1600

Tonsil Stomach Prostate

0.25
0.95
Intensity / arbitr. units

0.9 1.1
0.85 0.2
1
0.8
0.75 0.9
0.7 0.15
0.65 0.8
0.6
0.7
0.55 0.1
600 800 100012001400 1600 600 800 100012001400 1600 600 800 100012001400 1600
Wavenumber / cm-1
Figure 1. Mean energy sensitivity corrected Raman spectra from selected normal epithelial tissues.

Copyright  2002 John Wiley & Sons, Ltd. J. Raman Spectrosc. 2002; 33: 564–573
568 N. Stone et al.

Mean Raman spectrum from oesophagus tissue (all pathologies) at 830 (n = 1200)
0.8

1446
0.7

1001

1313
1335
1259
935
0.6

1083
1123

1655
853

1031
Intensity / arbitr. units

1209
828
0.5
524

1155
755
490

719

1170
781
643
621
669

1616
0.4

1548
1579
0.3

0.2

0.1
400 600 800 1000 1200 1400 1600 1800
Wavenumber / cm-1
Figure 2. Mean Raman spectrum from all samples used in the oesophagus study. Wavenumber positions of major peaks are
labelled.

large enough to make this method of discrimination poor for
spectral correlation with pathology.
It is likely that the extremely complex changes from
normal to cancerous tissue will be accompanied by many
molecular variations. Therefore, spectral analysis techniques
involving all the data in the spectra will be more likely
to describe the subtle but complex spectral variations.
Principal components (PCs) were used to describe the spectra
as a linear combination of weights and loading vectors.
Each loading vector or principal component described ever
decreasing variances in the spectra from the mean of the
dataset. PCA was performed to reduce the number of
variables in the analysis. The weights of the principal
components were used as the variables for entry into a
linear discriminant analysis (LDA) model, which maximized
the variance between groups and minimized the variance
within groups. The multivariate analysis routines were
performed using the PLS-toolbox (Eigenvector) on the
MATLAB platform. Spectral datasets were pre-processed
using Savitsky–Golay filters and mean-centering, following
correction for system energy sensitivity (see Experimental).
Prior to calculation of the linear discriminant functions to
separate the groups maximally, ANOVA was performed on
the PC weights to identify the diagnostically significant
components (PCs) of the spectra. Six PCs with >99.9%
significance of variance between groups were entered into
the LDA model. To separate three groups of data maximally,
two LD functions are required; for greater numbers of groups
(n), n
1 linear discriminant functions are required. All the
linear discriminant ‘diagnostic models’ were tested using
cross-validation or leave-one-out procedures, whereby each
spectrum is held back in turn, the model is calculated and
the withheld spectrum is projected on to the model. This
technique does not provide as good a test as projecting
completely new data on to the model, but with relatively
small numbers of patients and samples it will provide an
unbiased estimation of the model performance.
Plate 2 shows a two-dimensional scatter plot of linear dis-
criminant function weights for the optimized larynx model.
The resulting prediction performances for each group against
the ‘gold standard’ histopathology are shown in Table 2. Out-
come measures were calculated to allow evaluation of the
model in terms of sensitivity and specificity of in effect three
tests for the larynx model: either normal versus other groups,
dysplasia versus other groups or cancer versus other groups.
These are also displayed in Table 2. From scatter plots of
the linear discriminant function weights it could be seen
that group clustering was achieved successfully, although
some groups appeared to cluster into two sub-groups. This
is likely to be due to the evolving spectrometer configuration

Copyright  2002 John Wiley & Sons, Ltd. J. Raman Spectrosc. 2002; 33: 564–573
 Classification of epithelial pre-cancers and cancers 569

Table 1. Tentative assignments of major vibrational modes identified in oesophageal samples

