Artificial intelligence empowering the diagnosis and treatment of rotator cuff tears

Abstract

Background:

Artificial intelligence (AI) is being developed for image acquisition, quantitative tissue assessment, diagnosis, and clinical prediction in rotator cuff tears. The breadth of this literature and its methodological limitations have not been integrated across the complete care pathway.

Methods:

We searched PubMed, Embase, Web of Science, Scopus, and the Cochrane Library from inception to July 5, 2026 for English-language original studies applying AI to rotator cuff-related reconstruction, segmentation or quantitative imaging, diagnosis or classification, anatomical measurement, or prognosis. Two reviewers screened reports and extracted study, data, model, validation, reference standard, and performance characteristics. Diagnostic accuracy studies were assessed with QUADAS-2 and prediction-model studies with PROBAST. Because clinical tasks, modalities, models, reference standards, and reported metrics were not sufficiently comparable, evidence was synthesized narratively rather than pooled statistically.

Results:

Sixty-six reports were included: 35 diagnostic accuracy or classification studies, 20 segmentation or quantitative imaging studies, 9 prognosis or outcome prediction studies, 1 Magnetic Resonance Imaging reconstruction or image-optimization study, and 1 exploratory unsupervised-phenotyping study. Applications spanned Magnetic Resonance Imaging, radiography, ultrasound, computed tomography, clinical or multimodal data, motion video, and emerging sensing data. QUADAS-2 classified overall risk-of-bias as high in 27/36 diagnostic studies, unclear in 6/36, and low in 3/36. PROBAST classified all 8/8 completed prediction-model studies as high risk-of-bias; 1 protocol without model results was not rated. Promising internal performance was recurrent, but selection methods, imperfect or incompletely described reference standards, limited external validation, small or selected datasets, and analysis-stage weaknesses constrained clinical inference.

Conclusion:

AI has proof-of-concept value across rotator cuff imaging and decision support, but the current evidence does not establish routine clinical readiness or replacement of clinician judgment. Prospective multicenter studies, transparent reference standards, independent external validation, calibration and clinical-utility assessment, and standardized task-specific reporting are required.

1. Introduction

Rotator cuff tears (RCT) are a common source of shoulder pain and functional limitation. Diagnostic classification, treatment planning, and postoperative assessment depend on the integration of clinical examination with radiography, ultrasound, magnetic resonance imaging (MRI), computed tomography (CT), and longitudinal outcomes.^[1,2] This workflow is vulnerable to variation in image quality, acquisition time, reader experience, qualitative grading, and the reproducibility of anatomical measurements.^[3]

Artificial intelligence (AI), including deep learning (DL) and conventional machine learning, can address different stages of this pathway. Proposed applications include accelerated MRI reconstruction, automated segmentation of rotator cuff muscles and tendons, continuous quantification of atrophy and fatty infiltration, tear detection and classification, automated measurement of shoulder morphology, prediction of retear or recovery, and monitoring of rehabilitation or surgical performance.^[4,5] These tasks differ substantially in their intended use, input data, reference standard, and clinically meaningful performance criterion.^[6]

Previous reviews have emphasized selected diagnostic or imaging applications, while the primary literature has expanded across modalities and clinical tasks.^[7,8] Directly juxtaposing performance estimates can nevertheless be misleading: area under the receiver area under the curve (AUC), sensitivity, specificity, negative predictive value (NPV), accuracy, Dice similarity coefficient, measurement error, calibration, and decision utility answer different questions.^[9] In addition, a high estimate from internal testing may coexist with substantial risk of selection bias, reference standard bias, data leakage, overfitting, or limited applicability.^[10]

We therefore conducted a PRISMA 2020-guided systematic review of AI applications across the RCT care pathway. We aimed to classify the evidence by clinical task, summarize study and model characteristics, examine heterogeneity by modality, model family, and tear-related target, evaluate diagnostic studies with QUADAS-2 and prediction studies with PROBAST, and determine which methodological limitations currently restrict translation into routine care.

2. Methods

2.1. Protocol and reporting

This systematic review was constructed and reported in accordance with the PRISMA 2020 statement. The eligibility criteria, study selection process, data items, synthesis approach, and risk-of-bias procedures were specified before the final evidence synthesis.^[11] This systematic review has been registered with PROSPERO (CRD420261455708).

2.2. Information sources and search strategy

PubMed, Embase, Web of Science Core Collection, Scopus, and the Cochrane Library were searched from database inception to July 5, 2026. Search concepts combined controlled vocabulary, where available, and title, abstract, or topic terms for rotator cuff injury with terms for AI, machine learning, DL, neural networks, computer-aided diagnosis, automated segmentation, image reconstruction, and prediction models. The complete database-specific strategies are reproduced in Supplementary Material 1, Supplemental Digital Content 1. Reference lists of relevant articles and prior reviews were also inspected for potentially eligible reports.

2.3. Eligibility criteria

We included English-language original human studies that developed, validated, or evaluated an AI, machine learning, or DL method for a rotator cuff-related task: shoulder MRI reconstruction or optimization; segmentation or quantitative assessment of cuff muscles, tendons, or relevant shoulder structures; diagnosis or classification of RCTs using imaging, clinical records, examination, motion, or multimodal data; automated anatomical measurement relevant to rotator cuff pathology; prediction of postoperative retear, complications, function, recovery, resource use, or rehabilitation performance; or exploratory phenotyping directly related to rotator cuff disease.

