A setup for live AI support in interventional radiology

Abstract Artificial intelligence (AI) has the potential to support time-critical stroke treatment. In a previous study we demonstrated the feasibility of deep learning based automatic classification for thrombus detection during thrombectomies, a catheter-guided procedure to remove occlusions of cerebral vessels. However, this method has yet to be tested during a live intervention. In this work, we present a setup to integrate AI based support in an angiography suite. A classification PC was connected to the angiography by means of a real-time video connection as well as a research interface for control signals. We found that video conversion in real-time does not affect the classification result in comparison to offline classification of DICOM data. Analyzing 50 video streams of previous cases, the system could classify digital-subtraction angiography (DSA) sequences within 13.3 seconds on average. This processing time can further be reduced to an average of 7.9 seconds with GPU acceleration. Additionally, the system successfully classified two DSA sequences acquired during live thrombectomy, identifying the presence of thrombi in less than 5 seconds. So far, the classification result has only been displayed in the control room of the angiography suite to demonstrate feasibility. In the outlook, however, we also discuss how the result can be displayed directly on the angiography screen.