Real-Time Bone Fracture Detection Using MobileNetV2 and Explainable AI for Clinical Integration

Abstract

Bone fractures are among the most frequent injuries requiring immediate diagnosis, yet traditional X-ray analysis is time-consuming and reliant on expert interpretation. As medical AI advances, there is an increasing requirement in terms of effective and implementable diagnostic tools. The purpose of the study is to create a real-time, clinically practical system to detect a fracture combining lightweight deep learning, interpretability, and system-level integration. A convolutional neural network with MobileNetV2 architecture was trained on a stratified dataset of the elbow X-ray images, which have been divided into three classes: normal, hairline, and displaced fractures. Generalization and explainability were performed with data augmentation, two-phase fine-tuning, and Grad-CAM. This model had an accuracy of 89.26 percent, a precision of 91.52 percent, F1 score of 89.04 percent and a minimum false negative of 14 cases out of 1018 cases. The system is delivered using Docker on the AWS EC2 and available as a web interface implemented using Flask, which provides an opportunity to apply it in distant clinical facilities. The suggested pipeline merges both deep learning research and clinical practice domains because it provides a system allowing one to detect bone fractures quickly, interpretably, and scale up, and is the first of its kind to provide an entity that is accurate, can be used in real-time, and be deployable end-to-end.