Deep learning in abdominopelvic digital subtraction angiography: a systematic review of interventional radiology applications

Abstract

Purpose

Deep learning (DL) is increasingly explored in interventional radiology (IR) applications. This systematic review evaluates current DL-based applications for digital subtraction angiography (DSA) in abdominopelvic interventions, summarizes performance, and identifies gaps in the literature.

Materials and methods

Following PRISMA guidelines, we searched MEDLINE, Scopus, and Google Scholar for studies published up to February 1, 2025. English-language original articles assessing DL methods for automatic DSA image analysis were included, and study quality was evaluated with QUADAS-2.

Results

Nine studies were included. Two focused on hemorrhage detection, in which area under the curve (AUC) values ranged between 0.80–0.85. Four examined image enhancement, one performed vessel segmentation, and one applied classification of the anatomic location. Only a single study evaluated treatment response prediction, with an accuracy of 0.75. Most models were tested on small datasets from single centers, limiting their generalizability.

Conclusion

Preliminary studies show that DL can improve hemorrhage detection, image quality, and vessel segmentation in DSA. However, larger, prospectively validated datasets are warranted. Currently no FDA-approved DL tools exist for abdominal or pelvic DSA. Future efforts should explore advanced generative AI and multimodal approaches.