CMSCNet: a context based lightweight musculoskeletal ultrasound image segmentation method

Introduction

Musculoskeletal ultrasound (MSUS) is the application of ultrasound technology in the diagnosis and treatment of musculoskeletal diseases (1). It uses high-frequency ultrasound to obtain images of the musculoskeletal system, displaying the hierarchical relationship of soft tissues (such as muscles) and their internal structures. Compared to magnetic resonance imaging (MRI), computed tomography (CT) scans, and X-rays, it has the advantages of being convenient, real-time, non-invasive, and radiation-free. MSUS provides detailed images of muscle texture and echogenicity, making it a valuable tool for assessing muscle injuries, monitoring rehabilitation, and screening for muscle-related conditions like sarcopenia (2).

Skeletal muscle, comprising 40–45% of body weight, is vital for posture maintenance and physical activities through its functions of contraction and relaxation. Obtaining accurate skeletal muscle morphological parameters is of great significance in sports training, rehabilitation, and medical applications. Musculoskeletal models can be developed by estimating muscle strength through measurements of muscle fiber length and physiological cross-sectional area, which serve as valuable tools in treating gait disorders and conducting biomechanical research (3). Muscle morphological parameters can be used for rapid screening of sarcopenia (4,5). In sports, musculoskeletal parameters are effective indicators for predicting physical activity capabilities and evaluating training outcomes (6,7). Quantitative analysis of longitudinal MSUS images can yield important parameters reflecting muscle condition, including penniform angle, muscle thickness, and fiber length (8). It can also be used to observe the inflammatory state of muscles, providing morphological features that guide severity assessment and rehabilitation training.

However, ultrasound-based muscle evaluation is highly specialized and prone to subjective factors, such as the experience of the sonographer, which often influence measurement accuracy. The presence of complex noise in skeletal muscle ultrasound images also makes quantitative analysis time-consuming. Clinicians often rely on manual annotation of muscle thickness which is prone to error (9). Furthermore, pathological changes in muscles can obscure distinction between muscle texture and equipment noise, complicating diagnosis (10). Therefore, computer-aided methods are essential to reduce the time and effort required for analyzing MSUS image information.

Recent advancements in medical image segmentation, particularly through deep learning, have shown promise in addressing these challenges. Medical image segmentation divides images into clinically relevant regions, enabling efficient data processing and visualization to support medical decision-making. This technology has demonstrated strong clinical utility, exhibiting high specificity and sensitivity in diagnosing conditions such as partial-thickness and full-thickness supraspinatus tendon tears. It has also been validated by radiologists to improve the diagnostic accuracy of rotator cuff lesions (11), and has been proven effective in supporting the clinical diagnosis of early rheumatoid arthritis (12). Initially, image segmentation relied on traditional methods such as edge detection and region growing. Then the advent of deep neural networks has revolutionized the process by enabling automatic feature extraction. In 2015, the fully convolutional network (FCN) (13) pioneered deep learning for semantic segmentation by replacing fully connected layers with convolutional layers, allowing for heatmap outputs instead of class labels. Deconvolution layers were introduced to address the issue of small output size, while skip connections across different depth layers helped minimize detail loss. Medical semantic segmentation, however, faces more challenges due to noise, speckle, and limited annotated data. Jabbar (14) first applied convolutional neural networks (CNNs) to MSUS segmentation, demonstrating their effectiveness despite significant speckle noise. Ronneberger et al. proposed U-Net (15), a U-shaped network that captures both contextual and positional information, making it more suitable for medical image segmentation. Enhancements like attention gates (16), dilated convolutions and deep supervision (17) further improved performance by refining feature weights and expanding receptive fields, respectively. While deep learning algorithms automate feature extraction, their reliance on large annotated datasets remains a bottleneck, particularly since medical image annotation demands specialized clinical expertise. Recent research efforts, including generative adversarial networks (GANs) (18), contrastive learning, semi-supervised learning (19,20) and transformers (21), aim to address issues of data scarcity and long-range dependency. Despite these advancements, current algorithms often achieve high precision at the cost of increased parameters. Medical imaging systems have evolved from laboratory research to clinical applications, giving rise to a new paradigm known as point-of-care ultrasound (POCUS), which involves conducting tests and analyses directly at the patient’s bedside. POCUS, as an important bedside diagnostic tool, has been widely used for rapid hemodynamic and respiratory assessments in intensive care units due to its portability (22), ability to reduce patient evaluation time (23), and enhancement of overall care satisfaction (24). Moreover, POCUS holds great potential for use in resource-limited environments, such as in developing countries, remote areas, and emergency situations. Its small size, portability, and cost-effectiveness make it an invaluable medical tool.

