Multi-scale contrast based skin lesion segmentation in digital images

Segmentation is the first step in computer-aided prescreening of melanoma diagnosis. It is therefore important to perform an accurate segmentation to delineate efficiently the lesion borders. This paper proposes a contrast based algorithm for segmenting skin lesions in digital photographs. After a decomposition of the image into superpixels, we introduce a regional similarity which considers histogram based difference supported by occurrence probability calculation and spatial coherence. This metric is then used to compute the regional background connectivity as the overlapping rate with image borders which serves to discard efficiently background regions and estimate the regional contrast. Thereafter, we improve the contrast accuracy through a feature weighting scheme that controls the contribution of each feature according to its relevance. To deal with the scale variation of lesions and some areas in the image, we define a hierarchical inference process that optimizes the lesion/background distinctiveness. Our algorithm is simple, multi-scale structured and resistant to image artifacts. Experiments on two standard datasets show that our approach yields accurate lesion segmentation and outperforms some recent state-of-the-art methods that provide high segmentation rates.