A tibialis anterior muscle ultrasonic image intelligent detection method based on sequence modeling and feature enhancement

Abstract

This invention discloses an intelligent detection method for tibial anterior muscle ultrasound images based on sequence modeling and feature enhancement. The method introduces a Mamba model based on a state-space model to recursively model sequence features, effectively expanding the receptive field while maintaining linear computational complexity. A hierarchical shifting window attention module (HSW) is designed, utilizing a window self-attention mechanism to model features in local regions and combining it with a shifting window strategy to achieve cross-window information interaction, enhancing the ability to express local features. By fusing Mamba's global modeling capabilities with the local attention modeling capabilities of the HSW module, complementary expression of local texture information and global semantic information is achieved. A multi-scale feature pyramid structure is constructed, and the fused features are further enhanced by an extended state-space module, enabling the model to effectively utilize feature information at different scales, improving its adaptability to targets at different scales and enhancing detection stability.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and target detection technology, and relates to a high-precision intelligent detection method for tibialis anterior muscle ultrasound images. Background Technology

[0002] The tibialis anterior muscle is a major power muscle of the lower limb, responsible for ankle dorsiflexion and gait control. Ultrasound examination, with its advantages of being non-invasive, radiation-free, and capable of real-time imaging, has become an important tool for clinical examination of the tibialis anterior muscle. However, current clinical examinations mainly rely on manual operation, which is not only time-consuming and labor-intensive, but also significantly influenced by the operator's subjective experience. Therefore, intelligent examination of tibialis anterior muscle ultrasound images has significant clinical value.

[0003] In recent years, deep learning methods have been widely used in medical ultrasound image analysis, providing a new technical path for the automated and intelligent detection of ultrasound images. Convolutional neural networks can automatically extract hierarchical features of images, and YOLO-based target detection algorithms have been widely adopted in ultrasound image detection tasks due to their advantages of strong real-time performance and high detection accuracy, achieving significant results in various medical ultrasound image detection tasks.

[0004] However, speckle noise and blurred boundaries are common problems in ultrasound images of the tibialis anterior muscle, placing high demands on the feature extraction capabilities of detection methods. Traditional convolutional neural network models are limited by their local receptive fields, making it difficult to effectively model long-range global dependencies in images. This results in insufficient perception of the overall structure of the target region, affecting detection accuracy and robustness. To improve global feature modeling capabilities, sequence modeling methods based on State Space Models (SSMs) have received widespread attention in recent years. Among them, the Mamba model, by introducing a selective state space mechanism, achieves efficient modeling of long-distance dependencies while maintaining linear computational complexity, and has shown good application potential in medical image analysis tasks. Combining it with a target detection framework can improve the model's ability to express global contextual information to a certain extent. However, the Mamba model is mainly based on sequence modeling mechanisms, and its ability to characterize two-dimensional spatial structural information is limited. When dealing with targets with complex boundaries, its ability to extract local detailed features is insufficient. Therefore, developing a high-precision detection method that can balance global contextual modeling and local detailed feature extraction has significant clinical value for achieving intelligent detection of tibialis anterior muscle ultrasound images. Summary of the Invention

[0005] To address the shortcomings of existing detection technologies, such as insufficient global context modeling and limited feature extraction capabilities, which lead to low detection accuracy in tibialis anterior muscle ultrasound images, this invention provides an intelligent detection method for tibialis anterior muscle ultrasound images based on sequence modeling and feature enhancement. This method utilizes a Mamba model based on a spatial state model to perform long-distance dependency modeling of the global context information of the ultrasound image. Simultaneously, it introduces a HSW hierarchical shift window attention module to enhance local feature representation capabilities, thus achieving intelligent detection of tibialis anterior muscle ultrasound images.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] A smart detection method for tibialis anterior muscle ultrasound images based on sequence modeling and feature enhancement includes the following steps:

[0008] Step 1: Input the ultrasound image into the backbone feature extraction network, perform convolutional downsampling and serialization processing on the image, and use the state space model to perform global feature modeling and recursive update. The visual cue fusion module performs spatial downsampling and channel recombination to extract multi-scale features. The deep feature representation is output through the spatial pyramid pooling module.

[0009] Step 2: Design and introduce a hierarchical shifted window attention module (HSW) into the MHSW-YOLO backbone network. Input the enhanced feature map into the HSW module, use window self-attention and shifted window mechanism to model and enhance local texture features, and combine the global modeling capability of the state space model to fuse local features and global features, and output an enhanced feature map that fuses local texture information and cross-window context information.

