A method, device and computer equipment for classifying a cardiac ultrasound image section

Abstract

The application relates to a section classification method and device for a heart ultrasound image, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: performing feature extraction on the heart ultrasound image to obtain image features; performing semantic segmentation and position detection based on the image features to obtain a semantic segmentation graph and corner position information, the semantic segmentation graph comprising a plurality of pixel points and the tissues to which each pixel point belongs, and the corner position information representing the position of the heart of the heart ultrasound image, the heart comprising a plurality of tissues; converting the initial position information of each pixel point in the semantic segmentation graph into polar coordinate information based on the corner position information; and determining the section type of the heart ultrasound image based on the polar coordinate information of each pixel point and the tissue to which each pixel point belongs. The method can effectively solve the problems of weak generalization, insufficient classification granularity and limited quality control effect in current echocardiogram classification.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus, computer device, computer-readable storage medium, and computer program product for cross-sectional classification of cardiac ultrasound images. Background Technology

[0002] Echocardiography is the most widely used method for cardiac examination. During an echocardiogram, the sonographer follows clinical guidelines to manipulate the ultrasound probe and attempt to scan different sections of the heart from multiple angles. Therefore, using echocardiography requires a very high learning curve for the user. Automating the recording process by automatically identifying cardiac section types and assessing the image quality of the current section could greatly simplify the cardiac ultrasound examination procedure.

[0003] However, existing echocardiographic classification techniques, which directly use raw images as input, suffer from poor generalization. Furthermore, current methods do not explicitly utilize the angle and depth information of anatomical structures, making it impossible to distinguish between subdivided sections with similar structures, resulting in insufficient classification granularity. Finally, echocardiographic quality control based on existing target detection schemes does not fully utilize anatomical angle information, thus limiting its effectiveness. Therefore, there is an urgent need for an echocardiographic classification method that can generalize well and support fine-grained classification of the entire heart section. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer device, computer-readable storage medium, and computer program product for segmental classification of cardiac ultrasound images that can generalize well and support fine-grained classification of the entire heart segment, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for cross-sectional classification of cardiac ultrasound images, including:

[0006] Feature extraction is performed on cardiac ultrasound images to obtain image features;

[0007] Based on the image features, semantic segmentation and position detection are performed to obtain a semantic segmentation map and corner position information. The semantic segmentation map includes multiple pixels and the tissue to which each pixel belongs. The corner position information represents the position of the heart in the cardiac ultrasound image. The heart includes multiple tissues.

[0008] Based on the corner position information, the initial position information of each pixel in the semantic segmentation map is converted into polar coordinate information;

[0009] Based on the polar coordinate information of each pixel and the tissue to which each pixel belongs, the section type of the cardiac ultrasound image is determined.

[0010] In one embodiment, determining the section type of the cardiac ultrasound image based on the polar coordinate information of each pixel and the tissue to which each pixel belongs includes:

[0011] The semantic segmentation map after polar coordinate transformation is input into a feature extractor, and the feature extractor extracts features from the semantic segmentation map after polar coordinate transformation to obtain a target feature representation. The target feature representation is input into a section classifier, and the section classifier determines the section type of the cardiac ultrasound image based on the target feature representation.

[0012] In one embodiment, the step of extracting features from the semantic segmentation map after polar coordinate transformation using the feature extractor to obtain a target feature representation includes:

[0013] The semantic segmentation map after polar coordinate transformation is segmented into multiple local images of the same size by a feature extractor. For each local image, the local image is converted into an initial feature representation of a preset dimension. Based on the polar coordinate information of the pixels in the local image, the self-attention weights corresponding to the local image are determined. The initial feature representation and the self-attention weights are weighted to obtain the target feature representation.

[0014] In one embodiment, determining the section type of the cardiac ultrasound image based on the target feature representation using the section classifier includes:

[0015] The target feature representation is flattened and straightened to obtain a one-dimensional feature representation. For each section type, the weight value and bias value corresponding to the targeted section type are obtained. The one-dimensional feature representation and the weight value are weighted to obtain a linear combination feature representation. The linear combination feature representation and the bias value are added to obtain the activation value corresponding to the targeted section type. Based on the activation value corresponding to each section type, a section type probability distribution feature vector is obtained. Based on the section type probability distribution feature vector, the section type of the cardiac ultrasound image is determined.

[0016] In one embodiment, determining the section type of the cardiac ultrasound image based on the section type probability distribution feature vector includes:

[0017] The activation value corresponding to each section type is converted into the probability that the cardiac ultrasound image is classified into each section type; the section type of the cardiac ultrasound image is obtained based on the probability that the cardiac ultrasound image is classified into each section type.

[0018] In one embodiment, the target feature representation is input to a quality regressor, and a detection box is generated in the cardiac ultrasound image based on the target feature representation. The detection box is used to characterize the location of different tissue structures in the heart. The section quality evaluation rule corresponding to the section type of the cardiac ultrasound image is obtained. The section quality of the cardiac ultrasound image is determined according to the section quality evaluation rule and the detection box generated in the cardiac ultrasound image.

[0019] In one embodiment, the target feature representation is input to a quality evaluator, and the quality evaluator generates defect language text based on the target feature representation, the defect language text being used to describe the cross-sectional defects of the cardiac ultrasound image.

[0020] Secondly, this application also provides a section classification device for cardiac ultrasound images, comprising:

[0021] The extraction module is used to extract features from cardiac ultrasound images to obtain image features;

[0022] The module is used to perform semantic segmentation and location detection based on the image features to obtain a semantic segmentation map and corner location information. The semantic segmentation map includes multiple pixels and the tissue to which each pixel belongs. The corner location information represents the location of the heart in the cardiac ultrasound image. The heart includes multiple tissues.

[0023] The conversion module is used to convert the initial position information of each pixel in the semantic segmentation map into polar coordinate information based on the corner position information;

[0024] The determination module is used to determine the section type of the cardiac ultrasound image based on the polar coordinate information of each pixel and the tissue to which each pixel belongs.

