Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

vesselFM: A Foundation Model for Universal 3D Blood Vessel Segmentation

CVPR 2025 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Zero-shot robustness and transfer · 23 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Trains vesselFM, a segmentation foundation model for 3D blood vessels, on real annotated vascular data plus domain-randomized and flow-matching-generated image-mask pairs, achieving zero-, one- and few-shot generalization.

Problem

Supervised 3D blood vessel segmentation generalizes poorly across imaging modalities and protocols, requiring per-dataset voxel annotations, and existing medical segmentation foundation models fail on vessels.

Approach

Combine a large curated real vascular dataset with domain-randomized synthetic image-mask pairs (from corrosion-cast vascular graphs) and mask- and class-conditioned flow matching samples, training one UNet to generalize zero-shot to unseen vascular domains.

Key results

• Zero-shot vesselFM reaches 74.66 Dice on SMILE-UHURA and 67.49 Dice on BvEM, outperforming tUbeNet, VISTA3D, SAM-Med3D and MedSAM-2 across all four datasets.
• On MSD8 zero-shot, vesselFM scores 29.69 Dice, 5.86 points above VISTA3D, which was trained on MSD8-containing CT data.
• Ablation: adding Ddrand and Dflow to Dreal raises zero-shot Dice on SMILE-UHURA from 65.45 to 74.66 and clDice from 63.53 to 75.27.

Measured effects reported in the paper

Surprising findings

• vesselFM zero-shot results on SMILE-UHURA surpass few-shot results achieved by baseline models in Dice and clDice.
• Med-DDPM-generated data yields lower segmentation performance than flow-matching data (70.34 vs 74.66 Dice).

Limitations

• Connectivity of predicted vessels may need tailored post-processing.
• Scope limited to binary vessel segmentation; multi-class or instance segmentation not addressed.
• Zero-shot performance on MSD8 (29.69 Dice) and one-shot results on some datasets remain modest.

Experimental setup

Datasets: SMILE-UHURA BvEM OCTA MSD8 TubeTK TopCoW VesSAP DeepVesselNet HR-Kidney 3D-IRCADb-01 DeepVess CSD

Benchmarks: SMILE-UHURA MSD8 OCTA BvEM

Metrics: Dice clDice

Baselines: tUbeNet VISTA3D SAM-Med3D MedSAM-2

Builds on: Med-DDPM SynthSeg flow matching nnU-Net clDice

Closest papers by meaning

• VasTSD: Learning 3D Vascular Tree-state Space Diffusion Model for Angiography Synthesis (CVPR 2025 · similarity 0.87)
• DeNVeR: Deformable Neural Vessel Representations for Unsupervised Video Vessel Segmentation (CVPR 2025 · similarity 0.87)
• Learning Dynamic Collaborative Network for Semi-supervised 3D Vessel Segmentation (CVPR 2025 · similarity 0.87)
• VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging (CVPR 2025 · similarity 0.86)
• DeformCL: Learning Deformable Centerline Representation for Vessel Extraction in 3D Medical Image (CVPR 2025 · similarity 0.85)
• Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It (NeurIPS 2026 · similarity 0.84)