Brain network classification method based on functional topology and semantic double hypergraph representation learning

Abstract

The brain network classification method based on function topology and semantic double hypergraph representation learning disclosed by the application firstly acquires brain network data and performs preprocessing; function topology hypergraphs and semantic hypergraphs are respectively constructed by using the obtained function connection matrix and attribute information; hypergraph convolution networks are used to extract brain region representations under two types of hypergraph structures; the obtained brain region representations are adaptively fused to obtain fused nodes and input into a classifier, and finally the class prediction result of brain network classification is obtained. Through the above mode, the high-order interaction across brain regions is captured; the topology hypergraph is used to describe the explicit high-order topological relationship in the function connection correlation, and the stable topological mode in the original brain network is retained; the semantic hypergraph is based on node attribute coding and similarity modeling, and adaptively mines the potential high-order semantic association between brain regions. The application can effectively mine the high-order collaborative relationship between multiple brain regions in the functional brain network, thereby improving the brain network classification performance.

Description

Technical Field

[0001] This invention belongs to the field of brain science research technology, specifically involving a brain network classification method based on functional topology and semantic dual hypergraph representation learning. Background Technology

[0002] Functional magnetic resonance imaging (fMRI) can reflect the coordinated functional activities between brain regions through blood oxygenation level-dependent signals, making it an important non-invasive imaging technique for brain network research and assisted diagnosis. Functional brain networks based on resting-state fMRI typically use brain regions as nodes, constructing a functional connectivity matrix by calculating the statistical correlation between brain region time series data to characterize the coordinated activity relationships between brain regions. This type of method has been widely applied in research on mental disorders and neurodegenerative diseases. Traditional brain network analysis methods often rely on artificial graph theory indicators such as clustering coefficients, modularity, and node centrality to describe the topological organization of brain networks. While this provides a foundation for understanding brain functional abnormalities, relying solely on artificial features makes it difficult to fully extract brain network connectivity information.

[0003] In recent years, graph neural networks (Graph Neural Networks) and hypergraph neural networks (HNNs) have been increasingly applied to the learning of functional brain network representations. Graph Neural Networks can perform end-to-end message passing on the topology of brain region connections, enabling automatic learning of brain network features. Some methods further combine graph pooling, adaptive graph structure learning, information bottlenecks, or self-supervised learning mechanisms to reduce the impact of noisy connections and redundant information on model performance. Compared to ordinary graph structures, hypergraphs can connect multiple brain regions simultaneously through a single hyperedge, making them more suitable for characterizing high-order functional collaborations involving multiple brain regions. Therefore, hypergraph-based brain network modeling methods provide a new technical path for brain network classification and show certain advantages in representing high-order relationships. However, existing brain network modeling methods still have certain limitations. On the one hand, traditional graph neural networks are mainly based on binary connections between brain regions, making it difficult to fully represent high-order interactions between multiple brain regions. On the other hand, existing hypergraph construction methods are easily affected by noisy connections and individual differences. Furthermore, some methods mainly construct hypergraphs from a single functional topology perspective, failing to fully explore the potential semantic relationships of the overall connection patterns of brain region nodes.

[0004] To address this, this invention proposes a brain network classification method based on functional topology and semantic dual hypergraph representation learning, which can fully model high-order topological structures and latent semantic relationships, thereby improving brain network classification performance. Summary of the Invention

[0005] The purpose of this invention is to provide a brain network classification method based on functional topology and semantic dual hypergraph representation learning, which solves the problem that existing methods are unable to fully represent the high-order interaction relationships between multiple brain regions and the potential semantic relationships between brain region nodes, resulting in poor accuracy in brain network classification.

