1Department of Neurology, Chuncheon Sacred Heart Hospital, Hallym University College of Medicine, Chuncheon, Korea
2Institute of New Frontier Research Team, Hallym University College of Medicine, Chuncheon, Korea
Correspondence to Joo Hye Sung Department of Neurology, Chuncheon Sacred Heart Hospital, Hallym University College of Medicine, 77 Sakju-ro, Chuncheon 24253, Korea Tel: +82-33-240-5255 Fax: +82-33-241-8063 E-mail: centertruth@naver.com
• Received: June 29, 2025 • Revised: November 17, 2025 • Accepted: December 4, 2025
This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (https://creativecommons.org/licenses/by-nc/4.0/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.
Ultrasound (US) is a patient-friendly imaging modality that is well suited for structural assessments of skeletal muscle and surrounding tissues, and its use for both diagnostic and monitoring in neuromuscular medicine is steadily expanding. Despite its advantages, the broader clinical implementation of muscle US is still hinder by the obtained images and their interpretation varying between operators. Recent advances in artificial intelligence (AI)-particularly in deep learning (DL)-have led to significant innovations in medical imaging. These developments hold considerable promise for addressing the inherent limitations of muscle US and for improving its clinical applicability. This review summarizes key muscle US biomarkers, encompassing both qualitative visual assessments and quantitative parameters. We discuss their utility in the diagnosis of neuromuscular disorders and their potential role as responsive biomarkers for monitoring disease progression, based on findings from previous studies. We also introduce foundational concepts of AI and DL and review recent studies that have applied these technologies in muscle US for the automated segmentation of muscle boundaries, feature extraction, and disease classification. Finally, we highlight current challenges and outline future directions needed to fully realize the potential of AI-enhanced muscle US as a standardized and widely applicable tool for assessing neuromuscular diseases.
Muscle ultrasound (US) has become an increasingly valuable tool in neuromuscular medicine due to its noninvasiveness and patient friendliness while assessing muscles and surrounding tissues.1,2 In contrast to other commonly used modalities such as nerve conduction studies, electromyography, computed tomography, and magnetic resonance imaging, muscle US is painless, free of radiation, and can be conveniently performed at the bedside. By allowing the visualization of muscles throughout the body, muscle US enables clinicians to assess multiple target muscles flexibly and simultaneously.
Muscle US is increasingly utilized not only for diagnostic evaluations but also for monitoring disease progression. However, its broader clinical implementation is still hindered by several inherent challenges.1,3 The acquired images vary between operators and image interpretation often relies heavily on the examiner’s experience. Recent advances in artificial intelligence (AI)-particularly in deep learning (DL)-have revolutionized image acquisition and automated interpretation across various domains of medical imaging.4 Integrating these technologies with muscle US may help address existing limitations and improve consistency in clinical practice.
This review explores the expanding clinical applications of muscle US in neuromuscular disorders and discusses both its potential and limitations. We also introduce key concepts in AI and DL as they apply to muscle US and summarize recent studies that have implemented these technologies for automated segmentation and analysis.
MUSCLE US BIOMARKERS
Overview of muscle US
US imaging typically shows healthy muscle as exhibiting a “starry night” appearance in transverse views, where hypoechoic muscle fibers are interspersed with hyperechoic connective-tissue structures. In neuromuscular diseases, muscle fibers undergo degeneration to be progressively replaced by fibrotic and fatty tissue.5 These pathological transformations lead to changes in muscle size and internal architecture that are readily visualized in US imaging. Affected muscles often demonstrate a generalized increase in echogenicity with atrophy and a loss of their normal architecture.
