A cross-center ultrasound image segmentation method, system and storage medium

Abstract

The application discloses a cross-center ultrasonic image segmentation method and system and a storage medium, and belongs to the technical field of image segmentation. The method comprises the following steps: acquiring labeled source domain and unlabeled target domain ultrasonic images; shallow and deep features are extracted through a U-Net encoder, and are respectively input into a graph convolution network to obtain shallow and deep graph features, cross-center common features are obtained through attention fusion, and alignment of appearance and topological features is realized; the common features are input into a graph convolution encoder for graph self-representation learning, and combined with subspace contrast learning and reconstruction loss, the topological details of the anatomical structure are strengthened, and content-aware embedding representation is obtained; finally, the U-Net decoder is input, and the segmentation result is output by adopting Cauchy-Schwarz divergence clustering loss and target domain image reconstruction loss for constraint. The application significantly improves the generalization ability and boundary precision of cross-center ultrasonic image segmentation, and is especially suitable for precise segmentation of thin structures such as myocardium.

Description

Technical Field

[0001] This invention relates to a transcenter ultrasound image segmentation method, system, and storage medium, belonging to the field of image segmentation technology. Background Technology

[0002] Precise segmentation of cardiac structures such as the ventricles, atria, and myocardium is a core prerequisite for clinical analysis of ultrasound images, directly impacting the early diagnosis, disease grading, and treatment planning of cardiovascular diseases (such as cardiomyopathy and heart failure). However, ultrasound images collected in multicenter clinical studies exhibit significant heterogeneity: differences in ultrasound equipment models and scanning parameters (frequency, gain) across different centers lead to marked variations in image grayscale distribution, texture features, and contrast. Furthermore, the complex anatomy of the heart, blurred tissue boundaries, and the influence of respiratory movements and heart rate fluctuations further increase the difficulty of segmentation. Traditional single-center segmentation methods rely on the assumption of a uniform data distribution, exhibiting extremely poor generalization ability on multicenter data and failing to meet the needs of multicenter clinical studies.

[0003] To overcome the challenge of heterogeneity in cross-center data, researchers have proposed various cross-center ultrasound image segmentation methods, broadly categorized into traditional methods and deep learning methods. Traditional methods are further divided into feature-stitching, statistical modeling, and graphical modeling methods: Feature-stitching methods, such as wavelet transform-based fusion methods, stitch cross-center image features along dimensions and input them into the segmentation model, but they ignore the distribution differences of data from each center, cannot model the potential anatomical structural relationships between centers, and struggle to leverage the complementarity of cross-center data; Statistical modeling methods, such as Bayesian adaptive methods, construct joint latent variable models to characterize the distribution of cross-center data, but they can only capture data relationships globally and cannot reflect the subtle topological relationships of local anatomical structures (such as myocardial boundaries and atrial cavities); Graphical modeling methods, such as segmentation methods based on similarity network fusion, construct sample similarity networks for each center and obtain unified features through iterative fusion, but their accuracy is limited in complex cardiac structure segmentation due to the expressive power of manually designed features.

[0004] Despite the progress made by existing deep learning methods, significant limitations remain: First, they fail to fully exploit the specific features of ultrasound images from different centers, making it difficult to effectively adapt to differences in equipment and variations in scanning parameters. Second, there is insufficient synergistic optimization between feature representation and segmentation tasks, lacking specific constraints on the boundaries and topological relationships of cardiac anatomy, resulting in weak generalization ability for cross-center segmentation. Therefore, effectively exploring cross-center specific features and constructing feature representations suitable for segmentation tasks remain the core challenges for cross-center ultrasound cardiac structure segmentation. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a cross-center ultrasound image segmentation method, system and storage medium, which significantly improves the generalization ability and boundary accuracy of cross-center ultrasound image segmentation, and is especially suitable for the accurate segmentation of thin structures such as myocardium.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0007] In a first aspect, the present invention provides a method for cross-center ultrasound image segmentation, comprising:

[0008] A cross-center ultrasound image dataset is obtained, which includes source domain center data with segmentation labels and target domain center data without segmentation labels. Both the source domain center data and the target domain center data are ultrasound images containing cardiac anatomical structures, including the ventricles, atria, and myocardium. The ultrasound images containing the source domain center data and the target domain center data are respectively input into the U-Net encoder, and the shallow and deep features of the source domain center and the shallow and deep features of the target domain center are extracted accordingly. The shallow and deep features are used to characterize the texture morphology and spatial structure of the cardiac anatomy. The shallow and deep features of the source domain center and the shallow and deep features of the target domain center are respectively input into the graph convolutional network to obtain the shallow and deep graph features of the source domain center and the shallow and deep graph features of the target domain center. The graph features are used to model the anatomical topological relationships between cardiac anatomical structures. The shallow and deep graph features of the source domain center and the shallow and deep graph features of the target domain center are fused to obtain the multi-center common features of the source domain center and the target domain center. The common features are used to align the apparent features and topological features of the cardiac anatomy between different centers. The shared features of the source and target centers are input into a graph convolutional encoder to perform graph self-representation learning, thereby obtaining content-aware embedding representations of the source and target centers. These content-aware embedding representations are used to enhance the topological details of the cardiac anatomical structure. The content-aware embedding representations of the source domain center and the target domain center are input into the U-Net decoder, combined with the segmentation labels for clustering constraints, and the target domain image reconstruction loss constraints are used to output the final segmentation result, which is a pixel-level segmentation result of the cardiac anatomical structure.

[0009] Furthermore, the ultrasound image containing the source domain center data and the target domain center data is input into the U-Net encoder, and the shallow and deep features of the source domain center and the shallow and deep features of the target domain center are extracted accordingly, including: The first one containing cardiac anatomy The first source domain center ultrasound images and the first containing cardiac anatomy The first target domain center ultrasound images Input to the U-Net encoder, output by the shallow coding function of the U-Net encoder. The first source domain center shallow features and the The first target domain center shallow features The deep coding function output by the U-Net encoder is the first... The first source domain center Deep features and the The first target domain center Deep features The shallow features preserve the texture and grayscale distribution of the ventricles, atria, and myocardium, while the deep features preserve the higher-order semantics and spatial layout of the ventricles, atria, and myocardium. The formula is as follows: , ; , ; in, This is the shallow coding function of the U-Net encoder. For the deep encoding function and shallow features of the U-Net encoder All dimensions are Deep features All dimensions are , For image batch size, , These are the image height and width, respectively. This represents the number of shallow feature channels. This represents the number of deep feature channels.

[0010] Furthermore, the shallow and deep features of the source domain center and the shallow and deep features of the target domain center are respectively input into a graph convolutional network to obtain the shallow and deep graph features of the source domain center and the target domain center, respectively, including: shallow features of the source region center Shallow features of the target domain center Compressed to non-overlapping block aggregation After being flattened, the graph node features are input into the graph convolutional network, and then processed by the first... After layer graph convolution, the shallow node feature matrix at the center of the source domain is obtained. Shallow node feature matrix of the target domain center The final layer output is reshaped into a 4-dimensional shallow graph feature of the source domain center. Shallow graph features of the 4D target domain center The nodes in the graph correspond to local areas in the ultrasound image, and their adjacency relationships reflect the spatial proximity of the ventricular wall, atrial wall, and myocardial layer. Among them, the shallow node feature matrix , The calculation formula is as follows: ; ; In the formula: It is the ReLU activation function. This is the shallow normalized adjacency matrix of the source domain center. The shallow normalized adjacency matrix of the target domain center. The shallow features of the source domain center are input into the graph convolutional network and then... The output of the layer, The shallow features centered on the target domain are input into the graph convolutional network and then... The output of the layer, This represents the model parameter matrix for a shallow graph convolutional network. Deep features of the source domain center and deep features of the target domain center After being flattened, the graph node features are input into the graph convolutional network, and then processed by the... After layer graph convolution, the feature matrix of deep nodes at the center of the source domain is obtained. and the deep node feature matrix of the target domain center The final layer output is reshaped into a 4-dimensional deep graph feature at the center of the source domain. Deep graph features of the 4D target domain center The deep map features are used to encode the relative positions and topological relationships between the ventricles, atria, and myocardium; Among them, the deep node feature matrix , The calculation formula is as follows: ; ; in, The deep normalized adjacency matrix of the source domain center. The deep normalized adjacency matrix of the target domain center The deep features of the source domain center are input into the graph convolutional network and then... The output of the layer, The deep features centered on the target domain are input into the graph convolutional network. The output of the layer, This represents the model parameter matrix of a deep graph convolutional network.

