AI Ultrasound Improves Supraspinatus Tear Detection
Key Takeaways
- Among adults with shoulder pain at a Chinese center, an ultrasound-plus-clinical model distinguished MRI-defined supraspinatus tears from tendinopathy or bursitis better than clinical-only or ultrasound-only models.
- Adding a transverse US view without clinical inputs did not significantly improve discrimination over longitudinal US alone.
- The combined model detected all full-thickness tears in the internal test set; missed tears were partial-thickness.
- The combined model had the lowest Brier score, although calibration remained imperfect.
Jiang and colleagues assembled a single-center retrospective cohort at Wuxi Ninth People's Hospital Affiliated to Soochow University for their multimodal rotator cuff tear study. Among 947 adults presenting with shoulder pain or dysfunction, 370 had MRI-confirmed supraspinatus tears and 577 had MRI-confirmed supraspinatus tendinopathy or subacromial bursitis. Eligibility also required adequate ultrasound images and complete clinical information.
MRI was obtained within 60 days of the index ultrasound (US) examination. Every included examination provided longitudinal and transverse static views. The combined model fused features from both views with structured clinical variables covering symptoms, movement, and examination findings to estimate tear probability. Investigators compared clinical-only, longitudinal-US-only, dual-view-US-only, and dual-view-plus-clinical models. Training used stratified sampling and cross-validation confined to the training set; a held-out group of 284 underwent internal testing with training-derived parameters applied unchanged.
On the held-out test set, the combined model's area under the receiver operating characteristic curve (AUROC) was 0.935, versus 0.835 for dual-view US alone. DeLong testing found that the combined model discriminated tears significantly better than each comparator. These results distinguish MRI-defined supraspinatus tears from the selected nontear diagnoses, not from the full range of painful shoulder conditions.
Longitudinal US alone also outperformed the clinical-only model in discrimination, but adding a transverse view without clinical inputs did not produce a statistically significant gain over longitudinal US alone. SHapley Additive exPlanations (SHAP) ranked drop arm, external rotation lag, and empty can findings among the most influential clinical inputs. Gradient-weighted class activation mapping (Grad-CAM) qualitatively depicted model attention over the supraspinatus region; it did not verify tear localization.
The combined model's test sensitivity was 0.946 and specificity was 0.786. It detected all full-thickness tears in the test set, while its false negatives were limited to partial-thickness tears. Decision-curve analysis favored the combined configuration over the tested alternatives at assessed thresholds, but its modeled net benefit does not demonstrate improved patient outcomes.
All models showed imperfect calibration, particularly at higher predicted probabilities, although the combined configuration had the lowest Brier score. The classifier targeted supraspinatus tears against tendinopathy or bursitis, not tears in other cuff tendons or all causes of shoulder pain. Retrospective testing at one center among patients with adequate images and complete data does not establish external performance, prospective accuracy, MRI replacement, or clinical readiness. Examination interrater reliability was unmeasured; resized static frames can miss dynamic or fine image detail, and Grad-CAM heatmaps were not formally validated for precise localization.
The authors propose prospective multicenter evaluation, extension to additional rotator cuff pathology, and testing across different sonographers, ultrasound equipment, and populations. MRI-independent diagnosis remains a proposed application rather than an established alternative. Whether discrimination and calibration persist beyond this center remains an open validation question.
Clinician Questions
How was the classification threshold for MRI-referenced supraspinatus tear detection selected?
Investigators maximized the Youden index using pooled out-of-fold predictions from training-only cross-validation for the ultrasound-and-clinical model. They then applied the training-derived threshold unchanged to the held-out test set.
Who assigned the supraspinatus tear labels and acquired the ultrasound images for the classifier?
Board-certified radiologists assigned tear versus supraspinatus tendinopathy or subacromial bursitis labels from affected-shoulder MRI. Experienced musculoskeletal sonographers acquired the standardized shoulder scans, from which static views were used by the classifier.