Skeletal muscle ultrasonic image recognition method and device based on machine learning

Abstract

The invention provides a skeletal muscle ultrasonic image recognition method and device based on machine learning, and relates to the technical field of skeletal muscle ultrasonic image recognition. The method comprises the following steps: acquiring skeletal muscle ultrasonic images, and labeling skeletal muscle groups and health conditions corresponding to the skeletal muscle ultrasonic images to form a skeletal muscle image training set; training a convolutional neural network based on multi-scale feature extraction by using the skeletal muscle image training set to obtain a skeletal muscle ultrasonic image recognition model; and utilizing the trained skeletal muscle ultrasonic image recognition model to recognize the health condition of the corresponding skeletal muscle group in the skeletal muscle ultrasonic image. The method solves the problems that a conventional method can only perform a simple classification task for muscle-bone ultrasound, is limited by inherent defects of more noise, unclear tissue structure boundary and the like in an ultrasound image, and cannot complete complex lesion recognition and accurate segmentation of the lesion. According to the method, the analysis precision and efficiency of the complex muscle-bone ultrasonic image are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of skeletal muscle ultrasound image recognition, and particularly relates to a skeletal muscle ultrasound image recognition method and device based on machine learning. Background Art

[0002] Sports injuries are one of the common injuries in daily life and an important factor affecting the physical health of residents. The most common ones are acute and chronic skeletal muscle injuries and various aseptic inflammations after the injuries, which directly affect the daily work and life of residents. In the past, the evaluation indexes for acute and chronic skeletal muscle injuries mainly relied on the main examination methods for diagnosing muscle injuries, including imaging examinations such as electromyography (EMG), computed tomography (CT), magnetic resonance imaging (MRI), and ultrasound examination (US), as well as serum biochemical indexes to indirectly reflect the degree of the disease and the prognosis. However, clinically, since EMG is affected by many factors, the myoelectric signal is easily affected by adjacent muscles, the environment, and noise, etc., and only the change of the myoelectric signal can be observed, and the morphological information after muscle injury cannot be obtained. MRI has excellent tissue resolution for soft tissues and can accurately judge the location, degree, and scope of muscle injury, etc., and is suitable for evaluating muscle diseases, especially for showing traumatic and ischemic pre-diseases. However, MRI examination is not real-time, and the equipment is large, so it cannot be carried out as a routine examination item, and the examination cost is expensive. Therefore, it is not suitable for long-term follow-up after treatment and monitoring the rehabilitation degree of muscles.

[0003] At present, the research on artificial intelligence models for skeletal muscle ultrasound is still in the early stage. Some scholars have studied the two-dimensional ultrasound inductive content of muscle images through deep learning. According to the segmented images, a deep convolutional neural network is trained, but the diagnostic accuracy can only reach 55%. However, the research and development of its algorithm do not combine big data and artificial intelligence technologies, and the model data is insufficient, and the accuracy is difficult to be satisfactory. Due to the problems of low contrast, a large amount of noise (such as shadows, speckles, etc.), and blurred edges in ultrasound images, this undoubtedly poses a challenge to segmenting the target from ultrasound images. In the early stage of the research, Potonik et al. proposed an ultrasound image segmentation method combining spatial constrained kernel fuzzy clustering and edge-based active contours, and applied the distance regularized level set function. Giraldo et al. proposed a method based on Bayesian shape models and initialized the shape model by calculating the average shape of the nerve structures in the training set. However, these traditional methods rely on prior information to extract features, and have limited representation of the lesion area and a high error rate in complex backgrounds.

[0004] Traditional machine learning-based methods usually adopt relevant signal processing techniques to extract features from musculoskeletal ultrasound images and classify them. Traditional methods mainly include artificial neural networks, support vector machines, clustering algorithms, decision trees, etc. These methods rely on manual extraction and selection of features and use the extracted features to train the model to complete different classification tasks in various musculoskeletal ultrasounds.

[0005] Currently, traditional methods can only perform simple classification tasks for musculoskeletal ultrasounds. Limited by inherent defects such as more noise and unclear tissue structure boundaries in ultrasound images, they cannot complete complex lesion recognition and precise segmentation of lesions. At the same time, traditional methods are not completely data-driven. In many cases, they require real-time intervention by clinicians and rely on doctors' work experience, which also increases the misdiagnosis and missed diagnosis rates of patients' conditions clinically.

[0006] Therefore, there is an urgent need for a method for identifying skeletal muscle ultrasound images to solve the problems that current traditional methods can only perform simple classification tasks for musculoskeletal ultrasounds, are limited by inherent defects such as more noise and unclear tissue structure boundaries in ultrasound images, cannot complete complex lesion recognition and precise segmentation of lesions, and in many cases, traditional methods require real-time intervention by clinicians, rely on doctors' work experience, and also increase the misdiagnosis and missed diagnosis rates of patients' conditions clinically. Summary of the Invention

[0007] The purpose of the present invention is to provide a machine learning-based method and device for identifying skeletal muscle ultrasound images to solve the problem that current traditional methods can only perform simple classification tasks for musculoskeletal ultrasounds, are limited by inherent defects such as more noise and unclear tissue structure boundaries in ultrasound images, and cannot complete complex lesion recognition and precise segmentation of lesions.

[0008] To achieve the above purpose, the present invention adopts the following technical solutions: In the first aspect, this specification provides a machine learning-based method for identifying skeletal muscle ultrasound images, including: Obtain skeletal muscle ultrasound images and label the corresponding skeletal muscle groups and health conditions of the skeletal muscle ultrasound images, Form a skeletal muscle image training set; wherein the skeletal muscle ultrasound images include two-dimensional ordinary ultrasound images, tissue hardness maps, and confidence maps; Train a convolutional neural network based on multi-scale feature extraction using the skeletal muscle image training set to obtain a skeletal muscle ultrasound image recognition model; wherein the convolutional neural network based on multi-scale feature extraction includes an input layer, a multi-modal data fusion layer, a multi-scale feature extraction layer, and an output layer; wherein the input layer is used to receive several groups of skeletal muscle ultrasound images; the multi-modal data fusion layer is used to fuse the two-dimensional ordinary ultrasound image, tissue hardness map, and confidence map in each group of skeletal muscle ultrasound images at the feature level through physical association modeling, adaptive weight allocation, and noise suppression mechanisms to form a larger input tensor; the multi-scale feature extraction layer is used to accurately identify target regions of different shapes and sizes from complex skeletal muscle ultrasound training images based on a multi-scale parallel convolution structure and a Mamba module, and to accurately extract the features of the target regions using convolution kernels of different sizes. Use the trained skeletal muscle ultrasound image recognition model to identify the health status of the corresponding skeletal muscle group in the skeletal muscle ultrasound image.

