Brain network classification method and device based on hierarchical graph convolution and self-attention mechanism

Abstract

The invention discloses a brain network classification method and device based on hierarchical graph convolution and a self-attention mechanism, and the method comprises the steps: extracting a BOLD average time sequence of ROIs of the brain rs-fMRI of a subject, and taking the BOLD average time sequence as a graph node feature; calculating the similarity of the BOLD average time sequences of any two ROIs, constructing a functional connection matrix, and constructing an adjacent matrix of graph nodes through binarization processing; inputting each graph node feature and the corresponding adjacent matrix into a first-stage graph convolution network in a brain classification model to carry out graph convolution operation, and extracting a feature vector of the graph node; and constructing a group graph structure by using a self-attention mechanism, inputting the group graph structure into a second-stage graph convolution network in the brain classification model for graph convolution operation, extracting feature vectors of the group graph structure, inputting the feature vectors into a multi-layer perceptron in the brain classification model, predicting a classification result of the subject, and improving the classification precision of the brain network.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image classification, and in particular to a brain network classification method and device based on layered graph convolution and self-attention mechanism. Background Art

[0002] Alzheimer's disease (AD) is a common progressive neurodegenerative disease and an irreversible neurological disease. After the onset of the disease, patients will experience symptoms such as memory loss, inattention, forgetfulness, susceptibility to distraction, and slow response. The onset of AD is a slow process and there is currently no cure, but intervention in the mild cognitive impairment (MCI) stage can effectively slow down the progression of the disease. Therefore, it is very important to accurately treat and diagnose AD in the early stages, which is of great significance to the patients themselves, their families, and society. With the continuous development of medical imaging technology, resting-state functional magnetic resonance imaging (rs-fMRI) has been widely used in the diagnosis of brain diseases as a non-invasive method. rs-fMRI is based on the spontaneous neural activity of the brain. Even if the brain is in a resting state, its neurons will continue to spontaneously activity, which can be captured and recorded in rs-fMRI. By measuring changes in the blood oxygenation level-dependent (BOLD) signal between different brain regions, rs-fMRI can reveal the functional connectivity patterns of the brain at rest.

[0003] At present, classification methods based on deep learning models mainly include methods based on fully connected neural networks (FCNN), methods based on convolutional neural networks (CNN), and methods based on graph neural networks (GNN). Compared with traditional regular images, functional connection networks constructed from neuroimaging data are more difficult to process and represent due to their complex non-Euclidean features. In recent years, deep convolutional neural networks have been proven to be an effective method for processing brain network structures, but convolutional neural networks mainly focus on network input rather than network data. Graph Convolutional Networks (GCNs), as a popular graph embedding learning representation method, can well model the relationship between brain image data and has been widely used in brain network analysis tasks. However, previous brain network feature learning usually only analyzes a single brain network and ignores the hierarchical relationship between brain networks. Summary of the invention

[0004] In response to the problems existing in the prior art, the present invention provides a brain network classification method and device based on layered graph convolution and self-attention mechanism, which simultaneously considers the correlation between individual brain networks and individual brain networks in group brain networks to improve the accuracy of brain network classification.

[0005] To achieve the above technical objectives, the present invention adopts the following technical solution: a brain network classification method based on layered graph convolution and self-attention mechanism, specifically comprising the following steps:

[0006] Step S1, collecting the brain rs-fMRI of the subjects and marking the severity of AD patients;

[0007] Step S2, dividing the brain rs-fMRI of each subject into regions based on the AAL template, extracting the BOLD average time series of ROIs, and using the BOLD average time series of ROIs of each subject as the graph node feature;

[0008] Step S3, calculating the similarity of the BOLD average time series of any two ROIs of each subject by Pearson correlation coefficient, constructing a functional connectivity matrix, and constructing an adjacency matrix of the graph nodes by binarization;

[0009] Step S4, constructing a brain classification model based on layered graph convolution and self-attention mechanism, including: a first-stage graph convolution network, a second-stage graph convolution network and a multi-layer perceptron;

[0010] Step S5: Input each graph node feature and the corresponding adjacency matrix into the first-stage graph convolution network to perform graph convolution operation and extract the feature vector of the graph node;

