Bone Fracture Classification Using Deep Learning Models with Transfer Learning

Bone fracture can be defined as the complete or partial disruption of the integrity of bone tissue. Early and accurate diagnosis of fractures plays a decisive role in the effectiveness of treatment protocols and the recovery process of the patient. Today, X-rays, computed tomography (CT), and magnetic resonance imaging (MRI) have become routine in diagnosing bone fractures. Despite all these methods, rapid and accurate diagnosis of fractures, especially those that occur during sports activities, is still an important problem since sports injuries require urgent intervention. With the advancement of technology in the last decades, deep learning and transfer learning have become a common tool to provide rapid decision-making in medical diagnosis. In this study, the performance of three state-of-theart deep learning models, MobileNet V3-Small, EfficientNetV2Small and ShuffleNetV2, is evaluated in classifying simple and fragmented bone fractures using X-ray images. Comprehensive experiments are performed on 2384 images. The EfficientNetV2Small model achieved superior results with 0.984 accuracy and 0.997 AUC and outperformed other models. These findings show that deep learning models could assist experts in clinical applications in bone fracture detection.