In particular, the various types of muscular dystrophy are associated with a diffuse, homogeneously increased echogenicity and a loss of normal muscle architecture, often referred to as a “ground-glass” appearance. The increased US beam attenuation in more-advanced disease stages results in a brightly echogenic superficial layer with markedly reduced signal penetration into deeper tissue layers. A characteristic “moth-eaten” pattern may be seen in neurogenic disorders that consists of irregular, patchy hypoechoic areas representing residual viable motor units, which are surrounded by hyperechogenic regions indicative of chronic denervation and fibrosis.1
Visual assessments
Muscle US images are commonly evaluated semiquantitatively using the Heckmatt scale (Fig. 1),6 which is a four-point visual grading system based on the muscle grayscale level and the degree of attenuation affecting the visibility of underlying bone or fascia.
1) Grade 1: normal muscle appearance with typical echogenicity and clearly visible bone echoes. 2) Grade 2: mildly increased muscle echogenicity, with preservation of distinct bone echoes. 3) Grade 3: markedly increased muscle echogenicity accompanied by reduced bone echoes. And 4) grade 4: severely increased muscle echogenicity with complete loss of bone echoes.
According to previous reports, the sensitivity of using the Heckmatt scale for diagnosing neuromuscular diseases in muscle US ranges from 71% to 76%.7,8
Quantitative muscle US (QMUS)
QMUS provides more-objective and reproducible measurements of muscle characteristics.9,10 The muscle thickness (MT) and cross-sectional area (CSA) are commonly used to assess muscle atrophy. The overall echogenicity increases as muscle is replaced with fat and fibrous tissue. Echogenicity can be quantified in two ways. The more common method is to calculate the echo intensity (EI) as the mean grayscale value of pixels within a manually selected region of interest (ROI) (Fig. 2). However, EI is significantly influenced by demographic variables such as age, sex, body weight, and height, and so the z-score (corresponding to the number of standard deviations that a patient’s EI deviates from the reference mean for that muscle) is often used to enable comparisons between individuals. The second method, known as the calibrated backscatter technique, measures the intensity of the US signal (in decibels) backscattered from the tissue and received by the transducer.11 This value is calibrated using a tissue-mimicking phantom, enabling echogenicity measurements to be compared between different US systems.
Higher order muscle US features
Beyond first-order features such as EI, higher-order texture analysis techniques have been introduced to provide more-comprehensive assessments of tissue microstructure, which may be less affected by US-device settings. Among these, gray-level co-occurrence matrix (GLCM) analysis quantifies the spatial distribution and relationship between pixel intensities.1,4,12 The following GLCM-based features have been applied.
1) Contrast: quantifies the gray-level variation between neighboring pixels, with higher contrast values indicating increased textural heterogeneity. 2) Correlation: assesses the similarity of gray levels between adjacent pixels, with higher coefficients suggesting greater regional similarity, reflecting greater tissue homogeneity. 3) Energy: reflects the uniformity or smoothness of the texture, with higher values corresponding to more-regular and homogeneous textures, and lower values indicating increased randomness or heterogeneity. 4) Homogeneity: assesses local similarity by evaluating the closeness of gray levels in neighboring pixel pairs. And 5) entropy: represents the complexity or randomness of the pixel intensity distribution, with higher values being associated with greater textural irregularity.
Other higher-order texture features are described by the Nakagami distribution, which is defined using two parameters: m (shape) and ω (scale).13 This distribution reflects the scattering conditions of the tissue and is considered sensitive to microstructural changes.
Diagnostic value of QMUS
The diagnostic value of QMUS has been supported by several previous studies. In children with suspected neuromuscular disorders, QMUS has shown utility as a noninvasive screening tool. Pillen et al.9 showed that EI was effective at differentiating between neuromuscular and nonneuromuscular conditions, with a reported sensitivity of 92% and specificity of 90%. The optimal diagnostic threshold was defined as a z-score of 0.9 standard deviations above the normal mean in three or more muscles.
