A comparative study of algorithms for automatic segmentation of dermoscopic images

Melanoma is the most common as well as the most dangerous type of skin cancer. Nevertheless, it can be effectively treated if detected early. Dermoscopy is one of the major non-invasive imaging techniques for the diagnosis of skin lesions. The computer-aided diagnosis based on the processing of dermoscopic images aims to reduce the subjectivity and time-consuming analysis related to traditional diagnosis. The first step of automatic diagnosis is image segmentation. In this project, the implementation and evaluation of several methods were proposed for the automatic segmentation of lesion regions in dermoscopic images, along with the corresponding implemented phases for image preprocessing and postprocessing. The developed algorithms include methods based on different state of the art techniques. The main groups of techniques which have been selected to be studied and implemented are thresholding-based methods, region-based methods, segmentation based on deformable models, as well as a new proposed approach based on the bag-of-words model. The implemented methods incorporate modifications for a better adaptation to features associated with dermoscopic images. Each implemented method was applied to a database constituted by 724 dermoscopic images. The output of the automatic segmentation procedure for each image was compared with the corresponding manual segmentation in order to evaluate the performance. The comparison between algorithms was carried out regarding the obtained evaluation metrics. The best results were achieved by the combination of region-based segmentation based on the multi-region adaptation of the k-means algorithm and the subIngeniería de Sistemas Audiovisuale