Vascular Topology Rectification Network for Automated Retinal Artery/Vein Segmentation

Retinal artery/vein (A/V) segmentation is an important step in screening various diseases. However, accurate retinal A/V segmentation is still a challenging task, and even state-of-the-art methods often fail to respect the correct topology of the vascular structure, which is crucial for clinical application. For better preservation of topology and connectivity in fully automated A/V segmentation, we propose a two-stage deep learning framework that consists of a multi-task U-Net with a customized vessel aggregation module for initial A/V segmentation, and a vascular topology rectification network based on persistent homology. Extensive experiments on public DRIVE-AV and HRF-AV datasets show that the proposed method delivers competitive pixel-level accuracy and the best topology-level accuracy against state-of-the-art methods.