Imperial Journal of Interdisciplinary Research (IJIR)
Vol-3, Issue-1, 2017
ISSN: 2454-1362, http://www.onlinejournal.in
Automatic Segmentation System for
Melanoma Region in Dermoscopic Images
Karanvir Kaur1 & Er. Vikas Goyal2
1
Pursuing M.Tech , Electronics and communication engg.
ASRA College of Engg. And Tech. Bhawanigarh, Punjab
2
Associate Professor , Electronics and communication engg.
ASRA College of Engg. And Tech. Bhawanigarh, Punjab
Abstract— Malignant melanoma is the third most
frequent type of skin cancer and one of the most
malignant tumors. A malignant melanoma diagnosed
at an early stage can be cured without complications.
In this work, we address the problem of how to
separate the absence and presence of pigment
networks in a given dermoscopic image in two classes
called lesion region and normal skin. Our proposed
method is a robust, reliable, computer-aided
diagnostic tool for accurate detection of this
pigmented region. We have used xyz color space in
which lesion region becomes more dark then normal
skin which results in better segmentation than the
RGB colored input images. We have used ANN
classifier in final step which classifies the whole
image into two regions based on some thresholding
on the results output by ANN algorithm.
Keywords—Skin Cancer, xyz color space, Gaussian
Mixture Model, Artificial Neural Network.
I. Introduction
Skin cancer is a deadly disease and its accurate
detection plays a vital role for the diagnosis. Among
all skin cancers the most aggressive form of skin
cancer is melanoma [1]. Melanoma is found mostly in
white-skinned persons from age 15-34[2]. Ultraviolet
rays mainly UVA and UVB rays are the utmost
reasons for the growth of melanoma [2]. The first
instance of the melanoma was noticed in 1930 in the
United States [2]. After that the graph of melanoma
diagnosis increased day-by –day. No doubt the
growth of melanoma is very slow and if detected
early the survival rate of the person increases to 95%
for five years. Many publications reported the
automated diagnosis of the skin cancer by the image
processing due to the reasons that dermatological
results are not as accurate as well as requires lot of
training to dermatologists. The dermatoscopic criteria
uses the formula of ABCD(E) to differentiate
between a benign and melanoma where A stands for
asymmetry i.e. the two halves of the area differ in
shape, B stands for border irregularity i.e. the edges
of the area may be irregular or blurred and sometimes
show notches, C stands for color variation which
Imperial Journal of Interdisciplinary Research (IJIR)
indicates that color distribution in the lesion may be
uneven, D stands for diameter and the most of the
melanomas are at least 6mm in diameter and E stands
for evolution [3].
Skin has the giant biological structure. The chores of
skin are as follows:
 Protection from microbes
 Balancing of body temperature
 Permits the sensations of touch, heat and
cold
 Reduces harmful effects of UV radiation
 Produces vitamin D
Coatings of skin
 The Epidermis
 The Dermis
 Hypodermis
Epidermis
The outer coating of skin is known as epidermis. It is
soft and it repeatedly regenerates. The thickness of
epidermis varies on distinct parts of the body. Mainly
the thickness ranges between 0.5 and 4 mm and made
up of different types of cells known as keratinocytes,
corneocytes and melanocytes. The main functions of
the epidermis layer are that it acts as a physical
barrier between the body and the environment
moreover it also prevents the excessive water loss.
Dermis
The inner or lower layer of the two main layers of
cells which make up the skin is known as dermis. The
main functions of the dermis layer are producing
sweat and temperature balancing of the body.
Computer aided image analysis in melanoma
research
Page 1778
Imperial Journal of Interdisciplinary Research (IJIR)
Vol-3, Issue-1, 2017
ISSN: 2454-1362, http://www.onlinejournal.in

Image of suspicious lesion
Background compensation
Segmentation
Feature extraction
Classification
Fig.1 Automated melanoma diagnosis
Fig. 1 is representing the process of automated
diagnosis of melanoma, first of all input of the
suspicious image is given. Then shadows are
removed from the background by applying various
techniques in the process of background
compensation. After that division of regions in the
image is done in the segmentation. A brief
introduction of the segmentation is as follows.
Segmentation
The division of an image into meaningful
structures, image segmentation, is often an essential
step in image analysis, object representation,
visualization and many other image processing tasks.
The use of image segmentation is in locating
objects and boundaries. To change the representation
of an image into meaningful and easy to analyze is
the main aim of segmentation [3]. Segmentation is a
vital process in which partitioning and grouping is
done. Partitioning is done according to relevant
internal properties into regions/sequences. In
grouping, identification of sets of relevant tokens in
image is done.
Threshold
based
segmentation:
Threshold
segmentation is also known as intensity based
segmentation. The common techniques of threshold
based segmentation are histogram and the slicing
techniques. These techniques can be applied directly
but sometimes the preprocessing and the post
processing techniques are also used.
Advantages of thresholding: It is easy to implement.

