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

Area of Science:

• Orthopedic imaging analysis
• Computer vision in radiology
• Machine learning for medical diagnostics

Background:

• Accurate classification of radiographic views and anatomical localization are crucial for analyzing orthopedic radiographs.
• Current methods may not adequately address these components, potentially impacting downstream diagnostic model performance.
• Deep learning offers a promising approach to enhance radiograph analysis.

Purpose of the Study:

• To develop and validate a deep learning object detection model for classifying distal radius radiograph views (anterior-posterior, lateral, oblique).
• To localize the critical anatomical region relevant to distal radius fractures.
• To create a mobile application for clinical deployment of the developed model.

Main Methods:

• A dataset of 1593 deidentified distal radius radiographs was collected and annotated.
• A YOLOv5 object detection model was fine-tuned and trained on the annotated dataset.
• A Streamlit-based mobile application was developed for clinical use.

Main Results:

• The model achieved an overall accuracy of 97.3% in classifying radiographic views.
• Class-specific accuracies were 99% for AP, 100% for LAT, and 93% for OB views.
• Overall precision was 96.8% and recall was 97.5%, significantly outperforming random guessing (p < 0.001).

Conclusions:

• The deep learning model accurately classifies distal radius radiograph views and localizes relevant anatomy.
• Automated view classification and anatomical focus enhance the feature space for diagnostic models.
• This approach has the potential to improve the accuracy of downstream fracture classification models in clinical settings.