VesselSim trains a 3D segmentation model on synthetic vessels only, matching vascular foundation models zero-shot on real brain and kidney MR/CT

A stochastic, geometry-driven vascular simulation framework models recursive branching, curvature-controlled growth, and collision-aware topology, followed by domain-randomized intensity synthesis, generating 16,500 anatomically plausible 3D angiographic volumes. Where prior deep learning for vessel segmentation relies on real images with expert vascular annotations, this work shifts the training data source to purely synthetic volumes whose generation explicitly encodes vascular geometry and topology rather than simple image augmentation. The abstract specifies the three geometric elements of the simulation pipeline and the domain-randomized intensity synthesis, and states the generated volume count of 16,500; no distribution-comparison metrics against real data are given in the abstract.

A 3D U-Net is trained only on synthetic data, and a test-time adaptation strategy based on a self-supervised mask reconstruction decoder is introduced at inference to bridge the synthetic-to-real domain gap. The framework requires no real annotated data during training and uses test-time adaptation to fit unseen clinical scans without prior domain knowledge, differing from conventional transfer approaches that need target-domain data or labels. The abstract states that training uses synthetic data only and that adaptation happens at inference without prior domain knowledge; specific losses or hyperparameters of the adaptation module are not given.

Evaluated zero-shot on multiple real datasets spanning MR and CT and several anatomical regions including the brain and kidneys, performance is competitive with state-of-the-art vascular segmentation foundation models. The results suggest that learning vessel geometry from synthetic tubular structures yields robust cross-domain generalization, substantially reducing reliance on acquired medical imaging data and, more importantly, expert annotations. The abstract reports zero-shot evaluation across multiple datasets, modalities, and anatomical regions and concludes competitiveness with state-of-the-art foundation models; dataset names and quantitative metric values are not listed in the abstract.

The work targets medical image analysis and surgical planning scenarios that need 3D blood vessel segmentation, and applies where real annotated data are unavailable during training but test-time adaptation to unseen clinical scans is possible at inference; its evaluation spans MR and CT and several anatomical regions including the brain and kidneys, indicating a design goal of cross-modality, cross-site universal vessel segmentation rather than a single organ or imaging protocol. For readers, this suggests that in annotation-scarce vascular research, one can consider building training sets from synthetic geometric data and then aligning to real distributions via inference-time adaptation.

The abstract does not give specific dataset names, evaluation metric values, or statistical comparison details, so the strength of 'competitive with state-of-the-art foundation models' can only be understood from its zero-shot, cross-modality, cross-region evaluation setting; the concrete parameters of recursive branching, curvature-controlled growth, and collision-aware topology in the synthetic simulation, the ranges of domain randomization, and the implementation details of the self-supervised mask reconstruction decoder are not expanded in the abstract; additionally, the computational cost of test-time adaptation in clinical deployment and its sensitivity to input scan quality are questions readers may continue to watch.