Back to 2025 Abstracts
Upper Extremity Fracture Detection Using a Convolutional Neural Network with Multiview Incorporation
Lainey Grey Bukowiec, MD; Julia Todderud, BS; Austin F. Grove, BA; Anish Kanabar, MS; A. Noelle Larson, MD
Mayo Clinic, Rochester, MN
INTRODUCTION:Detecting fractures in pediatric radiographs using artificial intelligence (AI) and machine learning (ML) is challenging due to the presence of physes. These appear as gaps or radiolucent regions in bones, complicating the distinction from fractures. Their appearance and location vary with skeletal development, necessitating specialized strategies, such as incorporating prior knowledge about physes or developing algorithms to differentiate them from fractures. Large, well-annotated pediatric radiograph datasets are crucial for training robust ML models. This study aimed to develop a convolutional neural network (CNN)-based deep learning (DL) algorithm utilizing multiple views of the forearm, elbow, and humerus for automated fracture detection in pediatric patients.
METHODS:A total of 3507 upper extremity radiographs were reviewed. Patients with multiple views, including anteroposterior and lateral forearm, elbow, and humerus radiographs, as well as internal and external oblique elbow radiographs, were stratified into training, validation, or testing cohorts using a 75-8-17 split. The testing set included 206 fractures and 378 normal radiographs. Images were resized to 512x512 and augmented with random flipping, rotation, translation, and Gaussian noise. Ground truth was determined by three orthopedic surgery reviewers. Classification was performed using a DL model with a ConvNeXt-V2-tiny backbone pre-trained on ImageNet-1k.
Measured output metrics included precision, recall and F1 score.RESULTS:The model achieved >0.99 in precision, recall, and F1 score, and >99% accuracy in detecting fractures in the 584 test images.
DISCUSSION AND CONCLUSION:A highly accurate DL approach for detecting fractures in pediatric forearm, elbow, and humerus radiographs was developed. This classification system specifically addresses pediatric patients, considering the variability in bone and joint appearance due to development stages and differences in image acquisition quality. Despite significant clinical implications of missed fractures in children, most AI research and commercial efforts have focused on adults, highlighting the novelty and importance of this study. The algorithm aims to improve the accuracy of radiograph interpretation in pediatric emergency departments, reducing the risk of missed fractures, which can lead to severe consequences like compartment syndrome and delayed care. Additionally, automated interpretation can alleviate the burden of manual interpretation, mitigate issues with noisy metadata, and manage variability in image acquisition techniques.
Back to 2025 Abstracts
