Note: Descriptions are shown in the official language in which they were submitted.
CA 02396883 2002-03-27
WO 01/24699 PCT/CAOO/01187
NON-INVASIVE SCREENING OF SKIN DISEASES BY VISIBLE/NEAR-
INFRARED SPECTROSCOPY
FIELD OF THE INVENTION
The present invention relates generally to the field of spectroscopy.
More specifically, the present invention relates to a method for non-
invasively
diagnosing skin diseases using visible and near-infrared spectroscopy.
BACKGROUND OF THE INVENTION
Skin cancer is the most common human cancer. In 1999, it is
estimated that there will be 70000 new cases of skin cancer in Canada
(Canadian
Cancer Statistics: Toronto: National Cancer Institute of Canada, 1999) and
more
than 1 million new cases in the United States. The clinical diagnosis is often
difficult since many benign skin diseases resemble malignancies upon visual
examination. As a consequence, histopathological analysis of skin biopsies
remains the standard for confirmation of a diagnosis. However, the decision
must
be made as to which and how many suspicious skin diseases to biopsy.
A rapid, non-invasive technique that could be utilized for
characterization of skin diseases prior to biopsy would be useful.
Visible/infrared
(IR) spectroscopy may be that tool (Jackson et at, 1997, Biophys Chem 68:109-
125). The IR spectrum is divided into three regions: near-IR (700-2500 nm),
mid-IR
(2500-50000 nm) and far-IR (beyond 50000 nm). As light in the far-IR region is
completely absorbed by tissues, it is of little use for tissue analysis. Mid-
IR light is
absorbed by a variety of materials in skin, thus providing an insight into
skin
biochemistry. We have shown that biopsies from basal cell carcinoma (BCC),
squamous cell carcinoma (SCC) and melanocytic tumors have distinct mid-IR
signatures when compared to normal skin (McIntosh et at, 1999, J Invest
Dermatol
112:951-956; McIntosh et al, 1999, Biospectroscopy 5:265-275; Mansfield et at,
1999; App/ Spectroscopy, 53:1323-1330). However, the diagnostic potential of
mid-
IR spectroscopy in-vivo is limited, since complete absorption of mid-IR light
results
with samples greater than 10-15 m in thickness. In contrast, near-IR light is
scattered to a much greater extent than it is absorbed, making tissues
relatively
CA 02396883 2009-05-20
-2-
transparent to near-IR light, thus allowing the examination of much larger
volumes of
tissue and the potential for in-vivo studies.
The near-IR region is often sub-divided into the short (680-1100 nm)
and long (1100-2500 nm) near-IR wavelengths, based upon the technology
required
to analyze light in these wavelength regions. At shorter near-IR wavelengths,
the
heme proteins (oxy- and deoxyhemoglobin and myoglobin) and cytochromes
dominate the spectra, and their absorptions are indicative of regional blood
flow and
oxygen consumption. Long wavelength near-IR absorptions arise from overtones
and
combination bands of the molecular vibrations of C-H, N-H and O-H groups. The
absorption of near-IR light therefore provides information concerning tissue
composition (i.e. lipids, proteins) and oxygen delivery and utilization.
Acquisition of visible/near-IR data is straightforward. Visible and near-IR
light is brought from a spectrometer to the skin via a fiber optic cable. The
light
penetrates the skin, and water, hemoglobin species, cytochromes, lipids and
proteins
absorb this light at specific frequencies. The remaining light is scattered by
the skin,
with some light being scattered back to the fiber optic probe. The light is
collected by
the probe and transmitted back to the spectrometer for analysis. A plot of the
amount
of light absorbed at each wavelength (the spectrum) is computed. Measurements
are
rapid, non-destructive and non-invasive.
SUMMARY OF THE INVENTION
A method of diagnosing skin disease comprising: providing a patient having a
skin disease selected from the group consisting of dysplastic melanocytic
nevi; banal
nevi; lentigines; actinic keratoses; seborrheic keratoses; basal cell
carcinoma; and
malignant melanoma; emitting a first beam of visible or near-IR light into a
portion of
the skin afflicted with the skin disease; collecting and analyzing reflected
light from
the first beam, thereby producing a disease spectrum; emitting a second beam
of
visible or near-IR light into a control skin portion of the patient which is
not afflicted
with the skin disease; collecting and analyzing reflected light from the
second beam,
thereby producing a control spectrum; comparing the control spectrum and the
disease spectrum; and identifying the skin disease as dysplastic melanocytic
nevi;
CA 02396883 2009-05-20
_2a_
banal nevi; lentigines; actinic keratoses; seborrheic keratoses; basal cell
carcinoma;
or malignant melanoma based on said comparison, said method of diagnosing skin
disease having a rapid acquisition time of minutes.