[0011] Step S6: Use the self-attention mechanism to construct a group graph structure using the feature vectors of the extracted graph nodes, input the group graph structure into the second-stage graph convolution network for graph convolution operation, and extract the feature vector of the group graph structure;

[0012] Step S7, using the extracted feature vector of the group graph structure as the input of the multi-layer perceptron, and using the disease severity of each subject in the corresponding group as the label of the multi-layer perceptron to train the constructed brain classification model;

[0013] Step S8, repeating steps S5-S7 until the mean square error between the predicted classification and the label of the multilayer perceptron converges, thus completing the training of the brain classification model;

[0014] Step S9, obtain the brain rs-fMRI of a subject, repeat steps S2-S3, input the trained brain classification model, and predict the classification result of the AD patient.

[0015] Furthermore, the construction process of the functional connectivity matrix is ​​as follows:

[0016]

[0017] Among them, p ij represents the Pearson correlation coefficient between the i-th ROI brain region and the j-th ROI brain region of the subject, x i and x j Respectively represent the BOLD average time series signals of the i-th ROI brain region and the j-th ROI brain region of the subject, and Respectively represent x i and x j The mean of Represents x i and x j The standard deviation of .

[0018] Furthermore, the construction process of the adjacency matrix is ​​as follows: a threshold parameter ω is set to perform binarization processing on the constructed functional connectivity matrix, and the Pearson correlation coefficient in the functional connectivity matrix is ​​set to 1 if it is greater than the threshold parameter ω, otherwise it is set to 0, to obtain the adjacency matrix.

[0019] Furthermore, the first-stage graph convolution network adopts a cascaded architecture of three cascaded graph convolution layers, inputs each graph node feature and the corresponding adjacency matrix into the first-layer graph convolution unit for processing, and extracts the feature vector of the first-level graph node; inputs the feature vector of the first-level graph node into the second-layer graph convolution unit for processing, and extracts the feature vector of the second-level graph node; inputs the feature vector of the second-level graph node into the third-layer graph convolution unit for processing, and obtains the feature vector of the graph node. Further, the first-stage graph convolution network is composed of three consecutive graph convolution units, and the graph convolution layer in the first-layer graph convolution unit performs graph convolution operations on the graph node features and the corresponding adjacency matrix, introduces residual connections after the graph convolution operation, and uses activation functions to perform nonlinear activation on the features, and uses normalization layers to normalize the features, and the standardized features are further transformed using two layers of fully connected layers, wherein the activation function layer and dropout layer are applied after the first fully connected layer, and the residual connection module, activation function layer, dropout layer, and normalization layer are applied in sequence after the second fully connected layer. The first-stage graph convolution network has a second-layer graph convolution unit and a third-layer graph convolution unit with the same network structure, both of which are equipped with a graph convolution layer. The feature vector of the graph node extracted by the previous layer is subjected to graph convolution operation, and the activation function is used to activate the feature. The normalization layer is used to normalize the feature. The standardized feature is further transformed using two fully connected layers to extract the feature vector of the graph node. The activation function layer and the dropout layer are applied after the first fully connected layer, and the residual connection module, the activation function layer, the dropout layer and the normalization layer are applied in sequence after the second fully connected layer.

[0020] Furthermore, step S6 includes the following sub-steps:

[0021] Step S6.1, use the self-attention mechanism to calculate the correlation between the feature vectors of any two graph nodes as the weight of the edge corresponding to the graph node;

[0022] Step S6.2, set the edge threshold, set the edge weight greater than the edge threshold to 1 and retain the edge; otherwise, set it to 0 and delete the edge, and obtain the group graph structure and the adjacency matrix of the group graph structure;

[0023] Step S6.3, the feature vectors of all graph nodes are combined into a node matrix of the group graph structure, which together with the adjacency matrix of the group graph structure are used as the input of the second-stage graph convolutional network to extract the feature vector of the group graph structure.

[0024] Furthermore, the calculation process of the weight of the edge corresponding to the graph node in step S6.1 is:

[0025]

[0026] Among them, Q, K, and V represent the query matrix, key matrix, and value matrix obtained by linearly transforming the matrix composed of the feature vectors of the two graph nodes, respectively. k The embedding dimension of the matrix representing the feature vectors of the two graph nodes.