[0025] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0026] Feature extraction is performed on cardiac ultrasound images to obtain image features;

[0027] Based on the image features, semantic segmentation and location detection are performed to obtain a semantic segmentation map and corner location information. The semantic segmentation map includes multiple pixels and the tissue to which each pixel belongs. The corner location information represents the location of the heart in the cardiac ultrasound image. The heart includes multiple tissues.

[0028] Based on the corner position information, the initial position information of each pixel in the semantic segmentation map is converted into polar coordinate information;

[0029] Based on the polar coordinate information of each pixel and the tissue to which each pixel belongs, the section type of the cardiac ultrasound image is determined.

[0030] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0031] Feature extraction is performed on cardiac ultrasound images to obtain image features;

[0032] Based on the image features, semantic segmentation and position detection are performed to obtain a semantic segmentation map and corner position information. The semantic segmentation map includes multiple pixels and the tissue to which each pixel belongs. The corner position information represents the position of the heart in the cardiac ultrasound image. The heart includes multiple tissues.

[0033] Based on the corner position information, the initial position information of each pixel in the semantic segmentation map is converted into polar coordinate information;

[0034] Based on the polar coordinate information of each pixel and the tissue to which each pixel belongs, the section type of the cardiac ultrasound image is determined.

[0035] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0036] Feature extraction is performed on cardiac ultrasound images to obtain image features;

[0037] Based on the image features, semantic segmentation and position detection are performed to obtain a semantic segmentation map and corner position information. The semantic segmentation map includes multiple pixels and the tissue to which each pixel belongs. The corner position information represents the position of the heart in the cardiac ultrasound image. The heart includes multiple tissues.

[0038] Based on the corner position information, the initial position information of each pixel in the semantic segmentation map is converted into polar coordinate information;

[0039] Based on the polar coordinate information of each pixel and the tissue to which each pixel belongs, the section type of the cardiac ultrasound image is determined.

[0040] The aforementioned methods, devices, computer equipment, computer-readable storage media, and computer program products for cross-sectional classification of cardiac ultrasound images extract image features, effectively extracting the features of target anatomical structures from cardiac ultrasound images. This not only helps improve segmentation accuracy but also increases the model's generalization ability to different devices and imaging conditions, thereby reducing the impact of device differences.

[0041] Based on the image features, semantic segmentation and location detection are performed to obtain a semantic segmentation map and corner location information. The semantic segmentation map includes multiple pixels and the tissue to which each pixel belongs. The corner location information characterizes the location of the heart in the cardiac ultrasound image, and the heart comprises multiple tissues. Detailed anatomical structure labels are generated, and the heart's location in the image is accurately located. This solves the generalization problem that easily occurs when traditional methods rely solely on the original image for classification, ensuring that the algorithm can focus on the target structure itself rather than background noise or the imaging style of a specific machine. Simultaneously, the detection of fan-shaped corner points provides necessary geometric reference points for subsequent conversion to the polar coordinate system.

[0042] Based on the corner point location information, the initial position information of each pixel in the semantic segmentation map is converted into polar coordinate information, allowing the angle and depth information of the anatomical structure to be explicitly utilized. This is crucial for distinguishing fine-grained sections with similar anatomical structures but different angles, greatly improving the granularity and accuracy of classification. Furthermore, polar coordinate representation helps to better capture the spatial relationships of the internal structures of the heart, further enhancing the model's expressiveness.

[0043] Based on the polar coordinates of each pixel and the tissue to which each pixel belongs, the section type of the cardiac ultrasound image is determined. High-precision classification of cardiac ultrasound section types is achieved using polar coordinate data containing depth and angle information. Compared to existing technologies, this method can identify various standard and derived cardiac sections at a finer granularity, greatly simplifying the workflow of sonographers and improving the quality control level of automated recording.

[0044] By introducing anatomical semantic segmentation and polar coordinate transformation through the above steps, the problems of weak generalization, insufficient classification granularity, and limited quality control in the current echocardiography classification are effectively solved, providing strong technical support for achieving more intelligent and efficient cardiac ultrasound examinations. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is an application environment diagram of a cross-sectional classification method for cardiac ultrasound images in one embodiment;

[0047] Figure 2 This is a flowchart illustrating a cross-sectional classification method for cardiac ultrasound images in one embodiment.

[0048] Figure 3 This is a schematic diagram of polar coordinate transformation for a semantic segmentation graph in one embodiment;

[0049] Figure 4 Here is a feature extraction network graph from one embodiment;

[0050] Figure 5 This is a flowchart illustrating a cross-sectional classification method for cardiac ultrasound images in another embodiment;

[0051] Figure 6 This is a structural block diagram of a cross-sectional classification device for cardiac ultrasound images in one embodiment;

[0052] Figure 7 This is a structural block diagram of a cross-sectional classification device for cardiac ultrasound images in another embodiment;

[0053] Figure 8 This is a standard section of the left ventricular long axis and a derived section with increased depth of the left ventricular long axis in one embodiment;

[0054] Figure 9 Here are standard cross-sections of the apical four-chamber heart (centered on the left ventricle) and derived cross-sections of the apical four-chamber heart (centered on the right ventricle) in one embodiment;

[0055] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0057] The section classification method for cardiac ultrasound images provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or located in the cloud or on other network servers. Terminal 102 generates a section classification request for the cardiac ultrasound image and sends this request to server 104, so that server 104 can determine the section type of the cardiac ultrasound image based on the polar coordinate information of each pixel and the tissue to which each pixel belongs. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can be smartwatches, smart bracelets, head-mounted devices, etc. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides cloud computing services.

[0058] In one exemplary embodiment, such as Figure 2 As shown, a cross-sectional classification method for cardiac ultrasound images is provided, which can be applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 202 to 208. Wherein:

[0059] Step 202: Extract features from the cardiac ultrasound image to obtain image features.