However, mainstream medical imaging models, such as U-Net, are computationally intensive and parameter-heavy, making them difficult to deploy in resource-constrained clinical settings. This limits the widespread application of POCUS technology, especially on portable ultrasound devices. Therefore, the development of lightweight algorithms to meet real-time needs has become crucial for advancing POCUS technology. Lightweight algorithms can not only improve diagnostic efficiency and image interpretation capabilities for physicians but also enable patients to receive visualized diagnoses and treatments without visiting hospitals.

In this paper, we proposed a lightweight segmentation method for contextual pennate muscle ultrasound images, which accurately segments the muscle fascia in longitudinal ultrasound images and provides preliminary morphological parameters for analyzing muscle status. The position of muscle fascia is fundamental for calculating key metrics such as muscle thickness and pennation angle. This automated analysis method can reduce the workload of doctors and provide an effective basis for subsequent related assessments. The method selects U-Net, which performs well in medical image segmentation tasks, as the backbone network. By adding multi-layer perception modules in the encoder and decoder, our method reduces the parameter’s size and improves the computational efficiency. We have fully considered the significant difference in pixel values between the muscle fascicles and muscle fascia in MSUS images. Preliminary experiments found that multiple down-convolutions in the U-Net led to the loss of fascicle details. Therefore, this paper proposes adding a dense atrous convolution (DAC) module to extract contextual features from the images, reduce detail loss, and enhance segmentation accuracy. Furthermore, the integration of spatial and channel reorganization convolution modules reduces the extraction of redundant features in the segmentation task. Compared to other improved networks, our method can automatically, quickly, and accurately extract the fascia structure of pennate muscles, providing doctors with important data for multi-faceted analysis and evaluation of muscle morphology. Recent studies highlight the environmental implications of computationally intensive artificial intelligence (AI) models. For instance, large-scale models like DeepSeek require substantial energy consumption, contributing to increased carbon footprints (25). In resource-limited clinical settings, such as portable ultrasound devices, energy-efficient algorithms are critical not only for real-time performance but also for minimizing environmental impact. By reducing parameters and computational complexity, a context based lightweight musculoskeletal ultrasound image segmentation method (CMSCNet) aligns with sustainable AI practices, enabling low-power deployment without compromising accuracy. This is particularly vital in remote or emergency scenarios where energy resources are constrained.

Our main contributions are as follows:

• We introduce CMSCNet, a lightweight network for muscle fascia segmentation in ultrasound images, which demonstrates outstanding performance. This network also holds potential for the analysis of ultrasound images of other organs, and could contribute to the development of intelligent ultrasound devices in the future.
• We conducted a comprehensive study of existing medical image segmentation algorithms, combined with the advantages of U-Next, optimized the long-range dependency problem of U-Net networks, increased global information processing capabilities, reduced redundant feature extraction, and added dense dilated convolution modules to enable the network to handle detailed features such as muscle fibers better while reducing the number of parameters and achieving better performance.

We present this article in accordance with the CLEAR reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2523/rc).

Methods

Network architecture

We chose U-Net, a widely used backbone network in the field of medical segmentation. As shown in Figure 1, our network adopts an encoder-decoder architecture. Unlike the fully convolutional structure of U-Net, our network incorporates tokenized multilayer perceptron (MLP) layers, as used in U-Next (26), within the encoder and decoder. The tokenized MLP module in U-Next is inspired by the Swin Transformer (27), which improves the standard MLP. This feature extraction module significantly reduces the number of parameters and computational complexity compared to the traditional U-Net, thereby increasing inference speed. Additionally, our encoder and decoder include a spatial and channel reconstruction convolution module (SCConv) (28) to reduce redundant spatial feature extraction, and the DAC (29) module to capture contextual features of the image, enhancing both segmentation speed and accuracy.

Figure 1 The framework of our proposed method. BN, batch normalization; DAC, dense atrous convolution; GeLU, Gaussian error linear unit; MLP, multilayer perceptron; ReLU, rectified linear unit; SCConv, spatial and channel reconstruction convolution module.

