Harnessing Artificial Intelligence for Shoulder Ultrasonography: A Narrative Review.

Harnessing Artificial Intelligence for Shoulder Ultrasonography: A Narrative Review - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice J Imaging Inform Med . 2025 Sep 12;39(3):2228–2238. doi: 10.1007/s10278-025-01661-w Harnessing Artificial Intelligence for Shoulder Ultrasonography: A Narrative Review Wei-Ting Wu 1, 2 , Yi-Chung Shu 3 , Che-Yu Lin 3 , Consuelo B Gonzalez-Suarez 4, 5 , Levent Özçakar 6 , Ke-Vin Chang 1, 2, 7, ✉ Author information Article notes Copyright and License information 1 Department of Physical Medicine and Rehabilitation and Community and Geriatric, National Taiwan University Hospital, Bei-Hu Branch, Taipei, Taiwan 2 Department of Physical Medicine and Rehabilitation, National Taiwan University Hospital, College of Medicine, Taipei, Taiwan 3 Institute of Applied Mechanics, College of Engineering, National Taiwan University, Taipei, Taiwan 4 Research Center for Health Science, University of Santo Tomas, Manila, Philippines 5 Department of Physical Medicine and Rehabilitation, Our Lady of Lourdes Hospital, Manila, Philippines 6 Department of Physical and Rehabilitation Medicine, Hacettepe University Medical School, Ankara, Turkey 7 Center for Regional Anesthesia and Pain Medicine, Wang-Fang Hospital, Taipei Medical University, Taipei, Taiwan ✉ Corresponding author. Received 2025 Mar 18; Revised 2025 Jul 23; Accepted 2025 Aug 20; Collection date 2026 Jun. © The Author(s) under exclusive licence to Society for Imaging Informatics in Medicine 2025. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. PMC Copyright notice PMCID: PMC13230411  PMID: 40940587 Abstract Shoulder pain is a common musculoskeletal complaint requiring accurate imaging for diagnosis and management. Ultrasound is favored for its accessibility, dynamic imaging, and high-resolution soft tissue visualization. However, its operator dependency and variability in interpretation present challenges. Recent advancements in artificial intelligence (AI), particularly deep learning algorithms like convolutional neural networks, offer promising applications in musculoskeletal imaging, enhancing diagnostic accuracy and efficiency. This narrative review explores AI integration in shoulder ultrasound, emphasizing automated pathology detection, image segmentation, and outcome prediction. Deep learning models have demonstrated high accuracy in grading bicipital peritendinous effusion and discriminating rotator cuff tendon tears, while machine learning techniques have shown efficacy in predicting the success of ultrasound-guided percutaneous irrigation for rotator cuff calcification. AI-powered segmentation models have improved anatomical delineation; however, despite these advancements, challenges remain, including the need for large, well-annotated datasets, model generalizability across diverse populations, and clinical validation. Future research should optimize AI algorithms for real-time applications, integrate multimodal imaging, and enhance clinician-AI collaboration. Keywords: Rotator cuff, Ultrasound, Deep learning, Machine learning, Neural networks Introduction The integration of artificial intelligence (AI) into medical imaging has significantly enhanced diagnostic accuracy, consistency, and workflow efficiency. Among imaging modalities, ultrasound has especially benefited from AI advancements due to its real-time imaging capability and operator-dependent nature. In musculoskeletal medicine, the shoulder is a frequent site of pathology, where ultrasound plays a vital role in both diagnosis and intervention. Shoulder pain is one of the most common musculoskeletal complaints, affecting people across all age groups and substantially reducing productivity and quality of life. A systematic review by Luime et al. [ 1 ] reported a lifetime prevalence ranging from 6.7% to 66.7%. The causes are diverse—ranging from rotator cuff disorders and bursitis to adhesive capsulitis and degenerative joint disease [ 2 ]—all of which necessitate accurate imaging for effective treatment. Ultrasound has emerged as a preferred modality for shoulder evaluation due to its accessibility, dynamic imaging capability, and utility in guiding real-time interventions [ 3 , 4 ]. Compared to magnetic resonance imaging, ultrasound offers several practical advantages, including lower cost, portability, widespread availability, and real-time assessment. These features are particularly valuable in routine clinical settings or regions with limited access to advanced imaging. Furthermore, unlike magnetic resonance imaging, ultrasound allows bedside evaluation and can be immediately integrated into interventional procedures. Recent meta-analyses [ 5 , 6 ] have confirmed that ultrasound achieves diagnostic accuracy comparable to that of magnetic resonance imaging in detecting various shoulder pathologies. However, ultrasound remains highly operator-dependent, and diagnostic variability is a persistent challenge—especially given the complexity of musculoskeletal assessments and the