[0027] Furthermore, the graph convolution network of the second stage adopts a cascade architecture of three cascaded graph convolution layers, and inputs the node matrix of the group graph structure and the adjacency matrix of the group graph structure into the first layer of graph convolution units for processing, and extracts the feature vector of the group graph structure of the first layer of graph convolution units; inputs the feature vector of the group graph structure of the first layer of graph convolution units into the second layer of graph convolution units for processing, and extracts the feature vector of the group graph structure of the second layer of graph convolution units; inputs the feature vector of the group graph structure of the second layer of graph convolution units into the third layer of graph convolution units for processing, and extracts the feature vector of the group graph structure.

[0028] Furthermore, the network structures of the first-layer graph convolution unit, the second-layer graph convolution unit, and the third-layer graph convolution unit of the second-stage graph convolution network are the same, and all are provided with graph convolution layers. Graph convolution operation is performed on the feature vector of the group graph structure extracted in the previous layer, and the activation function is used to activate the features. The normalization layer is used to standardize the features, and the standardized features are further transformed using two layers of fully connected layers to extract the feature vector of the group graph structure, wherein the activation function layer and the dropout layer are applied after the first fully connected layer, and the residual connection module, the activation function layer, the dropout layer and the normalization layer are applied in sequence after the second fully connected layer.

[0029] Furthermore, the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the brain network classification method based on layered graph convolution and self-attention mechanism is implemented.

[0030] Compared with the prior art, the present invention has the following beneficial effects: the brain network classification method based on layered graph convolution and self-attention mechanism of the present invention constructs a brain classification model based on layered graph convolution and self-attention mechanism, wherein the first-stage graph convolution network is used to extract individual features. Subsequently, a group graph structure is constructed through a self-attention mechanism. By using a self-attention machine, the global dependencies between all individuals can be captured, and the weights between graph nodes can be dynamically allocated according to the correlation between individuals. Due to its parallel computing characteristics, the computing performance of the brain classification model can be improved; afterwards, the group graph structure is extracted through a second-stage graph convolution network. By jointly learning the brain network features of individuals and groups, individual specificity and group commonality can be utilized simultaneously, reducing the impact of individual differences on the brain classification model, and extracting rich feature expressions, thereby improving the classification performance of the brain classification model. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flowchart of the brain network classification method based on layered graph convolution and self-attention mechanism of the present invention. DETAILED DESCRIPTION

[0032] The technical solution of the present invention is further explained below in conjunction with the accompanying drawings.

[0033] like Figure 1 The flowchart of the brain network classification method based on layered graph convolution and self-attention mechanism of the present invention includes the following steps:

[0034] Step S1, collecting the brain rs-fMRI of the subjects, and marking the disease severity of the AD patients, including: late mild cognitive impairment LMCI, early mild cognitive impairment EMCI, Alzheimer's disease AD and normal NC.

[0035] Step S2: The brain rs-fMRI of each subject was preprocessed using the standard procedure in the FSLFEAT software package. The first three volumes were discarded for magnetization equilibrium before preprocessing, and then the remaining 137 volumes were processed according to the standard pipeline. The preprocessed brain fMRI was mapped with the automatic anatomical labeling AAL template to obtain 116 ROIs of brain rs-fMRI, thereby dividing the brain rs-fMRI of each subject into regions, extracting the BOLD average time series of ROIs, and using the BOLD average time series of ROIs of each subject as the graph node feature.

[0036] Step S3: Calculate the similarity of the BOLD average time series of any two ROIs of each subject through the Pearson correlation coefficient, so as to measure the functional connection strength between brain regions, construct a functional connection matrix, and construct an adjacency matrix of the graph nodes through binarization.

[0037] The construction process of the functional connectivity matrix in the present invention is:

[0038]

[0039] Among them, p ij represents the Pearson correlation coefficient between the i-th ROI brain region and the j-th ROI brain region of the subject, x i and x j Respectively represent the BOLD average time series signals of the i-th ROI brain region and the j-th ROI brain region of the subject, and Respectively represent x i and x j The mean of Represents xi and x j The standard deviation, p ij The range of is distributed between [-1, +1]. A value greater than 0 indicates a positive correlation, and a value less than 0 indicates a negative correlation. ij The closer the absolute value is to 1, the more relevant it is, and the closer it is to 0, the less relevant it is.