We excluded duplicate reports; non-English publications; reviews, editorials, letters, comments, and conference abstracts lacking sufficient original methods and results; animal or cadaveric studies; studies unrelated to rotator cuff injury; studies using only conventional non-AI image processing; and reports without an eligible model development, validation, performance, or clinically relevant outcome component. A study protocol was retained in the descriptive evidence map but was not assigned a completed model risk-of-bias rating.

2.4. Study selection

Two reviewers independently screened titles and abstracts and then assessed potentially eligible full texts. Disagreements were resolved by discussion, with consultation of a senior reviewer when required. The review team recorded counts at each stage and grouped full-text exclusions by the principal reason. The resulting selection process is shown in Figure 1.

Figure 1.

PRISMA 2020 flow diagram for new systematic reviews which included searches of databases and registers only.

2.5. Data extraction

Two reviewers extracted first author, publication year, study design, data source and modality, target condition or outcome, model family, sample or image volume, training-validation-test separation, validation strategy, reference standard or outcome ascertainment, primary performance metrics, and stated clinical value. Extracted values were checked against the full reports. Performance was retained in the metric reported by each study; no conversion between nonequivalent metrics was attempted.

2.6. Risk-of-bias and applicability assessment

Diagnostic accuracy and classification studies were assessed with QUADAS-2 across patient selection, index test, reference standard, and flow and timing. Prediction-model studies were assessed with PROBAST across participants, predictors, outcome, and analysis. Domain-level judgments were summarized as low, unclear, or high risk-of-bias using signaling questions and report-level support. The protocol without completed model results was not rated with PROBAST. Reconstruction, segmentation, measurement, and exploratory studies that did not estimate diagnostic accuracy or clinical prediction were not forced into either tool because doing so would imply applicability that the instruments do not have.^[12,13]

2.7. Data synthesis and heterogeneity

Clinical and methodological heterogeneity was anticipated because the studies addressed different points in the care pathway, used different modalities, enrolled different populations, defined tears and outcomes differently, and reported nonuniform metrics. We therefore performed a structured narrative synthesis stratified by clinical task, imaging or data modality, model family, and tear-related target when reported. Statistical pooling was not performed because sufficiently comparable effect estimates and denominators were unavailable within clinically coherent groups.

AUC, sensitivity, specificity, NPV, accuracy, Dice coefficient, measurement error, calibration, and clinical-utility measures were treated as noninterchangeable. Comparisons were restricted to studies reporting the same type of metric and were interpreted alongside prevalence, case selection, reference standard, and validation design. Statistical heterogeneity tests, meta-regression, funnel plots, and formal publication-bias tests were not applied because there was no pooled effect-size set for which these procedures would be valid. Possible reporting and publication biases were instead considered qualitatively, including selective reporting of favorable internal performance and the scarcity of negative or externally validated studies.

3. Results

3.1. Study selection

The searches identified 1506 records: 12 from the Cochrane Library, 380 from Embase, 246 from PubMed, 491 from Scopus, and 377 from Web of Science. After removal of 771 duplicates, 735 unique records remained. One non-English record was removed before screening, leaving 734 records for title and abstract screening. We excluded 320 reviews, conference abstracts, comments, and other non-original publications and sought 414 full-text reports; 139 could not be retrieved. Of 275 reports assessed for eligibility, 95 were unrelated to rotator cuff injury and 114 lacked relevant AI content. Sixty-six reports were included (Fig. 1).

3.2. Evidence taxonomy and study characteristics

The 66 reports comprised 35 diagnostic accuracy or classification studies, 20 segmentation or quantitative imaging studies, 9 prognosis or outcome prediction studies, 1 MRI reconstruction or image-optimization study, and 1 exploratory unsupervised-phenotyping study. The dominant modality was shoulder MRI (36 studies), followed by radiography (10), clinical or multimodal data (11 when clinical/multimodal and examination datasets were combined), CT (4), ultrasound (3), smartphone video or motion data (1), and microwave sensing or digital-twin data (1). Thirty-eight reports did not clearly state the design in the abstract or accessible report summary; 17 were retrospective, 5 prospective, 3 retrospective case-control, 2 multicenter, and 1 was a protocol. Detailed characteristics are provided in Tables 1 and 2 (One dual-purpose report contributed both a diagnostic model and a prognostic model, and 1 report contributed both segmentation and classification results; consequently, table and appraisal counts exceed the mutually exclusive primary-category counts.)

Table 1.

Characteristics of AI-based rotator cuff segmentation and quantitative imaging studies.

