Segmentation of Pigmented Melanocytic Skin Lesions Based on Learned Dictionaries and Normalized Graph Cuts

Pigmented melanocytic skin lesion pre-screening relies on the proper segmentation of the image regions affected by the skin lesion. This paper proposes a new pigmented melanocytic skin lesion segmentation algorithm for standard camera images. It is assumed that only one skin lesion is in each input image, and also is assumed that the skin lesion is placed at (or close to) the image center. Thus, the input is, at first, shading attenuated, and then converted to a three-channel color space that enhances the discrimination between healthy and unhealthy skin regions. Afterwards, a dictionary is generated for each image, which is compact and reconstructive, and represents the image patches. This dictionary is obtained in an unsupervised manner using a modified version of the Information-Theoretic Dictionary Learning (ITDL) method, which was originally proposed as supervised dictionary learning method. Normalized Graph Cuts is used to partition the set of projected patches in two groups, resulting in a binary mask that labels the pixels as corresponding to healthy or unhealthy image regions. Our preliminary experimental results obtained on a publicly available dataset are encouraging, and suggest that the proposed pigmented melanocytic skin lesion segmentation method provides, in average, a lower segmentation error rate than comparable state-of-the-art methods proposed in the literature.