VesselSim achieves zero-shot blood vessel segmentation without annotations

Blood vessel segmentation is critical for vascular disease care and surgical planning, but obtaining expert annotations for deep learning training remains a major bottleneck. To address this, researchers introduce VesselSim, a two-stage framework that removes the need for real annotated data. First, a stochastic, geometry-driven simulation models recursive branching, curvature-controlled growth, and collision-aware topology, combined with domain-randomized intensity synthesis to produce 16,500 anatomically plausible 3D angiographic volumes. Second, a 3D U-Net is trained exclusively on this synthetic data. To bridge the domain gap at inference time, a test-time adaptation strategy uses a self-supervised mask reconstruction decoder, enabling adaptation to unseen clinical scans without prior domain knowledge.

VesselSim was evaluated in a zero-shot setting on multiple real-world datasets spanning MR and CT across brain and kidney regions. Despite having never seen real medical images during training, its performance is competitive with state-of-the-art vascular segmentation foundation models. These findings suggest that learning vessel geometry from synthetic tubular structures generalizes robustly across domains, substantially reducing reliance on acquired medical imaging data and, critically, expert annotations. The work will be presented at MICCAI 2026 and is currently available on arXiv.