[0010] Step 3: Input the multi-scale features enhanced by the HSW module into the feature pyramid network for feature fusion. Introduce the extended state space module (XSSBlock) to further enhance the fused features. Input the enhanced multi-scale features into the detection head for classification and localization. Output the target category information and bounding box to obtain the trained MHSW-YOLO model. Finally, use the MHSW-YOLO model to detect the tibialis anterior muscle ultrasound image to be detected.

[0011] Compared with the prior art, the present invention has the following advantages:

[0012] 1. This invention introduces the Mamba model based on the state-space model to achieve recursive modeling of sequence features. While maintaining linear computational complexity, it effectively expands the receptive field, thereby enhancing the model's ability to express global structural information in ultrasound images and making up for the shortcomings of traditional convolution in modeling long-distance dependencies.

[0013] 2. This invention designs an HSW hierarchical shifting window attention module, which utilizes a window self-attention mechanism to model features of local regions and combines it with a shifting window strategy to achieve cross-window information interaction, thereby enhancing the ability to express local features. By integrating Mamba's global modeling capabilities with the local attention modeling capabilities of the HSW module, complementary expression of local texture information and global semantic information is achieved.

[0014] 3. This invention constructs a multi-scale feature pyramid structure and combines it with an extended state space module to further enhance the fused features, enabling the model to effectively utilize feature information at different scales and improve its adaptability and detection stability to targets at different scales. Attached Figure Description

[0015] Figure 1 This is a flowchart of an intelligent detection method for ultrasound images of the tibialis anterior muscle.

[0016] Figure 2 This is a labeled image of the tibialis anterior muscle using ultrasound.

[0017] Figure 3 This is a schematic diagram of the MHSW-YOLO network structure.

[0018] Figure 4 This is a comparison of the results of an ultrasound ablation experiment on the tibialis anterior muscle.

[0019] Figure 5 This is a comparison chart of Grad-CAM visualization analysis of ultrasound ablation experiments on the tibialis anterior muscle. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0021] This invention provides an intelligent detection method for tibialis anterior muscle ultrasound images based on sequence modeling and feature enhancement, such as... Figure 1 As shown, the method includes the following steps:

[0022] Step 1: Train the MHSW-YOLO model using ultrasound image data of the tibialis anterior muscle:

[0023] The ultrasound image is input into the backbone feature extraction network, where it undergoes convolutional downsampling and serialization. A state-space model is used for global feature modeling and recursive updating. A visual cue fusion module performs spatial downsampling and channel reconstruction to extract multi-scale features. Finally, a spatial pyramid pooling module outputs a deep feature representation. The specific steps are as follows:

[0024] Step 101: Input the tibialis anterior muscle ultrasound image into the initial feature extraction module (Simple Stem) of the Backbone layer of the MHSW-YOLO model. This module uses two sequentially connected downsampling convolutional blocks to extract features and downsample the input image. The process can be represented as follows:

[0025]

[0026] in, For the input image, , For convolutional blocks, It is a non-linear activation function. For batch normalization operations, This is the output shallow feature map. Through the above operations, the input image is downsampled by 4 times while retaining the fine-grained tissue texture information in the ultrasound image, resulting in a shallow feature map.

[0027] Step 102: After preliminary feature extraction, the shallow feature map is... The input is fed into multiple consecutively distributed Visual State Space Blocks (VSSBlock), which then convert the two-dimensional feature maps into... Expand into a one-dimensional sequence Furthermore, a state-space model is used to perform global context modeling of sequence features, outputting feature maps. The specific modeling steps are as follows:

[0028] First, the continuous state-space model is discretized, transforming the continuous parameters into discrete parameters, which are expressed in discrete form as follows:

[0029]

[0030] in, Let k be the discrete time step at position k. Let k be the feature vector at the k-th position in the sequence. The weight matrix is ​​a learnable matrix. Here is the state transition matrix. For the input control matrix, It is the identity matrix. This is the discretized state transition matrix. This is the discretized input mapping matrix.