wide spectrum of potential pathologies. AI, particularly deep learning techniques such as convolutional neural networks (CNNs), offers a promising solution by automating tasks like detection, classification, and segmentation [ 7 , 8 ]. These models have demonstrated usefulness in shoulder ultrasound, including identifying rotator cuff tears, quantifying tendon abnormalities, and delineating anatomical structures. Despite the rapid growth of AI applications in musculoskeletal ultrasound, no prior review has specifically examined its use in shoulder sonography. This narrative review addresses that gap by exploring AI-based approaches for shoulder ultrasound, highlighting recent advancements, clinical applications, and future directions in this evolving field. Strategy for Literature Search Although this article is a narrative review, a systematic search of electronic databases was conducted to identify relevant literature. Studies that utilized ultrasound to assess the shoulder joint, with images analyzed or processed by AI, were retrieved from PubMed, Scopus, and Web of Science. The search terms included “shoulder,” “rotator cuff,” “ultrasound,” “ultrasonography,” “artificial intelligence,” “machine learning,” “deep learning,” “convolutional neural network,” and “CNN.” The search period extended from the earliest available date to February 28, 2025. The inclusion criteria were:(1) original studies applying ultrasound and AI-related technology to imaging of the shoulder tendon, joint, or ligament, and (2) studies involving human participants. The exclusion criteria were (1) case reports or case series, (2) studies not conducted in vivo, (3) imaging approaches using plain radiography, computed tomography, or magnetic resonance imaging, and (4) studies performed on animal subjects. A total of nine studies were included, with their methodology and main findings summarized in Table 1 . Table 1. Comparative insights of the included studies Study Model(s) used Task Study population/demographics Comparative insight Lin et al. (2020) Faster R-CNN, CNN variants BPE classification Ultrasound images of BPE from an unspecified number of patients; focus on classification difficulty for inexperienced physicians Faster R-CNN used for ROI extraction; achieved 75% accuracy for three-class BPE classification using CNN architectures; emphasized the effect of dataset size and spatial context on ROI identification Triantafyllou et al. (2024) Random Forest, XGBoost, SVM Outcome prediction after calcification removal 84 patients undergoing US-PICT for calcific tendinopathy; external validation cohort included Combined radiomic and clinical features (e.g., max calcification diameter); Random Forest model achieved AUC of 0.88 (sensitivity: 0.90, PPV: 0.92); external testing AUC of 0.78 Tang et al. (2024) ResNet50, Mask R-CNN Standard plane recognition, segmentation 59,265 ultrasound images used for training; 1886 images used for segmentation testing Mask R-CNN achieved average IoU of 96.2% for tissue segmentation; ResNet50-based standard plane model reached 94.9% accuracy, with precision 96.4%, recall 95.4% and F1-score 95.9%; strong performance across all standard scanning planes Chiu et al. (2022) DenseNet-121 SSCT classification 2462 labeled images from 7836 patients evaluated by 18 physiatrists; focused on SSCT Longitudinal-transverse DenseNet-121 model reached 91.32% accuracy (sensitivity: 87.89%, specificity: 94.74%) for SSCT detection Lee et al. (2021) SMART-CA + IPLF RCT segmentation RCT images with and without speckle noise; dataset size unspecified; evaluated SMART-CA performance SMART-CA with IPLF improved precision (0.604), recall (0.942), Dice coefficient (0.736) in noisy images; outperformed state-of-the-art models; optimized for imbalanced datasets Ho et al. (2022) VGG19, InceptionV3, Xception, ResNet50, DenseNet121 RCT classification 194 ultrasound images from patients with RCTs; five-fold cross-validation applied DenseNet121 showed best performance among five CNNs (accuracy: 88.2%, sensitivity: 93.8%, specificity: 83.6%, and AUC: 0.832); Grad-CAM used to visualize key image features Wu et al. (2025) YOLOv7, YOLOv7-CBAM Real-time RCT detection 840 ultrasound images from 280 patients (equal number with torn and intact tendons); confirmed by MRI YOLOv7-CBAM model reached accuracy of 0.98, sensitivity of 0.98, F1-score of 0.980 and mAP of 0.993; improved diagnostic accuracy (88.9%) and interobserver reliability (0.71) Yu et al. (2023) Neural network + clustering Periarthritis pain staging 165 patients with scapulohumeral periarthritis; pain severity stratified by VAS Proposed clustering algorithm achieved accuracy of 0.833, sensitivity of 0.872 and specificity of 0.801; linked US-based capsule changes with pain stages; clinical factors (e.g., job strain) also correlated with pain severity Shu et al. (2023) CNN, STL-CNN, STL-CNN + AE Motion tracking in dynamic imaging 17 participants performed abduction/adduction for dynamic ultrasound motion tracking; cyclic task analyzed STL-CNN achieved MAE of 0.002–0.007 cm for minimum acromiohumeral