[0006] The technical solution adopted in this invention is a brain network classification method based on functional topology and semantic hypergraph representation learning. By constructing functional topology hypergraphs and semantic hypergraphs, it mines higher-order topological relationships and potential semantic associations between nodes in the functional brain network, thereby fully representing the collaborative activity patterns involving multiple brain regions. Simultaneously, it uses a hypergraph neural network to learn representations of brain region features under different hypergraph views, reducing the impact of noisy connections and redundant information on algorithm performance and improving the accuracy of brain network classification. Specifically, as follows: S1: Brain Network Data Acquisition and Preprocessing: Acquire functional magnetic resonance imaging (fMRI) data, and preprocess the fMRI data using DPABI, including discarding the first 10 time points, slice time correction, head movement correction, alignment of T1 images to functional image space, spatial normalization, etc. Subsequently, brain regions were divided using brain atlases. We identified regions of interest and extracted the mean oxygenation level dependent (BOLD) time series for each brain region. The functional connectivity matrix was obtained by calculating the Pearson correlation coefficient between the mean BOLD time series of any two brain regions. ,in To represent the number of brain regions, we define brain regions as nodes, then the set of nodes is represented as... ,in Indicates the first Each node has a set of node attributes defined as follows: , Represents a node Attribute vectors, Represents the functional connection matrix The Row vectors; S2: Utilize the obtained functional connection matrix and attribute information Construct functional topology hypergraphs respectively and semantic hypergraph ;in, Represents a set of nodes. Describes the set of superedges of a topological hypergraph. The hyperedge weight matrix represents the hypergraph's topological weights. Represents the set of superedges of a semantic hypergraph. The hyperedge weight matrix represents the semantic hypergraph. S3: For the obtained and Furthermore, HGNN+ hypergraph convolutional network is used to extract node features under the two types of hypergraph structures respectively. and ; S4: Analyze the obtained node features and Perform adaptive fusion to obtain the fused node representation. The data is then input into a classifier to obtain the brain network category prediction results.

[0007] In step S2, the k-nearest neighbor (KNN) algorithm is used to select the brain regions with the highest functional relevance for each node, constructing hyperedges in the topological view: For any node In the functional connection matrix Select with The front with the strongest connection There are nodes, and nodes Together they form a superedge :

[0008] in This indicates selecting the node with the strongest connection strength. 1 node Represents the functional connection matrix The Column vector. For example, when At that time, each hyperedge consists of one node and two nodes with the strongest connection strength. Therefore, for a hyperedge containing The functional brain network of each brain region node can be obtained super edge Ultimately, the functional hypergraph under the topology view can be represented as: , This is the hyperedge weight matrix, and the corresponding hypergraph incidence matrix is:

[0009] in, , No. Columns represent The first generated by the center A superedge. If the node Belongs to superedge Then the corresponding element in the correlation matrix takes the value of Otherwise, the value is 0; Indicates the superedge internal nodes The weight.

[0010] In step S2, the topological hypergraph can model local high-order relationships in brain networks. To uncover potential associations between overall connectivity patterns in brain regions, an attribute-adaptive hypergraph under the semantic view is further constructed.

[0011] For the obtained attribute matrix , No. The attribute vector of each node is A shared node encoder is used to perform preliminary feature extraction on attribute information, resulting in a feature matrix for hypergraph modeling under the semantic view. :

[0012] in, This represents a multilayer perceptron (MLP). For the first Hidden layer attributes of each brain region.

[0013] Obtaining the hidden layer representation of nodes Then, symmetrical node pair features are further constructed:

[0014] in, Indicates feature splicing, Describe the common semantic components of the hidden layer representations of two brain regions. Used to characterize the differences between the two in the latent semantic space; The original functional connection strength information is then retained.

[0015] Subsequently, the node pair features Inputting into an MLP, we obtain the similarity score between two nodes:

[0016] in, This represents an MLP used for node-to-relationship scoring; Furthermore, the relationship score is mapped to normalized soft association weights:

[0017] in, It is the Sigmoid function. Brain regions under semantic view and The association strength among brain regions that are grouped into the same higher-order functional group is determined by node attribute encoding and node pair relationship learning, and thus can reflect the potential attribute similarity and functional semantic relationship between brain regions.

[0018] For each node Choose the one with the highest similarity score. 1 node, and with Together they form a semantic hyperedge:

[0019] in, This indicates selecting the node with the strongest connection strength. 1 node Representation matrix The Column vectors; each brain region node generates a corresponding semantic hyperedge, and all semantic hyperedges together constitute the hyperedge set under the semantic view. Therefore, the functional hypergraph under the semantic view can be represented as , This is the hyperedge weight matrix, and the corresponding hypergraph incidence matrix is:

[0020] in, , No. Columns represent The first generated by the center A superedge. If the node Belongs to superedge Use the learned weights To indicate the strength of membership, the corresponding element in the association matrix takes the value of... ; Indicates the superedge internal nodes The weight.

