An Ensemble Deep Learning Approach for Automated Bone Fracture Detection in Medical Imaging

Recent advances in deep neural networks have significantly improved medical image analysis; however, detecting bone fractures in radiographic images remains challenging due to the complex skeletal structure, subtle fracture patterns, and limited annotated data. This study proposes an advanced ensemble deep learning framework that integrates two optimized architectures, EfficientNet-B4 and DenseNet-121, through a soft-voting fusion strategy to enable automated and precise bone fracture detection. The hybrid framework introduces adaptive weighting and optimized dense layers, which enhance feature discrimination and strengthen the network’s capacity to distinguish fine-grained fracture details. Moreover, transfer learning and fine-tuning techniques are employed to address data imbalance and improve model generalization across multiple anatomical regions. Comprehensive experiments conducted on the MURA dataset, consisting of radiographs from seven distinct anatomical regions, demonstrate that the proposed model achieves superior performance with 83.52% accuracy and 90.76% sensitivity, outperforming each individual baseline. The model’s robustness under different training configurations confirms its reliability and stability for clinical deployment. Overall, this research presents a novel ensemble-based diagnostic system that leverages complementary architectural strengths and adaptive feature fusion to achieve high diagnostic precision. The proposed method contributes not only to improving classification accuracy but also to establishing a scalable and interpretable framework for computer-aided fracture diagnosis, offering a practical step toward intelligent and reliable radiological decision support.