MOD-YOLOv8: a new YOLOv8-based model for meniscal tear detection in knee MRI

This study aimed to develop a deep learning–based approach for recognising meniscal tears on knee magnetic resonance imaging (MRI). A total of 600 examinations were randomly selected from the publicly available FastMRI dataset, yielding 828 labelled sagittal images for model development and evaluation. We proposed MOD-YOLOv8, an adaptation of the You Only Look Once version 8 (YOLOv8) architecture. Its performance was benchmarked against state-of-the-art (SOTA) YOLO models (YOLOv5, YOLOv6, YOLOv8, and YOLOv9t) using five-fold cross-validation and was also compared with RT-DETR and Detectron2 under the same dataset and evaluation protocol. The final metrics for the proposed model, including precision (P), recall (R), mAP 50 , and F1-score, were 0.8450, 0.8902, 0.9227, and 0.8670, respectively. Confidence intervals for P, R, mAP 50 , and F1-score were 0.8738 (95% CI [0.8020–0.9456]), 0.8846 (95% CI [0.8652–0.9041]), 0.8966 (95% CI [0.8507–0.9424]), and 0.8512 (95% CI [0.7995–0.9029]), respectively. These results demonstrate that MOD-YOLOv8 consistently outperforms the compared baselines while providing reliable detection performance. Heatmap-based visualisation indicated that the model focused on clinically relevant tear regions, supporting interpretability. Overall, MOD-YOLOv8 demonstrates superior accuracy, reliability, and potential clinical applicability for detecting meniscal tears on MRI.