[0009] In a second aspect, this specification provides a skeletal muscle ultrasound image recognition device based on machine learning, including: a skeletal muscle image training set formation module, a skeletal muscle ultrasound image recognition model training module, and a skeletal muscle ultrasound image recognition module, wherein: The skeletal muscle image training set formation module is used to obtain skeletal muscle ultrasound images and label the corresponding skeletal muscle groups and health statuses of the skeletal muscle ultrasound images to form a skeletal muscle image training set; wherein the skeletal muscle ultrasound images include two-dimensional ordinary ultrasound images, tissue hardness maps, and confidence maps. The skeletal muscle ultrasound image recognition model training module is used to train a convolutional neural network based on multi-scale feature extraction using the skeletal muscle image training set to obtain a skeletal muscle ultrasound image recognition model; wherein the convolutional neural network based on multi-scale feature extraction includes an input layer, a multi-modal data fusion layer, a multi-scale feature extraction layer, and an output layer; wherein the input layer is used to receive several groups of skeletal muscle ultrasound images; the multi-modal data fusion layer is used to fuse the two-dimensional ordinary ultrasound image, tissue hardness map, and confidence map in each group of skeletal muscle ultrasound images at the feature level through physical association modeling, adaptive weight allocation, and noise suppression mechanisms to form a larger input tensor; the multi-scale feature extraction layer is used to accurately identify target regions of different shapes and sizes from complex skeletal muscle ultrasound training images based on a multi-scale parallel convolution structure and a Mamba module, and to accurately extract the features of the target regions using convolution kernels of different sizes. The skeletal muscle ultrasound image recognition module is used to use the trained skeletal muscle ultrasound image recognition model to identify the health status of the corresponding skeletal muscle group in the skeletal muscle ultrasound image.

[0010] Based on the above technical solutions, the present specification can achieve the following technical effects: The present specification provides a method and device for identifying skeletal muscle ultrasound images based on machine learning. The method acquires skeletal muscle ultrasound images, annotates the corresponding skeletal muscle groups and health conditions of the skeletal muscle ultrasound images to form a skeletal muscle image training set; uses the skeletal muscle image training set to train a convolutional neural network based on multi-scale feature extraction to obtain a skeletal muscle ultrasound image recognition model; and uses the trained skeletal muscle ultrasound image recognition model to identify the health conditions of the corresponding skeletal muscle groups in the skeletal muscle ultrasound images. This method breaks through the limitations of traditional single-scale convolution, effectively improves the Dice coefficient of target segmentation under complex backgrounds (such as interlaced muscle bundles and speckle noise), balances the recognition accuracy and computational efficiency, thereby enhancing the recognition robustness of the target area under complex backgrounds, and also significantly improving the analysis accuracy and efficiency of complex musculoskeletal ultrasound images. It can effectively complete the classification, recognition and precise segmentation of different tissues in musculoskeletal ultrasound, and has the advantages of high classification accuracy, fast recognition speed and strong segmentation ability for different musculoskeletal tasks. At the same time, this method is completely data-driven, only requiring clinicians to collect clinical data of patients and perform corresponding annotations, without the need to interact with the algorithm in real time during clinical diagnosis, effectively assisting doctors in clinical disease diagnosis, reducing the misdiagnosis and missed diagnosis rates, and reducing the clinical burden on doctors. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 is a schematic diagram of a method for identifying skeletal muscle ultrasound images based on machine learning provided in an embodiment of the present specification; Figure 2 is a schematic diagram of a two-dimensional ultrasound image provided in an embodiment of the present specification; Figure 3 is a schematic diagram of a tissue hardness map provided in an embodiment of the present specification; Figure 4 is a schematic diagram of a confidence map provided in an embodiment of the present specification; Figure 5 is a schematic diagram of a comparison between ResNet and ConvNext Block provided in an embodiment of the present specification; Figure 6 is a schematic diagram of a neural network model of a skeletal muscle ultrasound image provided in an embodiment of the present specification; Figure 7 is a schematic diagram of a device for identifying skeletal muscle ultrasound images based on machine learning provided in an embodiment of the present specification. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0012] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. According to the following description and claims, the advantages and features of the present invention will be clearer. It should be noted that the accompanying drawings are all in a very simplified form and are applicable to non-precise scales, only for conveniently and clearly assisting in explaining the purpose of the embodiments of the present invention.

[0013] It should be noted that, in order to clearly illustrate the content of the present invention, the present invention specifically gives multiple embodiments to further illustrate different implementation manners of the present invention. Among them, the multiple embodiments are enumerative rather than exhaustive. In addition, for the sake of simplicity of description, the content already mentioned in the previous embodiments is often omitted in the subsequent embodiments. Therefore, the content not mentioned in the subsequent embodiments can be correspondingly referred to the previous embodiments.

[0014] Embodiment 1: This embodiment provides a method for identifying skeletal muscle ultrasound images based on machine learning. See Figures 1-6 ; First of all, it should be noted that in recent years, ultrasound elastography technology is a new type of ultrasound diagnostic technology, which has the advantages of real-time and rapid quantitative detection of lesions, portability and miniaturization of equipment, and non-invasive examination. This technology obtains information on the elasticity (hardness) of biological tissues, and the elasticity (or hardness) of biological tissues is closely related to the histopathology of lesions. However, ultrasound elastography technology has relatively high requirements for the technical level and diagnostic ability of operators. Usually, a qualified shear wave elastography ultrasound doctor needs to undergo long-term clinical training. With the development of artificial intelligence and big data technology and their application in the medical field, intelligent assisted diagnosis has become possible.

[0015] Although current deep learning methods have great potential advantages compared with traditional methods and machine learning methods, there is still much room for improvement due to inherent defects such as more noise and unclear tissue structure boundaries in ultrasound images. Therefore, in order to effectively extract features beneficial to downstream tasks such as musculoskeletal classification, detection, and segmentation from ultrasound images with poor imaging quality, it is necessary to redesign the architecture of the feature extraction model.

