Original title: UI-VISA: U-Net Initialized Vascular Image Segmentation Architecture
Asees Kaur, Suzanne S. Sindi, Erica M. Rutter · September 2026
Mapping blood vessels in medical X-ray scans is hard because vessels are thin, long, and branching. Common AI models, like U-NetA neural network model used for image segmentation that labels each pixel to find objects like vessels., look at pixels individually. Because they do not track how vessels connect, they often leave broken gaps in narrow branches. Another method, region growingAn algorithm that starts from seed points and expands to neighboring pixels to form connected regions., connects neighboring pixels well, but it needs good starting points to work properly. To solve this, researchers built UI-VISA. Their system lets U-Net take a first pass at spotting vessel pixels. It then feeds those predictions as starting points into a guided region growing step. This second step traces outward along connected paths, filling in missed details and cleaning up errors. In tests, UI-VISA preserved vessel connectivity significantly better than standalone U-Net and an older region-growing tool, keeping the delicate pathways intact.
You're drawing a detailed map of all the streets in a neighborhood. If you sketch each street separately, they might not connect properly where they should intersect. But if you first mark key intersections and then draw lines connecting them along the actual roads, the map stays whole. In this paper, a similar approach is used to map blood vessels: an AI method finds starting points, and then a process connects them to ensure no breaks.
For doctors studying blood vessels, accurate and connected maps are vital for surgeries and treatments. This approach reduces gaps in vessel outlines, which could lower the risk of errors in patient care.
The method was tested on only 26 DSA images, so it may not generalize to other types of medical images. The improvement in overall segmentation accuracy (Dice score) was not statistically significant, meaning the primary benefit is in connectivity rather than general overlap. It also depends on U-Net's initial predictions, which could limit performance if U-Net fails.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
UI-VISA: U-Net Initialized Vascular Image Segmentation Architecture — Asees Kaur, Suzanne S. Sindi, Erica M. Rutter
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.