International Journal of Computer Vision and Signal Proces sing, 3(1), 10-17(2013) ORIGINAL ARTICLE [619375]

International Journal of Computer Vision and Signal Proces sing, 3(1), 10-17(2013) ORIGINAL ARTICLE
A developed system for melanoma diagnosis
Nadia Smaoui
Control and Energy Management laboratory
National School of Engineers of Sfax, Tunisia
Souhir Bessassi
Higher Institute of Computer Science and Multimedia of Gabe s, Tunisia
ISSN: 2186-1390 (Online)
http://CenNSER.org/IJCVSPAbstract
In recent years, there has been a fairly rapid increase in the numbe r
of melanoma skin cancer patients. Melanoma, this deadliest form of s kin
cancer, must be diagnosed early for effective treatment. So, it is n eces-
sary to develop a computer-aided diagnostic system to facilitate its early
detection. In this paper, the proposed work is based on a combinat ion
of a segmentation method and an analytical method and aims to im-
prove these two methods in order to develop an interface that can assist
dermatologists in the diagnostic phase. As a first step, a sequence of
preprocessing is implemented to remove noise and unwanted struct ures
from the image. Then, an automatic segmentation approach locate s the
skin lesion. The next step is feature extraction followed by the ABCD
rule to make the diagnosis through the calculation of the TDV score. In
this research, three diagnosis are used which are melanoma, suspic ious,
and benign skin lesion. The experiment uses 40 images containing sus-
picious melanoma skin cancer. Based on the experiment, the accura cy
of the system is 92% which reflects its viability.
Keywords: melanoma, segmentation, skin cancer, ABCD rule
c/circlecopyrt2013, IJCVSP, CNSER. All Rights ReservedArticle History:
Received:22 March 2013
Revised:6 June 2013
Accepted:6 August 2013
Published Online:9 August 2013
1. INTRODUCTION
Skin cancer is considered as the most common form of
cancer worldwide. The incidence is considerably increas-
ing. For example in the US, at current rates, a skin cancer
will develop in one in five people during their lifetime [1].
Skin cancers can be classified into two major groups which
aremelanomaandnon-melanomaskincancers. Thesetype
of skin cancer (non-melanoma) is usually start in the basal
cells or squamous cells. Such cells are found at the base of
the outer layer of the skin. Approximately 1,200,000 non-
melanoma skin cancers develop in the US. The Exposure
of the skin to sunlight is considered as the main factor be-
hind the development of the most basal and squamous cell
cancers. Basal cell or squamous cell cancerscan be cured if
found and treated early. Melanoma is the most dangerous
type of skin cancer [1]. The World Health Organization
approximates that more than 70230 people a year in the
world die from too much sun, mostly from malignant skin
cancer [2]. Early detection of this cancer can help its cur-
ability. Melanoma arises from the cancerous growth in the
pigmented spots. Dermatologists can diagnose melanoma
∗Corresponding authorin about 80% of cases accordingto ABCD process [3]. Dig-
ital dermatoscopy could give dermatologists a closer look
at suspicious skin lesions. This, in turn, can help derma-
tologists to find suspicious lesions in an early step. To
measure and detect sets of features from dermoscopic im-
ages, the computerized analysis of these images can be ex-
tremely useful and helpful for dermatologists in order to
facilitate their diagnosis. Based on images obtained by
digital dermoscopy, our conclusive aim is to develop an
aided-diagnosticsystem for the identification ofearly stage
melanomas. This would enable supervised classification of
melanocytic lesions. The melanoma detection process is
composed of five steps that are the preprocessing, the seg-
mentation, the post-processing, the feature extraction by
the ABCD rule and finally the classification based on the
TotalDermatoscopicValue calculation(TDV). The ABCD
rule presentsthe followingfeatures which are the Asymme-
try, the BorderIrregularity,the Colorvariationandthe Di-
ameter. This latter must be greater than 6mm in the case
of melanoma. This paper is organized as follows. In Sec-
tion 2 we present a brief overviewof previous related work.
In Section 3, we illustrate our approach to melanoma skin
cancer detection. Experimental results and analysis are
presented and discussed in Section 4. Finally, we present
10

CNSER Int. J. Computer Vision Signal Process.
a brief conclusion.