CA 02396883 2002-03-28
=rinted:27-08-2001 ' FAX 12049425723 ISA21OREPLY M EP00967472.2 - PCTCA 00
01187
3
According to a second aspect of the invention, there is provided a
method comprising:
a) providing a patient having a skin disease;
b) ertlitt. g _a beam ofvisiblelnear lR light into a portion of the
skin afflicted with the skin disease;
c) collecting and analyzing reflected light from the beam, thereby
producing a disease spectrum;
d) emitting a beam of visible/near-!R light into a control skin
portion of the patient which is not afflicted with the skin disease;
e) collecting and analyzing reflected light from the beam, thereby
pi oducing a control spectrum;
f) performing a biopsy on the portion of the skin afflicted with the
skin disease;
g) classifying the skin disease based on the biopsy;
h) assigning the control spectrum and the disease spectrum to a
skin disease group based on the classification; and
creaf+ng a database y repeating
According to a third aspect of the invention, there is provided a
device for diagnosing skin diseases comprising: an emitter for emitting a beam
of
visiblefnear-IR light into a diseased skin portion and a control skin portion;
a
collector for collecting and analyzing reflected light from said beams; a
compiler for
producing a disease spectrum and a control spectrum from said analysis; an
analyzer for comparing the control spectrum and the disease spectrum and
identifying the skin disease based on said comparison-
BRIEF DESCRIPTION OF THE DRAWINGS
Figure 1 shows the mean normal control (n=378) and variance
spectrum. The origin of the major absorption bands are indicated. The variance
is
indicated by the shaded region.
Figure 2 shows paired t-test results comparing normal and skin
lesion near IR- spectra. The mean normalized spectra (blue and red traces) are
4
CA 02396883 2002-03-28
,rinted:27-08-2001 ' FAX 12049425723 ISA21OREPLY = EP00967472.2 - PCTCA 00
01187
38
shown overlaid on p-plot traces (black). The optical density scale refers to
the
spectra, while the p-value scales correspond to the p-plot traces.
Figure 3 shows the difference visiblelnear-IR spectra from skin
lesions. Difference spectra were obtained by subtracting each lesion-normal
painng for each group 8 11M1n fig 7 GCttlrliC tcel' roses rwiu } is%i, (r i) -
anic
lentigines (green), dysplastic nevi (black), banal nevi (pink) and seborrheic
keratoses (brown) are shown. The areas used for analysis of variance (ANOVA)
are shaded over the spectra.
Figure 4 shows optimal classification regions of visiblelnear-iR
CA 02396883 2002-03-27
WO 01/24699 PCT/CAOO/01187
4
spectra from skin lesions. Class average spectra are shown with the regions
for
optimal classification (GA-ORS) indicated in the darkly shaded regions and the
regions that were significant by ANOVA indicated in the lightly shaded
regions.
Three optimal regions were selected for dysplastic vs. banal nevi (a), five
regions
for actinic keratoses vs. actinic lentigines (b), five regions for actinic
keratoses vs.
seborrheic keratoses (c) and four regions for BCC vs. seborrheic keratoses
(d). No
regions were significant by ANOVA for b and c.
DESCRIPTION OF THE PREFERRED EMBODIMENTS
Unless defined otherwise, all technical and scientific terms used
herein have the same meaning as commonly understood by one of ordinary skill
in
the art to which the invention belongs. Although any methods and materials
similar
or equivalent to those described herein can be used in the practice or testing
of the
present invention, the preferred methods and materials are now described. All
publications mentioned hereunder are incorporated herein by reference.
DEFINITIONS
A "skin condition" is a dermatological disorder that manifests as a
rash, irritation or dry skin. Examples of skin conditions are psoriasis,
hives,
eczema, etc.
A "skin lesion" is a circumscribed abnormal area of the skin such as a
tumor, nodule or papule.
A "skin disease" is any abnormal area of the skin caused by disease.
Skin diseases include both skin conditions and skin lesions (but not injuries
due to
external insult such as cuts and burns).
"Actinic keratoses" are reddish, rough areas of damaged skin which
are considered pre-malignant. A small percentage of these lesions develop into
the
malignant tumor, squamous cell carcinoma.
"Basal cell carcinoma" or BCC refers to a slow-growing malignant
epithelial neoplasm. This type of cancer in usually "cured" by surgical
removal if
caught early.
CA 02396883 2002-03-27
WO 01/24699 PCT/CAOO/01187
"Actinic lentigines" are small benign pigmented lesions often referred
to as age or liver spots.