The process of thresholding is also fast.
Imperial Journal of Interdisciplinary Research (IJIR)
It performs best on several types of images
like documents and controlled lighting.
Disadvantages of thresholding: The only disadvantage of the thresholding is
that there is no assurance of the coherence of
the object. The results may contain holes or
the extraneous pixels.
Solution:
 To overcome this problem, morphological
operators can be used for the postprocessing.
Edge based segmentation: In this technique,
detection of the edges in the images are used for the
representation of the boundaries of the object and are
also used for the identification of these objects.
Common problems in Edge Based Detection:False Detections: Sometimes there is presence of
edge in the image even when there was no border.
Missed Detections: Many times there was no
detection of edge even when there a real border
exists.
Region based segmentation: Region based
segmentation is the opposite of the edge based
segmentation. This segmentation starts in the middle
of the object and then grows outwards till it meets the
boundaries of the object.
Advantages of region based segmentation:
These techniques perform very better in the noisy
images where there is difficulty in the detection of the
borders.
Disadvantages of region based segmentation:
Its biggest disadvantage is that either its output is
over segmented or under segmented.
II. LITERATURE REVIEW
In year 1999, screening algorithm of melanoma had
been described [4]. It shows the early signs of the
melanoma and the people having higher risk of the
melanoma were indicated.
Type of skin cancer were described by
[5].
They also explained the treatment options available at
that time. The treatments available at that time like
surgery, mono chemotherapy, combination therapy
and immunotherapy were described.
In 2010, the use of dermatoscopy by the US
dermatologists was described [6]. It gives the brief
view for the use of dermatoscopy and the limitations
of dermatoscopy due to the reasons of lack of training
and interest..
In 2011, a new method was discovered for the
classification of the pigmented skin lesions by using
the standard cameras [7]. The images were taken by
standard cameras after that each image went through
the series of processes such as: (1) Preprocessing (2)
Segmentation (3) Feature extraction (4) Lesion
Classification. In the preprocessing, shadows were
removed then segmentation was performed and in
Page 1779
Imperial Journal of Interdisciplinary Research (IJIR)
Vol-3, Issue-1, 2017
ISSN: 2454-1362, http://www.onlinejournal.in
feature extraction, the quantitative representation for
the lesion area was generated and at classification,
distinction between benign or melanoma was done.
In July 2013 a method was introduced to increase
the efficiency and to correct the illumination variation
named MSIM i.e. Multistage Illumination modeling
[8]. In the proposed algorithm, Monte Carlo sampling
was done to determine the non-parametric model of
the illumination and after that the reflectance
component of the photograph was calculated. The
specificity, sensitivity and accuracy of the output
images were quite high.
In 2015 [11] proposes an image classification
system for premature Skin cancer detection embraces
different stages: pre- and Post-processing,
segmentation, feature extraction and classifier. The
pre-processing resizes the image that improves the
Speed performance and removes the extra feature
such as the noise and fine hair. Post-processing
enhances the image eminence and sharpens the
framework of the cancer cell
In this paper, automatic melanoma recognition
was done using multi-stage neural networks [9] [11].
Gaussian mixture model was used for image
smoothening and neural networks were used for
automatic classification.
III. PROPOSED ALGORITHM
The main objective of the proposed algorithm is to
obtain automatic image annotation using Artificial
Neural Network and to undertake large set of skin
cancer and achieve highest possible accuracy. The
steps which are followed to achieve the mentioned
objectives are:
 Inputting the RGB image

Then conversion of RGB image is done into
xyz color space.

After that texture matrices are obtained and
texture representatives are obtained.

Then GMM modeling
matching of pixels is done.

Finally classification is done using ANN.

