HYPERINTENSE VESSEL SIGN DETECTION IN ACUTE ISCHEMIC STROKE

A DEEP LEARNING APPROACH USING FLAIR MRI

Abstract

The Hyperintense Vessel Sign (HVS) on FLAIR MRI is a subtle yet critical marker of large vessel occlusion (LVO) in acute ischemic stroke, with timely detection directly influencing eligibility for thrombolysis or thrombectomy. Manual HVS identification remains time-intensive and susceptible to inter-observer variability, particularly under high-pressure emergency conditions. A total of 72 patients were retrospectively recruited from Hospital Sultan Abdul Aziz Shah (HSAAS), Universiti Putra Malaysia (UPM), with FLAIR MRI acquired using a standardized protocol on a 3T scanner. A YOLO-based object detection architecture was trained to localize HVS regions using bounding box annotations in YOLO format. Performance was benchmarked against consensus annotations from three board-certified neuroradiologists as the gold standard. On the held-out test set, the model achieved a sensitivity (recall) of 0.54, precision of 0.57, F1-score of 0.51 at the optimal confidence threshold of 0.35, and an [email protected] of 0.47 ([email protected]–0.95: 0.16). The AI-based FLAIR analysis improved detection efficiency, reducing average triage decision times while maintaining diagnostic safety. Furthermore, the model showed strong potential for integration into stroke imaging triage pathways; its explainable AI (XAI) heatmaps enabled radiologists to effectively cross-verify AI-flagged HVS regions during time-critical scenarios. By combining patient-level data integrity, targeted on-the-fly augmentation, and explainability features, the proposed system offers a practical and deployable AI solution for stroke workflows in resource-limited or high-volume settings. Future work will focus on real-time clinical pipeline integration and extension to multi-modal imaging data for comprehensive stroke assessment.