2. Melanoma skin cancer detection: State
of the Art
Melanoma is considered as the most rapidly prolifer-
ating cancers in the world [4]. For that, there has been
a significant increase in concern in the development of
automatic digital dermatoscopic image analysis methods.
Among these methods, the color enhancement which is
based on the color correction with HSV model. In this ap-
proach, linear regression models are constructed for each
channel, whichallowsautomaticadjustmentofthe hueand
saturation [5]. A second approach is based on a scheme
of standardization based on two steps; removing the color
variationsandstrengtheningcontrastimages[6]. Although
this approach has in some cases good results, it remains
limited since lesions classification depends on other param-
eters like asymmetry and border irregularity.
Otherapproachesdevelopedin theliteraturesarebased
on feature extraction. The first describes a new method
based on graphs which will enable networks to extract pig-
ments from images of dermoscopy [7].The second imple-
ments the classification of skin lesions based on the char-
acteristics of ”granularity” [8]. The third proposes a clas-
sification with white areas on the images [9]. The last
method extracts the characteristics of curvature from 3D
images acquired using a photometric stereo system [9].
Another method is widelyadopted byresearchers. This
method consists, generally, of four stages which are: ac-
quisition, segmentation , feature extraction and classifica-
tion.The segmentation is the most difficult step since it af-
fects the precisionofthe subsequentstagesandbecause[4]:
(i) The transition between the lesion and the healthy skin
is usually of low contrast; (ii) the borders of the lesion are
usually fuzzy and irregular; (iii) the presence of complicat-
ing artifacts. To resolve these problems, several segmen-
tation algorithms have been developed. We can classify
these into three categories which are edge contour-based,
thresholding and region-based.
An effective thresholdingmethod proposedby Granaet
al. [10] is based on Otstu thresolding algorithm. It aims at
segmenting automatically the melanoma image. Then, the
developed method selects k points for spline-based interpo-
lation,which gives a smooth edge of lesion. Such methods
achieve good results when there is good contrast between
lesion and skin, but face problems when there is an overlap
of the modes of the two regions.
Edge/ contour-based approaches were used in [11][12].
Rubegni et al. [11] segmented dermoscopy images using
the zero-crossings of a LoG edge operator, while Zhou et
al. [12] used an improved snake model to detect lesion bor-
ders. Edge and contour-based approaches perform poorly
when the boundaries are not well dened, for instance when
the transition between skin and lesion is smooth. In such
Preprocessing
Post processing
Compute Total
Classification ABCD
Feature extractionDermatoscopic
ValueSegmentation
Process
Figure 1: Architecture System of Melanoma Skin Cancer Diagn osis
situations, the edges have gaps and the contour may leak
through them.
Region-based approaches have also been used. Some
examples include multi-scale region growing [13], fuzzy c-
means based on anisotropic mean shift [14], multi- resolu-
tion markov random elds [15]. The researchers presented
by Fatichah et al. [16][17] segmented the skin lesion us-
ing fuzzy region growing and the diagnoses are based on
ABCD feature which is used to diagnose melanoma skin
cancer based on Total Dermatoscopic Value (TDV).Based
ontheexperimentusing30samplesimages,theaccuracyof
the system proposed by Fatichah et al. [16] is 85% but the
accuracy using fuzzy region growing proposed by Fatichah
et al. [17] is 86.6%.
Despite region-based approaches have difculties, espe-
cially, when the lesion or the skin region are textured, this
category is widely used by researchers. For that, in this
paper we try to improve region growing algorithm for the
detection of lesion.
3. Architecture System of Melanoma Skin
Cancer Detection
The processing algorithm presented in Fig. 1 describes
the melanoma skin cancer detection. Dermatoscopic im-
agesused in this workarecollectedwith the Departmentof
Dermatology, University Hospital Hedi Chaker Sfax. They
are coded on 24 bits storing the three color components
r (red), g (green) and b (blue). These rgb images are
640×480 pixels in size. Unfortunately, the image acquisi-
tion leads to the permanent presence of illumination. The
existence of shadows, reflections on wet tissue or fatty re-
port more difficulties. In addition, artifacts such as hair
and diversity of skin color complicate the acquisition step
and make it difficult to reproduce. These constraints illu-
mination and noise justify the importance of the prepro-
cessing algorithm used thereafter.
