AI-Assisted Detection of Supraspinatus Tendon Pathologies Using a Hierarchical Deep Learning Model to Improve Clinical Applicability: Development and Evaluation Study.
Authors
Affiliations (12)
Affiliations (12)
- Institute of Clinical Medicine, National Yang Ming Chiao Tung University, No. 155, Sec. 2, Linong Street, Taipei, Taiwan, 886 2- 2871-2121 ext 23103.
- Department of Orthopaedics and Traumatology, Taipei Veterans General Hospital, Taipei, Taiwan.
- Department of Surgery, School of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan.
- Center for Fundamental Science, Kaohsiung Medical University, Kaohsiung, Taiwan.
- Biomedical Artificial Intelligence Academy, Kaohsiung Medical University, Kaohsiung, Taiwan.
- Institute of Statistics, National Yang Ming Chiao Tung University, Hsinchu, Taiwan.
- Department of Anaesthesiology, Taipei Veterans General Hospital, Taipei, Taiwan.
- Department of Medical Research, Kaohsiung Medical University Hospital, Kaohsiung, Taiwan.
- Department of Statistics and Data Science, Cornell University, Ithaca, NY, United States.
- Division of Nephrology, Department of Medicine, Taipei Veterans General Hospital, Taipei, Taiwan.
- Center for Intelligent Drug Systems and Smart Bio-devices, National Yang Ming Chiao Tung University, Hsinchu, Taiwan.
- Stem Cell Research Center, National Yang Ming Chiao Tung University, Taipei, Taiwan.
Abstract
Supraspinatus tendon pathologies are common causes of shoulder pain. Magnetic resonance imaging (MRI) is the reference imaging method but requires expert interpretation. Automated classification may improve diagnostic consistency and support musculoskeletal imaging workflows. This study aimed to develop and evaluate a hierarchical deep learning model to classify supraspinatus tendon status as intact tendons, tendinopathy/partial-thickness tears, or full-thickness tears. A total of 1192 shoulder MRI scans were analyzed. The hierarchical system consisted of a left-right orientation classifier, a full-thickness tear detector (model F), and a classifier for distinguishing intact tendons from tendinopathy/partial-thickness tears (model ITP). A flat 3-class model served as a baseline comparator. Performance was evaluated on both an internal test set and an independent external cohort. On the internal test set, the hierarchical system achieved a system-level sensitivity of 68.1% for tendinopathy/partial-thickness tears, outperforming the flat baseline (57.4%) while maintaining comparable sensitivity for full-thickness tears (hierarchical vs flat: 94.1% vs 95.1%). On the independent external cohort, the sensitivity for tendinopathy/partial-thickness tears was 45.5% for the hierarchical model and 18.2% for the flat baseline. The hierarchical model also showed a numerically higher balanced accuracy (hierarchical vs flat: 68.1% vs 64.5%), macro F1-score, and macro area under the curve, although its overall accuracy was lower (76.4% vs 79.8%). A hierarchical deep learning approach that mirrors clinical diagnostic reasoning may improve the recognition of tendinopathy and partial-thickness tears, a challenging category for nonspecialist readers. Given the overlapping CIs, these findings should be interpreted as indicative of a trend rather than definitive improvement. External validation supports feasibility across different MRI sources, though the predominance of data from a single institution limits generalizability and warrants further prospective evaluation.