QMUS has also been applied to the diagnosis of amyotrophic lateral sclerosis (ALS), particularly in distinguishing ALS from mimicking conditions and healthy controls.14 It was considered diagnostic for ALS when EI exceeded 1.5 standard deviations above the normal mean in at least two muscles, and fasciculations were detected in four or more muscles, yielding a sensitivity of 96% and specificity of 84%.15 Additional studies that have evaluated a range of US biomarkers including EI, GLCM-based texture features, and fasciculations demonstrate that various combinations of these parameters can facilitate the diagnosis of ALS.16-18
Biomarkers for disease progression
QMUS parameters have been evaluated as potential biomarkers for disease progression in muscular dystrophy. Several longitudinal studies have produced promising results supporting the use of US parameters in patients with Duchenne muscular dystrophy (DMD). MT and EI were found to increase significantly with age and showed longitudinal correlations with clinical measures.19-21 Zaidman et al.22 demonstrated the sensitivity of QMUS to disease progression over 6-, 12-, and 24-month intervals in boys with DMD, and found that US parameters could be used to detect progression before changes were detectable in functional tests. These findings support the feasibility of using QMUS as a noninvasive tool for assessing treatment effects in clinical trials.
Additionally, multiple muscles have been evaluated longitudinally in patients with facioscapulohumeral dystrophy23,24 and oculopharyngeal muscular dystrophy,25 revealing gradual increases in EI over time. These findings suggest that US parameters can serve as responsive biomarkers for tracking disease progression in these conditions. US parameters have also been investigated in longitudinal settings in ALS, with some studies finding modest associations between US findings and functional decline.26-30 While further validation is needed, these results indicate the potential utility of QMUS in monitoring disease progression in ALS.
Pitfalls of muscle US
Despite its many advantages, muscle US has several limitations that hinder its broader clinical application in neuromuscular medicine. It is a highly operator-dependent imaging modality and its accuracy is significantly influenced by the operator’s skill.1,31 Image acquisition requires careful control of the probe orientation, with the transducer needing to be positioned perpendicular to the tissue as possible (Fig. 3) since even slight tilting can cause substantial changes in EI. Minimal pressure must also be applied during scanning, since excessive compression can alter MT measurements.
Moreover, the interpretation of US images relies heavily on the examiner’s experience. Although quantitative analyses allow for more-objective assessments, these measurements are still affected by differences in device hardware, software settings, and acquisition parameters (Fig. 3). This technical variability restricts the comparability between institutions and often necessitates the establishment of local reference values. Additionally, US parameters are influenced by demographic variables, which further complicate standardization. As a result, muscle US is currently performed routinely only in specialized centers that have accumulated sufficient expertise in neuromuscular imaging and have access to large patient cohorts from which to generate their own normative datasets.1,32,33
Furthermore, the manual selection of the ROI is a common source of variability, and is also both time-consuming and prone to human error.1,34 Collectively these limitations highlight the need for more-objective, standardized and automated approaches to improve the reliability, efficiency, and usability of muscle US-AI offers a promising solution in this context.3,33
BASIC CONCEPTS OF AI AND DL
Definitions of AI, machine learning (ML), and DL
AI is a field of computer science that aims to replicate aspects of human cognitive function using computational systems. Traditional programming approaches rely on rule-based algorithms, in which explicit instructions are manually coded by humans to perform specific tasks. In contrast, ML-a subset of AI-enables systems to learn patterns from big data without explicit programming. DL-a subfield of ML-utilizes deep neural networks represented by artificial neural networks (ANNs) composed of multiple interconnected layers (Fig. 4).35,36 These multilayer structures allow the system to model increasingly abstract and complex representations of input data. DL has performed remarkably in tasks such as image classification, speech recognition, and natural language processing, and it is particularly well suited to the analysis of medical imaging data.4
ANNs
ANNs are computational models inspired by the structure and function of biological neurons (Fig. 4). In the human brain, a neuron receives electrical signals via its dendrites, integrates them in the cell body, and transmits its output via its axon. Analogously, an ANN is composed of nodes connected by edges, where each node receives inputs, applies weighted summation, and generates an output using an activation function.