"Dysplastic nevi " refer to atypical moles which are considered to be
pre-malignant or at greater risk of becoming malignant.
5 "Seborrheic keratoses" are common light brown to black skin growths
that are benign.
"Banal or benign nevi" are common benign moles.
The purpose of this study was to determine whether the information
obtained from visible/near-IR spectroscopy for a variety of skin diseases
would
prove to be sufficiently characteristic as to be diagnostic. Spectra from six
types of
skin lesion were collected, and univariate and multivariate techniques were
used to
determine whether differences existed between the skin lesions.
Specifically, visible/near-IR spectra were recorded for a number of
patients having skin lesions, as described below. In addition, a spectrum was
taken of an unaffected skin portion as a control from each patient. A biopsy
was
also performed on the skin lesion and the results of the biopsy were used to
assign
the skin lesion to a specific category. The disease spectra and the control
spectra
were then compared using statistical analysis as described below to detect
wavelength regions of significant difference between the control spectra and
the
lesion spectra. These results were then grouped by skin lesion category based
on
the biopsy results. As discussed below, the grouped spectra showed
characteristic
patterns in the differential spectra over a specific set of wavelengths. As a
consequence, these differences can be used to identify or diagnose a skin
disease
by comparing the visible/near-IR spectrum of a control region to a spectrum
taken
of the region of interest.
Specifically, the skin disease is diagnosed by emitting a beam of
visible/near-IR light into a portion of the skin afflicted with the skin
disease, and
collecting and analyzing reflected light from the beam, thereby producing a
spectrum of the diseased skin portion. The process is repeated for an
unaffected
region of skin, thereby providing a control spectrum. The control spectrum and
the
disease spectrum are then compared and the skin disease is identified based on
the comparison.
CA 02396883 2002-03-27
WO 01/24699 PCT/CAOO/01187
6
The skin disease is selected from the group consisting of dysplastic
melanocytic nevi; banal nevi; lentigines; actinic keratoses; seborrheic
keratoses;
basal cell carcinoma; and malignant melanoma.
The control spectrum and the disease spectrum may be compared at
wavelengths corresponding to visible/near-IR absorption by oxyhemoglobin,
deoxyhemoglobin, water, proteins, lipids or combinations thereof. The
wavelengths
may be selected from the group consisting of: 518-598 nm; 618-698 nm; 718-798
nm; 918-998 nm; 1158-1238 nm; 1418-1498 nm; 1718-1798 nm; and combinations
thereof.
In another embodiment of the invention, spectra are taken of affected
and control regions from several patients. A biopsy is then performed on each
of
the affected region, which is then used to positively identify the skin
condition. The
spectra are grouped according to skin condition, thereby forming a database.
The
control spectra and the disease spectra in each skin disease group in the
database
are then reduced to diagnostic wavelengths using a region selection algorithm.
This algorithm is then used to analyze spectra from other skin portions so
that the
disease afflicting the skin portion can be identified based solely on the
spectrum,
without performing a biopsy.
EXAMPLE I - SUBJECT SELECTION
A total of 195 cases were sampled from a study population of 153
(83 women and 70 men) referred to a dermatology clinic for definitive
diagnosis of
a skin disease, and for whom proper management necessitated a biopsy of their
lesion(s). Upon decision by the dermatologist that a biopsy(ies) was required,
the
patient was referred to the study nurse and the spectrum was recorded.
Subjects
were excluded from the study if they: 1) were using any skin medication on the
site
of the lesion, 2) were presently undergoing radiotherapy or chemotherapy, 3)
had
either Type I or Type 11 diabetes (which may alter blood flow in the skin).
Following
an explanation and discussion of the study, informed consent was obtained.
Ethical approval for this study was obtained from the Research Ethics Board of
the
National Research Council of Canada.
CA 02396883 2002-03-27
WO 01/24699 PCT/CAOO/01187
7
EXAMPLE II - ACQUISITION OF SPECTRA
Spectra were recorded in the 400-2500 nm range in 2 nm steps
using a commercial spectrometer (Foss NIRSystems Model 6500) equipped with a
bifurcated visible/near-IR fiber optic probe with a 7 mm active area. Each
reflectance spectrum was collected with a 10 nm slit width, and consisted of
32
scans, which were co-added to improve signal to noise. Prior to obtaining the
readings, the subject's skin and the end of the probe were cleansed with 70%
alcohol. The fiber optic probe was then positioned 0.5 mm from the measurement
site by measuring with a micrometer. For all 195 cases, three (3) visible/near-
IR
spectra were taken from: 1) the lesion and 2) an area of normal appearing skin
(the control site). Acquisition of each spectrum took 40 seconds.
