IJCSNS International Journal of Computer Science and Network Security, VOL.16 No.

4, April 2016 135

Classification of Dermoscopic Skin Cancer Images Using Color

and Hybrid Texture Features
Ebtihal Almansour and M. Arfan Jaffar††,
ebtihal.99@hotmail.com, arfan.jaffar@ccis.imamu.edu.sa
Al Imam Mohammad Ibn Saud Islamic University (IMSIU), Riyadh, Saudi Arabia

most fatal type of skin cancer caused by melanoma.As

Summary statistical studies in the United States, It shows that 76,690
Malignant Melanoma is one of the rare and the deadliest form of patients with melanoma and 9,480 of them passed away
skin cancer if left untreated. Death rate due to this cancer is three with the caused of melanoma in 2013 [5]. In Canada,
times more than all other skin-related malignancies combined. Melanoma occurred about 1.4% of all cancer deaths. There
Incidence rates of melanoma have been increasing, especially are around 6500 diagnosed with melanoma and 1,050 of
among young adults, but survival rates are high if detected early.
Unfortunately, the time and costs required for dermatologists to
them passed away in 2014 [4]. The expectation of
screen all patients for melanoma are prohibitively expensive. growing the melanoma in Canadians during their lifetime
There is a need for an automated system to assess a patient's risk is 1 in 73 women and 1 in 59 men; 1 in 395 women and 1
of melanoma using digital dermoscopy, that is, a skin imaging in 240 men will die of it [4]. During the past 25 years, the
technique widely used for pigmented skin lesion inspection. In incident rate of melanoma has been increased [4]. The 5-
this research, we aim to propose an intelligent automated method year relative survival rate for melanoma is 92% in women
for identification of the type of skin lesions using machine- and 85% in men [4].One of the most important factors to
learning techniques. Two types of texture feature have been used reduce the mortality rate of melanoma is detecting it early.
to perform classification of melanoma and non-melanoma. First But distinguishing the skin cancer from other benign
local information through Local Binary Pattern (LBP) on
different scales and Gray Level Co-Occurrence Matrix (GLCM)
pigmented skin lesions is a big challenge and not an easy
at different angles has been extracted as a texture features. These task even for dermatologists. Several clinical methods
features are robust due to scale invariant property of LBP and have been used to improve diagnostic accuracy, but
rotation invariant property of GLCM features. Global effective ways to extend the diagnoses to dermatologist are
information of different colors channels has been incorporated still lacking. Hence, the motivation of developing a
through four different moments extracted in six different color computer aided diagnose system was most evident these
spaces like RGB, HSV, YCbCr, NTSc, CIE L*u*v and CIE days. In this research, we will proposed a method to
L*a*b. Thus a fused hybrid texture local and color as global classify the pigmented dermoscopic images into melanoma
features has been proposed to classify the melanoma and non- and non-melanoma.
melanoma. Support vector machine has been used as a classifier
to classify melanoma and non-melanoma. Experiments have been
tested on well-known dataset dermis that is freely available on
the Internet. The proposed method has been compared with state 2. Related Work
of the art methods and shows better performance in comparison
to the existing methods. Stoecker et al. [6] have been used basic statistical
Key words: approaches, such as the gray-level co-occurrence matrix,
Feature extraction, glcm, lbp , svm , color feature to analyze texture in skin images. They found that texture
analysis approach could accurately and regions with a
smooth texture and that texture analysis can be applied to
1. Introduction both segmentation and classification of dermoscopy
images. Sonali et al. [7] combined Thresholding
Skin cancer is a type of dangerous diseases diagnosed segmentation technique to establishing boundaries in
around the world. It can be divided into melanoma and image with Fuzzy C-Means segmentation to fi\,m , lml,
non-melanoma [1]. Melanoma cancer is less common than nd final segmentation algorithm S. ManjuBharathi , S.
non-melanoma, however, the probability to spread on the Saraswathi .[8] proposed algorithm based on NC ratio
skin tissue and cause fatal is high [2]. Although the skin analysis in automatic cell segmentation. The experimental
cancer is death disease and may affect the human life, but result shows that high efficiency and accuracy of
it can be treated if detected in early stage. According to the segmentation process for cancer cell. S. Jeniva , C. Santhi
previous researches, if the cancer detected in early stage, [9] used the concept of texture distribution based on a
the treatment rate will be more than 90% while it will be learned model of natural skin and lesion textures. The
less than 50% if detected lately [3]. In recent statistical, the texture distribution metric captures the difference between