The simplest form of ANN is the perceptron, which mimics the all-or-none response of biological neurons using a unit step function.37 In AI systems, both input and output signals are represented numerically as vectors or matrices; for example, a color image can be represented as a three-dimensional matrix of red, green, and blue pixel values. The ANN functions as a mathematical mapping between the input and output and learns the optimal transformation during the training process.
While a single perceptron can only handle linearly separable problems, combining multiple perceptrons in a multilayer structure allows the network to model nonlinear relationships. These multilayer perceptrons serve as the foundational architecture for deep neural networks, which can extract increasingly abstract features from complex input data through successive hidden layers.
ANNs are trained to minimize a loss function, which quantifies the discrepancy between predicted outputs and true labels. Through iterative optimization, the network updates its internal parameters, or weights, to reduce the loss values. This is commonly achieved using gradient descent, where gradients of the loss values are calculated and propagated backward through the network using backpropagation.38
While it is essential to minimize the loss values during training, ensuring that the model performs well on unseen data is equally important. A key challenge is preventing overfitting, where the model memorizes training data rather than learning generalizable patterns. Various regularization techniques are employed to address this, including dropout, which randomly deactivates a subset of neurons during training to reduce coadaptation, and weight regularization methods such as L1 and L2 penalties, which constrain the complexity of the model and improve generalization.39
Convolutional neural networks (CNNs)
CNNs have a type of DL architecture that is particularly suitable for analyzing image data. CNNs are inspired by the way humans process visual information, focusing on spatial patterns that convey meaning rather on individual pixels. This approach allows CNNs to detect important features within an image while considering their spatial context.40
Convolution is the key operation in CNNs, which involves a small matrix called a kernel or filter moving systematically across the input image. This operation extracts local patterns such as edges, shapes, and textures, and produces a new representation known as a feature map that retains positional information. To reduce the dimensionality and emphasize the most-relevant features, pooling layers are applied to summarize values within a local region, often by selecting the maximum or average value.
By performing convolution across multiple layers and by repeating layers of convolution and pooling, the network progressively captures more abstract and complex features. These features are then passed to fully connected layers that compute the probability of each class, which enables the model to perform tasks such as classification or detection based on its learned representations.
AI-DRIVEN INNOVATIONS IN MUSCLE US
Recent progress in AI-particularly in ML and DL-has led to increasing interest in its potential applications in muscle US imaging.3 Among these, automated segmentation and feature extraction are gaining attention for their potential to improve the efficiency, objectivity, and reproducibility of imaging (Fig. 5). Several studies have introduced DL-based tools for use in muscle US imaging. For example, DeepACSA was developed by Ritsche et al.41 to automatically measure the anatomical CSA from panoramic US images of lower limb muscles obtained by multiple operators and US devices. Using a U-Net architecture with either standard or VGG16- based encoders, the model results were in close agreement with manual measurements and was associated with a significant reduction in the analysis time. Xin et al.42 introduced a U-Net-based model that successfully segmented the flexor digitorum superficialis, and outperformed novice physicians. Zhang et al.43 proposed MSF-Net employing multistage fusion and segmentation, which improved determinations of the MT, pennation angle, and fascicle length over using baseline models. Rivera et al.44 developed MyoVision-US using DeepLabV3 with ResNet-50, and reported strong concordance with expert annotations of MT, CSA, and EI among diverse patient groups.
Efforts have also been made to differentiate normal and pathological muscles using automated analysis methods. Marzola et al.34 used an ensemble model to segment muscle, calculate EI z-scores, and identify pathological muscle defined as an z-score >2 standard deviations. Their model achieved high segmentation accuracy and excellent agreement with manual z-score-based assessments. Zhou et al.45 developed a multitask model (MMA-Net) for simultaneous segmentation and abnormality classification, with their results demonstrating the potential for improved robustness and clinical applicability.