[0016] As Figure 1 shown, the method specifically includes the following steps: Step 102: Obtain skeletal muscle ultrasound images, and label the skeletal muscle groups and health conditions corresponding to the skeletal muscle ultrasound images to form a skeletal muscle image training set; wherein the skeletal muscle ultrasound images include two-dimensional ordinary ultrasound images, tissue hardness maps, and confidence maps; Among them, the two-dimensional ordinary ultrasound image is the muscle morphology information obtained by ultrasound imaging, which is used to reflect the gray-scale distribution of the tissue structure; see Figure 2 ; The tissue hardness map is a quantification of the elasticity (hardness) value of muscle tissue generated by elastography ultrasound imaging technology, and is used to evaluate the mechanical properties of the lesion area; see Figure 3 ; The confidence map is obtained by performing image processing (such as edge detection and region connectivity analysis) on two-dimensional conventional ultrasound images and tissue hardness maps. The confidence map includes high-confidence regions (such as clear muscle fibers) and low-confidence regions (such as shadows and artifacts); see Figure 4 ; In this embodiment, the muscles around the shoulder joint are used as the target muscle group for training and validating the model, which can be extended to the skeletal muscles of the whole body; Specifically, the categories of the skeletal muscle groups in this embodiment include: supraspinatus muscle, supraspinatus tendon, infraspinatus muscle, deltoid muscle; The health status includes healthy and abnormal.

[0017] Furthermore, this embodiment also provides a technical solution for preprocessing skeletal muscle ultrasound images, specifically: Transform and augment the original skeletal muscle ultrasound images to generate new skeletal muscle ultrasound image samples, increasing the size of the training set; Among them, the image transformation and augmentation methods include: rotation, translation, and cropping.

[0018] Step 104: Train a convolutional neural network based on multi-scale feature extraction using the skeletal muscle image training set to obtain a skeletal muscle ultrasound image recognition model; the convolutional neural network based on multi-scale feature extraction includes an input layer, a multi-modal data fusion layer, a multi-scale feature extraction layer, and an output layer; the input layer is used to receive several groups of skeletal muscle ultrasound images; the multi-modal data fusion layer is used to fuse the two-dimensional conventional ultrasound images, tissue hardness maps, and confidence maps in each group of skeletal muscle ultrasound images at the feature level through physical association modeling, adaptive weight allocation, and noise suppression mechanisms to form a larger input tensor; the multi-scale feature extraction layer is used to accurately identify target regions with different shapes and sizes from complex skeletal muscle ultrasound training images based on a multi-scale parallel convolutional structure and a Mamba module, and accurately extract the features of the target regions; It should be noted that in the field of computer vision, convolutional neural networks have always been the mainstream models. Inspired by the success of Transformer based on the Self-Attention structure in the field of natural language processing and achieving excellent performance in various datasets. Swin Transformer solves the problem of huge computational complexity, but still has its own problems: its implementation is very complex, especially the shifted LSA calculation method is difficult to deploy. The ConvNeXt model is obtained by ResNet borrowing the design ideas of Swin Transformer from five aspects. The CNN series has a clear and simple structure, a richer series of models, a richer hardware platform, and a more complete deployment pipeline. Therefore, the CNN series is still selected, and ConvNextBlock is more excellent. Its structure is as Figure 5 shown.

[0019] As Figure 6 shown, in this embodiment, a segmentation network model of a skeletal muscle ultrasonogram is described in detail for each part as follows: (1) Input layer Different from the traditional two-dimensional ultrasound images, this embodiment adopts an elastic ultrasound detection multi-modal data input mode. The input layer receives data from a group of ultrasound images (original two-dimensional ultrasound images, tissue hardness images, and confidence images). These data will be used as the input features of the network.

[0020] (2) Multi-modal data fusion layer Aiming at the problems of insufficient utilization of modal information, noise sensitivity, and static strategy limitations in traditional multi-modal fusion methods for skeletal muscle ultrasound images, this embodiment realizes precise fusion through physical association modeling, adaptive weight allocation, and noise suppression mechanisms.

[0021] First, construct a gray-scale-elastic correlation matrix between the tissue hardness image and the two-dimensional ultrasound image, quantify the statistical correlation of local elastic and texture features, and extract cross-modal principal components based on matrix decomposition (SVD) to generate a joint feature vector; Among them, the joint feature vector is expressed as: ; where i and j are coordinate information, and (i, j) represents the corresponding position of the pixel point on the graph. , respectively represent the tissue hardness and texture features (such as gray-level co-occurrence matrix energy) of the point (i, j). represents the average value of the elastic values in the corresponding area of the point (i, j) in the tissue hardness image. It represents the average texture feature value at the point (i, j) in the two-dimensional ultrasonic image; i = 1, 2, ……, 512, j = 1, 2, ……, 512; It should be noted that through physical correlation modeling, a gray-scale-elasticity correlation matrix between the tissue hardness map and the two-dimensional ultrasonic image is constructed, cross-modal principal components are extracted, and the tissue hardness value in the tissue hardness map is correlated with the skeletal muscle morphology in the two-dimensional ultrasonic image; A multi-modal dynamic weight assignment mechanism (MDWA) is introduced. Based on a lightweight gated network (MLP), according to the lesion types (including acute injury and chronic inflammation) and confidence distribution, the weights of each modality are dynamically assigned to suppress the noise in the low-confidence region and improve the classification accuracy; It should be noted that acute injury focuses on the two-dimensional ultrasonic morphological features, that is, the dynamic weight value of the two-dimensional ultrasonic will be higher; chronic inflammation focuses on the elastic information of the hardness map, and the dynamic weight value of the tissue hardness map will be higher; Finally, a spatial mask is generated through the confidence map to suppress the noise interference in the low-confidence region. The fusion formula is:

[0022] Among them, That is, Fused Features, which represents the comprehensive features processed by the multi-modal fusion layer, represents the binary confidence mask, ⊙ represents element-wise multiplication, 、 represent the dynamic weights corresponding to the two-dimensional ultrasonic image and the tissue hardness image respectively.

[0023] For the dynamic ultrasonic sequence, a Transformer encoder is introduced to capture the cross-frame elasticity-morphology evolution pattern, support the spatio-temporal feature analysis of the muscle contraction process (such as the elastic recovery rate), and obtain an elasticity-morphology correlation heat map; Among them, the elasticity-morphology correlation heat map highlights the identified damage points. For example, if one point is identified as damaged in 80 out of 100 similar pictures, then this damage point is color-coded, and no marking is done if there is no damage; the darker the color of the point, the higher the probability of damage; In clinical applications, the output elasticity-morphology correlation heat map can visually locate the core damage area, and the dynamic weight assignment reduces the doctor's manual intervention, providing a high-precision, adaptive, and anti-noise multi-modal fusion solution for the intelligent diagnosis of skeletal muscle ultrasound.