Manuscript received April 5, 2016

Manuscript revised April 20, 2016
136 IJCSNS International Journal of Computer Science and Network Security, VOL.16 No.4, April 2016

pair of texture distribution. Then, based on similarity, the important to important, combining these features together
images are divided into large number of smaller regions. may give us good classification results. The fourth and
This achieves higher segmentation accuracy. Ramya et al. final step is using Support Vector Machine (SVM)
[10] the segmentation approach proposed in this paper classifier in order to classify the input image into one of
focus on identifies skin cancer in epidermis layer of skin. the two classes melanoma and non-melanoma. In next
The nuclei regions; which located on epidermis layer; sections, a detailed description of our contribution is
segment using the K-means clustering algorithm based on mentioned. For more illustration, below diagram shows
space and some color information with k value equal to 3. our proposed method.
After that, local region recursive segmentation (LRRS)
algorithm which used intensity and size of nuclei as
parameter to filter the candidate nuclei regions is
performed to discover the region of nuclei. Final step is
applying local double ellipse descriptor (LDED) to
distinguish melanocytes from keratinocytes. This approach
has good performance even if the original image is
complex where background and foreground both have
similar appearance. Nidhalet al. [11] The proposed
Fig 1 Proposed Method
approach uses Wiener filter to remove noise such as hair
from original image then, used thresholding to segment the
skin cancer area from the whole image. Testing this 3.1 Feature Extraction
method is provided by comparing the result of
3.1.1 Color Feature
segmentation of this approach with the one done by
experts in medical filed and measures the distance between Color is the appearance of an object when exposed to light,
these two results by using HM and TDR gives high this definition will help us understand the chromatic
accuracy with 96.32%.Cheng Luet al.[12]This paper features such as color space. The aspect of color space can
proposed segmentation technique of the melanocytes in the be understood through studying the three primary colors
skin histopathological image. First, using mean shift and such as red, green and blue. The aspect of color mixing
local region recursive segmentation (LRRS) algorithm to can also be demonstrated by employing CMYK as well as
extract nuclei areas. Then, the local double ellipse HSL among others. The objective of using color feature
descriptor (LDED) integrates the feature of melanocytes technique is identifying the presented color in the
and provides parameters to identify the melanocytes. segmented lesion regions. This can be achieved by
Using 30 images with different factors as sample to test extracting four statistics usually called color moments or
this approach showing that this technique has the ability of color feature which is mean, standard deviation, variation
segmentation of melanocytes with over 80% sensitivity and skewness through individual channels of six various
rate and over 70% positive prediction rate. color spaces: RGB, HSV, YCbCr,NTSc, CIE L*u*v and
BinamrataBaralet al.[13]. The proposed technique showing CIE L*a*b 3, 7, 8 from that segmented lesion areas. Let P
in this paper for segmentation is based on Neuro-Fuzzy is the color channel, i is the image, N is number of pixels
model using decision-making. Segmentation is performed of image on a color space.Pj is the jth pixel of that color
with some features works as parameters. This approach channel P of an image i with N pixels in a color space.
gives good accuracy and quality. The definition of the four color features/ moments are
shown below:
• Moment 1- Mean is the average value of color
3. Proposed Method values in the channel which is calculated by
Our proposed method combines many steps and below expression,
techniques in order to get accurate and robust (1)
classification results. First, we collect the dermoscopic • Moment 2- Standard deviation is the square root
melanoma & non-melanoma images from DermIs. The of the variance of the distribution, which is
second step is to extract features from that segmented computed by,
image. Since these images have some texture
characteristics, we will use two common texture feature
extraction algorithms; Local Binary Pattern (LBP) and (2)
Gray Level Co-Occurrence Matrix (GLCM). In addition, • Moment 3- characterizes the degree of
the color is an important feature, which distinguish skin asymmetry of a distribution around its mean
lesion to others. Hence, extracting color feature is which is given by,
IJCSNS International Journal of Computer Science and Network Security, VOL.16 No.4, April 2016 137

