Topology-Preserving Image Segmentation
Overview
Topology-preserving image segmentation aims to recover not only which pixels belong to an object, but also its structural organization, such as connected vessel branches or intact road networks. Common approaches add skeleton, centerline, Betti-number, or persistent-homology objectives to neural networks, encode structure with graphs or geometric priors, and sometimes repair connectivity after prediction. Work spans retinal and coronary vessels, roads, cracks, and 3D anatomical structures, using measures such as clDice, Betti error, and connectivity scores alongside Dice and IoU. A central tension is improving structural correctness without sacrificing boundary or region accuracy, adding excessive computation, or introducing false connections—especially under domain shift and sparse annotation.
Latest papers
43 papers6 papers in the last four weeks, level with the four weeks before. 0.0% of all new papers.