Encoder-decoder architecture

During the encoding phase, we set up three 3×3 convolutions with a stride of 1 and padding of 1. After each convolution, batch normalization is applied, followed by 2×2 max-pooling with a stride of 2 for downsampling. In the decoding phase, we use bilinear interpolation instead of transposed convolution for upsampling while employing a symmetrical structure to the encoding phase and skip connections to compensate for the potential loss of detail due to multiple downsamplings in the encoding process. Although the convolutional pooling operations in the encoding phase effectively extract features, they may reduce spatial resolution, making it difficult for the network to recover detailed information. With skip connections, we directly pass the low-level features from the encoding phase to the decoding phase, combining them with the high-level features after upsampling. This way, the network can utilize high-level semantic information for decoding while also leveraging the detailed information in low-level features, thereby reconstructing image details and contours more effectively and achieving more accurate image segmentation results.

During the encoding and decoding processes, we added a tokenized MLP module, which enhances attention to specific positions by shifting the axes of specific channels in the feature map and mapping them to a high-dimensional space through fully connected layers. Next, we use depthwise separable convolution to further encode positional information, which reduces the number of parameters and improves operational efficiency. Finally, the fully connected layers restore the feature map to its original dimensions from the high-dimensional space, ensuring effective information transfer. Additionally, we add the SCConv module to reduce spatial and channel redundancy between features. Given that our network is designed to address the challenges of MSUS image segmentation, we particularly note that compared to MRI images, MSUS images have lower contrast, higher noise, and finer muscle tendon and muscle bundle lines. These factors make conventional medical image segmentation networks less effective in processing such images. To better address this issue, we replaced the tokenized MLP module in the bottleneck layer of the network with a DAC.

DAC module

As shown in Figure 2, the DAC module is inspired by Inception-ResNet-V2, and it achieves the goal of extracting features of different scales by combining atrous convolutions with different dilation rates. The DAC module has four cascades, each with a different number of atrous convolutions. The receptive fields of the four branches are 3, 7, 9, and 19, respectively, with each branch applying a 1×1 convolution to rectify linear activation. As the receptive field of each branch increases, this module can extract features of various sizes. Convolutions with larger receptive fields can capture large-scale objects and more abstract features, while those with smaller receptive fields can precisely extract the details of smaller objects. Therefore, a network with the DAC module is particularly suitable for the segmentation tasks of muscle ultrasound images.

Figure 2 Structure of dense atrous convolution module. C, channel; Conv, convolution; H, height; W, width.

SCConv module

Our network reduces a significant number of standard 3×3 convolutions, which, although greatly reducing the number of parameters, also diminishes the network’s feature extraction capability. We aim to add an efficient module that does not significantly increase the number of parameters while achieving better performance. As shown in Figure 3, the SCConv module consists of a spatial reconstruction unit (SRU) and a channel reconstruction unit (CRU). In Figure 3A, the SRU adopts a “separation-reconstruction” concept. The separation operation aims to separate feature maps rich in information from those with less information. We use the scaling factor γ in group normalization to assess the information content in different feature maps. Rich spatial information reflects greater variation in spatial pixels, resulting in a larger γ value. The normalized relevant weight W_γ is obtained through an equation, and the feature map re-weighted by W_γ is mapped using a sigmoid function and gated based on a threshold, where weights above the threshold are set as W₁ and those below as W₂. Finally, the input features are re-weighted to obtain the information-rich feature x1w and the information-poor redundant feature x2w. The module then uses cross-reconstruction operations to fully utilize the information in both feature maps, enhancing the flow of information. The SRU operation significantly suppresses redundancy in the spatial dimension, but some degree of information redundancy still exists in the channel features.

Figure 3 Structure of spatial and channel reconstruction convolution module. (A) The spatial reconstruction unit. (B) The channel reconstruction unit. Conv, convolution; GWC, group-wise convolution; PWC, point-wise convolution.

To reduce redundancy in the channel dimension, the CRU in Figure 3B adopts a “split-transform-fuse” strategy. The spatially refined feature data obtained in the previous step is split into two parts along the channel dimension. The upper part X_up acts as the main branch, using a combination of group convolution and pointwise convolution, which are efficient and low-cost, to extract rich representative features. The lower part X_low serves as a supplement for feature extraction, using only pointwise convolution to extract shallow features and reusing X_low to obtain more features without additional cost. The two parts are then summed and concatenated. In the fusion operation, the simplified SKNet method is used to adaptively fuse the features Y₁ and Y₂ output from the upper and lower stages. The CRU extracts rich feature information using lightweight convolution operations and reduces spatial information redundancy through simple operations and feature reuse.