[0060] Image features refer to information extracted from raw echocardiogram images that characterize the image content. This information is used in subsequent processing steps such as semantic segmentation and location detection. Understandably, in this application, image features are multi-scale features, with each scale corresponding to a level. For echocardiogram images, "image features" particularly emphasize characteristics related to cardiac anatomy, such as the location, size, and shape of ventricles, atria, valves, and other blood vessels. These features are crucial for accurate semantic segmentation (i.e., distinguishing different cardiac tissues).

[0061] Echocardiography images are two-dimensional grayscale images of the internal structures of the heart acquired using echocardiography equipment. These images reflect different sections of the heart, such as the long axis of the left ventricle and the apical four-chamber view, and can be static images or frames from dynamic video. Optionally, an ultrasound probe can be used to scan the heart from different angles and positions to obtain echocardiography images.

[0062] Specifically, the echocardiogram image is input into a feature extraction network. Understandably, the feature extraction network in this application includes at least an encoder, a decoder, and an output head. First, in the encoder stage, the encoder progressively reduces the spatial dimensions (width and height) of the image through a series of convolutional and pooling layers, while simultaneously increasing the number of feature channels. Each layer captures a higher-level feature representation. Furthermore, the encoder stage also saves the feature maps of each layer for use in the subsequent decoder stage.

[0063] Then, in the decoder stage, deconvolution (upsampling) is used to recover the spatial dimensions of the image, progressively reconstructing a high-resolution segmentation map. During this process, feature maps from the corresponding levels of the encoder are merged with the output of the current decoder level, ensuring that location information and details are preserved. Understandably, the merged outputs of the encoder and decoder at each level constitute the image features. Therefore, the image features are multi-scale features, thus ensuring the preservation of location information and details.

[0064] Step 204: Perform semantic segmentation and location detection based on image features to obtain a semantic segmentation map and corner location information. The semantic segmentation map includes multiple pixels and the tissue to which each pixel belongs. The corner location information represents the location of the heart in the cardiac ultrasound image. The heart includes multiple tissues.

[0065] Semantic segmentation maps, in particular, classify and label each pixel in a raw echocardiogram image, indicating which specific target structure (such as the ventricle, atrium, etc.) or background that the pixel belongs to. Understandably, semantic segmentation maps are used to clearly identify the specific location and extent of each anatomical structure in the image, providing a basis for subsequent polar coordinate transformation.

[0066] Optionally, the semantic segmentation map is an image of the same size as the original image, but the color or intensity values ​​no longer represent actual tissue density or blood flow velocity, but rather categories of different anatomical structures. For example, all parts belonging to the left ventricle might be red, the right ventricle blue, the mitral valve green, and so on; the background or other non-target areas are typically represented by black or other designated colors.

[0067] Corner location information refers to the positional information of three key points within a sector—the two endpoints and the vertex (i.e., the center of the sector). These three points uniquely determine the shape and orientation of the sector, helping to describe the geometric characteristics in cardiac ultrasound images. Understandably, corner location information is crucial for subsequent polar coordinate transformations because it provides the necessary geometric reference points, enabling the accurate calculation of the angle and distance of any given point relative to the sector vertex, which is essential for understanding the angles and depths of anatomical structures.

[0068] Tissues refer to the different anatomical structures or regions inside the heart, such as the ventricles, atria, valves, and great vessels. These structures constitute the main components of the heart. In a semantic segmentation map, each pixel is assigned a label that indicates the specific anatomical structure (i.e., tissue) to which the pixel belongs. For example, all pixels belonging to the left ventricle are labeled as one category, all pixels belonging to the right ventricle are labeled as another category, and so on.

[0069] Specifically, the feature extraction network includes two output heads. One output head ultimately generates a semantic segmentation map, in which each pixel is assigned a label indicating its associated anatomical structure or background. The other output head is specifically used to predict the positions of three corner points within a fan-shaped region. Optionally, the coordinates of these three points, i.e., the corner point position information, can be inferred directly from image features through a fully connected network.

[0070] Step 206: Based on the corner position information, convert the initial position information of each pixel in the semantic segmentation map into polar coordinate information.

[0071] The initial location information refers to the position of each pixel in the original cardiac ultrasound image without any transformation processing.

[0072] Polar coordinate information refers to the representation of each pixel after transforming it from a Cartesian coordinate system to a polar coordinate system. The polar coordinate system uses two parameters to define a point: the polar coordinate *r* (also called radial distance) and the polar angle *θ* (angle). In this application, the polar coordinate *r* represents the distance from the apex of the sector (usually the probe position) to a point, reflecting the "depth" of the cardiac structure at that point relative to the probe. The polar angle *θ* represents the angle formed by the line segment connecting the origin and the point from a reference direction (e.g., the positive x-axis), used to describe the directionality of the anatomical structure.

[0073] Specifically, refer to Figure 3 First, determine the origin and reference direction of the polar coordinate system. This includes using the vertex of one of the three corner points (including the polar origin, left pole, and right pole) as the origin of the polar coordinate system (i.e., the position of the ultrasound probe); and selecting a reference direction, typically the direction pointing horizontally to the right from the vertex as the 0-degree reference line. This step ensures the consistency and accuracy of all angle measurements. Understandably, corner point position information refers to the individual position information of the polar origin, left pole, and right pole.

[0074] Then, the polar radius r and polar angle θ of each pixel are calculated, including: for each pixel, the straight-line distance from it to the origin (i.e., the vertex) of the polar coordinate system is calculated, which is the polar radius r. This distance reflects the depth of the structure relative to the probe. The polar angle θ is then calculated from the reference direction to the angle formed by the line segment connecting the origin and the pixel. This angle describes the directionality of the anatomical structure.

