Advancing Ischemic Stroke Detection Through an In-depth Evaluation of YOLOv10 Models on Diffusion-Weighted Imaging Data

Creative Commons License

Bayram B., Ince S., Kilicarslan S., Veziroglu E., Celik O., Pacal I.

Journal of Soft Computing and Decision Analytics, cilt.4, sa.1, ss.16-31, 2026 (Scopus)

Özet

Rapid and accurate detection of acute ischemic stroke (AIS) is critical for minimizing irreversible tissue damage and improving patient outcomes. While deep learning has revolutionized medical imaging, a gap remains in balancing high diagnostic precision with the computational efficiency required for real-time clinical deployment. This study presents a systematic evaluation of the YOLOv10 architecture, utilizing its advanced dual-label assignment and non-maximum suppression (NMS)-free inference, to address this challenge using Diffusion-Weighted Imaging (DWI) data. Utilizing the multi-center ISLES 2022 dataset, five YOLOv10 variants (Nano to Extra-Large) were trained on 1,652 preprocessed DWI images using a rigorous experimental design that included transfer learning and extensive data augmentation to handle anatomical asymmetries and data imbalance. The results indicate that the YOLOv10l (Large) variant emerged as the optimal model for high-stakes diagnostics, achieving a superior precision of 0.933, recall of 0.783, and mAP50 of 0.887, significantly outperforming lighter variants in complex lesion delineation. Conversely, the YOLOv10n (Nano) variant demonstrated ultra-fast processing speeds (0.8 ms), highlighting its potential for resource-constrained environments such as mobile stroke units. These findings confirm that YOLOv10 offers a versatile framework for stroke detection, where YOLOv10l provides the robust accuracy necessary for hospital-grade decision support, and YOLOv10n offers a viable solution for real-time triage, underscoring the necessity of model selection based on specific clinical constraints.