distance; outperformed standard CNN (MAE: 0.081–0.333 cm); best for precise tracking of anatomical landmarks in dynamic motion analysis Open in a new tab Abbreviation: AE autoencoder, AI artificial intelligence, AUC area under the curve, BPE bicipital peritendinous effusion, CBAM Convolutional Block Attention Module, CNN convolutional neural network, DL deep learning, FCN fully convolutional network, F1-score F1 score (harmonic mean of precision and recall), Grad-CAM Gradient-weighted Class Activation Mapping, IoU intersection over union, IPLF integrated on positive loss function, MAE mean absolute error, mAP mean average precision, MRI magnetic resonance imaging, PPV positive predictive value, RCT rotator cuff tear, ROI region of interest, Seg. segmentation, SST supraspinatus tendon, SSCT supraspinatus calcific tendinopathy, STL-CNN self-transfer learning convolutional neural network, US ultrasound, US-PICT ultrasound-guided percutaneous irrigation of calcific tendinopathy, VAS visual analogue scale, YOLO You Only Look Once (object detection algorithm) Deep Learning for Grading Bicipital Peritendinous Effusion Bicipital peritendinous effusion (BPE) is the most common abnormality affecting the biceps tendon and is associated with a range of shoulder pathologies. Lin et al. (2020) [ 9 ] developed an automated system to detect and classify BPE using ultrasound imaging. The system utilized two CNN models: CNN-1 for detecting the presence of the biceps tendon and CNN-2 for classifying the severity of BPE (Fig. 1 ). The workflow began with a Faster Region-Based CNN (Faster R-CNN) to identify the region of interest. The proposed system achieved a classification accuracy of 75% across four severity levels: normal (< 1 mm), mild (1–2 mm), moderate (2–3 mm), and severe (> 3 mm). However, the model demonstrated limited discriminative performance between the normal and mild categories. Combining these two groups into a single class may improve diagnostic performance. Despite this limitation, the system shows promising potential to support clinicians in assessing BPE severity and distinguishing physiological findings from cases that may require further intervention. Fig. 1. Open in a new tab Architecture of CNN-1 for determining whether the figure contains the biceps long head tendon and CNN-2 for categorizing the severity of biceps peritendinous effusion. This figure was independently created by the authors based on the conceptual framework described in Lin et al. (2020), without reproducing any copyrighted material Machine Learning for Predicting Ultrasound-Guided Percutaneous Irrigation Success Calcific tendinopathy is a common source of shoulder pain, most frequently affecting the rotator cuff—particularly the supraspinatus tendon [ 10 ]. Triantafyllou et al. (2024) [ 11 ] investigated the use of ultrasound-guided percutaneous irrigation for treating rotator cuff calcifications in a cohort of 84 patients. The procedure involved the insertion of an 18-gauge needle connected to a syringe filled with 4 mL of 0.9% sodium chloride. Manual segmentation of the calcifications was performed using 3D Slicer software, outlining the anterior half of each lesion to create semilunar segmentations. Radiomic features predictive of treatment outcomes were extracted using the Least Absolute Shrinkage and Selection Operator (LASSO) method. Machine learning models were then applied to radiomic, clinical, and combined datasets, integrating key ultrasound-derived features with clinical variables such as the maximum diameter of calcification (Fig. 2 ). The top-performing model achieved an area under the curve (AUC) of 0.88 (95% CI: 0.73–0.99), with a positive predictive value of 0.92 and sensitivity of 0.90. These findings indicate that a radiomics-based machine learning model may be a valuable tool for predicting the outcomes of ultrasound-guided percutaneous irrigation. Nonetheless, as a retrospective study, the design limits the ability to control for confounding factors that may affect treatment response. Fig. 2. Open in a new tab Illustration of combining radiomic and clinical data to construct machine learning-based classification models for predicting the success of ultrasound-guided percutaneous irrigation in shoulder calcific tendinopathy. This figure was independently created by the authors based on the conceptual framework described in Triantafyllou et al. (2024), without reproducing any copyrighted material Deep Learning for Segmenting Shoulder Joint Sonographic Anatomy Shoulder joint ultrasonography presents specific challenges that require a thorough understanding of the anatomical distribution and characteristics of the local structures. Tang et al. (2024) [ 12 ] applied a deep learning framework for shoulder joint ultrasound imaging, incorporating a standard plane recognition module based on the ResNet50 network and an automatic tissue segmentation module utilizing the Mask R-CNN model. The Mask R-CNN framework integrates object detection, classification, and pixel-level segmentation to improve anatomical structure delineation (Fig. 3 ). The system consisted of two key modules: (1) a standard plane recognition module designed to automatically identify specific ultrasound standard planes and (2) a tissue segmentation module responsible for delineating anatomical structures within the identified planes. The study included 11 standard scanning planes. To improve model robustness, the training dataset also incorporated images of non-standard planes. The model achieved a test set accuracy of 94.9% for standard plane recognition, with an average precision of 96.4%, recall of 95.4%, and an F1 score of 95.9%. The automatic tissue segmentation model demonstrated strong performance, with an average intersection-over-union (IoU) value of 96.2%, indicating high segmentation accuracy. Mean IoU values exceeded 90.0% across all standard planes, confirming the model’s effectiveness in precisely delineating anatomical structures. The developed algorithm demonstrated that the model effectively shortens the learning curve and improves diagnostic quality, offering significant benefits to sonographers in clinical practice. The study is limited by the dataset used, which comprised only normal rotator cuff ultrasound images; as such, the classification and segmentation performance on abnormal shoulder joint images, such as frozen shoulder and rotator cuff injuries, remains uncertain. Fig. 3. Open in a new tab Mask R-CNN model structure diagram for automatic tissue segmentation of the shoulder joint. This figure was independently created by the authors based on the conceptual framework described in Tang et al. (2024), without reproducing any copyrighted material Deep Learning for Identifying Rotator Cuff Calcification and Tears B-mode ultrasound is a valuable modality for visualizing rotator cuff calcifications and guiding treatment decisions in patients with shoulder pain; however, its diagnostic accuracy is highly operator-dependent and requires substantial training. To address this limitation, Chiu et al. (2022) [ 13 ] conducted a retrospective study to evaluate the feasibility of a computer-aided system for detecting supraspinatus tendon calcifications. Only ultrasound images in longitudinal or transverse views of the supraspinatus tendon were included, and two physicians independently classified the presence or absence of calcifications. The researchers employed DenseNet-121, a pre-trained convolutional neural network, to develop their detection model. Among the models tested, a combined longitudinal-transverse CNN-based architecture yielded the best performance, with an accuracy of 91.32%, sensitivity of 87.89%, and specificity of 94.74%. These findings suggest that deep learning–based computer-aided systems can assist clinicians—particularly those with limited experience—in identifying supraspinatus tendon calcifications on ultrasound. A key limitation of the study is that morphological characteristics of the calcific deposits, such as size, shape, and echogenicity, were not considered in the analysis. A rotator cuff tear can significantly impair shoulder function, causing pain, weakness, and limited range of motion that interferes with daily activities. To improve ultrasound-based diagnosis, Lee et al. (2021) [ 13 ] proposed an innovative end-to-end fully CNN called the Segmentation Model Adopting a pre-trained Classification Architecture (SMART-CA) (Fig. 4 ). This model incorporates the Integrated Positive Loss Function (IPLF), a novel loss function designed to enhance segmentation accuracy, particularly for imbalanced datasets common in rotator cuff tendon tear detection. SMART-CA extracts distinctive features that conventional encoders often miss, resulting in superior segmentation performance. Compared to existing models (VGG19, GoogleNet, ResNet151, U-Net, SegNet, DenseNet, and FusionNet), SMART-CA achieved substantial gains in precision (0.604%, + 38.4%), recall (0.942%, + 14.0%), and Dice coefficient (0.736%, + 38.6%). These findings demonstrate that SMART-CA, optimized with IPLF, offers robust segmentation accuracy and resilience to image noise, making it a promising tool for diagnosing rotator cuff tears. However, future work should explore integrating SMART-CA with other state-of-the-art architectures beyond VGG19 and U-Net to further enhance diagnostic performance. Fig. 4. Open in a new tab Illustration of the trainable encoder and decoder performing segmentation to localize the rotator cuff tendon tear. This figure was independently created by the authors based on the conceptual framework described in Lee et al. (2021), without reproducing any copyrighted material Given the limited application of deep learning techniques in the diagnosis of rotator cuff tendon tears, Ho et al. (2022) [ 14 ] proposed a CNN–based method for automated classification and tear localization in ultrasound images. Their approach utilized transfer learning and fine-tuning to improve model performance, with hyperparameters optimized using Bayesian optimization. Among the five architectures tested, DenseNet121 demonstrated the highest performance, achieving an accuracy