[0024] (3) Multi-scale feature extraction layer Based on the convolutional layer, activation function, pooling layer, and fully connected layer of the traditional ConvNext, the multi-scale feature extraction layer adds a multi-scale parallel processing branch. The input data is processed in parallel through three different scales of convolutional kernels, namely 3×3, 5×5, and 7×7. The convolutional kernel combination is dynamically optimized by combining with the scale perception module. Each branch sequentially performs convolution, batch normalization, and non-linear activation, and then adjusts the dimension through channel concatenation and 1×1 convolution to achieve the feature fusion of local details and global structure; Based on this, in order to accurately identify target regions with different shapes and sizes in complex ultrasound images, this embodiment introduces a multi-scale feature extractor, which is different from the traditional convolutional operation that only uses a single-size convolutional kernel to extract features from ultrasound images. The multi-scale feature extractor introduces different sizes of convolutional kernels, embeds a dynamic multi-scale branch in the ConvNext feature extraction layer, performs parallel processing through 3×3, 5×5, and 7×7 multi-scale convolutional kernels, dynamically optimizes the convolutional kernel combination by combining with the scale perception module, and uses a cross-scale attention gating mechanism to weight and fuse multi-resolution features: after each branch sequentially performs convolution, batch normalization (BatchNorm), and non-linear activation (GELU), it reduces the dimension through channel concatenation and 1×1 convolution to achieve the adaptive fusion of local details and global structure. This design breaks through the limitations of traditional single-scale convolution, effectively improves the Dice coefficient of target segmentation under complex backgrounds (such as interlaced muscle bundles and speckle noise), balances the recognition accuracy and computational efficiency, thereby enhancing the recognition robustness of target regions under complex backgrounds and improving the performance of different downstream tasks of musculoskeletal ultrasound.

[0025] Embed the Mamba module hierarchically in the multi-scale feature pyramid of ConvNext. For low-resolution features, use long-sequence state space modeling to capture the continuity of the global muscle morphology; for high-resolution features, focus on local details through short-sequence block modeling, and at the same time introduce a learnable noise suppression factor to dynamically balance the state transition weight and suppress the interference of ultrasound speckle noise.

[0026] Specifically, this solution embeds the Mamba module in the feature extraction layer, combining the local perception advantage of convolution with the long-range dependence modeling ability of the state space model (SSM): First, serialize the feature map in the spatial dimension, and capture the global correlation between pixels through the SSM hidden state update; then stack the depthwise separable convolution (Depthwise Conv) to strengthen the local feature extraction, and add the input and the module output with a residual connection to avoid gradient disappearance. This design uses lightweight linear calculations (the complexity is reduced from O(N of the traditional Transformer 2) Reduced to O(N)), while improving the model's anti-noise performance, the inference speed is increased by 30%. Finally, the ConvNext network integrating multi-scale and Mamba modules can adaptively learn the hierarchical features of ultrasound images, and efficiently complete the classification task through pooling dimensionality reduction and fully connected mapping. The accuracy of the validation set reaches 88.2%, and both the precision and recall exceed 87.5%.

[0027] Based on this, to efficiently model the long-range dependencies between pixels in skeletal muscle ultrasound images and enhance the anti-noise performance, in this embodiment, Mamba modules are hierarchically embedded in the multi-scale feature pyramid of ConvNext. Combining the linear computational complexity of Mamba (O(N)) with the local perception advantage of convolution, this design achieves 58 FPS real-time inference at a resolution of 512×512, significantly improving the analysis accuracy and efficiency of complex musculoskeletal ultrasound images.

[0028] (4) Output layer The output layer is used to normalize the output of the fully connected layer using the Softmax function to determine the health status of the skeletal muscle group in the skeletal muscle ultrasound image to be measured.

[0029] At the same time, this embodiment is extended to a multi-task joint output architecture: simultaneously generating the recognition results of skeletal muscle types and the classification results of health status, and through collaborative training of weighted cross-entropy and Dice loss, achieving efficient collaboration of classification-segmentation tasks. By enhancing the decision interpretability through Grad-CAM heatmaps and lesion severity scores, the improved output layer has achieved comprehensive breakthroughs in accuracy, efficiency, and clinical practicability, forming a closed-loop system from data input to reliable decision-making, providing reliable technical support for intelligent diagnosis of skeletal muscle ultrasound.

[0030] Further, the loss function for the image classification problem uses a loss function based on cross-entropy. This loss function detects each voxel separately and compares the class prediction with the one-hot encoded vector of the target label. The specific formula is as follows:

[0031] where C refers to the category of the target tissue, refers to the predicted probability of category t, refers to the label of category t.

[0032] Step 106: Use the trained skeletal muscle ultrasound image recognition model to recognize the health status of the corresponding skeletal muscle group in the skeletal muscle ultrasound image.

[0033] In summary, in this embodiment, by acquiring skeletal muscle ultrasound images, annotating the corresponding skeletal muscle groups and health conditions of the skeletal muscle ultrasound images to form a skeletal muscle image training set; training a convolutional neural network based on multi-scale feature extraction using the skeletal muscle image training set to obtain a skeletal muscle ultrasound image recognition model; using the trained skeletal muscle ultrasound image recognition model to identify the health conditions of the corresponding skeletal muscle groups in the skeletal muscle ultrasound images. This method breaks through the limitations of traditional single-scale convolution, effectively improves the Dice coefficient of target segmentation under complex backgrounds (such as muscle bundle interlacing and speckle noise), balances the recognition accuracy and computational efficiency, thereby enhancing the recognition robustness of the target area under complex backgrounds, and also significantly improving the analysis accuracy and efficiency of complex musculoskeletal ultrasound images. It effectively completes the classification, recognition and precise segmentation of different tissues in musculoskeletal ultrasound, and has the advantages of high classification accuracy, fast recognition speed and strong segmentation ability for different musculoskeletal tasks. At the same time, this method is completely data-driven, only requiring clinicians to collect clinical data of patients and perform corresponding annotations, without the need to interact with the algorithm in real time during clinical diagnosis, effectively assisting doctors in clinical disease diagnosis, reducing the misdiagnosis and missed diagnosis rates, and reducing the clinical burden on doctors.

[0034] Embodiment 2: This embodiment provides a method for identifying skeletal muscle ultrasound images based on machine learning, which specifically includes the following steps: 1.1 Data annotation and augmentation The deep learning models used in medical image classification are generally supervised training, and data annotation of data is required to obtain the ground truth during the training process. We have more than 800 frames of skeletal muscle ultrasound images, and the specific data is shown in Table 1; Table 1 Dataset categories and scales

[0035] At the same time, each group of ultrasound images has its corresponding tissue hardness annotation image and its corresponding confidence image. Specific dataset image examples can be seen in Figure 2 、 Figure 3 、 Figure 4 ; To improve the generalization ability of the model and reduce the overfitting phenomenon, we augmented these data. Data augmentation generates new samples by transforming the original images, thereby increasing the size of the training set. Data augmentation methods include: Rotation: Rotating the image by a certain angle can increase the diversity of samples without changing the object category. Translation: Translating the image along the x-axis or y-axis to increase the position change of the object in the image. Cropping: Randomly cropping a part of the image to focus on the features of different regions. There are nearly 3000 groups of images in the final dataset.