3.1. Preprocessing
As mentioned, the image preprocessing is an important
step in image diagnosis. It is used to correct defects illu-
mination, eliminating noise and small spots and enhance
11

CNSER IJCVSP, 3(1),(2013)
a) b)
Figure 2: Preprocessing: (a) Original image (b) image given after
applying the median filter and the morphological closing
a)b)
Figure 3: Influence of the histogram adjustment on grayscale im-
age:(a) Filtered image converted to grayscale (b) Influence of the
histogram adjustment
the contours and contrast as much as possible without de-
grading the lesion. Preprocessing proposed in this work
is composed of three steps. It starts by using a median
filter aiming at cleaning the image by eliminating certain
defects. This spatial filter is based on calculating for each
pixel the median of the gray levels of the neighbors pix-
els. Since each component of the image can be viewed as
a grayscale image, we can apply this filter on each com-
ponent separately. Then, we performed a morphological
closing aiming at eliminating all artifacts such as hair. Ac-
cording to the images provided, it was found empirically
that setting a 5×5 median filter and a disk with a diameter
of 3 givesthe best results. Fig. 2 shows the influence of the
median filter and the morphological closing on the image.
Finally, we must improve the color between the le-
sion and healthy skin. Since our goal is to extract the
lesion apart, we improve the contrast of the image by
adjusting intensity values after converting the image to
grayscale. The influence of the Histogram adjustment on
the grayscale image is shown in Fig. 3.
3.2. Segmentation
To do a classificationof skin lesions, and particularlyto
distinguish melanoma from benign lesions, we must begin
by isolating the lesion from healthy skin that surrounds
each color image using a segmentation methodology. The
detection of this skin lesion is a critical problem in der-
matoscopic images because the transition between the le-
sion and the surrounding skin is difficult to detect accu-
rately. For this, segmentation method chosen must be pre-
list of the pixels P having the minimum value of the gray level
For each pixel P of the list,
calculate the average of the window 5 * 5 of its neighbors
—P-Avr—
Selecting the pixel having the smallest difference
with its neighbors as germP-Avr
Figure 4: Automatic method for the choice of the germ
cise. In this paper, the segmentation process used is region
growing. This segmentation technique begins with a pixel
as the seed point. Then it aggregates the pixels according
to two criteria: homogeneity and adjacency [18]. Several
studies have used the segmentation by region growing to
detect skin lesions. Fondn et al. [19] have described a
method based on selecting a region with the mouse. The
selectionofthis regioncanthen makethe choiceofthe seed
point and the stop criterion. Fatichah et al. [16] described
a method using the center of a homogenous area as the
germ and fixing the threshold randomly. These two man-
ual methods, based on a random selection of the threshold
and the germ, are not recommended. In fact, it will have
a bad Influence on the segmentation result.
In this work, we tried to adapt a simple method of
region growing for our application to detect skin lesions.
To optimize the result of the segmentation, the proposed
method is based on an automatic choice of the threshold
and the seed point from which and at each level of growth
we add the neighboring pixels having similar properties.
Since only one region must be segmented and the lesion
is always darker than the normal skin, the seed pixel will
be automatically selected as the pixel with the gray level
which is darker and more similar to its neighbors. This
characteristic is achieved by calculating the difference be-
tween the pixel P and the average of a 5×5 window of its
neighbors. This method is illustrated in the diagram pre-
sented in Fig. 4.
The result is shown in Fig. 5 where in Fig. 5(a) pixels
having the darkest gray level are selected. Then, the seed
pixel is detected in Fig. 5(b).
After selecting the seed pixel, the application of this
algorithm includes a specification of two variables that are
the size of the window and the threshold. The size of the
12

CNSER Int. J. Computer Vision Signal Process.
a) b)
Figure 5: The seed pixel selection
Pixel IntensityNumber of pixels
Figure 6: Triangle method
window is set to 3×3 pixels since it gives best result. The
threshold is calculated using the following steps:
*Finding the optimal threshold S of the image auto-
matically using the triangle method [20] which allows con-
structing a line between the histogram peak and the far-
thest end. Thethresholdisthemaximumdistancebetween
the line and the histogram as shown in Fig. 6. The choice
ofthismethodisbasedonthefactthatalltheimagesshow-
ing a lesionarestill presentedasa darkspot surroundedby
a lighter area (skin) as shown in Fig. 7(a). The histogram
of such an image is always given as presented in Fig. 7(b)
.