Author	Yr	Image data	Model development	Segmentation/measurement target	Data volume	Primary performance	Clinical value
Kim et al[14]	2019	Shoulder MRI	Fully convolutional CNN	Rotator cuff muscle atrophy/fatty infiltration quantification	240 patients	Dice: fossa 0.9718 ± 0.012; supraspinatus 0.9463 ± 0.047	Automated rotator cuff muscle atrophy/ fatty infiltration quantification
Hess et al[15]	2025	Shoulder MRI	Deep learning segmentation network	Rotator cuff tear-region segmentation/measurement	200 training MRIs; 59 evaluation MRIs	Tear Dice 0.58 ± 0.21; automatic Patte classification kappa 0.62	Automated rotator cuff tear-region segmentation/measurement
Levin et al[16]	2026	CT	Deep learning muscle segmentation model	Rotator cuff muscle atrophy/fatty infiltration quantification	952 shoulder CT scans	Dice: supraspinatus 0.92; subscapularis 0.92; infraspinatus 0.91; teres minor 0.81	Automated rotator cuff muscle atrophy/ fatty infiltration quantification
Wong et al[17]	2023	CT	2D/3D CNN ensemble	Rotator cuff or shoulder structure segmentation	55 MRI patients; 4 paired MRI/CT test patients	Best Dice: 0.86 (2D) and 0.82 (3D)	Automated rotator cuff or shoulder structure segmentation
Salhi et al[4]	2025	Shoulder MRI	Xception, InceptionV3, MobileNetV2	Rotator cuff muscle atrophy/ fatty infiltration quantification	383 patients; 1149 segmented muscle images	Best fused-model accuracy 91.1%; specificity 95.5%; F1 92.1%	Automated rotator cuff muscle atrophy/ fatty infiltration quantification
Alipour et al[18]	2024	Shoulder MRI	U-Net	Rotator cuff or shoulder structure segmentation	157 individuals	Median Dice: supraspinatus 0.89; subscapularis 0.86; infraspinatus 0.86; teres minor 0.78	Automated rotator cuff or shoulder structure segmentation
Li et al[19]	2025	Shoulder MRI	FPN + U-Net segmentation	Rotator cuff or shoulder structure segmentation	376 patients/5640 images for training; 94 patients/1410 images for testing	IoU 0.79 ± 0.01; Dice 0.75 ± 0.01 for tear-area segmentation	Automated rotator cuff or shoulder structure segmentation
Kim et al[20]	2024	Shoulder MRI	Deep neural network	Rotator cuff or shoulder structure segmentation	94 internal MRI scans plus 20 external MRI scans	External Dice 0.920; sensitivity 0.933; precision 0.912	Automated rotator cuff or shoulder structure segmentation
Hess et al[21]	2023	Shoulder MRI	nnU-Net	Rotator cuff or shoulder structure segmentation	111 training and 60 test MRI volumes from 76 patients across 19 centers	Mean Dice 0.91 ± 0.06; automated error flagging sensitivity 1.00/specificity 0.94	Automated rotator cuff or shoulder structure segmentation
Hess et al[22]	2026	Shoulder MRI	Voxel-wise deep learning regression	Rotator cuff muscle atrophy/ fatty infiltration quantification	75 training patients plus 24 validation patients	Whole-muscle fat-fraction error −0.5 ± 2.2% to 2.3 ± 3.9% versus Dixon MRI	Automated rotator cuff muscle atrophy/fatty infiltration quantification
Kim et al[23]	2022	Shoulder MRI	nnU-Net	Rotator cuff tear-region segmentation/ measurement	56 internal MRI scans plus 10 external MRI scans	Internal Dice: tendon 0.807; muscle 0.858; bone 0.978	Automated rotator cuff tear-region segmentation/ measurement
Lee et al[24]	2023	Shoulder MRI	CNN, 3D U-Net, U-Net	Rotator cuff tear-region segmentation/ measurement	303 patients	Dice 0.943; sensitivity 0.971; specificity 0.950	Automated rotator cuff tear-region segmentation/ measurement
Medina et al[25]	2021	Shoulder MRI	CNN, U-Net	Rotator cuff or shoulder structure segmentation	258 cases for Y-view selection; 1048 annotated Y-view images for segmentation	Dice (F1) score determined segmentation accuracy on 105 internal and 50 external test images.	Automated rotator cuff or shoulder structure segmentation
Conze et al[26]	2020	Shoulder MRI	Deep convolutional encoder-decoder	Rotator cuff or shoulder structure segmentation	24 shoulder examinations	Third, extended versions of deep convolutional encoder-decoder architectures using encoders pre- trained on non-medical data are proposed to improve the segmentation accuracy.	Automated rotator cuff or shoulder structure segmentation
Riem et al[27]	2023	Shoulder MRI	AI segmentation + multivariable regression	Rotator cuff muscle atrophy/ fatty infiltration quantification	47 paired CT/MRI cases for method development; 170 clinical MRIs for age/sex analysis	Scapular-length prediction mean absolute error 2.92%±1.68%; clinical quantification performed in 170 MRIs	Automated rotator cuff muscle atrophy/ fatty infiltration quantification
Ro et al[28]	2021	Shoulder MRI	AI/ML model (architecture not clear in abstract)	Rotator cuff muscle atrophy/ fatty infiltration quantification	240 randomly selected patients	Dice: fossa 0.97; supraspinatus 0.94	Automated rotator cuff muscle atrophy/ fatty infiltration quantification
Taghizadeh et al[29]	2021	CT	CNN	Rotator cuff or shoulder structure segmentation	103 CT scans from 95 patients	Mean Dice 0.88 ± 0.09	Automated rotator cuff or shoulder structure segmentation
Wang et al[30]	2024	Shoulder MRI	Attention-dense spatial-pyramid U-Net	Rotator cuff or shoulder structure segmentation	60 patients	Reported supraspinatus Dice approximately 0.91	Automated rotator cuff or shoulder structure segmentation
Hess et al[31]	2025	CT	nnU-Net + landmark network	Scapula/ shoulder morphology segmentation and measurement	Paired MRI/CT development data; 10-case held-out test set	Scapula Dice 0.87 ± 0.05; CT-referenced morphology errors also reported	Automated scapula/ shoulder morphology segmentation and measurement
Saavedra et al[32]	2024	Shoulder MRI	VGG-19, U-Net	Supraspinatus fatty infiltration grade	606 shoulders (900 T2-weighted Y-view MRIs)	Segmentation Dice 0.9441 ± 0.0031; classification AUC 0.9903 ± 0.0092	Automated supraspinatus fatty infiltration grade
Lee et al[33]	2021	Ultrasound	SMART-CA fully convolutional network	Rotator cuff tear-region segmentation/ measurement	59,265 shoulder ultrasound images; clinical experiments in 35 patients	Dice 0.736 overall and 0.484 with severe speckle noise	Automated rotator cuff tear-region segmentation/ measurement

Table 2.