[0011] Furthermore, the shallow and deep graph features of the source domain center and the shallow and deep graph features of the target domain center are fused to obtain the multi-center common features of the source domain center and the target domain center, including: Shallow graph features of the source region center and deep map features By splicing along the channel dimension, the shared features of multiple centers of the source domain center are formed. Attention weight matrix for source domain center is generated using a multilayer perceptron. The formula is as follows: ; in, For the Softmax function, The tanh activation function is used. , , , These are learnable parameters; Shallow map features at the center of the target domain and deep map features By splicing along the channel dimension, the shared features of multiple centers in the target domain are formed. An attention weight matrix for the target domain center is generated using a multilayer perceptron. The formula is as follows: ; The average common features are obtained by weighted summation and 3×3 convolution. The formula is as follows: ; in, The total number of source domain centers. It is a 3×3 convolution function; Introducing inter-center alignment loss The difference between shared features and average shared features across multiple centers is constrained. The center alignment loss promotes the consistency of ventricular, atrioventricular, and myocardial features extracted from different centers in the common space, as shown in the following formula: ; in, It is an L2 norm; and This is an adjustable parameter.

[0012] Furthermore, the shared features of the source domain centers and target domain centers are input into a graph convolutional encoder for graph self-representation learning to obtain content-aware embedding representations of the source domain centers and target domain centers, including: The shared features of multiple centers of source domain centers and target domain centers , Input graph convolution encoder The initial content-aware embedding representations of the source domain centers are obtained respectively. Initial content-aware embedding representation of the target domain center The content-aware embedding representation focuses on discriminative features of the ventricle, atrium, and myocardium, as shown in the following formula: ; ; in, An adjacency matrix with shared features among multiple centers. For GCN encoding functions; Introducing a subspace contrastive learning strategy, and Features before self-representation learning, used as source domain centers and target domain centers respectively; and features after self-representation learning, combined with those of source domain centers and target domain centers. , Construct contrast loss And integrate to obtain self-representation loss To optimize content-aware embedding representations, the self-representation learning method mines the intrinsic correlations between the ventricles, atria, myocardium, and their inter-ventricles through the self-representation coefficient matrix. Contrast loss is used to bridge different representations of the same anatomical structure. Simultaneously, a graph convolutional decoder reconstructs multi-center shared features. Introducing reconstruction loss Finally, the optimized content-aware embedding representation of the source domain center is obtained. Content-aware embedding representation of the target domain center ; Wherein, the contrast loss The formula is as follows: ; in, The cosine similarity function is used. , , , Indicates the first Features before learning the self-representation of the source domain center Indicates the first Features learned from the self-representation of each source domain center It is a self-representation coefficient matrix; The self-representation loss The formula is as follows: ; in, A function to find the diagonal matrix; Indicates the number of source domain centers; The reconstruction loss The formula is as follows: ; in, The reconstructed features of the source domain center are output by the graph convolutional decoder. The reconstructed features of the target domain center output by the graph convolutional decoder.

[0013] Furthermore, the content-aware embedding representations of the source domain centers and target domain centers are input into the U-Net decoder, clustering constraints are applied in conjunction with segmentation labels, and target domain image reconstruction loss constraints are used to output the final segmentation result, including: The optimized source domain center content-aware embedding representation Inputting a 3×3 convolutional layer and a softmax function yields the prediction probability matrix. ,in It is a 3×3 convolutional layer, and the prediction probability matrix reflects the probability that each pixel belongs to the ventricle, atrium, myocardium or background. Will After upsampling to the original image size, the first cross-entropy loss is used. Calculate pixel-by-pixel classification accuracy: ; in, The number of pixels in a source domain image. For the prediction probability matrix Upsample to the original image size to predict the probability matrix. For pixels The true label; Introducing Cauchy-Schwarz divergence As a clustering loss, the clustering loss enhances class separability by maximizing the difference between the probability distributions of each category of ventricle, atrium, and myocardium, as shown in the following formula: ; in, The number of segmentation categories is ventricles, atria, and myocardium. The affinity matrix is ​​calculated as follows: , For the prediction probability matrix The One channel; For the prediction probability matrix The One channel; It is the transpose symbol; The optimized source domain center content-aware embedding representation Input the U-Net decoder to obtain the predicted segmentation result of the source domain center. and using the second cross-entropy loss Calculate pixel-by-pixel classification accuracy: ; in, express The first in 1 pixel; Content-aware embedding representation of the target domain center Input to U-Net decoder to reconstruct the target domain image Introducing reconstruction loss The formula is as follows: ; Construct the total loss function The formula is used for joint optimization training of the U-Net encoder and U-Net decoder, as follows: ; in For loss weights; After training, the content-aware embedding representation of the target domain center is used. Input the trained U-Net decoder to obtain the predicted segmentation result of the target domain center, and output it as the final segmentation result.

[0014] In a second aspect, the present invention provides a transcenter ultrasound image segmentation system for implementing the transcenter ultrasound image segmentation method described in any of the preceding claims, the system comprising: The data acquisition module is used to acquire a cross-center ultrasound image dataset, which includes source domain center data with segmentation labels and target domain center data without segmentation labels. Both the source domain center data and the target domain center data are ultrasound images containing cardiac anatomical structures, including the ventricles, atria, and myocardium. The feature extraction module is used to input the ultrasound image containing the source domain center data and the target domain center data into the U-Net encoder respectively, and extract the shallow and deep features of the source domain center and the shallow and deep features of the target domain center. The shallow and deep features are used to characterize the texture morphology and spatial structure of the cardiac anatomy. A two-layer feature alignment module is used to input the shallow and deep features of the source domain center and the shallow and deep features of the target domain center into a graph convolutional network, respectively, to obtain the shallow and deep graph features of the source domain center and the shallow and deep graph features of the target domain center. The graph features are used to model the anatomical topological relationships between cardiac anatomical structures. The feature fusion module is used to fuse the shallow and deep graph features of the source domain center and the shallow and deep graph features of the target domain center to obtain the multi-center common features of the source domain center and the target domain center. The common features are used to align the apparent features and topological features of the cardiac anatomy structure between different centers. The graph self-representation learning module is used to input the multi-center common features of the source domain center and the target domain center into the graph convolutional encoder to perform graph self-representation learning, and obtain the content-aware embedding representation of the source domain center and the target domain center. The content-aware embedding representation is used to enhance the topological details of the cardiac anatomical structure. The clustering alignment module is used to input the content-aware embedding representations of the source domain center and the target domain center into the U-Net decoder, combine the segmentation labels for clustering constraints, and use the target domain image reconstruction loss constraints to output the final segmentation result, which is a pixel-level segmentation result of the cardiac anatomical structure.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0016] Fourthly, the present invention provides an electronic device, comprising:

[0017] Memory, used to store computer programs / instructions;

[0018] A processor for executing the computer program / instructions to implement the steps of any of the methods described above.