Peak
position/cm
1 Major assignments

490 Glycogen
524 S–S disulfide stretch in proteins
621 C–C twisting mode of phenylalanine
643 C–C twisting mode of tyrosine
669 C–S stretching mode of cystine
719 C–N (membrane phospholipid head)/nucleotide peak
755 Symmetric breathing of tryptophan
781 Cytosine/uracil ring breathing (nucleotide)
828 Out of plane ring breathing tyrosine/O–P–O stretch DNA
853 Ring breathing mode of tyrosine and C–C stretch of proline ring
920 C–C stretch of proline ring/glucose/lactic acid
935 C–C stretching mode of proline and valine and protein backbone (˛-helix
conformation)/glycogen
1001 Symmetric ring breathing mode of phenylalanine
1031 C–H in-plane bending mode of phenylalanine
1083 C–N stretching mode of proteins (and lipid mode to lesser degree)
1123 C–C stretching mode of lipids/protein C–N stretch
1155 C–C (and C–N) stretching of proteins (also carotenoids)
1170 C–H in-plane bending mode of tyrosine
1209 Tryptophan and phenylalanine
(C–C6 H5 ) mode
1240–1265 Amide III (C–N stretching mode of proteins, indicating mainly ˛-helix conformation)
1313 CH3 CH2 twisting mode of collagen/lipids
1335 CH3 CH2 wagging mode of collagen and polynucleotide chain (DNA–purine bases)
1446 CH2 bending mode of proteins and lipids
1548 Tryptophan
1579 Pyrimidine ring (nucleic acids) and heme protein
1603 C C in-plane bending mode of phenylalanine and tyrosine
1616 C C stretching mode of tyrosine and tryptophan
1655 Amide I (C O stretching mode of proteins, ˛-helix conformation)/C C lipid stretch

Table 2. Cross-validation classification results obtained using the

discriminant functions calculated from clinically significant principle
components of the larynx spectraa

Raman-predicted group membership (No.)

Classification Normal Dysplasia Carcinoma Total

Normal 100 5 7 112

Dysplasia 4 34 0 38
Carcinoma 7 0 42 49

Outcome measure Normal Dysplasia Carcinoma

Sensitivity/% 89 90 86
Specificity/% 87 93 95
a
Each case was classified by the functions derived from all cases other than that
case; 88.4% of cross-validated grouped cases are correctly classified using this
method.

Copyright  2002 John Wiley & Sons, Ltd. J. Raman Spectrosc. 2002; 33: 564–573
570 N. Stone et al.

that was employed during measurements of the laryngeal
spectral dataset. Raman spectra were acquired with differ-
ent resolutions, signal-to-noise ratios, calibration protocols
and laser stabilities. The use of a smoothing function to pre-
process the data was shown to reduce the effect of these
variations in the model.
Oesophagus
Oesophageal spectra were measured from homogeneous
samples of eight different pathological groups. The mean
normalized spectra from each pathological group are shown
in Plate 3 for visual comparison.
The spectra were combined into clinically significant
groups for construction of a multivariate linear discrim-
inant model in the same way as that described for the
larynx. However, following significant enhancement in laser
and spectrometer wavenumber repeatability afforded by
improvements in temperature stabilization, it was found to
be unnecessary to smooth/pre-process the oesophageal and
bladder data prior to calculation of principle components.
Both eight-group and three-group linear discriminant mod-
els were constructed and tested. The three-group model,
relevant for biopsy targeting applications, used spectra
measured from normal, Barrett’s oesophagus and Barrett’s
neoplasia. This required two linear discriminant functions
for optimum separation of the groups. The cross-validated
LD function weights are plotted in Plate 4. However, the
construction of an eight-group discriminant model for clas-
sification necessitated the use of seven LD functions and a
seven-dimensional decision space. In this case it is difficult to
visualize group separations using 2D scatter plots. However,
an attempt was made to include an informative selection
to allow visualization of some grouping within two dimen-
sions (Plate 7). The resulting prediction performances and
outcome measures for the three-group model are presented
in Table 3 and for the eight-group model in Table 4. Excellent
separation of samples into pathological groups was achieved
with 93.2 and 88.6% of cases correctly predicted by the three-
and eight-group models, respectively.
The greatest cross-over between groups occurred
between the intestinal metaplasia, high-grade dysplasia and
adenocarcinoma groups. It is these that cause pathologists
the greatest difficulty in classification. The cross-over is likely
to be due in part to both the difficulty in pathological
discrimination between these groups and the fact that
abnormal development of the epithelial tissue follows a
continuum with many biochemical similarities.
Utilization of tentative peak assignments for significant
positive and negative peaks in the difference spectra (not
shown) enabled some sense to be made of the biochemical
changes that accompany the progress towards adenocar-
cinoma in oesophageal tissue. Dysplastic and cancerous
tissues have been shown (tentatively) to exhibit greater DNA,
hydroxyapatite, urea/phenylalanine, ˛-helix and unordered
proteins, but lower glycogen, carbohydrates, protein disul-
fide bonding and carotenoids, than normal tissue. This
tends to corroborate known biochemical changes leading
to tumour develoment such as increased energy consump-
tion from cell division, increased nuclear/cytoplasmic ratio,
protein conformation changes and tissue calcification. The
histopathological groupings have been developed by pathol-
ogists utilizing architectural and morphological differences
to recognize from collective experience when a tissue has
undergone a change likely to be clinically significant. How-
ever, cells are undergoing a continuum of change from
Barrett’s mucosa to dysplasia to adenocarcinoma, and there-
fore full separation, either histologically or spectroscopically,
will be difficult.
Ten principal components (1, 2, 3, 4, 5, 6, 7, 8, 10 and
12) were shown to exhibit >99.9% significance of discrim-
ination between the eight pathological groups designated
from a consensus of opinion of three pathologists. The most
significant PC (6) described only 0.05% of the total variance

