Automated Multi-Layer Detection of Bone Fractures in X-Ray Images Using Transfer Learning

Deep learning-based fracture detection systems are gaining traction as they alleviate diagnostic burdens while enhancing radiological accuracy and efficiency. Due to the difficulty in identifying fractures in X-ray images even for experts, the need for a system that eases the identification of such fractures is very high. This paper presents a three-layer, automated fracture detection framework that uses advanced transfer learning techniques and transformer-based models. In the first layer, a Vision Transformer (ViT) model is fine-tuned to classify the input X-ray image into one of four bone types: hand, leg, hip, or shoulder. Once the bone type is identified, the second layer deploys a dedicated ResNet50-based binary classifier, trained specifically for that bone category, to determine whether a fracture is present. Following this, the third layer applies a Detection Transformer (DETR) to localize the fracture region accurately within the image. To address class imbalances in publicly available datasets, we extended the primary FracAtlas dataset with additional datasets, which resulted in a more balanced data collection, totaling 4,968 images. Our experiments demonstrate high classification accuracy (99.26%) in layer 1, strong binary fracture detection performance in layer 2 (accuracy varying across bone types, but reaching up to 91.67%), and precise localization in layer 3. Finally, we created a user-friendly graphical interface that integrates all three layers seamlessly, allowing users to upload an image, obtain bone classification, fracture detection, and annotated localization outputs, all within a single interface.