[0075] Finally, a new representation is created for each pixel, containing the polar radius r and polar angle θ. Optionally, this can be achieved by adding extra channels or metadata to the existing semantic segmentation map, or by directly saving this information in a new image format so that these spatial features can be reused in subsequent processing steps.

[0076] Step 208: Determine the section type of the cardiac ultrasound image based on the polar coordinate information of each pixel and the tissue to which each pixel belongs.

[0077] In echocardiography, the section type refers to the specific anatomical view or perspective represented by the cardiac ultrasound image. During an echocardiogram, different views of cardiac structures are obtained by adjusting the position and angle of the ultrasound probe; these specific combinations of angles and positions are called "sections." Each section type provides a unique way to observe different parts of the heart and their functions. For example, an echocardiogram can involve up to 27 two-dimensional sections, and an additional 55 Doppler modes can be derived from these basic sections, for a total of 72 sections. Each section has its specific criteria, such as the left ventricular long axis and the apical four-chamber view, used to assess specific cardiac structures such as valves, myocardial wall motion, and blood flow direction. (Reference) Figure 7 This image presents a standard view of the heart obtained by scanning the heart with an ultrasound probe at a specific angle, showing the long axis of the left ventricle, including structures such as the aortic valve, left ventricular wall, mitral valve, and part of the right ventricular outflow tract. This view is a fundamental and commonly used standard section in echocardiography for assessing cardiac function and structure. (Reference) Figure 8 Based on a standard left ventricular long-axis section, adjusting the position or parameters of the ultrasound probe allows for scanning deeper into the heart, resulting in a deeper field of view. This provides more information about the state of structures deep within the heart, such as a better view of the apex region of the left ventricle or a clearer view of certain structures located at the back of the heart. Such derived sections may be necessary for the diagnosis of certain diseases because they provide additional information that helps doctors make more accurate judgments.

[0078] Specifically, the polar coordinate-transformed cardiac ultrasound semantic segmentation image is input into a cardiac section prediction network. This cardiac section prediction network includes at least a feature extractor, a section classifier, a quality regressor, and a quality evaluator. The feature extractor processes the polar coordinate-transformed cardiac ultrasound semantic segmentation image, outputting a feature representation rich in contextual information (i.e., the target feature representation). Understandably, the target feature representation not only includes the spatial information of the original image but also incorporates depth and angle information. The section classifier processes the target feature representation output by the feature extractor, outputting the section type of the cardiac ultrasound image.

[0079] In one embodiment, a quality regressor processes the target feature representation output by the feature extractor to output the cross-sectional quality of the cardiac ultrasound image. A quality evaluator then processes the target feature representation output by the feature extractor to output a quality text description of the cross-sectional quality of the cardiac ultrasound image.

[0080] In one embodiment, the semantic segmentation map after polar coordinate transformation is input to a feature extractor, and the feature extractor extracts features from the semantic segmentation map after polar coordinate transformation to obtain a target feature representation; the target feature representation is input to a section classifier, and the section classifier determines the section type of the cardiac ultrasound image based on the target feature representation.

[0081] The feature extractor, as part of the cardiac segmentation prediction network, is used to automatically extract useful features from the input data. In this specific application scenario, the feature extractor is designed to handle cardiac ultrasound semantic segmentation maps after polar coordinate transformation.

[0082] A section classifier is a module specifically designed to identify and classify which standard or derived section an echocardiogram image belongs to. Understandably, based on feature representations provided by a feature extractor, the section classifier performs a multi-class classification task to determine whether an input echocardiogram image belongs to a specific standard section or a more fine-grained derived section.

[0083] Target feature representations refer to feature vectors or feature matrices rich in contextual information. They integrate anatomical structural information from the raw echocardiogram images, along with their angular and depth information, and are a high-level abstract representation that effectively supports subsequent tasks such as section classification. These feature representations are crucial for achieving fine-grained echocardiogram section classification because they contain the necessary details to distinguish different section types.

[0084] Specifically, the input data is first prepared, including ensuring that each pixel in the polar coordinate-transformed semantic segmentation image contains not only its initial position information (Cartesian coordinates) but also its calculated polar radius r and polar angle θ. Each pixel has also been assigned a label indicating the specific anatomical structure to which it belongs. The feature extractor then receives the polar coordinate-transformed semantic segmentation image as input. The feature extractor processes the polar coordinate-transformed echocardiogram semantic segmentation image, outputting a context-rich feature representation (i.e., the target feature representation). This target feature representation is then sent to the section classifier. Finally, the section classifier receives the target feature representation from the feature extractor. The section classifier processes the target feature representation and outputs the section type of the echocardiogram image.

[0085] By using anatomical semantic segmentation maps instead of the original images as input, the differences in imaging texture and background noise from different devices are eliminated, enabling the algorithm to better adapt to different brands and models of ultrasound equipment and enhancing the system's generalization ability. Furthermore, through polar coordinate transformation and semantic segmentation, the model can more accurately capture the angular and depth information of anatomical structures. This helps solve the problem of traditional methods struggling to distinguish fine-grained section types, thereby improving classification accuracy.

[0086] In one embodiment, a feature extractor segments the semantic segmentation map after polar coordinate transformation into multiple local images of the same size; for each local image, the local image is converted into an initial feature representation of a preset dimension; based on the polar coordinate information of the pixels in the local image, the self-attention weight corresponding to the local image is determined, and the initial feature representation and the self-attention weight are weighted to obtain the target feature representation.

[0087] In this context, a local image refers to a segmentation of an image into multiple small blocks of the same size when processing semantic segmentation maps after polar coordinate transformation. This is done to facilitate model processing and learning of finer-grained spatial information, and each such block is called a "local image".

[0088] The initial feature representation of a preset dimension refers to converting each local image into a vector of a fixed dimension (preset dimension). This vector is the "initial feature representation" of that local image. Optionally, this involves mapping each local image to a low-dimensional space using a specific encoding method (such as convolutional layers or linear transformations). For example, if the original local image is h×w in size, it can be converted into a vector of length d, where d is the preset dimension. Understandably, despite the dimensionality reduction, these initial feature representations still retain the key features of the local image, such as color and texture, and can reflect the relative positional relationships between pixels.

