Artificial intelligence revolution in shoulder magnetic resonance imaging: current evidence and future directions for rotator cuff diagnosis

Deep learning models for RCT detection and classification

Research over the past 5 years has produced a variety of AI models aimed at detecting RCTs on MRI [9,13,16]. The majority are deep learning models using CNN architectures trained on shoulder MRI datasets with labels for the presence or absence of tears (and in some cases tear type) [3,9,12]. Overall, these studies report high diagnostic performance, but with considerable heterogeneity depending on dataset size, tear type (partial vs. full), and validation strategy (internal vs. external). In a recent systematic review of MRI-based AI for RCTs (covering 2020–2024), Longo et al. [13] found that deep learning models achieved accuracies ranging from 71% to 100% for tear identification. Sensitivities and specificities were similarly impressive, both often in the 70%–100% range. Importantly, in the subset of studies that directly compared AI predictions to clinician interpretations, AI models performed broadly comparable to musculoskeletal radiologists, but external validation revealed lower performance by 5%–15% compared to internal testing, underscoring concerns regarding generalizability [12,13]. This suggests that state-of-the-art algorithms can approach the diagnostic capability of experienced musculoskeletal radiologists or orthopedic surgeons in detecting RCTs on MRI [9,12,13].

Several specific examples illustrate these capabilities. Shim et al. [17] developed a three-dimensional (3D) CNN for automatic classification of rotator cuff status (normal vs. partial tear vs. full-thickness tear) using shoulder MRI. The model showed an overall accuracy of 92.5% for binary tear detection and performed well in distinguishing full-thickness tears. However, sensitivity for partial tears was lower (approximately 72.5%) compared to the 100% for full-thickness tears, reflecting the clinical challenge of detecting subtle partial defects [17]. Yao et al. [18] similarly reported excellent performance for supraspinatus tear detection (their deep CNN achieved 94.7% sensitivity and 97.5% specificity) but noted challenges in correctly classifying low-grade partial tears, which were sometimes confused with tendinopathy. These findings emphasize the consistent difficulty of partial tear classification in both clinical and AI settings [17,18]. Lee et al. [3] developed a YOLOv8-based model to screen for RCTs across all three standard MRI planes (axial, coronal, sagittal). Their model achieved an AUC of 0.94 for detecting cuff tears when using all imaging planes together, significantly higher than using any single plane alone (AUC, approximately 0.70 on sagittal or coronal only). The all-plane model was extremely sensitive (98%) in identifying tears, highlighting the advantage of 3D contextual information. This supports the principle that multi-planar MRI evaluation improves diagnostic accuracy, a concept confirmed quantitatively by AI.

Another notable trend is the recurrence of certain network architectures. In the systematic review, the authors found that VGG-based CNNs were the most widely utilized AI models for RCT diagnosis. In fact, three groups that achieved near-perfect accuracy (98%–100%) all employed VGG-based networks. For example, Ni et al. [19] applied a VGG19-based model on a large dataset of more than 3,000 shoulder MRI cases, reporting 98% accuracy, 96% sensitivity, and 93% specificity in detecting supraspinatus tears [13]. Similarly, Key et al. [20] and Ro et al. [21] each demonstrated approximately 99%–100% accuracy with VGG-based architectures, though their results were derived from single-institution datasets with internal validation only, raising questions about potential overfitting [13]. That said, other architectures have also been explored: ResNet-50 and Inception-V3 have shown strong results for related tasks like classifying rotator cuff muscle fatty infiltration, and specialized detector networks (YOLO, U-Net variants) have been used for localization tasks [9,16,21,22]. Emerging interest in Vision Transformers (ViTs) and hybrid models has been noted, but published evidence in shoulder MRI remains scarce [2,13]. Overall, the current body of work indicates that deep learning can accurately identify RCTs on MRI, especially full-thickness tears, with performance approaching an AUC of 0.9–0.98 in many studies [2,5,9,11,13]. Longo et al. [13] reported that the best models achieved 94%–100% sensitivity and 93%–100% specificity for tear detection—an encouraging level for potential clinical utility. While these values are highly encouraging, they should be interpreted cautiously given that most studies were retrospective, from a single center, and relied primarily on internal validation. External validation was performed in only approximately 20% of the included studies, with reported performance decreases of 5%–15% compared to internal testing, underscoring concerns about generalizability.