[0019] Fifthly, the present invention provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0020] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0021] This invention provides a cross-center ultrasound image segmentation method, system, and storage medium. Ultrasound images containing source domain center data and target domain center data are input into a U-Net encoder to extract shallow and deep features, respectively. These features are then input into a graph convolutional network to obtain shallow and deep graph features, which are then fused to obtain multi-center shared features. This achieves synergistic alignment between apparent features and anatomical topological features, effectively offsetting the impact of cross-center equipment differences and scanning parameter heterogeneity. By inputting the multi-center shared features into the graph convolutional encoder for graph self-representation learning, content-aware embedding representations are obtained, enhancing the topological detail modeling of local anatomical structures and making feature representations more consistent with clinical anatomy. The content-aware embedding representations are input into the U-Net decoder, combined with segmentation labels for clustering constraints, and target domain image reconstruction loss constraints are used to accurately quantify the consistency between features and segmentation labels while preserving target domain feature details. Finally, accurate segmentation results are output, significantly improving the generalization ability, boundary accuracy, and adaptability to complex anatomical structures in cross-center ultrasound image segmentation. Attached Figure Description

[0022] Figure 1 This is a schematic flowchart of the cross-center ultrasound image segmentation method provided in an embodiment of the present invention;

[0023] Figure 2 This is a flowchart illustrating the two-layer feature alignment module provided in an embodiment of the present invention;

[0024] Figure 3 This is a flowchart illustrating the graph self-representation learning module provided in an embodiment of the present invention;

[0025] Figure 4 This is a flowchart illustrating the clustering alignment module provided in an embodiment of the present invention;

[0026] Figure 5 This is a visual comparison chart provided in the embodiments of the present invention. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0028] Example 1: This example introduces a transcenter ultrasound image segmentation method, including:

[0029] A cross-center ultrasound image dataset is obtained, which includes source domain center data with segmentation labels and target domain center data without segmentation labels. Both the source domain center data and the target domain center data are ultrasound images containing cardiac anatomical structures, including the ventricles, atria, and myocardium. The ultrasound images containing the source domain center data and the target domain center data are respectively input into the U-Net encoder, and the shallow and deep features of the source domain center and the shallow and deep features of the target domain center are extracted accordingly. The shallow and deep features are used to characterize the texture morphology and spatial structure of the cardiac anatomy. The shallow and deep features of the source domain center and the shallow and deep features of the target domain center are respectively input into the graph convolutional network to obtain the shallow and deep graph features of the source domain center and the shallow and deep graph features of the target domain center. The graph features are used to model the anatomical topological relationships between cardiac anatomical structures. The shallow and deep graph features of the source domain center and the shallow and deep graph features of the target domain center are fused to obtain the multi-center common features of the source domain center and the target domain center. The common features are used to align the apparent features and topological features of the cardiac anatomy between different centers. The shared features of the source and target centers are input into a graph convolutional encoder to perform graph self-representation learning, thereby obtaining content-aware embedding representations of the source and target centers. These content-aware embedding representations are used to enhance the topological details of the cardiac anatomical structure. The content-aware embedding representations of the source domain center and the target domain center are input into the U-Net decoder, combined with the segmentation labels for clustering constraints, and the target domain image reconstruction loss constraints are used to output the final segmentation result, which is a pixel-level segmentation result of the cardiac anatomical structure.

[0030] The cross-center ultrasound image segmentation method provided in this embodiment involves the following steps in its application:

[0031] In the domain adaptive segmentation task, including A dataset with segmentation labels for 1 clinical center is denoted as . A central data point without a delimiter label is denoted as the target domain. , The number of ultrasound image samples in the target domain. Represents the first of the target domains One ultrasound image sample; In each of the clinical centers, each sample in the data from each center has a segmentation label denoted as the source domain. The image dataset with one center is defined as ,in Indicates the first The first center One ultrasound image sample, For the first Number of ultrasound image samples per center express The segmentation labels are used to define each ultrasound image sample as an ultrasound image containing cardiac anatomical structures such as the ventricles, atria, and myocardium.

[0032] like Figure 1 As shown, the content-aware cross-center ultrasound image segmentation method and system includes a two-layer feature alignment module, a graph self-representation learning module, and a clustering alignment module. In the two-layer feature alignment module, the data containing the source domain center data is... The first source domain center ultrasound images The result is obtained by feeding the data into a convolutional neural network based on a U-Net encoder. The first source domain center shallow features , No. The first source domain center Deep features shallow features Deep features The shallow graph features of the source region center are obtained by feeding them into a graph convolutional neural network (GCN). Deep graph features of the source domain center The first one containing the target domain center data. The first target domain center ultrasound images The result is obtained by feeding the data into a convolutional neural network based on a U-Net encoder. The first target domain center shallow features , No. The first target domain center Deep features The shallow features preserve the texture and grayscale distribution of the ventricles, atria, and myocardium, while the deep features preserve the higher-order semantics and spatial layout of the ventricles, atria, and myocardium. Deep features The data are fed into a graph convolutional neural network (GCN) to obtain shallow graph features corresponding to the center of the target domain. Deep graph features of the target domain center The deep graph features are used to encode the relative positions and topological relationships between the ventricles, atria, and myocardium. Multi-center shared features of the source domain centers are obtained through weighted summation and fusion. Multi-center shared features of the target domain center An inter-center alignment loss is introduced to constrain the differences between the bilevel representations of each center and the shared features, thereby improving cross-center generalization ability. This inter-center alignment loss promotes consistency among the ventricular, atrioventricular, and myocardial features extracted from different centers in the common space. In the graph self-representation learning module, GCN is used to handle shared features from multiple centers. , Further processing enhances the anatomical structure association modeling by introducing a subspace contrastive learning strategy to strengthen anatomical structure similarity modeling. In the clustering alignment module, Cauchy-Schwarz (CS) divergence is introduced as the clustering loss to quantify the matching quality between content-aware features and segmentation labels. Additionally, to prevent feature loss in the target domain, the content-aware embedding representation of the target domain center is input into another U-Net decoder to obtain the reconstructed target domain image. .

[0033] The heterogeneity of cross-center ultrasound images is mainly reflected in apparent features such as texture and grayscale distribution, while cardiac anatomical structures (ventricular morphology, atrial position, myocardial thickness) have shared topological characteristics across centers. A single encoder struggles to simultaneously capture both apparent specificity and structural commonality: Convolutional Neural Networks (CNNs) excel at extracting high-dimensional image data and learning intrinsic apparent features, but struggle to capture anatomical relationships between samples; Generative Neural Networks (GCNs) effectively capture anatomical dependencies and extract content-aware embeddings by constructing graph structures between samples, but are less efficient in directly processing high-dimensional images and learning apparent features. Based on the complementary nature of these two approaches, such as... Figure 2 As shown, the dual-layer feature alignment module utilizes a Convolutional Neural Network (CNN) and a Graph Convolutional Neural Network (GCN) to model each center dataset, preserving apparent features while mining latent anatomical structures, achieving collaborative learning of center-specific representations. The CNN extracts shallow and deep features from the input image. The shallow features preserve the texture and grayscale distribution of the ventricles, atria, and myocardium, while the deep features preserve the higher-order semantics and spatial layout of the ventricles, atria, and myocardium. The formula is as follows:

[0034] , ;

[0035] , ;

[0036] in, This is the shallow coding function of the U-Net encoder, with the input being the first... Ultrasound images of each center Output For the first The first source domain center A shallow feature, with dimension 1 , For image batch size, , These are the height and width of the image, respectively. The number of channels represents the feature, preserving the texture and content features at each center. This is the deep coding function of the U-Net encoder, outputting... For the first The first source domain center Each deep feature has a dimension of [number]. , The number of channels represents the feature, preserving the content features of each center. Similarly, , Representing the target domain image samples respectively Shallow features and deep features.

