Deep learning-based war wound ultrasonic image diagnosis system and method

Abstract

The invention provides a battle wound ultrasonic image diagnosis system and method based on deep learning, and the method comprises the steps: carrying out the adaptive preprocessing of a to-be-processed battle wound ultrasonic image, and obtaining a preprocessed ultrasonic image; global context features of image textures and local high-frequency features of image edges are extracted from the preprocessed ultrasonic image, and then a semantic fusion feature map containing deep image semantic information is generated; determining feature correlation coefficients between different feature sub-regions in the semantic fusion feature map and a preset feature template based on a preset sliding window, and further determining a segmentation probability matrix when the preprocessed ultrasonic image is segmented; and performing regionalization segmentation on the preprocessed ultrasonic image according to the segmentation probability matrix to obtain a regionalization segmentation image corresponding to the war wound ultrasonic image. According to the technical scheme provided by the invention, precise regionalization segmentation can be performed on the war wound ultrasonic image under the structural heterogeneity and texture complexity of the war wound tissue.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ultrasound image processing, and more particularly, to a war injury ultrasound image diagnosis system and method based on deep learning. BACKGROUND

[0002] As a non-invasive and safe medical imaging examination technology, ultrasound imaging is widely used in clinical diagnosis, treatment monitoring and research field due to its advantages of device portability, simple operation and low cost. It uses high-frequency sound waves to obtain internal structural information through reflection and refraction in human tissues. However, in practical application, the quality of ultrasound images is affected by many factors such as probe selection, operation method and patient constitution, resulting in problems such as image noise and artifacts. In addition, ultrasound image analysis is highly dependent on the professional knowledge and experience of doctors, and is highly subjective and easily affected by individual judgment differences. Therefore, the development of ultrasound image processing is of great significance to improving image quality and diagnostic accuracy.

[0003] In existing ultrasound image processing, ultrasound image processing generally first improves image quality through filtering, contrast enhancement and other preprocessing, then extracts key features through edge detection, texture analysis and other methods, and then optimizes the target region through morphological operation or segmentation algorithm. Some combine machine learning to achieve automatic recognition and analysis. However, in war injury ultrasound image processing, the pathological morphology of war injury tissue (such as laceration, hematoma and internal contusion) shows high structural heterogeneity and texture complexity. On the war injury ultrasound image, this reflects the change of acoustic impedance difference between the injury area and healthy tissue. The edge often shows low contrast, weak boundary and discontinuous characteristics, rather than a uniform, continuous closed region. Therefore, when segmenting the injury area, the boundary segmentation is blurred. Therefore, how to accurately segment the war injury ultrasound image under the structural heterogeneity and texture complexity of war injury tissue has become a difficult problem in the industry. SUMMARY

[0004] The present application provides a war injury ultrasound image diagnosis system and method based on deep learning, which can accurately segment the war injury ultrasound image under the structural heterogeneity and texture complexity of war injury tissue.

[0005] In a first aspect, the present application provides a war injury ultrasound image processing method based on deep learning, comprising the following steps:

[0006] Obtain a war injury ultrasound image to be processed, and perform adaptive preprocessing based on deep learning prior knowledge on the war injury ultrasound image to obtain a preprocessed ultrasound image;

[0007] extract global context features of image texture and local high-frequency features of image edges from the preprocessed ultrasound image, and then generate a semantic fusion feature map containing deep image semantic information according to the global context features and the local high-frequency features;

[0008] determine a feature correlation coefficient between different feature sub-regions in the semantic fusion feature map and a preset feature template based on a preset sliding window, and then determine a segmentation probability matrix when segmenting the preprocessed ultrasound image according to the feature correlation coefficient corresponding to each feature sub-region;

[0009] segment the preprocessed ultrasound image according to the segmentation probability matrix to obtain a regionalized segmentation image corresponding to the war injury ultrasound image.

[0010] In some embodiments, the adaptive preprocessing of the war injury ultrasound image based on deep learning prior knowledge specifically includes:

[0011] construct a deep learning prior knowledge graph containing typical war injury ultrasound image artifact features and normal tissue features according to the imaging characteristics of war injury ultrasound images;

[0012] based on the deep learning prior knowledge graph, identify the artifact noise type of the obtained war injury ultrasound image, and then determine the artifact noise category existing in the war injury ultrasound image;

[0013] According to the identified artifact noise category, call the corresponding adaptive filtering algorithm to suppress noise of the war injury ultrasound image to obtain a preprocessed ultrasound image.

[0014] In some embodiments, the global context features of image texture and the local high-frequency features of image edges extracted from the preprocessed ultrasound image specifically include:

[0015] input the preprocessed ultrasound image into a pre-trained deep convolutional neural network, and extract global context features of image texture through a deep network of the deep convolutional neural network;

[0016] perform multi-scale Gaussian difference filtering on the preprocessed ultrasound image to generate a multi-scale edge response map;

[0017] extract local high-frequency features of image edges based on the multi-scale edge response map.