[0021] In step S3, for the obtained topological hypergraph and semantic hypergraph Furthermore, HGNN+ hypergraph convolutional network was used to extract brain region representations under two types of hypergraph structures. and ; First, the features of nodes belonging to the same hyperedge are aggregated to form the higher-order functional representation corresponding to that hyperedge. Then, this hyperedge representation is propagated to each node it connects to, updating the node representations and thus modeling the higher-order collaborative relationships in the functional brain network. The HGNN+ hypergraph convolution operation process is formalized as follows:

[0022]

[0023] in, This represents the hyperedge feature matrix obtained by aggregating node features. Represents the hypergraph incidence matrix. This represents the number of superedges; and They represent the first Layer and first Layer node feature matrix, This represents the number of nodes in the hypergraph. The dimension of the vertex features of this layer; and Let these represent the hyperedge degree matrix and the node degree matrix, respectively. This is the super-diagonal weight matrix. For learnable parameter matrix, This represents a nonlinear activation function. Through the above process, node features are first aggregated to the hyperedge level, and then propagated from the hyperedge to the node level, enabling nodes to receive contextual information from their respective higher-order functional groups.

[0024] During the propagation process, the correlation matrix It describes the membership relationships between vertices and hyperedges and determines which vertex features should be aggregated into the corresponding hyperedges. First, the model... Features of nodes belonging to the same hyperedge are aggregated to form a hyperedge-level representation. Subsequently, the hyperedge representation is propagated to the corresponding vertex based on the same incidence matrix, and combined with the learnable parameter matrix. The vertex features are updated to obtain the next layer of vertex representation. .

[0025] For topological hypergraphs and semantic hypergraph HGNN+ hypergraph convolutional network was used to extract brain region representations under two types of hypergraph structures. and :

[0026]

[0027] in, Represents the node attribute matrix; and These represent the hypergraph association matrices under the topological view and semantic view, respectively; These represent the brain region embedding vectors learned from the two views, respectively.

[0028] In step S4, two sets of node-level representations, a topological hypergraph and a semantic hypergraph, are obtained from step S3, denoted as follows: and A gating fusion mechanism is employed to adaptively fuse the representations of the two types of nodes. Considering that the functional topological hypergraph can provide relatively stable structural constraints, the fusion process is based on the topological view representation, and the strength of the supplementary semantic view information is controlled by gating weights.

[0029] For the Let there be nodes, and let their representations under the two hypergraph branches be respectively... and First, the node representations of the two types of hypergraphs are input into the gating network to obtain the node-level weights of the semantic hypergraph branches:

[0030] in, Represents the gate function. It is the Sigmoid function.

[0031] Then define the first The fusion of nodes is represented as The fused representation of all nodes is as follows :

[0032]

[0033] in, It is a learnable global residual coefficient used to constrain the overall enhancement of the semantic hypergraph representation to the topological hypergraph representation.

[0034] For the fused node representation The mean-max pooling strategy is used to further generate feature representations. and The two are then concatenated to obtain the sample-level feature vector. ; global features Obtained from the input MLP classifier , Indicates that the sample corresponds to The output scores for each category are then used; subsequently, the Softmax function is applied to calculate the category scores. Normalized to class prediction probability, i.e. ,in This refers to the brain network being classified into the first... Class probability, for The first in Each element; and cross-entropy loss is used. And comparative loss Conduct supervised training:

[0035]

[0036]

[0037] in, For the total loss, and These are the weight parameters; The number of categories is [number], and the categories include two types: normal and those with brain diseases. For real labels, To predict probabilities; For cosine similarity, and These are the graph-level representations of the topological hypergraph and semantic hypergraph of the same sample subject, respectively. For temperature parameters, It is an indicator function, when hour It is 1 if it is true, otherwise it is 0.

[0038] The beneficial effects of this invention are as follows: This invention fully considers the high-order collaborative relationships and node attribute association information between brain regions in functional brain networks. By constructing functional topological hypergraphs and semantic adaptive hypergraphs, it jointly models complex multi-brain region interaction patterns. Compared to traditional graph models that only utilize pairwise functional connections, this invention can more effectively capture the high-order connectivity structures in which multiple brain regions jointly participate in disease representation, thereby enhancing the ability to identify abnormal functional patterns related to brain networks. Secondly, this invention uses a functional topological hypergraph as the backbone and a semantic adaptive hypergraph as a supplementary enhancement branch. Through a gating fusion mechanism, it adaptively adjusts the contributions of the two types of hypergraph representations, enabling the model to retain stable topological information in the functional connectivity structure while also incorporating potential association information reflected by node attribute similarity. This design can reduce information bias caused by a single graph construction method and improve the accuracy of brain network classification. Attached Figure Description

[0039] Figure 1 This is a framework diagram of the brain network classification method based on functional topology and semantic dual hypergraph representation learning of the present invention. Detailed Implementation