how often different combinations of pixel luminance

(3) values (grey levels) occur in a specific pixel pairing of an
image. Using a two-dimensional gray-level co-occurrence
• Moment 4- Variance is the variation of the color matrix is commonly and widely used in the field of texture
distribution, which is calculated by below analysis In order to find the locative dependence of
expression. brightness (gray-level) values, which helps to find
valuable information about the neighboring pixels in an
(4) image.
The definition of co-occurrence matrix p of an image I of
The above described four features are calculated over size N*N is given below
every single channel which results in 72 color features
obtained by the following combination: (4 features) × (6
color spaces) × (3 channels in each color space). (6)
Here, the distance between the interested pixel and its
3.1.2 Texture Feature neighbor is defined as the offset (Δx ,Δy ).
Then, we can extract four different features from GLCM.
In this research, we combined two features, color and The feature of Entropy, Energy, Contrast and
texture. For texture feature, local binary pattern and gray Homogeneity are given below:
level co-occurrence metrics are combined.

3.1.2.1Local Binary Pattern (7)

Local images on the different representations of original

skin pigment image .The local Binary Pattern (LBP) is (8)
such type of a feature that transforms the image into an
array. Hence, we have applied LBP operator on every
(9)
pixel of the image in order to obtain the coded LBP image.

The idea of LBP is to compare each pixel on the image

(10)
with its neighbors. The procedure is as follows. Each pixel
is compared with its 3 x 3 neighborhood that is comprised 3.2 Classification
of eight other pixels. In that process the center pixel value
is subtracted by all the neighbors. The resulting negative We will use support vector machine (SVM) classifier to
values are labeled as 0, and all the others with 1. classify the input image into one of two groups: melanoma
Afterwards, for each pixel, the binary values, starting from and non-melanoma. We chose support vector machine
the one of its top-left neighbor, are concatenated in a (SVM) to be a classifier in this study since it interprets
clockwise direction, creating a new binary number. The image much quickly as well as effectively. The support
obtained decimal value is then used for labeling the given vector machine (SVM) classifier also has the ability for the
pixel and is referred to as LBP codes [20, 21]. detection and classification of the complex pattern of data
Given a pixel (xc, yc), LBP can be formally expressed in given. All the features of dataset images are being used for
decimal form: the purpose of training. All the data belongs to the same
group or classed sharing some features.
(5)

Where and are gray-level values of the central pixel 4. Results and Discussions
and surrounding pixels in the circle neighborhood with a 4.1 Dataset
radius . The function is defined to be 1 for all x ≥ 0
In this research, we evaluate the proposed method
and to be 0 for all x < 0.
performance by using a dataset of 69dermoscopic images.
3.1.2.2 Gray Level Co-Occurrence Matrix (GLCM) The images have been collected from Dermatology
Information System (DermIS) database [15].Those images
We use five of the classical statistical texture measures of belong to two main classes, melanoma and non-melanoma.
Haralick et al. [14]: entropy, energy, contrast, correlation The total number of melanoma images is 43 and non-
and homogeneity, which are derived from a grey level co- melanoma is 26.
occurrence matrix (GLCM). The GLCM is a tabulation of
138 IJCSNS International Journal of Computer Science and Network Security, VOL.16 No.4, April 2016