[0018] In some embodiments, the semantic fusion feature map containing deep image semantic information generated according to the global context features and the local high-frequency features specifically includes:

[0019] dimensionally adapt the global context feature, to generate an adapted global feature matching a dimension of the local high-frequency feature;

[0020] adopt an attention mechanism to calculate a feature correlation weight of the adapted global feature and the local high-frequency feature;

[0021] based on the feature correlation weight, weight and fuse the adapted global feature and the local high-frequency feature, to generate an initial fusion feature map;

[0022] through a residual connection module, perform feature enhancement on the initial fusion feature map, to obtain a semantic fusion feature map containing deep image semantic information.

[0023] In some embodiments, based on a preset sliding window, determine a feature correlation coefficient between different feature sub-regions in the semantic fusion feature map and a preset feature template, specifically including:

[0024] According to the preset sliding window, perform sliding sampling on the semantic fusion feature map, to generate a plurality of feature sub-regions;

[0025] Determine the feature vector corresponding to each feature sub-region;

[0026] Calculate the cosine similarity of the feature vector corresponding to each feature sub-region and the preset feature template;

[0027] Normalize each cosine similarity, to obtain the feature correlation coefficient between each feature sub-region and the preset feature template.

[0028] In some embodiments, according to the feature correlation coefficient corresponding to each feature sub-region, determine a segmentation probability matrix when segmenting the preprocessed ultrasound image, specifically including:

[0029] Map the feature correlation coefficient corresponding to each feature sub-region to the corresponding spatial position of the semantic fusion feature map, to generate an initial probability distribution matrix;

[0030] Perform spatial smoothing processing on the initial probability distribution matrix, to obtain a smoothed probability distribution matrix;

[0031] Using a flexible maximum function, perform normalization operation on the smoothed probability distribution matrix, to obtain a segmentation probability value of each pixel point belonging to a segmentation region;

[0032] According to the segmentation probability value of each pixel point, construct a segmentation probability matrix when segmenting the preprocessed ultrasound image.

[0033] In some embodiments, from a war injury ultrasound image storage file, obtain a war injury ultrasound image to be processed.

[0034] In a second aspect, the application provides a war injury ultrasound image diagnosis system based on deep learning, which is used to execute a war injury ultrasound image processing method based on deep learning, and comprises a war injury ultrasound image processing unit, which comprises:

[0035] An acquisition module is configured to acquire a war injury ultrasound image to be processed, and perform adaptive preprocessing on the war injury ultrasound image based on deep learning prior knowledge to obtain a preprocessed ultrasound image.

[0036] A processing module is configured to extract global context features of image texture and local high-frequency features of image edges from the preprocessed ultrasound image, and then generate a semantic fusion feature map containing deep image semantic information according to the global context features and the local high-frequency features.

[0037] The processing module is further configured to determine feature correlation coefficients between different feature sub-regions in the semantic fusion feature map and a preset feature template based on a preset sliding window, and then determine a segmentation probability matrix when segmenting the preprocessed ultrasound image according to the feature correlation coefficients corresponding to each feature sub-region.

[0038] An execution module is configured to perform regional segmentation on the preprocessed ultrasound image according to the segmentation probability matrix to obtain a regional segmentation image corresponding to the war injury ultrasound image.

[0039] In a third aspect, the application provides a computer device, which comprises a memory and a processor, the memory stores a code, and the processor is configured to acquire the code and execute the war injury ultrasound image processing method based on deep learning.

[0040] In a fourth aspect, the application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the war injury ultrasound image processing method based on deep learning.

[0041] The technical scheme provided by the embodiments of the application has the following beneficial effects:

[0042] The war injury ultrasound image diagnosis system and method based on deep learning provided in the application first acquires a war injury ultrasound image to be processed, and performs adaptive preprocessing based on deep learning prior knowledge on the war injury ultrasound image to obtain a preprocessed ultrasound image; secondly, global context features of image texture and local high-frequency features of image edges are extracted from the preprocessed ultrasound image, and then semantic fusion feature maps containing deep image semantic information are generated according to the global context features and the local high-frequency features; then, feature correlation coefficients between different feature sub-regions in the semantic fusion feature maps and preset feature templates are determined based on a preset sliding window, and then a segmentation probability matrix for segmenting the preprocessed ultrasound image is determined according to the feature correlation coefficients corresponding to each feature sub-region; finally, the preprocessed ultrasound image is regionally segmented according to the segmentation probability matrix to obtain a regional segmentation image corresponding to the war injury ultrasound image.

