Ankle joint ultrasound image enhancement method

Abstract

The application provides an ankle joint ultrasound image enhancement method, and belongs to the technical field of medical image analysis, and comprises the following steps: inputting an ankle joint ultrasound image into a feature extraction module to obtain a structured feature and a potential field feature; inputting the structured feature into a position difference feature generation module to obtain a position difference feature; inputting the potential field feature into a layer difference feature generation module to obtain a layer difference feature; and inputting the structured feature, the potential field feature, the position difference feature and the layer difference feature into an image construction module to obtain an ankle joint ultrasound image after image enhancement. The application solves the problems of dislocation of anatomical structures and distortion of tissue details in the enhanced image caused by the lack of spatial structure constraints and hierarchical division in the feature extraction process of the existing ultrasound image enhancement technology.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image analysis, and particularly relates to an ankle joint ultrasound image enhancement method. BACKGROUND

[0002] Medical image analysis is a key technology for disease diagnosis and condition monitoring at present, and is widely applied due to advantages such as non-invasiveness, real-time, convenience and low cost, and has important value for diagnosing diseases such as tendonitis, ligament tear and synovitis.

[0003] When an ultrasound wave is used for ankle joint exploration, due to reflection and attenuation of the ultrasound wave, the contrast, resolution and signal-to-noise ratio of the obtained ultrasound image are low, the texture and edge are blurred, the edge of the ankle joint tissue is blurred, the fine lesion texture features are missing, and the level between different soft tissue structures is poor, which cannot meet the needs of condition monitoring; the existing ultrasound image enhancement method adopts simple splicing through a full connection layer, cannot effectively distinguish the features of different depths and different acoustic characteristics of the tissue, causes the boundary of the ankle joint image to be blurred and the details to be distorted; the feature fusion method of the prior art lacks mapping space information cognitive ability, resulting in problems of anatomical structure position misplacement and different tissue detail confusion distortion in the enhanced ankle joint ultrasound image. SUMMARY

[0004] In view of the above problems in the prior art, the ankle joint ultrasound image enhancement method provided by the application solves the problems of anatomical structure misplacement and tissue detail distortion in the enhanced image caused by the lack of spatial structure constraint and hierarchical distinction in the feature extraction process of the existing ultrasound image enhancement technology.

[0005] In order to achieve the above purposes, the technical scheme adopted by the application is as follows: an ankle joint ultrasound image enhancement method, comprising the following steps:

[0006] inputting an ankle joint ultrasound image into a feature extraction module to obtain structured features and potential field features;

[0007] inputting the structured features into a position difference feature generation module to obtain position difference features;

[0008] inputting the potential field features into a layer difference feature generation module to obtain layer difference features;

[0009] inputting the structured features, the potential field features, the position difference features and the layer difference features into an image construction module to obtain an ankle joint ultrasound image after image enhancement.

[0010] Further, the feature extraction module comprises a plurality of feature extraction units connected in sequence, and the outputs of the plurality of feature extraction units are both structured features and potential field features.

[0011] The input of a first feature extraction unit of the plurality of feature extraction units is the ankle joint ultrasound image, and the input of a feature extraction unit other than the first feature extraction unit of the plurality of feature extraction units is the structured feature output by the previous feature extraction unit.

[0012] The further beneficial effect is that, through the connection of the plurality of feature extraction units, the ankle joint ultrasound image can be progressively and multiple times deepened for feature extraction, the feature expression is gradually enriched and refined, the information correlation of each region in the ankle joint ultrasound image is enhanced, the image information fault problem is alleviated, and the progressive input data is provided for the generation of the position difference feature and the layer difference feature, so that the position difference feature and the layer difference feature are more accurate and reliable.

[0013] Further, the structured feature and the potential field feature output by the last feature extraction unit of the plurality of feature extraction units are denoted as the last structured feature and the last potential field feature, and are input to the image construction module.

