Advanced real-time detection of acute ischemic stroke using YOLOv12, YOLOv11, and YOLO-NAS: a comparative study for multi-class classification
Marwa El-Geneedy, Hossam El-Din Moustafa, Hatem Khater, Seham Abd-Elsamee, Samah A. Gamel
Abstract
Acute ischemic stroke (AIS) remains a leading cause of mortality and disability worldwide, demanding diagnostic tools that are both accurate and fast for timely intervention. This study presents a comparative evaluation of three state-of-the-art object detection models-YOLOv12, YOLOv11, and YOLO-NAS-for multi-class AIS detection in magnetic resonance imaging (MRI). The dataset, comprising four categories (Normal, PD-Patient, Acute Ischemic Stroke, and Control), was preprocessed with normalization, resizing, and augmentation, then split into training (70%), validation (20%), and testing (10%). Models were trained and evaluated on identical data, with performance measured by precision, recall, mean average precision at IoU 0.5 (mAP@50), and inference speed. YOLOv11 achieved the highest mAP@50 (98.5%) and balanced precision (95.4%) and recall (96.6%), making it the most reliable across classes. YOLOv12 performed comparably (mAP@50 98.3%, precision 95.2%, recall 96.0%) with slightly slower inference, while YOLO-NAS offered the fastest speed (154 FPS) but lower precision (76.3%. Results highlight the trade-offs between detection accuracy and processing speed, providing guidance for selecting YOLO-based architectures suited to specific clinical workflows such as emergency stroke care. The real-time implementation, accessible via Roboflow, demonstrates the feasibility of deploying these models for rapid, automated AIS detection in clinical settings.