Automatic Multiple Regions Segmentation of Dermoscopy Images

Skin lesion segmentation is one of the most important steps in automated early skin cancer detection, since the accuracy of the following steps significantly depends on it. In this paper, a two-stage approach based on Mean Shift and spectral graph partitioning algorithms is proposed. This method effectively extracts lesion borders. Moreover, a distinctive advantage of this approach is extracting the region of interest levels that is not addressed in pervious state of the art methods. In the first stage, the image is segmented to regions using Mean Shift algorithm. In the second stage, a graph-based representation is used to demonstrate the structure of the extracted regions and their relationships. Afterwards a clustering process is applied, considering the neighborhood system and analyzing the color and texture distance between regions. The proposed method is applied to 170 dermoscopic images and evaluated with two different metrics. This evaluation has performed by means of the segmentation results provided by an experienced dermatologist as the ground truth. Experiments demonstrate that in this method, challenging features of skin lesions are handled as might be expected when compared to five state of the art methods.