[0040] The construction process of the adjacency matrix is ​​as follows: a threshold parameter ω is set to binarize the constructed functional connection matrix, and the Pearson correlation coefficient in the functional connection matrix is ​​set to 1 if it is greater than the threshold parameter ω, otherwise it is set to 0, and the adjacency matrix is ​​obtained.

[0041] Step S4, construct a brain classification model based on layered graph convolution and self-attention mechanism, including: a first-stage graph convolution network, a second-stage graph convolution network and a multi-layer perceptron, wherein the first-stage graph convolution network is used to extract individual features. Then, a group graph structure is constructed through a self-attention mechanism. By using a self-attention machine, the global dependencies between all nodes can be captured, and the weights between nodes can be dynamically allocated according to the correlation between nodes. Due to its parallel computing characteristics, the brain network classification computing performance can be improved; then, the group graph structure is used to extract group features through a second-stage graph convolution network, considering the deeper relationship between group brain networks, thereby combining individual brain network features with group brain network features. By jointly learning the brain network features of individuals and groups, individual specificity and group commonality can be utilized at the same time, reducing the impact of individual differences on the brain classification model, extracting rich feature expressions, and improving the classification performance of the brain classification model.

[0042] Step S5, input each graph node feature and the corresponding adjacency matrix into the first-stage graph convolutional network for graph convolution operation, extract the feature vector of the graph node, and convert the graph structure into the feature representation of the graph node by aggregating the information of the graph node and its neighbors, so that the graph information can be converted into a vector form while retaining the local neighborhood information of the graph node; specifically, the first-stage graph convolutional network adopts a cascade architecture of three cascaded graph convolutional layers (Graph Convolutional Layer, GCL), inputs each graph node feature and the corresponding adjacency matrix into the first-layer graph convolution unit for processing, performs preliminary transformation and fusion on the node features, and thus extracts the feature vector of the first-level graph node; inputs the feature vector of the first-level graph node into the second-layer graph convolution unit for processing, and deepens the feature vector obtained in the previous step through nonlinear transformation and feature aggregation mechanism to extract the feature vector of the second-level graph node; inputs the feature vector of the second-level graph node into the third-layer graph convolution unit for processing to extract the feature vector of the graph node.

[0043] In the present invention, the first-stage graph convolution network is composed of three consecutive graph convolution units. The graph convolution layer in the first-layer graph convolution unit is used as a core component to perform graph convolution operations on graph node features and corresponding adjacency matrices. After the graph convolution operation, residual connections are introduced, and activation functions are used to perform nonlinear activation on features, and normalization layers are used to perform feature normalization. The standardized features are further transformed using two fully connected layers, wherein an activation function layer and a dropout layer are applied after the first fully connected layer, and a residual connection module, an activation function layer, a dropout layer, and a normalization layer are applied in sequence after the second fully connected layer; the second-layer graph convolution unit of the first-stage graph convolution network has the same network structure as the third-layer graph convolution unit, and both are provided with a graph convolution layer, and a graph convolution operation is performed on the feature vectors of the graph nodes extracted from the previous layer. Similarly, a combination of activation functions, normalization layers, and two fully connected layers is used to further refine features, wherein the processing flow after the fully connected layer is consistent with that of the first layer, thereby ensuring the consistency and coherence of the first-stage graph convolution network in processing features at different levels.

[0044] Step S6: Use the self-attention mechanism to construct a group graph structure using the feature vectors of the graph nodes extracted in the first stage, input the group graph structure into the second-stage graph convolution network for graph convolution operation, and extract the feature vector of the group graph structure. By using the self-attention machine, the global dependency between all nodes can be captured, and the weights between nodes can be dynamically allocated according to the correlation between the nodes. Due to its parallel computing characteristics, the computing performance of the brain classification model can be improved; including the following sub-steps:

[0045] Step S6.1, use the self-attention mechanism to calculate the correlation between the feature vectors of any two graph nodes as the weight of the corresponding edge of the graph node;

[0046] The calculation process of the weight of the edge corresponding to the graph node in the present invention is:

[0047]

[0048] Among them, Q, K, and V represent the query matrix, key matrix, and value matrix obtained by linearly transforming the matrix composed of the feature vectors of the two graph nodes, respectively. k The embedding dimension of the matrix representing the feature vectors of the two graph nodes.

