NON-INVASIVE MELANOMA DIAGNOSIS USING

MULTISPECTRAL IMAGING

Ianisse Quinz

an Su

arez

, Pedro Latorre Carmona

, Pedro Garc

ıa Sevilla

, Enrique Boldo

Filiberto Pla

, Vicente Garc

ıa Jim

enez

, Rafael Lozoya

and Guillermo P

erez de Luc

ıa

Institute of New Imaging Technologies, Jaume I University, Castell

on, Spain

Fundaci

on de la Comunidad Valenciana, Hospital Provincial de Castell

on, Castell

on, Spain

Keywords:

Melanoma detection, Multispectral images, Support Vector Machines, Class imbalance, SMOTE.

Abstract:

The early analysis of pigmented skin lesions is important for clinicians in order to recognize malignant

melanoma. However, it is difﬁcult to differentiate it from benign skin lesions due to their similarity based on

their appearance. Since melanoma has a tendency to grow inside the skin and the depth of penetration of light

into the skin is wavelength dependent, a multispectral imaging acquisition and processing approach to clas-

sify pigmented lesions as melanoma seems appropriate. This paper presents a method to diagnose melanoma

lesions over a group of 26 samples acquired with a multispectral system, where 6 of them are melanomas, and

the other 20 are other types of pigmented lesions. A Leave-One-Out strategy is used to create the training/test

set. The classiﬁcation imbalance problem inherent to this dataset is alleviated using a SMOT E technique. The

random component of the SMOT E methodology is dealt with running it 25 times and a Qualiﬁed Majority

Voting (QMV ) scheme is used to do the ﬁnal classiﬁcation, using SV M. Results show this strategy allows to

obtain competitive classiﬁcation quality results.

1 INTRODUCTION

Melanoma is becoming an important health problem

worldwide. Just in Europe, 32107 cases for men and

35324 cases for women were detected in 2008, which

represents an Age Standardized Rate (European) per

100000 of 11.5 and 11.4, respectively (ECO, 2011).

Similar statistics appear in the United States (Jemal

et al., 2010).

The current procedure for the detection of skin

cancers in general is a clinical examination followed

by a tissue biopsy and histopathology. However, there

are several limitations associated to the process, and a

recent clinical study showed that approximately 40%

of suspicious biopsied lesions were classiﬁed as be-

nign by pathologists (Mogensen and Jemec, 2007).

Therefore, there is an urgent need for a real-time, non-

invasive diagnostic method that can reduce morbidity

and mortality for these cancers.

On the other hand, it is also crucial to diagnose

melanoma early in its evolution, because it has been

proven that its prognosis is directly related to the

depth of the lesion. The ﬁrst attempts to make an

early diagnosis a reality came in 1985 with the ABCD

diagnosis procedure (Friedman et al., 1985; Rigel and

Friedman, 1993). Later, Abbasi (Abbasi et al., 2004)

revised this procedure adding a new parameter, E of

Evolving. Incorporation of the ABCD methodology

into the evaluation campaigns was made in the mid

1980s. Since then, mass screenings have been under-

taken. In the 1990s light-based visual technologies

were incorporated to augment the early diagnosis ca-

pability.

The depth of penetration of light into the skin is

wavelength dependent (Rigel et al., 2010). Informa-

tion found at different depths is useful in order to dif-

ferentiate benign pigmented skin lesions from malig-

nant lesions. Here lies the potential to use multispec-

tral imaging systems as noninvasive methods to help

distinguish melanoma from other lesions.

Traditionally, multispectral imaging systems have

been used in the ﬁeld of remote sensing images of the

Earth taken from satellites, but this technology is in-

creasingly being introduced in other application ﬁelds

such as medicine, biology, and art, among others.

Medicine, and particularly dermatology, seems to

be a natural application ﬁeld, because it can provide

tools of special interest for new types of diagnostic

tests and control treatments. It also has the advantage

that the necessary data acquisition application is fast,

386

Quinzán Suárez I., Latorre Carmona P., García Sevilla P., Boldo E., Pla F., García Jiménez V., Lozoya R. and Pérez de Lucía G..

NON-INVASIVE MELANOMA DIAGNOSIS USING MULTISPECTRAL IMAGING.

DOI: 10.5220/0003843803860393

In Proceedings of the 1st International Conference on Pattern Recognition Applications and Methods (PRARSHIA-2012), pages 386-393

ISBN: 978-989-8425-98-0

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

simple and, above all, non-invasive. Besides, multi-

spectral images, in fact, may give us useful informa-

tion that can not be found in other parts of the wave-

length spectrum.

