Researchers have built a version of the Segment Anything Model that's actually good at finding blood vessels.
The new system, called ReG-SAM, adapts SAM for 2D vessel segmentation in medical images by pulling in a reference graph database of vessel shapes instead of relying on manual prompts or ground-truth masks at inference time. It generates two kinds of embeddings: graph prompt embeddings that capture overall spatial layout, and vascular prototype embeddings that capture fine, modality-specific vessel detail. Because real masks aren't available when the model is actually used, the researchers built a database of vessels sorted by imaging modality and trained the model to estimate those embeddings from database samples instead. The team tested it across 19 datasets spanning different anatomies and imaging types, and it consistently beat existing baselines, including ones that get manual prompts as a head start.
Vessel segmentation is a workhorse task in medicine, feeding into everything from stroke diagnosis to surgical planning, but most tools are built for one modality and fall apart on thin, hard-to-trace vessels. A model that generalizes across anatomies and imaging types without needing manual clicks at inference time would cut a real bottleneck in clinical image analysis.
It's a research paper tested on benchmark datasets, not a hospital-ready product, so the real test is whether it holds up outside the paper's own evaluation setup.