[0031] Subsequently, the sequence features are updated linearly and recursively, and the update process is expressed as follows:

[0032]

[0033] in, For the input sequence features, It is in a hidden state. Here is the state transition matrix. For the input mapping matrix, For the output matrix, This serves as the output feature. A recursive update mechanism is used to effectively model long-range dependencies in the sequence. The above recursive state update process based on discretized parameters can be equivalently converted into a convolutional form, specifically expressed as:

[0034]

[0035] in, The convolution kernel generated for the state space parameters, Given the input sequence, This is the output sequence.

[0036] Finally, to enhance feature representation and stabilize deep network training, the VSSBlock module introduces a gated residual structure, the computation of which is expressed as:

[0037]

[0038] in, Given the input feature sequence, For layer normalization, For the state space modeling module, , For learnable gating parameters, For activation function, This is the output feature map. By introducing a gating mechanism to adaptively weight the residual features, the information flow path can be dynamically adjusted according to the saliency of the input features, thereby enhancing the expression of key features and improving the model's ability to model complex ultrasonic texture structures.

[0039] In this step, the VSSBlock module mainly consists of a local feature extraction unit, a sequence state space modeling unit, and a feedforward network unit. First, the local feature extraction unit extracts local texture and spatial structure information from the input features to enhance the representation of local region features. Then, the two-dimensional features are sequentially unfolded according to the spatial dimension, and the SSM state space model is used to model long-distance dependencies of the sequence features. Finally, the feedforward network unit performs nonlinear mapping and channel interaction on the features to further enhance feature representation. The entire module adopts a residual connection structure and combines it with a DropPath random path discarding strategy to improve network training stability and model generalization ability.

[0040] Step 103: Obtain the output feature map after state-space modeling and gated residual structure enhancement. Subsequently, a visual cue fusion module (VisionClueMerge) is introduced to further reduce the feature representation space resolution and expand the receptive field to enhance semantic expressiveness. The specific steps are as follows:

[0041] First, input features The space is divided into 2×2 regions, and features of each sub-region are extracted. , , , This reduces the spatial resolution to the original level. .

[0042] Subsequently, the features of each sub-region are spliced ​​together along the channel dimension to rearrange spatial information along the channel dimension.

[0043] Finally, a 1×1 convolution is introduced to linearly map the concatenated features, achieving channel information fusion and dimensionality compression to obtain the output feature map. The specific process is as follows:

[0044]

[0045] in, Features of sub-regions This is for channel splicing operations. This is the output feature map after spatial downsampling and channel reconstruction.

[0046] By using the aforementioned spatial downsampling and channel recombination, the spatial resolution of feature representation is reduced, while effectively preserving local structural information and achieving feature scale transformation, thereby enhancing feature representation capabilities.

[0047] In this step, the VisionClueMerge module mainly consists of a spatial rearrangement unit and a feature fusion unit. By performing spatial downsampling and channel recombination on the output features, feature scale transformation is achieved, and multi-scale feature representation is gradually constructed. This can reduce the amount of computation while more completely preserving local structural information.

[0048] In the backbone network, the VSSBlock and VisionClueMerge modules are connected in an alternating stacking manner to perform long-distance dependency modeling, spatial downsampling, and channel reorganization on features at the current scale. As the network depth increases, the spatial resolution of the feature maps decreases while the channel dimension gradually increases, thereby achieving a transition from shallow features to deep, highly semantic features and gradually constructing a multi-scale feature representation with rich semantic information.

[0049] Step 104: Input the deep features after feature scale transformation into the Spatial Pyramid Pooling (SPPF) module. Through multiple cascaded max pooling operations, the receptive field is gradually expanded while keeping the spatial size of the feature map unchanged, thereby fusing multi-scale contextual information to obtain a high-level feature representation for subsequent feature modeling.

[0050] Step 2: Design and introduce a hierarchical shifted window attention module (HSW) into the MHSW-YOLO backbone network. Input the enhanced feature map into the HSW module, utilize window self-attention and shifted window mechanisms to model and enhance local texture features, and combine the global modeling capability of the state space model to fuse local and global features, outputting an enhanced feature map that integrates local texture information and cross-window context information. The specific steps are as follows:

[0051] Step 201: In the deep feature extraction stage of the backbone network, the enhanced feature map is input into the HSW module for in-local window self-attention calculation. The specific steps are as follows:

[0052] First, the HSW module will input the feature map. The feature map is saved as a residual branch and normalized to stabilize subsequent attention calculations. The feature map is resized using a window completion unit. To overcome the limitation of local window attention modeling only within the window, a shift window mechanism is introduced in the HSW module, and the feature map is shifted in the cyclic translation unit. This ensures computational efficiency while enabling cross-window information interaction, outputting an enhanced feature map that fuses local texture information and cross-window context information. Simultaneously, an attention mask is introduced to constrain information propagation between different regions and avoid invalid connections.

