[ieee 2014 international conference on computing for sustainable global development (indiacom) - new...

4
GLCM and Multi Class Support Vector Machine based Automated Skin Cancer Classification Ritesh Maurya 1 , Surya Kant Singh 2 , Ashish K. Maurya 3 , Ajeet Kumar 4 1,3,4 Department of Computer Science & Engineering Shri Ramswaroop Memorial University, U.P., India 2 Department of Computer Science & Engineering GLA University, Mathura, India 1 [email protected], 2 [email protected], 3 [email protected], 4 [email protected] Abstract-It is utmost important to early detect the skin cancers.Proper diagnosis is critical for survival of the patient .Biopsy method of detection is much painful .We have proposed an automated system for detection and classification of one of the skin four types of skin cancers: Melanoma ,Basal cell carcinoma, actinic Keratosis, Squamous cell carcinoma. There are a certain features of these types of skin cancers, which can be extracted using proper feature extraction algorithm. The features of skin lesions are extracted normalized symmetrical Grey Level Co-occurrence Matrices (GLCM).GLCM based texture features are extracted from each of the four classes and given as input to the Multi-Class Support vector machine which is used for c1assification purpose. It c1assifies the given data set into one of the four skin cancer classes. The accuracy of our proposed method is 81.43%. Keywords- Grey Level Co-occurrence Matrices, multi-class Support Vector Machine ,Gray level,Texture features,Color- coherence vector,Global color Histogram I. INTRODUCTION Skin cancer is a deadly disease. Melanoma and Non- Melanoma are two main types of skin cancers. Each type of skin cancer is different from the other skin cancers in certain characteristics. In this paper we approach a technique for the classification of four different types of skin cancers i.e. Melanoma, Basal cell carcinoma, actinic Keratosis, Squamous cell carcinoma. Each type of skin cancer is having a certain distinguishing characteristics. In this paper we have devised an approach for classifying four types of skin. Co occurrence Matrix (GLCM) Rajeswari. S et. al [21] used a similar approach for MRI brain images, in which the image is first preprocessed using median filter and then GLCM is used to extract features from an image, for classifying images into benign, malign and normal multi- class support vector machine is used .GLCM defines the probability of the occurrence of the combination of one grey tone at a specified distance and in a direction. N. K. Al abbadi et. al. [3] proposed a method for skin texture recognition using both color & texture features and classification using three layer neural networks. In [2] J.M. Rubegni et. al. proposed a method of diagnosis of pigmented skin lesions based on digital dermoscopy analysis. In [12] F. Smach et. al. proposed a system for face recognition and uses features such as skin color, texture, shape of the face and classifies these images using MLP neural network. Many diagnosis systems aided by computer have been suggested for use with trained dermatologists with their accuracy comparable to that of dermatologists [15]. The huge number of variations in the structure as well as colors of lesions implies that cell carcinoma, actinic Keratosis, Squamous cell carcinoma. Each type of skin cancer is having a certain distinguishing characteristics. In this paper we have devised an approach for classifying four types of skin cancers such as Melanoma, Basal cell carcinoma, actinic Keratosis, Squamous cell carcinoma Melanoma and Non-Melanoma are two major categories of skin cancers. Malignant melanoma is of several sub-types, and basal cell carcinoma, squamous cell carcinomas are two main types of non-melanoma skin cancers. Some of the distinguishing characteristics of these skin cancers are as follows: Basal cell carcinomas have a resemblance of pale moles, can be described as smoothy, symmetric in appearance and wart- like bumps. They may be either in flesh color, pale or can be reddish also. Squamous cell carcinomas generally begin as Pre-cancerous lesions with flat areas of scaling and skin- redness, known as actinic keratoses, having a hard white scale with red as a base. In the growth phase of squamous cell carcinomas, they become deeper red in color, with the enlarge diameter [modified from 1]. Melanomas’ appearance is much more variable. “ABCDE” method is generally used technique for the classification of melanoma appearance [18]. In [4] approach developed by Haralick based on statistical method to extract texture patterns using Grey Level diagnostic techniques must take a number of features into account. A significant number of studies have proven that lesion quantification maybe of prime importance in clinical practice, because skin lesions can be identified based on these quantifiable features extracted from the images of lesions ([13], [14], [15], [16]). The “2CSVM”algorithm [17] is applied to generate a classifier by placing different weights on positive and negative samples. Many classifiers such as ANN and SVM have been used in detection of malignant melanoma, which is two-class classification problem ([14, 20,16, 8]). 444 978-93-80544-12-0/14/$31.00 c 2014 IEEE