Furthermore, the diagnosis and classification of various neuromuscular diseases have been explored through conventional ML, DL, and hybrid approaches. Noda et al.32 developed a real-time AI system that classified neurogenic, myogenic, and normal conditions by applying a random-forest classifier to texture features of the biceps brachii. Burlina et al.33 applied an ALEXnet-based DL model to inflammatory myopathy classification, which performed better than conventional methods while requiring no manual input. Katakis et al.46 employed the TMUNet vision transformer-based model to automatically segment the CSA and estimate EI from US images of multiple muscles, and achieved close agreement with manual measures and >84% accuracy in group classifications of healthy, elderly, and sarcopenic muscles. Yik et al.47 used an AI-assisted tool (MuscleSound®, Denver, CO, USA) to quantify intramuscular adipose tissue (IMAT) and proposed the IMAT index for sarcopenia diagnosis in surgical patients, with an area under the receiver operating characteristic curve of 0.727 showing good reliability for both novice and expert users. Liao et al.48 introduced a hybrid method combining clustering-based texture analysis and CNNs-including VGG16 and VGG19-to classify ambulatory function and disease severity in DMD patients, and achieved high predictive accuracy.
REMAINING CHALLENGES AND FUTURE DIRECTIONS
Developing robust AI models for muscle US requires large, multicenter datasets derived from diverse patient populations, disease types, and clinical stages. The advancement of high-performance computing resources will also be critical to support real-time processing and the training of large-scale models. By capturing fine-grained textural information beyond human perception, DL technologies facilitate the differentiation of various tissue compositions and pathological alterations.3,33
Explainability
While DL has demonstrated impressive performance, the integration of explainable AI (XAI) remains essential to ensure transparency and support clinical applicability.4 The numerous layers and complex representations in DL architectures result in their decision-making processes not being readily intelligible to humans. There is a particular need for XAI in the medical field, where understanding model behavior is vital for the integration of AI into clinical workflows. XAI techniques including attribution-based methods such as saliency maps and heat maps can be utilized to identify the image features that influence model outputs.49 Class activation mapping (CAM) approaches such as grad-CAM and score- CAM highlight relevant anatomical regions and textural patterns that contribute most strongly to model predictions. These visualization methods also provide insight into incorrect classifications, thereby aiding further refinement of the models.
Generalizability
It is necessary to standardize imaging protocols and annotation practices to ensure data consistency between centers. The heterogeneity in patient populations, US devices, and acquisition protocols in multicenter environments can lead to substantial domain shifts, which may compromise the generalizability of AI models. To mitigate these biases, diverse data collection, domain adaptation, harmonization, and multicenter imaging standardization are essential.49 Expanding collaborative research through federated learning may further enable secure, multicenter model development without compromising patient privacy.4,50
CONCLUSION
The clinical relevance of muscle US in neuromuscular medicine is increasing and its utility is being further improved by the integration of AI. DL-based approaches enable the automated extraction of conventional and advanced imaging features and improve the consistency of interpretations among examiners. To facilitate the broader clinical adoption of AI-assisted muscle US, future work should focus on developing standardized imaging protocols, XAI frameworks, and multicenter data collaboration.
Notes
Conflict of Interest
The author certifies that there is no conflict of interest with any financial organization regarding the material discussed in the manuscript.
Funding
This research was supported by the Bio & Medical Technology Development Program of the National Research Foundation (NRF) funded by the Korean government (MSIT) (grant number RS-2023-00223501).
Fig. 1.
Representative transverse ultrasound images of the tibialis anterior muscle, each corresponding to a different grade on the Heckmatt scale: (A) grade 1, showing normal muscle echogenicity with a clearly visible interosseous membrane; (B) grade 2, showing mildly increased echogenicity with preservation of a distinctly visible interosseous membrane; (C) grade 3, showing markedly increased echogenicity with reduced visibility of the interosseous membrane; and (D) grade 4, showing severely increased echogenicity with no visualization of the interosseous membrane.