[0037] pass Non-overlapping block aggregation: This embodiment uses mean pooling to aggregate shallow features. , from Compress to Simultaneously, each block is mapped to a node in the graph; neighborhood node features are fused using a normalized adjacency matrix, and the number of channels is increased from [previous value] to [previous value] using a weight matrix. arrive Transformation:

[0038] ;

[0039] ;

[0040] in, The shallow features of the source domain center are input into the graph convolutional network and then... The output of the layer, Shallow feature input map of the target domain center, convolutional network, and then... The output of the layer, Using the ReLU activation function, This is the shallow normalized adjacency matrix of the source domain center. The shallow normalized adjacency matrix of the target domain center is calculated as follows: , It is the identity matrix. for The degree matrix; The shallow adjacency matrix of the source domain center. The shallow zone self-loop adjacency matrix is ​​the center of the source domain. This represents the model parameter matrix for a shallow graph convolutional network. , for degree matrix, The shallow adjacency matrix of the target domain center. The shallow band self-loop adjacency matrix is ​​the center of the target domain. Indicates the number of GCN layers. The shallow features of the source domain center are input into the graph convolutional network and then... Layer output The shallow features centered on the target domain are input into the graph convolutional network and then... The output of the layer, for Compress to The feature vector after flattening for Compress to The flattened vector of features. The node feature matrix output from the last layer of the GCN is reshaped into a shallow graph feature matrix of the target's 4-dimensional source region center. Shallow graph features of the 4D target domain center , shape This completes the conversion from a "graph" to a "feature map".

[0041] Deep features of the source domain center and deep features of the target domain center The purpose of feeding the data into GCN is to supplement shallow content modeling capabilities and enhance the content awareness of the representation.

[0042] ;

[0043] ;

[0044] in The deep feature input map of the source domain center is convolved by the network and then... Layer output, The deep feature input map of the target domain center is convolutionally processed by the network. Layer output, The deep normalized adjacency matrix of the source domain center is calculated as follows: , , for degree matrix, The deep adjacency matrix of the source domain center. The deep zone self-loop adjacency matrix is ​​the center of the source domain. Indicates the number of GCN layers. The deep features of the source domain center are input into the graph convolutional network and then... Layer output, for The flattened vector, This represents the model parameter matrix of a deep graph convolutional network. The normalized adjacency matrix is ​​calculated as follows: , , for degree matrix, The deep adjacency matrix of the target domain center. The deep band self-loop adjacency matrix is ​​the center of the target domain; The deep features centered on the target domain are input into the graph convolutional network. Layer output, for The flattened vector. The node feature matrix output from the last layer of the GCN is reshaped into a 4-dimensional deep graph feature matrix centered on the source region. Deep graph features of the 4D target domain center Its shape is The deep map features are used to encode the relative positions and topological relationships between the ventricles, atria, and myocardium.

[0045] Shallow graph features of the source region center and deep map features By splicing along the channel dimension, the shared features of multiple centers of the source domain center are formed. , Representing a real-valued matrix, an attention weight matrix for the source domain center is constructed using a multilayer perceptron (MLP). The calculation method is as follows Shallow map features centered on the target domain. and deep map features By splicing along the channel dimension, the shared features of multiple centers in the target domain are formed. , Representing a real-valued matrix, an attention weight matrix for the target domain center is constructed using a multilayer perceptron (MLP). The calculation method is as follows .in, For the Softmax function, The tanh activation function is used. , , , For learnable parameters, Dimensions and The number of channels is the same. Dimensions and The number of channels is consistent, which is used to dynamically allocate weights to each channel.

[0046] Subsequently, the average shared features are obtained by weighted summation and fusion: , in for Convolution function, used to convert the number of channels become , This represents the average shared characteristics after fusion.

[0047] To enhance feature consistency among centers and improve cross-center generalization ability, an inter-center alignment loss is introduced. The difference between shared features and average shared features of multiple centers is constrained. The center alignment loss promotes the consistency of ventricular, atrioventricular, and myocardial features extracted from different centers in the common space, as shown in the following formula:

[0048] ;

[0049] in, It is an L2 norm; and Adjustable parameters to balance alignment loss between deep and shallow layers.

[0050] Multi-center shared features of the merged source domain centers Multicenter shared features with the target domain center Although it integrates deep and shallow structural information, it still has shortcomings in anatomical topological detail modeling. Deep modeling of common features, fully exploring the topological structure between pixels, improves feature quality. For example... Figure 3 As shown, the graph self-representation learning module utilizes GCN to learn shared features from multiple centers. , Further processing was performed to enhance the modeling of anatomical structures:

[0051] ; ;

[0052] in, For GCN encoding functions, An adjacency matrix with shared features among multiple centers. The initial content-aware embedding representation centered on the source domain further enhances the topological consistency of the features. The initial content-aware embedding representation is centered on the target domain, focusing on discriminative features of the ventricle, atrium, and myocardium. A subspace contrastive learning strategy is introduced: the features of each sample before learning from its self-representation are used. Features learned from self-representation ,Will Features of this sample are considered positive sample pairs, while features of the sample and other samples are considered negative sample pairs. Anatomical structure similarity modeling is enhanced through contrastive loss. The contrast loss for each sample is:

[0053] ;

[0054] in The cosine similarity function is used. , Indicates the first Features learned from the self-representation of the source domain center Indicates the first Features before learning the self-representation of the source domain center , The features after learning the self-representation of the target domain. Let be the self-representation coefficient matrix. After integrating the contrastive loss, the self-representation loss is:

[0055] ;

[0056] in, It is the L2 norm. Represents the number of samples in a batch, with constraints. To avoid sample self-representation and improve the generalization ability of the representation, the self-representation learning uses the self-representation coefficient matrix to mine the intrinsic correlations between the ventricles, atria, myocardium and between them, and contrast loss is used to bring together different representations of the same anatomical structure.

[0057] To avoid self-representation learning from destroying the anatomical structure of features, a GCN decoder is introduced to reconstruct shared features. , This forms a closed-loop optimization:

[0058] ;

[0059] in, The reconstructed features of the source domain center are output by the graph convolutional decoder. The reconstructed features of the target domain center are output by the graph convolutional decoder. , It is used to constrain the GCN encoding process to align with the global target, repair structural information that may be damaged in self-representation learning, and enhance feature robustness.

[0060] Obtain the optimized content-aware embedding representation of the source domain center Content-aware embedding representation of the target domain center Subsequently, it is necessary to further constrain its adaptability to the segmentation task, ensuring that the features can accurately distinguish the boundaries of the ventricles, atria, and myocardium, and meet the segmentation objective of "similar features of similar structures and separation of features of dissimilar structures." For example... Figure 4 As shown, the clustering alignment module introduces Cauchy-Schwarz (CS) divergence as the clustering loss to quantify the matching quality between features and segmentation labels, guiding features to optimize in the direction of suitable segmentation. The clustering loss enhances class separability by maximizing the difference between the probability distributions of each category (ventricle, atrium, myocardium), as shown in the following formula:

[0061] ;

[0062] in The number of segmentation categories is [number], which includes ventricles, atria, and myocardium. For the prediction probability matrix The aisle, For the prediction probability matrix The aisle, Indicates transpose. , for Convolutional layers (will) Mapped to Dimensional segmentation probability); The affinity matrix is ​​calculated as follows: The prediction probability matrix reflects the anatomical similarity between pixels, indicating the probability that each pixel belongs to the ventricle, atrium, myocardium, or background. Upsampled to the original image size Using the first cross-entropy loss Calculate pixel-by-pixel classification accuracy:

[0063] ;

[0064] in, The number of pixels in a source domain image. For the prediction probability matrix Upsample to the original image size to predict the probability matrix. and Pixels The predicted probability and the true label.

