AutoRCuff: CNN-Autoencoder-Based Intelligent Detection of Rotator Cuff Tendon Tears from Ultrasound Imaging

Abstract-Rotator cuff muscle tears are among the most prevalent musculoskeletal injuries, and ultrasound imaging serves as an effective diagnostic tool. However, the interpretation of these scans requires specialized expertise, often leading to significant delays in diagnosis. This study introduces an AI-driven approach to accelerate the detection of full-thickness rotator cuff tears, reducing assessment time from months to mere minutes. The proposed method consists of two key steps: first, segmentation of the humeral cortex and subacromial bursa, followed by classification of tears based on these identified regions. Automated segmentation in ultrasound imaging poses challenges due to speckle noise, low contrast, and image artifacts. To overcome these, we employ a CNN-based autoencoder that directly predicts the boundary contour points of relevant anatomical structures instead of traditional pixel-wise semantic segmentation. This approach enhances interpretability by focusing on clinically significant landmarks rather than relying on a black-box classifier. The study utilized a dataset of 206 patients, comprising 10,080 training images and 2,520 evaluation images. The proposed segmentation model outperformed the conventional UNet, achieving a Dice coefficient of 94.2% and a Hausdorff Distance of 2.8 mm, compared to UNet’s 90.5% DC and 6.8 mm HD. Following segmentation, a VGG-16-based classification model achieved an accuracy of 81.0%, with a sensitivity of 78.5% and specificity of 76.2%. The implementation of AI-powered ultrasound for rotator cuff tear detection has the potential to facilitate early and precise diagnosis, significantly improving patient outcomes. This automated system can be deployed in primary care settings such as general practitioner clinics and emergency departments, empowering lightly trained personnel to perform initial assessments efficiently.