Importantly, the evidence base consists primarily of retrospective diagnostic studies (level III), frequently from single institutions and using internal validation, which limits generalizability [13]. These studies generally used one of two reference standards for labeling ground truth: expert readings or surgical findings [5,13]. All studies in the systematic review did establish some form of ground truth for training and testing. The majority (eight studies) had musculoskeletal radiologists provide or verify the labels, and six studies used an orthopedic surgeon assessment. Notably, only five studies (of 19 reviewed) used intra-operative arthroscopy findings as the ground truth for RCTs [13]. Because arthroscopy is the clinical gold standard, reliance on radiologist-based labels introduces potential bias and variability, which can propagate into model training. This heterogeneity in reference standards highlights the need for standardized labeling protocols and more consistent use of surgical confirmation where feasible [6,13]. Encouragingly, in those studies that did compare AI output to arthroscopy, performance remained high (for example, a model by Hahn et al. [6] detected biceps tendon tears with 95% accuracy against arthroscopic findings). However, such examples are exceptions rather than the rule. Most published work has not undergone prospective validation or multicenter testing, and external validations remain scarce. When performed, external validation typically revealed modest but important declines in performance (5%–15%), reflecting the challenges of applying AI across diverse scanners and patient populations [2,13,15]. A few studies have performed external tests – for instance, Cui et al. [23] externally validated a radiomics-based model on a separate cohort, but this is not yet routine [15].

Radiomics and hybrid approaches

In parallel with deep learning, some researchers have explored radiomics and classical machine learning for rotator cuff diagnosis [2,24]. Radiomics refers to extraction of a large number of quantitative features from images (such as texture, intensity histogram, shape descriptors) and using those features to build predictive models. This approach can capture subtle patterns that might not be apparent visually [24]. In the shoulder, radiomics has been applied to both tendon and muscle analysis. Cui et al. [23] recently reported a shoulder MRI-based radiomics model for diagnosing supraspinatus tendon tears and grading their severity on MRI. They extracted dozens of texture features from the tendon region on T2 MRI and trained a classifier to distinguish no tear, partial tear, and full-thickness tear. According to their report, the radiomics model showed good diagnostic consistency and could stratify partial tears by size (mild vs. severe partial) in correlation with surgical findings. This type of severity staging might add clinical value, since partial tears exist on a spectrum and management decisions often hinge on tear depth [2,25].

Another area of radiomics application is muscle degeneration secondary to chronic cuff tears. Fatty infiltration and atrophy of rotator cuff muscles (like the supraspinatus and infraspinatus) are key prognostic factors for surgical outcomes. Traditional measurement of these requires manual grading (Goutallier classification) on MRI slices [21,26]. AI can assist here: several teams have built models to automatically segment rotator cuff muscles and quantify fat content [15,21,26]. For example, another study developed a deep learning U-Net to segment the rotator cuff muscles on axial MRI, achieving mean Dice scores up to 0.93–0.95 for supraspinatus muscle delineation [15]. Using such segmentations, one can compute the muscle volume or fat fraction [15,21]. Saavedra et al. [16] trained CNN models (including VGG-19 and ResNet-50) to classify supraspinatus fatty infiltration grades; notably, their best model (VGG-19) reached 97.3% accuracy for distinguishing high vs. low fatty infiltration, demonstrating the feasibility of AI-driven muscle quality assessment. These approaches illustrate that radiomics and deep learning can complement each other—radiomics offering interpretability and deep learning providing high performance—supporting a potential hybrid strategy for comprehensive evaluation of both tendons and muscles [2,15,21,24].