[0065] The optimized source domain center content-aware embedding representation Input the U-Net segmentation decoder to obtain the predicted segmentation result. Using the second cross-entropy loss Calculate pixel-by-pixel classification accuracy:

[0066] ;

[0067] in, The predicted segmentation result for the source domain center. express The first in Each pixel. Furthermore, to prevent feature loss in the target domain, a content-aware embedding representation of the target domain center is used. Input another U-Net decoder to obtain the reconstructed target domain image Calculate the reconstruction loss:

[0068] ;

[0069] During training, affinity matrix Prediction probability matrix Form an optimization loop with the feature encoder: Structural optimization promotes Update to more accurate segmentation labels. The gradient generated by CS divergence guides the encoder to generate more discriminative features, thereby optimizing the encoder. Ultimately, this makes the feature boundaries clearer and the topology more reasonable.

[0070] Construct the total loss function The formula is used for joint optimization training of the U-Net encoder and U-Net decoder, as follows:

[0071] ;

[0072] in To balance the loss contribution of each module, a loss weight is assigned.

[0073] After training, the content-aware embedding representation of the target domain center is used. Input the trained U-Net decoder to obtain the predicted segmentation result of the target domain center, and output it as the final segmentation result.

[0074] The advantages of this embodiment are as follows:

[0075] 1. The dual-layer feature alignment module in this embodiment breaks through the limitations of traditional single-dominant feature extraction. It proposes a dual-branch collaborative architecture of Convolutional Neural Network (CNN) and Graph Convolutional Network (GCN) to achieve synchronous alignment of appearance features and topological features. Unlike methods that rely solely on CNN to capture pixel-level appearance information, the dual-layer feature alignment module fuses the deep features extracted by CNN with the anatomical topological features modeled by GCN, constructing a two-dimensional feature alignment mechanism. This mechanism specifically addresses the heterogeneity of equipment and scanning parameters in cross-center medical images, adapting to the segmentation of complex anatomical structures. Especially for thin tissues such as myocardium, it can improve the consistency of boundary segmentation through topological feature constraints, providing high-quality feature input for subsequent modules.

[0076] 2. The graph self-representation learning module in this embodiment introduces a graph self-representation mechanism based on subspace reconstruction, deeply combining feature learning with detail enhancement. Unlike traditional graph models that rely solely on preset topological relationships, the graph self-representation learning module combines feature self-representation reconstruction with subspace contrast loss to enhance local details of anatomical structures such as myocardial margins and atrial spaces, achieving adaptive feature optimization. This provides features rich in topological details for the subsequent clustering alignment module, reducing the interference of invalid features on the alignment effect, and enabling the model to maintain stable performance even on samples with motion artifacts and uneven grayscale.

[0077] 3. The clustering alignment module in this embodiment optimizes the feature-label matching and cross-domain feature preservation mechanism by introducing a dual-constraint clustering alignment strategy of Cauchy-Schwarz (CS) divergence supervision and target domain reconstruction. Compared with the traditional supervision method using cross-entropy loss, the clustering alignment module introduces CS divergence loss to accurately quantify the consistency between feature distribution and labels. At the same time, it uses target domain image reconstruction loss, which solves the problems of easy loss of target domain feature information and inaccurate supervision signal during cross-domain alignment, maintains the stability of feature distribution, and avoids distortion of anatomical structural features during alignment.

[0078] The following example illustrates the content involved in the above embodiments.

[0079] The following experiment was used in this embodiment:

[0080] (I) Experimental setup:

[0081] All methods are implemented using Python 3.8.19 and PyTorch 1.12.1. The segmentation head adopts the U-Net architecture, with a ResNet-18 encoder and a transposed convolutional decoder. Hyperparameter settings are as follows: the learning rate for both Model 1 and Model 2 is set to... The Adam optimizer was used; the training epochs were 300 each, and an early stopping strategy was employed to mitigate overfitting. The early stopping strategy had a preset tolerance epoch limit of 50. The CNN contained two convolutional layers with output channels of 64 and 128, and the GCN was a 2-layer structure. The loss weights... The values ​​were set to 0.2, 0.3, 0.1, 0.2, 0.2, and 0.5 to balance the contributions of each module. The experimental hardware was an NVIDIA RTX 3090 GPU (24GB VRAM).

[0082] The evaluation metrics for this embodiment are as follows:

[0083] The performance was evaluated using commonly used metrics in the field of medical image segmentation, including: Dice similarity coefficient (Dice), intersection-over-union ratio (IoU), and Hausdorff distance (HD95). Dice and IoU measure the overlap of segmented regions (values ​​range from 0 to 1, with values ​​closer to 1 being better); HD95 measures the deviation of the segmentation boundary (unit: pixels, with lower values ​​being better). Simultaneously, the model's adaptive capability was evaluated through cross-center generalization experiments: in the training set, one center had a segmentation label and one center did not; the mean of the metrics on the test set was calculated to assess cross-center performance.

[0084] The dataset for this embodiment is as follows:

[0085] All methods were validated on two central echocardiography datasets, including two mainstream ultrasound devices: GE and Siemens. Central heterogeneity was reflected in differences in device models, scanning frequencies, and gain parameters. In this embodiment, the scanning frequency ranged from 2 to 10 MHz. Each dataset contained complete annotations of the ventricles, atria, and myocardium, with a uniform image size of 512×512. Both central datasets were annotated by an ultrasound physician. All datasets included three scanning modes: B-mode ultrasound, M-mode ultrasound, and color Doppler, with sample sizes of 488 and 510 patients, respectively.

[0086] The comparative experiment in this embodiment is as follows:

[0087] Comparison method:

[0088] 1. Basic deep learning method: The standard U-Net is used as the baseline model, trained only with source domain data and without domain adaptation strategy;

[0089] 2. Domain Adaptive Adversarial Alignment Method: U-Net + DomainAdain (domain adaptive adversarial training) is used as the representative adversarial alignment strategy;

[0090] 3. Domain Adaptive Attention Alignment Method: DANet (Domain Adaptive Attention Mechanism) is used as the representative attention alignment strategy;

[0091] 4. Federated learning method: FedAvg+U-Net, which achieves cross-center data privacy protection training through parameter aggregation, and is adapted to clinical cross-center scenarios.

[0092] Table 1. Quantitative results of cross-center splitting for each method

[0093]

[0094] As shown in Table 1, the quantitative results demonstrate that our proposed method achieves the best performance across all core metrics, with overall performance significantly superior to all comparable methods. Specifically, the average Dice coefficient reaches 85.8%, a significant improvement of 25.8% compared to the baseline model U-Net (60.0%), 4.1% compared to the domain adaptation method DANet (81.7%), and 4.8% compared to the federated learning method FedAvg+U-Net (81.0%). The average IoU reaches 77.2%, an improvement of 29.0% compared to U-Net and 4.9% compared to DANet, exhibiting a highly significant advantage in overlap reduction. The average HD95 is reduced to 4.1 pixels, a reduction of 8.5 pixels (67.5%) compared to U-Net and 1.7 pixels (29.0%) compared to DANet, achieving a qualitative improvement in boundary segmentation consistency. This fully validates the superior performance of our proposed method in cross-center scenarios.