[0043] It can be seen that the application can accurately regionally segment the war injury ultrasound image under the structural heterogeneity and texture complexity of war injury tissue. First, through adaptive preprocessing based on deep learning prior knowledge, the typical noise distribution of the war injury ultrasound image can be targeted for optimization, avoiding the interference of invalid information on feature analysis. Secondly, by extracting global context features of image texture and local high-frequency features of image edges and fusing to generate semantic fusion feature maps, the image semantic information at the global level is integrated, and the key texture details at the local level are strengthened, realizing the deep coupling of features of different scales and solving the problem that a single feature cannot fully represent the depth semantics of the image. Then, the semantic fusion feature maps are traversed based on the preset sliding window, the feature correlation coefficients of each feature sub-region are calculated by matching with the war injury feature template library, and the segmentation probability matrix is generated through spatial mapping, smoothing processing and normalization, which converts the abstract feature correlation into a pixel-level segmentation possibility distribution, making the segmentation boundary of the war injury ultrasound image more accurate and avoiding the segmentation ambiguity caused by the structural heterogeneity and texture complexity of war injury tissue. Finally, the preprocessed ultrasound image is regionally segmented according to the segmentation probability matrix to obtain a regional segmentation image corresponding to the war injury ultrasound image, effectively determining the semantic attribution of the pixel region, filtering out irrelevant background information and noise interference, highlighting the segmentation features of image segmentation, and solving the ambiguity of ultrasound image semantic segmentation in the war injury scene. In summary, the technical solution provided by the application can accurately regionally segment the war injury ultrasound image under the structural heterogeneity and texture complexity of war injury tissue. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is an exemplary flowchart of a war injury ultrasound image processing method based on deep learning according to some embodiments of the application;

[0045] Figure 2 is an exemplary flowchart of determining a semantic fusion feature map according to some embodiments of the present application;

[0046] Figure 3 is a structural schematic diagram of a war trauma ultrasound image processing unit according to some embodiments of the present application;

[0047] Figure 4 is a structural schematic diagram of a computer device for implementing a deep learning-based war trauma ultrasound image processing method according to some embodiments of the present application. DETAILED DESCRIPTION

[0048] In order to better understand the technical solutions of the present application, the technical solutions of the present application will be described in detail below in combination with the drawings of the specification and specific embodiments.

[0049] Reference Figure 1 The figure is an exemplary flowchart of a deep learning-based war trauma ultrasound image processing method according to some embodiments of the present application, which mainly includes the following steps:

[0050] In step S101, a war trauma ultrasound image to be processed is obtained, and adaptive preprocessing based on deep learning prior knowledge is performed on the war trauma ultrasound image to obtain a preprocessed ultrasound image.

[0051] In specific implementation, the war trauma ultrasound image to be processed is obtained from a war trauma ultrasound image storage file, the war trauma ultrasound image to be processed refers to a war trauma ultrasound image that needs to be processed, and the war trauma ultrasound image storage file refers to a file that stores the war trauma ultrasound image to be processed, which will not be described here. Through the acquisition of the war trauma ultrasound image to be processed, data basis can be provided for the processing of the war trauma ultrasound image.

[0052] In some embodiments, the adaptive preprocessing based on deep learning prior knowledge of the war trauma ultrasound image to obtain a preprocessed ultrasound image is implemented by the following steps:

[0053] A deep learning prior knowledge graph containing typical war trauma ultrasound image artifact features and normal tissue features is constructed according to the imaging characteristics of the war trauma ultrasound image;

[0054] Based on the deep learning prior knowledge graph, the acquired war trauma ultrasound image is identified for artifact noise type, and then the artifact noise category existing in the war trauma ultrasound image is determined;

[0055] According to the identified artifact noise category, a corresponding adaptive filtering algorithm is called to suppress noise of the war trauma ultrasound image to obtain a preprocessed ultrasound image.

[0056] In a specific implementation, first, a large amount of war trauma ultrasound image data is collected, and the artifact region and normal tissue region in the war trauma ultrasound image data are manually labeled. Then, existing feature extraction algorithms are used to extract artifact features and normal tissue features of the corresponding regions, and a structured deep learning prior knowledge graph is constructed according to the categories. The imaging characteristics that are prone to occur in the war trauma ultrasound image during imaging include artifact features and normal tissue features. The artifact features include acoustic shadow and speckle. The normal tissue features include, but are not limited to, gray scale distribution and texture structure. The deep learning prior knowledge graph refers to a structured data set containing feature information of various artifact features and normal tissue features in the war trauma ultrasound image, which can be used to guide subsequent image processing. Second, based on the deep learning prior knowledge graph, a pattern recognition method (such as a support vector machine) is used to extract scanning features of each region in the war trauma ultrasound image. The features of each region in the war trauma ultrasound image are compared and matched with the artifact features in the deep learning prior knowledge graph by using cosine similarity. The artifact category corresponding to the maximum cosine similarity is taken as the artifact noise category (such as acoustic shadow artifact and speckle artifact) of the war trauma ultrasound image. Different artifact features correspond to different artifact noise categories. The artifact noise category refers to an image interference type that does not belong to the real tissue structure in the war trauma ultrasound image. Finally, according to the identified artifact noise category, a corresponding adaptive filtering algorithm is called. When the artifact noise category is speckle artifact, an existing speckle noise reduction algorithm (such as wide filtering) is used to suppress noise of the war trauma ultrasound image to obtain a preprocessed ultrasound image. When the artifact noise category is acoustic shadow artifact, an existing compensation algorithm based on region growing is used to suppress noise of the war trauma ultrasound image to obtain a preprocessed ultrasound image. In addition, other adaptive filtering algorithms can also be used to suppress noise of the war trauma ultrasound image, which is not limited here.