After acquisition of visible/near-IR spectra, a biopsy of the lesion was
taken. Biopsies were sent to the pathologist, and hematoxylin and eosin
stained
sections of formalin fixed, paraffin embedded slides were evaluated. Based on
the
histopathology, spectra were grouped into one of six lesion categories: 1)
actinic
keratoses (33 cases, 99 spectra), 2) BCC (32 cases, 96 spectra), 3) dysplastic
melanocytic nevi (13 cases, 39 spectra), 4) actinic lentigines (12 cases, 36
spectra), 5) banal common acquired nevi (22 cases, 19 intradermal and 3
compound nevi, 66 spectra) and 6) seborrheic keratoses (18 cases, 54 spectra).
A
total of 130 cases were thus included in the data set. The remaining 65 cases
either did not fit into one of the above categories or the patient declined to
have a
biopsy after the measurements. The histopathology was the "gold standard" by
which spectra were classified.
EXAMPLE III - SPECTRAL PROCESSING AND ANALYSIS
Significant noise was apparent in the 1850-2400 nm region due to
the strong absorption of light by water in that spectral range. Prior to data
analysis
spectra were therefore truncated to 400-1840 nm, leaving a total of 720 data
points
per spectrum. Spectra were pooled according to the above 6 lesion categories.
Spectra were pre-processed by normalizing to their total area and offset
correcting.
The mean and standard deviation spectrum for each lesion category
was generated by calculating the mean (+/- SD) intensity at each of the 720
CA 02396883 2002-03-27
WO 01/24699 PCT/CAOO/01187
8
spectral data points for each category. Any spectrum that lay outside 2
standard
deviations from the mean for each lesion group was removed from the study. It
is
interesting to note that in all instances spectra that lay outside 2 standard
deviations were associated with patient movement as recorded by the study
nurse.
The remaining spectral database consisted of 94 (of 99) actinic keratosis
spectra,
90 (of 96) BCC spectra, 38 (of 39) dysplastic nevus spectra, 33 (of 36)
actinic
lentigo spectra, 63 (of 66) banal nevi and 49 (of 54) seborrheic keratosis
spectra.
Mean spectra for individual lesions were then calculated, which resulted in 33
actinic keratoses, 34 BCC, 13 dysplastic nevi, 12 actinic lentigines, 22 banal
nevi
and 18 seborrheic keratoses spectra.
The same procedure was followed for control spectra. A total of 378
spectra from 390 possible control spectra (acquired from 130 sites) were found
to
lie within 2 standard deviations of the mean spectrum. Once again, control
spectra
that lay outside 2 standard deviations from the mean were associated with
patient
movement. Control spectra for each control site were then averaged, resulting
in
130 control spectra.
For each of the six skin lesion categories, paired t-tests (Statistica
5.1. StatSoft, Tulsa, OK) were applied to find significant differences between
lesion
spectra and control skin spectra. The resulting p-values were plotted against
wavelength, as discussed below. Subsequently each mean control spectrum was
subtracted from each mean lesion spectrum in a pair-wise fashion to emphasize
differences between spectra. This resulted in one difference spectrum for each
case, representing spectral differences between the lesion and control site.
Based
upon t-test results, seven regions were selected in which to perform repeated
measures analysis of variance (ANOVA) on difference spectra, as discussed
below. Fisher's least significant difference (LSD) and Duncan's multiple range
tests
were performed post hoc (Statistica 5.1, StatSoft), as discussed below.
In addition to univariate statistical tests, data was subjected to
multivariate analysis. In the first step of the multivariate analysis, an
optimal region
selection genetic algorithm (GA-ORS) (Nikulin et al, 1998, NMR Biomed 11:209-
216) was applied to determine the 3-5 most discriminatory regions of the
difference
spectra. The data sets were then reduced to only those wavelength regions and
CA 02396883 2002-03-27
WO 01/24699 PCT/CAOO/01187
9
linear discriminant analysis (LDA) was performed using a "leave-one-out" cross
validation strategy (Eysel et al, 1997, Biospectroscopy 3:161-167; Mansfield
et al,
1999, Vib Spectrosc 19:33-45). LDA returns a value ranging between 0 (not
belonging) and 1 (belonging) to each spectrum in a data set, indicating the
membership in each class. Thus, the values returned provide an indication of
the
likelihood of a spectrum belonging to each class. Each spectrum is then
allocated
to the class to which it most belongs.
