Title: Comparative Analysis of Deep Learning Architectures for Bone Fracture Detection: MobileNetV2 vs. ResNet50
Authors: Fatima M. Salman, Samy S. Abu-Naser
Volume: 10
Issue: 1
Pages: 39-51
Publication Date: 2026/01/28
Abstract:
Bone fracture detection is a critical task in orthopedic radiology, where timely and accurate diagnosis significantly impacts patient recovery. Manual interpretation of X-ray images can be challenging due to the subtle nature of certain fractures and the high workload of clinicians. This study explores the efficacy of Deep Learning models in automating the detection of bone fractures. We implemented and compared two prominent architectures: ResNet50 and a proposed optimized MobileNetV2.To address the challenges of class imbalance and limited clinical data, we utilized advanced Data Augmentation techniques and Dropout regularization. The models were evaluated on a dataset of 425 unseen test images. The results demonstrate that the proposed MobileNetV2 model significantly outperformed the baseline, achieving a remarkable 99% accuracy, a 1.00 Recall for fracture detection, and a Macro F1-score of 0.95. In contrast, while ResNet50 achieved 94% accuracy, it exhibited severe bias and failed to generalize to normal cases. Furthermore, we employed Grad-CAM (Gradient-weighted Class Activation Mapping) to provide visual interpretability, confirming that the model's "attention" aligns with actual anatomical fracture sites. These findings suggest that the lightweight MobileNetV2 architecture is not only highly accurate but also suitable for integration into portable medical devices and real-time clinical decision support systems, offering a reliable tool for enhancing diagnostic precision in emergency settings.