Learning to Reconnect: Graph-based Path Classification for Restoring Retinal Vessel Segmentation Connectivity

Accurate segmentation of retinal blood vessels in 2D fundus images is essential for reconstructing complete vascular trees and enabling downstream analyses that require a reliable vascular topology. However, most deep learning approaches rely on pixel-wise segmentation, which can lead to vessel discontinuities, particularly in thin or low-contrast regions. These fragmentation errors can affect the structural integrity of the reconstructed vascular network and impact subsequent modeling tasks. Tracking-based methods and automated post-processing approaches have been proposed to reconnect fragmented segments, but they often rely on accurately identifying beforehand which points to reconnect along the vascular skeleton. In practice, determining which endpoints should be connected remains a challenging problem and typically requires manual intervention. To address this limitation, we introduce an automated framework designed to identify vessel point pairs that should be reconnected. Instead of relying on handcrafted geometric criteria, we formulate reconnection as a path classification problem: given a candidate path between two vessel points in a binary mask, the model determines whether a direct vessel segment exists between them. By encoding complete paths as single entities, our approach leverages image information along the entire potential connection. Experimental results on retinal images from the FIVES dataset show that our approach effectively identifies disconnected segments in predicted vascular masks, substantially outperforming geometric baselines and providing a reliable guidance for downstream reconnection methods. Code is publicly available at https://github.com/oscarmorand/EVAPORE/tree/shapeMI.

accepted for Shape in Medical Imaging (ShapeMI) @ Medical Image Computing and Computer Assisted Intervention (MICCAI)