EXAMPLE IV - RESULTS
The mean control (i.e. from normal skin) visible/near-IR spectrum is
shown in Fig 1. Spectra are plotted showing the amount of light absorbed by
the
skin at each wavelength between 400-1840 nm. Each peak in the spectrum can be
assigned to a specific compound found in the skin. Visually, strong absorption
bands arising from O-H groups of water dominate the spectrum. However, much
information is present in the weaker spectral features. For instance, the
relatively
strong absorption feature at -550 nm arises from hemoglobin species and
provides information relating to the oxygenation status of tissues. Further
information on tissue oxygenation can be obtained from analysis of a weak
absorption feature at 760 nm, arising from deoxyhemoglobin (Stranc et al,
1998, Br
J Plast Surg 51:210-217). Compositional information can be obtained from an
analysis of two absorption bands between 1700-1800 nm associated with C-H
groups of skin lipids. In addition, a series of weak absorption bands arising
from
protein N-H groups is found in close proximity (usually overlapped by) the
strong
water absorptions. In addition to information on tissue composition (lipid,
protein
and water content) and tissue oxygenation, information on tissue
architecture/optical properties can be obtained from the spectra. Changes in
tissue
architecture/optical properties may affect the basic nature of the interaction
of light
with the tissue. For example changes in the character of the epidermis (i.e.
dehydration) may result in more scattering of light from the surface, reducing
penetration of light into the skin in a wavelength dependant manner. Also,
different
tumor densities (i.e. nodular vs. diffuse) may result in more scattering of
light from
CA 02396883 2002-03-27
WO 01/24699 PCT/CAOO/01187
the surface. Such phenomena would be manifest in spectra as changes in the
slope of the spectral curves, especially in the 400-780 nm region.
The variance observed at each point in each of the spectrum (n=378)
(variance spectra) is also plotted in Fig 1. The variance spectra appear
essentially
5 identical in form to the mean spectrum, the major difference being a slight
offset.
Variance is essentially constant across the spectral range used. This suggests
that
spectra are highly reproducible, with only slight differences in absorption
intensity
observed across the spectrum (most likely due to small differences in probe
placement).
10 Mean spectra for each type of lesion are shown in Fig 2. No obvious
qualitative differences were observed in spectral groups. To assess whether
significant differences existed between control and abnormal skin, paired t-
tests
were applied at each wavelength. The resulting p-values were plotted against
wavelength (p-plots). In Fig 2 mean normalized lesion spectra (red traces) and
control spectra (blue traces) are shown overlaid on corresponding p-plots
(black
traces). Several areas of the resulting p-plot contained contiguous regions of
statistically significant p-values (p<0.05). Each lesion-normal comparison
exhibited
a slightly different p-plot, and therefore, a distinct pattern of
significance.
Based upon the p-plots, the following regions were chosen in which
to perform repeated measures ANOVA on difference spectra: 1) 518-598 nm, 2)
618-698 nm, 3) 718-798 nm, 4) 918-998 nm, 5) 1158-1238 nm, 6) 1418-1498 nm,
7) 1718-1798 nm (shaded regions in Fig 3). Fisher's LSD and Duncan's Multiple
Range tests, multiple comparison tests that are designed to correct for
multiple
pair-wise comparisons, were performed post-hoc. As shown in Table I, both LSD
and Duncan's tests showed various significant inter-group differences between
the
lesion groups, depending on the region tested. Spectra from dysplastic nevi
were
significantly different from actinic keratoses, BCC, lentigines, banal nevi
and
seborrheic keratoses in a number of spectral regions. In addition, BCC spectra
were significantly different from banal nevi and seborrheic keratoses in three
spectral regions, and seborrheic keratoses were different from lentigines in
one
spectral region.
Two class LDAs were performed on the following comparisons: 1)
CA 02396883 2002-03-27
WO 01/24699 PCT/CAOO/01187
11
dysplastic vs. banal nevi, 2) dysplastic nevi vs. lentigines, 3) actinic
keratoses vs.
lentigines, 4) actinic keratoses vs. seborrheic keratoses, 5) BCC vs.
seborrheic
keratoses, 6) BCC vs. banal nevi and 7) dysplastic nevi vs. seborrheic
keratoses.
Prior to performing the LDA, optimal regions were identified by the GA-ORS
algorithm. Figure 4 shows the optimal regions for comparisons 1, 3, 4 and 5.
LDA
resulted in an overall accuracy of 97.7-72.4% compared to a clinical accuracy
(by
visual examination) of 100-78.0% and are shown in Table II. For each
comparison
in Table II, the numbers in rows represent the histopathological
classification, while
results in columns represent the calculated classification.