Materials

Dataset

The ultrasound image data used in the experiments is sourced from an open-source dataset documented in Cronin’s paper (30). The dataset was collected using four different ultrasound devices and includes images from different populations (athletes, elderly individuals, and young healthy subjects) of the medial and lateral gastrocnemius, vastus lateralis, and tibialis anterior muscle. It also includes different types of movements and contractions. The dataset contains 574 aponeurosis images and 309 fascicle images. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Data augmentation

Instance segmentation tasks require a large amount of annotated data for training. However, because annotating muscle ultrasound data is difficult, there is a limited availability of datasets. Therefore, to better extract features from muscle images and improve model performance, we used three common data augmentation methods: random horizontal flip, random vertical rotation, and random rotation (−10, 10). Figure 4 shows the data images and augmentation results. To enable the model to better learn positional and content information from the images, we also added an advanced data augmentation method, CutMix (31), as shown in Figure 5. This method, proposed in 2019, generates new training data by cutting and exchanging parts of two randomly selected images. The above four data augmentation methods can avoid overfitting during training while not altering the data distribution, effectively improving the model’s generalization ability. We divided the enhanced MSUS dataset into a training set and a test set in an 8:2 ratio. During model training, we will use the training dataset to train the model and the test dataset to evaluate the model’s performance to obtain the best segmentation model.

Figure 4 Result of data augmentation. (A) Original image. (B) Random horizontal flip. (C) Random vertical rotation. (D) Random rotation.

Figure 5 Result of CutMix. (A) Original image A. (B) Original image B. (C) CutMix.

Results

Training

Implementation environment

Our proposed method is developed and implemented based on the PyTorch library. The platform we used is equipped with an Intel Xeon Gold 5118 (2.30GHz) processor, running the Ubuntu 20.04.6 LTS operating system, and an NVIDIA GeForce RTX 2080 Ti graphics processing unit (GPU).

In this study, the model training employs the Adam optimizer, which performs 100k iterations with a batch size of 8 and an initial learning rate of 10⁻⁵, and a decay coefficient of 10⁻⁸. The optimizer adjusts the learning rates of each parameter dynamically through adaptive moment estimation (first-order moment β1 =0.9, second-order moment β2 =0.999), which is suitable for optimizing non-stationary objective functions. To enhance the model’s convergence, we introduce the Cosine Annealing Warm Restarts learning rate scheduler, with core parameters set as an initial period T0 =1 and a period multiplier Tmult =2. This exponentially increasing period strategy allows for rapid exploration of the parameter space in the early stages of training and extends the annealing period in the later stages for fine convergence. The minimum learning rate ηmin = 5×10⁻⁵ prevents the training from stalling due to excessive learning rate decay. Additionally, we employ the automatic mixed precision (AMP) training mechanism, which uses GradScaler to dynamically scale the loss function (with an initial scaling factor of 216). While maintaining the computational speed advantage of FP16, this mechanism effectively avoids gradient underflow issues, and its adaptive scaling mechanism can automatically detect gradient overflow and adjust the scaling factor. This combination of parameters ensures numerical stability while balancing training efficiency and model performance.

Evaluation metrics

In this paper, we use the number of parameters, accuracy, F1 score, and intersection over union (IoU) score as metrics to evaluate the effectiveness of the model.

Since the purpose of this paper is to propose a lightweight algorithm, the number of parameters is considered an important metric. The number of parameters refers to the total number of parameters required during model training, including the number of layers in the neural network and the number of neurons in each layer. The number of parameters affects the model’s complexity and performance. A higher number of parameters makes the model more complex, which can lead to overfitting; a lower number of parameters makes the model simpler, which may cause underfitting.

Accuracy measures the proportion of correctly predicted areas, including both the correctly predicted target areas and the correctly predicted background areas, relative to all areas. Therefore, it focuses on overall accuracy—the closer the accuracy is to 1, the more accurate the prediction (32). The F1 score is the harmonic mean of precision and recall, combining the advantages of both to evaluate the model’s performance (33).

The IoU score is a metric that measures the overlap between the predicted result and the ground truth, expressed as the ratio of their intersection to their union. The IoU score ranges from 0 to 1, where a value closer to 1 indicates a higher degree of overlap and a greater similarity between the predicted results and the ground truth (34).