Upload: ajeet

Post on 20-Feb-2017

213 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: [IEEE 2014 International Conference on Computing for Sustainable Global Development (INDIACom) - New Delhi, India (2014.3.5-2014.3.7)] 2014 International Conference on Computing for

GLCM and Multi Class Support Vector Machine based Automated Skin Cancer Classification

Ritesh Maurya1, Surya Kant Singh2, Ashish K. Maurya3, Ajeet Kumar4

1,3,4 Department of Computer Science & Engineering Shri Ramswaroop Memorial University, U.P., India

2Department of Computer Science & Engineering GLA University, Mathura, India

[email protected], [email protected], [email protected], [email protected]

Abstract-It is utmost important to early detect the skin cancers.Proper diagnosis is critical for survival of the patient .Biopsy method of detection is much painful .We have proposed an automated system for detection and classification of one of the skin four types of skin cancers: Melanoma ,Basal cell carcinoma, actinic Keratosis, Squamous cell carcinoma. There are a certain features of these types of skin cancers, which can be extracted using proper feature extraction algorithm. The features of skin lesions are extracted normalized symmetrical Grey Level Co-occurrence Matrices (GLCM).GLCM based texture features are extracted from each of the four classes and given as input to the Multi-Class Support vector machine which is used for c1assification purpose. It c1assifies the given data set into one of the four skin cancer classes. The accuracy of our proposed method is 81.43%. Keywords- Grey Level Co-occurrence Matrices, multi-class Support Vector Machine ,Gray level,Texture features,Color-coherence vector,Global color Histogram

I. INTRODUCTION Skin cancer is a deadly disease. Melanoma and Non-

Melanoma are two main types of skin cancers. Each type of skin cancer is different from the other skin cancers in certain characteristics. In this paper we approach a technique for the classification of four different types of skin cancers i.e. Melanoma, Basal cell carcinoma, actinic Keratosis, Squamous cell carcinoma.

Each type of skin cancer is having a certain distinguishing characteristics. In this paper we have devised an approach for classifying four types of skin.

Co occurrence Matrix (GLCM) Rajeswari. S et. al [21] used a similar approach for MRI brain images, in which the image is first preprocessed using median filter and then GLCM is used to extract features from an image, for classifying images into benign, malign and normal multi-class support vector machine is used .GLCM defines the probability of the occurrence of the combination of one grey tone at a specified distance and in a direction.

N. K. Al abbadi et. al. [3] proposed a method for skin texture recognition using both color & texture features and classification using three layer neural networks. In [2] J.M. Rubegni et. al. proposed a method of diagnosis of pigmented skin lesions based on digital dermoscopy analysis. In [12] F. Smach et. al. proposed a system for face

recognition and uses features such as skin color, texture, shape of the face and classifies these images using MLP neural network.

Many diagnosis systems aided by computer have been suggested for use with trained dermatologists with their accuracy comparable to that of dermatologists [15]. The huge number of variations in the structure as well as colors of lesions implies that cell carcinoma, actinic Keratosis, Squamous cell carcinoma.

Each type of skin cancer is having a certain distinguishing characteristics. In this paper we have devised an approach for classifying four types of skin cancers such as Melanoma, Basal cell carcinoma, actinic Keratosis, Squamous cell carcinoma Melanoma and Non-Melanoma are two major categories of skin cancers. Malignant melanoma is of several sub-types, and basal cell carcinoma, squamous cell carcinomas are two main types of non-melanoma skin cancers. Some of the distinguishing characteristics of these skin cancers are as follows: Basal cell carcinomas have a resemblance of pale moles, can be described as smoothy, symmetric in appearance and wart-like bumps. They may be either in flesh color, pale or can be reddish also. Squamous cell carcinomas generally begin as Pre-cancerous lesions with flat areas of scaling and skin-redness, known as actinic keratoses, having a hard white scale with red as a base. In the growth phase of squamous cell carcinomas, they become deeper red in color, with the enlarge diameter [modified from 1]. Melanomas’ appearance is much more variable. “ABCDE” method is generally used technique for the classification of melanoma appearance [18]. In [4] approach developed by Haralick based on statistical method to extract texture patterns using Grey Level diagnostic techniques must take a number of features into account. A significant number of studies have proven that lesion quantification maybe of prime importance in clinical practice, because skin lesions can be identified based on these quantifiable features extracted from the images of lesions ([13], [14], [15], [16]). The “2CSVM”algorithm [17] is applied to generate a classifier by placing different weights on positive and negative samples. Many classifiers such as ANN and SVM have been used in detection of malignant melanoma, which is two-class classification problem ([14, 20,16, 8]).