EXAMPLE VI - DISCUSSION
The visible/near-IR spectra of skin presented here exhibit strong
absorption bands from water and a number of weak, but consistent, absorption
bands arising from oxy- and deoxy-hemoglobin, lipids and proteins. However,
visual examination of spectra did not show distinct differences in these
spectral
features that could be used to distinguish between spectra of skin diseases
and
healthy skin. Univariate statistics were therefore applied in order to
determine
whether differences existed between skin lesions and healthy skin.
Subsequently,
multivariate statistics (LDA) were performed in an attempt to objectively
classify
spectra.
As control spectra were acquired from a normal site for each lesion,
paired t-tests were performed on spectra from each disease grouping. The
results
demonstrated that each of the skin lesions studied differed significantly from
normal skin in a number of contiguous regions in the visible/near-IR region.
Although comparisons were only made between each skin lesion group and
control skin, each p-plot exhibited a slightly different pattern of
significance,
suggesting that significant spectral differences existed between the different
types
of skin lesions.
To assess whether statistical differences did indeed occur between
the different types of skin lesions, ANOVA was performed on difference
spectra.
Fisher's LSD and Duncan's multiple range tests were applied post hoc. Spectral
sub-regions were identified for these analyses. Results demonstrated that
CA 02396883 2002-03-27
WO 01/24699 PCT/CAOO/01187
12
significant differences existed between spectra of the different types of
lesions in
all regions tested, except the visible region of the spectrum (region I, 518-
598 nm).
However, differences in no one spectral region were sufficient to allow
differentiation between all of the lesion groups.
Some general comments may be made concerning the nature of the
spectral differences identified by univariate statistics. At least two of the
spectral
regions exhibiting significant differences (by ANOVA) are associated with
absorption bands from hemoglobin species. Specifically, the region 718-798 nm
contains the absorption of deoxyhemoglobin, while the region 918-998 nm
contains a broad absorption associated with oxyhemoglobin. Thus, significant
differences between lesion and control spectra in these regions may be
indicative
of changes in oxygenation or blood flow. The regions 1158-1238 nm and 1418-
1498 nm contain significant absorption bands from water, and possibly some
contribution from protein N-H groups. Thus, it appears as if changes in the
amount
or structure of water in tissues occur between some types of lesion and
control
tissues. Finally, spectral bands attributed primarily to C-H groups of skin
lipids
populate the region 1718-1798 nm. Significant differences between spectra in
this
region may imply differences in the amount or structure of skin lipids.
Application of univariate statistics showed that significant differences
not only exist between spectra of healthy skin and the six lesions studied,
but also
between spectra of the lesions. Whilst this is encouraging, significant
differences
are not necessarily diagnostic differences. To assess whether there were
spectral
differences with diagnostic value, a pattern recognition technique, genetic
algorithm guided linear discriminant analysis (GA-LDA), was applied to the
data.
GA-LDA makes use of the fact that clinical information is available regarding
the
spectroscopic data (i.e. biopsy reports). This information is used to train an
LDA
algorithm to recognize the particular combinations of peak frequencies,
absorption
bandwidths, relative intensities, etc. that are characteristic of spectra from
a
particular clinical grouping. The trained LDA algorithm can then be applied to
unknown spectra, and the unknown spectra are partitioned into one of the
clinical
groupings based upon the spectral pattern found. The advantage of LDA is that
a
combination of spectral regions (which perhaps on their own do not contain
CA 02396883 2002-03-27
WO 01/24699 PCT/CAO0/01187
13
sufficient information to allow diagnosis), rather than individual regions,
are used to
achieve a diagnosis.
Specifically, the genetic algorithm starts at one end of an N-point
spectrum by selecting a window consisting of M << N adjacent data points.
Typically, M = 10-12. Discriminant analysis is carried out with these M points
as
local attributes, and the average classification accuracy on the test subsets
is
recorded. The window is advanced by M/2 data points along the spectrum and the
process is repeated. When the spectra are fully traversed, the nonoverlapping
subregions are sorted in decreasing order of accuracy. If the best subregion
found
satisfies a prescribed accuracy (typically > 90%), the subregion selection
process
is terminated. If this does not occur, the next stage is initiated. Typically,
the best
6-8 subregions are tested in all possible combinations. The most parsimonious
combination that satisfies the accuracy criterion provides the feature set for
the
final classifier. The linear discriminant analysis program takes the regions
selected
by the algorithm and identifies the hyperplane that optimally separates the
sets of
points corresponding to the spectral classes of interest. Specifically, class
assignment of any given spectrum involves computing its distance from all
class
centroids (i.e. the representative class average spectrum) and allocating it
to the
class whose centroid is nearest. Thus, for each spectrum, a value ranging
between
0 (not belonging) and 1 (belonging) is given, indicating the membership in
each
class, with the sum of the membership values for all classes being unity. The
value
returned therefore provides an indication of the likelihood of the spectrum
belonging to each class. Thus, for spectra arising from BCC, an ideal LDA
would
return values of 1 for the BCC class and 0 for the other classes in the
comparison.