The aim of this paper is to present a method to

diagnose melanoma lesions over a group of samples

acquired with a multispectral system, where there is a

class imbalance problem alleviated using a technique

called SMOTE (Chawla et al., 2002). The structure

of the paper is as follows: Section 2 describes the

sate of the art in the diagnosis of melanoma and other

pigmented skin lesions. Section 3 explains multi-

spectral image acquisition and processing tools used.

Section 4 explains the classiﬁcation method used and

the SMOT E technique, and results and discussion are

presented in Section 5. Conclusions are given in Sec-

tion 6.

2 MELANOMA DIAGNOSIS

As stated in Section 1, early melanoma detection has

traditionally been based on the ABCD/ABCDE sys-

tems, which provide a set of standards to facilitate vi-

sual recognition of these tumors. ABCDE stands for:

• A. Asymmetry: Melanoma lesions are typically

irregular (asymmetric), benign moles are usually

circular (symmetrical).

• B. Border: Melanoma lesions often have irregular

edges, whereas benign moles have smooth edges.

• C. Colour: Injuries have many brown and black

textured shapes. In the case of benign moles, usu-

ally only a brown spot appears.

• D. Diameter: Typically more than 6 millimeters

in diameter. The diameter in the case of moles is

usually much lower.

• E. Evolving: Comparison of the evolution of the

lesion with time.

However, in the early stages of melanoma the above

features are hardly visible and can lead to misdiagno-

sis. There have been various attempts to automatically

obtain the ABCD characteristics. The ﬁrst problem in

these systems is the segmentation accuracy (Carrara

et al., 2005). Once the image has been segmented,

both morphological (Binder et al., 1998) as well as

texture features (Deshabhoina et al., 2003) have been

used. Segmentation following the ABCD rules has

been done on grey scale images as well as for each

band in multispectral images (Tomatis et al., 2003).

In the last decade, melanoma detection has in-

cluded other parameters of study apart from the

ABCD criteria, for example colour and spectral in-

formation. Cheng (Cheng et al., 2008) extracted a

group of 17 features, the ﬁrst 2 of geometric type,

and the rest related to a relative colour space made

subtracting the average (R,G,B) values of the nor-

mal skin from the corresponding values of the lesion

image. After this, colour parameters related to the

mean colour and its standard deviation, as well as

skewness, energy and entropy, were assessed. R. Joe

Stanley (Stanley et al., 2007) also used a classiﬁca-

tion strategy based on using a relative colour space.

In this case, they built a colour histogram based on

this space and classiﬁed the lesion as melanoma or

non-melanoma depending on the percentage of pix-

els considered as melanoma. A pixel is considered

as melanoma if it falls in the same histogram area

where training melanoma pixels are located. On the

other hand, Stoecker (Stoecker et al., 2011) used tex-

ture measures obtained from the co-occurrence ma-

trix and color features based on the RGB and XY Z

colour spaces (and their corresponding normalized

versions). Tenenhaus (Tenenhaus et al., 2010) de-

veloped a method to classify images as melanoma

or non-melanoma lesions under uncontrolled illumi-

nation conditions creating a 47 dimensional vector,

formed by 5 geometric and 42 colorimetric features.

The geometric features were obtained using a multi-

scale classiﬁcation scheme, mimicking dermatolo-

gists’ approach who ﬁrst look at the image as a whole

and then concentrate on local details to precisely lo-

calize the border.

Kuzmina (Kuzmina et al., 2011a; Kuzmina et al.,

2011b) developed a system consisting of a Nuance

EX (CRi) multispectral camera that acquired images

in the [450,950]nm spectral range using an internal

liquid crystal polarizer and three halogen lamps sur-

rounding the surface to be imaged. After image ac-

quisition and using a three-chromophore absorption

(regression) model they obtained the relative concen-

trations of oxy-hemoglobin, deoxy-hemoglobin and

melanin and considering the feature space of deoxy-

hemoglobin and melanin, they classiﬁed lesions as

melanoma or non-melanoma. Diebele (Diebele et al.,

2011) proposed a melanoma differentiation param-

eter based on the different behaviour of melanoma

from other lesions in terms of the Optical Density

(OD) value at 540, 650 and 950nm, using the ac-

quisition system proposed in (Kuzmina et al., 2011a;

Kuzmina et al., 2011b).

Garcia−Uribe (Uribe et al., 2011) developed an

oblique incidence diffuse reﬂectance spectrometer to

conduct in-vivo measurements of the optical prop-

erties of three different types of pigmented skin le-

sions (melanoma, dysplastic and common nevi). Both

NON-INVASIVE MELANOMA DIAGNOSIS USING MULTISPECTRAL IMAGING

387

absorption and reduced scattering coefﬁcient spec-

tra were estimated from the images acquired in the

[455,765]nm range for 144 pigmented skin lesions

including 16 melanomas. Obtaining the extinction

coefﬁcients of oxy-hermoglobin, deoxy-hemoglobin

and melanin allows to assess the total concentration

of hemoglobin and the oxygen saturation, using this

last ﬁgure as classiﬁcation criteria for melanoma de-

tection.