*Calculating the average of the region where N is the
LesionSkin areaSkin intensity levels
Lesion intensity levels
Figure 7: Choice of the triangle method:a) An image presenti ng a
lesion,b) Corresponding histogramnumber of pixels belonging to the region:
Avr=/summationtextvalues of the pixels belonging to the regions
N
(1)
*The threshold is calculated by measuring the difference
between the average and the optimal threshold
Threshold =|Avr−S| (2)
Then, the growth process used the following steps:
*Step 1: Adding the seed point neighboring pixels belong-
ing to a window of 3×3 pixels to a list of neighbors.
*Step 2: Calculating the difference between the graylevel
of each neighboring pixel and the region average.
*Step 3: Adding the neighboring pixel of the seed point,
with the smallest difference, in the region and mark it as
a seed pixel.
*Step 4: Updating the average Avr, the threshold and the
gray level difference.
*Step 5: Deleting the pixel from the list of neighbors
*Step 6: Returning tostep 1 eachtime the difference found
in step 4 is less than the threshold.
Hence, the approachisto growthe regionaroundthe start-
ing pixel. This aggregation does not stop until the differ-
ence exceeds a certain threshold:
|I(x,y)−Avr|>=threshold (3)
3.3. Post processing
To remove isolated pixels within the region of the le-
sion by the segmentation, a further step of morphological
closing is produced. In our work, the disc diameter of this
morphological closing is set to 5 as it gives the best result.
Finally, the resulting image is a binary mask used to sepa-
rate the lesion from the skin. It is then superimposed with
the original image. Fig. 8 illustrates this step.
3.4. ABCD feature extraction
The ABCD rule was introduced by Stolz et al. [21]
and used by dermatologists in recognition process of skin
lesions to assess the risk of malignity of a pigmented le-
sion. This method is able to provide a more objective
and reproducible diagnostic of skin cancers in addition to
its speed of calculation.It is based on four parameters: A
(Asymmetry) concerns the result of evaluation of lesions
asymmetry, B (Border) estimates the character of lesions
border, C (Color) identifies the number of colors present
in the investigated lesion, and D (Diameter).
3.4.1. Asymmetry
The asymmetry of the lesion is the most important in-
dicator characterizing the malignancy. This characteristic
is considered in terms of form and color. To facilitate the
task of calculating the score of asymmetry, we translated
the lesion to the center of the image. Then we rotate the
lesion to align the major axis horizontally as presented in
Fig. 9.
13

CNSER IJCVSP, 3(1),(2013)
a) b)
c) d)
Figure 8: The resultof the segmentation: (a) originalimage (b) lesion
detected by region growing (c) lesion detected after post-p rocessing
(d) segmented image
a) b)
Figure 9: The division of the lesion: (a) according to the maj or axis
(b) According to the minor axis
Now it is easy to divide the image according to the
major axis and the minor axis of the lesion (Fig. 9). To
measuretheasymmetryin termsofform, wearebasingour
calculation on the asymmetry index presented as follows
[22].
AI=2/summationdisplay
k=1∆Ak
AL(4)
wherekis the major axis and the minor axis ∆ Akis
the area of non-overlapping zone and ALis the total area
of the lesion.
To measure the asymmetry in terms of color, our cal-
culations are based on the histograms of the three RGB
components of each part of the lesion and the chi-square
distance.
D(h1,h2) =N/summationdisplay
i=1(h1(i)−h2(i))2
h1(i)+h2(i)(5)
whereh1 andh2 are the two histograms of size N.
Figure 10: Calculating the border score: B=3
Finally, the asymmetry score Ais calculated as the av-
erage between the asymmetry score in terms of form and
that in terms of color.
