Abstract
The usage of artificial intelligence and machine learning has significantly strengthened computer-aided medical diagnostics, and fine-tuning models and architectures for medical detection purposes has become a common occurrence. Bone fracture detection is one of the applications where accurate localization of fractures is crucial for proper treatment. In this study, we propose a hybrid ResYOLO11 architecture that combines ResNet50's feature extraction capability and YOLO11's detection efficiency in a single model. The proposed architecture uses ResNet layers in the backbone and YOLO11 modules like C3K2, SPPF, and C2PSA to enhance the spatial feature representation, improve the classification precision and detection robustness. The architecture model was trained and evaluated on the public dataset of GRAZPEDWRI-DX, using precision, recall, mAP@50, and mAP@50-95 as performance metrics. The ResYOLO11 architecture achieved precision scores of 0.935, 0.944, 0.945, 0.956, and 0.963, and mAP@50 scores of 0.970, 0.974, 0.977, 0.982, and 0.986 across the nano, small, medium, large, and extra-large variants of the model, respectively. The inference time is 0.012, 0.014, 0.016, 0.019, and 0.026 seconds, respectively, for each model. Quantitative analysis show that ResYOLO11 achieved up to 4.2% higher mAP50 and 6.1% higher mAP50-95 compared to standard YOLO11 variants and was 24% faster in detecting fractures. This comparison showcases the architecture's potential for assisting orthopedic specialists in accurately identifying fractures and supporting clinical decision-making by providing a clinically robust and computationally efficient solution for computer-aided fracture diagnosis.