Sorg (Sorg et al., 2005) used hyperspectral imag-

ing to study oxygen transportation using red ﬂuores-

cent protein (RFP) to identify all tumor cells and

green ﬂuorescent protein (GFP) to identify hypoxic

cells. In this way, maps were created, and (using

hemoglobin saturation) were able to classify each

pixel in the image as tumor cells (where only ﬁnd

RFP) or hypoxic tumor cells (where they ﬁnd RFP

and GFP).

Rajaram (Rajaram et al., 2010) used a CCD cam-

era, a xenon lamp for white illumination, a pulsed ni-

trogen laser at 337nm, and a pulsed nitrogen laser at

445nm for melanoma detection. White illumination

was used in combination with a model of diffuse re-

ﬂectance spectroscopy to obtain parameters related to

the morphology of the tissue under analysis. Laser il-

lumination at 337 and at 445nm was used to activate

two components (NADH and FAD) which are related

to the tissue metabolic activity.

Mazzoli (Mazzoli et al., 2010) used a camera

modiﬁed to contain a LED lighting system able to ac-

quire images at 950nm and a Monte-Carlo simulator

able to infer the depth of the lesion from the acquired

images (at 950nm).

Nagaoka (Nagaoka et al., 2011) developed a hy-

perspectral imaging system using an imaging spec-

trograph (ImSpector V 8E, Specim) and a CCD cam-

era (iXon) with a 1.3nm step width resolution, cov-

ering the [380,780]nm interval. Melanoma detec-

tion was obtained assessing the spectral angle be-

tween the spectral curve corresponding to normal skin

and the curve of the skin lesion. From this angle, a

probability distribution was assessed and used as the

melanoma discrimination index. In particular, statis-

tically signiﬁcant differences between the melanoma

and non-melanoma groups were analyzed using the

Mann−Whitney U− test.

Dhawan (Dhawan et al., 2005) developed an op-

tical imaging system based on multispectral transil-

lumination to provide images of skin lesions show-

ing sub-surface pigmentation and vascular architec-

ture based blood volume information, comparing their

method with the Epiluminescence Light Microscopy

method. Melanoma detection was made using image

segmentation techniques based on the CIE −L

∗

colour space.

Raposio (Raposio and et. al, 2007) developed a

system for the detection of melanoma using a mul-

tispectral linear camera that acquired images in the

[400,1000]nm range. They applied Principal Com-

ponent Analysis (PCA) to the image and selected the

ﬁrst component. Then they assessed the local minima

of the image histogram and used them as threshold

values to segment the different image regions.

Melanoma detection is currently made not only

using colour cameras or multispectral or hyperspec-

tral systems. New ﬁelds of research are being opened

in ultrasound imaging, mRNA−based signature, and

electrical bioimpedance. For a complete survey of

current and new commercial and non-commercial

systems used the reader is referred to Table 1 (and

details therein) of (Rigel et al., 2010).

3 IMAGE ACQUISITION

A set of two systems, consisting of two different cam-

eras and two different Liquid Crystal Tunable Filters

(LCT F) were used to acquire the multispectral im-

ages of the patients’ skin lesions. Since they are tun-

able, the number of wavebands can be changed. A

spectral resolution of 10nm was ﬁxed. The use of the

two ﬁlters is justiﬁed in order to cover the wavelength

range from 400nm to 1100nm, and the use of the two

cameras avoids the need to remove the ﬁlter from the

camera to cover the whole wavelength range. The ﬁrst

camera was a CCD QImaging Retiga EX camera (12-

bit, Monochrome Cooled camera without IR Filter).

The sensor resolution is 1036 ×1360, down-sampled

to 516 ×676 pixels. The second camera was a Marlin

F080B model. The V IS ﬁlter was attached to the Mar-

lin camera, covering the [400, 720]nm spectral range.

The NIR ﬁlter was attached to the QImaging camera

and it covered the [650,1100]nm range (see Figure 1).

In both cases, a Canon TV zoom lens (whose focal

distance varies between 8 and 48mm) was used before

the ﬁlter, and a Macro Schneider system (SCHNEI-

DER, 2011) was used between the ﬁlter and the cam-

era.

