Dermoscopic Image Segmentation: A Comparison of Methodologies

Abstract

An accurate segmentation of pigmented lesions may improve classification results of Computer Aided Diagnosis (CAD) tools. Thus, finding a reliable segmentation methodology becomes crucial. During the past few years, many segmentation methodologies of dermoscopic images have been proposed. In this paper, a comparison between three methodologies is presented: semantic segmentation with SegNet, histogram-based segmentation via convex optimization and segmentation based on a Fully Convolutional Network (FCN). As a result of evaluating the segmentation results for 600 dermoscopic images from the Test set of ISIC-2017 database, the semantic segmentation provides a 90.12% of accuracy, followed by segmentation using histograms and Fully Convolutional Network, with 86,47% and 81,70% of accuracy, respectively.