Weakly and semi-supervised deep level set network for automated skin lesion segmentation

Abstract

In this paper, we proposed an end-to-end deep convolutional neural model to implement weakly and semi-supervised learning in order to resolve insufficient training data with pixel-wised annotation. Entire model consists of two branches. At first, the segmentor is trained by small amount of training data with high-level annotation serving certain ability of semantic segmentation. In addition, deep level set and conditional random field branches are responsible for converting image-level annotation to pixels, which provide sufficient data to retrain the segmentor holding excellent capability for lesion segmentation. Finally, experiments benchmarked our proposed method with the state-of-the-art models demonstrating superior performance and generalization.