[0089] Self-attention weights refer to the self-attention weight matrix, used to measure the importance of different pixels or local images. It allows the model to dynamically adjust the influence of each part on the final output. Intuitively, it utilizes the polar radius r and polar angle θ of each pixel to calculate the importance of each pixel within a local image. The directionality and depth information of anatomical structures are considered to ensure the model can focus on key regions.

[0090] Specifically, the semantic segmentation map after polar coordinate transformation is first input into a feature extractor, which then segments the semantic segmentation map into multiple small patches (i.e., local images) of the same size. Each patch represents a part of the original image, helping the model to capture local features more precisely. Optionally, the segmentation can be uniform or non-uniform based on key regions of the anatomical structure to ensure that important structures are not over-segmented.

[0091] Then, for each local image, a specific encoding method (such as convolutional layers or linear transformations) is used to map it to a vector space of fixed dimensions (preset dimension) to obtain an initial feature representation. Understandably, the choice of the preset dimension depends on the model design and computational resource constraints, ensuring that the local image features are fully expressed while maintaining computational efficiency.

[0092] Next, the polar radius r and polar angle θ of each pixel in the local image are used to calculate the self-attention weights. Polar coordinate information helps the model understand the specific location and orientation of each pixel within the heart structure. Different pixels correspond to different self-attention weight matrices, thereby determining the importance of each pixel within each local image.

[0093] Finally, for each local image, its initial feature representation is multiplied by the corresponding self-attention weight matrix for weighting. This means that pixels or local images deemed more important will have a greater influence in subsequent tasks. The weighted feature representation, i.e., the target feature representation, more accurately reflects the important structural information in the local image while preserving the global context.

[0094] By segmenting the semantic segmentation map after polar coordinate transformation into multiple local images of the same size, each small patch can be analyzed and processed independently. This method captures more refined anatomical information, improving the understanding of different regions of the heart. Furthermore, introducing self-attention weights and weighting them with the initial feature representation allows the model to dynamically adjust the influence of each part on the final output. This helps highlight key anatomical structures and reduce the influence of irrelevant or redundant information, thereby improving classification accuracy.

[0095] In one embodiment, the target feature representation is flattened and straightened to obtain a one-dimensional feature representation; for each section type, the weight value and bias value corresponding to the targeted section type are obtained; the one-dimensional feature representation and the weight value are weighted to obtain a linear combination feature representation; the linear combination feature representation and the bias value are added to obtain the activation value corresponding to the targeted section type; based on the activation value corresponding to each section type, a section type probability distribution feature vector is obtained; based on the section type probability distribution feature vector, the section type of the cardiac ultrasound image is determined.

[0096] In this context, a one-dimensional feature representation refers to a one-dimensional vector obtained by flattening and straightening the target feature representation (which is usually multi-dimensional). Understandably, despite the dimensionality transformation, all information in the original feature representation is retained, ensuring that important anatomical features are not lost. Weight values ​​are a set of numerical parameters corresponding to each facet type, used to measure the importance of different features. The bias value is an additional parameter added to the linear combination feature representation to adjust the position of the output. The linear combination feature representation is a new feature representation obtained by weighting the one-dimensional feature representation and the weight values. The activation value is the result of adding the bias value to the linear combination feature representation. The facet type probability distribution feature vector is the probability distribution obtained by applying an activation function to the activation values ​​of all facet types.

[0097] Specifically, the target feature representation is first input into a section classifier, which flattens the multidimensional target feature representation into a one-dimensional vector, i.e., a one-dimensional feature representation. This step simplifies the input format, making subsequent linear transformations and calculations easier to implement. It ensures that important anatomical structural information is not lost during the flattening process. For each possible section type, a set of corresponding weight values ​​W is pre-trained and stored. i and bias value b i , where i represents different face types.

[0098] Next, for each section type i, the linear combination feature representation Z is calculated. i =X×W i The contribution of each feature to a specific facet type in the one-dimensional feature representation is measured by calculating the linear combination feature representation corresponding to each facet type. Here, X is the one-dimensional feature representation, and Z... i This represents the linear combination feature representation for each facet type. For each facet type i, calculate the activation value ai = Z. i +b i This step comprehensively considers the impact of all features and their weights, as well as the role of bias values, to form a preliminary score or rating for a specific aspect type.

[0099] Finally, an activation function is applied to transform all activation values ​​into a normalized probability distribution pi= / Where j represents all possible slice types. This step generates a probability distribution reflecting the relative likelihood that the input image belongs to each slice type. Finally, the output slice type probability distribution feature vector p = [p1, p2, ..., p...] is generated. n Let p be the number of facet types. The facet type with the highest probability is selected as the final classification result. That is, find the maximum value p in p. max The corresponding index i represents the predicted section type of the cardiac ultrasound image. Optionally, if the maximum probability value is below a certain threshold, it can be marked as uncertain or requiring further investigation.

[0100] By weighting the one-dimensional feature representation with weight values ​​and combining them with bias values ​​to calculate activation values, this approach allows the model to flexibly handle different types of echocardiographic images, maintaining good classification performance even with data acquired using new equipment or under different conditions. The uncertainty of the classification results can be assessed using the probability distribution. A low maximum probability value suggests potential ambiguity or anomalies, requiring further investigation or confirmation, thus enhancing the system's robustness and reliability.

[0101] In one embodiment, the activation value corresponding to each section type is converted into the probability that the cardiac ultrasound image is classified into each section type; the section type of the cardiac ultrasound image is obtained based on the probability that the cardiac ultrasound image is classified into each section type.

