Hybrid EfficientNet Models with Attention Mechanisms for Enhanced Bone Fracture Detection and Classification Using X-ray Images

Abstract

Bone fractures pose a significant diagnostic challenge, requiring swift and precise detection. This study introduces a modified EfficientNetB03 model enhanced with a self-attention mechanism to effectively detect and classify bone fractures in X-ray images sourced from the Kaggle Bone Fracture dataset. The model improves focus on subtle fracture details while maintaining computational efficiency. Preprocessing techniques were applied to enhance image quality, while data augmentation combined with transfer learning further boosted performance. Both EfficientNetB03 and the Attention Mechanism were trained independently, and later integrated to design the proposed Hybrid model. The accuracy achieved was 0.99 for EfficientNetB03, 0.91 for the Attention Mechanism, and 0.99 for the Hybrid model. The classification report highlights strong results across two classes: Class 0 (non-fracture) achieved a precision of 0.99 and a recall of 0.88, while Class 1 (fracture) attained a precision of 0.95 and a perfect recall of 1.00. F1-scores stood at 0.93 for Class 0 and 0.97 for Class 1, with support values of 372 and 925 instances, respectively. Overall, the proposed hybrid model achieved an accuracy range from 0.96 to 0.98 with cross-validation, and up to 0.99 accuracy without cross-validation. The model serves as an aid for medical professionals in delivering timely and accurate fracture detection and classification.