Characteristics of AI-based rotator cuff diagnostic and classification studies.

Author	Yr	Input data	Model development	Target diagnosis/classification	Data volume	Primary diagnostic performance	Clinical value
Sezer et al[34]	2019	Shoulder MRI	CapsNet, CNN	Rotator cuff/ supraspinatus tear or pathology	1006 PD-weighted shoulder MRI images	Overall classification accuracy 94.75%	AI-assisted rotator cuff/ supraspinatus tear or pathology
Miao et al[35]	2025	Clinical/ multimodal data	DSFF-3DNet, YOLO	Rotator cuff/ supraspinatus tear or pathology	1014 patients plus 2 independent external test sets	AUC: 0.9788 internal; 0.8806 and 0.8447 on two external test sets	AI-assisted rotator cuff/ supraspinatus tear or pathology
Iio et al[36]	2023	Shoulder radiography	Deep CNN	Rotator cuff/ supraspinatus tear or pathology	2803 shoulder radiographs (2241 train; 280 validation; 282 test)	AUC 0.82; sensitivity 91.3%; NPV 92.7%	AI-assisted rotator cuff/ supraspinatus tear or pathology
Hashimoto et al[37]	2024	Shoulder radiography	CNN	Rotator cuff/ supraspinatus tear or pathology	1169 shoulder radiographs	AUC 0.88; sensitivity 92%; specificity 69%; accuracy 86%	AI-assisted rotator cuff/ supraspinatus tear or pathology
He et al[38]	2025	Shoulder MRI	Multimodal radiomics	Subscapularis tendon tear	324 development/internal patients plus 60 external patients	Best external/internal MRI sequence performance included AUC up to 0.844	AI-assisted subscapularis tendon tear
Ni et al[39]	2025	Shoulder MRI	ResNet-34	Rotator cuff/ supraspinatus tear or pathology	1150 internal patients plus 224 external patients	Internal AUCs up to 0.99; external combined-sequence AUCs 0.95–0.99	AI-assisted rotator cuff/ supraspinatus tear or pathology
Key et al[40]	2022	Shoulder MRI	VGG19, support vector machine, artificial neural network	Rotator cuff/ supraspinatus tear or pathology	468 MRI images	Accuracy >99% with the KNN classifier	AI-assisted rotator cuff/ supraspinatus tear or pathology
Engin et al[41]	2025	Shoulder MRI	ResNet-50	Rotator cuff/ supraspinatus tear or pathology	Retrospective MRI cohort; exact total not clearly stated in abstract	Sensitivity 93.75–100% but specificity 18.75–62.5%; partial-vs-full accuracy 34.38%	AI-assisted rotator cuff/ supraspinatus tear or pathology
Esfandiari et al[42]	2026	Shoulder MRI	MobileNetV2, SqueezeNet, CNN	Rotator cuff/ supraspinatus tear or pathology	150 subjects (75 normal; 75 pathological)	Accuracy 94.5%; sensitivity 94.6%; specificity 93.4%	AI-assisted rotator cuff/ supraspinatus tear or pathology
Esfandiari et al[43]	2023	Shoulder MRI	MobileNetV2, SqueezeNet, CNN	Rotator cuff/ supraspinatus tear or pathology	150 shoulder MRI images (balanced classes)	Accuracy 92.67%; sensitivity 91.75%; specificity 92.22%	AI-assisted rotator cuff/ supraspinatus tear or pathology
Alike et al[44]	2023	Shoulder radiography	AI/ML model (architecture not clear in abstract)	Rotator cuff/ supraspinatus tear or pathology	974 patients (828 train; 89 internal validation; 57 external validation)	AUC 0.902; sensitivity 0.880; specificity 0.812; accuracy 0.836	AI-assisted rotator cuff/ supraspinatus tear or pathology
Axenhus et al[45]	2026	Shoulder radiography	CNN	Cuff tear arthropathy/ glenoid erosion	6733 radiographs for development; 561 validation images	AUC 0.95 for Hamada grades 3–4; 0.81–0.91 for other grades/erosion	AI-assisted cuff tear arthropathy/ glenoid erosion
Bautista et al[46]	2026	Shoulder radiography	VGG19, CNN	Rotator cuff calcific tendinopathy	4268 development radiographs; 480 internal and 308 external validation radiographs	Internal AUC 0.956–0.961; external AUC 0.940–0.942	AI-assisted rotator cuff calcific tendinopathy
Darevsky et al[47]	2023	Smartphone video/ motion data	logistic regression	Rotator cuff/ supraspinatus tear or pathology	Human RC-tear/control cohort assessed by smartphone motion capture; small sample	AUC 0.944 ± 0.051; accuracy >90%	AI-assisted rotator cuff/ supraspinatus tear or pathology
Chiu et al[48]	2022	Ultrasound	CNN	Rotator cuff calcific tendinopathy	133,619 source ultrasound images from 7836 consecutive patients; 2462 selected images used for modeling	Accuracy 91.32%; sensitivity 87.89%; specificity 94.74%	AI-assisted rotator cuff calcific tendinopathy
Lin et al[49]	2023	Shoulder MRI	Multisequence 3D ResNet-50	Rotator cuff/ supraspinatus tear or pathology	11,925 shoulder MRI examinations from two institutions	AUC: supraspinatus 0.93; infraspinatus 0.89; subscapularis 0.90	AI-assisted rotator cuff/ supraspinatus tear or pathology
Guo et al[50]	2023	Shoulder MRI	Xception, CNN	Rotator cuff/ supraspinatus tear or pathology	701 MRI examinations/2804 images plus 69-surgery-patient clinical test set	AUC 0.921 (surgery test) and 0.882 (internal test)	AI-assisted rotator cuff/ supraspinatus tear or pathology
Ho et al[51]	2022	Ultrasound	VGG19, ResNet-50, Xception, DenseNet121	Rotator cuff/ supraspinatus tear or pathology	194 ultrasound images	AUC 0.832; accuracy 88.2%; sensitivity 93.8%; specificity 83.6%	AI-assisted rotator cuff/ supraspinatus tear or pathology
