Real-Time AI Implementation in Middle Meningeal Artery Embolization for Subdural Hematoma

INTRODUCTION/PURPOSE Real‐time artificial intelligence (AI) embolization tracking may improve operator awareness and procedural safety during neurointerventions. Its application in middle meningeal artery embolization (MMAE) for subdural hematoma (SDH) has not been systematically evaluated. We aimed to implement a standardized protocol for Neuro‐Vascular Assist in MMAE and assess its accuracy and potential clinical impact. MATERIALS AND METHODS Fourteen MMAE procedures were performed using AI integration with a biplane angiographic system. Neuro‐Vascular Assist tracked the entrance of the liquid embolic material in manually set regions of interest (ROI). ROIs were placed according to a predefined protocol: the reflux region, the ipsilateral orbit, and the superior sagittal sinus. We classified AI notifications as true positives (TP), false positives (FP), and false negatives (FN). TPs were assessed for clinical relevance, which was defined as injections halted within 5 seconds of a TP notification. Accuracy was measured by precision [TP/(TP+FP)] and recall [TP/(TP+FN)]. In addition, adjusted recall was calculated by reclassifying “late notified FN” as TP. These were alerts that appeared after the operator had already stopped the injection, indicating that the AI correctly identified the event, but with slower detection compared to the physician. Results Of 14 procedures, 2 were excluded (one due to coil embolization and one with MMA originating from the ophthalmic artery), leaving 12 evaluable cases (six on the right side). The ROI placement protocol was successfully implemented in all evaluable procedures without workflow disruption. A total of 56 TPs, 4 FPs, and 26 FNs were recorded, yielding an overall precision of 93.3%, a recall of 68.3%, and an adjusted recall of 82.9%. TP alerts were clinically relevant in 64.3% of cases, representing 60% of all AI notifications. Conclusion Neuro‐Vascular Assist demonstrated promising preliminary results analyzing MMAE in SDH with high precision, improved adjusted recall, and meaningful clinical relevance metrics. The feasibility of the proposed protocol was confirmed, supporting further investigation of its potential role in enhancing procedural safety and clinical outcomes.