of 88.2%, sensitivity of 93.8%, specificity of 83.6%, and an area under the curve score of 0.832. To enhance interpretability, gradient-weighted class activation mapping was employed to visualize the image regions most relevant to the classification. A key limitation of the study was its reliance on data from a single center and a single imaging modality, with a relatively small proportion of subjects without rotator cuff tendon tears. To enhance the accuracy of one-stage object detection of rotator cuff tendon tear, Wu et al. (2025) [ 15 ] conducted a study involving individuals with chronic shoulder pain who underwent both ultrasound and magnetic resonance imaging. Participants were divided into two balanced groups based on the presence or absence of tendon tears to ensure robust model training. The researchers applied transfer learning using a pre-trained Yolov7 model and an improved version augmented with the Convolutional Block Attention Module (CBAM). Model performance was evaluated using mean average precision, sensitivity, and the F1-score (Fig. 5 ). On the test set, the Yolov7 model achieved an accuracy of 0.96, while the Yolov7-CBAM model improved performance further to 0.98. The integration of the AI-assisted Yolov7-CBAM model not only increased diagnostic accuracy from 80.86% to 88.86%, but also improved interobserver reliability among physicians from 0.49 to 0.71. Clinically, applying the trained Yolov7-CBAM model to ultrasound videos of the rotator cuff tendon enabled real-time and accurate detection during inference, demonstrating the model’s potential for practical deployment. A limitation noted was that in some cases, the contours of tendon tears were poorly defined, making segmentation difficult. Since effective segmentation depends on clearly delineated object boundaries, its application was initially attempted but proved unsuccessful due to issues such as gradient vanishing during training. Fig. 5. Open in a new tab Visual representation of the original ultrasound image with a rotator cuff tendon tear, highlighting object detection and Grad-CAM heatmaps focused on the hypoechoic and anechoic tendon defect. This figure was independently created by the authors based on the conceptual framework described in Wu et al. (2020), without reproducing any copyrighted material Application of Artificial Intelligence for Differentiating Scapulohumeral Periarthritis Scapulohumeral periarthritis is characterized by the gradual onset of shoulder pain, which tends to worsen at night and is accompanied by restricted shoulder mobility. Yu et al. (2023) [ 15 ] conducted a study involving patients diagnosed with scapulohumeral periarthritis to investigate the posterior structures of the shoulder joint (Fig. 6 ). The researchers introduced an intelligent clustering analysis algorithm that incorporated musculoskeletal ultrasound parameters. The findings revealed a correlation between posterior capsule thickness and pain severity. In patients with mild pain, the posterior capsule was thickened (2.02 ± 0.72 mm), while in those with moderate pain, it thinned progressively (1.01 ± 0.38 mm). In the severe pain group, the thickness slightly increased but remained reduced overall (1.21 ± 0.42 mm). When validated on an independent dataset, the model achieved an accuracy of 83.3%, a sensitivity of 87.2%, and a specificity of 80.1% for identifying the condition. These results highlight the potential of combining ultrasound imaging with AI for diagnosing and staging scapulohumeral periarthritis. A noted limitation was the reduced generalizability in patients with moderate to severe symptoms, which hindered accurate ultrasound assessment of the posterior capsule. Nonetheless, when shoulder abduction could be performed during the examination, visualization of the posterior glenohumeral structures improved considerably. Fig. 6. Open in a new tab Ultrasound image of the posterior glenohumeral capsule. This figure was independently created by the authors based on the conceptual framework described in Yu et al. (2023), without reproducing any copyrighted material. IST, infraspinatus tendon; PL, posterior labrum; G, glenoid; H, humerus Deep Learning for Delineating Subacromial Motion Unlike static ultrasonography, dynamic ultrasonography enables real-time visualization of muscle and tendon movements and has been increasingly utilized in shoulder assessments. Shu et al. (2023) [ 15 ] investigated the application of a deep learning algorithm to extract subacromial motion metrics from dynamic ultrasound imaging (Fig. 7 ). The study involved participants performing cyclic shoulder abduction and adduction in the scapular plane, during which a deep learning model tracked the trajectory of the humeral greater tubercle relative to the lateral acromion. Fig. 7. Open in a new tab Illustration of AI-based image analysis showing the subacromial motion metrics of the greater tubercle (blue point) relative to the lateral acromion (red point). This figure was independently created by the authors based on the conceptual framework described in Shu