Topology-Aware Correction of Retinal Vessel Graphs Using Graph Neural Networks

Abstract

Vessel graphs are commonly used to study the structure of retinal vessels, but graphs obtained from vessel segmentation often contain topology errors, including broken vessels and spurious connections. These errors arise from limitations in pixel level processing and negatively impact graph-based analysis.Our work addresses this problem by formulating topology correction as a graph-level learning task. We represent connectivity errors as missing or incorrect edges and correct them through a candidate edge prediction task. A perturbation framework is introduced to simulate realistic structural errors, enabling supervised learning across various scenarios. A graph neural network is used to encode both the geometry and the graph structure of the vessels, enabling the prediction of valid vessel connections. Experimental results show that this approach improves the structural consistency of vessel graphs and enhances the accuracy of connectivity modeling.