Table 3. Cross-validation classification results obtained using the three-group

linear discriminant model calculated from clinically significant principle
components of the oesophagus spectraa

Raman-predicted group membership (No.)

Classification Normal Barrett’s Neoplasia Total

Normal 301 1 8 310

Barrett’s 0 122 23 145
Neoplasia 3 7 152 162

Outcome Measures Normal Barrett’s Neoplasia

Sensitivity/% 97 84 94
Specificity/% 99 98 93
a
Each case was classified by the functions derived from all cases other than that
case; 93.2% of cross-validated grouped cases are correctly classified using this
method.

Copyright  2002 John Wiley & Sons, Ltd. J. Raman Spectrosc. 2002; 33: 564–573
 Classification of epithelial pre-cancers and cancers 571

Table 4. Cross-validation classification results obtained from the eight-group linear discriminant model calculated from
clinically significant principle components of the oesophagus spectraa

Raman-predicted group membership (No.)

Classification Normal CB FB IM HGD Adeno Sq. dys SCC Total

Normal 292 0 1 4 9 0 4 0 310

CB 0 12 0 0 0 0 0 0 12
FB 0 0 45 2 0 0 0 0 47
IM 0 0 4 66 16 0 0 0 86
HGD 2 0 3 0 60 3 0 0 68
Adeno 0 0 5 1 19 69 0 0 94
Sq. Dys. 0 0 0 0 0 0 15 0 15
SCC 0 0 0 0 0 0 0 10 10

Outcome measure Normal CB FB IM HGD Adeno Sq. dys SCC

Sensitivity/% 94 100 96 77 88 73 100 100

Specificity/% 99 100 98 99 92 99 99 100
a
Each case was classified by the functions derived from all cases other than that case; 88.6% of cross-validated grouped
cases are correctly classified using this method.

Table 5. Cross-validation classification results obtained from the five-group linear

discriminant model calculated from clinically significant principle components of the
bladder spectraa

Raman-predicted group membership (No.)

Classification Normal CIS LGC MGC HGC Total

Normal 14 0 0 1 0 15
CIS 1 33 4 2 0 40
LGC 0 2 53 0 0 55
MGC 1 3 0 36 6 46
HGC 0 0 1 0 39 40

Outcome measure Normal CIS LGC MGC HGC

Sensitivity/% 93 83 96 78 98
Specificity/% 99 97 96 98 96
a
Each case was classified by the functions derived from all cases other than that case;
89.3% of cross-validated grouped cases are correctly classified using this method.

from the mean, indicating that subtle modifications in bio-
chemistry precede and accompany significant pathological
changes to the tissue.
Although most principal components are made up of a
complex combination of tissue components, PC3 can be seen
to represent glycogen, which is known to be the main energy
source found in human cells. Normal cells will have this in
abundance whereas in abnormal cells it will be in deficit.
This appears to be demonstrated in both difference spectra
and principal component analyses.
Bladder
Raman spectra of bladder specimens were measured and
grouped according to histopathology opinion. The mean
spectra for each group are shown in Plate 5. Subtle dif-
ferences can be seen. These have been exploited with
the use of a principal component fed linear discriminant
model (as described above) to separate maximally the
five pathology groups exhibited by the specimens mea-
sured. Four LD functions were calculated and the weights
of the first two are plotted in Plate 6 to demonstrate
group clustering. The cross-validation prediction perfor-
mance and the outcome measures are outlined in Table 5.
Excellent group separation is achieved, although there is
some overlap between carcinoma in situ and the low-
grade cancers and between the moderate- and high-grade
cancers. This may be due in part to the biochemical contin-
uum between the increasing grades of malignancy and the