[0036] 1.2 Design of the neural network model In the field of computer vision, convolutional neural networks have always been the mainstream models. Inspired by the success of the Transformer based on the Self-Attention structure in the field of natural language processing, Microsoft Research Asia proposed the Swin Transformer model in 2021

[13] and achieved excellent performance in various datasets. Swin Transformer solves the problem of huge computational complexity, but it still has its own problems: its implementation is very complex, especially the shifted LSA calculation method is difficult to deploy.

[0037] The ConvNeXt model is obtained by ResNet borrowing the design idea of Swin Transformer from five aspects. The CNN series has a clear and simple structure, a richer series of models, a richer hardware platform, and a more complete deployment pipeline, so the CNN series is still selected, and the ConvNextBlock is more excellent, and its structure is as Figure 5 shown.

[0038] Figure 6 The following shows the segmentation network model for skeletal muscle ultrasound images, which mainly consists of the following parts: (1) Input layer: Different from traditional two-dimensional ultrasound images, this solution adopts an elastic ultrasound detection multi-modal data input mode, and the input layer receives data from a group of ultrasound images (original image, tissue hardness image, and confidence image). Original image: It is a two-dimensional ordinary ultrasound image, which obtains muscle morphology information through B-mode ultrasound imaging and reflects the gray-scale distribution of the tissue structure. Tissue hardness map: Generated based on ultrasound elastography technology, it quantifies the elastic (hardness) value of muscle tissue and is used to evaluate the mechanical properties of the lesion area. Confidence map: Generated through image processing algorithms (such as edge detection, region connectivity analysis), it identifies high-confidence regions (such as clear muscle fibers) and low-confidence regions (such as shadows, artifacts) in the image. These data will be used as the input features of the network.

[0039] (2) Multi-modal data fusion layer Aiming at the problems of insufficient utilization of modal information, noise sensitivity, and static strategy limitations in traditional multi-modal fusion methods for skeletal muscle ultrasound images, this solution proposes a dynamic cross-modal attention fusion network (DCAF-Net), which realizes precise fusion through physical association modeling, adaptive weight allocation, and noise suppression mechanisms.

[0040] Device differences are eliminated through grayscale normalization and elasticity value standardization, and affine transformation is used to achieve multi-modal spatial alignment (unified to a resolution of 512×512). Multi-modal dynamic weight allocation (MDWA) is innovatively introduced, and the fusion weights of each modality are dynamically adjusted based on the lesion type (e.g., acute injury focuses on morphology, and chronic inflammation focuses on elasticity) to improve the classification accuracy. Combined with the adaptive preprocessing engine (APE), the noise and contrast are analyzed by a convolutional autoencoder, and the filter and enhancement parameters are adaptively selected, which improves the Dice coefficient of low-contrast data segmentation; at the same time, confidence-guided input optimization (CGO) is adopted to suppress the noise interference in low-confidence regions with a binary mask, reducing the false detection rate, and laying a high-quality and noise-resistant data foundation for subsequent feature extraction and multi-task output.

[0041] Specifically, it includes the following steps: First, construct the gray-scale elasticity correlation matrix of the hardness map and B-ultrasound image, quantify the statistical correlation of local elasticity and texture features, and extract cross-modal principal components based on matrix decomposition (SVD) to generate a joint feature vector (Ci,j). The calculation formula is: ; Second, design a lightweight gating network (MLP), dynamically allocate the weights of each modality according to the lesion type (classification label) and confidence distribution (e.g., acute injury focuses on the morphological features of B-ultrasound, and chronic inflammation focuses on the elasticity information of the hardness map), and generate a spatial mask through the confidence map to suppress the noise interference in low-confidence regions. The fusion formula is: .

[0042] For dynamic ultrasound sequences, a Transformer encoder is introduced to capture the cross-frame elasticity-morphology evolution pattern, supporting the spatio-temporal feature analysis of the muscle contraction process (such as the elasticity recovery rate). In clinical applications, the output elasticity-morphology correlation heat map can visually locate the core area of the injury, and the dynamic weight allocation reduces the manual intervention of doctors, providing a high-precision, adaptive, and noise-resistant multi-modal fusion solution for the intelligent diagnosis of skeletal muscle ultrasound.

[0043] (3) Feature extraction layer: The feature extraction layer of this embodiment is based on the ConvNext network architecture. By introducing the collaborative optimization of the multi-scale parallel convolution structure and the Mamba module, the feature learning ability of skeletal muscle ultrasound images is significantly improved. Specifically, on the basis of the convolutional layer, activation function (GELU), pooling layer and fully connected layer of the traditional ConvNext, the feature extraction layer adds a multi-scale parallel processing branch: the input data is respectively processed in parallel through convolutional kernels of three different scales of 3×3, 5×5, and 7×7. Each branch sequentially performs convolution, batch normalization (BatchNorm) and non-linear activation, and then adjusts the dimension through channel splicing and 1×1 convolution to achieve the feature fusion of local details (small scale) and global structure (large scale), thereby enhancing the recognition robustness of the target area under complex backgrounds.

[0044] Furthermore, this solution embeds the Mamba module in the feature extraction layer, combining the local perception advantage of convolution and the long-range dependence modeling ability of the state space model (SSM): First, the feature map is serialized in the spatial dimension, and the global correlation between pixels is captured through the SSM hidden state update; then, the depthwise separable convolution (Depthwise Conv) is superimposed to strengthen the local feature extraction, and the input and the module output are added by residual connection to avoid gradient disappearance. This design improves the inference speed by 30% while enhancing the anti-noise performance of the model through lightweight linear calculations (the complexity is reduced from O(N2) of the traditional Transformer to O(N)). Finally, the ConvNext network that fuses the multi-scale and Mamba modules can adaptively learn the hierarchical features of ultrasound images, and efficiently complete the classification task through pooling dimensionality reduction and fully connected mapping. The accuracy of the validation set reaches 88.2%, and both the precision and recall exceed 87.5%.