[0053] Subsequently, the window partitioning unit divides the feature map according to the preset window size. The system is divided into multiple non-overlapping local window regions. Within each local window, multi-head self-attention computation is performed, achieving local feature modeling while reducing computational complexity. Query, key, and value are generated through linear mapping for subsequent calculation of self-attention weights, as shown below:

[0054] , ,

[0055] in, , , For learnable weight matrix, This is the set of all local windows obtained by partitioning the input feature map. , , This represents the query vector, key vector, and value vector.

[0056] Finally, attention weights are obtained by calculating the similarity between the query and the key, and the value vectors are weighted and summed to achieve self-attention modeling of features within the window. The specific calculation process is as follows:

[0057]

[0058] in, Scaling factor , , To represent query vectors, key vectors, and value vectors, This represents the normalization operation. Through the self-attention calculation within the window described above, the model can effectively capture the correlation between pixels within a local region, thereby enhancing its ability to represent the texture structure of ultrasound images.

[0059] Step 202: After completing the attention calculation, the attention output features of each window are obtained. The window restoration unit rearranges and concatenates the features of each local window according to their corresponding spatial positions and removes the padded areas, restoring the complete feature map through the window restoration operation. On this basis, the attention output feature map is residually connected to the residual branch, and channel feature transformation and nonlinear enhancement are performed through the normalization layer and the feedforward feature enhancement unit RGBlock. Then, the output features of the HSW module are obtained through residual connection, thereby realizing the fusion of local texture features and cross-window context information, while improving feature expression ability and network training stability, and providing richer semantic expression for subsequent feature fusion and object detection.

[0060] Step 3: Input the enhanced multi-scale features into the feature pyramid network for feature fusion. Introduce an extended state space module (XSSBlock) to further enhance the fused features. Input the enhanced multi-scale features into the detection head to complete classification and localization, outputting the target category and bounding box, thus obtaining the trained MHSW-YOLO model. Finally, use the MHSW-YOLO model to detect the tibialis anterior muscle ultrasound image to be detected. The specific steps are as follows:

[0061] Step 301: Input the features of different scales output by the backbone network into the feature pyramid network, and fuse the multi-scale features through a dual path of top-down and bottom-up.

[0062] In the top-down path, high-level semantic features are upsampled level by level and concatenated with low-level features of the corresponding scale to obtain feature representations that combine high resolution and rich semantic information. The process is as follows:

[0063]

[0064] in, For upsampling operation, This is for channel splicing operations. Characterized by high-level features This corresponds to low-level features at the corresponding scale. These are the features after top-down fusion.

[0065] In the bottom-up path, low-level features are downsampled using convolutions with a stride of 2 and then fused with high-level features at the corresponding scale. This integrates detailed information from high-resolution features into high-level features, improving the completeness of feature representation, as shown below:

[0066]

[0067] in, For downsampling operation, This is for channel splicing operations. This is a feature of the previous layer of fusion. For the current layer features, These are the features resulting from bottom-up fusion.

[0068] Through the aforementioned dual-path fusion mechanism of top-down and bottom-up, bidirectional transmission of high-level semantic information and low-level spatial detail information is achieved, thereby obtaining a fused multi-scale feature map.

[0069] Step 302: After completing multi-scale feature fusion, an extended state space module (XSSBlock) is introduced to further enhance the fused features. This module normalizes the input features and performs long-distance dependency modeling on the features through state space branches. The modeling results are then fused with the input features through residual connections. A local space modeling module is also introduced to interact with the features in spatial dimensions, further enhancing the expressive power of local structures. Furthermore, a multilayer perceptron branch is introduced into the module to perform nonlinear mapping on the features and fuse them with the backbone features through residual connections, thereby improving feature expressive power and model stability, as detailed below:

[0070]

[0071] in, For input features, Modeling the state space, Modeling local space, For multilayer perceptron mapping, This refers to the enhanced output features.

[0072] Through the aforementioned multi-branch collaborative modeling mechanism, the global and local information of the fused features are jointly enhanced, resulting in high-quality feature representations for subsequent detection tasks.

