Description of visual content in dermoscopy images using joint histogram of multiresolution local binary patterns and local contrast

Abstract

Melanoma is a deadly form of skin lesion for which the mortality rate can be significantly reduced, if detected at an early stage. Clinical findings have shown that an early detection of melanoma can be done by an inspection of visual characteristics of some specific regions (lesions) of the skin. This paper proposes a pattern recognition system that includes three vital stages to conform the analysis of skin lesions by the clinicians: segmentation, feature extraction, and classification. Segmentation is performed using active contours with creasness features. The feature extraction phase consists of a variant of local binary pattern (LBP) in which joint histogram of LBP pattern along with the contrast of the patterns are used to extract scale adaptive patterns at each pixel. Classification was performed using support vector machines. Experimental results demonstrate the superiority of the proposed feature set over several other state-of-the-art texture feature extraction methods for melanomas detection. The results indicate the significance of contrast of the pattern along with LBP patterns.