[0040] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0041] Example 1 The brain network classification method proposed in this invention, based on functional topology and semantic dual-hypergraph representation learning, such as... Figure 1 As shown, it includes the following steps: S1: Brain Network Data Acquisition and Preprocessing: Acquire functional magnetic resonance imaging (fMRI) data, use AAL brain atlas to divide brain regions, and obtain functional connectivity matrices. ,in Let the number of brain regions be represented, and let each brain region be defined as a node. Then the set of nodes is: ,in Indicates the first Each node has a set of node attributes defined as follows: , Represents a node Attribute vectors; S2: Utilize the obtained functional connection matrix and attribute information Construct functional topology hypergraphs respectively and semantic hypergraph ;in, Represents a set of nodes. Describes the set of superedges of a topological hypergraph. The weight matrix represents the topological hyperedge. Represents the set of superedges of a semantic hypergraph. The weight matrix representing the semantic hyperedge; S3: For the obtained and Furthermore, HGNN+ hypergraph convolutional network was used to extract brain region features under two types of hypergraph structures. and ; S4: Representation of the obtained brain regions and Perform adaptive fusion to obtain the fused node representation. The data is then input into a classifier to obtain the brain network category prediction results.

[0042] Example 2 Based on Example 1, in step S2, according to the obtained functional connection matrix The k-nearest neighbor algorithm is used to select the brain regions with the highest functional relevance to each node, constructing a set of hyperedges in the topological view. ; For any node In the functional connection matrix Select with The front with the strongest connection There are nodes, and nodes Together they form a superedge :

[0043] in This indicates selecting the node with the strongest connection strength. 1 node Represents the functional connection matrix The Column vector. For the ABIDE-II dataset, set... Each hyperedge consists of one node and four nodes with the strongest connection strength. Therefore, for a hyperedge containing... A functional brain network of brain regions can yield 116 hyperedges. .

[0044] In a preferred embodiment of the present invention, in step S2, the set of hyperedges under the calculated topological view is... Calculate the hypergraph inclination matrix under the topological view as follows: :

[0045] in, , Represents a node With super-edge The subordinate relationship between them, the first Columns represent The first generated by the center A super edge; if node Belongs to superedge Then the corresponding element in the correlation matrix takes the value of Otherwise, the value is 0; Indicates the superedge internal nodes The weight.

[0046] In a preferred embodiment of the present invention, in step S2, the hyperedge set under the obtained topological view is... and the correlation matrix Constructing the functional hypergraph under the topology view is ,in, A set of nodes defined for 116 brain regions. Represents the hyperedge weight matrix; In a preferred embodiment of the present invention, in step S2, in order to explore the potential associations between overall connectivity patterns in brain regions, an attribute-adaptive hypergraph under the semantic view is further constructed, and the obtained attribute matrix is... A shared node encoder is used to perform preliminary feature extraction on attribute information, resulting in a feature matrix for hypergraph modeling under the semantic view. :

[0047] in, This represents a multilayer perceptron (MLP). For the first Hidden layer attributes of each brain region.

[0048] In a preferred embodiment of the present invention, in step S2, the obtained node hidden layer representation is utilized. Construct symmetric node pair features :

[0049] in, Indicates feature splicing, Describe the common semantic components of the hidden layer representations of two brain regions. Used to characterize the differences between the two in the latent semantic space; The original functional connection strength information is then retained.

[0050] In a preferred embodiment of the present invention, in step S2, the calculated node pair features are... Input a two-layer multilayer perceptron (MLP) structure to obtain the similarity score between two nodes. :

[0051] in, This represents an MLP used for node-to-relationship scoring; In a preferred embodiment of the present invention, in step S2, the relationship is scored. Mapped to normalized soft correlation weights :

[0052] in, It is the Sigmoid function. Brain regions under semantic view and The association strength among those grouped into the same higher-order functional group is determined by node attribute encoding and node pair relationship learning, and thus can reflect the potential attribute similarity and functional semantic relationship between brain regions.

[0053] In a preferred embodiment of the present invention, in step S2, for each node Choose the one with the highest similarity score. 1 node, and with Together they form a semantic hyperedge :

[0054] in, This indicates selecting the node with the strongest connection strength. 1 node Representation matrix The Column vectors; thus, all semantic hyperedges together constitute the set of hyperedges under the semantic view. For the ABIDE-II dataset, set... Each hyperedge consists of a node and the node with the highest similarity score to it; In a preferred embodiment of the present invention, in step S2, the set of hyperedges under the calculated topological view is... Calculate the hypergraph inclination matrix under the topological view as follows: :

[0055] in, , No. Columns represent The first generated by the center A superedge. If the node Belongs to superedge Use the learned weights To indicate the strength of membership, the corresponding element in the association matrix takes the value of... ; Indicates the superedge internal nodes The weight.