[0073] Step 303: After completing multi-scale feature enhancement, input feature maps of different scales into the detection head for target prediction. The detection head inputs features at different scales. Perform convolution mappings separately, and output the class probability and bounding box regression parameters at the corresponding positions, as shown below:

[0074] ,

[0075] in, For the first Input feature maps at various scales, and These are respectively represented as convolutional mapping operations used for classification prediction and bounding box regression. For category prediction results, For activation function, The bounding box regression results are presented. By stitching together the multi-scale prediction results, target classification prediction and bounding box regression are performed, outputting the target's category information and bounding box location. A loss function is constructed based on the difference between the prediction results and the ground truth annotations, and the model parameters are iteratively updated using the backpropagation algorithm until the model converges, thus obtaining the trained MHSW-YOLO model. Finally, the ultrasound image of the tibialis anterior muscle to be detected is input into the trained model to achieve target detection.

[0076] In this step, the XSSBlock module is an extended structure based on VSSBlock. By continuously stacking multiple state space modeling units, it further enhances the global context modeling capability and cross-layer semantic interaction capability of fused features, thereby improving the feature representation capability of complex ultrasonic texture structures and target regions.

[0077] Example

[0078] Ultrasound images of the tibialis anterior muscle from multiple subjects under different scanning angles were collected to construct an original dataset. This dataset was then divided into training, validation, and test sets in a 7:2:1 ratio. Figure 2 As shown, the tibialis anterior muscle region in the ultrasound image was labeled to obtain labeled data for model training. The tibialis anterior muscle ultrasound image data was then input into the MHSW-YOLO model for training.

[0079] Step 1: Train the MHSW-YOLO model using ultrasound image data of the tibialis anterior muscle. The specific network structure is as follows: Figure 3As shown, the input image undergoes convolutional downsampling and preliminary feature extraction through a backbone feature extraction network to obtain shallow feature representations containing rich texture information. Furthermore, a global dependency model is introduced using a spatial pyramid pooling layer to model the features, enhancing the model's ability to express long-range structural information in ultrasound images. The output deep features are processed through a spatial pyramid pooling layer to expand the receptive field and fuse multi-scale contextual information.

[0080] Step 2: Input the extracted features into the HSW module, perform self-attention modeling of local region features through window partitioning and shifting mechanisms, and realize information interaction between different regions, thereby improving the model's ability to represent the boundary and detailed texture of the tibialis anterior muscle.

[0081] Step 3: The enhanced features are input into a multi-scale feature fusion network to fuse features at different levels, allowing high-level semantic information and low-level spatial detail information to complement each other, thus obtaining a feature representation with multi-scale expressive capabilities. The fused features are then input into the detection head, which outputs bounding boxes containing confidence, category, and location information. Based on the loss function, a backpropagation algorithm is used to iteratively optimize the model parameters, resulting in the trained MHSW-YOLO tibialis anterior muscle ultrasound image detection model. The ultrasound image of the tibialis anterior muscle to be detected is then input into the model to achieve ultrasound image detection of the tibialis anterior muscle. The comparison results of MHSW-YOLO with other models for tibialis anterior muscle ultrasound image detection are shown below. Figure 4 As shown, the MHSW-YOLO model proposed in this invention exhibits high consistency with the actual annotations, high detection confidence, and the ability to more accurately locate the tibialis anterior muscle region. Through... Figure 5 The visualization results of the Grad-CAM method model shown indicate that the MHSW-YOLO model enhances the response to the target region and weakens the response to the background region, demonstrating the model's effective representation ability of target features. The model's performance in detection was quantified using mean accuracy (mAP@[0.5:0.95]), F1 coefficient, Matthews correlation coefficient (MCC), and accuracy metrics, and compared with YOLOv8, YOLOv8 with the HSW module, and Mamba-YOLO. As shown in Table 1, the MHSW-YOLO model proposed in this invention achieves the best results in all quantitative metrics for the tibialis anterior muscle ultrasound image detection task, indicating that this method can effectively improve the detection accuracy and feature representation ability of tibialis anterior muscle ultrasound images.

[0082] Table 1 Comparison of evaluation indicators for ablation experiments using ultrasound images of the tibialis anterior muscle.