Hou et al[52]	2026	Shoulder radiography	XGBoost, random forest, logistic regression, support vector machine	Rotator cuff/ supraspinatus tear or pathology	342 patients (180 tears; 162 intact)	XGBoost AUC 0.87 training and 0.74 testing	AI-assisted rotator cuff/ supraspinatus tear or pathology
Asci et al[53]	2026	Shoulder MRI	Pa-ViT, Vision Transformer, CNN	Rotator cuff/ supraspinatus tear or pathology	2447 T2-weighted coronal shoulder MRI examinations	Accuracy 91%; macro-F1 0.91; partial-tear AUC 0.903	AI-assisted rotator cuff/ supraspinatus tear or pathology
Kang et al[54]	2021	Shoulder radiography	Deep CNN	Subscapularis tendon tear	2779 development radiographs; 340 arthroscopy reference and 627 MRI reference test cases	AUC 0.83 (arthroscopy reference) and 0.82 (MRI reference)	AI-assisted subscapularis tendon tear
Kim et al[55]	2020	Shoulder radiography	Deep CNN	Rotator cuff/ supraspinatus tear or pathology	6793 development radiograph series; 1095 temporal test series	Sensitivity 97.3%; NPV 96.6%; LR −0.06	AI-assisted rotator cuff/ supraspinatus tear or pathology
Li et al[56]	2023	Clinical and examination data	XGBoost, random forest	Rotator cuff/ supraspinatus tear or pathology	1684 patients; 417 arthroscopy-confirmed RCTs	XGBoost AUC 0.92; accuracy 0.85; Brier score 0.15	AI-assisted rotator cuff/ supraspinatus tear or pathology
Lu et al[57]	2014	Shoulder MRI	logistic regression, decision tree	Rotator cuff/ supraspinatus tear or pathology	169 patients	This study used predictive data mining and Bayesian theory to improve the accuracy of diagnosing rotator cuff tears by clinical examination alone.	AI-assisted rotator cuff/ supraspinatus tear or pathology
Fei et al[58]	2024	Shoulder MRI	Radiomics	Postoperative retear	Retrospective diagnostic cohort and postoperative retear subset; see article for cohort splits	Diagnosis AUC 0.979 validation; retear nomogram AUC 0.808 ± 0.081 validation	AI-assisted postoperative retear
Matcuk et al[59]	2020	Shoulder MRI	CART/ random forest/ logistic models	Rotator cuff/ supraspinatus tear or pathology	200 MRI patients (120 RCT; 80 normal)	CART model accuracy up to 99.5% for separating RCT and non-RCT groups	AI-assisted rotator cuff/ supraspinatus tear or pathology
Oeding et al[60]	2024	Clinical and examination data	XGBoost	Subscapularis tendon tear	Surgical cohort with MRI-confirmed/operative subscapularis status; exact total in full-text	C-statistic 0.84; accuracy 0.85; F1 0.87	AI-assisted subscapularis tendon tear
Oettl et al[61]	2025	Shoulder MRI	Vision transformer, CNN	Traumatic versus degenerative tear	99 patients (79 train; 20 validation)	Accuracy 0.75 ± 0.08; sensitivity 0.80 ± 0.08; specificity 0.71 ± 0.11	AI-assisted traumatic versus degenerative tear
Borzooei et al[62]	2024	Microwave sensing/ digital-twin data	SVM on simulated microwave scattering data	Rotator cuff/ supraspinatus tear or pathology	Numerically simulated shoulder models; no clinical patient cohort	100% classification accuracy in simulated experiments; no clinical validation	AI-assisted rotator cuff/ supraspinatus tear or pathology
Yao et al[63]	2022	Shoulder MRI	CNN, U-Net	Rotator cuff/ supraspinatus tear or pathology	200 retrospectively sampled MRI examinations	AUC 0.943; sensitivity 85%; specificity 85%; segmentation Dice reported	AI-assisted rotator cuff/ supraspinatus tear or pathology
Ni et al[64]	2024	Shoulder MRI	VGG16	Rotator cuff/ supraspinatus tear or pathology	3087 internal patients plus 573 external patients	VGG16 AUC 0.92; multitask classifier AUCs 0.97–0.99	AI-assisted rotator cuff/ supraspinatus tear or pathology
Jhan et al[65]	2026	Shoulder radiography	ResNet-50, InceptionV3	Rotator cuff/ supraspinatus tear or pathology	587 radiographs (406 train; 101 validation; 80 independent test)	Independent test AUROC 0.93; accuracy 81%; sensitivity 100%; specificity 63%	AI-assisted rotator cuff/ supraspinatus tear or pathology
Shim et al[66]	2020	Shoulder MRI	Voxception-ResNet 3D CNN	Rotator cuff/ supraspinatus tear or pathology	2124 patients	Binary accuracy 92.5%; sensitivity 0.92; specificity 0.86	AI-assisted rotator cuff/ supraspinatus tear or pathology
Saavedra et al[67]	2023	Shoulder MRI	VGG-19, ResNet-50	Supraspinatus fatty infiltration grade	606 shoulders (900 T2-weighted Y-view MRIs)	AUC 0.861–0.991 across Goutallier binary tasks	AI-assisted supraspinatus fatty infiltration grade
Saavedra et al[32]	2024	Shoulder MRI	VGG-19, U-Net	Supraspinatus fatty infiltration grade	606 shoulders (900 T2-weighted Y-view MRIs)	Segmentation Dice 0.9441 ± 0.0031; classification AUC 0.9903 ± 0.0092	AI-assisted supraspinatus fatty infiltration grade
Zhao et al[68]	2024	Shoulder radiography	Random forest and comparator ML models	Rotator cuff/ supraspinatus tear or pathology	89 osteoporosis patients (63 RCT; 26 without RCT)	RCT diagnosis AUC 0.95 on test set; tear-type test accuracy 73.684%	AI-assisted rotator cuff/ supraspinatus tear or pathology

