Abstract
Background/Objectives: Tibial plateau fractures are difficult to detect using X-ray imaging due to limited three-dimensional visibility. This study evaluated the performance of four You Only Look Once (YOLO) deep learning models trained on different radiographic views for fracture detection. Methods: A total of 1489 knee X-rays were collected from a tertiary referral hospital, comprising 727 fracture images and 762 non-fracture images. YOLOv4, YOLOv5, YOLOv8, and YOLOv9 were each trained using anteroposterior (AP), lateral, and combined views. Results: YOLO models trained on AP views consistently outperformed those trained on other views. YOLOv9 trained on AP images achieved the highest accuracy, specificity, precision, F1-score, and area under the curve (AUC) of 0.99, with both sensitivity and negative predictive value (NPV) at 1.00. YOLOv8 trained on AP views reached 0.97 across all metrics with an AUC of 0.98. YOLOv5 trained on AP images achieved an accuracy and F1-score of 0.98, a sensitivity and NPV of 0.97, and an AUC of 1.00. YOLOv4 trained on AP views showed slightly lower performance, with an accuracy and F1-score of 0.96 and an AUC of 1.00. External validation confirmed the strong generalizability of AP-trained models, particularly YOLOv9, which reached an accuracy of 0.87, a sensitivity of 1.00, a specificity of 0.75, a precision of 0.80, an NPV of 1.00, an F1-score of 0.88, and an AUC of 0.93. Artificial intelligence-assisted YOLO models showed strong potential in detecting tibial plateau fractures. Conclusions: Models trained on AP views consistently achieved better diagnostic accuracy than those using other views. Among all, YOLOv9 delivered the best results, highlighting the benefits of newer deep learning architectures.