Introduction: Supraspinatus tendinopathy is a frequent cause of shoulder pain, affecting both quality of life and functional ability. Ultrasound offers benefits such as lower cost, greater accessibility, and real-time evaluation; however, its accuracy depends on the availability of a skilled operator. Using a deep learning (DL) model for automated supraspinatus tear classification and diagnosis may be beneficial in settings with limited resources. This study aims to utilise DL to classify supraspinatus tears based on ultrasound images.
Method: The DL model was trained using a combination of 2D longitudinal and transverse deidentified US images of the supraspinatus tendon, using approximately 600 annotated images with non-tear, partial- or full-thickness tears. Split: 85% training and 15% validation. There were three categories: non tear, partial tear and full tear. The included images had been identified and reported by a radiologist and tear sizes were confirmed by an orthopaedic surgeon with arthroscopy.
Results: The DL protype reached high performance metrics for detection and classification of supraspinatus tears on US images of approximately 93 % accuracy across all classifications. Final Metrics: Accuracy: 0.93,Precision: 0.93, Recall: 0.93, F1-Score: 0.93.
Conclusion: This DL model has the potential to act as a triaging tool for supraspinatus tears and address the gap between urgent diagnostic needs and lack of expertise/skills in under-resourced settings.
Take home message: Deep learning demonstrates significant potential in accurately triaging shoulder tears and supporting diagnostic processes. Implementing such a model may contribute to reduced healthcare costs and improved patient outcomes.