[0045] (4) Output layer: The output layer usually uses the Softmax function to normalize the output of the fully connected layer to obtain the probability distribution of each category. These probability values can be used to determine the category to which the input image belongs. At the same time, this solution is innovatively extended to a multi-task joint output architecture: synchronously generate the recognition results of skeletal muscle types and the classification results of health status, and realize the efficient collaboration of classification-segmentation tasks through the collaborative training of weighted cross-entropy and Dice loss. By enhancing the decision interpretability through the Grad-CAM heatmap and the lesion severity score, the improved output layer has achieved a comprehensive breakthrough in accuracy, efficiency and clinical practicability, forming a closed-loop system from data input to reliable decision-making, providing reliable technical support for the intelligent diagnosis of skeletal muscle ultrasound.

[0046] 1.3 Loss function For the loss function for image classification problems, the commonly used one is the loss function based on cross-entropy. This loss function detects each voxel separately and compares the class prediction with the one-hot encoded vector of the target label. The specific formula is as follows:

[0047] where C refers to the class of the target tissue, refers to the predicted probability of class t, refers to the label of class t.

[0048] 1.4 Training Techniques In addition to the design of the network architecture, the training process also affects the final performance. This is mainly related to the hyperparameter settings associated with the optimization strategy.

[0049] We used an NVIDIA GeForce RTX 3090 with 24GB of video memory to train the network and employed a training method similar to that of DeiT and Swin Transformer, including: (a) Training ResNets was extended from the initial 90 epochs to 300 epochs.

[0050] (b) The AdamW optimizer was used during the training process.

[0051] (c) New regularization schemes, including Random Depth and Label Smoothing.

[0052] 1.5 Fine-tuning Strategy For the fine-tuning process, the following strategy was followed: (a) Most of the weights of the original pre-trained ConvNext model were retained, and only the last few layers were adjusted. This is because the lower layers are usually responsible for extracting general features, while the higher layers are responsible for learning task-related features.

[0053] (b) A smaller learning rate was used for training to avoid destroying the useful information in the pre-trained model.

[0054] (c) Early stopping was used during the training process to prevent overfitting.

[0055] 1.6 Evaluation Criteria To evaluate the classification effect of the network on dermoscopic images, this experiment used accuracy, precision, and recall to jointly evaluate the classification results.

[0056] In the experiment, positive samples are those that are consistent with the true labels, while negative samples are those that are inconsistent with the true labels. Among them, TP is the number of samples in which the model predicts the positive class as the positive class, FN is the number of samples in which the model predicts the positive class as the negative class, FP is the number of samples in which the model predicts the negative class as the positive class, and TN is the number of samples in which the model predicts the negative class as the negative class.

[0057] Accuracy refers to the probability that the correctly classified samples account for the total number of samples, and can be expressed as: ; Precision refers to the proportion of samples that the model correctly predicts as the positive class among all samples predicted as the positive class, and is expressed as: ; Recall refers to the proportion of samples that the model correctly predicts as the positive class among all samples that are actually the positive class, and can be expressed as: ; 1.7 Result Analysis By training and validating the neural network, we obtained the following results: (a) Training process: The accuracy of the model on the training set gradually increased, and the value of the loss function gradually decreased. This indicates that the model can continuously optimize the parameters to adapt to the data distribution during the learning process.

[0058] (b) Validation process: The augmented training set was divided into a training set and a test set at a ratio of 4:1, and the model was subjected to 5-fold cross-validation after training. Table 2 shows the average results of the five-fold cross-validation of the model. At the same time, the performance of each category on the validation set was statistically analyzed, and the results are shown in Table 3. Since the number of datasets for different muscle groups is different, it has an impact on their accuracy. Generally, the performance of the model on the validation set reached a relatively high level, and all evaluation indicators performed well. This shows that the model has good generalization ability.

[0059] Table 2 Average Results of Cross-Validation

[0060] Table 3 Average Results of Predictions for Different Categories

[0061] In summary, in this embodiment, by obtaining skeletal muscle ultrasound images, annotating the skeletal muscle groups and health conditions corresponding to the skeletal muscle ultrasound images, a skeletal muscle image training set is formed; using the skeletal muscle image training set to train a convolutional neural network based on multi-scale feature extraction to obtain a skeletal muscle ultrasound image recognition model; using the trained skeletal muscle ultrasound image recognition model to identify the health conditions of the corresponding skeletal muscle groups in the skeletal muscle ultrasound images. This method breaks through the limitations of traditional single-scale convolution, effectively improves the Dice coefficient of target segmentation under complex backgrounds (such as interlaced muscle bundles and speckle noise), balances the recognition accuracy and computational efficiency, thereby enhancing the recognition robustness of the target area under complex backgrounds, and also significantly improving the analysis accuracy and efficiency of complex musculoskeletal ultrasound images. It can effectively complete the classification, recognition, and precise segmentation of different tissues in musculoskeletal ultrasound, and has the advantages of high classification accuracy, fast recognition speed, and strong segmentation ability for different musculoskeletal tasks. At the same time, this method is completely data-driven, only requiring clinicians to collect the clinical data of patients and perform corresponding annotations, without the need to interact with the algorithm in real time during clinical diagnosis, effectively assisting doctors in clinical disease diagnosis, reducing the misdiagnosis and missed diagnosis rates, and reducing the clinical burden on doctors.

[0062] Embodiment 3: This embodiment provides a skeletal muscle ultrasound image recognition device based on machine learning. Refer to Figure 7 , including: a skeletal muscle image training set formation module 701, a skeletal muscle ultrasound image recognition model training module 702, and a skeletal muscle ultrasound image recognition module 703, where: The skeletal muscle image training set formation module 701 is configured to obtain skeletal muscle ultrasound images, and annotate the skeletal muscle groups and health conditions corresponding to the skeletal muscle ultrasound images to form a skeletal muscle image training set; where the skeletal muscle ultrasound images include two-dimensional ordinary ultrasound images, tissue hardness maps, and confidence maps; The skeletal muscle ultrasound image recognition model training module 702 is configured to use the skeletal muscle image training set to train a convolutional neural network based on multi-scale feature extraction to obtain a skeletal muscle ultrasound image recognition model; where the convolutional neural network based on multi-scale feature extraction includes an input layer, a multi-modal data fusion layer, a multi-scale feature extraction layer, and an output layer; where the input layer is used to receive several groups of skeletal muscle ultrasound images; the multi-modal data fusion layer is used to fuse the two-dimensional ordinary ultrasound images, tissue hardness maps, and confidence maps in each group of skeletal muscle ultrasound images at the feature level through physical association modeling, adaptive weight assignment, and noise suppression mechanisms to form a larger input tensor; the multi-scale feature extraction layer is used to accurately identify target areas of different shapes and sizes from complex skeletal muscle ultrasound training images based on a multi-scale parallel convolution structure and a Mamba module, and accurately extract the features of the target areas; The skeletal muscle ultrasound image recognition module 703 is used to recognize the health status of the corresponding skeletal muscle group in the skeletal muscle ultrasound image by using the trained skeletal muscle ultrasound image recognition model.

