VesselSim: learning 3D blood vessel segmentation without expert annotations

Researchers introduce VesselSim, a framework that trains 3D blood vessel segmentation models entirely on synthetic, unannotated data rather than requiring expert-labeled medical images. The system combines geometric vascular simulation with domain adaptation techniques to achieve competitive performance with state-of-the-art models on real clinical scans across multiple imaging modalities and anatomical regions.

VesselSim addresses a fundamental bottleneck in medical AI development: the scarcity and expense of expert-annotated training data. Traditional deep learning approaches for blood vessel segmentation require thousands of manually labeled medical images, a resource-intensive process that slows research progress and limits model availability in resource-constrained settings. This work demonstrates that synthetic data generation, combined with intelligent adaptation strategies, can substantially reduce this dependency.

The framework operates through two complementary innovations. First, a stochastic simulation engine generates anatomically plausible synthetic vessel networks by modeling realistic biological properties like recursive branching, curvature-controlled growth, and collision-aware topology. This produces 16,500 diverse training examples without manual annotation. Second, the model incorporates test-time adaptation via self-supervised learning, allowing it to adjust to real clinical images at inference without requiring domain knowledge or fine-tuning datasets.

The implications extend beyond vessel segmentation. This approach validates a broader principle: synthetic data paired with robust adaptation strategies can enable zero-shot generalization across diverse real-world medical imaging scenarios. For healthcare institutions and researchers with limited annotation budgets, this reduces barriers to deploying specialized segmentation models. The technique's effectiveness across different anatomical regions and imaging modalities (MR, CT) suggests applicability to other segmentation tasks in medical imaging.

Looking forward, the accessibility benefits could accelerate medical AI deployment in underserved regions while reducing the annotation burden on radiologists. The framework also offers potential for rapid model customization to institution-specific imaging equipment and protocols through efficient adaptation mechanisms.

• →VesselSim eliminates the need for expert-annotated data by training entirely on synthetically generated 3D vessel structures
• →Test-time adaptation via self-supervised learning bridges the domain gap between synthetic training and real clinical images
• →The model achieves competitive performance with state-of-the-art foundation models despite zero prior exposure to real clinical data
• →Framework generalizes across multiple imaging modalities (MR, CT) and anatomical regions, reducing annotation burden in medical AI
• →Approach validates synthetic data generation as a viable strategy for overcoming data scarcity in specialized medical imaging tasks