A Comparative Study of Different Methods for Pigmented Lesion Classification Based on Color and Texture Features

Abstract

This paper presents a method that classifies color dermoscopic images into their different dermatologic patterns. CSGV (composite subband gradient vector) is used to represent each pattern. CSGV is obtained from the gradient vectors generated from the different sub-images in a wavelet decomposition. Classification results are compared with those obtained applying Gabor filters and Markov Random Field (MRF). Classification is performed with a fuzzy-ARTMAP. Performance is analysed in L*a*b* and RGB color spaces. L*a*b* provides the best results (81,25%).There are two main contributions in this work. The first one is to combine features from already existing methods with color information, in the RGB or L*a*b* representations. The second contribution is that a large study of different features applied to the specific problem of dermoscopic pattern classification has been implemented.