Semi-supervised Retinal Vessel Segmentation Through Point Consistency

Retinal vessels usually serve as biomarkers for early diagnosis and treatment of ophthalmic and systemic diseases. However, collecting and labeling these clinical images require extensive costs and thus existing models are commonly based on extremely limited labeled data for supervised segmentation of retinal vessels, which may hinder the effectiveness of deep learning methods. In this paper, we propose a novel point consistency-based semi-supervised (PCS) framework for retinal vessel segmentation, which can be trained both on annotated and unannotated fundus images. It consists of two modules, one of which is the segmentation module predicting the pixel-wise vessel segmentation map like a common segmentation network. Otherwise, considering that retinal vessels present tubular structures and hence the point set representation enjoys its prediction flexibility and consistency, a point consistency (PC) module is introduced to learn and express vessel skeleton structure adaptively. It inputs high-level features from the segmentation module and produces the point set representation of vessels simultaneously, facilitating supervised segmentation. Meanwhile, we design a consistency regularization between point set predictions and directly predicted segmentation results to explore the inherent segmentation perturbation of the point consistency, contributing to semi-supervised learning. We validate our method on a typical public dataset DRIVE and provide a new large-scale dataset (TR160, including 160 labeled and 120 unlabeled images) for both supervised and semi-supervised learning. Extensive experiments demonstrate that our method is superior to the state-of-the-art methods.