[0095] The performance differences between different methods are significant, clearly reflecting the impact of technical approaches on cross-center generalization ability: Standard U-Net, as the baseline model, relies solely on source domain data for training and lacks any domain adaptation strategy. It cannot withstand the heterogeneity of devices and scanning parameters in cross-center data, resulting in an average Dice of only 60.0% and an HD95 as high as 12.6 pixels. Its segmentation accuracy and boundary consistency are both poor, consistent with the normal performance level of baseline models in cross-center scenarios. U-Net+DomainAdain and DANet, by introducing domain alignment strategies, effectively reduce the difference in feature distribution between the source and target domains, increasing the average Dice to 80.2% and 81.7% respectively, and significantly reducing HD95 to 6.3 pixels and 5.8 pixels. ANet's attention alignment mechanism can more accurately focus on key anatomical features, outperforming adversarial training strategies and validating the effectiveness of attention-guided domain alignment. FedAvg+U-Net, through federated parameter aggregation, suits cross-center privacy requirements, achieving an average Dice of 81.0%, slightly lower than DANet (0.7%). This is mainly due to a small amount of feature information loss during parameter aggregation, reflecting a trade-off between privacy protection and performance. This method, through CNN and GCN dual-layer feature fusion and content-aware modeling, balances appearance feature extraction and topological association capture, comprehensively surpassing existing methods, especially in boundary segmentation accuracy and adaptability to complex structures. It solves the problem that existing methods struggle to balance appearance heterogeneity and topological integrity.

[0096] All methods exhibited a pattern of "highest accuracy in ventricular segmentation and lowest accuracy in myocardial segmentation," reflecting the impact of anatomical complexity on segmentation performance. This pattern was even more pronounced on the benchmark model U-Net. Ventricular segments, with their large volume and clear boundaries, generally achieved relatively high segmentation accuracy, with our method reaching 89.5%, a 24.2% improvement over U-Net (65.3%) and a 3.2% improvement over DANet (86.3%). Atrial segmentation was slightly more affected by artifacts, but our method achieved 84.3%, a 25.6% improvement over U-Net (58.7%) and a 4.1% improvement over DANet (80.2%). Myocardial segmentation was the most challenging, with thin layers and blurred boundaries. The average Dice of all compared methods was below 79%, with U-Net at only 56.0%. Our method achieved 83.6%, a 27.6% improvement over U-Net, a 5.1% improvement over DANet, and a 5.7% improvement over FedAvg+U-Net, demonstrating the most significant advantage. This result demonstrates that the content-aware module of this method can effectively capture the topological relationships of thin tissues such as myocardium, mitigating the effects of boundary blurring, artifact interference, and transcentric heterogeneity, and is more adaptable to the segmentation of complex anatomical structures.

[0097] like Figure 5The image shows a visual comparison, where (a) is the source domain ultrasound image, (b) is the segmentation label of the source domain ultrasound image, (c) is the target domain ultrasound image, (d) is the segmentation label of the target domain ultrasound image, (e) is the segmentation image of this method, (f) is the segmentation image of FedAvg+U-Net, (g) is the segmentation image of DANet, and (h) is the segmentation image of U-Net. A direct comparison of (a) and (c) reveals significant heterogeneity between the two datasets: the source domain image (a) shows high brightness in the ventricular region and strong contrast with surrounding tissues, with continuous and discernible myocardial edges; the target domain image (c) has a generally low grayscale value, artifacts appear at the junction of the atrium and myocardium, and some details are masked by noise. This heterogeneity directly leads to a decrease in the segmentation performance of traditional methods, providing real-world clinical support for testing the generalization ability of each method. Figure 5 As shown in (e), the segmentation results of this method highly match the true labels of the target domain (d). The ventricular region is complete without any missing segments, the atrial boundaries accurately conform to the metal standard, and the segmentation of thin tissues such as myocardium is continuous without any breaks, with no obvious over-segmentation or under-segmentation. This is due to the fact that the two-layer feature alignment module offsets the influence of device heterogeneity, the graph self-representation learning module strengthens the association of anatomical structures, and the clustering alignment module ensures accurate matching between features and labels. Figure 5 As shown in (f), FedAvg+U-Net can roughly segment the core ventricular structure, but there are significant deviations in the segmentation of the atria and myocardium. Fragmented segmentation occurs in the myocardial region, and the atrial boundary is offset from the true label (d) by approximately 2-3 pixels. This is mainly due to the loss of some local structural features during the federated parameter aggregation process, resulting in limited adaptability to cross-center heterogeneity. Figure 5 As shown in (g), DANet performs better than FedAvg+U-Net and U-Net in segmentation, but still has some detail deficiencies. Slight oversegmentation occurs at the edges of the atrial region, and the distinction between the myocardium and surrounding tissues is insufficient, especially in the artifact region of the target domain image (c), where the segmentation boundaries are blurred, reflecting that its attention alignment mechanism fails to fully capture the subtle topological relationships of the anatomical structures. Figure 5 As shown in (h), U-Net has the worst segmentation performance and cannot adapt to cross-center heterogeneity. The ventricular region shows some incorrect segmentation, significantly differing from the true label (d). Due to the lack of domain adaptation and structure awareness mechanisms, models trained solely on source domain data cannot cope with differences in device and parameter settings in the target domain, resulting in weak generalization ability.

[0098] The ablation experiment in this embodiment is as follows:

[0099] To accurately verify the effectiveness and contribution of the core modules of this method, seven ablation variants (V1-V7) were constructed based on this method, and experiments were designed using the controlled variable method of "removing / replacing core components". All ablation experiments shared the same dataset, hyperparameter configuration, and hardware environment as the main experiment to ensure consistency of experimental conditions.

[0100] Table 2 Module Design Ablation Experiment Description

[0101]

[0102] Table 3 Quantitative results of ablation experiments

[0103]

[0104] Based on the performance degradation (average Dice difference between the complete model and the ablation variant) in Table 3, the contribution of each core mechanism can be quantified, ranked as follows: two-layer feature alignment module (4.6%) > graph self-representation learning module (2.8%) > contrastive learning constraint (2.3%) > CS divergence loss (1.8%) > attention-weighted fusion (1.5%) > target domain reconstruction loss (1.3%). Among them, the two-layer feature alignment module has the highest contribution; its removal caused the average Dice to drop from 85.8% to 81.2%, proving that the two-layer feature alignment module is the core key to achieving cross-center feature alignment and improving segmentation performance.

[0105] The two-layer feature alignment module (V1) showed a 4.6% decrease in average Dice and a 5.3% decrease in IoU after removal, while HD95 increased from 4.1 pixels to 5.6 pixels (a 36.6% increase). This indicates that the module's collaborative design, which uses CNN to capture appearance features and GCN to model related features, effectively aligns cross-center feature distributions, significantly improving the model's generalization ability and serving as the core support for performance improvement.

[0106] GCN structural modeling (V2): After removing GCN, the average Dice decreased by 3.8%, and HD95 increased to 5.3 pixels (an increase of 29.3%), especially with a 5.2% decrease in Dice for myocardial segmentation. This demonstrates that GCN can accurately capture the topological relationships of anatomical structures and significantly improve boundary segmentation accuracy, which is crucial for segmenting small tissues such as myocardium.

[0107] Attention-weighted fusion (V3): After replacing attention fusion with equal weights, the average Dice decreased by 1.5%. This indicates that the module can dynamically adjust feature weights and filter out invalid feature interference. Although its contribution is low, it can stably optimize the fusion effect.