et al. (2023), without reproducing any copyrighted material. DEL, deltoid; SS, supraspinatus; ACR, acromion; HH, humeral head Two model architectures were evaluated: a CNN and a self-transfer learning–based CNN (STL-CNN), each tested with and without an autoencoder. The primary outcome was the mean absolute error of the vertical acromiohumeral distance compared to manually annotated ground truth data. Results showed that the CNN alone yielded significantly higher errors in estimating the relative horizontal and vertical positions of the anatomical landmarks. In contrast, both STL-CNN and STL-CNN with autoencoder substantially reduced error rates. These findings indicate that STL-CNN, particularly when integrated with an autoencoder, offers superior accuracy in extracting subacromial motion metrics from dynamic ultrasound data. The developed algorithm accurately captured the minimum vertical acromiohumeral distance—the most clinically valuable parameter—highlighting its strong potential for future diagnostic applications. However, as this was a pilot study involving only asymptomatic participants, the findings may have limited generalizability to populations with shoulder disorders. Discussion This review underscored the expanding role of deep learning in shoulder ultrasound, with strong performance in segmenting and classifying pathologies such as BPE and rotator cuff tears, as well as identifying key anatomical structures. Machine learning has also surpassed logistic regression in predicting treatment outcomes, including rotator cuff calcification removal. Notably, one study [ 16 ] demonstrated its utility in analyzing subacromial motion metrics from dynamic imaging, highlighting its potential for real-time differentiation of subacromial impingement. Our AI-focused review found that deep learning models, particularly CNNs, have been increasingly applied to the segmentation and classification of shoulder pathologies. Studies have demonstrated the effectiveness of deep learning in detecting BPE and rotator cuff tears using ultrasound, with automated systems achieving substantial accuracy. For instance, Lin et al. [ 9 ] employed a two-stage CNN approach, integrating Faster R-CNN for region-of-interest detection and classification models for grading BPE severity. Similarly, deep learning models such as DenseNet-121 have been successfully applied to identify supraspinatus tendon calcifications [ 13 ], with models trained on both transverse and longitudinal views demonstrating superior diagnostic performance compared to single-view models. In segmentation tasks, advanced architectures like SMART-CA [ 17 ], incorporating pre-trained classification networks and novel loss functions, have outperformed traditional models in detecting rotator cuff tears. Additionally, hybrid models combining CNNs with attention mechanisms, such as the Yolov7-CBAM model [ 15 ], have significantly enhanced diagnostic precision and interobserver reliability in differentiating intact from torn tendons. These findings emphasize the importance of incorporating diverse imaging perspectives and optimized training strategies to improve model performance in ultrasound-based shoulder pathology assessment. In our review, Mask R-CNN demonstrated superior performance in segmenting ultrasound images of shoulder structures—particularly the rotator cuff—outperforming models such as YOLOv8-seg and HRNet-seg. In the study by Tang et al. [ 12 ], Mask R-CNN achieved a high intersection-over-union of 96.2%, surpassing YOLOv8-seg and HRNet-seg, particularly in segmenting smaller, intricate structures such as the supraspinatus and subscapularis tendons. This performance is attributed to Mask R-CNN’s ability to conduct both object detection and pixel-level segmentation [ 18 , 19 ], allowing for precise delineation of overlapping anatomical features. While YOLOv8-seg excels in speed and efficiency [ 20 ], it may struggle with fine-grained segmentation, particularly in complex structures like tendons. HRNet-seg [ 21 ], known for its multi-resolution feature extraction, is effective at capturing detailed spatial relationships but does not match Mask R-CNN’s segmentation accuracy in highly detailed anatomical structures. Therefore, Mask R-CNN’s strong pixel-level segmentation capabilities make it well-suited for the fine-grained analysis required in shoulder ultrasound imaging. Moreover, our review found that the application of deep learning in shoulder ultrasound imaging has advanced considerably. Various models have been employed, including U-Net, FusionNet, DenseNet121, VGG19, InceptionV3, Xception, ResNet50, and YOLOv7, each offering distinct strengths. U-Net and FusionNet are particularly effective for segmentation tasks, while DenseNet121, VGG19, and ResNet50 perform well in classification. YOLOv7 is notable for real-time detection, although its accuracy depends on factors such as image quality and resolution. Our review also highlighted the application of deep learning in analyzing subacromial motion metrics from dynamic shoulder imaging. AI-based models enable