[0102] Specifically, the activation function is applied to transform all activation values ​​into a normalized probability distribution pi= / Where j represents all possible slice types. This step generates a probability distribution reflecting the relative likelihood that the input image belongs to each slice type. Finally, the output slice type probability distribution feature vector p = [p1, p2, ..., p...] is generated. n Let p be the number of facet types. The facet type with the highest probability is selected as the final classification result. That is, find the maximum value p in p. max The corresponding index i represents the predicted section type of the cardiac ultrasound image. Optionally, if the maximum probability value is below a certain threshold, it can be marked as uncertain or requiring further investigation.

[0103] By transforming activation values ​​into a probability distribution, not only is the most likely slice type provided, but also other possibilities and their probabilities are given. This not only improves classification accuracy but also provides a quantitative assessment of the reliability of the classification results.

[0104] In one embodiment, the target feature representation is input into a quality regressor, and a detection box is generated in the cardiac ultrasound image based on the target feature representation. The detection box is used to characterize the location of different tissue structures in the heart. The section quality evaluation rules corresponding to the section type of the cardiac ultrasound image are obtained. The section quality of the cardiac ultrasound image is determined according to the section quality evaluation rules and the detection box generated in the cardiac ultrasound image.

[0105] The quality regressor is a module responsible for evaluating the quality of cardiac ultrasound images and providing a quantitative assessment. It processes the target feature representation output by the feature extractor to estimate a continuous value as a scoring metric to reflect the quality of the current slice. Optionally, it outputs a numerical value representing a quantitative assessment of the current slice quality (e.g., a score between 1 and 5).

[0106] A bounding box is a rectangular box that marks the location of a specific anatomical structure in an echocardiogram image. Each bounding box corresponds to a specific tissue structure in the heart, such as a ventricle, atrium, or valve. They clearly define the specific location of these structures in the image.

[0107] Slice quality assessment rules refer to a set of standards or criteria established for each possible type of cardiac ultrasound slice to measure the quality of that type of image. Understandably, specific scoring criteria are set for different slice types, such as image sharpness, visibility of anatomical structures, and other factors. These rules guide how to assess image quality based on the detection box and other features.

[0108] Section quality refers to the overall quality level of a cardiac ultrasound image after considering various factors. It is represented by one or more numerical values, which intuitively reflect the image's quality grade. High-quality images mean a clearer and more complete display of anatomical structures, which is beneficial for accurate medical diagnosis.

[0109] Specifically, the target feature representation is first input into a quality regressor, which then generates multiple bounding boxes in the echocardiogram image based on the target feature representation. Each bounding box corresponds to a specific cardiac anatomical structure (such as the ventricle, atrium, valve, etc.) and explicitly indicates the specific location of that structure in the image.

[0110] Then, based on the section type previously determined by the section classifier, the section quality evaluation rule corresponding to that section type is retrieved from a pre-defined rule base. Understandably, the section quality evaluation rule defines the scoring criteria for that section type, such as factors like image sharpness and anatomical structure visibility.

[0111] Finally, combining the slice quality evaluation rules and the anatomical information provided by the detection box, the quality regressor comprehensively considers various factors, such as image sharpness, anatomical integrity, and positional accuracy. It calculates one or more continuous values ​​as scoring indicators to reflect the slice quality of the cardiac ultrasound image. Optionally, slice quality is expressed as a score within a numerical range (e.g., a score between 1 and 5). If the score is below a preset threshold, the image quality is marked as low, indicating the need for further examination or re-acquisition.

[0112] By employing slice quality assessment rules corresponding to the slice type of cardiac ultrasound images, it ensures that different physicians or systems use the same standards when evaluating the same type of images. This reduces subjective differences and enhances the consistency and reliability of assessment results. Furthermore, assessment based on target feature representations considers not only the presence or absence of anatomical structures but also incorporates depth and angle information, enabling the effective differentiation of slices that are structurally similar but differ slightly in angle or position.

[0113] In one embodiment, a target feature representation is input to a quality evaluator, which then generates a defect language text based on the target feature representation. The defect language text is used to describe cross-sectional defects in the cardiac ultrasound image.

[0114] The quality evaluator is a module used to describe the cross-sectional defects of cardiac ultrasound images. It receives the target feature representation from the feature extractor as input and outputs a description of the defects in the cardiac ultrasound cross-section, i.e., defect language text, through a series of processing steps.

[0115] The defect language text refers to a textual description generated by the quality evaluator based on the target feature representation. This text is used to point out problems or areas for improvement in the current cardiac ultrasound section. Specifically, the cardiac section prediction network also includes a quality evaluator, which generates a textual description of the defects in the current section. Optionally, the target feature representation is input into the quality evaluator, which uses a trained section defect language model employing an encoder-decoder architecture or a conditional generative adversarial network. The section defect language model generates descriptive text about the defects in the current section based on information from the target feature representation. For example, the output might be a textual description pointing out problems or areas for improvement in the current section, such as "misaligned mitral valve" or "apex misalignment."

[0116] In one embodiment, reference Figure 4 This describes the processing procedure of the feature extraction network.

[0117] In the encoder section: the input image size is 572x572, which is processed through a 3x3 convolutional layer for feature extraction. Then, a max-pooling layer is used for downsampling, reducing the size to 282x282. Further downsampling is performed through convolutional and max-pooling layers, resulting in sizes of 140x140, 68x68, and 32x32 respectively.

[0118] In the bottleneck layer: In the deepest downsampling layer, features are extracted through a convolutional layer with a size of 32x32.

[0119] In the decoder section: upsampling is performed through an upsampling layer, changing the size to 64x64. The upsampled feature map is concatenated with the feature map of the corresponding layer in the encoder, and then feature fusion is performed through a convolutional layer. Upsampling continues through upsampling layers and feature fusion, with the sizes successively changing to 128x128, 256x256, and 512x512.