[0108] Graph self-representation learning (V4): After removal, the average Dice decreased by 2.8%, with the most significant decrease in myocardial segmentation accuracy (3.7%). This demonstrates its ability to enhance topological detail modeling and improve the representation of complex structural features.

[0109] Contrastive learning constraint (V5): After removing the subspace contrastive loss, the average Dice decreases by 2.3%, and the clustering of similar structural features decreases. This demonstrates that contrastive learning can enhance structural similarity modeling and further reduce inter-domain differences.

[0110] CS divergence loss (V6): After replacing it with cross-entropy loss, the average Dice decreased by 1.8%, and the feature-label matching accuracy decreased. This indicates that CS divergence can more accurately quantify the consistency between feature distribution and labels, and is superior to the supervision effect of traditional cross-entropy loss.

[0111] Target domain reconstruction loss (V7): After removal, the average Dice decreases by 1.3%, making it the module with the lowest contribution, but it can help preserve the feature details of the target domain.

[0112] Quantitative results show that the performance of the complete model is not a simple summation of the modules, but rather achieves efficient complementarity through module collaboration. The performance degradation after removing the top three core modules (two-layer alignment, graph self-representation, and contrastive learning) individually was 4.6%, 2.8%, and 2.3%, respectively, with a cumulative degradation of 9.7%. The sum of the individual degradation of each module was basically consistent with the cumulative degradation, with no additional performance loss, indicating good compatibility between modules.

[0113] Example 2:

[0114] This embodiment provides a transcenter ultrasound image segmentation system, including:

[0115] The data acquisition module is used to acquire a cross-center ultrasound image dataset, which includes source domain center data with segmentation labels and target domain center data without segmentation labels. Both the source domain center data and the target domain center data are ultrasound images containing cardiac anatomical structures, including the ventricles, atria, and myocardium. The feature extraction module is used to input the ultrasound image containing the source domain center data and the target domain center data into the U-Net encoder respectively, and extract the shallow and deep features of the source domain center and the shallow and deep features of the target domain center. The shallow and deep features are used to characterize the texture morphology and spatial structure of the cardiac anatomy. A two-layer feature alignment module is used to input the shallow and deep features of the source domain center and the shallow and deep features of the target domain center into a graph convolutional network, respectively, to obtain the shallow and deep graph features of the source domain center and the shallow and deep graph features of the target domain center. The graph features are used to model the anatomical topological relationships between cardiac anatomical structures. The feature fusion module is used to fuse the shallow and deep graph features of the source domain center and the shallow and deep graph features of the target domain center to obtain the multi-center common features of the source domain center and the target domain center. The common features are used to align the apparent features and topological features of the cardiac anatomy structure between different centers. The graph self-representation learning module is used to input the multi-center common features of the source domain center and the target domain center into the graph convolutional encoder to perform graph self-representation learning, and obtain the content-aware embedding representation of the source domain center and the target domain center. The content-aware embedding representation is used to enhance the topological details of the cardiac anatomical structure. The clustering alignment module is used to input the content-aware embedding representations of the source domain center and the target domain center into the U-Net decoder, combine the segmentation labels for clustering constraints, and use the target domain image reconstruction loss constraints to output the final segmentation result, which is a pixel-level segmentation result of the cardiac anatomical structure.

[0116] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0117] Example 3:

[0118] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described in Embodiment 1.

[0119] Example 4: This example provides an electronic device, including:

[0120] Memory, used to store computer programs / instructions;

[0121] A processor for executing the computer program / instructions to implement the steps of any of the methods described in Embodiment 1.

[0122] Example 5: This example provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in any one of Examples 1.

[0123] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0124] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0125] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0126] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0127] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the pending claims.

Claims

1. A method for segmenting transcenter ultrasound images, characterized in that, include: A cross-center ultrasound image dataset is obtained, which includes source domain center data with segmentation labels and target domain center data without segmentation labels. Both the source domain center data and the target domain center data are ultrasound images containing cardiac anatomical structures, including the ventricles, atria, and myocardium. The ultrasound images containing the source domain center data and the target domain center data are respectively input into the U-Net encoder, and the shallow and deep features of the source domain center and the shallow and deep features of the target domain center are extracted accordingly. The shallow and deep features are used to characterize the texture morphology and spatial structure of the cardiac anatomy. The shallow and deep features of the source domain center and the shallow and deep features of the target domain center are respectively input into the graph convolutional network to obtain the shallow and deep graph features of the source domain center and the shallow and deep graph features of the target domain center. The graph features are used to model the anatomical topological relationships between cardiac anatomical structures. The shallow and deep graph features of the source domain center and the shallow and deep graph features of the target domain center are fused to obtain the multi-center common features of the source domain center and the target domain center. The common features are used to align the apparent features and topological features of the cardiac anatomy between different centers. The shared features of the source and target domain centers are input into the graph convolutional encoder to perform graph self-representation learning, thereby obtaining the content-aware embedding representations of the source and target domain centers. The content-aware embedding representations of the source domain center and the target domain center are input into the U-Net decoder, combined with the segmentation labels for clustering constraints, and the target domain image reconstruction loss constraints are used to output the final segmentation result, which is a pixel-level segmentation result of the cardiac anatomical structure.

2. The transcenter ultrasound image segmentation method according to claim 1, characterized in that, The ultrasound image containing the source domain center data and the target domain center data is input into the U-Net encoder, and the shallow and deep features of the source domain center and the target domain center are extracted accordingly, including: The first one containing cardiac anatomy The first source domain center ultrasound images and the first containing cardiac anatomy The first target domain center ultrasound images Input to the U-Net encoder, output by the shallow coding function of the U-Net encoder. The first source domain center shallow features and the The first target domain center shallow features The deep coding function output by the U-Net encoder is the first... The first source domain center Deep features and the The first target domain center Deep features The shallow features preserve the texture and grayscale distribution of the ventricles, atria, and myocardium, while the deep features preserve the higher-order semantics and spatial layout of the ventricles, atria, and myocardium. The formula is as follows: , ； , ； in, This is the shallow coding function of the U-Net encoder. For the deep encoding function and shallow features of the U-Net encoder All dimensions are Deep features All dimensions are , For image batch size, , These are the image height and width, respectively. This represents the number of shallow feature channels. This represents the number of deep feature channels.

3. The transcenter ultrasound image segmentation method according to claim 2, characterized in that, The shallow and deep features of the source domain center and the shallow and deep features of the target domain center are respectively input into a graph convolutional network to obtain the shallow and deep graph features of the source domain center and the target domain center, respectively, including: shallow features of the source region center Shallow features of the target domain center Compressed to non-overlapping block aggregation After being flattened, the graph node features are input into the graph convolutional network, and then processed by the first... After layer graph convolution, the shallow node feature matrix at the center of the source domain is obtained. Shallow node feature matrix of the target domain center The final layer output is reshaped into a 4-dimensional shallow graph feature of the source domain center. Shallow graph features of the 4D target domain center The nodes in the graph correspond to local areas in the ultrasound image, and their adjacency relationships reflect the spatial proximity of the ventricular wall, atrial wall, and myocardial layer. Among them, the shallow node feature matrix , The calculation formula is as follows: ； ； In the formula: It is the ReLU activation function. This is the shallow normalized adjacency matrix of the source domain center. The shallow normalized adjacency matrix of the target domain center. The shallow features of the source domain center are input into the graph convolutional network and then... The output of the layer, The shallow features centered on the target domain are input into the graph convolutional network and then... The output of the layer, This represents the model parameter matrix for a shallow graph convolutional network. Deep features of the source domain center and deep features of the target domain center After being flattened, the graph node features are input into the graph convolutional network, and then processed by the... After layer graph convolution, the feature matrix of deep nodes at the center of the source domain is obtained. and the deep node feature matrix of the target domain center The final layer output is reshaped into a 4-dimensional deep graph feature at the center of the source domain. Deep graph features of the 4D target domain center The deep map features are used to encode the relative positions and topological relationships between the ventricles, atria, and myocardium; Among them, the deep node feature matrix , The calculation formula is as follows: ； ； in, The deep normalized adjacency matrix of the source domain center. The deep normalized adjacency matrix of the target domain center The deep features of the source domain center are input into the graph convolutional network and then... The output of the layer, The deep features centered on the target domain are input into the graph convolutional network. The output of the layer, This represents the model parameter matrix of a deep graph convolutional network.