[0057] It should be noted that the preprocessed ultrasound image in the present application refers to the war trauma ultrasound image after artifact noise suppression. By determining the preprocessed ultrasound image, the inherent artifacts (such as speckle and acoustic shadow) in the war trauma ultrasound image are removed and the contrast and clarity of the war trauma ultrasound image are optimized, so that the key features such as texture and edge of the war trauma tissue in the image are more prominent and easy to identify, thereby providing a more reliable basis for subsequent extraction of global context features of image texture and local high-frequency features of image edge from the image.

[0058] In step S102, global context features of image texture and local high-frequency features of image edge are extracted from the preprocessed ultrasound image, and then semantic fusion feature maps containing deep image semantic information are generated according to the global context features and the local high-frequency features.

[0059] In some embodiments, extracting the global context features of image texture and the local high-frequency features of image edges from the preprocessed ultrasound image is achieved by the following steps:

[0060] inputting the preprocessed ultrasound image into a pre-trained deep convolutional neural network, and extracting the global context features of image texture through a deep network of the deep convolutional neural network;

[0061] performing multi-scale Gaussian difference filtering on the preprocessed ultrasound image to generate a multi-scale edge response map;

[0062] extracting the local high-frequency features of image edges based on the multi-scale edge response map.

[0063] In a specific implementation, first, a deep convolutional neural network (such as an existing visual geometry group network) pre-trained on a large-scale war trauma ultrasound image dataset is selected, and the preprocessed ultrasound image is adjusted according to the network input size requirement. The deep convolutional neural network extracts features from the preprocessed ultrasound image layer by layer through a deep network structure composed of multiple convolutional layers and pooling layers, to extract the distribution features and morphological features of image texture in the preprocessed ultrasound image, and then constructs global context features of image texture according to all the extracted features. The global context features refer to a high-dimensional feature set reflecting the distribution features and morphological features of image texture in the preprocessed ultrasound image. Second, the preprocessed ultrasound image is subjected to Gaussian blur processing using Gaussian kernel functions of different scales, to obtain multiple groups of blurred ultrasound images of different scales. Then, the blurred ultrasound images of adjacent scales are subjected to difference operation, to generate a multi-scale edge response map. The multi-scale edge response map refers to an image that responds to edge structures in the preprocessed ultrasound image under different spatial scales, and can highlight edges of different sizes. Finally, based on the multi-scale edge response map, a gradient operator (such as a Sobel operator) is used to calculate the gradient amplitude and direction of the pixel points in the multi-scale edge response map. The pixel points with a gradient amplitude greater than a preset gradient threshold are retained through threshold screening, and then the local high-frequency features of image edges are extracted. The gradient threshold can be set according to actual requirements, which is not limited here.

[0064] It should be noted that the local high-frequency feature in the present application refers to a feature reflecting the local details of the image edge in the pre-processed ultrasound image. Through determination of the global context feature and the local high-frequency feature of the image texture, the global context feature can reflect the morphology of the image texture in the pre-processed ultrasound image, providing information support at the global level for the regional segmentation of the pre-processed ultrasound image, and the local high-frequency feature can accurately capture the detailed change information of the image edge. The combination of the two can realize multi-dimensional feature representation from the whole to the local, avoiding the loss of local details caused by relying only on global features, and preventing overall judgment deviation caused by only focusing on local features, thereby laying a comprehensive and accurate feature foundation for subsequent generation of semantic fusion feature maps containing deep image semantic information.

[0065] In some embodiments, with reference to Figure 2 As shown in the figure, the figure is an exemplary flow chart for determining a semantic fusion feature map according to some embodiments of the present application. In the present embodiment, the generation of a semantic fusion feature map containing deep image semantic information according to the global context feature and the local high-frequency feature can be realized by the following steps:

[0066] In step S1021, the global context feature is subjected to dimension adaptation processing to generate an adapted global feature matching the dimension of the local high-frequency feature;

[0067] In step S1022, an attention mechanism is used to calculate the feature correlation weight of the adapted global feature and the local high-frequency feature;

[0068] In step S1023, the adapted global feature and the local high-frequency feature are weighted and fused based on the feature correlation weight to generate an initial fusion feature map;

[0069] In step S1024, the initial fusion feature map is subjected to feature enhancement through a residual connection module to obtain a semantic fusion feature map containing deep image semantic information.

[0070] In a specific implementation, first, in view of the difference in dimensions between the global context features and the local high-frequency features, a feature mapping method (such as through a fully connected layer) is used to linearly transform the global context features, adjust the feature dimensions of the global context features to be consistent with the dimensions of the local high-frequency features, and generate adaptive global features, which are global context features that can be used for subsequent fusion operations with the local high-frequency features; second, an attention mechanism (such as by calculating the cosine similarity between the adaptive global features and the local high-frequency features) is used to quantify the correlation strength between the two, and then a flexible maximum function (softmax function) is used to convert the correlation strength of each element into a feature correlation weight, which is a weight value used to measure the correlation degree of the adaptive global features and the local high-frequency features at different positions; then, the adaptive global features and the local high-frequency features are element-wise weighted and summed according to the feature correlation weight, to generate an initial fusion feature map, which is a feature map that preliminarily integrates the adaptive global features and the local high-frequency features; finally, an element-level addition is performed between the initial fusion feature map and the features input into the residual connection module (such as the adaptive global features or the local high-frequency features) through a residual connection module, to supplement the details lost in the feature fusion process and enhance the representation ability of the fusion feature map, thereby obtaining a semantic fusion feature map containing deep image semantic information.