444978-93-80544-12-0/14/$31.00 c©2014 IEEE

Page 2: [IEEE 2014 International Conference on Computing for Sustainable Global Development (INDIACom) - New Delhi, India (2014.3.5-2014.3.7)] 2014 International Conference on Computing for

In this paper we proposed an approach for extracting the features of a lesions by utilizing the power of texture feature extraction and also used multi-class Support Vector Machine (MSVM) which is implemented as a combination of binary Support Vector Machines (SVMs) for the training and classification purpose for automatic detection as well as classification of four different types of skin cancers discussed earlier [6].

The rest of the paper is organized as follows. Section 2 briefly reviews Global Color Histogram (GCH) and Color Coherence Vector (CCV) which is used in proposed frame work. Section 3 reports extensive experimental results and Section 4 concludes the paper.

II. PROPOSED FRAMEWORK

A.Texture Feature Extraction using GLCM Texture features of skin cancer images are used to check the accuracy and efficiency of our proposed approach.

From Gray Level Co-occurrence matrix (GLCM) various features are the features that are used for the skin cancer classification such as autocorrelation, Contrast , Correlation, Energy, Entropy, homogeneity, maximum probability , sum of squares, sum average, sum variance .The process is shown in figure 1.

Gray Level Co-occurrence matrix (GLCM): There are some distinguished features that seperates each class of skin cancers. Feature extraction algorithms extracts an distinguish features from skin cancer images of each class. After feature extraction we got features that are useful in processing[7].We have used feature extraction technique i.e. Gray Level Co-occurrence Matrix (GLCM).The RGB image first get converted into gray scale image and given as input to the GLCM, in which the number of rows and columns is equal to the number of gray levels. GLCM computes the frequency of certain number of gray levels reappearing at different position in an image[9]. GLCM feature extraction maps the gray level co-occurrence probabilities in different angular position based on spatial relationship between different combination of pixels. The following features are extracted from matrix based on GLCM: autocorrelation, Contrast, Energy, Entropy, homogeneity. Feature Explanation Formula Contrast It measures the

coarseness of texture

Energy It measures the image textural uniformity

Entropy It measures the degree of disorder or

Homogeneity

It measures the distribution of element in

Table1. Explanation and Formulas of common selected GLCM Feature.

B.Tranning and Classification An approach [6] is presented which can combine more than two SVM classifiers for more than two class classification. Multi-class classification is performed by combining the results of various binary SVM classifiers.

Binary Classifiers are treated as base learners. Multi-class problem can be mapped into two-class problem by using dive and conquer approach. N × (N-1)/2 binary classifiers will be needed for the class classification of N different classes.

Figure1. Block Diagram for Proposed Frame Work

According to the author, ixj binary classifier treats the pattern of I class positive and the jth class as negative. After that the minimum distance is calculated from the binarypattern ID of each class and generated vector (binary).Test image will belong to the class, from which the generated ID has minimum distance with the class ID.

