Topology-aware Unsupervised Domain Adaptation for Curvilinear Structure Segmentation

My presentation will focus on TopoUDA, a framework I developed to improve how segmentation models generalize to new, unlabeled clinical data. Rather than just adjusting for visual appearance differences, TopoUDA explicitly tackles topological mismatches, like broken connectivity and branching in vascular networks, which often lead to fragmented results. I will walk through how the framework integrates topological reasoning at multiple stages using a Curvilinear-Aware Discriminator, a Hierarchical Topological Refiner based on persistent homology, and a Hybrid Trust Mechanism, ultimately demonstrating its strong performance on real-world clinical datasets like retinal imaging and OCTA.