VesselSim: Synthetic Data for 3D Blood Vessel Segmentation
VesselSim is a dual-phase system designed for 3D segmentation of blood vessels, which does not require real annotated data for training. Initially, a stochastic, geometry-based simulation creates 16,500 anatomically realistic 3D angiographic volumes. In the subsequent phase, a 3D U-Net is trained exclusively on this synthetic dataset. To connect the gap between synthetic and real images, a self-supervised mask reconstruction decoder is employed during testing. This approach tackles the difficulty of acquiring expert vascular annotations necessary for deep learning applications in medical image analysis.
Key facts
- VesselSim is a two-stage framework for 3D blood vessel segmentation.
- It eliminates the need for real annotated data during training.
- First stage: stochastic, geometry-driven vascular simulation generates 16,500 synthetic 3D angiographic volumes.
- Simulation models recursive branching, curvature-controlled growth, and collision-aware topology.
- Domain-randomized intensity synthesis is used for realistic appearance.
- Second stage: a 3D U-Net is trained solely on synthetic data.
- Test-time adaptation uses a self-supervised mask reconstruction decoder.
- The method targets vascular disease care and surgical planning.
Entities
—