[0120] In the output layer, a final upsampling layer is used to resize the image to 388x388. A 1x1 convolutional layer then generates the final segmentation map, also 388x388 in size. This architecture allows the feature extraction network to effectively perform image segmentation while maintaining high resolution.

[0121] In one exemplary embodiment, such as Figure 5 As shown, steps 502 to 510 are included. Wherein:

[0122] Step 502: Extract features from the cardiac ultrasound image to obtain image features;

[0123] Step 504: Perform semantic segmentation and location detection based on image features to obtain a semantic segmentation map and corner location information. The semantic segmentation map includes multiple pixels and the tissue to which each pixel belongs. The corner location information represents the location of the heart in the cardiac ultrasound image. The heart includes multiple tissues.

[0124] Step 506: Based on the corner position information, convert the initial position information of each pixel in the semantic segmentation map into polar coordinate information;

[0125] Step 508: The semantic segmentation map after polar coordinate transformation is input into the feature extractor, which segments the semantic segmentation map into multiple local images of the same size. For each local image, it is converted into an initial feature representation of a preset dimension. Based on the polar coordinate information of the pixels in the local image, the self-attention weights corresponding to the local image are determined, and the initial feature representation and the self-attention weights are weighted to obtain the target feature representation. The target feature representation is input into the section classifier, which flattens and straightens the target feature representation. The process involves: obtaining a one-dimensional feature representation; for each section type, acquiring the corresponding weight and bias values; weighting the one-dimensional feature representation and weight values ​​to obtain a linear combination feature representation; summing the linear combination feature representation and bias values ​​to obtain the activation value corresponding to the corresponding section type; obtaining the section type probability distribution feature vector based on the activation values ​​corresponding to each section type; converting the activation values ​​corresponding to each section type into the probability that the echocardiogram image is classified into each section type; and obtaining the section type of the echocardiogram image based on the probability that the echocardiogram image is classified into each section type.

[0126] Step 510: Input the target feature representation into the quality regressor, and generate a detection box in the echocardiogram image based on the target feature representation. The detection box is used to characterize the location of different tissue structures in the heart. Obtain the section quality evaluation rules corresponding to the section type of the echocardiogram image. Determine the section quality of the echocardiogram image based on the section quality evaluation rules and the detection box generated in the echocardiogram image. Input the target feature representation into the quality evaluator, and generate defect language text based on the target feature representation. The defect language text is used to describe the section defects of the echocardiogram image.

[0127] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0128] Based on the same inventive concept, this application also provides a cardiac ultrasound image section classification device for implementing the above-described method for section classification of cardiac ultrasound images. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the cardiac ultrasound image section classification device provided below can be found in the limitations of the cardiac ultrasound image section classification method described above, and will not be repeated here.

[0129] In one exemplary embodiment, such as Figure 6 As shown, a section classification device 600 for cardiac ultrasound images is provided, including: an extraction module 602, an acquisition module 604, a conversion module 606, and a determination module 608, wherein:

[0130] Extraction module 602 is used to extract features from cardiac ultrasound images to obtain image features;

[0131] The module 604 is used to perform semantic segmentation and location detection based on image features to obtain a semantic segmentation map and corner location information. The semantic segmentation map includes multiple pixels and the tissue to which each pixel belongs. The corner location information represents the location of the heart in the cardiac ultrasound image. The heart includes multiple tissues.

[0132] The conversion module 606 is used to convert the initial position information of each pixel in the semantic segmentation map into polar coordinate information based on the corner position information;

[0133] The determination module 608 is used to determine the section type of the cardiac ultrasound image based on the polar coordinate information of each pixel and the tissue to which each pixel belongs.

[0134] In one embodiment, the determination module 608 is used to input the semantic segmentation map after polar coordinate transformation to the feature extractor, and to extract features from the semantic segmentation map after polar coordinate transformation to obtain the target feature representation; input the target feature representation to the section classifier, and to determine the section type of the cardiac ultrasound image based on the target feature representation through the section classifier.

[0135] In one embodiment, the determining module 608 is used to segment the semantic segmentation map after polar coordinate transformation into multiple local images of the same size through a feature extractor; for each local image, the local image is converted into an initial feature representation of a preset dimension; based on the polar coordinate information of the pixels in the local image, the self-attention weight corresponding to the local image is determined, and the initial feature representation and the self-attention weight are weighted to obtain the target feature representation.

[0136] In one embodiment, the determining module 608 is used to flatten and straighten the target feature representation to obtain a one-dimensional feature representation; for each section type, obtain the weight value and bias value corresponding to the targeted section type; perform weighted processing on the one-dimensional feature representation and the weight value to obtain a linear combination feature representation; add the linear combination feature representation and the bias value to obtain the activation value corresponding to the targeted section type; obtain the section type probability distribution feature vector according to the activation value corresponding to each section type; and determine the section type of the cardiac ultrasound image according to the section type probability distribution feature vector.

[0137] In one embodiment, the determining module 608 is used to convert the activation value corresponding to each section type into the probability that the cardiac ultrasound image is classified into each section type; and to obtain the section type of the cardiac ultrasound image based on the probability that the cardiac ultrasound image is classified into each section type.

[0138] In one embodiment, the section classification device for cardiac ultrasound images further includes a quality module 610, which is used to input the target feature representation into a quality regressor, and generate a detection box in the cardiac ultrasound image based on the target feature representation through the quality regressor. The detection box is used to characterize the location of different tissue structures in the heart; obtain the section quality evaluation rules corresponding to the section type of the cardiac ultrasound image; and determine the section quality of the cardiac ultrasound image according to the section quality evaluation rules and the detection box generated in the cardiac ultrasound image.

[0139] In one embodiment, the section classification device for cardiac ultrasound images further includes an evaluation module 612 for inputting a target feature representation to a quality evaluator and generating defect language text based on the target feature representation through the quality evaluator. The defect language text is used to describe section defects in the cardiac ultrasound images.