[0049] Step S6.2, set the edge threshold, set the edge weight greater than the edge threshold to 1 and retain the edge; otherwise, set it to 0 and delete the edge, and obtain the group graph structure and the adjacency matrix of the group graph structure;

[0050] Step S6.3, the feature vectors of all subjects are combined into a node matrix of a group graph structure, and the node matrix of the group graph structure is used together with the adjacency matrix of the group graph structure as the input of the second-stage graph convolution network to extract the feature vector of the group graph structure; specifically, the second-stage graph convolution network also adopts the cascade architecture of three cascaded graph convolution layers. In this stage, the node matrix of the group graph structure and the adjacency matrix of the group graph structure are input into the first-layer graph convolution unit for processing, the relevant information between the subjects is integrated, and the feature vector of the group graph structure of the first-layer graph convolution unit is extracted; the feature vector of the group graph structure of the first-layer graph convolution unit is input into the second-layer graph convolution unit for processing, the information from the wider neighborhood of the first layer is integrated, and the feature vector of the group graph structure of the second-layer graph convolution unit is extracted; the feature vector of the group graph structure of the second-layer graph convolution unit is input into the third-layer graph convolution unit for processing, the most critical and global features in the group graph structure are captured, and the feature vector of the group graph structure is obtained.

[0051] In the present invention, the network structures of the first-layer graph convolution unit, the second-layer graph convolution unit, and the third-layer graph convolution unit of the second-stage graph convolution network are the same, and all are provided with graph convolution layers. Graph convolution operation is performed on the feature vector of the group graph structure extracted by the previous layer, and the activation function is used to activate the features, and then the normalization layer is used to standardize the features. The standardized features are further transformed using two layers of fully connected layers to extract the feature vector of the group graph structure, wherein the activation function layer and the dropout layer are applied after the first fully connected layer, and the residual connection module, the activation function layer, the dropout layer and the normalization layer are applied in sequence after the second fully connected layer.

[0052] Step S7: Using the extracted feature vector of the group graph structure as the input of the multi-layer perceptron, and using the disease severity of each subject in the corresponding group as the label of the multi-layer perceptron, to train the constructed brain classification model.

[0053] Step S8, repeat steps S5-S7 until the mean square error between the predicted classification and the label of the multilayer perceptron converges, thus completing the training of the brain classification model.

[0054] Step S9, obtain the brain rs-fMRI of a subject, repeat steps S2-S3, input the trained brain classification model, and predict the classification result of the subject.

[0055] The data set used in the present invention is derived from the public Alzheimer's Disease Neuroimaging Initiative (ADNI) database, covering resting-state functional magnetic resonance imaging (rs-fMRI) data of 174 participants. These data include 31 patients with Alzheimer's disease (AD), 45 individuals with late mild cognitive impairment (LMCI), 50 patients with early mild cognitive impairment (EMCI), and 48 normal control group (NC) members. Each participant received multiple scans over a period of 6 months to 1 year. Specifically, the scanning frequencies of the EMCI, LMCI, AD, and NC groups were 165, 145, 99, and 154 times, respectively, with a total of 563 scan records.

[0056] In the present invention, a five-fold cross-validation strategy was used to test the effectiveness of the proposed four-classification method (distinguishing AD, NC, EMCI, and LMCI). Specifically, the overall data set was evenly divided into five parts, one of which was selected as the test set in each cycle, and the remaining parts were combined as training sets. In each cross-validation cycle, 20% of the training set was also divided as the validation set to determine the optimal parameter configuration of the model. In order to enhance the generalization performance of the model, each scan of each subject was independently treated as a sample, but all scans of the same subject were guaranteed to maintain consistent category identification. Finally, the performance of the model was comprehensively evaluated by calculating the overall classification accuracy and the classification accuracy of each category. The brain network method proposed in the present invention was compared with the following methods: (1) Baseline: This method first calculated the Pearson correlation coefficient PCC of the BOLD signal to construct a static functional connectivity network FCN. Then, the static FCN was feature screened using the t-test (with a p value set to less than 0.05), and the support vector machine SVM technology was used to further perform feature selection and classification tasks. (2) SVM: First, a fixed functional connection network FCN is constructed for each subject, and then the local clustering coefficient is extracted from its network as the feature value. In the feature selection stage, the t-test method is used; in the classification stage, the support vector machine SVM with default settings is used for classification. (3) GCN: This method constructs a functional connection network as input, then uses continuous convolution operations to extract features, and uses MLP for classification. (4) HGCN: This method constructs a functional connection network as input, first performs graph convolution for feature extraction, then uses non-imaging information as the edge of the group brain network, and then inputs it into the graph convolution for node classification.