At present, there are fewer radiomics studies in shoulder MRI than pure deep learning studies, but they offer some explainability advantages. The handcrafted features can sometimes be mapped to known tissue properties, and simpler models (like logistic regression on radiomic features) can be easier to interpret [2,24]. Nevertheless, radiomics models are highly sensitive to MRI acquisition protocols (e.g., field strength, sequence type, slice thickness), and lack of standardization increases risk of overfitting, especially without external validation [24]. A “black box” deep CNN might outperform radiomics given enough data, but radiomics might perform better on small datasets by incorporating prior knowledge [2,24]. Thus, radiomics should not be seen as a competitor to deep learning, but as a complementary approach that can improve robustness in limited-data scenarios. The current evidence suggests that both approaches are viable, and some studies have used them in combination [2,24,25]. Kim et al. [11], for example, used a weighted combination of deep learning outputs and radiomic features to diagnose RCTs, aiming to harness the strengths of each method. As AI in shoulder MRI progresses, hybrid models that combine segmentation, radiomics, and deep CNN analysis are likely to play an increasing role, moving toward multi-task systems capable of detecting tears, quantifying secondary changes, and predicting outcomes within a single framework [2,15,21,24,26].

LLM-based reporting and decision support

While most AI efforts for shoulder MRI have centered on image analysis, there is a parallel interest in applying natural language AI to musculoskeletal care. LLMs like ChatGPT have captured public imagination with their ability to generate human-like text, and researchers are beginning to examine their utility in medicine. One application relevant to radiology is AI-generated reporting – using models to draft a clinical report from either images or structured inputs (findings).

At present, LLM-driven report generation for shoulder MRI remains conceptual and largely unvalidated, though analogous systems have been tested in other imaging fields [2,27]. Current evidence has instead tested LLMs on ancillary tasks around RCT care—such as patient education or clinical decision support—rather than direct image reading [2,27]. For instance, Kolac et al. [2] tested four popular LLMs (ChatGPT-3.5, GPT-4, Google Gemini, and Microsoft Bing’s AI) using 50 frequently asked questions about RCTs. The models produced responses that were generally readable and mostly correct, although approximately 40% were incomplete or inaccurate. The median DISCERN score (a validated measure of health information quality) was only 40/80, with 56% of responses rated “sufficient” in quality. This indicates that current LLMs provide understandable language (median readability 83%) but variable informational reliability, underscoring the need for expert oversight [2,27]. Another study explored ChatGPT-4’s ability to interpret postoperative clinical images and answer patient questions after rotator cuff repair. Orthopedic clinicians rated 80% of its responses as satisfactory or better in clarity and correctness, but limitations in nuance and safety were noted. The conclusion was that LLMs can ease communication but are not yet reliable for independent medical advice [28].

Across available evaluations, LLM performance in musculoskeletal contexts shows the following: readability, high (8th-grade level; 80%–85%); informational reliability: moderate (50%–60% sufficient quality); clinical safety: dependent on expert oversight; not diagnostic. Overall, the literature suggests that current LLM applications in shoulder MRI are exploratory and supportive rather than diagnostic. Their most immediate potential lies in drafting radiology reports, summarizing literature, or generating patient education material, tasks where their readability is valuable but where accuracy must be monitored. Future directions include fine-tuning LLMs on orthopedic datasets, coupling them with vision models for multimodal reporting, and developing evaluation frameworks to ensure factual correctness. Until such advances occur, LLM use in musculoskeletal imaging should be considered assistive and speculative, always requiring human verification [2,27].

