On Construction of Tibial Plateau Fracture Detection in Different Radiographic Views Using YOLO Models.

Area of Science:

• Radiology
• Artificial Intelligence
• Deep Learning

Background:

• Tibial plateau fractures present diagnostic challenges with conventional X-ray imaging due to limited 3D visualization.
• Deep learning, specifically You Only Look Once (YOLO) models, offers potential for enhanced fracture detection.

Purpose of the Study:

• To evaluate the performance of four YOLO deep learning models (YOLOv4, YOLOv5, YOLOv8, YOLOv9) in detecting tibial plateau fractures.
• To compare the efficacy of models trained on different radiographic views (anteroposterior, lateral, combined).

Main Methods:

• A dataset of 1489 knee X-rays (727 fracture, 762 non-fracture) was utilized.
• YOLOv4, YOLOv5, YOLOv8, and YOLOv9 models were trained using anteroposterior (AP), lateral, and combined radiographic views.
• Model performance was assessed using metrics including accuracy, sensitivity, specificity, precision, F1-score, and area under the curve (AUC).

Main Results:

• Models trained on AP views consistently outperformed those trained on lateral or combined views.
• YOLOv9 trained on AP images achieved the highest performance across all metrics (accuracy, specificity, precision, F1-score, AUC) with values up to 0.99 and 1.00 for sensitivity and NPV.
• External validation demonstrated strong generalizability, with AP-trained YOLOv9 achieving an accuracy of 0.87 and AUC of 0.93.

Conclusions:

• Training deep learning models on AP radiographic views significantly enhances diagnostic accuracy for tibial plateau fractures.
• YOLOv9, particularly when trained on AP views, demonstrates superior performance, highlighting the advantages of advanced deep learning architectures.
• AI-assisted YOLO models show considerable promise for improving the detection of tibial plateau fractures.