[0014] The further beneficial effect is that, after multiple deepening processing of the plurality of feature extraction units, the structured feature and the potential field feature output by the last feature extraction unit are the most representative features among all the structured features and potential field features output by the feature extraction module, which contain the most abundant semantic information and spatial information. Taking the structured feature and the potential field feature as the input of the image construction module can improve the accuracy of the structure and tissue details in the ankle joint ultrasound image after image enhancement, and alleviate the image reconstruction distortion problem caused by insufficient feature levels.

[0015] Further, the workflow of the first feature extraction unit of the plurality of feature extraction units includes:

[0016] According to the ankle joint ultrasound image, a first basic feature map is obtained by using a 3*3 convolution layer and a RELU function for processing;

[0017] According to the ankle joint ultrasound image, a second basic feature map is obtained by using a 1*1 convolution layer and a RELU function for processing;

[0018] The first basic feature map and the second basic feature map are subtracted in corresponding elements to obtain a subtraction feature map, and the subtraction feature map is processed by a 3*3 convolution layer and a RELU function to obtain a processed subtraction feature map;

[0019] The first basic feature map is sequentially subjected to a pooling operation for down-sampling, a RELU function processing and up-sampling to obtain an up-sampling feature map, and the up-sampling feature map and the second basic feature map are multiplied in corresponding elements to obtain a structured feature;

[0020] The structured features are processed using the sigmoid function to obtain the processed structured features;

[0021] The processed structured features, the processed subtracted feature map, and the first basic feature map are added element by element to obtain the fused feature map. The fused feature map is then processed by the sigmoid function to obtain the potential field features.

[0022] The workflow of the feature extraction units other than the first feature extraction unit is the same as that of the first feature extraction unit, but the inputs are all structured features output from the previous feature extraction unit that are sequentially connected.

[0023] The further beneficial effects mentioned above are as follows: In the feature extraction unit, the subtracted feature map is obtained by subtracting the feature maps, which can enhance the contrast of details. The downsampling and upsampling can capture multi-scale contextual information, making the structured features more spatially correlated and enabling the potential field features to better represent the diffusion trend of the ankle joint tissue boundary, thereby improving the ability of the potential field features to represent complex ankle joint tissue structures.

[0024] Furthermore: the position difference feature generation module includes several parallel first global max pooling layers, and the several first global max pooling layers are configured one-to-one with several feature extraction units in the feature extraction module;

[0025] The input to the position difference feature generation module is several structured features output by the feature extraction module, and the output is a position difference feature.

[0026] The inputs to the plurality of first global max pooling layers are the structured features output by the plurality of feature extraction units in the feature extraction module;

[0027] The first global max pooling layers of the position difference feature generation module are all used to compress the channel size. That is, the first global max pooling layers are used to perform global max pooling processing along the channel direction of the feature map.

[0028] The further beneficial effects mentioned above are as follows: by processing structured features at different levels through multiple first global max pooling layers corresponding one-to-one with feature extraction units, it is possible to extract and integrate multi-scale spatial information of ankle ultrasound images. By compressing channel size, representative features can be retained, so that the final fused positional difference features have the ability to spatially locate the anatomical structure of the ankle joint, avoiding enhancement errors and deformation of tissue structures.

[0029] Further: The workflow of the position difference feature generation module includes:

[0030] Several structured features are obtained, and these structured features are concatenated to obtain integrated structured features.

[0031] Based on the integrated structured features, the first transformation features are obtained by processing them using a 3×3 convolutional layer and the ReLU function;

[0032] Several structured features are input into several first global max pooling layers, and several two-dimensional matrices are output. Several two-dimensional matrices are concatenated to obtain a set of two-dimensional matrices.

[0033] Based on the set of two-dimensional matrices, the second transformation feature is obtained by processing with a 3×3 convolutional layer and the ReLU function;

[0034] Subtracting the corresponding elements of the first transformation feature from the first transformation feature yields the position difference feature.

[0035] The further beneficial effects mentioned above are: through dual-path parallel processing and feature difference extraction, it is possible to extract the difference information of multi-scale structured features, so that the generated positional difference features can accurately characterize the relative positional relationship of anatomical structures, provide reliable spatial constraints for ankle joint image reconstruction, and improve the structural accuracy of enhanced ankle joint images.