4.2 Feature Extraction and Classifications Table 1 Performance Results for Different Feature Sets
Accuracy Specificity Sensitivity
For simplicity of discussion and analysis, the following (%) (%) (%)
notation is used throughout this section: µ σ µ σ µ σ
• SL{set of Local Binary Pattern features 85.24 2.32 72.56 2.05 88.23 2.34
GLCM
describing texture}.
• SG {set of GLCM features}. LBP 87.35 1.04 75.17 2.02 90.24 1.49
• ST {set of texture features by appending SL and
SG (ST=SL U SG) . 83.50 1.24 91.18 1.25 78.14 2.13
Color
• SC {set of Color chromatic features obtained
Texture 87.98 1.34 77.53 2.42 93.14 1.82
from four different color spaces}
• STC {set of texture color features by appending ST
and SC (STC=ST U SC) Proposed 90.32 2.36 85.84 1.23 93.97 2.05
Method
All these different feature sets have been used for training
to the SVM and for testing. Three different measures like
Analyzing the Results of GLCM Features SG
accuracy, sensitivity and specificity has been measured to
Interesting metrics were observed for the SGfeature set.
check the performance of all these different feature sets.
The results show that it attains very high sensitivity
The performance of classifiers is calculated and analyzed
(consistently around 90%) but conversely low specificity
by the following performance measures.
(consistently below 80%). This behavior is due to the
effects exhibited by using a linear classifier in a low
 Accuracy: Number of classified mass / Number of dimensional feature space on a small data set. This can be
total mass= (TP + TN)/(TP + TN + FP + FN)
explained by considering the following case.
 Sensitivity: Number of correct classified
Analyzing the Results of Local Binary Patterns SL
malignant mass / Number of total malign masses
Interesting metrics were also observed for the SL feature
=(TP) /(TP + FN)
set. The results show that it attains very high specificity
 Specificity: Number of correct classified benign (consistently around 90%) but conversely low sensitivity
masses / Number of total benign masses (consistently below 75%). This behavior is due to the
effects exhibited by using a linear classifier in a low
= (TN) /(TN + FP) dimensional feature space on a small data set. This can be
explained by considering the following case.
Where TP is True positive, FP is false positive FN is false
negative and TN is true negative. The performance of a Analyzing the Results of Texture Features ST
classifier could be estimated in terms of the number of true The results show that ST feature set achieves very high
positives and false positives. The performance validation accuracy around 90 % sensitivity (consistently around
measures are accuracy, sensitivity, and specificity. 90%) and specificity (consistently more than 95%).
The results using the classification scheme with the feature
sets SG, SL, ST are summarized in table 1. The sensitivity Analyzing the Results of Color Features SC
and specificity scores were therefore obtained over 10 The most remarkable finding for ST feature set is its very
independent classification runs. The mean and standard high accuracy around 90 % sensitivity (consistently around
deviation of the accuracy metrics across the 10 trials were 90%) and specificity (consistently more than 95%).
used to show the consistency of the results. The best mean
results in table 10 are bolded for each metric (i.e., Analyzing the Results of Texture Color Features STC
sensitivity, specificity, accuracy). The following ST feature set demonstrates very high accuracy around
observations can be made from the results. ST consistently 90 % sensitivity (consistently around 90%) and specificity
attains slightly higher accuracy, sensitivity and specificity (consistently more than 95%).
metrics than the individual feature sets. Scatter plots in
below figure shows different features behaviors. Like 1st
and 2nd rows show that features are overlapped with each
other's and difficult to classify as comparative to the
features shown in 3th row where features are separated
clearly and less overlapped.
IJCSNS International Journal of Computer Science and Network Security, VOL.16 No.4, April 2016 139