Copyright  2002 John Wiley & Sons, Ltd. J. Raman Spectrosc. 2002; 33: 564–573
572 N. Stone et al.

difficulty in distinguishing fully between these groupings
with histopathology.
Potential confounding factors
Figure 3 displays Raman spectra of potential contaminants
to tissue spectra, that may be confounding factors in spectral
diagnostic model prediction. Plots of bile, mucus, spittle
and coffee combined, blood and endoscope lubricant, all
measured under the same conditions as the tissue spectra, are
shown for comparison of signal strength with an oesophagus
spectrum. The tissue spectra measured to date (without
washing) were not unduly affected by these contaminants,
although the strengths of the signals indicate that they should
not be ignored from analysis altogether.

Bile Nasal Mucus

10 0.1
Intensity / arbitr. units

9 0.08

8 0.06

7 0.04

6 0.02

5 0
400 600 800 1000 1200 1400 1600 1800 400 600 800 1000 12001400 1600 1800

Mucus Spittle
0.11 0.5
Intensity / arbitr. units

0.1
0.45
0.09
0.4
0.08
0.35
0.07

0.06 0.3
400 600 800 1000 1200 1400 1600 1800 400 600 800 1000 12001400 1600 1800
Wavenumber / cm -1 Wavenumber / cm -1

Blood Cells Whole Blood

0.5 45
Intensity / arbitr. units

0.45 0.4
0.35
0.4
0.3
0.35
0.25
0.3
0.2
0.25 0.15
0.2 0.1
400 600 800 1000 12001400 1600 1800 400 600 800 1000 12001400 1600 1800

Endoscope Lubricant Average Oesophagus

0.18 0.9
Intensity / arbitr. units

0.8
0.16
0.7
0.14 0.6
0.5
0.12 0.4
0.3
0.1
0.2
0.08 0.1
400 600 800 1000 12001400 1600 1800 400 600 800 1000 12001400 1600 1800
Wavenumber / cm -1 Wavenumber / cm -1
Figure 3. Raman spectra of possible contaminants of in vivo and in vitro tissue Raman spectra, measured in 30 s and corrected for
system energy sensitivity. An oesophagus spectrum is included for comparison.

