Skin lesion segmentation of dermoscopy images using U-Net

Abstract

Skin cancer is one of the most threatening cancers as reported and has been on the increase over the past 10 years. The traditional methods of skin cancer segmentation are time-consuming and inefficient. U-Net is a powerful and accurate way of self-segmentation in the medical field. In order to solve this problem, this paper proposes a U-Net skin cancer segmentation system that can provide results and feedback quickly, accurately and intelligently. It is composed of two parts: Skin Image Analysis Module and Skin Image Segmentation Module. In the skin image analysis module, the system learns segmentation from the training set images, and verifies the correctness of learning from part of the images. In Skin Image Segmentation Module, the system segments all the images in the test set folder. Among several experiments, the system using GPU training and learning with 100 images of ISIC dataset after 10 epochs has the training accuracy of 0.9085, while the validation accuracy is 0.9536. The system allows users to upload their lesion images to a test folder to obtain reliable segmentation results in a timely manner, thereby improving the survival rate of potential patients.