Experiments

Ablation study

In this section, to study the relative effectiveness of each part of the proposed network, we conducted four ablation experiments as shown in Figure 6, to verify the performance of the U-Next network, the SCConv module, and the DAC module. First, we added an SCConv module after three layers of standard convolution in U-Next (UN-onelayer). Compared to the original U-Next, the IoU score increased to 0.7148 after adding one layer of the SCConv module. However, when we modified the bottleneck of the network to use the DAC module in parallel with the residual multi-kernel pooling (RMP) module (CE-onelayer), the IoU score dropped. This may be due to the small variation in the size of the muscle fascia, as this module encodes global information with four different receptive field sizes, which is more suitable for detecting objects of varying sizes. When we set the output channel of the DAC module to 512 (Nospp-onelayer), the number of parameters increased significantly, and the IoU score saw only a small improvement. When only one layer of the SCConv module and DAC module were retained in the network, with the output channel set to 256 (little para no spp), the model achieved the best performance, with an IoU score of 0.7227, which is 1.2% higher than U-Next. The results of the ablation experiments are shown in Table 1. The visualization results of the ablation study are shown in Figure 7.

Figure 6 Structure diagram of the ablation study downsampling. Conv, convolution; CRU, channel reconstruction unit; DAC, dense atrous convolution; MLP, multilayer perceptron; SRU, spatial reconstruction unit.

Figure 7 Visualization results of the ablation study.

Performance comparison

We compared our network with the traditional U-Net and baseline to prove its effectiveness. As shown in Table 2, our network has 90% fewer parameters than the traditional U-Net, but the IoU score is only 0.2% lower. Compared to U-Next, our network’s IoU score improved by 1.2%. Compared to CE-Net (29), our network has only 1/10 of the parameters but the IoU score increased by 1.16%.

As shown in Figure 8, our network captures more details of the muscle aponeurosis and effectively focuses on the junctions between the aponeurosis and muscle fibers, showing minimal differences from the ground truth and achieving very high accuracy. Our proposed algorithm achieves high segmentation accuracy (IoU: 0.7227) while maintaining computational efficiency. This balance is critical for clinical translation. For instance, CMSCNet’s lightweight design allows deployment on portable ultrasound devices, enabling real-time muscle assessment in emergency settings or rural clinics. Compared to resource-intensive models like CE-Net (29M parameters), CMSCNet reduces hardware dependency without compromising diagnostic utility, making it a practical tool for automated muscle parameter extraction in diverse healthcare environments.

Figure 8 Comparison of our network with other state-of-art algorithms.

On a single NVIDIA GeForce RTX 2080 Ti GPU [with 10 GB of video random access memory (VRAM)], our network achieves a training speed of 67 batches per second and an inference speed of 74.4 batches per second, as shown in Figure 9.

Figure 9 Comparison of inference and training speed. GPU, graphics processing unit.

Discussion

Through the analysis of the experimental results, we can observe that the superiority of CMSCNet lies in its ability to automatically extract features from ultrasound images using deep learning, while reducing the number of parameters and achieving a lightweight network structure. This not only enhances the segmentation performance but also maintains computational efficiency. In previous studies focused on obtaining musculoskeletal morphology and parameters from ultrasound images, traditional image segmentation algorithms were predominantly used. For instance, the team led by Caresio used techniques such as first-order derivative Gaussian (FODG) filters, standard Hough transforms, and differential Gaussian filters to achieve an automatic measurement of muscle thickness in longitudinal ultrasound images (35). This algorithm highlights muscle bundles more effectively. Two years later, the same team employed multi-scale filters to enhance connective tissue for identifying upper and lower aponeuroses, achieving automatic detection of transverse ultrasound images and measuring the visible cross-sectional area of muscles (36). To obtain more morphological parameters, Jabbar et al. (37) proposed an edge detection method based on a fuzzy inference system, which is used to automatically measure muscle thickness, perimeter, and other muscle parameters. Zhou et al. (38) started with image segmentation, taking the coherence of fascicle orientation as a feature, and applied the Lucas-Kanade optical flow algorithm to track manually marked points along the fascicle paths in the segmented images, thereby obtaining fascicle changes. Although these methods can extract parameters and analyze muscle morphology, they still require manual feature extraction from the images, and thus do not achieve true “automation”.