[0036] Furthermore: the layer difference feature generation module includes several parallel second global max pooling layers, and the number of the several second global max pooling layers is set to correspond one-to-one with the number of feature extraction units in the feature extraction module;

[0037] The input to the layer difference feature generation module is several potential field features output by the feature extraction module, and the output is a layer difference feature.

[0038] The inputs to the plurality of second global max pooling layers are the potential field features output by the plurality of feature extraction units in the feature extraction module;

[0039] The second global max pooling layers of the layer difference feature generation module are all used to compress the spatial size. That is, the second global max pooling layers are used to perform global max pooling processing along the spatial direction of the feature map.

[0040] The further beneficial effects mentioned above are: by processing potential field features at different levels through multiple global max pooling layers, it is possible to retain complete channel information while compressing spatial size, so that the generated layer difference features can characterize the acoustic characteristics of ankle joint tissues at different depths, achieve accurate tissue layer differentiation in the image reconstruction process, and avoid feature confusion between different structures of the ankle joint.

[0041] Furthermore, the workflow of the layer difference feature generation module includes:

[0042] Several potential field features are obtained, and these features are spliced ​​together to obtain integrated potential field features.

[0043] Several potential field features are input into several second global max pooling layers, and several one-dimensional vectors are output. Several one-dimensional vectors are concatenated to obtain a set of one-dimensional vectors.

[0044] Based on a set of one-dimensional vectors, the first deep feature is obtained by processing it using a 1×1 convolutional layer and the ReLU function;

[0045] Based on the integrated potential field features, the data are processed sequentially through a 3×3 convolutional layer, a second global max pooling layer, and a sigmoid function to obtain the second depth features.

[0046] Subtracting the corresponding elements of the first depth feature and the second depth feature yields the layer difference feature.

[0047] The further beneficial effects mentioned above are as follows: the layer difference feature generation module splices several potential field features through a dual-path processing structure and processes them separately to obtain several one-dimensional vectors, which can complete the extraction of global channel features and vertical spatial features, so that the obtained layer difference features can characterize the acoustic characteristics of ankle joint tissues at different depths, thereby improving the tissue discrimination of ankle joint image reconstruction.

[0048] Furthermore: the image construction module includes a 3×3 convolutional layer and a ReLU function connected in sequence;

[0049] The image construction module takes structured features, potential field features, positional difference features, and layer difference features as inputs and outputs an enhanced ultrasound image of the ankle joint.

[0050] The further beneficial effects mentioned above are: by using the features of the ankle joint image that can be represented by structured features, potential field features, positional difference features and layer difference features, the image construction module can reconstruct the ankle joint image while keeping the model simple, and the enhanced ankle joint ultrasound image has clear ankle joint tissue edges and texture details.

[0051] Furthermore, the workflow of the image construction module includes:

[0052] Multiply the last structured feature and the positional difference feature element by element to obtain the spatial constraint feature;

[0053] Multiply the last potential field feature and the layer difference feature element by element to obtain the layer distinction feature;

[0054] The spatial constraint features and hierarchical differentiation features are added element by element to obtain the fused reconstruction features;

[0055] Based on the fusion reconstruction features, a 3×3 convolutional layer and the ReLU function were used for processing to obtain an enhanced ultrasound image of the ankle joint.

[0056] The further beneficial effects mentioned above are as follows: the features obtained by multiplying the last structured feature and the positional difference feature by corresponding elements can better constrain the spatial structure of the ankle ultrasound image; the features obtained by multiplying the last potential field feature and the layer difference feature by corresponding elements can better distinguish the layers of the ankle ultrasound image; and the addition of corresponding elements integrates the accuracy of spatial position and the differences in layered tissues, so that the final enhanced ankle ultrasound image can retain the anatomical structure of the ankle joint while having better detail clarity and avoiding image distortion.

[0057] The beneficial effects of this invention are:

[0058] Through the cascaded structure and multi-path processing mechanism of the feature extraction module, multi-scale feature extraction of ankle ultrasound images is realized; it can combine local details and global contextual information of ankle ultrasound images to solve the problems of noise interference and unclear images in ultrasound images.