[0056] In a preferred embodiment of the present invention, in step S2, based on the obtained set of hyperedges under the semantic view... and the correlation matrix Constructing a hypergraph under a semantic view is ,in, A set of nodes defined for 116 brain regions. Represents the hyperedge weight matrix; Example 3 Based on Example 2, in step S3, for the obtained topological hypergraph and semantic hypergraph Furthermore, HGNN+ hypergraph convolutional network was used to extract brain region features under two types of hypergraph structures. and ; First, the features of nodes belonging to the same hyperedge are aggregated to form the higher-order functional representation corresponding to that hyperedge. Then, this hyperedge representation is propagated to each node it connects to, updating the node representations and thus modeling the higher-order collaborative relationships in the functional brain network. The HGNN+ hypergraph convolution operation process is formalized as follows:

[0057]

[0058] in, This represents the hyperedge feature matrix obtained by aggregating node features. Represents the hypergraph incidence matrix. This represents the number of superedges; and They represent the first Layer and first Layer node feature matrix, This represents the number of nodes in the hypergraph. The dimension of the vertex features of this layer; and Let these represent the hyperedge degree matrix and the node degree matrix, respectively. The hyperdiagonal weight matrix is ​​set to a unit diagonal matrix. For learnable parameter matrix, This represents a nonlinear activation function. Through the above process, node features are first aggregated to the hyperedge level, and then propagated from the hyperedge to the node level, enabling nodes to receive contextual information from their respective higher-order functional groups. During the propagation process, the correlation matrix... It describes the membership relationships between vertices and hyperedges and determines which vertex features should be aggregated into the corresponding hyperedges. First, the model... Features of nodes belonging to the same hyperedge are aggregated to form a hyperedge-level representation. Subsequently, the hyperedge representation is propagated to the corresponding vertex based on the same incidence matrix, and combined with the learnable parameter matrix. The vertex features are updated to obtain the next layer of vertex representation. .

[0059] Example 4 In a preferred embodiment of the present invention, based on Example 3, for the topological hypergraph and semantic hypergraph HGNN+ hypergraph convolutional network was used to extract brain region representations under two types of hypergraph structures. and :

[0060]

[0061] in, Represents the node attribute matrix; and These represent the hypergraph association matrices under the topological view and semantic view, respectively; These represent the brain region embedding representations learned from the two views, respectively.

[0062] Example 5 Based on Example 4, in step S4, the two sets of node-level representations of the calculated topological hypergraph and semantic hypergraph are respectively denoted as... and For the first Let there be nodes, and let their representations under the two hypergraph branches be respectively... and First, the node representations of the two types of hypergraphs are input into the gating network to obtain the node-level weights of the semantic hypergraph branches. :

[0063] in, Represents the gate function. It is the Sigmoid function.

[0064] In a preferred embodiment of the present invention, in step S4, the first... The fusion of nodes is represented as Then, the fused representation of all nodes is calculated as follows: :

[0065]

[0066] in, It is a learnable global residual coefficient used to constrain the overall enhancement of the semantic hypergraph representation to the topological hypergraph representation.

[0067] In a preferred embodiment of the present invention, in step S4, the fused node representation... The mean-max pooling strategy is used to further generate feature representations. and The two are then concatenated to obtain the sample-level feature vector. .

[0068] In a preferred embodiment of the present invention, in step S4, global features are... Obtained from the input MLP classifier , Indicates that the sample corresponds to The output scores are divided into two categories: patients and healthy individuals. A Softmax function is then used to convert the category scores into a single output score. Normalized to class prediction probability, i.e. ,in This refers to the brain network being classified into the first... Class probability, for The first in Each element; and cross-entropy loss is used. And comparative loss Conduct supervised training:

[0069]

[0070]

