Improvement in the diagnosis of melanoma and dysplastic lesions by introducing ABCD-PDT features and a hybrid classifier

Melanoma and dysplastic lesions are pigmented skin lesions whose accurate classification is of great importance. In this paper, we have proposed a computer-aided diagnosis (CAD) system to improve the diagnostic ability of the conventional ABCD (asymmetry, border irregularity, color, and diameter) analysis. We introduced features extracted by local analysis of range of intensity variations within the lesion that describe pigment distribution and texture (PDT) features. The statistical distribution of pigmentation at a specified direction and distance was analyzed through grey level co-occurrence matrix (GLCM). Some other quantitative features were also extracted by computing neighborhood grey-tone difference matrix. These were correlated with human perception of texture. A hybrid classifier was designed for classification of melanoma, dysplastic, and benign lesions. Log-linearized Gaussian mixture neural network (LLGMNN), K-nearest neighborhood (KNN), linear discriminant analysis (LDA), and support vector machine (SVM) construct the hybrid classifier. The proposed system was evaluated on a set of 792 dermoscopy images and the diagnostic accuracies of 96.8%, 97.3%, and 98.8% for melanoma, dysplastic, and benign lesions were achieved, respectively. The results indicate that PDT features are promising features which in combination with the conventional ABCD features are capable of enhancing the classification performance of the pigmented skin lesions.