Automated lesion border detection of dermoscopy images using spectral clustering

Skin lesion segmentation is one of the most important steps for automated early skin cancer detection since the accuracy of the following steps significantly depends on it. In this paper we present a novel approach based on spectral clustering that provides accurate and effective segmentation for dermoscopy images. In the proposed method, an optimized clustering algorithm has been provided which effectively extracts lesion borders using spectral graph partitioning algorithm in an appropriate color space, considering special characteristics of dermoscopy images. The proposed segmentation method has been applied to 170 dermoscopic images and evaluated with two metrics, by means of the segmentation results provided by an experienced dermatologist as the ground truth. The experiment results of this approach demonstrate that, complex contours are distinguished correctly while challenging features of skin lesions such as topological changes, weak or false contours, and asymmetry in color and shape are handled as might be expected when compared to four state of the art methods.