Automated Detection of Bone Fractures in X-ray Images Using Deep Learning and Ensemble Learning

Abstract

Bone fractures are among the most common injuries worldwide and pose a significant challenge for accurate diagnosis, with error rates reaching up to 10%, potentially leading to health complications and delayed treatment. This study aims to develop and evaluate deep learning models for enhancing the accuracy and efficiency of fracture detection, utilizing three primary frameworks: conventional Convolutional Neural Networks (CNNs), the VGG19 architecture, and DenseNet121, with transfer learning leveraging CheXNet-pretrained weights optimized for medical imaging. Preprocessing techniques and hyperparameter optimization using the Hyperband algorithm were applied, and ensemble learning through soft voting was employed to integrate model outputs. The models were trained and evaluated on the FracAtlas dataset, which comprises over 4,000 X-ray images. Results indicated that the conventional CNN achieved an accuracy of 82.89%, although fracture recall was limited to 13%. Both VGG19 and DenseNet121 improved performance balance, achieving area under the curve (AUC) values of 0.79 and 0.81, respectively. The ensemble learning model achieved a performance close to that of the individual models. These findings demonstrate that deep learning can effectively support fracture diagnosis, particularly when incorporating transfer learning. However, challenges such as data imbalance and clinical case variability continue to affect model performance. This study represents a step toward the development of more reliable clinical decision support systems for fracture detection.