Dermoscopic Lesion Segmentation via Optimal Color Channel Fusion

Death caused by various kinds of cancer is on rise and skin cancer is one of the most common one worldwide. Due to the importance of early detection, dermoscopy is adopted for visualizing skin lesions and computer-aided detection benefits from these dermoscopic images for better diagnosis results. One of the most important phase of automated skin lesion detection or classification is segmentation, but it is a very challenging task because of several artifacts existing on these images. In this paper, a new method to improve skin lesion segmentation from the existing deep network architectures is proposed, based on the fusion of various results produced by existing models on different color channels. Experimental validations demonstrate that the proposed method increases the average accuracy, on lesion segmentation in terms of Sorensen-Dice and Jaccard indices, when compared to conventional techniques.