Fig. 2.
Transverse US images of the biceps brachii (A), flexor carpi radialis (B), abductor pollicis brevis (C), and vastus intermedius (D). Green boundaries indicate representative regions of interest for an echo-intensity analysis. Measurement consistency is ensured by using clearly identifiable anatomical landmarks as reference points. US, ultrasound.
Fig. 3.
Transverse US images of the midforearm. (A, B) Differences in muscle EI and texture at the same anatomical location in the same subject resulting from different imaging settings when using the same US device. The effect of probe tilting on image quality is demonstrated by comparing images acquired with the probe positioned perpendicular (i.e., at 90°) to the tissue surface (B) and with the probe intentionally tilted (C). US, ultrasound; EI, echo intensity.
Fig. 4.
Schematic comparison between biological and artificial neurons and the structure of a deep neural network. (A) A biological neuron receives input signals via its dendrites, integrates them in the soma, and transmits output signals via its axon to synaptic terminals. (B) An artificial neuron mimics this process by computing a weighted sum of inputs and applying an activation function to produce an output. (C) Deep neural networks are artificial neural networks composed of multiple hidden layers and designed to solve complex nonlinear problems.
Fig. 5.
Representative example of muscle segmentation using a deep learning (DL)-based model. (A) Original transverse US image of the suprahyoid region. (B) Expert manual annotation outlining the digastric, mylohyoid and geniohyoid. (C) Automated segmentation generated by the DL model, illustrating the delineation of each suprahyoid muscle layer. (D) Segmentation masks for the individual muscles: digastric (left), mylohyoid (middle), and geniohyoid (right). The automated outputs show strong concordance with expert tracings, demonstrating the model’s ability to accurately identify and segment multilayer suprahyoid muscle structures. US, ultrasound.
References
1. Wijntjes J, van Alfen N. Muscle ultrasound: present state and future opportunities. Muscle Nerve 2021;63:455-466.
2. Walker FO, Cartwright MS, Alter KE, Visser LH, Hobson-Webb LD, Padua L, et al. Indications for neuromuscular ultrasound: expert opinion and review of the literature. Clin Neurophysiol 2018;129:2658-2679.
3. Taha MA, Morren JA. The role of artificial intelligence in electrodiagnostic and neuromuscular medicine: current state and future directions. Muscle Nerve 2024;69:260-272.
8. Brandsma R, Verbeek RJ, Maurits NM, van der Hoeven JH, Brouwer OF, den Dunnen WF, et al. Visual screening of muscle ultrasound images in children. Ultrasound Med Biol 2014;40:2345-2351.
10. Arts IM, Pillen S, Schelhaas HJ, Overeem S, Zwarts MJ. Normal values for quantitative muscle ultrasonography in adults. Muscle Nerve 2010;41:32-41.
11. Zaidman CM, Holland MR, Anderson CC, Pestronk A. Calibrated quantitative ultrasound imaging of skeletal muscle using backscatter analysis. Muscle Nerve 2008;38:893-898.
13. Oelze ML, Mamou J. Review of quantitative ultrasound: envelope statistics and backscatter coefficient imaging and contributions to diagnostic ultrasound. IEEE Trans Ultrason Ferroelectr Freq Control 2016;63:336-351.
15. Arts IM, Overeem S, Pillen S, Kleine BU, Boekestein WA, Zwarts MJ, et al. Muscle ultrasonography: a diagnostic tool for amyotrophic lateral sclerosis. Clin Neurophysiol 2012;123:1662-1667.
16. Grimm A, Prell T, Décard BF, Schumacher U, Witte OW, Axer H, et al. Muscle ultrasonography as an additional diagnostic tool for the diagnosis of amyotrophic lateral sclerosis. Clin Neurophysiol 2015;126:820-827.
