Segmentation of Skin Lesion Images Using Combination of Texture and Color Information

If skin cancer is detected in the early stages, the survival rate is very high. So, computer-aided diagnosis (CAD) systems are being developed to help dermatologists in early and accurate diagnosis. A common CAD system is composed of three steps: 1) segmentation, 2) feature extraction, 3) classification. Segmentation is the first and most important step in the auto diagnosis systems. The purpose of this paper is to introduce a new method based on geometric active contours that combines texture and color information to separate the lesion area from healthy skin. Combination of texture and color information can play an important role in distinguishing between lesion and healthy skin pixels. The innovation of this paper is the way that, color and texture information are combined together to define the speed function and the use of texture features in the form of an image. In this method, in order to use the color information more effectively two color spaces CIE L*a*b* and CIE L*u*v*, have been adopted. For the texture features extraction, several methods of texture analysis including Gabor, GLCM, local entropy filter, local range filter and local standard deviation filter have been used. To evaluate the proposed method, two databases including dermoscopy images, were used: The ISIC2017 database (including 2750 data) and the PH2 database (including 200 data). Then, the results were compared with the recent works on these two databases. Experimental results showed that, the proposed algorithm has the highest accuracy (97.92% for PH2 database and 94.78% for ISIC 2017 test data), sensitivity (97.83% for PH2 database and 90.11% for ISIC 2017 test data) and specificity (99.45% for PH2 and 98.53% for ISIC 2017 test data) in comparison with recent state-of-the-art algorithms.