3.4.2. Border Irregularity
Generally, the borders of the benign lesions are clearly
defined. The irregularityof the bordercan usually report a
cancer during growth and propagation. For an evaluation,
the lesions are divided into eight sectors as presented in
Fig. 10. Within each sector,a strong and sharp cut pat-
tern at the periphery receives a score of 1. In contrast,
a gradual approach, indistinct within each segment has a
score of 0. Thus, the maximum irregular border score is
8, and the minimum one is 0. This calculation is based on
the Euclidean distance and the standard derivation in each
sector.
Di=/radicalbig
(x2−x1)2+(y2−y1)2 (6)
wherex2 andy2 are the coordinates of the center of the
lesion.x1 andy1 are the coordinates of pixel i.
Distance =N/summationdisplay
i=1Di (7)
WithNis the number of pixels in the edge belonging to
the considered area. Diis the Euclidean distance between
the center of the lesion and the pixel i.
Then the standard deviation is calculated for each sec-
tor with the following equation:
s= (1
NN/summationdisplay
i=1(xi−x)2)1/2(8)
wherenis the number of elements in the sector.
Within each sector, if the deviation exceeds a certain
threshold, the score edge is 1. This threshold is set empir-
ically to 30.
3.4.3. Colors
Melanomas are characterized by the presence of six dif-
ferent colors, namely, white, red, light brown, dark brown,
blue-gray and black. For each present color, we add 1 to
the score. When the six colors are present, the maximum
score is 6. The minimum score is 1.To verify the existence
14

CNSER Int. J. Computer Vision Signal Process.
Figure 11: Color score calculation for a lesion:C=2(light b rown and
dark brown)
Figure 12: The scales used to measure the actual diameter
of each color in the lesion,the image was converted from
the RGB color space to the CIE Labcolor space because
the distance between two colors in the RGB color space
does not reflect the ”difference” perceived by the naked
eye contrary to the CIE Labcolor space. This difference
can be calculated using the Euclidean distance delta E.
∆E=/radicalbig
(L1−L2)2+(a1−a2)2+(b1−b2)2(9)
whereL1,a1 andb1 are the components of the CIE Lab
colorimetric space of the desired color and L2,a2 andb2
those of each pixel of the image.
The smallest perceptible difference is known ’Just Notice-
able Difference’ (JND) [23]. It is fixed to 2.3. In practice,
to consider two colors are equivalent, tolerance can be set
at 3 * JND. The Fig. 11 shows the result for a lesion.
3.4.4. Diameter
Since the dermatoscope gives the possibility of tracing
the marks on the image during acquisition (Fig. 12), we
are based on these scales to calculate the actual diameter
of the lesion.
We note that the lesion diameter greater than 6 mil-
limeters is more likely to be malignant than a small lesion.
3.5. Classification
After obtaining the values of the four characteristics,
the Total Dermatoscopic value (TDV) is calculated basedon the following formula where each one of the presented
characteristics is multiplied by a weighting factor:
TDS= 1.3∗A+0.1∗B+0.5∗C+0.5∗D(10)
This score contributes to detect benign or malignant pig-
mented skin lesion as presented as follows:
∗TDS <4.75: Benign skin lesion
∗4.75≤TDS≤5.45: Suspicious
In this case, the intervention of the doctor is required by
removing the suspicious lesion (biopsy) followed by its mi-
croscopic examination (histology). In general, dermatolo-
gists prefer to immediately remove the entire lesion, but in
some cases, a partial biopsy may be performed.
∗TDS >5.45: Melanoma
4. Experimental results and analysis
In the experiment, the dermatoscopic image suspected
as melanoma is used as input data. To evaluate the perfor-
manceofourdevelopedsystem, thisresearchuse datasetof
40 dermatocospic images. The measures used for the eval-
uation are accuracy, sensibility and specificity measures
presented as follows,
Accuracy =(Number of detected lesions )
(Number of images in the dataset )∗100
(11)
Sensivity =TP
TP+FN(12)
Specificity =TN
TN+FP(13)
whereTPis the number of true positives, FNthe number
offalsenegatives, TNthenumberoftruenegative,and FP
the number of false positives. A positive result is a lesion
which is classified as malignant and a negative result is a
lesionthatisclassifiedasbenign. Thesensitivityandspeci-
ficity measurethe percentagequoted classificationsspecific
for malignant and benign cases, respectively. our work is
based on 40 images (27 malignant and 13 benign lesions).