[0063] Optionally, the two-dimensional ordinary ultrasound image is the muscle morphology information obtained by ultrasound imaging, which is used to reflect the gray-scale distribution of the tissue structure; The tissue hardness map is the elastic value of the quantified muscle tissue generated by the elastographic ultrasound imaging technology, which is used to evaluate the mechanical properties of the lesion area; The confidence map is obtained by processing the two-dimensional ordinary ultrasound image and the tissue hardness map. The confidence map includes high-confidence regions and low-confidence regions.

[0064] Optionally, the categories of skeletal muscle groups include: supraspinatus muscle, supraspinatus tendon, infraspinatus muscle, deltoid muscle; the health status includes healthy and abnormal.

[0065] Optionally, the multi-modal fusion layer is specifically used for: Construct a gray-scale-elasticity correlation matrix of the tissue hardness image and the two-dimensional ultrasound image, quantify the statistical correlation between local elasticity and texture features, and extract cross-modal principal components based on matrix decomposition to generate a joint feature vector; Among them, the joint feature vector is expressed as: ; where i and j are coordinate information, and (i, j) represents the position corresponding to the current pixel point on the graph, , respectively represent the tissue hardness and texture features (such as gray-level co-occurrence matrix energy) of the point (i, j); represents the average value of the elastic value in the area corresponding to the point (i, j) in the tissue hardness image, represents the average value of the texture features of the point (i, j) in the two-dimensional ultrasound image; Introduce a multi-modal dynamic weight allocation mechanism, based on a lightweight gating network, and dynamically allocate the weights of each modality according to the lesion type and confidence distribution; Generate a spatial mask through the confidence map to suppress the noise interference in the low-confidence region. The fusion formula is:

[0066] where, represents the comprehensive feature processed by the multi-modal fusion layer, represents the binary confidence mask, ⊙ represents element-wise multiplication, , respectively represent the dynamic weights corresponding to the two-dimensional ultrasound image and the tissue hardness image.

[0067] Optionally, based on the convolutional layer, activation function, pooling layer, and fully connected layer of the traditional ConvNext, the multi-scale feature extraction layer adds a multi-scale parallel processing branch. The input data is processed in parallel through convolutional kernels of three different scales: 3×3, 5×5, and 7×7. The scale-aware module is combined to dynamically optimize the convolutional kernel combination. Each branch sequentially performs convolution, batch normalization, and non-linear activation, and then adjusts the dimension through channel concatenation and 1×1 convolution to achieve the feature fusion of local details and global structure. The Mamba module is hierarchically embedded in the multi-scale feature pyramid of ConvNext. For low-resolution features, long-sequence state space modeling is used to capture the continuity of the global muscle morphology. For high-resolution features, short-sequence block modeling is used to focus on local details. At the same time, a learnable noise suppression factor is introduced to dynamically balance the state transition weights and suppress the interference of ultrasound speckle noise.

[0068] Optionally, the output layer is used to normalize the output of the fully connected layer using the Softmax function to determine the health status of the skeletal muscle group in the skeletal muscle ultrasound image to be measured.

[0069] Optionally, the device further includes an image preprocessing module, where the image preprocessing module is used for: Transform and augment the original skeletal muscle ultrasound image to generate new skeletal muscle ultrasound image samples, increasing the size of the training set. The image transformation and augmentation methods include rotation, translation, and cropping.

[0070] In summary, in this embodiment, by acquiring the skeletal muscle ultrasound image, annotating the corresponding skeletal muscle group and health status of the skeletal muscle ultrasound image, a skeletal muscle image training set is formed; the convolutional neural network based on multi-scale feature extraction is trained using the skeletal muscle image training set to obtain a skeletal muscle ultrasound image recognition model; the trained skeletal muscle ultrasound image recognition model is used to identify the health status of the corresponding skeletal muscle group in the skeletal muscle ultrasound image. This method breaks through the limitations of traditional single-scale convolution, effectively improves the Dice coefficient of target segmentation under complex backgrounds (such as muscle bundle interlacing and speckle noise), balances the recognition accuracy and computational efficiency, thereby enhancing the recognition robustness of the target area under complex backgrounds, and also significantly improving the analysis accuracy and efficiency of complex musculoskeletal ultrasound images. It can effectively complete the classification, recognition, and precise segmentation of different tissues in musculoskeletal ultrasound, and has the advantages of high classification accuracy, fast recognition speed, and strong segmentation ability for different musculoskeletal tasks. At the same time, this method is completely data-driven, only requiring clinicians to collect the clinical data of patients and perform corresponding annotations, without the need to interact with the algorithm in real time during clinical diagnosis, effectively assisting doctors in clinical disease diagnosis, reducing the misdiagnosis and missed diagnosis rates, and reducing the clinical burden on doctors.

[0071] The above description has been made of specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the acts or steps recited in the claims may be performed in a different order than in the embodiments and still achieve the desired results. Additionally, a particular order shown or a sequential order depicted in the drawings is not necessarily required to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0072] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A method for identifying skeletal muscle ultrasound images based on machine learning, characterized in that, Including: Obtain skeletal muscle ultrasound images, and label the skeletal muscle groups and health conditions corresponding to the skeletal muscle ultrasound images to form a skeletal muscle image training set; Wherein the skeletal muscle ultrasound images include two-dimensional ordinary ultrasound images, tissue hardness maps, and confidence maps; Use the skeletal muscle image training set to train a convolutional neural network based on multi-scale feature extraction to obtain a skeletal muscle ultrasound image recognition model; wherein the convolutional neural network based on multi-scale feature extraction includes an input layer, a multi-modal data fusion layer, a multi-scale feature extraction layer, and an output layer; wherein the input layer is used to receive several groups of skeletal muscle ultrasound images; the multi-modal data fusion layer is used to fuse the two-dimensional ordinary ultrasound images, tissue hardness maps, and confidence maps in each group of skeletal muscle ultrasound images at the feature level through physical association modeling, adaptive weight allocation, and noise suppression mechanisms to form a larger input tensor; the multi-scale feature extraction layer is used to accurately identify target regions of different shapes and sizes from complex skeletal muscle ultrasound training images based on a multi-scale parallel convolutional structure and a Mamba module, using convolutional kernels of different sizes, and accurately extract the features of the target regions; Use the trained skeletal muscle ultrasound image recognition model to identify the health conditions of the skeletal muscle groups corresponding to the skeletal muscle ultrasound images.

