A Computer Vision and Machine Learning Approach to Classify Views in Distal Radius Radiographs

Recent advancements in computer vision and machine learning have significantly enhanced the capacity to analyze orthopedic radiographs, a vital tool in medical diagnostics. Despite these technological strides, one often overlooked yet crucial step is the precise classification of radiographic views and the localization of relevant anatomical regions. These factors are instrumental in the efficacy of subsequent diagnostic models, particularly in detecting fractures accurately. This study introduces a deep learning-based object detection model, paired with a mobile application, aimed at classifying distal radius radiographs into three standard views: anterior-posterior (AP), lateral (LAT), and oblique (OB). Alongside this classification, the model localizes the key anatomical area most pertinent to distal radius fractures. The dataset comprised 1,593 anonymized radiographs collected from a single institution over a period spanning 2021 to 2023, distributed fairly evenly among the three view categories: 544 AP, 538 LAT, and 521 OB. Annotation of each image was meticulously performed using Labellerr software by drawing bounding boxes that encompassed the anatomical region from the second digit metacarpophalangeal (MCP) joint to the distal third of the radius. These annotations were verified by an experienced orthopedic surgeon to ensure accuracy and clinical relevance. Leveraging this annotated dataset, a YOLOv5 object detection model was fine-tuned and trained with a train/validation/test split of 70/15/15. The resulting model demonstrated impressive performance metrics, achieving an overall accuracy of 97.3%. Breaking this down, the model reached class-specific accuracies of 99% for AP views, 100% for LAT views, and 93% for OB views. The precision and recall rates were similarly high, at 96.8% and 97.5%, respectively. Statistical analysis confirmed the model's performance was significantly better than random guessing (p < 0.001, binomial test). To facilitate clinical integration, a user-friendly mobile application was developed using Streamlit, enabling easy deployment within healthcare settings. This automated classification tool serves to reduce the feature space by isolating only the pertinent anatomical regions within radiographs. This focused approach is expected to enhance the accuracy of downstream fracture classification models by minimizing distractions from irrelevant anatomical structures. In sum, this research demonstrates a practical, high-performing solution for automated view classification and anatomical localization in distal radius radiographs. By concentrating subsequent diagnostic efforts on the correctly identified and localized regions, the approach holds promise for improved fracture detection and better patient outcomes, marking a valuable contribution to computer-assisted orthopedic imaging.