The illumination system was a Fiber-Lite DC −

950 system (Dolan−Jenner industries) with a 150W

Quartz halogen lamp, and a 3250

◦

Kelvin colour tem-

perature, connected to the camera optics through a

ﬁber optic ring light guide. A white cylinder was

attached to the optical ring. This was done in order

to homogenize and concentrate light, signiﬁcantly de-

creasing the exposure times required for each spectral

band. It also controlled the distance between the cam-

era and the lesion and prevented external light coming

ICPRAM 2012 - International Conference on Pattern Recognition Applications and Methods

388

Figure 1: Hardware for image acquisition.

into. The cylinder was 20cm long and 7.5cm in diam-

eter. At 20cm distance, an approximately 8 ·8 cm

square Field-of-View was achieved.

3.1 Camera Calibration

The acquisition time per band used for the LCT F ﬁl-

ters was obtained using an ideal reﬂectance diffuser

object, called spectralon, which is built to reﬂect the

same amount of light in the whole wavelength spec-

trum. This time was adjusted so that the image ac-

quired of the spectralon were the same for all the im-

age pixels and for all bands.

In order to accomplish this aim, an automatic cal-

ibration method was implemented. For one band, the

acquisition time and grey level value are considered to

increase linearly, if illumination and distance to object

do not vary. The method consists of looking for two

(time,greylevel) pairs not allocated at extreme val-

ues, and then the exposure time is assessed based on

the linear relationship assumed for an expected grey

value.

3.2 Image Registration

Image registration is the process to obtain a mapping

that allows one image to be superimposed over an-

other image. Multispectral image acquisition is time

consuming, because an image per band has to be ac-

quired. During this time interval, involuntary move-

ments of the patient may happen, and therefore image

registration among spectral bands is a requirement.

A method based on the maximization of the mutual

information between two images (Maes et al., 1997)

(Pluim et al., 2003) was applied. In particular, let us

consider two images, A and B, of size (M +1)·(N +1)

pixels. Registering image B against A can be done

maximizing the following information criterion:

I(A,

B) =

∑

p(a,b) ·log

p(a,b)

p(a) · p(b)

(1)

where:

p(a,b) =

M ·N

∑

δ(a,A(x,y)) ·δ(b,

B(x,y)) (2)

where

B is the transformed version of B, a and b are

the possible grey level values of the image and x and

y, the pixel positions. δ is the delta function. p(a, b)

is the joint probability distribution of the grey levels

of both images, and p(a) (p(b)) is the probability dis-

tributions grey level a (b). The maximization of the

mutual information was achieved using the SIMPLEX

method (Press et al., 1992).

4 CLASSIFICATION

Multispectral image datasets related to melanoma

classiﬁcation are intrinsically high dimensional and

low cardinality spaces where one of the classes

(melanoma) usually contains a lower number of sam-

ples than the others. Therefore, solving this problem

has to account for the relationships between the num-

ber of data points and the dimensionality of the fea-

ture spaces, on the one hand, and on the imbalance

among the classes, on the other hand.

4.1 Support Vector Machines

Support Vector Machines (SV Ms) is a classiﬁcation

technique that is considered robust against datasets

of high dimensionality and low cardinality (Chang

et al., 1999). Given a labeled training data set

{(x

),..., (b f x

)}, where x

∈ R

and y

∈

{−1,+1} and a nonlinear mapping f , usually to a

higher dimensional space, f : R

→ H , the SVM

method solves:

min

w,ξ

{

kwk

+ C

∑

} (3)

constrained to:

·(hf (x

),wi+ b) ≥ 1 −ξ

∀i = 1,..., n (4)

≥ 0 ∀i = 1,. . . ,n (5)

where w and b deﬁne a linear classiﬁer in the feature

space. Transformation f is assessed in such a way

that the classiﬁcation in the higher (transformed) di-

mensional space may be easier. The regularization

parameter C controls the generalization capability of

the classiﬁer, and ξ

are positive variables controlling

the permitted errors.

4.2 Class Imbalance

Class imbalance constitutes one of the problems that

has recently received most attention in research ar-

eas such as Machine Learning, Pattern Recognition

NON-INVASIVE MELANOMA DIAGNOSIS USING MULTISPECTRAL IMAGING

389

and Data Mining (Fern

andez et al., 2011). A two-

class data set is said to be imbalanced if one of the

classes (the minority one) is represented by a very

small number of instances in comparison to the other

(the majority) class (He and Garcia, 2009). It has

been observed that class imbalance often leads to poor

classiﬁcation performance, especially for the minor-

ity classes because these are often biased towards the

majority class. This issue is particularly important in

real-world applications where it is costly to misclas-

sify examples of the minority class.

