JMIR Publications

ABSTRACT

Background:

Supraspinatus tendon pathologies are a leading cause of shoulder dysfunction, requiring timely, accurate diagnosis to guide treatment. Magnetic resonance imaging (MRI) is the gold standard for detecting such injuries and requires expert interpretation.

Objective:

This study aimed to develop an AI-based model to classify supraspinatus tendon status on coronal T2-weighted shoulder MRI into three clinically relevant categories: intact, tendinopathy or partial-thickness tears, and full-thickness tears.

Methods:

In total, 1,192 cases comprising 22,529 coronal T2-weighted MRI slices derived from 65 institutions were included. A hierarchical deep learning framework was constructed using three sequential 3D ResNet-18 models. The system included a Left–Right Classifier to standardize anatomical orientation, followed by two binary classification models: one to detect full-thickness tears (Model F), and another to distinguish intact tendons from tendinopathy or partial tears (Model ITP). Score-weighted Class Activation Mapping was used to support interpretation (Score-CAM).

Results:

The Left–Right Classifier achieved balanced accuracy (99.4%). Model F reached 94.8% identifying full-thickness tears, and Model ITP achieved 82.6% classifying intact versus partial/tendinopathy. The final hierarchical model yielded balanced accuracy (81.1%), outperforming conventional flat classification in initial comparisons.

Conclusions:

The new AI-based hierarchical model demonstrated strong performance classifying supraspinatus tendon pathologies and results aligned closely with clinical reasoning. This new model has the potential to improve diagnostic consistency, support non-specialist clinicians, and enhance efficiency in musculoskeletal imaging workflows. Clinical Trial: Not applicable