[0071] It should be noted that the semantic fusion feature map in this application refers to a feature map that can reflect deep image semantic information in the war injury ultrasound image. The existing technology has the limitation that when global features and local features are simply spliced or weighted fused, feature redundancy and key information dilution easily occur. The present scheme ensures the accurate correlation of the two types of features in a unified space and channel dimension through dimension adaptation, dynamically quantifies the cross-dimension correlation strength of the global context and the local high-frequency features with the help of the feature correlation weight, makes the fusion process more focused on preserving the features related to image semantics in the war injury ultrasound image, and then supplements the weakened global distribution trend or local edge details in the fusion process through residual connection, finally forms a semantic fusion feature map containing deep image semantic information, to accurately preserve the edge gray mutation and subtle texture in the local level of the war injury ultrasound image, and realizes the deep recognition of the semantic information in the war injury ultrasound image.

[0072] In step S103, the feature correlation coefficients between different feature sub-regions in the semantic fusion feature map and the preset feature template are determined based on the preset sliding window, and then the segmentation probability matrix when the preprocessed ultrasound image is segmented is determined according to the feature correlation coefficients corresponding to each feature sub-region.

[0073] In some embodiments, determining the feature correlation coefficients between different feature sub-regions in the semantic fusion feature map and the preset feature template based on the preset sliding window is implemented by the following steps:

[0074] sliding sampling on the semantic fusion feature map according to a preset sliding window to generate a plurality of feature sub-regions;

[0075] determining a feature vector corresponding to each feature sub-region;

[0076] calculating a cosine similarity of the feature vector corresponding to each feature sub-region and a preset feature template;

[0077] normalizing each cosine similarity to obtain a feature correlation coefficient between each feature sub-region and the preset feature template.

[0078] In a specific implementation, first, a fixed-size sliding window (such as 5x5 pixels) is set, and the sliding window is slid along the horizontal and vertical directions on the semantic fusion feature map in a preset step (such as 1 pixel). Each time the sliding window is slid, a corresponding region is intercepted as a feature sub-region, and a plurality of feature sub-regions are generated. The feature sub-region refers to a local feature region intercepted from the semantic fusion feature map by the sliding window. Second, for each feature sub-region, a feature vector extraction method is used to extract the mean, variance, and entropy of the pixel gray value in each feature sub-region, and the mean, variance, and entropy are spliced into a one-dimensional vector by row to generate a feature vector corresponding to each feature sub-region. The feature vector refers to a feature used to quantitatively represent the feature sub-region. Third, a preset feature template is called from a preset war injury feature template library (the preset war injury feature template library includes preset feature templates constructed by historical war injury ultrasound images. The preset feature templates are obtained by image processing of historical war injury ultrasound images based on an existing machine learning model, and details are not described herein. The preset feature template refers to a standard template image including the shape and texture of a historical war injury ultrasound image, and details are not described herein). A cosine similarity of the feature vector corresponding to each feature sub-region and a feature vector corresponding to the preset feature template is calculated. The feature vector corresponding to the preset feature template can be determined in the same manner as the feature vector corresponding to the feature sub-region, and details are not limited herein. The cosine similarity refers to a similarity between the feature vector corresponding to the feature sub-region and the feature vector corresponding to the preset feature template, and is used to reflect the similarity between the image features of the current war injury ultrasound image and the standard image features. Finally, a normalization method (such as linear transformation of the cosine similarity to the interval [0, 1]) is used to process the calculated cosine similarity, so that the result is in a unified numerical range, and a feature correlation coefficient between each feature sub-region and the preset feature template is obtained.

[0079] It should be noted that the feature correlation coefficient in the present application refers to a value for measuring the correlation degree of the feature sub-region and the preset feature template. By determining the feature correlation coefficient, the correlation degree of each feature sub-region in the semantic fusion feature map and the preset feature template can be quantified. On the one hand, the local region highly related to the typical preset feature template in the image can be accurately positioned by the size of the correlation coefficient, thereby providing a quantitative basis for the regional segmentation of the war trauma ultrasound image.

[0080] In some embodiments, determining the segmentation probability matrix for segmenting the preprocessed ultrasound image according to the feature correlation coefficients of the respective feature sub-regions is implemented by the following steps:

[0081] mapping the feature correlation coefficients of the respective feature sub-regions to the corresponding spatial positions of the semantic fusion feature map to generate an initial probability distribution matrix;

[0082] performing spatial smoothing processing on the initial probability distribution matrix to obtain a smoothed probability distribution matrix;

[0083] performing a normalization operation on the smoothed probability distribution matrix by using a flexible maximum function to obtain the segmentation probability value of each pixel point belonging to the segmentation region;

[0084] constructing a segmentation probability matrix for segmenting the preprocessed ultrasound image according to the segmentation probability values of the respective pixel points.

