A large-scale comparative study of YOLO-based detectors for ischemic stroke lesion localization on diffusion-weighted...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Ischemic stroke lesion detection on diffusion-weighted imaging (DWI) is difficult due to lesion characteristics like small size, faintness, irregularity, or multiplicity.
• Accurate localization of these lesions is crucial for timely diagnosis and treatment.

Purpose of the Study:

• To provide a controlled, lesion-level benchmark of recent YOLO-based object detection models for localizing ischemic stroke lesions on DWI.
• To compare the performance of various YOLO detector configurations under a unified evaluation protocol.

Main Methods:

• A private clinical dataset of 300 patients with approximately 2,200 DWI images was used, with expert-annotated bounding boxes for lesions.
• Twenty-four YOLO detector configurations (YOLOv10, YOLO11, YOLO12, YOLO13, YOLO26) were trained and evaluated using standardized transfer learning, augmentation, and patient-level partitioning.
• Performance metrics included precision, recall, mAP@50, mAP@50-95, inference time, parameter count, GFLOPs, and bootstrap confidence intervals.

Main Results:

• YOLO26x achieved the highest strict-localization accuracy (0.8287 precision, 0.6867 recall, 0.7940 mAP@50, 0.5094 mAP@50-95).
• YOLO11x showed the highest precision, while YOLO12x demonstrated the highest recall.
• YOLO26s offered a favorable accuracy-efficiency balance with comparable mAP@50-95 at a lower computational cost.
• Overlapping bootstrap confidence intervals suggest caution when interpreting small performance differences between top detectors.

Conclusions:

• The study offers a clinically relevant comparison of YOLO-based detectors for DWI ischemic stroke lesion localization.
• Detector selection should balance strict localization accuracy with computational efficiency for optimal clinical application.