Advanced real-time detection of acute ischemic stroke using YOLOv12, YOLOv11, and YOLO-NAS: a comparative study for...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Neurology

Background:

• Acute ischemic stroke (AIS) is a major global health concern, necessitating rapid and accurate diagnostic methods.
• Magnetic resonance imaging (MRI) is crucial for AIS detection, but timely interpretation remains challenging.

Purpose of the Study:

• To comparatively evaluate the performance of YOLOv12, YOLOv11, and YOLO-NAS object detection models for multi-class AIS detection in MRI.
• To assess the trade-offs between detection accuracy and inference speed for clinical applicability.

Main Methods:

• A dataset of four MRI categories (Normal, PD-Patient, AIS, Control) was preprocessed and split into training, validation, and testing sets.
• YOLOv12, YOLOv11, and YOLO-NAS models were trained and evaluated using precision, recall, mAP@50, and inference speed metrics.

Main Results:

• YOLOv11 achieved the highest mAP@50 (98.5%) with excellent precision (95.4%) and recall (96.6%).
• YOLOv12 showed comparable results (mAP@50 98.3%) but with slightly slower inference.
• YOLO-NAS provided the fastest inference (154 FPS) but with significantly lower precision (76.3%).

Conclusions:

• YOLOv11 and YOLOv12 are highly reliable for accurate AIS detection in MRI, with YOLOv11 being the top performer.
• The choice of YOLO model depends on the clinical workflow, balancing the need for speed versus accuracy in emergency stroke care.
• Real-time deployment via platforms like Roboflow is feasible for automated AIS detection.