[0083]

Claims

1. A method for intelligent detection of tibialis anterior muscle ultrasound images based on sequence modeling and feature enhancement, characterized in that... The method includes the following steps: Step 1: Input the ultrasound image into the backbone feature extraction network, perform convolutional downsampling and serialization processing on the image, and use the state space model to perform global feature modeling and recursive update. The visual cue fusion module performs spatial downsampling and channel recombination to extract multi-scale features. The deep feature representation is output through the spatial pyramid pooling module. Step 2: Design and introduce a hierarchical shifted window attention module (HSW) into the MHSW-YOLO backbone network. Input the enhanced feature map into the HSW module, use window self-attention and shifted window mechanism to model and enhance local texture features, and combine the global modeling capability of the state space model to fuse local features and global features, and output an enhanced feature map that fuses local texture information and cross-window context information. Step 3: Input the multi-scale features enhanced by the HSW module into the feature pyramid network for feature fusion. Introduce the extended state space module XSSBlock to further enhance the fused features. Input the enhanced multi-scale features into the detection head for classification and localization. Output the target category information and bounding box to obtain the trained MHSW-YOLO model. Finally, use the MHSW-YOLO model to detect the tibialis anterior muscle ultrasound image to be detected.

2. The intelligent detection method for tibialis anterior muscle ultrasound images based on sequence modeling and feature enhancement according to claim 1, characterized in that... The specific steps of step 1 are as follows: Step 101: Input the ultrasound image of the tibialis anterior muscle into the Simple Stem initial feature extraction module of the Backbone layer of the MHSW-YOLO model to perform feature extraction and downsampling on the input image. The process is as follows: in, For the input image, , For convolutional blocks, It is a non-linear activation function. For batch normalization operations, This is the output shallow feature map; Step 102: After preliminary feature extraction, the shallow feature map is... The input is fed into multiple consecutively distributed visual state space modules, VSSBlock, which then convert the two-dimensional feature maps... Expand into a one-dimensional sequence Furthermore, a state-space model is used to perform global context modeling of sequence features, outputting feature maps. ; Step 103: Obtain the output feature map after state-space modeling and gated residual structure enhancement. Subsequently, the visual cue fusion module VisionClueMerge is introduced to further reduce the feature representation space resolution and expand the receptive field to enhance semantic expressiveness.

3. The intelligent detection method for tibialis anterior muscle ultrasound images based on sequence modeling and feature enhancement according to claim 2, characterized in that... The specific modeling steps in step 102 are as follows: First, the continuous state-space model is discretized, transforming the continuous parameters into discrete parameters, which are expressed in discrete form as follows: in, Let k be the discrete time step at position k. Let k be the feature vector at the k-th position in the sequence. The weight matrix is ​​a learnable matrix. Here is the state transition matrix. For the input control matrix, It is the identity matrix. This is the discretized state transition matrix. The input mapping matrix is ​​the discretized form. Subsequently, the sequence features are updated linearly and recursively, and the update process is expressed as follows: in, For the input sequence features, It is in a hidden state. Here is the state transition matrix. For the input mapping matrix, For the output matrix, For output features; The above recursive state update process based on discretized parameters can be equivalently converted into a convolutional form, specifically expressed as: in, The convolution kernel generated for the state space parameters, Given the input sequence, The output sequence; Finally, the VSSBlock module introduces a gated residual structure, the calculation of which is expressed as follows: in, Given the input feature sequence, For layer normalization, For the state space modeling module, , For learnable gating parameters, For activation function, This is for outputting feature maps.

4. The intelligent detection method for tibialis anterior muscle ultrasound images based on sequence modeling and feature enhancement according to claim 1, characterized in that... The specific steps of step 103 are as follows: First, input features The space is divided into 2×2 regions, and features of each sub-region are extracted. , , , This reduces the spatial resolution to the original level. ; Subsequently, the features of each sub-region are spliced ​​together along the channel dimension to rearrange spatial information along the channel dimension. Finally, a 1×1 convolution is introduced to linearly map the concatenated features, achieving channel information fusion and dimensionality compression to obtain the output feature map. The specific process is as follows: in, Features of sub-regions This is for channel splicing operations. This is the output feature map after spatial downsampling and channel reconstruction.