[0140] In another embodiment, such as Figure 7 As shown, Figure 7The following is a structural block diagram of a section classification device for cardiac ultrasound images in another embodiment, including: an extraction module 602, an acquisition module 604, a conversion module 606, and a determination module 608. The section classification device 600 for cardiac ultrasound images further includes a quality module 610, used to input the target feature representation into a quality regressor, and through the quality regressor, generate detection boxes in the cardiac ultrasound image based on the target feature representation. The detection boxes are used to characterize the positions of different tissue structures in the heart. It also acquires section quality evaluation rules corresponding to the section type of the cardiac ultrasound image; and determines the section quality of the cardiac ultrasound image based on the section quality evaluation rules and the detection boxes generated in the cardiac ultrasound image. An evaluation module 612 is also included, used to input the target feature representation into a quality evaluator, and through the quality evaluator, generate defect language text based on the target feature representation. The defect language text is used to describe section defects in the cardiac ultrasound image.

[0141] The modules in the aforementioned cardiac ultrasound image section classification device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0142] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 8 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores data related to the section classification of echocardiogram images. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a section classification method for echocardiogram images.

[0143] Those skilled in the art will understand that Figure 8The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0144] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0145] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0146] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0147] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0148] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0149] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0150] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for cross-sectional classification of cardiac ultrasound images, characterized in that, The method includes: Feature extraction is performed on cardiac ultrasound images to obtain image features; Based on the image features, semantic segmentation and position detection are performed to obtain a semantic segmentation map and corner position information. The semantic segmentation map includes multiple pixels and the tissue to which each pixel belongs. The corner position information represents the position of the heart in the cardiac ultrasound image. The heart includes multiple tissues. The corner position information includes the position information of the vertices and two endpoints of the fan-shaped scanning area formed by the ultrasound probe in the cardiac ultrasound image. Based on the corner point position information, the initial position information of each pixel in the semantic segmentation map is converted into polar coordinate information; the polar coordinate information includes polar radius and polar angle, the polar radius represents the radial distance of the pixel relative to the vertex, and the polar angle represents the angle formed by the line segment connecting the vertex and the pixel from the reference direction; Based on the polar coordinate information of each pixel and the tissue to which each pixel belongs, the section type of the cardiac ultrasound image is determined.

2. The method according to claim 1, characterized in that, Determining the section type of the cardiac ultrasound image based on the polar coordinate information of each pixel and the tissue to which each pixel belongs includes: The semantic segmentation map after polar coordinate transformation is input into the feature extractor, and the feature extractor performs feature extraction on the semantic segmentation map after polar coordinate transformation to obtain the target feature representation; The target feature representation is input into a section classifier, and the section classifier determines the section type of the cardiac ultrasound image based on the target feature representation.

3. The method according to claim 2, characterized in that, The step of extracting features from the semantic segmentation map after polar coordinate transformation using the feature extractor to obtain the target feature representation includes: The feature extractor divides the semantic segmentation map after polar coordinate transformation into multiple local images of the same size. For each local image, the local image is converted into an initial feature representation of a preset dimension; Based on the polar coordinate information of pixels in the targeted local image, the self-attention weights corresponding to the targeted local image are determined, and the initial feature representation and the self-attention weights are weighted to obtain the target feature representation.

4. The method according to claim 2, characterized in that, The step of determining the section type of the cardiac ultrasound image based on the target feature representation using the section classifier includes: The target feature representation is flattened and straightened to obtain a one-dimensional feature representation. For each aspect type, obtain the weight and bias values ​​corresponding to that aspect type; The one-dimensional feature representation and the weight value are weighted to obtain a linear combination feature representation; The linear combination feature representation and the bias value are added together to obtain the activation value corresponding to the type of section being targeted. Based on the activation value corresponding to each facet type, the feature vector of the facet type probability distribution is obtained; The section type of the cardiac ultrasound image is determined based on the probability distribution feature vector of the section type.

5. The method according to claim 4, characterized in that, Determining the section type of the cardiac ultrasound image based on the section type probability distribution feature vector includes: The activation value corresponding to each slice type is converted into the probability that the cardiac ultrasound image is classified into each slice type. The section type of the cardiac ultrasound image is obtained based on the probability that the image is classified into each section type.

6. The method according to claim 2, characterized in that, The method further includes: The target feature representation is input into a quality regressor, and a detection box is generated in the cardiac ultrasound image based on the target feature representation. The detection box is used to characterize the location of different tissue structures in the heart. Obtain the section quality evaluation rules corresponding to the section type of the cardiac ultrasound image; The section quality of the cardiac ultrasound image is determined based on the section quality evaluation rules and the detection box generated in the cardiac ultrasound image.

7. The method according to claim 6, characterized in that, The method further includes: The target feature representation is input into a quality evaluator, and the quality evaluator generates defect language text based on the target feature representation. The defect language text is used to describe the cross-sectional defects of the cardiac ultrasound image.

8. A section classification device for cardiac ultrasound images, characterized in that, The device includes: The extraction module is used to extract features from cardiac ultrasound images to obtain image features; The module is used to perform semantic segmentation and location detection based on the image features to obtain a semantic segmentation map and corner position information. The semantic segmentation map includes multiple pixels and the tissue to which each pixel belongs. The corner position information represents the position of the heart in the cardiac ultrasound image. The heart includes multiple tissues. The corner position information includes the position information of the vertices and two endpoints of the fan-shaped scanning area formed by the ultrasound probe in the cardiac ultrasound image. A conversion module is used to convert the initial position information of each pixel in the semantic segmentation map into polar coordinate information based on the corner position information; the polar coordinate information includes polar radius and polar angle, the polar radius represents the radial distance of the pixel relative to the vertex, and the polar angle represents the angle formed by the line segment connecting the vertex and the pixel from the reference direction; The determination module is used to determine the section type of the cardiac ultrasound image based on the polar coordinate information of each pixel and the tissue to which each pixel belongs.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images