[0085] In a specific implementation, first, the feature correlation coefficients corresponding to each feature sub-region are mapped according to the spatial coordinate positions of the corresponding feature sub-regions in the semantic fusion feature map, that is, each pixel point covered by each feature sub-region is assigned the feature correlation coefficient of the feature sub-region, to form an initial probability distribution matrix, which is a matrix containing initial correlation probabilities of each pixel point and is directly generated by mapping the feature correlation coefficients according to the spatial positions; second, a spatial smoothing algorithm (such as a Gaussian filter) is used to process the initial probability distribution matrix, and the initial correlation probabilities of each pixel point and its neighboring pixel points are weighted and averaged (the weights can be set according to actual requirements, for example, decrease with increasing distance) to obtain a smoothed probability distribution matrix, which is a probability distribution matrix after spatial smoothing processing; then, a flexible maximum value function (softmax function) is applied to the probability value of each pixel point in the smoothed probability distribution matrix, and the probability values of all pixel points are normalized after exponential operation, so that the probability value of each pixel point falls within the interval [0, 1] and the sum of the probability values of all pixel points is 1, to obtain a segmentation probability value of each pixel point belonging to a segmentation region, wherein the segmentation probability value of a pixel point belonging to a segmentation region is a numerical value quantifying the likelihood of the pixel point belonging to the segmentation region, and the segmentation region is a region that needs to be segmented from the preprocessed ultrasound image; finally, the segmentation probability values of all pixel points are arranged according to their spatial positions in the semantic fusion feature map to construct a segmentation probability matrix for segmenting the preprocessed ultrasound image.

[0086] It should be noted that the segmentation probability matrix in this application refers to a matrix used to guide the segmentation of the preprocessed ultrasound image. The segmentation probability matrix provides a precise and quantitative basis for the segmentation and identification of the preprocessed ultrasound image. The probability value of each pixel point in the matrix intuitively reflects the likelihood of the pixel point belonging to a segmentation region. The segmentation probability matrix can not only quickly locate high-probability segmentation regions by setting a probability threshold, but also determine the boundaries and overall morphology of the segmentation by relying on the spatial distribution characteristics of the probability values, thereby avoiding the problems of blurred boundaries or missed regions caused by relying on a single threshold in traditional segmentation methods, and effectively improving the objectivity and accuracy of the segmentation results.

[0087] In step S104, the preprocessed ultrasound image is regionally segmented according to the segmentation probability matrix to obtain a regional segmentation image corresponding to the war injury ultrasound image.

[0088] In some embodiments, the regional segmentation of the preprocessed ultrasound image according to the segmentation probability matrix to obtain a regional segmentation image corresponding to the war injury ultrasound image is implemented by the following steps:

[0089] A first segmentation value and a second segmentation value for the regional segmentation of the preprocessed ultrasound image are determined based on the segmentation probability matrix.

[0090] obtaining the segmentation probability value of the corresponding pixel point in the pre-processed ultrasound image from the segmentation probability matrix;

[0091] dividing the pixel region with the segmentation probability value greater than or equal to the second segmentation value into a first-level segmentation region, dividing the pixel region with the segmentation probability value between the first segmentation value and the second segmentation threshold into a second-level segmentation region, and dividing the pixel region with the segmentation probability value less than or equal to the first segmentation value into a third-level segmentation region, thereby obtaining the regionalized segmentation image corresponding to the war injury ultrasound image.

[0092] In a specific implementation, first, the maximum probability value and the minimum probability value in the segmentation probability matrix are extracted, the probability interval length is determined by the difference between the maximum probability value and the minimum probability value, 1 / 3 times the probability interval length is taken as the reference probability segmentation value, the sum of the minimum probability value and 1 times the reference probability segmentation value is taken as the first segmentation value, and the sum of the minimum probability value and 2 times the reference probability segmentation value is taken as the second segmentation value, wherein the first segmentation value and the second segmentation value refer to the first threshold value and the second threshold value for segmenting the pre-processed ultrasound image; second, the segmentation probability value of the corresponding pixel point in the pre-processed ultrasound image is obtained from the segmentation probability matrix; third, the pixel region with the segmentation probability value greater than or equal to the second segmentation value is divided into a first-level segmentation region, the pixel region with the segmentation probability value between the first segmentation value and the second segmentation threshold is divided into a second-level segmentation region, and the pixel region with the segmentation probability value less than or equal to the first segmentation value is divided into a third-level segmentation region, thereby obtaining the regionalized segmentation image corresponding to the war injury ultrasound image.

