Tendon And Muscle Texture Assessment Using Ultrasound Imaging With Machine Learning

PURPOSE: Ultrasonography (US) is a widely used imaging modality for the diagnosis and treatment of musculoskeletal disease. It is particularly well suited for the assessment of Achilles tendons given their superficial anatomy and variance in treatment based on the type of injury or tendinopathy. Given the paucity of available quantitative data, the aim of this study is to determine the efficacy of machine learning processing as an adjunct for the diagnosis and monitoring of Achilles tendon pathology. A successful machine learning approach would enable future studies exploring the effectiveness of the various treatment options for tendon injury and tendinopathy. METHODS: Images of both diseased and healthy Achilles tendons were obtained in long and short axis views using high resolution B-mode US. All images were collected following a consistent protocol by one of two examiners. US scans were processed using image texture analysis algorithms like the Grayscale Level Co-occurrence Matrix (GLCM), and then used for training a deep convolutional neural network to characterize tendon status, similar to a machine learning framework previously used in a study identifying myofascial trigger points in muscle tissue. RESULTS: The study is currently collecting US scans (70 collected, 100 planned by May 2024). Preliminary analysis of the tendon scans has confirmed feasibility of texture comparison and tissue status classification. The scans undergo feature extraction and preprocessing before being used to train the machine learning model. Preliminary results for a binary classification of healthy and unhealthy Achilles tendon scans has shown the machine learning methods achieving 85% accuracy, 80% precision, 77% recall, and 78% F-score. CONCLUSION: The data and analysis evidence will enable planning of a future interventional clinical trial for selected tendinopathy treatment with multiple treatment options. The study to date has derived analysis of the correlation between US image features and tendinopathy pathologies which will be the basis for developing machine learning based tools for point of care assessment and longitudinal monitoring.