Data level methods have been developed for deal-

ing with the class imbalance problem. These meth-

ods consist of balancing the original data set, either

by over-sampling the minority class and/or by under-

sampling the majority class until the problem classes

are approximately equally represented. Both strate-

gies can be applied in any learning system since they

act as a preprocessing phase, thus allowing the system

to receive the training instances as if they belonged to

a well-balanced data set. By using this strategy, any

bias of the learning system towards the majority class

due to the skewed class priors will hopefully be elim-

inated.

Several researchers have reported that over-

sampling in general obtains more accurate results than

the under-sampling methods (Batista et al., 2004),

(He et al., 2005), (Hulse et al., 2007). The simplest

method to increase the size of the minority class cor-

responds to random over-sampling, that is, a non-

heuristic method that balances the class distribution

through the random replication of positive examples.

Nevertheless, since this method replicates existing ex-

amples in the minority class, overﬁtting is more likely

to occur. Chawla et al (Chawla et al., 2002) proposed

an over-sampling technique that generates new syn-

thetic minority samples by interpolating between sev-

eral preexisting positive examples that lie close to-

gether. It ﬁrst ﬁnds the k nearest neighbors (k-NN)

belonging to the minority class for each positive ex-

ample and then, the synthetic examples are generated

in the direction of some or all of the nearest neigh-

bors. This method, called Synthetic Minority Over-

sampling TEchnique (SMOTE), allows the classiﬁer

to build larger decision regions that contain nearby

samples from the minority class. Depending upon the

amount of over sampling required, neighbors from the

k nearest neighbors are randomly chosen. Synthetic

prototypes are generated in the following way: take

the difference between the feature vector (instance)

under consideration and its nearest neighbor. Multi-

ply this difference by a random number between 0 and

1, and add it into the feature vector under considera-

tion.

Figure 2: Optical Density curves of melanoma and a pig-

mented benign lesion.

5 EXPERIMENTAL RESULTS

A group of 26 samples (each one with the two avail-

able V IS and NIR components) of skin lesions was ac-

quired by the multispectral system. All of them were

preliminary classiﬁed as suspected pigmented lesions,

and all of them were biopsied. Biopsy determined that

6 of them were melanoma and the other 20 were other

types of lesions.

For each lesion, a region of interest was manually

selected, and the mean value (normalized between 0

and 1) for that region and for all bands was used as

feature vector. Since there is an overlapping spectral

interval in the [650,720]nm region between each pair

of images, a proportionality factor was applied to the

NIR spectral curve. This factor consisted in the mean

over the overlapping spectral interval of the ratios of

the V IS and NIR mean (over the region of interest)

signals.

Figure 2 shows the Optical Density, deﬁned as:

OD = −log[

Skin

Spectr

] (where I

Skin

is the mean curve ob-

tained using a robust mean technique, over all the im-

ages in the database for the same class of lesion, and

Spectr

is the mean curve over a region of interest of

an image acquired of a spectralon) for a melanoma

lesion, and for a pigmented benign lesion. Figure 3

shows the different spectral behaviour (at 650, 800

and 900nm) of a melanoma (ﬁrst row) and a non

melanoma (second row) lesion.

A Leave-One-Out strategy was selected (due to

the low number of samples in the dataset) to create

the training and testing sets formed by the feature vec-

tors. In order to classify them, an SV M classiﬁer with

a Kernel Radial Basis Function (RBF) was used. A

grid search using equally spaced steps in the logarith-

mic space of the SVM tuning parameters (C,ξ), where

ICPRAM 2012 - International Conference on Pattern Recognition Applications and Methods

390

Figure 3: (First row) Spectral images of melanoma. (Sec-

ond row) Spectral images of pigmented lesion, in both cases

acquired at 650, 800 and 900nm.

= ξ,∀i was made to select and ﬁx the best parame-

ters, using a 5-fold cross-validation strategy. In order

to do a quantitative assessment of the classiﬁcation

quality, a typical two-class problem (melanoma or

non-melanoma) confusion matrix as shown in Table

1 was deﬁned. Classiﬁcation results in this case are

for the original dataset (without considering the class

imbalance problem), where Predicted positive means

the case when SVM identiﬁes a lesion as melanoma,

Predicted negative when SV M considers a lesion as

non melanoma, Positive class when the hospital certi-

ﬁes the lesion is melanoma, and Negative class when

it is not. We obtained: T P = 1, T N = 17, FP = 3

and FN = 5 (the results appear in the table as well).

On the other hand, FN is probably, with T P the two

most important values of the four of Table 1. FN = 5

in our case means that SV M has let pass through 5

melanoma lesions, without identifying them correctly.

FP = 3 means that three lesions were wrongly identi-

ﬁed as melanoma.

Table 1: Confusion matrix for the original dataset.

