Abstract
Advances in computer vision and machine learning have augmented the ability to analyze orthopedic radiographs. A critical but underexplored component of this process is the accurate classification of radiographic views and localization of relevant anatomical regions, both of which can impact the performance of downstream diagnostic models. This study presents a deep learning object detection model and mobile application designed to classify distal radius radiographs into standard views-anterior-posterior (AP), lateral (LAT), and oblique (OB)- while localizing the anatomical region most relevant to distal radius fractures. A total of 1593 deidentified radiographs were collected from a single institution between 2021 and 2023 (544 AP, 538 LAT, and 521 OB). Each image was annotated using Labellerr software to draw bounding boxes encompassing the region spanning from the second digit MCP joint to the distal third of the radius, with annotations verified by an experienced orthopedic surgeon. A YOLOv5 object detection model was fine-tuned and trained using a 70/15/15 train/validation/test split. The model achieved an overall accuracy of 97.3%, with class-specific accuracies of 99% for AP, 100% for LAT, and 93% for OB. Overall precision and recall were 96.8% and 97.5%, respectively. Model performance exceeded the expected accuracy from random guessing (p < 0.001, binomial test). A Streamlit-based mobile application was developed to support clinical deployment. This automated view classification step reduces feature space by isolating only the relevant anatomy. Focusing subsequent models on the targeted region can minimize distraction from irrelevant areas and improve the accuracy of downstream fracture classification models.