As will be appreciated by one knowledgeable in the art, the above is
intended as an illustrative example. Other suitable analytical methods may
also be
used.
GA-LDA was applied to difference spectra from benign and
premalignant/ malignant lesion groups. Some of the more difficult visual
diagnoses
were successfully distinguished. All LDA comparisons save one resulted in an
accuracy rate greater than 80%. Although the clinical (visual) diagnostic
accuracy
rate in this particular study was high (greater than 78%), other studies
report
CA 02396883 2002-03-27
WO 01/24699 PCT/CAOO/01187
14
clinical diagnostic accuracy rates of 42-65% (Pichter et al, 1991, Br J
Dermatol 125
(Suppi 38):93-97; Hallock and Lutz, 1998, Plast Reconstr Surg 101:1255-1261).
The LDA results presented here compare favorably with such studies. Spectral
regions that contained diagnostic information were not the same as those
identified
by ANOVA, perhaps reflecting the fact that LDA uses combinations of regions
(each of which on it's own may not show significant differences between
classes)
to enable diagnosis. However, many spectral regions identified by GA-LDA
suggest essentially the same biochemical basis for distinguishing between
classes
as by ANOVA. For example regions around 760 nm (deoxyhemoglobin), 900 nm
(oxyhemoglobin) and 1200 nm (water) allowed discrimination between actinic
keratoses and actinic lentigines. However, in some cases the biophysical basis
underlying the diagnostic regions remains unclear.
The ANOVA and LDA results are both positive steps towards the
differential diagnosis of skin cancer. For example, from a clinical
perspective, it is
particularly noteworthy that dysplastic nevi exhibited a highly significant
difference
(p<0.001) from almost all other lesion groups across most of the regions
tested by
ANOVA. In addition, classification between dysplastic and banal nevi had the
highest accuracy of all classifications (97.7%), with classification between
dysplastic nevi and lentigines close behind (92%). Although there is debate
over
the propensity of dysplastic nevi to develop into malignant melanoma, the
accurate
and early diagnosis of dysplastic nevi is a significant development in the
recent
emphasis placed on melanoma detection. The differentiation of the pre-
malignant
(Callen et al, 1997, J Am Acad Dermatol 36:650-653) actinic keratosis from an
early SCC, seborrheic keratosis or lentigo is of clinical import and ANOVA was
not
successful in this regard. However, LDA differentiated actinic keratoses from
lentigines and seborrheic keratoses with an accuracy of 88.9% and 84.3%,
respectively. It has been suggested that clinicians focus more on the features
of
seborrheic keratoses for differential diagnosis of skin cancer (Marks et al,
1997, J
Am Acad Dermatol 36:721-726), as seborrheic keratosis is perhaps the most
common lesion considered in the differential diagnosis of melanoma in older
persons (Rivers and Gallagher, 1995, Cancer 75:661-666). Our results showed
significant differences between seborrheic keratoses, dysplastic nevi, BCC and
CA 02396883 2002-03-27
WO 01/24699 PCT/CAOO/01187
lentigines by ANOVA.
EXAMPLE VII - CONCLUSIONS
This is the first extensive visible/near-IR spectroscopic study of the
5 non-inflammatory skin lesions most commonly encountered in a general
dermatology clinic. The visible/near-IR spectroscopic technique has clear
potential
for the non-invasive diagnosis of skin diseases, differentiating between
normal skin
and a variety of common skin lesions. More importantly, it appears that
visible/near-IR spectroscopy holds promise for the discrimination of malignant
from
10 benign skin tumors.
Visible/near-IR spectroscopy could form the basis of a clinical
method to diagnose skin diseases. It is rapid (i.e. acquisition time of
minutes),
simple to perform and non-invasive. Measurements are accurate and
reproducible.
Collection of spectra causes little or no patient discomfort, does not alter
the basic
15 physiology of the skin, poses no hazard to the patient and does not
interfere with
any other standard clinical diagnostic practices. The test could be performed
by a
non-specialist and, therefore, might be a useful tool for pre-screening skin
diseases.
While the preferred embodiments of the invention have been
described above, it will be recognized and understood that various
modifications
may be made therein, and the appended claims are intended to cover all such
modifications which may fall within the spirit and scope of the invention.
