Related Experiment Video
Updated: Jul 10, 2026
06:45
Automated Joint Space Detection Improves Bone Segmentation Accuracy
Published on: November 28, 2025
AI-Assisted Detection of Supraspinatus Tendon Pathologies Using a Hierarchical Deep Learning Model to Improve
Kun-Hui Chen1,2,3, Jacky Chung-Hao Wu4,5,6, Hsin-Yu Chang6
1Institute of Clinical Medicine, National Yang Ming Chiao Tung University, No. 155, Sec. 2, Linong Street, Taipei, Taiwan, 886 2- 2871-2121 ext 23103.
JMIR Medical Informatics
|July 8, 2026
Summary
A hierarchical deep learning model shows promise for classifying supraspinatus tendon pathologies, potentially improving diagnosis of tendinopathy and partial-thickness tears on MRI. Further validation is needed for widespread clinical use.
Area of Science:
- Musculoskeletal imaging
- Artificial intelligence in radiology
- Deep learning for medical diagnosis
Background:
- Supraspinatus tendon pathologies are frequent causes of shoulder pain.
- Magnetic resonance imaging (MRI) is the gold standard for diagnosis but requires specialized interpretation.
- Automated classification systems can enhance diagnostic consistency and streamline musculoskeletal imaging workflows.
Purpose of the Study:
- To develop and evaluate a hierarchical deep learning model for classifying supraspinatus tendon status.
- The model aims to differentiate between intact tendons, tendinopathy/partial-thickness tears, and full-thickness tears.
- To compare the hierarchical model's performance against a standard flat classification model.
Main Methods:
- A hierarchical deep learning system was designed, including orientation classification, full-thickness tear detection, and intact vs. tendinopathy/partial-thickness tear classification.
- The system was trained and evaluated on 1192 shoulder MRI scans.
- Performance was assessed using internal test sets and an independent external cohort, comparing against a flat 3-class model.
Main Results:
- The hierarchical model demonstrated superior sensitivity for tendinopathy/partial-thickness tears (68.1%) compared to the flat model (57.4%) on the internal test set.
- Sensitivity for full-thickness tears was comparable between the hierarchical (94.1%) and flat (95.1%) models.
- On the external cohort, the hierarchical model showed significantly higher sensitivity (45.5%) for tendinopathy/partial-thickness tears than the flat model (18.2%), with improved balanced accuracy and macro F1-score.
Conclusions:
- A hierarchical deep learning approach may enhance the detection of tendinopathy and partial-thickness tears, particularly for less experienced readers.
- External validation suggests feasibility across different MRI sources, but generalizability is limited by single-institution data predominance.
- Further prospective studies are warranted to confirm these findings and evaluate clinical utility.