3.3. Image reconstruction, segmentation, and quantitative imaging

AI-assisted decision-making in the diagnosis and treatment of shoulder joint disorders relies on tissue segmentation techniques based on DL and other AI technologies, which enable clinical applications such as automatic 3D reconstruction and quantitative analysis of shoulder joint tissues.^[24,30] This area also represents a key focus for the clinical application of AI-assisted RCTs.

Table 1 summarizes the characteristics of recent studies on shoulder joint segmentation.^[4,14–33] It is worth noting that, building on the segmentation of all rotator cuff muscles achieved by Riem et al^[69] the study conducted a quantitative analysis of the rotator cuff muscles in 170 healthy individuals using a proprietary algorithm. Through an objective analysis of the 3-dimensional volume of the rotator cuff muscles and the degree of fatty infiltration in patients of different ages and genders, the study revealed the relative relationship between the volume of the rotator cuff muscles and the degree of fatty infiltration and age-related change^[27,70,71]: The volume of the supraspinatus, teres minor, and infraspinatus muscles decreases with age; as people age, the relative contribution of the supraspinatus decreases, while that of the infraspinatus increases; the FI of the infraspinatus muscle increases with age. This study explored the clinical applications of DL-assisted medical image segmentation: specifically, the rapid quantification of image data from large patient cohorts: to reveal objective changes in tissue.

The research by Ro et al^[28] and Kim et al^[14] applied DL to the quantitative assessment of supraspinatus atrophy severity, as shown in Figure 2.^[28] Both studies were conducted using sagittal T1-weighted images. However, unlike Kim et al Ro et al performed automated measurements of fatty infiltration in the supraspinatus muscle using the Otsu thresholding method,^[28] based on the imaging characteristics of sagittal T1-weighted images. A comparison with the traditional Goutallier grading system revealed that this evaluation of fatty infiltration based on the Otsu threshold method differs statistically across Goutallier grades.^[28]

Figure 2.

Workflow for automated assessment of supraspinatus atrophy on sagittal oblique T1-weighted shoulder MRI. (A) Loading an MR image into ITK-SNAP for manual annotation. (B) Selection of the scapular Y-view and identification of the supraspinatus fossa and muscle (upper row), followed by manual annotation of the fossa (red) and muscle (yellow) (lower row). Arrowheads indicate adjacent vessels. (C) Architecture of the fully convolutional network used to segment the supraspinatus fossa and muscle. Adapted from Figures 3 and 4 of Ro et al^[28], Scientific Reports. 2021;11:15065, https://doi.org/10.1038/s41598-021-93026-w, under the Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0/). Panels have been rearranged and relabeled. MR = magnetic resonance; MRI = magnetic resonance imaging.

The reconstruction evidence suggested that DL denoising or reconstruction can shorten shoulder MRI acquisition while preserving interpretable image quality, but the evidence base was small and task-specific. Segmentation and quantitative imaging studies predominantly used U-Net-derived or other convolutional architectures to delineate rotator cuff muscles, tendons, fat, bone, or related shoulder structures. Dice coefficients and overlap measures were appropriate within segmentation tasks, whereas measurement agreement and absolute error were more informative for derived anatomical quantities. These estimates were not compared with diagnostic AUCs or predictive accuracy.^[72]

MRI dominated segmentation studies, reflecting its ability to depict muscle volume, tendon morphology, atrophy, and fatty infiltration.^[31,73] CT studies emphasized bone or muscle morphology, and ultrasound studies explored real-time scanning or tear-related tissue delineation.^[28–30,33] Several systems achieved high internal overlap, yet dataset size, annotation procedures, acquisition heterogeneity, and external validation were often limited.^[14–17] Consequently, automated 3-dimensional quantification is promising for reproducible phenotyping but remains dependent on scanner, sequence, annotation, and population generalizability.^[4,18–22]