In objective evaluations, CMSCNet achieved favorable results in terms of accuracy, F1 score, and IoU score, which not only demonstrate the overall effectiveness of the network but also validate the effectiveness of each module. In subjective evaluations, our segmentation results were able to capture the details at the junction between muscle aponeurosis and muscle fibers more effectively than other advanced segmentation algorithms, and were closer to the ground truth. However, as shown in Figure 10, our segmentation results were less satisfactory when dealing with high echogenicity muscle fibers, where the echogenicity exceeds that of the aponeurosis and the boundaries are blurred. First, the performance of the ultrasound device, the skill level of the operator, and the diversity of the target structures can all affect imaging clarity and contrast, thereby impacting segmentation results. For example, Figure 10A is a typical case where the image quality is poor, with no clear boundary between the segmentation object and the background, and black shadows in the middle of the image. These factors affect the segmentation quality, as the deep fascia does not appear as a continuous straight line as shown in the ground truth but as segmented short lines. In Figure 10B, our segmentation results mistakenly included some muscle fibers as part of the segmented object, indicating that when the grayscale values of the foreground and background are too similar, networks relying on image grayscale values for segmentation may misclassify. Therefore, we will further improve this issue in our future research.

Figure 10 The segmentation performance on the public dataset is poor. (A) Poor-quality image with ambiguous boundaries and artifacts affecting fascia continuity; (B) misclassification of muscle fibers due to similar foreground-background grayscale values.

Our study has several limitations that warrant further investigation:

• Limited dataset diversity and quality: the currently available public datasets for leg muscle ultrasound images are small in scale and primarily sourced from older ultrasound systems. This may limit the model’s generalizability to modern high-resolution images or diverse populations (e.g., pediatric or pathological cases). To address this, we are constructing our own MSUS dataset, which will include a diverse population and be collected by professional sonographers using high-definition imaging devices. In the future, we also plan to collaborate with hospitals to obtain high-quality images. Additionally, we plan to incorporate semi-supervised learning (e.g., contrastive learning) and synthetic data generation (via GANs) to mitigate annotation bottlenecks.
• Device-specific performance variability: differences in image contrast and noise levels across ultrasound systems may degrade segmentation accuracy. Future work will incorporate domain adaptation techniques to harmonize feature distributions between devices, ensuring robustness in clinical deployment.
• Restricted parameter extraction scope: CMSCNet currently analyzes only longitudinal ultrasound images, limiting its ability to measure key parameters like muscle cross-sectional area. We plan to extend the framework to transverse image segmentation and incorporate optical flow algorithms for dynamic fascicle length tracking in video sequences.
• Loss function sensitivity: while the current model achieves high accuracy, boundary delineation in high-echogenicity regions (e.g., Figure 10) remains challenging. A comparative study of loss functions (e.g., Dice loss, Tversky loss) will be conducted to improve edge detection.

Conclusions

This paper introduces CMSCNet, a lightweight MSUS segmentation framework based on U-Net. It integrates tokenized MLP modules, DAC modules, and spatial-channel reconstruction convolutions to achieve high-precision automated fascial segmentation. Extensive evaluations on a public leg muscle ultrasound dataset demonstrate its effectiveness: CMSCNet achieves an IoU of 0.7227 with only 10% of U-Net’s parameters and enables real-time processing at 74.4 batches per second. Subjective analysis further confirms its ability to capture fine anatomical details, such as aponeurosis-muscle fiber junctions, with continuity comparable to manual annotations.

CMSCNet’s lightweight design allows it to be deployed on POCUS devices for real-time muscle assessment in emergency departments, rehabilitation centers, and sports medicine settings. The model operates efficiently on low-cost hardware (e.g., a single GPU), significantly reducing dependence on high-end computational infrastructure. Moreover, CMSCNet’s automated segmentation capability can quickly process tasks such as muscle thickness measurement. Beyond technical advancements, CMSCNet addresses key clinical and operational challenges. It eliminates time-consuming manual measurements (e.g., muscle thickness, pennation angle) through automated fascial segmentation, reducing diagnostic delays in emergencies such as acute muscle tears. The algorithm provides objective metrics to reduce inter-observer variability, thereby enhancing the consistency of longitudinal monitoring of sarcopenia or rehabilitation progress. Compared to U-Net, CMSCNet reduces the parameter count by 90%, significantly lowering computational costs. If deployed on portable ultrasound devices, it could reduce energy consumption by over 90%, aligning with eco-friendly healthcare initiatives (39).

Future work will focus on expanding CMSCNet to transverse ultrasound analysis, optimizing its compatibility with edge devices through model quantization, and validating its clinical utility through multi-center trials. These efforts aim to bridge the gap between AI innovation and practical healthcare delivery, especially in resource-limited settings.

Acknowledgments

An earlier version of this abstract was disseminated as a preprint on SSRN. We acknowledge the support of SSRN for facilitating early access to our research findings.