[0057] Table 1 shows the experimental results of the above methods in the set multi-classification task. It can be seen from Table 1 that for the ADNI dataset, the accuracy index of the present invention is higher than that of the other methods, especially in terms of accuracy.NC , Accuracy lMCI , Accuracy AD The above results are significantly better than other methods, indicating that the present invention combines the characteristics of individual brain networks with those of group brain networks by considering the deeper relationship between group brain networks to enhance the classification ability of brain networks.

[0058] Table 1 Performance of all methods on AD vs. lMCI vs. eMCI vs. NC classification (%)

[0059]

[0060] In one technical solution of the present invention, an electronic device is also provided, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the brain network classification method based on layered graph convolution and self-attention mechanism is implemented.

[0061] In the embodiments disclosed in the present application, the computer storage medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. The computer storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the above. More specific examples of computer storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0062] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0063] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A brain network classification method based on layered graph convolution and self-attention mechanism, characterized in that: The specific steps include: Step S1, obtaining the subject's brain rs-fMRI and marking the subject's disease severity; Step S2, dividing the brain rs-fMRI of each subject into regions based on the AAL template, extracting the BOLD average time series of ROIs, and using the BOLD average time series of ROIs of each subject as the graph node feature; Step S3, calculating the similarity of the BOLD average time series of any two ROIs of each subject by Pearson correlation coefficient, constructing a functional connectivity matrix, and constructing an adjacency matrix of the graph nodes by binarization; Step S4, constructing a brain classification model based on layered graph convolution and self-attention mechanism, including: a first-stage graph convolution network, a second-stage graph convolution network and a multi-layer perceptron; Step S5: Input each graph node feature and the corresponding adjacency matrix into the first-stage graph convolution network to perform graph convolution operation and extract the feature vector of the graph node; Step S6: Use the self-attention mechanism to construct a group graph structure using the feature vectors of the extracted graph nodes, input the group graph structure into the second-stage graph convolution network for graph convolution operation, and extract the feature vector of the group graph structure; Step S7, using the extracted feature vector of the group graph structure as the input of the multi-layer perceptron, and using the disease severity of each subject in the corresponding group as the label of the multi-layer perceptron to train the constructed brain classification model; Step S8, repeating steps S5-S7 until the mean square error between the predicted classification and the label of the multilayer perceptron converges, thus completing the training of the brain classification model; Step S9, obtain the brain rs-fMRI of a subject, repeat steps S2-S3, input the trained brain classification model, and predict the classification result of the subject.

2. A brain network classification method based on layered graph convolution and self-attention mechanism according to claim 1, characterized in that: The construction process of the functional connectivity matrix is ​​as follows: Among them, p ij represents the Pearson correlation coefficient between the i-th ROI brain region and the j-th ROI brain region of the subject, x i and x j Respectively represent the BOLD average time series signals of the i-th ROI brain region and the j-th ROI brain region of the subject, and Respectively represent x i and x j The mean of Represents x i and x j The standard deviation of .

3. A brain network classification method based on layered graph convolution and self-attention mechanism according to claim 2, characterized in that: The construction process of the adjacency matrix is ​​as follows: a threshold parameter ω is set to perform binarization processing on the constructed functional connection matrix, and the Pearson correlation coefficient in the functional connection matrix is ​​set to 1 if it is greater than the threshold parameter ω, otherwise it is set to 0, to obtain the adjacency matrix.

4. The brain network classification method based on layered graph convolution and self-attention mechanism according to claim 1, characterized in that: The first-stage graph convolutional network adopts a cascaded architecture of three cascaded graph convolutional layers. Each graph node feature and the corresponding adjacency matrix are input into the first-layer graph convolution unit for processing, and the feature vector of the first-level graph node is extracted; the feature vector of the first-level graph node is input into the second-layer graph convolution unit for processing, and the feature vector of the second-level graph node is extracted; the feature vector of the second-level graph node is input into the third-layer graph convolution unit for processing, and the feature vector of the graph node is obtained.

