A large-scale comparative study of YOLO-based detectors for ischemic stroke lesion localization on diffusion-weighted imaging
Neuroscience, cilt.612, ss.134-151, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 612
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.neuroscience.2026.07.050
- Dergi Adı: Neuroscience
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, BIOSIS, Chemical Abstracts Core, EMBASE, MEDLINE, Academic Search Ultimate (EBSCO)
- Sayfa Sayıları: ss.134-151
- Anahtar Kelimeler: Deep learning, Diffusion-weighted imaging, Ischemic stroke, Medical image analysis, Stroke localization, You Only Look Once (YOLO)
- İstanbul Gelişim Üniversitesi Adresli: Evet
Özet
Ischemic stroke lesion localization on diffusion-weighted imaging (DWI) remains challenging because acute lesions may be small, faint, irregular, or multifocal. This study presents a controlled lesion-level benchmark of recent YOLO-based detectors for ischemic stroke localization on DWI. A private clinical cohort of 300 patients and approximately 2,200 lesion-containing axial trace-weighted DWI images was annotated with expert-consensus bounding boxes. Twenty-four detector configurations from YOLOv10, YOLO11, YOLOv12, YOLOv13, and YOLO26 were trained and evaluated under a unified protocol using patient-level partitioning, standardized transfer learning, a common augmentation policy, and identical evaluation settings. Performance was assessed using precision, recall, mAP@50, mAP@50–95, inference time, parameter count, GFLOPs, and patient-level bootstrap confidence intervals. The detector families showed distinct operating profiles. YOLO26x produced the highest strict-localization point estimate, achieving 0.8287 precision, 0.6867 recall, 0.7940 mAP@50, and 0.5094 mAP@50–95. YOLO11x achieved the highest precision, indicating more conservative positive detections, whereas YOLO12x achieved the highest recall, reflecting stronger lesion retrieval. YOLO26s reached a closely comparable mAP@50–95 value with markedly lower computational cost, suggesting a favorable accuracy-efficiency balance. Bootstrap confidence intervals overlapped among several leading detectors, indicating that small numerical margins should be interpreted cautiously. The study provides a clinically grounded comparison of YOLO-based DWI lesion localization and shows that detector selection should consider strict localization accuracy and computational efficiency together.