[0093] It should be noted that the regional segmentation image in the present application refers to the image obtained by dividing the pixel group with similar deep semantic features in the preprocessed ultrasound image. Due to the fact that the war injury ultrasound image is easily disturbed by the imaging environment (such as the battlefield moving scene and the device shaking), noise and artifacts are easily generated, and the boundary between the damaged tissue and the normal tissue is often unclear, it is difficult to identify the semantic attribution of the pixel region by directly segmenting the original war injury ultrasound image. The segmentation probability value is divided into three levels (first, second and third segmentation regions), which is a targeted design based on the pathological characteristics of the war injury tissue. Specifically, the first segmentation region (high probability region) corresponds to the pixel region with the highest segmentation probability value, which usually represents the core semantic region of the war injury ultrasound image; the second segmentation region (medium probability region) represents the transition region with medium probability value, which usually corresponds to the image edge region of the war injury ultrasound image, and the third segmentation region (low probability region) is the normal tissue or irrelevant background with the lowest probability value. This three-level division strategy overcomes the limitations of traditional binary segmentation in complex war injury scenes, and can more finely depict the heterogeneity gradient of damaged tissues. Therefore, by performing regional segmentation on the war injury ultrasound image, the semantic attribution of the pixel region can be effectively determined, irrelevant background information and noise interference can be filtered out, the segmentation features of image segmentation can be highlighted, the technical problems of semantic ambiguity and unclear segmentation target of the ultrasound image in the war injury scene can be solved, and the effectiveness of the segmentation processing of the war injury ultrasound image can be significantly improved.

[0094] In addition, another aspect of the present application provides a war injury ultrasound image diagnosis system based on deep learning in some embodiments. The system comprises a war injury ultrasound image processing unit, which is described with reference to Figure 3 The figure is a structural schematic diagram of a war injury ultrasound image processing unit according to some embodiments of the present application. The war injury ultrasound image processing unit comprises an acquisition module 201, a processing module 202 and an execution module 203, which are described as follows:

[0095] The acquisition module 201 is mainly used for acquiring the war injury ultrasound image to be processed, and performing adaptive preprocessing based on deep learning prior knowledge on the war injury ultrasound image to obtain a preprocessed ultrasound image.

[0096] The processing module 202 is mainly used for extracting the global context features of image texture and the local high-frequency features of image edge from the preprocessed ultrasound image, and then generating a semantic fusion feature map containing deep image semantic information according to the global context features and the local high-frequency features.

[0097] The processing module 202 is further configured to determine a feature correlation coefficient between different feature sub-regions in the semantic fusion feature map and a preset feature template based on a preset sliding window, and then determine a segmentation probability matrix for segmenting the preprocessed ultrasound image according to the feature correlation coefficient corresponding to each feature sub-region.

[0098] The execution module 203 is mainly configured to perform regional segmentation on the preprocessed ultrasound image according to the segmentation probability matrix to obtain a regional segmentation image corresponding to the war injury ultrasound image.

[0099] In addition, the present application further provides a computer device, which comprises a memory and a processor, the memory stores codes, and the processor is configured to acquire the codes and execute the above-mentioned war injury ultrasound image processing method based on deep learning.

[0100] In some embodiments, with reference to Figure 4 The figure is a structural schematic diagram of a computer device for implementing the war injury ultrasound image processing method based on deep learning according to some embodiments of the present application. The war injury ultrasound image processing method based on deep learning in the above-mentioned embodiments can be implemented by the computer device shown in the figure, which comprises at least one processor 301, a communication bus 302, a memory 303, and at least one communication interface 304. Figure 4

[0101] The processor 301 can be a general central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more circuits for controlling the execution of the war injury ultrasound image processing method based on deep learning in the present application.

[0102] The communication bus 302 can be used to transmit information between the above-mentioned components.

[0103] ​The memory 303 can be a read-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM), or other type of dynamic storage device that can store information and instructions, an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disk storage, a magneto-optical disk, a magnetic disk or other magnetic storage device, or any other medium capable of storing desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited thereto. The memory 303 can exist independently of the processor 301 and be connected to the processor 301 through the communication bus 302. The memory 303 can also be integrated with the processor 301.

[0104] The memory 303 is configured to store program codes for implementing the solutions of the present application, and the processor 301 is configured to control the execution of the program codes. The processor 301 is configured to execute the program codes stored in the memory 303. The program codes can include one or more software modules. The determination of the deep learning-based war injury ultrasound image processing method in the above embodiments can be implemented by one or more software modules in the program codes in the processor 301 and the memory 303.

[0105] The communication interface 304 is configured to communicate with other devices or communication networks, such as an Ethernet, a radio access network (RAN), a wireless local area network (WLAN), etc., using any transceiver-like device.

[0106] In specific implementations, as an example, the computer device can include multiple processors, each of which can be a single-CPU processor or a multi-CPU processor. The processor herein can refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).

[0107] The computer device described above can be a general-purpose computer device or a special-purpose computer device. In a specific implementation, the computer device can be a desktop computer, a laptop computer, a network server, a personal digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. The embodiments of the present application do not limit the type of the computer device.

[0108] In addition, the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the war injury ultrasound image processing method based on deep learning.

[0109] Although the preferred embodiments of the present application have been described, those skilled in the art who are familiar with the basic inventive concept can make additional changes and modifications to the embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0110] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A deep learning-based method for processing war trauma ultrasound images, used in a war trauma ultrasound image diagnostic system for processing war trauma ultrasound images, characterized in that, The method includes the following steps: Acquire the war trauma ultrasound image to be processed, and perform adaptive preprocessing based on deep learning prior knowledge on the war trauma ultrasound image to obtain the preprocessed ultrasound image. Global context features of image texture and local high-frequency features of image edges are extracted from the preprocessed ultrasound image, and then a semantic fusion feature map containing deep image semantic information is generated based on the global context features and the local high-frequency features. Based on a preset sliding window, the feature correlation coefficients between different feature sub-regions in the semantic fusion feature map and the preset feature template are determined, and then the segmentation probability matrix for segmenting the preprocessed ultrasound image is determined according to the feature correlation coefficients corresponding to each feature sub-region. The preprocessed ultrasound image is regionalized based on the segmentation probability matrix to obtain the regionalized segmented image corresponding to the combat trauma ultrasound image.