[0071] in, For the total loss, and These are the weight parameters; For the number of categories, For real labels, To predict probabilities; For cosine similarity, and These are the graph-level representations of the topological hypergraph and semantic hypergraph of the same sample subject, respectively. For temperature parameters, It is an indicator function, when hour It is 1 if it is true, otherwise it is 0. Example 6 The implementation of this embodiment for the brain network classification task is as follows: Using ABIDE-II (18 sites / 1054 subjects) as the data source, ABIDE-II is a public dataset on autism spectrum disorder (ASD), including 497 ASD patients and 557 healthy controls (HC) subjects. The raw fMRI data from ABIDE-II was preprocessed using DPABI. After preprocessing, the brain was divided into 116 regions of interest using the AAL brain atlas, and the BOLD time series of each brain region was extracted. A functional connectivity matrix was constructed by calculating the Pearson correlation coefficient between any two BOLD time series of brain regions. ,in Represents the number of brain regions, with the node set as follows: The node attribute set is defined as follows: , Represents a node Attribute vectors; To construct the topological hypergraph and semantic hypergraph, based on the obtained functional connectivity matrix With k=5, the KNN algorithm is used to select the four neighboring brain regions that are most functionally relevant to each node, constructing a total of 116 topological hyperedges, thus obtaining the hyperedge set of the topological hypergraph. According to the hyperedge set Calculate the hypergraph incidence matrix , Represents a node With super-edge The hierarchical relationship between nodes, if nodes Belongs to superedge Then the corresponding element in the correlation matrix takes the value of Otherwise, the value is 0. To uncover potential associations among overall brain region connectivity patterns, an attribute-adaptive hypergraph under the semantic view is further constructed using a two-layer perceptron. For attribute matrix Preliminary feature extraction is performed, with the nonlinear activation function set to ReLU, to obtain the feature matrix used for hypergraph modeling under the semantic view. , For the first Hidden layer attribute representation of individual brain regions; using node hidden layer representation Construct node pairs features , node pair features Input two-layer perceptron In this process, the non-linear activation function is set to ReLU, and the similarity score between the two nodes is obtained. Using the Sigmoid function to Mapped to normalized soft correlation weights , Brain regions under semantic view and The association strength between brain regions grouped into the same higher-order functional group is determined by both node attribute encoding and node pair relationship learning, thus reflecting the potential attribute similarity and functional semantic relationships between brain regions. Similarity score The larger the value, the stronger the semantic similarity between the two brain regions, and the higher the corresponding soft association weight. The closer to 1, the higher the similarity score. The smaller the value, the higher the soft association weight. The closer it is to 0. For example, when the similarity score is... Sometimes, , representing a node and There is a strong correlation. Then, based on the obtained... With k=2, the KNN algorithm is used to select the node with the highest similarity score for each node, constructing a total of 116 semantic hyperedges, thus obtaining the set of hyperedges of the semantic hypergraph. According to the hyperedge set Calculate the hypergraph incidence matrix , Represents a node With super-edge The hierarchical relationship between nodes, if nodes Belongs to superedge Then the corresponding element in the correlation matrix takes the value of Otherwise, the value is 0. The final constructed hypergraph is the topological hypergraph. and semantic hypergraph .

[0072] Then, HGNN+ hypergraph convolutional network was used to... and Feature extraction is performed to obtain brain region representations. and And calculate the fused representation of all nodes as follows: The mean-max pooling strategy is used to further generate feature representations. and The two are then concatenated to obtain the sample-level feature vector. Further global features Obtained from the input MLP classifier , Indicates that the sample corresponds to The output scores for each category are then used; subsequently, the Softmax function is applied to calculate the category scores. Normalized to class prediction probability, i.e. ,in This refers to the brain network being classified into the first... Class probability, for The first in The training phase is based on the PyTorch framework. The ABIDE-II dataset is divided into training, testing, and validation sets in a 7:2:1 ratio. The model is trained using the training set and trained with the Adam optimizer at a batch size of 16. The initial learning rate is 1×10⁻⁶. 4. The maximum number of training iterations is 150. The best model is selected using the validation set, and the final output is the class classification result of the brain network. The performance is validated using four commonly used metrics: accuracy (ACC), sensitivity (SEN), specificity (SPE), and F1 score. To verify the performance of the constructed model, GCN, GAT, BrainGNN, HGNN+, and ALTER methods were selected for performance comparison. Experimental results show that the proposed TSHG algorithm achieves the best classification performance across all four metrics.

[0073] Table 1 shows the comparison results between the proposed TSHG algorithm and other methods.

[0074] Table 1 shows that the proposed TSHG algorithm (the method of this invention) outperforms other comparative methods in all four metrics on the ABIDEII dataset, achieving the best performance in the ASD (autism disorder) and HC (healthy individuals) classification tasks. Compared to the best-performing baseline HGNN+, the TSHG algorithm improves ACC and F1 by 5.59 and 6.12 percentage points, respectively; compared to the sensitivity results of GAT and HGNN+, the TSHG algorithm improves SEN by approximately 7.07 percentage points. This indicates that the method of this invention not only improves the overall classification accuracy but also enhances the ability to identify ASD patient samples. The method of this invention simultaneously introduces a functional topological hypergraph and a semantic adaptive hypergraph, enabling the model to incorporate potential semantic association information while preserving stable functional connectivity patterns. Therefore, it maintains relatively small fluctuations in all four metrics, demonstrating better robustness.