3.4. Diagnosis and classification

Diagnostic and classification studies used MRI, radiography, ultrasound, clinical information, motion data, and multimodal combinations to detect tears, classify tendon involvement or severity, or support screening.^[18,44–47] Deep convolutional networks were common in imaging studies,^[74] whereas conventional machine learning methods and radiomics were more frequent when structured clinical features or engineered image descriptors were used.^[34–37] Reported AUC, sensitivity, specificity, NPV, and accuracy were summarized within their own metric strata in Table 2.^[32,34–68]

MRI studies generally targeted supraspinatus or multi-tendon tear detection and grading; radiographic studies evaluated indirect structural signatures or multimodal combinations; ultrasound studies addressed tear classification in operator-dependent images. Tear definitions ranged from any tear to partial- or full-thickness lesions and individual tendon involvement. Because thresholds, prevalence, reference standards, and validation sets differed, the review did not rank modalities or architectures by a single performance value. Apparent high internal discrimination should be interpreted as task-specific proof-of-concept rather than evidence of clinical equivalence to expert assessment.^[42]

3.5. Anatomical measurement, prognosis, and other applications

Automated measurement studies assessed clinically relevant quantities such as the critical shoulder angle, acromial index, scapular morphology, muscle volume, tear size, and fatty infiltration.^{[6,7,21,75–78]} These applications may reduce reader-dependent variation, but 2-dimensional positioning, segmentation accuracy, and reference-measurement reproducibility remain important sources of error.^[18,43,44]Nine studies addressed prognosis or outcome prediction, including retear, complications, clinically meaningful recovery, cost or resource use, and postoperative outcomes.^{[3,5,79–87]}

Models used clinical, imaging, operative, or multimodal predictors. Discrimination was sometimes favorable in internal validation; however, calibration, handling of missing data, predictor selection, sample size, and independent validation were frequently inadequate or incompletely reported.^[45,46] Other exploratory applications included unsupervised-phenotyping, rehabilitation-exercise recognition, motion assessment, and surgical-skill evaluation. These studies broaden the care pathway map but do not yet constitute a uniform evidence base for clinical deployment.^[3,6,78]

3.6. Narrative subgroup synthesis

By modality, MRI supported the broadest range of reconstruction, segmentation, diagnosis, and quantitative phenotyping tasks; radiography was used mainly for screening and automated morphology; ultrasound studies were fewer and faced acquisition-operator variability; and clinical or multimodal data supported prediction and interpretable decision support.^[19] By model family, convolutional and other DL systems predominated in image-based tasks, while conventional machine learning, boosting, and radiomics were more common for tabular clinical predictors or engineered features.^[4,48,88] No architecture could be declared superior across tasks because training sets, targets, and validation designs were not comparable.

By target phenotype, studies distinguished any RCT, tendon-specific lesions, partial-versus full-thickness tears, tear severity, muscle degeneration, and postoperative retear or recovery. Evidence for partial- versus full-thickness subgroups was too inconsistently reported for statistical subgroup pooling. The narrative comparison therefore identifies where evidence exists and where target definitions must be standardized before meaningful cross-study comparisons can be made.^[50,89,90]

3.7. Risk of bias and applicability

Among 36 diagnostic accuracy or classification studies assessed with QUADAS-2, overall risk of bias was high in 27/36, unclear in 6/36, and low in 3/36. Patient selection was high risk in 23 studies, unclear in 4, and low in 9; the index test was high risk in 19, unclear in 1, and low in 16; the reference standard was high risk in 14, unclear in 7, and low in 15; and flow and timing were high risk in 20, unclear in 4, and low in 12. Common concerns were convenience or case-control sampling, incomplete threshold prespecification, reference standards that were imperfect or insufficiently described, and incomplete or selectively analyzed participant flow. Detailed study-level QUADAS-2 judgments are provided in Supplementary Table S1, Supplemental Digital Content 2.

All 8/8 completed prediction-model studies assessed with PROBAST were at high overall risk of bias, and all had high risk in the analysis domain. Recurrent concerns included limited effective sample size, data-driven predictor selection, incomplete handling of missing data, internal validation without adequate optimism correction, insufficient calibration reporting, and absence of independent external validation. One protocol was not rated because no completed model results were available. These findings mean that favorable performance estimates are likely to be optimistic and that applicability to unselected routine care populations remains uncertain. Detailed study-level PROBAST judgments are provided in Supplementary Table S2, Supplemental Digital Content 3.

4. Discussion

4.1. Principal findings

This review maps 66 reports across the RCT care pathway and shows that AI research has moved beyond isolated tear detection to include acquisition, segmentation, quantitative phenotyping, anatomical measurement, prognosis, rehabilitation, and procedural assessment. The strongest recurring signal was proof-of-concept technical performance within internally defined tasks. The most important counter-signal was methodological fragility: QUADAS-2 judged 27/36 diagnostic studies at high overall risk, 6/36 unclear, and only 3/36 low, while PROBAST judged all 8/8 completed prediction models at high risk.