18. Tsuji Y, Noto YI, Shiga K, Teramukai S, Nakagawa M, Mizuno T. A muscle ultrasound score in the diagnosis of amyotrophic lateral sclerosis. Clin Neurophysiol 2017;128:1069-1074.
19. Jansen M, van Alfen N, Nijhuis van der Sanden MW, van Dijk JP, Pillen S, de Groot IJ. Quantitative muscle ultrasound is a promising longitudinal follow-up tool in Duchenne muscular dystrophy. Neuromuscul Disord 2012;22:306-317.
20. Zaidman CM, Malkus EC, Connolly AM. Muscle ultrasound quantifies disease progression over time in infants and young boys with duchenne muscular dystrophy. Muscle Nerve 2015;52:334-338.
21. Hu J, Jiang L, Hong S, Cheng L, Wang Q, Peng X, et al. Quantitative muscle ultrasound in children with Duchenne muscular dystrophy: comparing to magnetic resonance imaging. J Clin Ultrasound 2023;51:674-679.
23. Goselink RJM, Schreuder THA, Mul K, Voermans NC, Erasmus CE, van Engelen BGM, et al. Muscle ultrasound is a responsive biomarker in facioscapulohumeral dystrophy. Neurology 2020;94:e1488-e1494.
24. Vincenten SCC, Teeselink S, Voermans NC, van Engelen BGM, Mul K, van Alfen N. Establishing the role of muscle ultrasound as an imaging biomarker in facioscapulohumeral muscular dystrophy. Neuromuscul Disord 2023;33:936-944.
26. Lee CD, Song Y, Peltier AC, Jarquin-Valdivia AA, Donofrio PD. Muscle ultrasound quantifies the rate of reduction of muscle thickness in amyotrophic lateral sclerosis. Muscle Nerve 2010;42:814-819.
28. Martínez-Payá JJ, Del Baño-Aledo ME, Ríos-Díaz J, Tembl-Ferrairó JI, Vázquez-Costa JF, Medina-Mirapeix F. Muscular echovariation: a new biomarker in amyotrophic lateral sclerosis. Ultrasound Med Biol 2017;43:1153-1162.
29. Martínez-Payá JJ, Ríos-Díaz J, Medina-Mirapeix F, Vázquez-Costa JF, Del Baño-Aledo ME. Monitoring progression of amyotrophic lateral sclerosis using ultrasound morpho-textural muscle biomarkers: a pilot study. Ultrasound Med Biol 2018;44:102-109.
30. Pathak S, Caress JB, Wosiski-Kuhn M, Milligan C, Williams D, Cartwright MS. A pilot study of neuromuscular ultrasound as a biomarker for amyotrophic lateral sclerosis. Muscle Nerve 2019;59:181-186.
33. Burlina P, Billings S, Joshi N, Albayda J. Automated diagnosis of myositis from muscle ultrasound: exploring the use of machine learning and deep learning methods. PLoS One 2017;12:e0184059.
34. Marzola F, van Alfen N, Doorduin J, Meiburger KM. Deep learning segmentation of transverse musculoskeletal ultrasound images for neuromuscular disease assessment. Comput Biol Med 2021;135:104623.
41. Ritsche P, Wirth P, Cronin NJ, Sarto F, Narici MV, Faude O, et al. DeepACSA: automatic segmentation of cross-sectional area in ultrasound images of lower limb muscles using deep learning. Med Sci Sports Exerc 2022;54:2188-2195.
42. Xin C, Li B, Wang D, Chen W, Yue S, Meng D, et al. Deep learning for the rapid automatic segmentation of forearm muscle boundaries from ultrasound datasets. Front Physiol 2023;14:1166061.
43. Zhang M, Zhao L, Wang X, Lo WLA, Wen J, Li L, et al. Automatic extraction and measurement of ultrasonic muscle morphological parameters based on multi-stage fusion and segmentation. Ultrasonics 2024;137:107187.