[0075] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A brain network classification method based on functional topology and semantic dual hypergraph representation learning, characterized in that, Functional topological hypergraphs and semantic hypergraphs are constructed using brain network data to mine higher-order topological relationships and potential semantic associations between nodes in functional brain networks, thereby fully representing collaborative activity patterns involving multiple brain regions. At the same time, brain region features under different hypergraph views are represented and learned through hypergraph neural networks.

2. The brain network classification method based on functional topology and semantic dual hypergraph representation learning according to claim 1, characterized in that, The specific steps are as follows: S1: Brain Network Data Acquisition and Preprocessing fMRI brain imaging data was acquired, preprocessed using DPABI, and then brain regions were divided using brain atlases. We identified regions of interest and extracted the mean blood oxygenation level dependent on the BOLD time series for each brain region. The functional connectivity matrix was obtained by calculating the Pearson correlation coefficient between the mean BOLD time series of any two brain regions. ; in, Let the number of brain regions be represented, and each brain region be defined as a node. Then the set of nodes is: ,in Indicates the first Each node has a set of node attributes defined as follows: , Represents a node Attribute vectors, Represents the functional connection matrix The Row vectors; S2: Utilize the obtained functional connection matrix and attribute information Construct functional topology hypergraphs respectively and semantic hypergraph ;in, Represents a set of nodes. Describes the set of superedges of a topological hypergraph. The hyperedge weight matrix represents the hypergraph's topological weights. Represents the set of superedges of a semantic hypergraph. The hyperedge weight matrix represents the semantic hypergraph. S3: For the obtained and Hypergraph Neural Network (HGNN) and Hypergraph Convolutional Network (HCN) are used to extract node features for two types of hypergraph structures. and ; S4: Analyze the obtained node features and Perform adaptive fusion to obtain the fused node representation. The data is then input into a classifier to obtain the brain network category prediction results.

3. The brain network classification method based on functional topology and semantic dual hypergraph representation learning according to claim 2, characterized in that, In step S1, preprocessing includes discarding the first 10 time points, slice time correction, head motion correction, alignment of T1 image to functional image space, and spatial normalization.

4. The brain network classification method based on functional topology and semantic dual hypergraph representation learning according to claim 2, characterized in that, In step S2, the KNN algorithm is used to select the neighboring brain regions with the highest functional relevance for each node, constructing hyperedges in the topological view: For any node In the functional connection matrix Select with The front with the strongest connection There are nodes, and nodes Together they form a topological hyperedge ; in, This indicates selecting the node with the strongest connection strength. 1 node Represents the functional connection matrix The Column vector; for containing The functional brain network of each brain region node can obtain super edge Ultimately, the functional hypergraph under the topology view is represented as follows: , This is the hyperedge weight matrix, and the corresponding hypergraph incidence matrix is: in, , No. Columns represent The first generated by the center A super edge; if node Belongs to superedge Then the corresponding element in the correlation matrix takes the value of Otherwise, the value is 0; Indicates the superedge internal nodes The weights; In step S2, the topological hypergraph can model local high-order relationships in the brain network. In order to explore the potential associations between the overall connection patterns of brain regions, an attribute adaptive hypergraph under the semantic view is further constructed.

5. The brain network classification method based on functional topology and semantic dual hypergraph representation learning according to claim 4, characterized in that, In step S2, for the obtained attribute matrix , No. The attribute vector of each node is A shared node encoder is used to perform preliminary feature extraction on attribute information, resulting in a feature matrix for hypergraph modeling under the semantic view. : in, This represents a multilayer perceptron (MLP). For the first Hidden layer attributes of individual brain regions; After obtaining the hidden layer representation of the nodes, we further construct symmetric node pair features: in, Indicates feature splicing, Describe the common semantic components of the hidden layer representations of two brain regions. Used to characterize the differences between the two in the latent semantic space; The original functional connection strength information is then retained; Subsequently, the node pair features Inputting into an MLP, we obtain the similarity score between two nodes: in, This represents an MLP used for node-to-relationship scoring; Map the relationship score to normalized soft association weights: in, It is the Sigmoid function. Brain regions under semantic view and The association strength of brain regions that are grouped into the same higher-order functional group is determined by node attribute encoding and node pair relationship learning, and thus can reflect the potential attribute similarity and functional semantic relationship between brain regions. For each node Choose the one with the highest similarity score. 1 node, and with Together they form a semantic hyperedge: in, This indicates selecting the node with the strongest connection strength. 1 node Representation matrix The Column vectors; thus, each brain region node generates a corresponding semantic hyperedge, and all semantic hyperedges together constitute the hyperedge set under the semantic view. Therefore, the functional hypergraph under the semantic view can be represented as , This is the hyperedge weight matrix, and the corresponding hypergraph incidence matrix is: in, , No. Columns represent The first generated by the center A super edge; if node Belongs to superedge Use the learned weights To indicate the strength of membership, the corresponding element in the association matrix takes the value of... ; Indicates the superedge internal nodes The weight.