[0059] The positional difference feature generation module establishes a precise spatial constraint mechanism through global max pooling and feature difference calculation. This mechanism can accurately determine the relative positions of various anatomical structures in the anchored ankle joint, prevent tissue misalignment and deformation during image reconstruction, and ensure that the enhanced image conforms to the actual anatomical structure.

[0060] The layer difference feature generation module extracts depth difference information through dual-path processing, enabling the differentiation of acoustic properties of different tissues. It is particularly suitable for distinguishing adjacent soft tissues such as tendons and synovium, alleviating the boundary ambiguity problem common in traditional methods, making tissue layers more distinct and lesions easier to identify.

[0061] Each module of this invention has no complex neural network structure and is a lightweight network structure, which significantly reduces computational complexity while ensuring the enhancement effect and improves the efficiency and accuracy of ultrasound image enhancement. Attached Figure Description

[0062] Figure 1 This is a diagram of the neural network structure used for ankle ultrasound image enhancement. Detailed Implementation

[0063] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0064] Example 1

[0065] like Figure 1 The diagram shows a neural network structure for ankle joint ultrasound image enhancement, which enables the implementation of an ankle joint ultrasound image enhancement method, including the following steps:

[0066] Ankle ultrasound images are input into the feature extraction module to obtain structured features and potential field features;

[0067] The structured features are input into the position difference feature generation module to obtain the position difference features;

[0068] Input the potential field features into the layer difference feature generation module to obtain the layer difference features;

[0069] By inputting structured features, potential field features, positional difference features, and layer difference features into the image construction module, an enhanced ultrasound image of the ankle joint is obtained.

[0070] To address the problems of anatomical misalignment and tissue detail distortion in enhanced images caused by the lack of spatial structural constraints and hierarchical differentiation in the feature extraction process of existing ultrasound image enhancement technologies, this invention employs a feature extraction module to extract structured features and potential field features in parallel, capturing the spatial correlation of tissues in ankle ultrasound images and improving the accuracy of image enhancement. A positional difference feature generation module establishes spatial constraints to determine the spatial structure of tissues in ankle ultrasound images, while a layered difference feature generation module achieves hierarchical differentiation of ankle joint tissues. Finally, the image construction module integrates structured features, potential field features, positional difference features, and layered difference features for reconstruction, significantly improving tissue boundary clarity while maintaining anatomical accuracy, resulting in richer and more accurate image details.

[0071] In one embodiment of the present invention, to address the problems of complex tissue structures, blurred edges, and easy loss of spatial information in ankle ultrasound images, the feature extraction module of the present invention adopts a multi-level cascaded architecture design. This module includes several sequentially connected feature extraction units. The first feature extraction unit directly processes the input ankle ultrasound image, and subsequent feature extraction units perform progressive feature extraction on the structured features output by the previous feature extraction unit. This achieves in-depth mining and multi-scale feature extraction of ankle ultrasound images. The extracted potential field features characterize the diffusion trend of information in various regions of the ankle ultrasound image, forming a feature map that reflects the direction and rate of information diffusion. The extracted structured features focus on the correlation of information in various regions of the ankle ultrasound image, which can alleviate the problem of information discontinuity in ultrasound images.

[0072] The feature extraction module of the present invention includes a plurality of feature extraction units connected in sequence, and the outputs of the plurality of feature extraction units are structured features and potential field features;

[0073] The input of the first feature extraction unit in a plurality of feature extraction units is an ankle ultrasound image. The inputs of the feature extraction units other than the first feature extraction unit are all structured features output from the previous feature extraction unit, which are sequentially connected.

[0074] The structured features and potential field features output by the last feature extraction unit among several feature extraction units are denoted as the last structured feature and the last potential field feature, and serve as the input to the image construction module.

[0075] In a specific embodiment of the present invention, the workflow of the first feature extraction unit among several feature extraction units includes:

[0076] Based on the ankle ultrasound image, the first basic feature map was obtained by processing it with a 3×3 convolutional layer and the ReLU function;

[0077] Based on the ankle ultrasound image, a second basic feature map was obtained by processing it with a 1×1 convolutional layer and the ReLU function;

[0078] The first basic feature map is subtracted from the second basic feature map element by element to obtain the subtracted feature map. The subtracted feature map is then processed by a 3×3 convolutional layer and the ReLU function to obtain the processed subtracted feature map.

