Patient-Level Classification of Rotator Cuff Tears on Shoulder MRI Using an Explainable Vision Transformer Framework

Area of Science:

• Orthopedics
• Radiology
• Artificial Intelligence

Background:

• Diagnosing rotator cuff tears (RCTs) with MRI is challenging due to complex anatomy and interobserver variability.
• Slice-centric Convolutional Neural Networks (CNNs) struggle with volumetric context for accurate RCT grading.
• This study addresses the need for improved automated RCT classification.

Purpose of the Study:

• Develop and validate the Patient-Aware Vision Transformer (Pa-ViT) for automated, patient-level RCT classification.
• Enhance diagnostic sensitivity, especially for subtle partial-thickness tears.
• Provide an explainable deep-learning framework for clinical decision support.

Main Methods:

• Utilized a retrospective dataset of 2447 T2-weighted coronal shoulder MRI scans.
• Employed a Vision Transformer (ViT-Base) backbone within a Weakly-Supervised Multiple Instance Learning (MIL) paradigm.
• Trained the model with weighted cross-entropy loss and benchmarked against CNNs and traditional classifiers.

Main Results:

• Pa-ViT achieved 91% overall accuracy and a 0.91 macro-averaged F1-score, outperforming VGG-16 (87%).
• Demonstrated superior performance for partial-thickness tears (ROC AUC: 0.903).
• Attention Rollout visualizations confirmed reliance on anatomical features, not artifacts.

Conclusions:

• Pa-ViT effectively models long-range dependencies, offering a robust alternative to CNNs for RCT diagnosis.
• The framework provides a clinically viable, explainable tool for enhanced diagnostic sensitivity.
• Pa-ViT shows promise for improving the detection of subtle partial-thickness rotator cuff tears.