6. The brain network classification method based on functional topology and semantic dual hypergraph representation learning according to claim 5, characterized in that, In step S3, For the obtained topological hypergraph and semantic hypergraph Furthermore, HGNN+ hypergraph convolutional network was used to extract brain region representations under two types of hypergraph structures. and ; First, the features of nodes belonging to the same hyperedge are aggregated to form the higher-order functional representation corresponding to that hyperedge. Then, this hyperedge representation is propagated to each node it connects to, updating the node representations and thus modeling the higher-order collaborative relationships in the functional brain network. The HGNN+ hypergraph convolution operation process is formalized as follows: in, This represents the hyperedge feature matrix obtained by aggregating node features. Represents the hypergraph incidence matrix. This represents the number of superedges; and They represent the first Layer and first Layer node feature matrix, This represents the number of nodes in the hypergraph. The dimension of the vertex features of this layer; and Let these represent the hyperedge degree matrix and the node degree matrix, respectively. This is the super-diagonal weight matrix. For learnable parameter matrix, This represents a nonlinear activation function. Through the above process, node features are first aggregated to the hyperedge level, and then propagated from the hyperedge to the node level, enabling the node to receive contextual information from its higher-order functional group.

7. The brain network classification method based on functional topology and semantic dual hypergraph representation learning according to claim 6, characterized in that, During the propagation process, the hypergraph correlation matrix It describes the membership relationship between vertices and hyperedges and determines which vertex features should be aggregated into the corresponding hyperedges; first, the model based on Features of nodes belonging to the same hyperedge are aggregated to form a hyperedge-level representation. Subsequently, the hyperedge representation is propagated to the corresponding vertex based on the same incidence matrix, and combined with the learnable parameter matrix. The vertex features are updated to obtain the next layer of vertex representation. ; For topological hypergraphs and semantic hypergraph HGNN+ hypergraph convolutional network was used to extract brain region representations under two types of hypergraph structures. and : in, Represents the node attribute matrix; and These represent the hypergraph association matrices under the topological view and semantic view, respectively; These represent the brain region features learned from the two views.

8. The brain network classification method based on functional topology and semantic dual hypergraph representation learning according to claim 7, characterized in that, In step S4, two sets of node-level representations, a topological hypergraph and a semantic hypergraph, are obtained from step S3, denoted as follows: and A gated fusion mechanism is used to adaptively fuse the two types of node representations. Considering that the functional topology hypergraph can provide relatively stable structural constraints, the fusion process is based on the topology view representation and the strength of the supplementary semantic view information is controlled by the gate weight. For the Let there be nodes, and let their representations under the two hypergraph branches be respectively... and First, the node representations of the two types of hypergraphs are input into the gating network to obtain the node-level weights of the semantic hypergraph branches; in, Represents the gate function. It is the Sigmoid function; Then define the first The fusion of nodes is represented as The fused representation of all nodes is as follows : in, These are learnable global residual coefficients used to constrain the overall enhancement of the semantic hypergraph representation to the topological hypergraph representation; For the fused node representation The mean-max pooling strategy is used to further generate feature representations. and The two are then concatenated to obtain the sample-level feature vector. ; global features Obtained from the input MLP classifier , Indicates that the sample corresponds to The output scores for each category are then used; subsequently, the Softmax function is applied to calculate the category scores. Normalized to class prediction probability, i.e. ,in This refers to the brain network being classified into the first... Class probability, for The first in Each element; and cross-entropy loss is used. And comparative loss Conduct supervised training: in, For the total loss, and These are the weight parameters; The number of categories is [number], and the categories include two types: normal and those with brain diseases. For real labels, To predict probabilities; For cosine similarity, and These are the graph-level representations of the topological hypergraph and semantic hypergraph of the same sample subject, respectively. For temperature parameters, It is an indicator function, when hour It is 1 if it is true, otherwise it is 0.

Images