The approach can be understood by taking example of the four class classification problem. Let there be the four classes x, y, x and w. In order to classify four classes we need n(n-1)/2 binary classifiers ,i.e 6 as shown in TABLE I ((i.e., x×y, x×z, x×w y×z, y×w, w×z) will be used as base learners, and each image will be given as input to all of the six binary classifiers as a training image, from the outputs generated by each classifiers we compute the unique ID for each class as shown in TABLE I. In order to compute the entries a binary comparison is made for all of the six classifiers and +1 is entered for which the input image

Training set

SSC Data-

set

Feature extraction By

GLCM

Feature extraction By

GLCM

Image pre Preprocessing

Testing set

Images

Training By MSVM

Classification by MSVM

Skin Cancer class detected

Converting images

into gray scale

2014 International Conference on Computing for Sustainable Global Development (INDIACom) 445

Page 3: [IEEE 2014 International Conference on Computing for Sustainable Global Development (INDIACom) - New Delhi, India (2014.3.5-2014.3.7)] 2014 International Conference on Computing for

belongs and -1 for the class and 0 is entered for the remaining third class. For example-binary comparison between classes x×y, if the input image belongs to the class x ,then +1 is assigned to it and -1 for class y and 0 for class w and z both. Same computation is done for all of the six binary classifiers and the corresponding entries are computed.

TABLE2 UNIQUE ID OF EACH CLASS

x×y x×z x×w y×z y×w w×zX +1 +1 +1 0 0 0 Y -1 0 0 -1 +1 0 Z W

0 0

-1 0

0 -1

+1 0

0 -1

-1+1

Each binary classifier will result a binary response for any

input example. Let’s say if the outcomes for the binary classifier x×y, x×z, x×w y×z, y×w, w×z are 0, -1, +1,0,0,+1 respectively then the input example will belongs to that class which have the minimum distance from the vector [0, -1, +1,0,0,+1 ].

Multi class support vector machine with the combination of multiple binary classifiers are used to train and classify the random input images.

III. EXPERIMENTS AND SIMULATION

A.Data-Set Preperation We have collected the skin cancer images from

http://dermnetnz.org of about 359 images and named our dataset as Skin Cancer Classification (SCC).The resolution of each image was 150*112, these images are grouped into four classes Melanoma, Basal-cell Carcinoma, Actinic Keratosis, Squamous-cell Carcinoma. The image in each of the following class is 77, 84,101 and 101 images respectively. These images are used to train the Multi-class support vector machine and also for classification purpose. Random image is given as input, which is classified the MSVM into one of the four classes. Some of the sample images from our database are shown in Fig. 2.

B.Result and Analysis All the experiments were done in MATLAB 7.12.The

GLCM features were extracted. The accuracy of our proposed method is 81.43%.for 75 training images which is far more better than the experiments performed by us using color coherence vector and global color histogram approach [5].The accuracy of our previously designed system was 76.38% for RGB format images 75 training images using color coherence vector-MSVM based classification and 71.84% with Global color histogram-MSVM based classification [5]. It can be clearly seen in fig.3 that approach used in this paper GLCM+MSVM gives accuracy of 81.43%. Recently Jaleel et al [11] conducted experiments using GLCM as a feature extraction method and Neural Network as a classifier for classification of only two classes-cancerous

and non-cancerous and achieves the accuracy of 88% for 55 training images, while we have classified images into four classes by combining the output of six support vector machines for 75 training images of each class and achieves the overall accuracy of 81.43%.

Figure2. Four images of each class Melanoma, Basal cell carcinoma,

Squamous cell carcinoma, and Actinic Keratosis are shown consecutively.

Figure .3 % Accuracy of GLCM-MSVM based classification for different

number of training images %Accuracy for each of the four classes is shown in fig.4 shown below. The %accuracy for class1 with 75 training images is 81.23%, for class 2 the % accuracy is 83.324% with the same no. of training images, and 82.213% and 84.12% for class 4 consecutively for 75 training images.

35 40 45 50 55 60 65 70 7550

60

70

80

90

no. of training images per skin class

Acc

urac

y(%

)

446 2014 International Conference on Computing for Sustainable Global Development (INDIACom)

Page 4: [IEEE 2014 International Conference on Computing for Sustainable Global Development (INDIACom) - New Delhi, India (2014.3.5-2014.3.7)] 2014 International Conference on Computing for

Figure .4 % Accuracy for each class for different no. of training Images

IV. CONCLUSIONS It can be easily concluded that the proposed system of skin cancer classification into one of the four given types of skin cancer is very useful for detection of skin cancers of multiple class with a good accuracy. Our proposed system is capable of detecting multiple cancers by the use of multiclass support vector machine, which works in a combination of multiple binary support vector machines. Other systems can also be developed in combination of various features.

