AI Assistance for Detecting Active Bleeding During DSA Procedures

Résumé succinct

Clinically, DSA-VDet can continuously indicate suspected bleeding regions throughout a complete angiographic sequence and provide a second-reader opinion to the operator. It is expected to reduce the time required for frame-by-frame review and localization of culprit vessels, reduce missed detection of small, multiple, or unusually located bleeding sites, and improve consistency among physicians with different levels of experience. Faster culprit-vessel localization may also reduce repeat angiography, radiation exposure, and unnecessary superselective catheterization. These potential benefits will be tested in the prospective randomized trial; the study does not assume that the model will necessarily improve clinical outcomes. From a translational perspective, the study will establish a standardized temporal DSA database spanning multiple centers, devices, and anatomical regions, together with bleeding-site annotation standards, a model suitable for real-time deployment, and a pathway for clinical evaluation. These resources may provide a reproducible technical foundation for regional collaboration in emergency interventional care and AI-assisted interpretation in lower-resource hospitals. The core value of this study is therefore to address the 'single-region, weak-temporal, algorithm-centered, clinically under-evaluated' limitations of existing models. Using multicenter DSA videos from multiple anatomical regions, the study will leverage contrast kinetics and motion information to improve detection of small bleeding sites and will progress from retrospective validation through controlled physician interpretation to a prospective randomized trial, providing a continuous evaluation from technical performance to clinical benefit.