Graph weighting scheme for skin lesion segmentation in macroscopic images

[Display omitted] •High level guidance for lesion segmentation based on background and foreground priors.•Edge weighting process that combines efficiently color and texture features.•Global and local information for feature relevance evaluation.•Graph ranking for contrast evaluation.•Fusion strategy using logical operations on contrast-maps for lesion segmentation. Melanoma is the least common skin cancer but the most severe and lethal. Due to the expensive cost of medical screening, there is a need to develop automated computer-aided diagnosis (CAD) of melanoma in macroscopic images. Segmentation of skin lesions is then required as an initial processing step in these systems. In this paper, we propose a graph-based skin lesion segmentation algorithm. We start by roughly determining lesion and background templates. We then develop a graph that ranks regional similarities to both lesion and skin cues involving several features. The contribution of each feature is controlled according to its regional discriminative power. Finally, we segment the lesion using a strategy that combines the two ranking operations. A comparative evaluation is performed on both melanoma and benign lesions selected from three challenging databases. Our ranking process can suppress skin regions and highlights effectively the complete lesion. Quantitative evaluation using several metrics shows that our algorithm achieves consistently better segmentation performance compared to ten state-of-the-art methods. It yields accurate segmentation and enhances applicability on complicated images with different artifacts. Its high performance indicates that it is suitable for practical use as part of CAD systems.