4. The transcenter ultrasound image segmentation method according to claim 3, characterized in that, The shallow and deep graph features of the source domain center and the shallow and deep graph features of the target domain center are fused to obtain the multi-center common features of the source domain center and the target domain center, including: Shallow graph features of the source region center and deep map features By splicing along the channel dimension, the shared features of multiple centers of the source domain center are formed. Attention weight matrix for source domain center is generated using a multilayer perceptron. The formula is as follows: ； in, For the Softmax function, The tanh activation function is used. , , , These are learnable parameters; Shallow map features at the center of the target domain and deep map features By splicing along the channel dimension, the shared features of multiple centers in the target domain are formed. An attention weight matrix for the target domain center is generated using a multilayer perceptron. The formula is as follows: ； The average common features are obtained by weighted summation and 3×3 convolution. The formula is as follows: ； in, The total number of source domain centers. It is a 3×3 convolution function; Introducing inter-center alignment loss The difference between shared features and average shared features across multiple centers is constrained. The center alignment loss promotes the consistency of ventricular, atrioventricular, and myocardial features extracted from different centers in the common space, as shown in the following formula: ； in, It is an L2 norm; and This is an adjustable parameter.

5. The transcenter ultrasound image segmentation method according to claim 4, characterized in that, The shared features of the source and target domain centers are input into a graph convolutional encoder for graph self-representation learning to obtain content-aware embedding representations of the source and target domain centers, including: The shared features of multiple centers of source domain centers and target domain centers , Input graph convolution encoder The initial content-aware embedding representations of the source domain centers are obtained respectively. Initial content-aware embedding representation of the target domain center The content-aware embedding representation focuses on discriminative features of the ventricle, atrium, and myocardium, as shown in the following formula: ； ； in, An adjacency matrix with shared features among multiple centers. For GCN encoding functions; Introducing a subspace contrastive learning strategy, and Features before self-representation learning, used as source domain centers and target domain centers respectively; and features after self-representation learning, combined with those of source domain centers and target domain centers. , Construct contrast loss And integrate to obtain self-representation loss To optimize content-aware embedding representations, the self-representation learning method mines the intrinsic correlations between the ventricles, atria, myocardium, and their inter-ventricles through the self-representation coefficient matrix. Contrast loss is used to bridge different representations of the same anatomical structure. Simultaneously, a graph convolutional decoder reconstructs multi-center shared features. Introducing reconstruction loss Finally, the optimized content-aware embedding representation of the source domain center is obtained. Content-aware embedding representation of the target domain center ; Wherein, the contrast loss The formula is as follows: ； in, The cosine similarity function is used. , , , Indicates the first Features before learning the self-representation of the source domain center Indicates the first Features learned from the self-representation of each source domain center It is a self-representation coefficient matrix; The self-representation loss The formula is as follows: ； in, A function to find the diagonal matrix; Indicates the number of source domain centers; The reconstruction loss The formula is as follows: ； in, The reconstructed features of the source domain center are output by the graph convolutional decoder. The reconstructed features of the target domain center output by the graph convolutional decoder.

6. The transcenter ultrasound image segmentation method according to claim 5, characterized in that, The content-aware embedding representations of the source domain centers and target domain centers are input into the U-Net decoder, combined with segmentation labels for clustering constraints, and the target domain image reconstruction loss is used as a constraint to output the final segmentation result, including: The optimized source domain center content-aware embedding representation Inputting a 3×3 convolutional layer and a softmax function yields the prediction probability matrix. ,in It is a 3×3 convolutional layer, and the prediction probability matrix reflects the probability that each pixel belongs to the ventricle, atrium, myocardium or background. Will After upsampling to the original image size, the first cross-entropy loss is used. Calculate pixel-by-pixel classification accuracy: ； in, The number of pixels in a source domain image. For the prediction probability matrix Upsample to the original image size to predict the probability matrix. For pixels The true label; Introducing Cauchy-Schwarz divergence As a clustering loss, the clustering loss enhances class separability by maximizing the difference between the probability distributions of each category of ventricle, atrium, and myocardium, as shown in the following formula: ； in, The number of segmentation categories is ventricles, atria, and myocardium. The affinity matrix is ​​calculated as follows: , For the prediction probability matrix The One channel; For the prediction probability matrix The One channel; It is the transpose symbol; The optimized source domain center content-aware embedding representation Input the U-Net decoder to obtain the predicted segmentation result of the source domain center. and using the second cross-entropy loss Calculate pixel-by-pixel classification accuracy: ； in, express The first in 1 pixel; Content-aware embedding representation of the target domain center Input to U-Net decoder to reconstruct the target domain image Introducing reconstruction loss The formula is as follows: ； Construct the total loss function The formula is used for joint optimization training of the U-Net encoder and U-Net decoder, as follows: ； in For loss weights; After training, the content-aware embedding representation of the target domain center is used. Input the trained U-Net decoder to obtain the predicted segmentation result of the target domain center, and output it as the final segmentation result.

7. A transcenter ultrasound image segmentation system, characterized in that, include: The data acquisition module is used to acquire a cross-center ultrasound image dataset, which includes source domain center data with segmentation labels and target domain center data without segmentation labels. Both the source domain center data and the target domain center data are ultrasound images containing cardiac anatomical structures, including the ventricles, atria, and myocardium. The feature extraction module is used to input the ultrasound image containing the source domain center data and the target domain center data into the U-Net encoder respectively, and extract the shallow and deep features of the source domain center and the shallow and deep features of the target domain center. The shallow and deep features are used to characterize the texture morphology and spatial structure of the cardiac anatomy. A two-layer feature alignment module is used to input the shallow and deep features of the source domain center and the shallow and deep features of the target domain center into a graph convolutional network, respectively, to obtain the shallow and deep graph features of the source domain center and the shallow and deep graph features of the target domain center. The graph features are used to model the anatomical topological relationships between cardiac anatomical structures. The feature fusion module is used to fuse the shallow and deep graph features of the source domain center and the shallow and deep graph features of the target domain center to obtain the multi-center common features of the source domain center and the target domain center. The common features are used to align the apparent features and topological features of the cardiac anatomy structure between different centers. The graph self-representation learning module is used to input the multi-center common features of the source domain center and the target domain center into the graph convolutional encoder to perform graph self-representation learning, and obtain the content-aware embedding representation of the source domain center and the target domain center. The content-aware embedding representation is used to enhance the topological details of the cardiac anatomical structure. The clustering alignment module is used to input the content-aware embedding representations of the source domain center and the target domain center into the U-Net decoder, combine the segmentation labels for clustering constraints, and use the target domain image reconstruction loss constraints to output the final segmentation result, which is a pixel-level segmentation result of the cardiac anatomical structure.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the steps of the method described in any one of claims 1-6.

9. An electronic device, characterized in that, include: Memory, used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of the method according to any one of claims 1-6.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1-6.

Images