Abstract
OBJECTIVE: This study aims to develop an artificial intelligence system for the accurate detection and classification of intertrochanteric femoral fractures (types A1-A3 according to the AO/OTA classification) in X-ray images, focusing on improving precision and optimizing computational efficiency. METHODS: This study adopted a retrospective design, using 976 X-ray image datasets collected from hospital archives. The images were preprocessed, annotated by orthopedic specialists, and divided into training and test sets. The model was improved by replacing the traditional convolutional downsampling modules in YOLOv8 with Average Pooling Downsampling (ADown) modules to enhance feature extraction for small fracture targets. Model training incorporated data augmentation techniques and was evaluated using metrics such as precision, recall, and mean Average Precision (mAP). RESULTS: The proposed YOLOv8-ADown model achieved an overall mAP50 of 81.7%, higher than the 80.5% of the original YOLOv8. The detection precision for A1, A2, and A3 type fractures increased by 7.3, 3.5, and 7.8%, respectively. Furthermore, the number of model parameters was reduced by 12.3%, and computational complexity (FLOPs) was decreased by 9.8%, demonstrating potential for deployment on edge devices. CONCLUSION: The YOLOv8-ADown model provides an efficient solution for fracture detection and is expected to assist in clinical diagnosis. Future work should address data collection challenges and conduct multi-center validation.