Predicted positive Predicted negative

Positive class True Positive (TP)= 1 False Negative (FN)= 5

Negative class False Positive (FP)= 3 True Negative (TN)= 17

With these measures, the Geometric mean (Gm)

(Kubat and Matwin, 1997) and the Area Under the

Curve (AUC) (Sokolova et al., 2006) were obtained.

The geometric mean is deﬁned as:

Gm =

√

·a

−

(6)

where:

T P

T P + FN

(7)

Figure 4: Plot of the ﬁrst two (highest variance) PCA com-

ponents of the spectral curves of the melanoma and non

melanoma classes (left) before and (right) after applying the

SMOTE technique.

and

−

T N

T N + FP

(8)

The Area Under the Curve (AUC) is deﬁned as:

AUC =

) + (a

−

)

(9)

giving Gm = 0.376 and AUC = 0.508. In order to

analyze the impact of the imbalance between classes

on the classiﬁcation performance, we applied the

SMOTE technique on our dataset. Figure 4 shows the

effect of the SMOT E technique on a group of points

represented onto the ﬁrst two components (with the

highest variance) of the Principal Component Analy-

sis (PCA) decomposition of the feature vectors of the

melanoma and non melanoma classes of the dataset.

The generation of the synthetic prototypes in

SMOTE was made considering k = 3 neighbours.

Due to this random selection, the SMOT E technique

was applied 25 times for each one of the training

datasets created using the Leave-One-Out method.

Since the result of the classiﬁcation on each of the

25 times for each dataset could be different, a Qual-

iﬁed Majority Voting (QMV ) scheme was applied.

This means that a minimum percentage is required

to take a sensitive decision. In our case, if in more

than 70% of the 25 cases, the lesion were classiﬁed as

non melanoma, the scheme would classify the lesion

as non melanoma. In this case, the confusion matrix

gives the results that appear in Table 2.

Table 2: Confusion matrix after the application of SMOT E.

Predicted positive Predicted negative

Positive class True Positive (TP)= 6 False Negative (FN)= 0

Negative class False Positive (FP)= 3 True Negative (TN)= 17

The Geometric mean (Gm) and the Area Under the

Curve (AUC) gave the following results: Gm = 0.922

and AUC = 0.925. Gm and AUC results show that

detection of melanoma, when applying the SMOT E

technique, is feasible. Besides, FN = 0 means that

the proposed technique has not let pass through any

melanoma lesion. However, the number of False

NON-INVASIVE MELANOMA DIAGNOSIS USING MULTISPECTRAL IMAGING

391

Positives remained equal (FP = 3), which means

that three lesions were again wrongly identiﬁed as

melanoma lesions.

6 CONCLUSIONS

This paper proposes a non-invasive melanoma detec-

tion methodology based on the acquisition of multi-

spectral images in the [400,1100]nm spectral range.

The classiﬁcation imbalance problem inherent to this

dataset was alleviated using a SMOTE technique.

The random component of the SMOT E methodology

was dealt with running it 25 times and a Qualiﬁed Ma-

jority Voting (QMV ) scheme was used to do the ﬁ-

nal classiﬁcation (using SVM). This strategy allowed

to obtain good classiﬁcation results (Gm = 0.922 and

AUC = 0.925), as compared with the unbalanced case

(Gm = 0.376 and AUC = 0.508). Also, the number of

False Negatives achieved was zero (FN = 5 for the

unbalanced case). Nevertheless, the number of False

Positives was not reduced by the SMOT E technique.

In any case, the number of samples of the dataset used

is still quite low and therefore, we can only consider

these as preliminary results. With more data points,

future work could include the application of feature

selection and texture characterization techniques.

ACKNOWLEDGEMENTS

This work was supported by the Spanish Ministry

of Science and Innovation under the projects Con-

solider Ingenio 2010CSD2007 −00018, and EODIX

AYA 2008−05965 −C04 −04/ESP, by the Fundaci

Caixa-Castell

o through the project P11B2007 −48,

and by the Generalitat Valenciana through the project

PROMETEO/2010/028.

REFERENCES

Abbasi, N. R., Shaw, H. M., and Riegel, D. S. (2004). Early

diagnosis of cutaneous melanoma: revisiting the abcd

criteria. Journal of the American Medical Association,

292:2771–2776.

Batista, G. E. A. P. A., Prati, R. C., and Monard, M. C.

(2004). A study of the behavior of several methods for

balancing machine learning training data. SIGKDD

Explorations Newsletter, 6(1):20–29.

Binder, M., Kittler, H., Seeber, A., SteinerA, A., Pe-

hamberger, H., and Wolff, K. (1998). Epiluminis-

cence microscopy-based classiﬁcation of pigmented

skin lesions using computerized image analysis and

an artiﬁcial neuronal network. Melanoma Research,

8(3):261–266.