real-time assessment of movement abnormalities, providing quantitative data that can aid in diagnosing impingement syndromes and shoulder instability. STL-CNN, in particular, demonstrated superior performance by refining learned features through shared convolutional layers and specialized branches, effectively reducing overfitting. It accurately identified the moment of minimal acromiohumeral distance and delivered the most precise landmark estimations in the test dataset. Moreover, AI-driven motion tracking holds promise for enhancing rehabilitation by quantifying movement patterns and guiding personalized treatment strategies. Although the model shows considerable potential, further validation is needed to assess its effectiveness in differentiating pathological conditions such as subacromial impingement syndrome and glenohumeral osteoarthritis. In addition, this review showcased the value of machine learning models in predicting treatment outcomes, including the success of calcification removal in rotator cuff tendinopathy [ 11 ]. Unlike traditional logistic regression, which relies on predefined features and may fail to capture complex patterns in ultrasound images, machine learning techniques can extract hierarchical radiomic features, improving predictive accuracy. In our review, we found that Triantafyllou et al. [ 11 ] applied various machine learning methods to analyze radiomic and clinical data. Their findings showed that combining wavelet-based radiomic features with clinical variables, such as maximum calcification diameter, enhanced outcome prediction. This supports the potential of radiomics-based machine learning models to improve clinical decision-making through a more nuanced, data-driven approach. Despite recent advances, the application of AI to shoulder ultrasound remains challenging. Key obstacles include the requirement for large, annotated datasets, which are time-consuming to produce and subject to variability in image acquisition. The generalizability of AI models across different ultrasound machines, clinical institutions, and operator techniques also poses a significant concern. Additionally, many AI models function as “black boxes,” lacking interpretability, which hampers clinical trust and regulatory approval. Successful integration into clinical workflows necessitates rigorous validation through prospective, multi-center trials and the use of transparent, standardized performance metrics. The adoption of AI in shoulder ultrasound may be hindered by regulatory, ethical, and workflow challenges. Despite promising retrospective results, few tools have been prospectively validated or approved for clinical use. Widespread implementation requires clinician trust, integration with existing ultrasound picture archiving and communication systems or electronic medical records, and clear legal accountability for diagnostic errors. The present narrative review has several limitations. The included studies show considerable heterogeneity in dataset sizes, preprocessing techniques, and model architectures, which complicates direct comparisons. Additionally, the reliance on retrospective datasets may introduce selection bias, limiting the generalizability of the findings to broader clinical populations. The reviewed studies primarily focused on rotator cuff tendon tears, calcifications, and BPE. However, other common causes of shoulder pain, such as subdeltoid bursitis and acromioclavicular joint synovitis, were insufficiently addressed. Given their clinical relevance, future studies should incorporate a broader range of pathologies to offer a more comprehensive understanding of ultrasound findings in patients with shoulder pain. Conclusions Deep learning has shown great potential in shoulder ultrasound by improving diagnostic accuracy and efficiency. Its ability to analyze complex image patterns can enhance the detection of shoulder pathologies, such as rotator cuff tears and impingement. However, challenges remain, including limited datasets, variability in image acquisition, and the need for clinical validation. Future research should refine deep learning models, expand shoulder ultrasound datasets, and integrate AI-assisted analysis into clinical practice. With continued advancements, deep learning may significantly enhance the diagnostic utility of shoulder ultrasound. Funding The present study was funded by National Taiwan University Hospital, Bei-Hu Branch (grant nos. 113-BH002 and 113-BH004), the Ministry of Science and Technology (grant nos. MOST 106-2314-B-002-180-MY3, 109-2314-B-002-114- MY3, 111-2314-B-002–161, NSTC 113-2314-B-002-208-MY2, NSTC 113-2314-B- 16 002-209-MY2 and NSTC 114-2923-B-002 -002 -MY3), and the Taiwan Society of Ultrasound in Medicine. Declarations Conflict of Interest The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Luime JJ, Koes BW, Hendriksen IJ, et al. Prevalence and incidence of shoulder pain in the general population; a systematic review. Scand J Rheumatol 2004;33(2):73–81. