Segmentation of dermoscopy image using adversarial networks

Abstract

Skin lesion image segmentation task has many difficulties due to the hair occlusion or low contrast. Traditional methods based image quality are powerless when processing these kinds of skin images. In this paper, a segmentation architecture based adversarial networks is proposed. The architecture is consist of a segmentation network based U-net and a discrimination network linked by certain convolutional layers. Both two networks are trained alternately and at last segmentation network gets a high segmentation accuracy. We test our architecture on dataset PH2 and dataset obtained from “Skin Lesion Analysis Toward Melanoma Detection” challenge which was hosted by ISBI 2016 conference and we achieved segmentation average accuracy of 0.97, dice coefficient of 0.94 which outperform other existed segmentation network, including winner of ISBI 2016 challenge for skin melanoma segmentation.