Improving Deep Learning-Based Grading of Partial-thickness Supraspinatus Tendon Tears with Guided Diffusion...

Area of Science:

• Orthopedics and Sports Medicine
• Radiology
• Artificial Intelligence in Medicine

Background:

• Partial-thickness supraspinatus tendon (SST) tears are common, requiring accurate grading for effective treatment.
• Current grading methods can be subjective and time-consuming.
• Deep learning offers potential for objective and efficient tear assessment.

Purpose of the Study:

• To develop and validate a deep learning system for grading partial-thickness SST tears.
• To utilize guided diffusion-based data augmentation to improve model performance.
• To compare the system's accuracy and efficiency against experienced musculoskeletal radiologists.

Main Methods:

• A retrospective study of 1150 patients with confirmed SST tears.
• MRI images were augmented using a guided diffusion model to address data imbalance.
• A ResNet-34 model was trained and validated for Ellman grading on bursal-sided and articular-sided tears.
• Performance was evaluated using AUC and compared to three radiologists.

Main Results:

• The deep learning system achieved high AUCs (up to 0.99) for both tear types across MRI sequences.
• The model demonstrated significantly superior grading performance compared to radiologists (P<0.001).
• Guided diffusion augmentation improved model robustness, and the system reduced evaluation time.

Conclusions:

• The deep learning system offers a reliable and efficient tool for grading partial-thickness SST tears.
• The system achieves radiologist-level accuracy with enhanced consistency and speed.
• This AI approach shows promise for improving rotator cuff tear diagnosis.