Table 2. Comparing the proposed results with existing results [3] R. Anderson, A. Parrish, "The optics of human skin", The
Authors Accuracy Sensitivity Specificity Journal ofInvestigative Dermatology, Vol. 77, pp 13-19,
(%) (%) (%) 1981.
[4] “http://www.phac-aspc.gc.ca/cd-
Cavalcanti 79.61 83.19 74.71
mc/cancer/melanoma_skin_cancer-cancer_peau_melanome-
et. al [16] eng.php”
Amelard 87.38 90.76 82.76 [5] Public Health Agency of Canada, "Melanoma skin cancer,"
et. al [18] http://www.phac-aspc.gc.ca/cd-mc/cancer/melanoma skin
Amelard 81.17 96.64 65.06 cancer-cancer peaumelanome-eng.php, 2013, Accessed: 29
Apr 2013.
et. al [17]
[6] Stoecker, Williarn V., Chang-Sung Chiang, and Randy H.
Proposed 90.32 85.84 93.97 Moss. "Texture in skin images: comparison of three
Method methods to determine smoothness." Computerized medical
imaging and graphics 16, no. 3 (1992): 179-190.
[7] SONALI RAGHUNATH JADHAV, D.K.KAMAT,
4.3 Discussion and Limitation “SEGMENTATION BASED DETECTION OF SKIN
CANCER”, IRF International Conference, 2014
One over-arching conclusion can be drawn from the [8] Manjubharathi, S., and S. Saraswathi. "CANCER CELL
SEGMENTATION AND DETECTION USING NC
experimental results: Color is important to distinguish
RATIO ANALYSIS." in International Journal of Research
melanoma and non-melanoma. So,using different color in Engineering and Technology , 2014
channels moments will extract important features.Texture [9] Jeniva, S., and C. Santhi."An Efficient Skin Lesion
captures relevant information for melanoma detection. We Segmentation Analysis Using Statistical Texture
showed that a small set of GLCM can increase Distinctiveness." algorithms (2015).
classification performance when combined with a large set [10] P. Ram ya, “Skin Histopathological Image Used LDED
of LBP features, The performance of the GLCM Based Algarithmwith Automated Detect ion of Melanocytes
themselves are not as powerful as we would have hoped, Skin”, Middle-East Journal of Scientific Research 20 (8):
however they do present one significant advantage over 986-991, 2014.
large LBP feature sets: a small GLCM set requires much [11] NidhalKhdhair El Abbadi and Abbas Hussien Miry, “ AUT
OMAT IC SEGMENTATION OF SKIN LESIONS USING
less data to adequately populate the multi-dimensional HISTOGRAM T HRESHOLDING”, Science Publications,
feature space. A larger data set may allow the LBP set to Journal of Computer Science 10 (4): 632-639, ISSN: 1549-
perform better, as the classifier could be trained on more 3636, 2014
representative class distributions. [12] Cheng Lu, Muhammad Mahmood, NareshJha,
MrinalMandal, “Automated Segmentat ion of the
Melanocytes in Skin Histopathological Images”, IEEE
5. Conclusion JOURNAL OF BIOMEDICAL AND HEALTH
INFORMATICS, VOL. 17, NO.2, MARCH 2013.
In this research, we have proposed a new feature [13] BinamrataBaral, SandeepGonnade, ToranVerma , “ Lesion
extraction technique for classification of dermoscopic Segmentation in Dermoscopic Images Using Decision
images into melanoma and non-melanoma. Two types of Based Neuro Fuzzy Model”, BinamrataBaral et al, /
(IJCSIT) International Journal of Computer Science and
features have been used, color and texture. For texture Information Technologies, Vol. 5 (2) , 2014, 2546-2552.
features, GLCM and LBP have been used. Combining [14] Haralick, Robert M., KarthikeyanShanmugam, and Its'
these features improves the accuracy of the classification HakDinstein. "Textural features for image
results. In this way, our proposed technique has been able classification." Systems, Man and Cybernetics, IEEE
to better classify dermoscopic images into Melanoma and Transactions on 6 (1973): 610-621.
Non-Melanoma groups. In order to evaluate the usefulness [15] "DermIS." DermIS.Web.
and performance of proposed model, experimentation is [16] Cavalcanti, Pablo G., and Jacob Scharcanski. "Automated
performed on standard dataset of dermIS. The experiments prescreening of pigmented skin lesions using standard
showed good results for the proposed methodology. Both cameras." Computerized Medical Imaging and Graphics 35,
no. 6 (2011): 481-491.
qualitative and quantitative error measures are used to [17] Amelard, R., Glaister, J., Wong, A., &Clausi, D.
assess the system performance. (2015).High-Level Intuitive Features (HLIFs) for Intuitive
Skin Lesion Description. Biomedical Engineering, IEEE
References Transactions on, 62(3), 820-831.
[1] “http://www.cancer.org/cancer/skincancer/index.” [18] Amelard, R., A. Wong, and D. A. Clausi. "Extracting
[2] Robert S. Porter, ”The Merck Manual”, July 20, 2011. & D. Morphological High-level Intuitive Features (HLIF) for
L. Kasper, E. Braunwald, and A. Fauci, “Harrison’s Enhancing Skin Lesion Classification." 2012 Annual
Principles of Internal Medicine”, 17th ed., McGraw-Hill, International Conference of the IEEE Engineering in
New York (2008). (put 6 and 7 together) Medicine and Biology Society (2012)