5. A brain network classification method based on layered graph convolution and self-attention mechanism according to claim 4, characterized in that: The first-stage graph convolution network consists of three consecutive graph convolution units. In the first layer of graph convolution units, the graph convolution layer performs graph convolution operations on the graph node features and the corresponding adjacency matrix. After the graph convolution operation, residual connections are introduced, and activation functions are used to perform nonlinear activation on the features. The normalization layer is used to normalize the features, and the standardized features are further transformed using two fully connected layers. The activation function layer and dropout layer are applied after the first fully connected layer, and the residual connection module, activation function layer, dropout layer and normalization layer are applied in sequence after the second fully connected layer. layer; the network structure of the second-layer graph convolution unit of the first-stage graph convolution network is the same as that of the third-layer graph convolution unit, both of which are equipped with a graph convolution layer. Graph convolution operation is performed on the feature vector of the graph node extracted by the previous layer, and the activation function is used to activate the feature. The normalization layer is used to standardize the feature. The standardized feature is further transformed using two fully connected layers to extract the feature vector of the graph node. Among them, the activation function layer and the dropout layer are applied after the first fully connected layer, and the residual connection module, the activation function layer, the dropout layer and the normalization layer are applied in sequence after the second fully connected layer.

6. The brain network classification method based on layered graph convolution and self-attention mechanism according to claim 1, characterized in that: Step S6 includes the following sub-steps: Step S6.1, use the self-attention mechanism to calculate the correlation between the feature vectors of any two graph nodes as the weight of the corresponding edge of the graph node; Step S6.2, set the edge threshold, set the edge weight greater than the edge threshold to 1 and retain the edge; otherwise, set it to 0 and delete the edge, and obtain the group graph structure and the adjacency matrix of the group graph structure; Step S6.3, the feature vectors of all graph nodes are combined into a node matrix of the group graph structure, which together with the adjacency matrix of the group graph structure are used as the input of the second-stage graph convolutional network to extract the feature vector of the group graph structure.

7. A brain network classification method based on layered graph convolution and self-attention mechanism according to claim 6, characterized in that: The calculation process of the weight of the edge corresponding to the graph node in step S6.1 is: Among them, Q, K, and V represent the query matrix, key matrix, and value matrix obtained by linearly transforming the matrix composed of the feature vectors of the two graph nodes, respectively. k The embedding dimension of the matrix representing the feature vectors of the two graph nodes.

8. The brain network classification method based on layered graph convolution and self-attention mechanism according to claim 1, characterized in that: The graph convolution network of the second stage adopts a cascade architecture of three cascaded graph convolution layers, and inputs the node matrix of the group graph structure and the adjacency matrix of the group graph structure into the first layer of graph convolution units for processing, and extracts the feature vector of the group graph structure of the first layer of graph convolution units; inputs the feature vector of the group graph structure of the first layer of graph convolution units into the second layer of graph convolution units for processing, and extracts the feature vector of the group graph structure of the second layer of graph convolution units; inputs the feature vector of the group graph structure of the second layer of graph convolution units into the third layer of graph convolution units for processing, and extracts the feature vector of the group graph structure.

9. A brain network classification method based on layered graph convolution and self-attention mechanism according to claim 8, characterized in that: The network structures of the first-layer graph convolution unit, the second-layer graph convolution unit, and the third-layer graph convolution unit of the second-stage graph convolution network are the same, and all are equipped with graph convolution layers. Graph convolution operation is performed on the feature vector of the group graph structure extracted in the previous layer, and the activation function is used to activate the features. The normalization layer is used to standardize the features, and the standardized features are further transformed using two fully connected layers to extract the feature vector of the group graph structure. Among them, the activation function layer and the dropout layer are applied after the first fully connected layer, and the residual connection module, the activation function layer, the dropout layer and the normalization layer are applied in sequence after the second fully connected layer.

10. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the brain network classification method based on layered graph convolution and self-attention mechanism as described in any one of claims 1 to 9 is implemented.

Images