Multi-scale Fully Convolutional DenseNets for Automated Skin Lesion Segmentation in Dermoscopy Images

Abstract

This paper addresses the problem of automated skin lesion segmentation in dermoscopy images. We propose a novel Multi-Scale Fully Convolutional DenseNets (MSFCDN) for skin lesion segmentation. The MSFCDN adopts fully convolutional architecture, which after training, can perform semantic segmentation of an image with arbitrary size. We conduct extensive experiments on ISBI 2017 “Skin Lesion Analysis Towards Melanoma Detection” Challenge dataset. Our method achieves an average Dice coefficient of 86.9% and an average accuracy of 95.3% for skin lesion segmentation.