Copyright  2002 John Wiley & Sons, Ltd. J. Raman Spectrosc. 2002; 33: 564–573

CONCLUSIONS
This study has demonstrated the facility of Raman spec-
troscopy to distinguish between the pathological state of
three distinct epithelial tissues. This indicates the generic
nature of a Raman spectroscopic tool for tissue classification
and detection of disease in different epithelial tissues.
Improvements have been made to previously published
data on the larynx,18 by optimizing the model using spectral
filtering prior to multivariate analysis. This effectively
provided a method of smoothing out the spectral variations
due to spectrometer configuration changes and an unstable
laser wavelength. The bladder and oesophagus models
did not require this procedure to achieve high-quality
discrimination. This was due to the use of a stable,
optimized spectrometer with repeatable calibration protocols
throughout acquisition of the datasets.
This study was the first to separate spectrally so many
pathology groups in the oesophagus and thus indicate
the facility of Raman spectroscopy for tissue classification,
rather than just simple disease detection. Furthermore,
the first example of Raman spectral separation of bladder
tissue specimens into their respective pathology groups was
demonstrated using a Raman/linear discriminant model.
There are difficulties associated with the full acceptance
of this technique as a medical tool, either for biopsy targeting
or in vitro pathology classification. The use of an imperfect
gold standard to calibrate and evaluate new diagnostic
modalities will always cause problems. However, this is
currently the best technique available, and therefore the
use of strict protocols and reliance on a minimum of a
consensus of three expert histopathologists will at least tend
to reduce the errors from histopathological discrepancies.
Further work has been carried out to evaluate these problems
elsewhere.23
The use of much greater sample numbers will be required
to build and test models to include full inter-sample and
inter-spectrometer reproducibility. All spectra measured for
this study are traceable back to histological sections and
correction files, and therefore they can be included in any
future reassessment. Work is ongoing to increase sample
numbers for model construction and testing, and evaluation
of the model prediction performance using heterogeneous
samples is being performed.
Acknowledgements
The authors thank Bryan Warren and Karel Geoes, Registry
pathologists, for their assistance in providing a consensus of
histopathological opinion on the tissue specimens used in this study.
The Gloucestershire Royal Hospital Histopathology Department
provided sample preparation and welcome advice. Bob Bennett
at Renishaw deserves thanks for his collaboration and expertise
in system optimization for Raman tissue measurements. Funding
Copyright  2002 John Wiley & Sons, Ltd.
Classification of epithelial pre-cancers and cancers 573
was provided by MedLINK UK [DoH, Dti, Renishaw, Keymed
(Olympus)].
REFERENCES
1. Coleman MP, Babb P, Damiecki P, Grosclaude P, Honjo S,
Jones J, Knerer G, Pitard A, Quin M, Sloggett A, De Stavola B.
Cancer Survival Trends in England and Wales, 1971–1995:
Deprivation and NHS Region. Office for National Statistics: Studies
in Medical and Population Subjects, vol. 61. HM Stationery
Office: London, 1999.
2. Robin PE, Reid A, Powell DJ, McConkey CC. Clin. Otolaryngol.
1991; 16: 198.
3. Hesketh PJ, Clapp RW, Doos WG, Spechler SJ. Cancer 1989; 64:
526.
4. Blot WJ, Devesa SS, Fraumeni JF. J. Am. Med. Assoc. 1993; 270:
1320.
5. Hameetemaan W, Tygat GNJ, Houthoff HJ, van den Tweel JG.
Gastroenterology 1989; 96: 1249.
6. Cancer Statistics Registrations: Registrations of Cancer Diagnosed
in 1993, England and Wales. Office for National Statistics, Series
MB1, vol. 26. HM Stationery Office: London, 1999; 40–41.
7. Sanfey H, Hamilton SR, Smith RRL, Cameron JL. Surg. Gynecol.
Obstet. 1985; 161: 570.
8. Reid BJ, Rubin CE. Gastroenterology 1985; 88: 1152.
9. Ridell RH. In Barrett’s Esophagus: Pathophysiology, Diagnosis, and
Management, Spechler SJ, Goyal GK (eds). Elsevier: New York,
1985; 188–197.
10. Muir C, Weiland L. Cancer 1995; 75(suppl): 147.
11. Van Sandick JW, van Lanshot JJ, Kuiken BW, Tytgat GN,
Offenhaus GJ, Obertop H. Gut 1998; 43: 216.
12. Melville DM, Jass JR, Merson BC, Pollock Dj, Richman PI,
Shepherd NA, Ritchie JK, Love SB, Lennard-Jones JE. Hum.
Pathol. 1998; 20: 1008.
13. Jensen P, Krogsgaard MR, Christiansen J, Braendstrup O,
Johansen A, Olsen J. Dis. Colon Rectum 1995; 38: 195.
14. Reid BJ, Haggitt RC, Rubin CE, Roth G, Surawicz CM, Van
Belle G, Lewin K, Weinstein WM, Antonioli DA, Goldman H,
MacDonald W, Owen D. Hum. Pathol. 1998; 19: 166.
15. Blackwell KE, Calcaterra TC, Fu Y-S. Ann. Otol. Rhinol. Laryngol.
1995; 104: 596.
16. Mahadevan-Jansen A, Mitchell MF, Ramanujam N, Malpica A,
Thomsen S, Utzinger U, Richards-Kortum R. Photochem.
Photobiol. 1998; 68: 123.
17. Bakker Schut TC, Van Dekken H, Tilanus HW, Bruining HA,
Puppels GJ. In Spectroscopy of Biological Molecules: Modern Trends,
Carmona P, Navarro R, Hernanz A (eds). Kluwer: Dordrecht,
1997; 455–456.
18. Stone N, Stavroulaki P, Kendall C, Birchall M, Barr H.
Laryngoscope 2000; 110: 1756.
19. Gniadecka M, Wulf HC, Mortensen NN, Nielsen OF, Chris-
tensen DH. J. Raman Spectrosc. 1997; 28: 125.
20. Boustany NN, Crawford JM, Manoharan R, Dasari RR, Feld MS.
Lab. Invest. 1999; 79: 1201.
21. Shim MG, Song LMWM, Marcon NE, Wilson BC. Photochem.
Photobiol. 2000; 72: 146.
22. Shanmugaratriam K. Histological Typing of Tumours of the Upper
Respiratory Tract and Ear. Springer: Berlin, 1991.
23. Kendall CA, Stone N, Shepherd NA, Geboes K, Warren BF,
Bennett R, Barr H. Gastroenterology in press.
J. Raman Spectrosc. 2002; 33: 564–573