CA 02396883 2002-03-27
WO 01/24699 PCT/CA00/01187
16
TABLE I. Statistically significant p values (p<0.05) from Duncan's multiple
range
and Fisher's least significant difference (LSD) tests for the seven regions
tested.
Region Significant comparisons p Significant comparisons p
(nm) (Duncan's) value (Fisher's LSD) value
No significance Not No significance Not
518-598 applic applic
-able -able
Dysp. nevi vs actinic 0.007 Dysp. nevi vs actinic 0.005
keratoses keratoses
II. Dysp. nevi vs BCC 0.001 Dysp. nevi vs BCC 0.001
618-698 Dysp. nevi vs lentigines 0.001 Dysp. nevi vs lentigines 0.002
BCC vs banal nevi 0.010 BCC vs banal nevi 0.001
BCC vs seborrheic 0.005 BCC vs seborrheic 0.001
keratoses keratoses
Lentigines vs seborrheic 0.042 BCC vs actinic keratoses 0.021
kerat.
III. Dysp. nevi vs BCC 0.014 Dysp. nevi vs BCC 0.008
718-798 D s . nevi vs lentigines 0.006 D s . nevi vs lentigines 0.014
IV. Dysp. nevi vs BCC 0.010 Dysp. nevi vs BCC 0.005
918-998 BCC vs banal nevi 0.032 BCC vs banal nevi 0.006
BCC vs seborrheic 0.047 BCC vs seborrheic 0.018
keratoses keratoses
Dysp. nevi vs actinic 0.005 Dysp. nevi vs actinic 0.003
keratoses keratoses
V. Dysp. nevi vs BCC 0.001 Dysp. nevi vs BCC 0.001
1158-1238 Dysp. nevi vs lentigines 0.001 Dysp. nevi vs lentigines 0.002
Dysp. vs banal nevi 0.022 Dysp. vs banal nevi 0.026
Seborrheic keratoses vs 0.019 Seborrheic keratoses vs 0.005
BCC BCC
Seborrheic kerat. vs 0.022 Seborrheic kerat. vs 0.032
lentigines lentigines
Dysp. nevi vs actinic 0.001 Dysp. nevi vs actinic 0.001
VI. keratoses keratoses
1418-1498 Dysp. nevi vs BCC 0.001 Dysp. nevi vs BCC 0.001
Dysp. nevi vs lentigines 0.001 Dysp. nevi vs lentigines 0.001
Dysp. vs banal nevi 0.001 Dysp. vs banal nevi 0.001
Dysp. nevi vs seborrheic 0.002 Dysp. nevi vs seborrheic 0.007
kerat. kerat.
Dysp. nevi vs actinic 0.001 Dysp. nevi vs actinic 0.001
VII. keratoses keratoses
1718-1798 Dysp. nevi vs BCC 0.001 Dysp. nevi vs BCC 0.001
Dysp. nevi vs lentigines 0.001 Dysp. nevi vs lentigines 0.001
Dysp. vs banal nevi 0.001 Dysp. vs banal nevi 0.001
Dysp. nevi vs seborrheic 0.001 Dysp. nevi vs seborrheic 0.001
kerat. kerat.
CA 02396883 2002-03-27
WO 01/24699 PCT/CAOO/01187
17
TABLE II. Linear discriminant analysis (LDA) results.
Dysplastic Banal Accuracy Accuracy
Nevi nevi by LDA by clinician
Dysplastic nevi 13 a 0 100 91.7
97.7 89.6
Banal nevi 1 21 95.5 87.5
Dysplastic Actinic
Nevi lentigines
Dysplastic nevi 12 1 92.3 100
92.0 100
Actinic lenti ines 1 11 91.7 100
Actinic Actinic
Keratoses lentigines
Actinic keratoses 31 2 93.9 96.0
89.9 78.0
Actinic lentigines 3 9 75.0 60.0
Actinic Seborrheic
Keratoses keratoses
Actinic keratoses 31 2 93.9 96.0
84.3 94.4
Seborrheic 6 12 66.7 92.8
keratoses
BCC Seborrheic
keratoses
BCC 31 1 96.9 96.8
81.8 94.8
Seborrheic 6 12 66.7 92.8
keratoses
BCC Banal nevi
BCC 31 1 96.9 100
81.5 91.1
Banal nevi 9 13 59.1 82.3
Dysplastic Seborrheic
Nevi keratoses
Dysplastic nevi 8 5 61.5 100
72.4 100
Seborrheic 3 15 83.3 100
keratoses
a Numbers in rows represent the histopathological classification, while
results in
columns represent the calculated LDA classification. The numbers in bold are
therefore correct classifications. Numbers in bold italics are overall
accuracy.
