Semi-ECNet: Edge-Consistency Based Semi-Supervised Retinal Vessel Segmentation Network

Blood vessel segmentation plays an important role in the diagnosis and treatment of retinal diseases. The performance of supervised deep-learning-based segmentation methods is dependent on the training labels, which brings a great burden to surgeons. Semi-supervised methods can solve the problem partly, but recently proposed algorithms hardly consider the complexity of the tree structures in retinal images, especially fine peripheral bronchi. Thus, we propose a novel edge-consistency based semi-supervised retinal vessel segmentation algorithm, named Semi-ECNet. Specifically, Semi-ECNet first generates two kinds of vessel maps, including an edge constraint map and a pixel-wise probability map in the model-prediction stage. Then for the loss-consistency stage, we adopt the Sobel operator and propose a novel loss strategy for the consistency constraints among these maps and the ground truth. Extensive experiments on a publicly available dataset demonstrate that our Semi-ECNet effectively leverages unlabeled data, and outperforms other state-of-the-art semi-supervised segmentation methods by introducing this innovative edge-consistency strategy.