Footnote

Reporting Checklist: The authors have completed the CLEAR reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2523/rc

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2523/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Ozçakar L, Tok F, De Muynck M, Vanderstraeten G. Musculoskeletal ultrasonography in physical and rehabilitation medicine. J Rehabil Med 2012;44:310-8. [Crossref] [PubMed]
2. Wang Q, Huang QH, Yeow JTW, Pickering MR, Saarakkala S. Quantitative Analysis of Musculoskeletal Ultrasound: Techniques and Clinical Applications. Biomed Res Int 2017;2017:9694316. [Crossref] [PubMed]
3. Larsson L, Grimby G, Karlsson J. Muscle strength and speed of movement in relation to age and muscle morphology. J Appl Physiol Respir Environ Exerc Physiol 1979;46:451-6. [Crossref] [PubMed]
4. Wei M, Meng D, Guo H, He S, Tian Z, Wang Z, Yang G, Wang Z. Hybrid Exercise Program for Sarcopenia in Older Adults: The Effectiveness of Explainable Artificial Intelligence-Based Clinical Assistance in Assessing Skeletal Muscle Area. Int J Environ Res Public Health 2022;19:9952. [Crossref] [PubMed]
5. Zhu S, Lin W, Chen S, Qi H, Wang S, Zhang A, Cai J, Lai B, Sheng Y, Ding G. The correlation of muscle thickness and pennation angle assessed by ultrasound with sarcopenia in elderly Chinese community dwellers. Clin Interv Aging 2019;14:987-96. [Crossref] [PubMed]
6. Farup J, Kjølhede T, Sørensen H, Dalgas U, Møller AB, Vestergaard PF, Ringgaard S, Bojsen-Møller J, Vissing K. Muscle morphological and strength adaptations to endurance vs. resistance training. J Strength Cond Res 2012;26:398-407. [Crossref] [PubMed]
7. Wen Z, Xu X, Zhu S. Correlations between gastrocnemius morphology parameters and physical activity capacity in elderly females under high-frequency ultrasound. Chinese Journal of Tissue Engineering Research 2025;29:1058-63.
8. Guo JY, Zheng YP, Xie HB, Koo TK. Towards the application of one-dimensional sonomyography for powered upper-limb prosthetic control using machine learning models. Prosthet Orthot Int 2013;37:43-9. [Crossref] [PubMed]
9. 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. [Crossref] [PubMed]
10. Zaidman CM, Wu JS, Kapur K, Pasternak A, Madabusi L, Yim S, Pacheck A, Szelag H, Harrington T, Darras BT, Rutkove SB. Quantitative muscle ultrasound detects disease progression in Duchenne muscular dystrophy. Ann Neurol 2017;81:633-40. [Crossref] [PubMed]
11. Gupta R, Elamvazuthi I, Dass SC, Faye I, Vasant P, George J, Izza F. Curvelet based automatic segmentation of supraspinatus tendon from ultrasound image: a focused assistive diagnostic method. Biomed Eng Online 2014;13:157. [Crossref] [PubMed]
12. Chang CW, Chang CY, Zhu YX, Wang ST. Wrist joint synovial hypertrophy and effusion detection in musculoskeletal ultrasound images using self-attention U-net. Multimedia Tools and Applications 2024;83:89317-34.
13. Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation. Conference on Computer Vision and Pattern Recognition. IEEE; 2015:3431-40.
14. Jabbar SI, Day CR, Heinz N, Chadwick EK. Using convolutional neural network for edge detection in musculoskeletal ultrasound images. 2016 International Joint Conference on Neural Networks (IJCNN). IEEE; 2016:4619-26.
15. Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF. editors. Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. Cham: Springer International Publishing; 2015:234-41.
16. Michard H, Luvison B, Pham QC, Morales-Artacho AJ, Guilhem G. AW-Net: Automatic muscle structure analysis on B-mode ultrasound images for injury prevention. Proceedings of the 12th ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics. Gainesville, FL, USA: ACM; 2021:1-9.
17. Kuok CP, Yang TH, Tsai BS, Jou IM, Horng MH, Su FC, Sun YN. Segmentation of finger tendon and synovial sheath in ultrasound image using deep convolutional neural network. Biomed Eng Online 2020;19:24. [Crossref] [PubMed]
18. Cronin NJ, Finni T, Seynnes O. Using deep learning to generate synthetic B-mode musculoskeletal ultrasound images. Comput Methods Programs Biomed 2020;196:105583. [Crossref] [PubMed]
