Study On The Recognition Of Pigmented Skin Lesions Based On Support Vector Machines

The judgment of categories for skin lesions is essential to the treatment. With the application of computer and image processing technology in the fields of medicine, it's possible for us to establish an intelligent recognition system. The main technology of this system is multi-class classification for skin lesion images.As a new general pattern recognition method, SVM demonstrates unique advantage in small sample, nonlinear and high-dimensional pattern recognition. It can overcome the problems of over-fitting and local minima which often exist in traditional recognition methods. SVM has been widely concerned in medical applications.Traditional SVM is based on two-class classification. Unclassifiable regions exist in conventional SVM with the two-class problems extending to the multi-class problems using "One-against-One" or "One-against-Rest" strategy. In this paper, the principle of SVM is studied. Its multi-classification process is discussed. A new multi-class SVM algorithm based on fuzzy membership function is proposed.The main work in this paper includes two aspects: (1) Feature extraction of skin lesions; (2) Solving the blind spots problem by improving traditional SVM.First, images of symptoms are converted from RGB to HSV space, and are segmented in V domain. According to ABCD-rule, we extract 17 feature parameters including maximum diameter of the region, circularity etc.Secondly, we construct multi-classification based on SVM. In order to solve the blind spots problems, fuzzy membership function is introduced in this paper. As different samples have different contributions to the classification, blind spots can be classified by giving each sample a corresponding membership.Experimental results show that the proposed fuzzy support vector machine algorithm for the recognition of pigment skin lesions is effective and is conducive to such a clinical diagnosis of skin lesions.