[0079] The first basic feature map is sequentially processed by pooling downsampling, ReLU function processing and upsampling to obtain an upsampled feature map. The upsampled feature map and the second basic feature map are multiplied element-wise to obtain the structured features.

[0080] The structured features are processed using the sigmoid function to obtain the processed structured features;

[0081] The processed structured features, the processed subtracted feature map, and the first basic feature map are added element by element to obtain the fused feature map. The fused feature map is then processed by the sigmoid function to obtain the potential field features.

[0082] Meanwhile, the workflow of the feature extraction units other than the first feature extraction unit is the same as that of the first feature extraction unit, but the inputs are all structured features output from the previous feature extraction unit, which are connected sequentially, so as to realize progressive feature extraction of ankle ultrasound images and improve the accuracy of image enhancement.

[0083] In one embodiment of the present invention, a number of structured features and a number of potential field features can be obtained through the feature extraction unit. The structured features are input to the position difference feature generation module for processing to obtain position difference features. The region constraint information in the position difference features can anchor the relative positions of key structures in the ankle ultrasound image during the reconstruction process, avoiding structural misalignment during image construction and avoiding enhancement errors that do not conform to the anatomical structure. The potential field features are input to the layer difference feature generation module for processing to obtain layer difference features. The depth difference information in the layer difference features is helpful to more accurately distinguish the performance differences of different types of tissues in the ankle ultrasound image and avoid distortion of the enhanced ultrasound image due to feature confusion.

[0084] In a specific embodiment of the present invention, the position difference feature generation module includes a plurality of first global max pooling layers arranged in parallel, and the plurality of first global max pooling layers are configured in a one-to-one correspondence with a plurality of feature extraction units in the feature extraction module;

[0085] The input to the position difference feature generation module is several structured features output by the feature extraction module, and the output is a position difference feature.

[0086] The inputs to several first global max pooling layers are the structured features output by several feature extraction units in the feature extraction module;

[0087] Several first global max pooling layers in the position difference feature generation module are used to compress channel size.

[0088] In a specific embodiment of the present invention, the workflow of the position difference feature generation module includes:

[0089] Several structured features are obtained, and these structured features are concatenated to obtain integrated structured features.

[0090] Based on the integrated structured features, the first transformation features are obtained by processing them using a 3×3 convolutional layer and the ReLU function;

[0091] Several structured features are input into several first global max pooling layers, and several two-dimensional matrices are output. Several two-dimensional matrices are concatenated to obtain a set of two-dimensional matrices.

[0092] Based on the set of two-dimensional matrices, the second transformation feature is obtained by processing with a 3×3 convolutional layer and the ReLU function;

[0093] Subtracting the corresponding elements of the first transformation feature from the first transformation feature yields the position difference feature.

[0094] In a specific embodiment of the present invention, the layer difference feature generation module includes a plurality of parallel second global max pooling layers, and the number of the plurality of second global max pooling layers is set to correspond one-to-one with the plurality of feature extraction units in the feature extraction module;

[0095] The input to the layer difference feature generation module is several potential field features output by the feature extraction module, and the output is a layer difference feature.

[0096] The inputs to several second global max pooling layers are the potential field features output by several feature extraction units in the feature extraction module;

[0097] Several second global max pooling layers in the layer difference feature generation module are used to compress spatial dimensions.

[0098] In a specific embodiment of the present invention, the workflow of the layer difference feature generation module includes:

[0099] Several potential field features are obtained, and these features are spliced ​​together to obtain integrated potential field features.

[0100] Several potential field features are input into several second global max pooling layers, and several one-dimensional vectors are output. These one-dimensional vectors are then concatenated, that is, the first and last ends of several one-dimensional vectors are concatenated sequentially to obtain a set of one-dimensional vectors.

