An incremental approach to pigmented skin lesion segmentation with classification refinements in uncertain regions.

Skin lesion segmentation in dermatoscopic images is difficult because there are large inter variations in shape, size, color, and texture between lesions and skin types. Hence, computational features learned from a training set of lesion images may not be applicable to other lesion images. In this paper, we propose an incremental method for lesion segmentation. It leverages the Expectation-Maximization algorithm to find an initial segmentation. A new adaptive method is proposed to define two types of segmented regions: the high-confident and the low-confident. We train a support vector machine, using computational features from the high-confident regions, to further refine segmentation and, hence, achieve improved results for the low-confident regions. Validation experiments of our proposed method are performed on 319 dermatoscopy images and we have achieved good results with precision and recall to be 0.864 and 0.875 respectively.