Carrara, M., Tomatis, S., Bono, A., Bartoli, C., Moglia, D.,

Lualdi, M., Colombo, A., Santinami, M., and March-

esini, R. (2005). Automated segmentation of pig-

mented skin lesions in multispectral imaging. Physics

in Medicine and Biology, 50:345–357.

Chang, C. I., Q. Du, T. L. S., and Althouse, M. L. G. (1999).

A joint band prioritization and band-decorrelation

approach to band selection for hyperspectral image

classiﬁcation. IEEE Trans. Geosc. Remote Sens.,

37(6):2631–2641.

Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer,

W. P. (2002). SMOTE: Synthetic minority over-

sampling technique. J. Artif. Intell. Res., 16:321–357.

Cheng, Y., Swamisai, R., Umbaugh, S. E., Moss, R. H.,

Stoecker, W. V., Teegala, S., and Srinivasan, S. K.

(2008). Skin lesion classiﬁcation using relative color

features. Skin Research and Technology, 14:53–64.

Deshabhoina, S. V., Umbaugh, S. E., Stoecker, W. V., Moss,

R. H., and Srinivasan, S. K. (2003). Melanoma and

seborrheic queratosis differentiation using texture fea-

tures. Melanoma Research, 9(4):348–356.

Dhawan, A. P., D’Alessandro, B., Patwardhan, S., and Mul-

lani, N. (2005). An over-sampling expert system for

learing from imbalanced data sets. In Proc. of the In-

ternational Conference on Neural Networks and Brain

(ICNN & B ’05), volume 1, pages 537–541.

Diebele, I., Kuzmina, I., Kapostinsh, J., Derjabo, A., and

Spigulis, J. (2011). Melanoma-nevus differentiation

by multispectral imaging. In Proc. of SPIE-OSA

Biomedical Optics, volume 8087, pages 80872G1–

80872G6.

ECO (2011). Cancer: Melanoma of skin. In European

Cancer Observatory http://eu-cancer.iarc.fr/cancer-

11-melanoma-of-skin.html,en.

Fern

andez, A., Garc

ıa, S., and Herrera, F. (2011). Address-

ing the classiﬁcation with imbalanced data: Open

problems and new challenges on class distribution.

In Corchado, E., Kurzynski, M., and Wozniak, M.,

editors, Hybrid Artiﬁcial Intelligent Systems, volume

6678 of Lecture Notes in Computer Science, pages 1–

10.

Friedman, R. J., Rigel, D. S., and Kopf, A. W. (1985). Early

detection of malignant melanoma: the role of physi-

cian examination and self-examination of skin. CA: A

Cancer Journal for Clinicians, 35:130–151.

He, G., Han, H., and Wang, W. (2005). An over-sampling

expert system for learing from imbalanced data sets.

In Proc. of the International Conference on Neural

Networks and Brain (ICNN & B ’05), volume 1, pages

537–541.

He, H. and Garcia, E. (2009). Learning from imbalanced

data. IEEE Transactions on Knowledge and Data En-

gineering, 21(9):1263–1284.

Hulse, J. V., Khoshgoftaar, T. M., and Napolitano, A.

(2007). Experimental perspectives on learning from

imbalanced data. In Proc. of the 24th international

conference on Machine learning (ICML’07), pages

935–942.

ICPRAM 2012 - International Conference on Pattern Recognition Applications and Methods

392

Jemal, A., Siegel, R., Xu, J., and Ward, E. (2010). Cancer

statistics, 2010. CA: A Cancer Journal for Clinicians,

60:277–300.

Kubat, M. and Matwin, S. (1997). Addressing the curse of

imbalanced training sets: one-sided selection. In 14th

ICML, pages 179–186.

Kuzmina, I., Diebele, I., Jakovels, D., Spigulis, J., Valeine,

L., Kapostinsh, J., and Berzina, A. (2011a). To-

wards noncontact skin melanoma selection by multi-

spectral imaging analysis. Journal of Biomedical op-

tics, 16(6):0605021–0605023.

Kuzmina, I., Diebele, I., Valeine, L., Jakovels, D., Kempele,

A., Kapostinsh, J., and Spigulis, J. (2011b). Multi-

spectral imaging analysis of pigmented and vascular

skin lesions: results of a clinical trial. In Proc. of

SPIE, volume 7883, pages 7883121–7883127.

Maes, F., Collignon, A., Vandermeulen, D., Marchal, G.,

and Suetens, P. (1997). Multimodality image regis-

tration by maximization of mutual information. IEEE

Trans. Med. Imaging, 16(2):187–198.