19. Hu X, Zeng D, Xu X, Shi Y. Semi-supervised contrastive learning for label-efficient medical image segmentation. In: De Bruijne M, Cattin PC, Cotin S, Padoy N, Speidel S, Zheng Y, Essert C. editors. Medical Image Computing and Computer Assisted Intervention - MICCAI 2021. Cham: Springer International Publishing; 2021:481-90.
20. Bai Y, Chen D, Li Q, Shen W, Wang Y. Bidirectional copy-paste for semi-supervised medical image segmentation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023:11514-24.
21. ChenJLuYYuQLuoXAdeliEWangYLuLYuilleALZhouY. TransUNet: Transformers make strong encoders for medical image segmentation.2021. arXiv:2102.04306.
22. Zieleskiewicz L, Muller L, Lakhal K, Meresse Z, Arbelot C, Bertrand PM, et al. Point-of-care ultrasound in intensive care units: assessment of 1073 procedures in a multicentric, prospective, observational study. Intensive Care Med 2015;41:1638-47. [Crossref] [PubMed]
23. Szwamel K, Polański P, Kurpas D. Experiences of family physicians after a CME ultrasound course. Family Medicine & Primary Care Review 2017;19:62-9.
24. Bornemann P, Bornemann G. Military family physicians' perceptions of a pocket point-of-care ultrasound device in clinical practice. Mil Med 2014;179:1474-7. [Crossref] [PubMed]
25. Strubell E, Ganesh A, McCallum A. Energy and policy considerations for modern deep learning research. Proceedings of the AAAI Conference on Artificial Intelligence 2020;34:13693-6.
26. Valanarasu JMJ, Patel VM. UNeXt: MLP-based rapid medical image segmentation network. In: Wang L, Dou Q, Fletcher PT, Speidel S, Li S. editors. Medical Image Computing and Computer Assisted Intervention - MICCAI 2022. Cham: Springer Nature Switzerland; 2022:23-33.
27. LiuZLinYCaoYHuHWeiYZhangZLinSGuoB.Swin Transformer: Hierarchical Vision Transformer using Shifted Windows.2021. arXiv:2103.14030.
28. Li J, Wen Y, He L. Scconv: Spatial and channel reconstruction convolution for feature redundancy. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE; 2023:6153-62.
29. Gu Z, Cheng J, Fu H, Zhou K, Hao H, Zhao Y, Zhang T, Gao S, Liu J. CE-Net: Context Encoder Network for 2D Medical Image Segmentation. IEEE Trans Med Imaging 2019;38:2281-92. [Crossref] [PubMed]
30. CroninNJFinniTSeynnesO. Fully automated analysis of muscle architecture from B-mode ultrasound images with deep learning.2020. arXiv:200904790.
31. Yun S, Han D, Oh SJ, Chun S, Choe J, Yoo Y. Cutmix: Regularization strategy to train strong classifiers with localizable features. Proceedings of the IEEE/CVF International Conference on Computer Vision. IEEE; 2019:6023-32.
32. Swets JA. Measuring the accuracy of diagnostic systems. Science 1988;240:1285-93. [Crossref] [PubMed]
33. LiptonZCElkanCNarayanaswamyB. Thresholding classifiers to maximize F1 score.2014. arXiv:14021892.
34. Rezatofighi H, Tsoi N, Gwak J, Sadeghian A, Reid I, Savarese S. Generalized intersection over union: A metric and a loss for bounding box regression. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE; 2019:658-66.
35. Caresio C, Salvi M, Molinari F, Meiburger KM, Minetto MA. Fully Automated Muscle Ultrasound Analysis (MUSA): Robust and Accurate Muscle Thickness Measurement. Ultrasound Med Biol 2017;43:195-205. [Crossref] [PubMed]
36. Salvi M, Caresio C, Meiburger KM, De Santi B, Molinari F, Minetto MA. Transverse Muscle Ultrasound Analysis (TRAMA): Robust and Accurate Segmentation of Muscle Cross-Sectional Area. Ultrasound Med Biol 2019;45:672-83. [Crossref] [PubMed]
37. Jabbar SI, Day CR, Chadwick EK. Using fuzzy inference system for detection the edges of musculoskeletal ultrasound images. 2019 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE). IEEE; 2019:1-7.
38. Zhou GQ, Zheng YP. Automatic Fascicle Length Estimation on Muscle Ultrasound Images With an Orientation-Sensitive Segmentation. IEEE Trans Biomed Eng 2015;62:2828-36. [Crossref] [PubMed]
39. Schwartz R, Dodge J, Smith NA, Etzioni O, Green AI. Communications of the ACM 2020;63:54-63.