VesselSim: learning 3D blood vessel segmentation without expert annotations

Erin Rainville1, Melissa Ananian1, Tristan Mirolla1, Hassan Rivaz2, Yiming Xiao 1
1Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada, 2Department of Electrical and Computer Engineering, Concordia University, Montreal, Canada
GitHub Link
Project teaser figure

VesselSim enables zero-shot blood vessel segmentation across modalities and organs without any real training images. The full generated data is available for use here, the generation code is available here and the manuscript is available here.

Abstract

Blood vessel segmentation is a core task in medical image analysis for the care of vascular diseases and surgical planning, yet the challenges of providing expert vascular annotations pose a major obstacle for the progress of related deep learning techniques. To address this, we propose VesselSim, a two-stage framework for universal 3D blood vessel segmentation that eliminates the need for real annotated data during training. First, we introduce a stochastic, geometry-driven vascular simulation framework that models recursive branching, curvature-controlled growth, and collision-aware topology, followed by domain-randomized intensity synthesis to generate 16,500 anatomically plausible 3D angiographic volumes. Second, a 3D U-Net is trained solely on this synthetic data. To bridge the domain gap from synthetic to real images at inference time, we introduce a test-time adaptation strategy via a self-supervised mask reconstruction decoder, enabling adaptation to unseen clinical scans without prior domain knowledge. We evaluate VesselSim in a zero-shot setting on multiple real-world datasets spanning MR and CT across several anatomical regions, including the brain and kidneys. Despite being trained exclusively on synthetic data, VesselSim achieves performance competitive with state-of-the-art vascular segmentation foundation models. These findings suggest that learning vessel geometry from synthetic tubular structures is effective for robust cross-domain generalization, substantially reducing the reliance on acquired medical imaging data and more importantly, expert annotations.

Method Overview

  • Synthetic Vascular Training Data: A stochastic vessel generation framework produces diverse 3D vascular networks and domain-randomized angiographic images without requiring real annotated data.
  • 3D U-Net Vessel Segmentation: A 3D U-Net is trained entirely on synthetic vascular volumes using Dice, Cross-Entropy, and centerline-aware losses to preserve thin vessel structures and connectivity.
  • Self-Supervised Reconstruction Branch: A masked in-painting decoder reconstructs randomly corrupted image regions, providing auxiliary self-supervision and improving domain robustness.
  • Test-Time Adaptation (TTA): During inference, reconstruction-based self-supervision adapts encoder normalization parameters to unseen clinical scans without requiring annotations.
  • Cross-Modality Generalization: The framework is evaluated in a zero-shot setting across CTA, MRA, and HiP-CT datasets spanning different anatomies and imaging domains.
Method overview figure

Results

We evaluate on two publicly available medical image repositories, HiP-CT and TopCoW. HiP-CT consists of three ultra–high-resolution 3D human kidney volumes acquired using Hierarchical Phase-Contrast Tomography. TopCoW includes 125 computed tomography angiographies (CTA) and 125 magnetic resonance angiographies (MRA) of the human brain. For TopCow CTA, manual annotations are provided for the Circle of Willis. Volumes are cropped to the bounding box of this region to avoid penalizing predictions outside the territory. For TopCoW MRA , we use the VesselVerse extension, which expands the vessel annotations to the full brain. Thus allowing us to assess robustness across organ systems (kidney, Circle of Willis, full brain) and across imaging contrasts (HiP-CT, CTA and MRA).

Comparison of segmentation performance across datasets. VesselSim denotes our final model with CE + Dice + cbDice + Reconstruction TTA. Best results are shown in bold, second best results are underlined.

Model TopCoW CTA TopCoW MRA HiP-CT
Dice clDice Dice clDice Dice clDice
Foundation Model Baselines
VesselFM 52.9±9.5 58.9±10.2 48.3±8.1 46.9±9.1 35.1±24.5 25.6±19.5
SAM-Med3D 6.6±4.4 9.1±6.0 3.1±2.6 2.6±2.2 8.0±21.6 3.1±8.9
UniverSeg 19.7±8.6 29.4±12.3 48.2±5.1 47.0±6.1 4.7±21.1 0.0±0.0
Ablation Studies
Base UNet 42.7±12.2 60.6±12.7 70.1±3.9 73.1±5.1 34.3±23.0 23.0±42.7
+ cbDice 46.8±12.9 62.9±13.6 72.6±3.5 75.7±4.4 31.1±23.7 41.2±25.1
+ recon 43.1±14.1 60.6±13.3 71.9±3.9 74.6±5.2 36.0±22.7 41.3±23.1
VesselSim 48.7±11.9 64.0±13.0 71.8±3.3 76.2±4.1 31.6±21.5 42.3±23.8
+ finetuned 59.5±7.7 70.4±7.3 77.3±3.3 83.2±3.6 46.8±33.2 50.7±27.8

Segmentation Visualization

Maximum intensity projections of the input scan and different model predictions. From top to bottom: HiP-CT, TopCoW CTA, TopCoW MRA. For the predictions, true positives are shown in blue, false positives in red, and false negatives in green. SAM-Med3D visualizations are omitted due to consistently low segmentation performance across datasets.

Results figure

BibTeX

The VesselSim paper has been accepted into the MICCAI 2026 conference. The peer reviewed, finalized copy will be available in October 2026. The unofficial version is currently available on arXiv. 
 @misc{rainville2026vesselsim,
      title={VesselSim: learning 3D blood vessel segmentation without expert annotations},
      author={Erin Rainville and Melissa Ananian and Tristan Mirolla and Hassan Rivaz and Yiming Xiao},
      year={2026},
      eprint={2605.26277},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.26277},
}