Mazzoli, A., Munaretto, R., and Scalise, L. (2010). Prelim-

inary results on the use of a noninvasive instrument

for the evaluation of the depth of pigmented skin le-

sions: numerical simulations and experimental mea-

surements. Lasers Med. Sci., 25:403–410.

Mogensen, M. and Jemec, G. (2007). Diagnosis of non-

melanoma skin cancer/keratinocyte carcinoma: a re-

view of diagnostic accuracy of nonmelanoma skin

cancer diagnostic tests and technologies. Dermatol.

Surg., 33:1158–1174.

Nagaoka, T., Nakamura, A., Okutani, H., Kiyohara, Y., and

Sota, T. (2011). A possible melanoma discrimina-

tion index based on hyperspectral data: a pilot study.

Skin Research and Technology, (DOI:10.1111/j.1600-

0846.2011.00571.x):1–10.

Pluim, J. P. W., Maintz, J. B. A., and Viergever, M. A.

(2003). Mutual-information-based registration of

medical images: A survey. IEEE Trans. Med. Imag-

ing, 22(8):986–1004.

Press, W., Teukolsky, S. A., Vetterling, W. T., and Flannery,

B. P. (1992). Numerical Recipes in C: The Art of Sci-

entiﬁc Computing. Cambridge University Press.

Rajaram, N., Aramil, T. J., Lee, K., Reichenberg, J. S.,

Nguyen, T. H., and Tunnell, J. W. (2010). Design

and validation of a clinical instrument for spectral di-

agnosis of cutaneous malignancy. Applied Optics,

49(2):142–152.

Raposio, E. and et. al (2007). Spectrophotometric technol-

ogy for the early detection of cutaneous melanoma.

International Journal of Simulation Systems, Science

& Technology, 8(4):46–54.

Rigel, D. S. and Friedman, R. J. (1993). The rationale of the

abcds of early melanoma. J. Am. Acadm. Dermatol.,

29:1060–1061.

Rigel, D. S., Russak, J., and Friedman, R. (2010). The evo-

lution of melanoma diagnosis: 25 years beyond the

abcds. CA: A Cancer Journal for Clinicians, 60:301–

316.

SCHNEIDER (2011). Industrial optics: Oem. In http://

www.schneiderkreuznach.com.

Sokolova, M., Japkowicz, N., and Szpakowicz, S. (2006).

Beyond accuracy, f-score and roc: A family of dis-

criminant measures for performance evaluation. In

Sattar, A. and Kang, B.-h., editors, AI 2006: Ad-

vances in Artiﬁcial Intelligence, volume 4304 of Lec-

ture Notes in Computer Science, pages 1015–1021.

Sorg, B. S., Moeller, B. J., Donovan, O., Cao, Y., and

Dewhirst, M. W. (2005). Hyperspectral imaging of

hemoglobin saturation in tumor microvasculature and

tumor hipoxia development. Journal Biomedical Op-

tics, 10(4):044004.

Stanley, R. J., Stoecker, W. V., and Moss, R. H. (2007). A

relative color approach to color discrimination for ma-

lignant melanoma detection in dermoscopy images.

Skin Research and Technology, 13:62–72.

Stoecker, W. V., Wronkiewiecz, M., Chowdhury, R., Stan-

ley, R. J., Xu, J., Bangert, A., Shrestha, B., Cal-

cara, D. A., Rabinovitz, H. S., Oliviero, M., Ahmed,

F., Perry, L. A., and Drugge, R. (2011). Detection

of granularity in dermoscopy images of malignant

melanoma using color and texture features. Comput-

erized Medical Imaging and Graphics, 35:144–147.

Tenenhaus, A., Nkengne, A., Horn, J.-F., Serruys, C.,

Giron, A., and Fertil, B. (2010). Detection of

melanoma from dermoscopic images of naevi ac-

quired under uncontrolled conditions. Skin Research

and Technology, 16:85–97.

Tomatis, S., Bono, A., Bartoli, C., Carrara, M., LualdiM,

M., Tragni, G., and Marchesini, R. (2003). Automated

melanoma detection, multispectral imaging and neu-

ronal network approach for classiﬁcation. Melanoma

Research, 30(2):212–221.

Uribe, A. G., Smith, E. B., Zou, J., Duvic, M., Prieto,

V., and Wang, L. V. (2011). In-vivo characteriza-

tion of optical properties of pigmented skin lesions

including melanoma using oblique incidence diffuse

reﬂectance spectrometry. Journal of Biomedical op-

tics, 16(2):0205011–0205013.

NON-INVASIVE MELANOMA DIAGNOSIS USING MULTISPECTRAL IMAGING

393