Real-world implementation: regulatory-approved tools and clinical adoption

As of 2025, we are witnessing the first waves of AI tools reaching clinical practice in imaging [2,27]. In the shoulder MRI domain, the initial implementations have focused on image acquisition and reconstruction rather than autonomous diagnosis [6,27]. A prime example is the use of deep learning to accelerate MRI scans and improve image quality. Major MRI manufacturers have integrated AI-based reconstruction algorithms into their scanners, allowing much faster scans without sacrificing diagnostic detail. For shoulder MRI, this is a significant benefit; conventionally, a shoulder exam might take 15–20 minutes, but with AI acceleration, high-quality exams can be performed in 5–7 minutes [6]. This reduces patient discomfort and motion artifacts while increasing scanner throughput [6,27].

A recent prospective study validated a 7-minute shoulder MRI protocol enhanced by Siemens Healthineers’ Deep Resolve Sharp (Food and Drug Administration [FDA] 510(k) K213693, updated K232322). In 121 patients, ultrafast deep learning-based scans were compared with standard images and correlated with arthroscopic findings. Seven musculoskeletal radiologists reported significantly reduced noise and sharper anatomy, with no loss of diagnostic performance for rotator cuff pathology [27]. Similarly, GE Healthcare’s AIR Recon deep learning (FDA 510(k) K201233) and Philips’ SmartSpeed Precise (FDA 510(k) K250123) have received regulatory clearance and are now widely deployed. These products have independent peer-reviewed validation showing improved signal-to-noise ratio and preserved diagnostic accuracy, supporting their integration into daily clinical MRI practice [6,27]. Beyond acceleration, we are starting to see AI tools for automated analysis in the musculoskeletal realm, though regulatory approvals are limited. To date, no fully automated RCT diagnostic software has received FDA clearance (database search verified as of September 30, 2025) [6]. However, related products targeting muscle quantification have emerged. One example is Springbok Analytics’ “MuscleView,” an AI platform that analyzes muscle health from MRI scans. In 2024, MuscleView received FDA 510(k) clearance (K241017) for providing 3D visualizations and quantitative metrics of muscle volume and fatty infiltration. Independent validation demonstrated accurate supraspinatus fat quantification compared with manual segmentation, supporting its utility in chronic cuff tear assessment. Such quantitative outputs can complement radiologist reports and aid surgical planning [2,6].

Table 1 summarizes the regulatory-cleared commercial AI tools relevant to shoulder MRI, including product names, developers, applications, validation evidence, and official regulatory identifiers [6,27]. While most focus on image quality enhancement, they represent the first step toward broader adoption of diagnostic AI. Tools designed specifically for automated cuff tear detection remain at the research stage, and their clinical adoption will depend on multicenter validation, workflow integration, and regulatory approval [2,6].

Table 1.

Commercial AI software for shoulder MRI and related applications [2,3,6,13,27]

Technical perspective: datasets, algorithms, and challenges in shoulder MRI AI

Data characteristics are a fundamental consideration in developing AI for shoulder MRI. Published datasets vary widely in size, from fewer than 100 MRIs to several thousand cases, and many are retrospective, single-institution cohorts prone to imbalance (e.g., more tear-positive than -negative cases) [2,7,15]. Ground-truth labeling is heterogeneous across studies: some rely on musculoskeletal radiologist consensus, while others use surgical confirmation. Only a minority incorporate intraoperative arthroscopy as the reference standard, which is considered the gold standard for RCT diagnosis. This variability can introduce bias—models trained solely on radiologist labels can only replicate human-level subjectivity rather than true surgical findings [5,6].

Annotation quality and protocol standardization are critical. Best practices include adjudication by multiple experts or semi-automated labeling to reduce interobserver variability. However, detailed annotation (tear margins, retraction length, muscle segmentation) is time-consuming, and some studies use indirect labeling strategies (assigning “tear” or “no tear” per MRI exam without slice-level annotation), which can introduce label noise [7,15].

Algorithmic trends have evolved from early 2D CNNs toward 3D CNNs and multi-planar ensemble networks, which better leverage volumetric information from MRI. For example, 3D ResNet-based models can integrate context across slices, though they demand larger datasets and higher computational resources. Transfer learning from ImageNet or similar large datasets is widely employed to overcome limited medical data. Attention mechanisms and transformer-based architectures are emerging, offering potential to capture long-range dependencies in MRI sequences [7,15].