2. The method as described in claim 1, characterized in that, The war trauma ultrasound images are subjected to adaptive preprocessing based on deep learning prior knowledge to obtain preprocessed ultrasound images, specifically including: Based on the imaging characteristics of war trauma ultrasound images, a deep learning prior knowledge graph containing artifact features and normal tissue features of typical war trauma ultrasound images is constructed. Based on the deep learning prior knowledge graph, artifact noise type identification is performed on the acquired war trauma ultrasound images to determine the category of artifact noise present in the war trauma ultrasound images. Based on the identified artifact noise category, the corresponding adaptive filtering algorithm is called to suppress noise in the war trauma ultrasound image, resulting in a preprocessed ultrasound image.

3. The method as described in claim 1, characterized in that, Extracting global contextual features of image texture and local high-frequency features of image edges from the preprocessed ultrasound image specifically includes: The preprocessed ultrasound image is input into a pre-trained deep convolutional neural network, and the global context features of the image texture are extracted through the deep network of the deep convolutional neural network. The preprocessed ultrasound image is subjected to multi-scale Gaussian difference filtering to generate a multi-scale edge response map; The local high-frequency features of the image edges are extracted based on the multi-scale edge response map.

4. The method as described in claim 1, characterized in that, Generating a semantic fusion feature map containing deep image semantic information based on the global context features and the local high-frequency features specifically includes: The global context features are subjected to dimension adaptation processing to generate adapted global features that match the dimensions of local high-frequency features; An attention mechanism is used to calculate the feature association weights between the adaptive global features and the local high-frequency features; Based on the feature association weights, the adaptive global features and the local high-frequency features are weighted and fused to generate an initial fused feature map; The initial fused feature map is enhanced by the residual connection module to obtain a semantic fused feature map containing deep image semantic information.

5. The method as described in claim 1, characterized in that, Determining the feature correlation coefficients between different feature sub-regions in the semantic fusion feature map and the preset feature template based on a preset sliding window specifically includes: Multiple feature sub-regions are generated by sliding sampling on the semantic fusion feature map according to a preset sliding window; Determine the feature vectors corresponding to each feature sub-region; Calculate the cosine similarity between the feature vector corresponding to each feature sub-region and the preset feature template; The cosine similarity is normalized to obtain the feature correlation coefficient between each feature sub-region and the preset feature template.

6. The method as described in claim 1, characterized in that, The segmentation probability matrix for segmenting the preprocessed ultrasound image is determined based on the feature correlation coefficients corresponding to each feature sub-region. Specifically, this includes: The feature correlation coefficients corresponding to each feature sub-region are mapped to the corresponding spatial locations of the semantic fusion feature map to generate an initial probability distribution matrix; The initial probability distribution matrix is ​​spatially smoothed to obtain a smoothed probability distribution matrix; The smooth probability distribution matrix is ​​normalized using a flexible maximum value function to obtain the segmentation probability value of each pixel belonging to the segmented region; A segmentation probability matrix is ​​constructed based on the segmentation probability value of each pixel when segmenting the preprocessed ultrasound image.

7. The method as described in claim 1, characterized in that, Retrieve the war trauma ultrasound images to be processed from the war trauma ultrasound image storage file.

8. A deep learning-based combat trauma ultrasound image diagnostic system, used to execute the deep learning-based combat trauma ultrasound image processing method as described in any one of claims 1 to 7, wherein the deep learning-based combat trauma ultrasound image diagnostic system includes a combat trauma ultrasound image processing unit, characterized in that, The war trauma ultrasound image processing unit includes: The acquisition module is used to acquire the war trauma ultrasound image to be processed, and to perform adaptive preprocessing based on deep learning prior knowledge on the war trauma ultrasound image to obtain a preprocessed ultrasound image. The processing module is used to extract global context features of image texture and local high-frequency features of image edges from the preprocessed ultrasound image, and then generate a semantic fusion feature map containing deep image semantic information based on the global context features and the local high-frequency features. The processing module is further configured to determine the feature correlation coefficients between different feature sub-regions in the semantic fusion feature map and the preset feature template based on a preset sliding window, and then determine the segmentation probability matrix for segmenting the preprocessed ultrasound image based on the feature correlation coefficients corresponding to each feature sub-region. The execution module is used to perform regional segmentation on the preprocessed ultrasound image according to the segmentation probability matrix to obtain the regional segmentation image corresponding to the combat trauma ultrasound image.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing code, and the processor being configured to retrieve the code and execute the deep learning-based war trauma ultrasound image processing method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the deep learning-based ultrasonic image processing method for combat trauma as described in any one of claims 1 to 7.

Images