A Robust Deep Learning Framework for Automated Bone Fracture Detection and Classification from X-Ray Images

Abstract

Bone fractures are a common medical condition that require accurate and timely diagnosis to prevent complications and ensure appropriate treatment. Manual interpretation of radiographic X-ray images can be challenging, particularly in complex cases, and may be affected by human error and workload. Recent advances in deep learning have demonstrated significant potential for automated fracture detection and classification. This study presents a comparative deep learning framework for automated bone fracture detection from X-ray images using Vision Transformer (ViT-B/16) and ResNet-50 architectures. A Bone Fracture Multi-Region X-ray dataset containing 10,580 grayscale images with a resolution of 224 × 224 pixels was used for training, validation, and testing. The dataset includes radiographic images from multiple anatomical regions, including the upper limbs, lower limbs, lumbar spine, hips, and knees. Appropriate preprocessing and data augmentation techniques were applied to improve data consistency and model generalization. The experimental results demonstrated that both models achieved strong performance in binary fracture classification. ViT-B/16 achieved the best performance, with a validation accuracy of 98.91% and a test accuracy of 100%. It achieved a precision, recall, and F1-score of 1.00, indicating correct classification of all test images. ResNet-50 achieved a test accuracy of 98.62%, with a precision of 0.99, recall of 0.98, and F1-score of 0.99. These results demonstrate that both architectures are effective for automated bone fracture detection, while ViT-B/16 achieved superior classification performance on the evaluated dataset. The findings highlight the potential of transformer-based and convolutional deep learning architectures to support automated bone fracture detection from radiographic X-ray images.