[0101] Based on a set of one-dimensional vectors, the first deep feature is obtained by processing it using a 1×1 convolutional layer and the ReLU function;

[0102] Based on the integrated potential field features, the data are processed sequentially through a 3×3 convolutional layer, a second global max pooling layer, and a sigmoid function to obtain the second depth features.

[0103] Subtracting the corresponding elements of the first depth feature and the second depth feature yields the layer difference feature.

[0104] In one embodiment of the present invention, positional difference features are obtained through a positional difference feature generation module, layer difference features are obtained through a layer difference feature generation module, and the last structured feature and the last potential field feature are obtained through the last feature extraction unit in the feature extraction module. Based on the last structured feature, the last potential field feature, the positional difference feature, and the layer difference feature, the present invention uses an image reconstruction module to reconstruct the ankle joint ultrasound image. Through multi-feature collaborative reconstruction, the last structured feature provides deep semantic information, the last potential field feature preserves tissue diffusion characteristics, the positional difference feature ensures spatial accuracy, and the layer difference feature ensures the differentiation of different tissues in the ankle joint. This results in an enhanced ankle joint image that maintains anatomical accuracy while possessing higher detail reproduction, enabling better differentiation of the continuity of tendon edges, the boundaries between ligaments and surrounding tissues, and fine structures such as the synovium in the ankle joint image.

[0105] In a specific embodiment of the present invention, the image construction module includes a 3×3 convolutional layer and a ReLU function connected in sequence; the input of the image construction module is structured features, potential field features, positional difference features and layer difference features, and the output is an enhanced ankle joint ultrasound image.

[0106] In a specific embodiment of the present invention, the workflow of the image construction module includes:

[0107] Multiply the last structured feature and the positional difference feature element by element to obtain the spatial constraint feature;

[0108] Multiply the last potential field feature and the layer difference feature element by element to obtain the layer distinction feature;

[0109] The spatial constraint features and hierarchical differentiation features are added element by element to obtain the fused reconstruction features;

[0110] Based on the fusion reconstruction features, a 3×3 convolutional layer and the ReLU function were used for processing to obtain an enhanced ultrasound image of the ankle joint.

[0111] The beneficial effects of this invention are as follows: This invention solves the problems of structural distortion and detail blurring in ankle ultrasound image enhancement through a multi-module collaborative deep learning architecture. The feature extraction module achieves multi-scale feature deepening through cascaded units, the positional difference feature module establishes spatial constraints to ensure the accuracy of anatomical structures, the layer difference feature module achieves precise differentiation of tissue layers, and finally, the multi-feature fusion of the image reconstruction module improves image quality while maintaining the correct anatomical structure, enhances the clarity of tissue boundaries and the richness of details, reduces noise interference, and improves the overall visual effect of ankle ultrasound images, providing more reliable imaging evidence for the clinical diagnosis of ankle joints.

