Author(s): Prottasha Modak and Ashoke Kumar Modak

Abstract: Accurate and efficient bone fracture detection remains a critical challenge in clinical practice, with implications for patient care outcomes and healthcare resource utilization. This research presents an innovative approach combining advanced image processing techniques with a hybrid deep learning architecture for automated bone fracture detection. The proposed system integrates a custom-designed CNN-Vision Transformer hybrid model with an optimized preprocessing pipeline, addressing key limitations in existing automated detection systems. Our methodology was evaluated using a comprehensive dataset of 10,000 X-ray images encompassing various fracture types and anatomical locations, with ground truth labels provided by experienced radiologists. The system achieved exceptional performance metrics, demonstrating 0.945 sensitivity and 0.932 specificity in fracture detection, surpassing previous benchmarks and approaching expert-level accuracy. Notable improvements include enhanced detection of subtle fractures such as hairline fractures (0.912 sensitivity) and robust performance across different anatomical locations, with particularly strong results in femur (0.956 accuracy) and tibia (0.942 accuracy) fractures. The system's average processing time of 230.3ms per image represents a 40% improvement over existing methods, making it suitable for real-time clinical applications. External validation on an independent dataset of 1,000 images confirmed the system's generalizability, maintaining consistent performance (0.928 accuracy) across different clinical settings. Implementation of Grad-CAM visualization demonstrated high interpretability, with the model focusing on clinically relevant regions in 94.3% of cases. These results suggest significant potential for improving diagnostic efficiency and accuracy in clinical settings, while the system's interpretability and rapid processing capabilities make it particularly suitable for integration into existing healthcare workflows.

DOI: 10.33545/26633582.2025.v7.i1a.152

Pages: 25-31 | Views: 486 | Downloads: 223

Download Full Article: Click Here