44. Rivera ZC, González-Seguel F, Horikawa-Strakovsky A, Granger C, Sarwal A, Dhar S, et al. Development of an artificial intelligence powered software for automated analysis of skeletal muscle ultrasonography. Sci Rep 2025;15:14936.
45. Zhou L, Liu S, Zheng W. Automatic analysis of transverse musculoskeletal ultrasound images based on the multi-task learning model. Entropy (Basel) 2023;25:662.
46. Katakis S, Barotsis N, Kakotaritis A, Tsiganos P, Economou G, Panagiotopoulos E, et al. Muscle cross-sectional area segmentation in transverse ultrasound images using vision transformers. Diagnostics (Basel) 2023;13:217.
47. Yik V, Kok SSX, Chean E, Lam YE, Chua WT, Tan WJ, et al. Diagnosing sarcopenia with AI-aided ultrasound (DINOSAUR)-A pilot study. Nutrients 2024;16:2768.
48. Liao AH, Wang CH, Wang CY, Liu HL, Chuang HC, Tseng WJ, et al. Computer-aided diagnosis of Duchenne muscular dystrophy based on texture pattern recognition on ultrasound images using unsupervised clustering algorithms and deep learning. Ultrasound Med Biol 2024;50:1058-1068.
49. Castiglioni I, Rundo L, Codari M, Di Leo G, Salvatore C, Interlenghi M, et al. AI applications to medical images: from machine learning to deep learning. Phys Med 2021;83:9-24.
Muscle ultrasound as a promising tool: clinical applications and the emerging role of deep learning
Fig. 1. Representative transverse ultrasound images of the tibialis anterior muscle, each corresponding to a different grade on the Heckmatt scale: (A) grade 1, showing normal muscle echogenicity with a clearly visible interosseous membrane; (B) grade 2, showing mildly increased echogenicity with preservation of a distinctly visible interosseous membrane; (C) grade 3, showing markedly increased echogenicity with reduced visibility of the interosseous membrane; and (D) grade 4, showing severely increased echogenicity with no visualization of the interosseous membrane.
Fig. 2. Transverse US images of the biceps brachii (A), flexor carpi radialis (B), abductor pollicis brevis (C), and vastus intermedius (D). Green boundaries indicate representative regions of interest for an echo-intensity analysis. Measurement consistency is ensured by using clearly identifiable anatomical landmarks as reference points. US, ultrasound.
Fig. 3. Transverse US images of the midforearm. (A, B) Differences in muscle EI and texture at the same anatomical location in the same subject resulting from different imaging settings when using the same US device. The effect of probe tilting on image quality is demonstrated by comparing images acquired with the probe positioned perpendicular (i.e., at 90°) to the tissue surface (B) and with the probe intentionally tilted (C). US, ultrasound; EI, echo intensity.
Fig. 4. Schematic comparison between biological and artificial neurons and the structure of a deep neural network. (A) A biological neuron receives input signals via its dendrites, integrates them in the soma, and transmits output signals via its axon to synaptic terminals. (B) An artificial neuron mimics this process by computing a weighted sum of inputs and applying an activation function to produce an output. (C) Deep neural networks are artificial neural networks composed of multiple hidden layers and designed to solve complex nonlinear problems.
Fig. 5. Representative example of muscle segmentation using a deep learning (DL)-based model. (A) Original transverse US image of the suprahyoid region. (B) Expert manual annotation outlining the digastric, mylohyoid and geniohyoid. (C) Automated segmentation generated by the DL model, illustrating the delineation of each suprahyoid muscle layer. (D) Segmentation masks for the individual muscles: digastric (left), mylohyoid (middle), and geniohyoid (right). The automated outputs show strong concordance with expert tracings, demonstrating the model’s ability to accurately identify and segment multilayer suprahyoid muscle structures. US, ultrasound.
Fig. 1.
Fig. 2.
Fig. 3.
Fig. 4.
Fig. 5.
Muscle ultrasound as a promising tool: clinical applications and the emerging role of deep learning