Claims

1. A method for enhancing ultrasound images of the ankle joint, characterized in that, Includes the following steps: Ankle ultrasound images are input into the feature extraction module to obtain structured features and potential field features; The structured features are input into the position difference feature generation module to obtain the position difference features; Input the potential field features into the layer difference feature generation module to obtain the layer difference features; The structured features, potential field features, positional difference features, and layer difference features are input into the image construction module to obtain an enhanced ultrasound image of the ankle joint. The feature extraction module includes several feature extraction units connected in sequence, and the outputs of the several feature extraction units are structured features and potential field features; The input of the first feature extraction unit among the plurality of feature extraction units is an ankle ultrasound image, and the inputs of the feature extraction units other than the first feature extraction unit are all structured features output by the previous feature extraction unit in sequence. The workflow of the first feature extraction unit among the plurality of feature extraction units includes: Based on the ankle ultrasound image, the first basic feature map was obtained by processing it with a 3×3 convolutional layer and the ReLU function; Based on the ankle ultrasound image, a second basic feature map was obtained by processing it with a 1×1 convolutional layer and the ReLU function; The first basic feature map is subtracted from the second basic feature map element by element to obtain the subtracted feature map. The subtracted feature map is then processed by a 3×3 convolutional layer and the ReLU function to obtain the processed subtracted feature map. The first basic feature map is sequentially processed by pooling downsampling, ReLU function processing and upsampling to obtain an upsampled feature map. The upsampled feature map and the second basic feature map are multiplied element-wise to obtain the structured features. The structured features are processed using the sigmoid function to obtain the processed structured features; The processed structured features, the processed subtracted feature map, and the first basic feature map are added element by element to obtain the fused feature map. The fused feature map is then processed by the sigmoid function to obtain the potential field features. The workflow of the feature extraction units other than the first feature extraction unit is the same as that of the first feature extraction unit, but the inputs are all structured features output from the previous feature extraction unit that are sequentially connected. The workflow of the position difference feature generation module includes: Several structured features are obtained, and these structured features are concatenated to obtain integrated structured features. Based on the integrated structured features, the first transformation features are obtained by processing them using a 3×3 convolutional layer and the ReLU function; Several structured features are input into several first global max pooling layers, and several two-dimensional matrices are output. Several two-dimensional matrices are concatenated to obtain a set of two-dimensional matrices. Based on the set of two-dimensional matrices, the second transformation feature is obtained by processing with a 3×3 convolutional layer and the ReLU function; Subtract the corresponding elements from the first transformation feature to obtain the position difference feature; The workflow of the layer difference feature generation module includes: Several potential field features are obtained, and these features are spliced ​​together to obtain integrated potential field features. Several potential field features are input into several second global max pooling layers, and several one-dimensional vectors are output. Several one-dimensional vectors are concatenated to obtain a set of one-dimensional vectors. Based on a set of one-dimensional vectors, the first deep feature is obtained by processing it using a 1×1 convolutional layer and the ReLU function; Based on the integrated potential field features, the data are processed sequentially through a 3×3 convolutional layer, a second global max pooling layer, and a sigmoid function to obtain the second depth features. Subtracting the corresponding elements of the first depth feature and the second depth feature yields the layer difference feature.

2. The ankle joint ultrasound image enhancement method according to claim 1, characterized in that, The structured features and potential field features output by the last feature extraction unit among the plurality of feature extraction units are denoted as the last structured features and the last potential field features, and are the inputs of the image construction module.

3. The ankle joint ultrasound image enhancement method according to claim 1, characterized in that, The position difference feature generation module includes several parallel first global max pooling layers, and the several first global max pooling layers are configured one-to-one with several feature extraction units in the feature extraction module; The input to the position difference feature generation module is several structured features output by the feature extraction module, and the output is a position difference feature. The inputs to the plurality of first global max pooling layers are the structured features output by the plurality of feature extraction units in the feature extraction module; The first global max pooling layers of the position difference feature generation module are all used to compress the channel size.

4. The ankle joint ultrasound image enhancement method according to claim 1, characterized in that, The layer difference feature generation module includes several parallel second global max pooling layers, and the number of the several second global max pooling layers is set to correspond one-to-one with the number of feature extraction units in the feature extraction module; The input to the layer difference feature generation module is several potential field features output by the feature extraction module, and the output is a layer difference feature. The inputs to the plurality of second global max pooling layers are the potential field features output by the plurality of feature extraction units in the feature extraction module; The second global max pooling layers of the layer difference feature generation module are all used to compress spatial dimensions.

5. The ankle joint ultrasound image enhancement method according to claim 1, characterized in that, The image construction module includes a 3×3 convolutional layer and a ReLU function connected in sequence; The image construction module takes structured features, potential field features, positional difference features, and layer difference features as inputs and outputs an enhanced ultrasound image of the ankle joint.

6. The ankle joint ultrasound image enhancement method according to claim 5, characterized in that, The workflow of the image construction module includes: Multiply the last structured feature and the positional difference feature element by element to obtain the spatial constraint feature; Multiply the last potential field feature and the layer difference feature element by element to obtain the layer distinction feature; The spatial constraint features and hierarchical differentiation features are added element by element to obtain the fused reconstruction features; Based on the fusion reconstruction features, a 3×3 convolutional layer and the ReLU function were used for processing to obtain an enhanced ultrasound image of the ankle joint.

Images