Deploying Lightweight AI Diagnostic Tools in Resource-Limited Healthcare Settings: A Case Study in Bone Fracture Detection

Abstract

Bone fractures are among the most common musculoskeletal injuries requiring timely diagnosis, yet accurate X-ray interpretation remains challenging in resource-limited healthcare settings due to radiologist shortages. This study develops and evaluates a Computer-Aided Diagnosis (CAD) system for automated binary fracture detection using Convolutional Neural Networks (CNNs). A publicly available X-ray dataset was rigorously preprocessed and augmented to mitigate overfitting. Multiple architectures, including custom CNNs and transfer learning models (EfficientNetB0, VGG16), were trained and compared. The optimal lightweight model was subsequently integrated into a secure, Flask-based web application featuring real-time inference and automated data privacy protocols. Experimental results demonstrate that transfer learning significantly outperforms custom architectures, with EfficientNetB0 achieving the highest diagnostic performance (93.8% accuracy, 0.96 AUC-ROC, and 92.7% sensitivity). However, the lightweight custom CNN achieved a highly competitive 89.5% accuracy with a rapid CPU inference time of ~120ms, making it ideal for practical deployment. Future studies should focus on multi-class fracture typing, external validation across diverse clinical datasets, and the integration of Explainable AI (XAI) to enhance radiologist trust.