Semi-supervised Skin Lesion Segmentation via Transformation Consistent Self-ensembling Model

• DOI: 10.48550/arxiv.1808.03887
• Publication Date: 2018-01-01
• Authors: Xiaomeng Li; Lequan Yu; Hao Chen; Chi‐Wing Fu; Pheng‐Ann Heng
• Keywords: FOS: Computer and information sciences; 03 medical and health sciences; 0302 clinical medicine; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition

Automatic skin lesion segmentation on dermoscopic images is an essential component in computer-aided diagnosis of melanoma. Recently, many fully supervised deep learning based methods have been proposed for automatic segmentation. However, these approaches require massive pixel-wise annotation from experienced dermatologists, which very costly and time-consuming. In this paper, we present a novel semi-supervised method by leveraging both labeled unlabeled data. The network optimized the weighted combination common loss inputs only regularization segmentation, where Our encourages consistent prediction using outputs network-in-training under different regularizations, so that it can utilize To data, our predictions same input regularizations. Aiming problem, enhance effect pixel-level introducing transformation, including rotation flipping, scheme self-ensembling model. With 300 training samples, sets new record benchmark International Skin Imaging Collaboration (ISIC) 2017 challenge. Such result clearly surpasses fully-supervised state-of-the-arts are trained with 2000