2. The method according to claim 1, characterized in that The two-dimensional ordinary ultrasound image is the muscle morphology information obtained by ultrasound imaging, which is used to reflect the gray-scale distribution of the tissue structure; The tissue hardness map is the elastic value of the quantified muscle tissue generated by elastic ultrasound imaging technology, which is used to evaluate the mechanical properties of the lesion area; The confidence map is obtained by processing the two-dimensional ordinary ultrasound image and the tissue hardness map, and the confidence map includes high-confidence regions and low-confidence regions.

3. The method according to claim 1, wherein The categories of skeletal muscle groups include: supraspinatus muscle, supraspinatus tendon, infraspinatus muscle, deltoid muscle; The health conditions include healthy and abnormal.

4. The method according to claim 1, wherein The multi-modal fusion layer is specifically used for: Construct a gray-scale-elasticity correlation matrix of the tissue hardness image and the two-dimensional ultrasound image, quantify the statistical correlation of local elasticity and texture features, and extract cross-modal principal components based on matrix decomposition to generate a joint feature vector; Among them, the combined feature vector is expressed as: ; wherein, i and j are coordinate information, and (i, j) represents the corresponding position of the current pixel point on the graph. and respectively represent the tissue hardness and texture features of the point (i, j). represents the average value of the elastic values within the region corresponding to the point (i, j) in the tissue hardness image. represents the average value of the texture features of the point (i, j) in the two-dimensional ultrasound image. Introduce a multi-modal dynamic weight allocation mechanism, based on a lightweight gating network, and dynamically allocate the weights of each modality according to the lesion type and confidence distribution; Generate a spatial mask through the confidence map to suppress the noise interference in the low-confidence region, and the fusion formula is: ； Among them, represents the comprehensive feature processed by the multi-modal fusion layer, represents the binarized confidence mask, and ⊙ represents element-wise multiplication, and respectively represent the dynamic weights corresponding to the two-dimensional ultrasound image and the tissue hardness image.

5. The method according to claim 1, characterized in that On the basis of the convolutional layer, activation function, pooling layer, and fully connected layer of the traditional ConvNext, the multi-scale feature extraction layer adds a multi-scale parallel processing branch. The input data is processed in parallel through convolutional kernels of three different scales: 3×3, 5×5, and 7×7. Combine the scale perception module to dynamically select the convolutional kernel combination. Each branch sequentially performs convolution, batch normalization, and non-linear activation, and then adjusts the dimension through channel splicing and 1×1 convolution to achieve the feature fusion of local details and global structure; Embed the Mamba module hierarchically in the multi-scale feature pyramid of ConvNext. For low-resolution features, adopt long-sequence state-space modeling to capture the continuity of the global muscle morphology; for high-resolution features, focus on local details through short-sequence block modeling, and at the same time introduce a learnable noise suppression factor to dynamically balance the state transition weights and suppress the interference of ultrasound speckle noise.

6. The method according to claim 1, characterized in that, The output layer is used to normalize the output of the fully connected layer using the Softmax function to determine the health status of the skeletal muscle group in the skeletal muscle ultrasound image to be measured.

7. The method according to claim 1, wherein After obtaining the skeletal muscle ultrasound image, the method further includes: Performing transformation augmentation on the original skeletal muscle ultrasound image to generate new skeletal muscle ultrasound image samples and increase the size of the training set; Among them, the image transformation augmentation methods include: rotation, translation, and cropping.

8. A skeletal muscle ultrasound image recognition device based on machine learning, for implementing the method according to any one of claims 1-7, characterized in that, Including: A skeletal muscle image training set formation module, a skeletal muscle ultrasound image recognition model training module, and a skeletal muscle ultrasound image recognition module, where: The skeletal muscle image training set formation module is used to obtain skeletal muscle ultrasound images and label the corresponding skeletal muscle groups and health statuses in the skeletal muscle ultrasound images to form a skeletal muscle image training set; among them, the skeletal muscle ultrasound images include two-dimensional ordinary ultrasound images, tissue hardness maps, and confidence maps; The skeletal muscle ultrasound image recognition model training module is used to train a convolutional neural network based on multi-scale feature extraction using the skeletal muscle image training set to obtain a skeletal muscle ultrasound image recognition model; among them, the convolutional neural network based on multi-scale feature extraction includes an input layer, a multi-modal data fusion layer, a multi-scale feature extraction layer, and an output layer; the input layer is used to receive several groups of skeletal muscle ultrasound images; the multi-modal data fusion layer is used to fuse the two-dimensional ordinary ultrasound image, tissue hardness map, and confidence map in each group of skeletal muscle ultrasound images at the feature level through physical association modeling, adaptive weight allocation, and noise suppression mechanisms to form a larger input tensor; the multi-scale feature extraction layer is used to accurately identify target regions of different shapes and sizes from complex skeletal muscle ultrasound training images based on a multi-scale parallel convolution structure and the Mamba module, and accurately extract the features of the target regions; The skeletal muscle ultrasound image recognition module is used to use the trained skeletal muscle ultrasound image recognition model to recognize the health status of the corresponding skeletal muscle group in the skeletal muscle ultrasound image.

9. The device according to claim 8, characterized in that, The two-dimensional ordinary ultrasound image is the muscle morphology information obtained through ultrasound imaging and is used to reflect the gray-scale distribution of the tissue structure; The tissue hardness map is the elastic value of the quantified muscle tissue generated through elastographic ultrasound imaging technology and is used to evaluate the mechanical properties of the lesion area; The confidence map is obtained by performing image processing on the two-dimensional ordinary ultrasound image and the tissue hardness map, and the confidence map includes high-confidence high regions and low-confidence regions.

10. The device according to claim 8, characterized in that, Based on the convolutional layer, activation function, pooling layer, and fully connected layer of the traditional ConvNext, the multi-scale feature extraction layer adds a multi-scale parallel processing branch. The input data is processed in parallel through convolutional kernels of three different scales: 3×3, 5×5, and 7×7. The scale-aware module is combined to dynamically optimize the convolution kernel combination. Each branch sequentially performs convolution, batch normalization, and non-linear activation, and then adjusts the dimension through channel splicing and 1×1 convolution to achieve the feature fusion of local details and global structures. The Mamba module is hierarchically embedded in the multi-scale feature pyramid of ConvNext. For low-resolution features, long-sequence state space modeling is used to capture the continuity of the global muscle morphology. For high-resolution features, short-sequence block modeling is used to focus on local details. At the same time, a learnable noise suppression factor is introduced to dynamically balance the state transition weights and suppress the interference of ultrasound speckle noise.

Images