5. The intelligent detection method for tibialis anterior muscle ultrasound images based on sequence modeling and feature enhancement according to claim 1, characterized in that... The specific steps of step 2 are as follows: Step 201: In the deep feature extraction stage of the backbone network, the enhanced feature map is input into the HSW module to perform self-attention calculation within the local window; Step 202: After completing the attention calculation, the attention output features of each window are obtained. The window restoration unit rearranges and stitches the local window features according to their corresponding spatial positions and removes the padding areas. The window restoration operation restores the complete feature map. On this basis, the attention output feature map is connected to the residual branch, and channel feature transformation and nonlinear enhancement are performed through the normalization layer and the feedforward feature enhancement unit RGBlock. The output features of the HSW module are then obtained through residual connection, thereby realizing the fusion of local texture features and cross-window context information.

6. The intelligent detection method for tibialis anterior muscle ultrasound images based on sequence modeling and feature enhancement according to claim 5, characterized in that... The specific steps of step 201 are as follows: First, the HSW module will input the feature map. The feature map is saved as a residual branch and normalized. The size of the feature map is adjusted by the window completion unit. A shift window mechanism is introduced in the HSW module, and the feature map is shifted in the cyclic translation unit to output an enhanced feature map that integrates local texture information and cross-window context information. At the same time, an attention mask is introduced to constrain the information propagation between different regions. Subsequently, the window partitioning unit divides the feature map according to the preset window size. The system is divided into multiple non-overlapping local window regions. Within each local window, multi-head self-attention computation is performed, generating queries, keys, and values ​​through linear mapping for subsequent calculation of self-attention weights, as shown below: ， ， in, , , For learnable weight matrix, This is the set of all local windows obtained by partitioning the input feature map. , , To represent the query vector, key vector, and value vector; Finally, attention weights are obtained by calculating the similarity between the query and the key, and the value vectors are weighted and summed to achieve self-attention modeling of features within the window. The specific calculation process is as follows: in, Scaling factor , , To represent query vectors, key vectors, and value vectors, This indicates a normalization operation.

7. The intelligent detection method for tibialis anterior muscle ultrasound images based on sequence modeling and feature enhancement according to claim 1, characterized in that... The specific steps of step 3 are as follows: Step 301: Input the features of different scales output by the backbone network into the feature pyramid network, and fuse the multi-scale features through a dual path of top-down and bottom-up. Step 302: After completing the multi-scale feature fusion, the extended state space module XSSBlock is introduced to further enhance the fused features. This module normalizes the input features and performs long-distance dependency modeling on the features through state space branches. The modeling results are then fused with the input features through residual connections. A local space modeling module is introduced to perform spatial dimension information interaction on the features. A multilayer perceptron branch is introduced to perform nonlinear mapping on the features and fuse them with the backbone features through residual connections, as detailed below: in, As input features, Modeling the state space, Modeling local space, For multilayer perceptron mapping, For the enhanced output features; Step 303: After completing the multi-scale feature enhancement, input the feature maps of different scales into the detection head for target prediction.

8. The intelligent detection method for tibialis anterior muscle ultrasound images based on sequence modeling and feature enhancement according to claim 7, characterized in that... The specific steps of step 301 are as follows: In the top-down path, high-level semantic features are upsampled level by level and concatenated with low-level features of the corresponding scale to obtain feature representations that combine high resolution and rich semantic information. The process is as follows: in, For upsampling operation, This is for channel splicing operations. Characterized by high-level features This corresponds to low-level features at the corresponding scale. Features obtained after top-down fusion; In the bottom-up path, low-level features are downsampled using convolutions with a stride of 2, and then fused with high-level features at the corresponding scale. This integrates detailed information from high-resolution features into high-level features, as shown below: in, For downsampling operation, This is for channel splicing operations. This is a feature of the previous layer of fusion. For the current layer features, These are the features resulting from bottom-up fusion.

9. The intelligent detection method for tibialis anterior muscle ultrasound images based on sequence modeling and feature enhancement according to claim 7, characterized in that... In step 303, the detection head inputs features at different scales. Perform convolution mappings separately, and output the class probability and bounding box regression parameters at the corresponding positions, as shown below: ， in, For the first Input feature maps at various scales, and These are respectively represented as convolutional mapping operations used for classification prediction and bounding box regression. For category prediction results, For activation function, The bounding box regression results are used to perform target classification prediction and bounding box regression by stitching together the multi-scale prediction results, outputting the target's category information and bounding box position; a loss function is constructed based on the difference between the prediction results and the ground truth annotations, and the model parameters are iteratively updated through the backpropagation algorithm until the model converges, thus obtaining the trained MHSW-YOLO model; finally, the ultrasound image of the tibialis anterior muscle to be detected is input into the trained model to achieve target detection.

Images