Multi-task learning is an emerging trend: instead of training separate models for tear detection, segmentation, or muscle analysis, one shared encoder can power multiple tasks simultaneously (e.g., “RC-MTL” models for classification and segmentation). This strategy both enhances feature learning and produces clinically useful composite outputs (tear probability, size, segmentation map) [2,7,15].

Domain shift remains a key challenge. MRI protocols differ by institution (T1 vs. T2-weighted, fat suppression vs. non–fat suppression, 1.5T vs. 3T field strength), and models often show performance decrease when tested externally. Data harmonization, augmentation, or federated learning approaches have been suggested to mitigate these effects [2,7,15,17]. Explainability and clinician trust are essential for adoption. Many models incorporate saliency maps (e.g., Grad-CAM) to highlight the tendon region influencing AI decisions. These often align with clinical reasoning but occasionally expose “shortcut learning” (e.g., focus on the acromion instead of the tendon). Combining segmentation with classification can enforce anatomically grounded predictions [7,15,17].

Computational efficiency is another practical consideration. While modern GPUs allow CNN-based MRI inference in less than one minute, lighter models (e.g., MobileNet) are being explored for deployment on scanner consoles or PACS workstations. Seamless AI models without delayed workflow will be critical for clinical adoption [2,7,15].

Validation and generalizability

Most current AI studies remain retrospective and from a single center. Future research must include multi-center datasets, diverse MRI scanners, and standardized imaging protocols to ensure generalizability across clinical environments [9,12,13]. External validation and prospective clinical trials are also necessary to evaluate whether AI integration improves diagnostic accuracy, treatment planning, and surgical outcomes [9,12,13]. Such validation will determine not only the performance of AI but also its safety, reproducibility, and medico-legal accountability.

Workflow integration

AI tools should be integrated into clinical workflows in measurable, outcome-oriented ways. This includes automated case triage, pre-populated quantitative tear measurements, and preliminary report drafting that shortens interpretation time [2,7,15,25]. Pilot studies should evaluate how these tools affect radiologist efficiency, interobserver consistency, and surgical planning turnaround times. The most practical path forward involves AI use as an assistive “co-pilot” within existing PACS and reporting systems, rather than as a standalone diagnostic entity.

Ethical and regulatory oversight

Establishing robust ethical and regulatory frameworks is essential. Clear standards for algorithm transparency, clinician accountability, and patient consent are required before broad clinical use [2,7,15,25]. AI systems must be validated for fairness, demonstrating equivalent performance across patient demographics to prevent hidden bias. Regulatory agencies and professional societies should also provide explicit guidelines for AI validation and reporting.

Explainability and clinical adoption

For successful clinical adoption, AI systems must provide interpretable outputs, such as tear-probability maps or segmentation overlays, and be accompanied by clinician training in AI literacy [2,7,15,25]. Transparent visualization fosters trust and allows radiologists and surgeons to validate AI-generated insights within their own clinical reasoning.

Multimodal and patient-centered directions

Future development should also explore multimodal AI models that integrate MRI with clinical and functional data (e.g., age, activity level, pain scores) to provide personalized treatment recommendations [2,7,15,25]. At the same time, patient-facing AI applications, such as fine-tuned orthopedic LLMs that generate plain-language summaries of MRI findings, can improve communication and health literacy—but only when used under expert supervision [6,27].

Collaboration as the clinical endpoint

Ultimately, the most sustainable future for AI in shoulder MRI involves collaboration. AI should augment—not replace—radiologists and surgeons by providing quantitative tear metrics, highlighting key findings, and generating draft reports while maintaining human oversight [2,7,15,25]. This balanced human–AI partnership, supported by strong validation, seamless workflow integration, and ethical governance, represents the most realistic path for translating AI into daily clinical practice.