At last, accuracy assessment of the
algorithm is done by checking the sensitivity
and specificity.
and
posterior
A. Block diagram of the proposed algorithm
Fig. 2 represents the block diagram and the block
diagram explains the operation of proposed
algorithm. First of all RGB image is converted into
XYZ color space and then XYZ color space is
converted into the Lab color space because Lab color
space is easy to visualize by the human perception.
After that texture matrices are obtained and then
texture representation is done by using the Fuzzy C
Imperial Journal of Interdisciplinary Research (IJIR)
Means algorithm. Then the Gaussian mixture models
are used and the posterior matching of the pixels is
done. Then the artificial neural networks are used for
the final classification. Before using artificial neural
networks first of all they are trained to get the desired
results. At last, sensitivity and the specificity are
checked.
CONVERSION OF RGB IMAGE INTO XYZ
COLOR SPACE
OBTAINING THE TEXTURE MATRICES
GENERATION OF TEXTURE
REPRESENTATIVES
GMM MODELLING AND POSTERIOR
MATCHING OF PIXELS
FINAL CLASSIFICATION USING ANN
ACCURACY ASSESSMENT OF THE
ALGORITHM
Fig. 2 Block diagram of proposed algorithm
IV. RESULTS
Inputting the image in MATLAB
In this step input of the melanoma image is provided
and its ground truth is obtained.
Page 1780
Imperial Journal of Interdisciplinary Research (IJIR)
Vol-3, Issue-1, 2017
ISSN: 2454-1362, http://www.onlinejournal.in
(a)
(B)
Fig.3 In the image (a) ground truth image of original image
(b) there is original image
A. Conversion of RGB image into xyz color space
CIE considered the tristimulus values for red, green,
and blue to be undesirable for creating a standardized
color model. Instead, they used a mathematical
formula to convert the RGB data to a system that uses
only positive integers as values [10]. The
reformulated tristimulus values were indicated as
XYZ. These values do not directly correspond to red,
green, and blue, but are approximately so. The curve
for the Y tristimulus value is equal to the curve that
indicates the human eye's response to the total power
of a light source. For this reason the value Y is called
the luminance factor and the XYZ values have been
normalized so that Y always has a value of 100.
Obtaining the XYZ tristimulus values is only part of
defining the color. The color itself is more readily
understood in terms of hue and chroma [10]. To make
this possible, CIE used the XYZ tristimulus values to
formulate a new set of chromaticity coordinates that
are denoted xyz.
Fig. 4 Original image divided into three channels x,y,z
B. Obtaining the texture matrices
The texture vector contains pixels in a neighborhood
of size n centered on the pixel of interest. Let s be a
pixel location (x, y) in the photograph. Then, the
vector ts represent the n × n × a texture patch
centered at pixels.
The process of extracting a texture vector for the
images by having multiple channels a separate vector
is obtained for each channel and concatenated
sequentially. In the proposed algorithm a matrix of 25
pixels is made and centre point is taken as 3 and twotwo neighbouring pixels are taken to make a matrix.
A texture vector is obtained from the matrix.
Imperial Journal of Interdisciplinary Research (IJIR)
C. Generation of texture representatives
In this step, fuzzy clustering has been used on a
single channel of the image, which converts the
whole image into fixed number of clusters, one can
vary the number of clusters to get better results
according to the content present in the images. First
of all, all clusters are taken separately in which a
pixel has been chosen as texture representative of that
cluster. One can choose any pixel from that cluster,
but getting pixel automatically we use center of mass
of that cluster as texture representative. The
importance of this selection is that the pixel that will
be chosen will always come from the center of cluster
and not the boundary region of the cluster. GMM
modeling and posterior matching of pixels.Mixture
models come under the classification of the density
models. These models are encompassed of
component functions mainly of Gaussian functions.
Multimodal densities are formed by the combination
of these functions. They are employed to model the
texture vectors of chosen texture representatives in
order to perform tasks for final improved
segmentation. After generating these models, every
pixel’s texture vector has been compared with all the
GMM’s using posterior maximum log likelihood,
