Vessel segmentation is a fundamental task in medical image analysis, supporting diagnosis and treatment planning. The complex morphology of vessels and the variability of imaging conditions make it difficult to build a universal segmentor that works across anatomies and modalities. Existing deep learning approaches are usually tailored to specific datasets and lack generalization. Although the Segment Anything Model (SAM) has shown strong generic segmentation ability, its original design does not exploit vascular topology and struggles with fine‑scale vessels, resulting in suboptimal performance.
To address these issues, we introduce ReG‑SAM, a SAM‑based framework specifically designed for 2D vessel segmentation. The key idea is to employ a reference graph set to enrich vascular representations. From reference masks we derive two modality‑aware embeddings: Graph Prompt Embeddings (GPE) that encode global spatial information of the graphs, and Vascular Prototype Embeddings (VPE) that capture fine‑grained, modality‑specific vessel characteristics using multi‑scale feature maps and vascular masks.
Because ground‑truth vessel masks are unavailable at inference time, we construct a modality‑wise vascular database and propose two reference‑graph‑guided representation learning schemes. These schemes estimate GPE and VPE from database samples instead of relying on true masks. Extensive experiments on 19 public datasets demonstrate that ReG‑SAM consistently outperforms all baselines, including those using manual prompts, especially on thin‑vessel segmentation.
Review: ReG‑SAM leverages dual embeddings driven by reference graphs to fill the gap left by SAM in modeling vascular details. Its cross‑modality and cross‑anatomy robustness makes it a promising, scalable solution for medical vessel segmentation.