Abstract
3D blood vessel segmentation is critical for medical diagnosis but suffers from significant domain gaps across imaging modalities. We present vesselFM, a foundation model designed for universal vascular segmentation. By training on a diverse mixture of real clinical data, domain-randomized synthetic volumes, and high-fidelity generative samples produced via flow matching, vesselFM achieves robust generalization. Our results demonstrate state-of-the-art performance in zero-shot, one-shot, and few-shot scenarios across modalities including MRI, CT, and electron microscopy, significantly outperforming existing specialized and general-purpose medical foundation models.
Introduction
Segmenting intricate 3D vascular structures is a fundamental task in medical imaging, essential for managing conditions such as stroke, aneurysms, and neurodegenerative diseases. However, the high variability in imaging modalities—ranging from macro-scale CT and MRI to micro-scale electron microscopy—presents a daunting "domain gap." Each modality introduces unique artifacts, noise profiles, and spatial resolutions, making it difficult for models trained on one dataset to generalize to others.
While general-purpose foundation models like the Segment Anything Model (SAM) have shown promise, they often fail to capture the thin, connected, and complex topology of blood vessels. vesselFM addresses this by learning a universal geometric representation of vasculature through massive data diversity.
Methodology: The Power of Data Diversity
The core innovation of vesselFM lies in its training strategy, which combines three distinct data regimes to build a robust "tubular inductive bias."
1. Curated Real Data ($D_{real}$)
The model is trained on the largest compiled collection of annotated 3D vessel datasets to date, encompassing 23 datasets from various organs (brain, liver, kidney) and species (human, rat, mouse). This provides the necessary anatomical ground truth.
2. Domain Randomization ($D_{drand}$)
To prevent overfitting to specific scanner characteristics, a synthetic pipeline generates vascular "skeletons" wrapped in randomized textures, intensities, and common imaging artifacts (e.g., bias fields, Gibbs noise, and dropout).
3. Generative Flow Matching ($D_{flow}$)
High-fidelity synthetic data is generated using a mask-conditioned Flow Matching model. Unlike traditional diffusion models, flow matching utilizes straighter probability paths, resulting in synthetic images that closely mimic the complex textures of real clinical scans.
Key Results and Discussion
vesselFM demonstrates remarkable zero-shot generalization capabilities, often outperforming models specifically fine-tuned for a target modality.
- Zero-Shot Performance: On human liver CT (MSD8), vesselFM achieved a Dice score of 29.69, surpassing VISTA3D (23.83), a model trained specifically on CT data.
- Cross-Modality Robustness: In the challenging domain of mouse brain electron microscopy (BvEM), vesselFM reached a Dice score of 67.49 in a zero-shot setting, whereas MedSAM-2 only achieved 10.92.
- Efficiency: With 31.4M parameters, the model is computationally efficient, processing $128^3$ volumes in approximately 30.6 ms on an NVIDIA A100 GPU.
A notable emergent property is the model's strong "tubular bias." It occasionally segments non-vascular tube-like structures (e.g., axons or ribs), indicating it has learned a fundamental geometric primitive for "tubes" rather than just memorizing vessel textures.
Conclusion
vesselFM represents a significant step toward a universal tool for vascular analysis. By integrating real, randomized, and generative data, it overcomes the domain gap that has long hindered 3D vessel segmentation. Future work will focus on improving graph-level connectivity and extending the model to instance segmentation for individual branch analysis.
Summary updated