REFRENCES

[1] http://dermnetnz.org/ [2] [2] Blackledge, J. M.; Dubovitskiy, D. A. (2009): Texture

classification using fractal geometry for the diagnosis of skin cancers, in Proceedings of EG UK Theory and Practice of Computer Graphics, UK, pp. 1-8.

[3] Al. Abadi, N. K.; Dahir, N. S.; Alkareem, Z. A. (2008): Skin texture recognition using neural network, in Proceedings of the International Arab Conference on Information Technology, Tunisia, December 16-18, pp.1-4.

[4] [Haralick, R.M. (1979): Statistical and structural approaches to Texture, Proceedings of IEEE, 67(5), pp. 784-804.

[5] Surya Singh, Maurya Ritesh ,”Skin Cancer classification using Multi Class Support Vector Machine in color space”, CERA-2013

[6] A. Rocha, C. Hauagge, J. Wainer and D. Siome, “Automatic Fruit and Vegetable Classification from Images,”

[7] Fakhry M. Khellah, “Texture Classification Using Dominant Neighborhood Structure”, IEEE Transactions on Image Processing, 2011, pp 3270-3279.

[8] Rubegni, P. et al. (2002): Automated Diagnosis on Pigmented Skin Lesions, International Journal on Cancer, 101, pp. 576-580.

[9] Mariam, A.Sheha,Mai, S.Mabrouk, Amr Sharawy, “Automatic Detection of Melanoma Skin Cancer using Texture Analysis”, International Journal of Computer Applications, Volume 42, 2012.

[10] Shubhankar Ray and Andrew Chan, “Automatic feature extraction from wavelet coefficients using Genetic Algorithms”, IEEE Transactions on Image Processing, pp. 980-1025, June 2000.

[11] J Abdul Jaleel, Sibi Salim, Aswin.R.B ,”Computer Aided Detection 01 Skin Cancer”, 2013 International Conference on Circuits, Power and Computing Technologies [ICCPCT-2013]

[12] Smach, F. et. al. (2006): Design of a neural network classifier for face detection, Journal of Computer Science, 2(3), pp. 257-260.

[13] Green, A., Martin, N., Pfitzner, J., O'Rourke, M., Knight N., “Computer image analysis in the diagnosis of melanoma,” Journal of the American Academy of Dermatology, vol. 31, no. 6, pp. 958-964, Dec. 1994.

[14] Ganster, H., Pinz, P., Rohrer, R., Wildling, E., Binder, M., Kittler, H., “Automated melanoma recognition,” IEEE Transactions on Medical Imaging, vol. 20, no. 3, pp. 233 -239, 2001.

[15] Kjoelen, A., Thompson, M., Umbaugh, S., Moss, R., Stoecker, W.,“Performance of artificial intelligence methods In automated detection of melanoma,” IEEE Engineering in Medicine and Biology, vol. 14, no. 4, pp. 411-416, 1995

[16] Maglogiannis, I., Zafiropoulos, Kyranoudis, C., “Intelligent segmentation and classification of pigmented skin lesions in dermatological images,”, SETN 2006, LNAI 3955, Springer-Verlag Berlin Heidelberg, pp. 214 – 223, 2006.

[17] K. Veropoulos, C. Campbell and N. Cristianini, “Controlling the Sensitivity of Support Vector Machines”, In Proceedings of IJCAI Workshop Support Vector Machines, May, 1999.

[18] Kiran Ramlakhan ,Yi Shang,Department ,“A Mobile Automated Skin Lesion Classification System” ,Computer Science,University of Missouri Columbia, Missouri

[19] Ning Situ , Xiaojing Yuan, Ji Chen, and George Zouridakis “Malignant Melanoma Detection by Bag-of-Features Classification”

[20] S. M. Rajpara, A. P. Botello, J. Townend, and A. D. Ormerod, “Systematic review of dermoscopy and digital dermoscopy/artificial intelligence for the diagnosis of melanoma,” British Journal of Dermatology, vol. 161, pp. 591–604, 2009.

[21] Rajeswari. S , Theiva Jeyaselvi. K, “Support Vector Machine Classification For MRI Images”, International Emerging Trends in Computer and Electronics Engineering Dubai,Journal of Electronics and Computer Science Engineering Volume1, Number 3, March 24-25, 2012

2014 International Conference on Computing for Sustainable Global Development (INDIACom) 447