Circle of Willis-Guided Localization for Simultaneous Detection and Classification of Large Vessel Occlusions in Brain CTA

Abstract Large vessel occlusions (LVOs) are blockages in the brain’s major arteries that can cause severe neurological damage. Rapid and accurate detection using computed tomography angiography (CTA) is critical for timely stroke treatment. Here, we present a fully automated approach that detects LVOs and classifies the affected vessel simultaneously. Our method incorporates a spatial prior by using Circle of Willis (CoW) segmentation as additional input, guiding the model to anatomically relevant regions. We evaluated the two strategies, the global approach using the full CTA volume, and the local one focused on CoW regions. Both achieved high performance. Detection sensitivity was 0.97 at 0.20 false positives per image for the global approach, and 0.97 at 0.13 false positives for the local approach. Classification accuracy reached 94% and 91% for global and local strategies, respectively. Importantly, the local approach was 3.3 $$\\times $$ faster, offering a computationally efficient solution, a critical advantage in acute stroke care, where every minute impacts patient outcomes.