The experimental results of total dermatocospic images is
described in the following Table.
Table 1: Accuracy result
Measure Accuracy Sensitivity specificity
Experimental Result 92.5% 88.88% 92.3%
To facilitate the use of our application, we have im-
plemented a graphical user interface presented in Fig. 13:
Figure 13 shows the main interface of our application
which allows the user to select an image using the ’open’
15

CNSER IJCVSP, 3(1),(2013)
Figure 13: The Main user interface of the system
Figure 14: The interface containing the listofpatients int he database
command in the ’File’ menu. Then he can use the ’Seg-
ment’ command to detect the lesion from the healthy skin
and display the results for the analysis and diagnosis. The
user can either use the ’Save’ command from the ’Diagnos-
tics’ menu to save the image in the database, or the ’View’
command to view the results already achieved. The ’View’
command opens the interface shown in Fig. 14.
The ’Search’button allowsthe userto searchthe record
with the ID typed in the text box to display all the details
such as image, name of the patient, the characteristics of
thelesion(Asymmetry, Border,colors,DiameterandTDS)
and the diagnostic in the interface of Fig. 15.
To evaluate our developed interface, we have tested the
40 images collected with the Department of Dermatology
of the University Hospital Hedi Chaker Sfax, Tunisia.
Fig. 16 shows the results found by our interface com-
pared against the diagnosis of dermatologist for two im-
ages.
As was mentioned, the accuracy of the system is about
92%. Based on the experimental results, the small value
Figure 15: The TDV result user of the interface
Graphical interface Dermatologist di-
agnisis
Benign
Melanoma
Figure 16: Graphical interface results
of color variation causes, in some cases false diagnosis. To
improve this accuracy, we must tuning the calculation of
this parameter. We can use for example a color clustering
model for dermoscopic images.
5. Conclusions
The objective of our work has involved the detection
of melanoma based on region growing segmentation and
the ABCD rule used for the detection of malignancy of
pigmented skin lesion. The use of region growing in our
method based on automatic selection of the seed pixel and
the threshold ensures the best results and avoids overlap
between the lesion and healthy skin. The result of the seg-
mentation was then used in the next step by the ABCD
rule in order to classify pigmented skin lesion as benign or
malignant. There are three diagnosis that are used in this
research i.e. melanoma, suspicious, and benign skin lesion.
The experiment uses 40 samples of dermatoscopic images
containing a lesion that is suspicious melanoma skin can-
cer. Based on the experiment, the accuracy of the system
is 92% with 4 false diagnosis of the 40 samples. For illus-
tration, we have developed a graphical interface in order
to facilitate the diagnostic task for the dermatologists.
16

CNSER Int. J. Computer Vision Signal Process.
References
[1] S. Stechschulte, C. Ricotti, C. J. Cockerell, Advances i n diagnos-
tic testing for skin cancer, TOUCH BRIEFINGS (2008) 73–76.
[2] R.Lucas, T.McMichael, W.Smith, B.Armstrong, Solar ult ravio-
let radiation.global burden of disease from solar ultravio let ra-
diation, Environmental Burden of Disease. Series 13 (2006) 117.
[3] D.Rigel, J.Russak, R.Friedman, The evolution of melano ma di-
agnosis, A Cancer Journal for Clinicians 60 (2010) 301316.
[4] F. Xie, A. Bovik, Automatic segmentation of dermoscopy i m-
ages using self-generating neural networks seeded by genet ic al-
gorithm, Pattern Recognition, ELSEVIER (2013) 1012–1013.
[5] H. Iyatomi, M. E. Celebi, G. Schaefer, M. Tanaka, Automat ed
color calibration method for dermoscopy images, Computeri zed
Medical Imaging and Graphics (2011) 89–98.
[6] G. Schaefer, M. I. Rajab, M. E. Celebi, H. shi Iyatomi, Col our
and contrast enhancement for improved skin lesion segmenta –
tion, Computerized Medical Imaging and Graphics (2011) 99–
104.
[7] M. Sadeghi, M. Razmara, T. K. Lee, M. Atkins, A novel metho d
for detection of pigment network in dermoscopic images usin g
graphs, Computerized Medical Imaging and Graphics (2011)
137–143.