4.2. Interpretation by clinical task

Reconstruction and segmentation may improve efficiency and reproducibility before any diagnostic decision is made. Their value should be tested through acquisition robustness, annotation reliability, failure analysis, and the accuracy of clinically relevant derived measurements, not only image similarity or Dice overlap.^[9,91] Diagnostic models require representative sampling, prespecified thresholds, credible reference standards, and external testing.^[92] Prediction models require adequate events per predictor, missing data handling, calibration, optimism correction, and decision-curve or clinical-utility analysis. Treating these tasks as a single performance continuum obscures the distinct evidence required for each intended use.^[93]

4.3. Heterogeneity, metrics, and publication-bias

Extreme heterogeneity arose from modality, acquisition protocol, model architecture, tear definition, reference standard, prevalence, sampling frame, validation design, and performance reporting.^[10] AUC describes discrimination across thresholds; sensitivity and specificity describe threshold-dependent classification; NPV varies with prevalence; accuracy can be distorted by class imbalance; Dice measures spatial overlap; and calibration assesses agreement between predicted and observed risk.^[94] These quantities cannot be substituted for 1 another. For this reason, we stratified the narrative and tables by task and retained each metric in its original meaning.^[95]

The absence of a coherent common effect measure precluded defensible meta-analysis, statistical heterogeneity testing, meta-regression, funnel plots, or formal publication-bias tests. This does not imply absence of reporting bias. The predominance of small retrospective development studies, favorable internal results, selective presentation of the best-performing model, and scarcity of negative external validations suggest that the published literature may over represent optimistic performance. Readers should therefore interpret numerical maxima as hypothesis-generating rather than as pooled estimates of expected clinical performance.

4.4. Bias implications and clinical translation

Selection bias may inflate discrimination when clear cases and controls are sampled instead of consecutive patients. Index test bias can arise when thresholds or model versions are selected after viewing test performance.^[12] Reference standard bias is particularly important when labels depend on a single report, non-blinded review, or heterogeneous surgical and imaging criteria. Flow and timing bias can result from exclusions after model evaluation or from incomplete application of the reference standard. In prediction studies, analysis bias can produce optimistic estimates through small event counts, inadequate resampling, leakage between development and test sets, or omission of calibration.

Accordingly, the current evidence supports continued development and prospective evaluation, not autonomous replacement of clinical judgment. Clinically credible systems should specify their intended use population and workflow position, report failure cases and uncertainty, undergo independent external and temporal validation, assess calibration and net benefit where relevant, and be evaluated for effects on clinician decisions, time, patient outcomes, and equity.^[96,97]

4.5. Strengths and limitations of this review

Strengths include searches across 5 major databases, explicit selection counts, classification of all 66 reports, study-level tables, tool-specific risk-of-bias assessment, and a synthesis that respects the nonequivalence of performance metrics. The review also separates diagnostic accuracy from clinical prediction and does not apply QUADAS-2 or PROBAST outside their intended domains.

Limitations remain. One non-English record was excluded, 139 reports could not be retrieved, and gray-literature searching was not comprehensive, so relevant evidence may have been missed. Study reporting was frequently incomplete, and some classifications relied on the available report rather than author-supplied clarifications. No meta-analysis was performed because comparable effect estimates were insufficient. Formal publication-bias tests were therefore unavailable, and qualitative inference about selective reporting remains uncertain. Finally, rapid methodological evolution means that this evidence map will require updating.

4.6. Future directions for clinical translation

Future research on the intersection of AI and RCTs will be grounded in real-world clinical settings and practical workflows, with a focus on developing capabilities for whole-organ diagnosis and advanced classification.^[18] In this process, it is necessary to conduct in-depth research into the development and clinical validation of AI-assisted tools to fully understand how the diagnostic performance and efficiency gains demonstrated by AI algorithms in isolated research settings can be successfully translated into applications in routine clinical practice.^[76] Such research will provide a more comprehensive demonstration of the feasibility of AI technology in real-world clinical settings, thereby further facilitating its smooth transition into clinical practice.

5. Conclusion

AI has demonstrated significant application potential in multiple aspects of RCT management, including image acquisition, tissue quantification, automated diagnosis, anatomical parameter measurement, and postoperative prognosis prediction. It improves the efficiency of imaging evaluation, boosts diagnostic accuracy, and increases the objectivity of clinical decision-making.

However, current evidence primarily stems from retrospective, single-center studies with internal validation. Significant gaps remain in external validation, real-world clinical robustness, 3-dimensional automated measurement, and continuous quantitative assessment. Across 66 included reports, AI showed proof-of-concept potential in rotator cuff image reconstruction, segmentation and quantification, diagnosis, anatomical measurement, prognosis, and rehabilitation-related assessment. Even so, the QUADAS-2 tool rated 27/36 diagnostic studies as having high overall risk, 6/36 as having unclear risk, and only 3/36 as having low overall risk; the PROBAST tool rated all 8/8 completed prediction studies as having high overall risk.

Consequently, AI is not yet mature enough to widely replace clinical judgment. Multicenter, prospective, standardized studies are needed to advance its development from algorithmic performance validation to practical clinical implementation.

Author contributions

Data curation: Haitao Wang.

Funding acquisition: Liyan Zhang.

Project administration: Liyan Zhang.

Supervision: Liyan Zhang.

Software: Deming Guo.

Visualization: Deming Guo, Yingqi Gao, Weixin Xian.

Formal analysis: Yingqi Gao, Weixin Xian.

Writing – original draft: Haitao Wang.

Writing – review & editing: Deming Guo.