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Meislin RJ, Sperling JW, Stitik TP. Persistent shoulder pain: epidemiology, pathophysiology, and diagnosis. Am J Orthop (Belle Mead NJ) 2005;34(12 Suppl):5–9. [ PubMed ] [ Google Scholar ] 3. Wu WT, Lin CY, Shu YC, Chen LR, Özçakar L, Chang KV. Subacromial Motion Metrics in Painful Shoulder Impingement: A Dynamic Quantitative Ultrasonography Analysis. Arch Phys Med Rehabil 2023;104(2):260-9. [ DOI ] [ PubMed ] [ Google Scholar ] 4. Chang KV, Mezian K, Naňka O, Wu WT, Lin CP, Özçakar L. Ultrasound-guided interventions for painful shoulder: from anatomy to evidence. J Pain Res 2018;11:2311-22. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Roy JS, Braën C, Leblond J, et al. Diagnostic accuracy of ultrasonography, MRI and MR arthrography in the characterisation of rotator cuff disorders: a systematic review and meta-analysis. Br J Sports Med 2015;49(20):1316-28. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Farooqi AS, Lee A, Novikov D, et al. Diagnostic Accuracy of Ultrasonography for Rotator Cuff Tears: A Systematic Review and Meta-analysis. Orthop J Sports Med 2021;9(10):23259671211035106. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Yamashita R, Nishio M, Do RKG, Togashi K. Convolutional neural networks: an overview and application in radiology. Insights into Imaging 2018;9(4):611-29. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Dinescu SC, Stoica D, Bita CE, et al. Applications of artificial intelligence in musculoskeletal ultrasound: narrative review. Front Med (Lausanne) 2023;10:1286085. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Lin BS, Chen JL, Tu YH, et al. Using Deep Learning in Ultrasound Imaging of Bicipital Peritendinous Effusion to Grade Inflammation Severity. IEEE J Biomed Health Inform 2020;24(4):1037-45. [ DOI ] [ PubMed ] [ Google Scholar ] 10. Kim MS, Kim IW, Lee S, Shin SJ. Diagnosis and treatment of calcific tendinitis of the shoulder. Clin Shoulder Elb 2020;23(4):210-6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Triantafyllou M, Vassalou EE, Klontzas ME, Tosounidis TH, Marias K, Karantanas AH. Ultrasound radiomics predict the success of US-guided percutaneous irrigation for shoulder calcific tendinopathy. Jpn J Radiol 2025:Epub ahead of print. [ DOI ] [ PubMed ] 12. Tang R, Li Z, Jiang L, et al. Development and Clinical Application of Artificial Intelligence Assistant System for Rotator Cuff Ultrasound Scanning. Ultrasound Med Biol 2024;50(2):251-7. [ DOI ] [ PubMed ] [ Google Scholar ] 13. Chiu PH, Boudier-Revéret M, Chang SW, Wu CH, Chen WS, Özçakar L. Deep Learning for Detecting Supraspinatus Calcific Tendinopathy on Ultrasound Images. J Med Ultrasound 2022;30(3):196–202. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Ho TT, Kim GT, Kim T, Choi S, Park EK (2022) Classification of rotator cuff tears in ultrasound images using deep learning models. Med Biol Eng Comput 60(5):1269–1278 [ DOI ] [ PubMed ] [ Google Scholar ] 15. Wu KT, Chen PC, Chou WY, Chang CD, Lien JJ. Diagnostic Accuracy and Interobserver Reliability of Rotator Cuff Tear Detection with Ultrasonography are Improved with Attentional Deep Learning. Arthroscopy 2024:Epub ahead of print. [ DOI ] [ PubMed ] 16. Shu YC, Lo YC, Chiu HC et al (2023) Deep learning algorithm for predicting subacromial motion trajectory: Dynamic shoulder ultrasound analysis. Ultrasonics 134 [ DOI ] [ PubMed ] 17. Lee K, Kim JY, Lee MH, Choi CH, Hwang JY. Imbalanced Loss-Integrated Deep-Learning-Based Ultrasound Image Analysis for Diagnosis of Rotator-Cuff Tear. Sensors (Basel) 2021;21(6):2214. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Couteaux V, Si-Mohamed S, Nempont O, et al. Automatic knee meniscus tear detection and orientation classification with Mask-RCNN. Diagn Interv Imaging 2019;100(4):235-42. [ DOI ] [ PubMed ] [ Google Scholar ] 19. Zhang Y, Tian Z, Lei Y, et al. Automatic multi-needle localization in ultrasound images using large margin mask RCNN for ultrasound-guided prostate brachytherapy. Phys Med Biol 2020;65(20):205003. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Casas E, Ramos L, Romero C, Rivas-Echeverría F. A comparative study of YOLOv5 and YOLOv8 for corrosion segmentation tasks in metal surfaces. Array 2024;22:100351. [ Google Scholar ] 21. Cheng Z, Fu D. Remote Sensing Image Segmentation Method based on HRNET. In: IGARSS 2020 - 2020 IEEE International Geoscience and Remote Sensing Symposium . Articles from Journal of Imaging Informatics in Medicine are provided here courtesy of Springer ACTIONS View on publisher site PDF (1.3 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Application of artificial intelligence to the electrocardiogram#679465
• Physiognomic Artificial Intelligence#964388
• Artificial Intelligence for Diagnostic Support in Neuropathology / Katherine J. Hewitt ; Gutachter: Jakob N. Kather, Daniela Aust#669407
• Artificial intelligence and predictive tools in non-muscle invasive bladder cancer: a narrative review of current insights and advances.#362183
• The application of artificial intelligence in adrenal imaging: current state of knowledge, challenges, and future directions.#392822