[8] W. V. Stoecker, M. Wronkiewiecz, R. Chowdhury, R. J. Stan ley,
J. Xu, A. Bangert, B. Shrestha, D. A. Calcara, H. S. Rabinovit z,
M. Oliviero, F. Ahmed, L. A. Perry, R. Drugge., Detection of
granularity in dermoscopy images of malignant melanoma usi ng
color and texture features, Computerized Medical Imaging a nd
Graphics (2011) 144–147.
[9] A. Dalal, R. H. Moss, R. J. Stanley, W. V. Stoecker, K. Gupt a,
D. A. Calcara, J. Xu, B. Shrestha, R. Drugge, J. M. Mal-
ters, L. A. Perry, Concentric decile segmentation of white a nd
hypopig- mented areas in dermoscopy images of skin lesions
allows discrimination of malignant melanoma, Computerize d
Medical Imaging and Graphics (2011) 148–154.
[10] C. Grana, G. Pellacani, R. Cucchiara, S. Seidenari, A ne w al-
gorithm for border description of polarized light surface m icro-
scopic images of pigmented skin lesions, IEEE Transactions on
Medical Imaging 22 (2003) 959964.
[11] P. Rubegni, A. Ferrari, G. Cevenini, D. Piccolo, M. Burr on, Dif-
ferentiation between pigmented spitz naevus and melanoma b y
digital dermoscopy and stepwise logistic discriminant ana lysis,
Melanoma Research 11 (2001) 3744.
[12] H. Zhou, G. Schaefer, M. C. nad F. Lin, T. Liu, Gradient ve ctor
?ow with mean shift for skin lesion segmentation, Computeri zed
Medical Imaging and Graphics 35 (2011) 121127.
[13] W. S. M.E. Celebi, A. Aslandogan, Unsupervised border d etec-
tion in dermoscopy images, Skin Research and Technology 13
(2007) 454462.
[14] H. Zhou, G. Schaefer, A. Sadka, M. Celebi, Anisotropic m ean
shift based fuzzy c-means segmentation of dermoscopy image s,
IEEE Journal of Selected Topics in Signal Processing 3 (2009 )
2634.
[15] J. Gao, J. Zhang, M. Fleming, A novel multiresolution co lor
image segmen- tation technique and its application to derma to-
scopic image segmentation, Proceedings of the IEEE Interna –
tional Conference on Image Processing Vancouver, BC, Canad a
(2000) 2634.
[16] B.Amalian, C.Fatichah, M.R.Widyanto, Abcd feature ex trac-
tion for melanoma skin cancer diagnosis, Proceedings of the 9th
International Conference on Advanced Computer Science and
Information System (2009) 224228.
[17] M.R.Widyanto, C.Fatichah, B.Amaliah, Skin lesion
detection using fuzzy region growing and abcd fea-
ture extraction for melanoma skin cancer diagnosis,
http://www.its.ac.id/personal/files/pub/ (2013) 4554–
chastine–Full.
[18] H.Laguel, Deploiement sur une plateforme de visualisa tion, d’un
algorithme cooperatif pour la segmentation d’images irm ba se
sur les systemes multiagents, University of Science and Tec hnol-
ogy Houari Boumediene,Algeria (2010) 1–104.[19] I.Fondn, C.Serrano, B.Acha, Segmentation ofskin canc er images
based on multistep region growing, MVA2007 IAPR Conference
on Machine Vision Applications (2007) 339–342.
[20] S.Latt, G.Zack, W.Rogers, Automatic measurement of si ster
chromatid exchange frequency, Thejournal of histochemist ry
and cytochemistry Volume 25 (1977) 741–753.
[21] W.Stolz, Abcd rule of dermatoscopy: a newpractical met hod for
early recognition of malignant melanoma, European Journal of
Dermatology, Volume 4 (1994) 521–527.
[22] M.R., Widyanto, C. Fatichah, B. Amaliah, Abcd feature e xtrac-
tion for melanoma skin cancer diagnosis, Proceedings of the 9th
International Conference on Advanced Computer Science and
Information System (2009) 224–228.
[23] O.Kakde, K.Bhoyar, Color image segmentation based on j nd
color histogram, International Journal of Image Processin g
(IJIP), Volume 3 (2010) 283–292.
17