which puts every pixel to the closest texture Gaussian
model.
After Gaussian modeling whole image is
divided into two parts, one is pure lesion region and
the other is skin region. This has been chosen
according to maximum negative log likelihood values
as they contain the darkest skin lesion regions. Our
results are also obtained from the Gaussian mixture
model but they are not so accurate .
After this RGB values from two corresponding
classes are concatenated separately and make ready to
feed in neural network for training process.
D. Final classification using ANN
After the extraction of the feature vector which is
comprised of two element classes namely lesion and
the normal skin, proposal of the neural object
recognizer was done which is used to learn the
relationship between the 3-layered artificial neural
network system. The patterns are learned by Error
Back Propagation Algorithm.
Neural network consist of 3 layers namely input
layer, hidden layer and output layer. There are 3 input
units, 1 hidden unit and the 2 output units.
After training, the neural network can be tested for
the whole image results in final classification.
Page 1781
Imperial Journal of Interdisciplinary Research (IJIR)
Vol-3, Issue-1, 2017
ISSN: 2454-1362, http://www.onlinejournal.in
Fig. 6: Image showing sensitivity and specificity
graphs for input images
V. CONCLUSION
In this paper, an efficient technique is applied for the
detection of skin cancer. The facts of the applied
scheme:
The results of the applied technique are very accurate.
About 95 % sensitivity is achieved.
Proposed algorithm can be applied to large data sets.
Fig. 5 (a) Input image (b) Results after applying ANN.
E. Accuracy assessment of the algorithm
The objective of this algorithm is to measure
sensitivity and specificity.
The formulas of sensitivity and specificity are given
below, where TP is the number of true positive
pixels, FP is the number of false positive pixels, TF is
the number of true negative pixels and FN is the
number of false negative pixels.
Sensitivity = (TP/(TP+FN)).
Specificity= (TN/(TN+FP))
F. GRAPH OF ALL IMAGES
Table 1: showing sensitivity
evaluating accuracy
Image
TP
FN
Number
Image1
6930
Image2
4151
Image3
8069
Image4
4487
Image5
5519
Table 2: showing
evaluating accuracy
Image
TN
Number
Image1
8789
Image2
11690
Image3
7429
Image4
11564
Image5
10263
619
15
386
266
476
specificity
parameters
for
Sensitivity
0.9180
0.9963
0.9543
0.9440
0.9206
parameters
FP
Specificity
46
528
500
67
126
0.9947
0.9567
0.9369
0.9942
0.9878
for
Imperial Journal of Interdisciplinary Research (IJIR)
References:
[1] Hoffmann K, Gambichler T, Rick A, Kreutz M,
Anschuetz M, Grunendick T, “Diagnostic And Neural
Analysis Of Skin Cancer (DANAOS)” British Journal
of Dermatology 2003
[2] Jeffrey L. Melton, Jason R. Swanson, “Anatomy
And Histology Of Normal Skin – Contents” Loyola
University Dermatology Medical Education Website
[3] H. Zhang , J. Fritts and A. Goldman, “Image
Segmentation Evaluation: A Survey Of Unsupervised
Methods,”
Computer
Vision
And
Image
Understanding 110 (2008) ,pp- 260–280, July, 2008.
[4] K.. Freedberg, A. Geller, D. Miller, R. Lew, and
H. Koh, “Screening For Malignant Melanoma: A
cost- effectiveness analysis,” J.Amer. Acad.
Dermatol., vol. 41, no. 5, pp. 738–745, November,
1999.\
[5] A. Jerants, J. Johnson, C. Sheridan, and T.
Caffrey, “Early detection and treatment of skin
cancer,” Amer. Family Physics, vol. 62, no. 2, pp. 1–
6, July, 2000.
[6]
H. C. Engasser and E. M.Warshaw,
“Dermatoscopy Use By US Dermatologists:A CrossSectional Survey.” J. Amer. Acad. Dermatol., vol. 63,
no. 3,pp. 412–419, 2010.
[7] P. G. Cavalcanti and J. Scharcanski, “Automated
Prescreening Of Pigmented Skin Lesions Using
Standard Cameras,” Comput. Med. Imag. Graph.,vol.
35, no. 6, pp. 481–491, Sep. 2011.
[8] J. Glaister, R. Amelard, A. Wong, and D. A.
Clausi, “MSIM: Multi-Stage Illumination Modeling
Of Dermatological Photographs For Illumination
corrected Skin Lesion Analysis,” IEEE Trans.
Biomed. Eng., vol. 60, no. 7,pp. 1873–1883, Jul.
2013.
[9]
Nikhil Cheerla, Debbie Frazier, “Automatic
Melanoma Detection Using Multi-Stage Neural
Networks” International Journal of Innovative
Research in Science,Engineering and Technology
[10] Fairchild,“ Color And Image Appearance
Models” John Wiley and Sons, 2005, 340−3855
[11] A. Mehta; A. S. Parihar; N. Mehta,” Supervised
classification of dermoscopic images using optimized
fuzzy clustering based Multi-Layer Feed-forward
Neural Network.” Computer, Communication and
Control (IC4), 2015 International Conference on
Year: 2015 Pages: 1 - 6,
Page 1782