Brain network variable graph convolution analysis method driven by brain cognition

Abstract

The invention discloses a brain network variable graph convolution analysis method driven by brain cognition, which belongs to the technical field of medical image processing, and comprises the following steps: S1, aiming at DTI modal data and fMRI data of a brain network, carrying out brain community division by adopting an overlapped adaptive brain network community detection method based on deep learning; s2, defining an adaptive community detection score to obtain an optimal brain region affiliation matrix; and S3, based on the optimal brain region affiliation matrix, establishing an adaptive variable graph convolutional network based on a brain community structure by adopting a one-way message passing mechanism to perform brain network classification. According to the method, a graph convolution receptive field is dynamically adjusted in combination with a community structure of a brain network, the brain structure and functional data are fused, community detection is optimized by utilizing a Bernoulli-Poisson probability model, directed information flow in a brain interval is simulated by introducing a one-way message passing mechanism, a remote dependency relationship and high-order topological characteristics in the brain interval are effectively captured, and the method is suitable for the remote detection of the brain network. And the accuracy and individualization ability of brain network analysis are obviously improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a brain network variable graph convolution analysis method driven by brain cognition. Background Art

[0002] In the field of brain network analysis, existing technologies mainly focus on extracting topological features of brain networks and applying deep learning methods such as graph convolutional neural networks (GCNs) to analyze the structure and function of brain networks. The topological feature extraction techniques for shallow brain networks usually rely on small-world network characteristics, such as metrics like clustering coefficient and degree centrality. These features can help describe the connection degree and information transmission ability of each node in the brain network. However, most of these methods are limited to the extraction of low-order topological features and are difficult to capture more complex high-order topological structures in the brain network.

[0003] With the development of deep learning technologies, more and more studies have started to adopt methods such as graph convolutional neural networks (GCNs) to extract high-order features of brain networks. For example, GCNs can capture the non-linear structure of a graph through the aggregation of adjacent nodes, thereby revealing the complex relationships between nodes in the brain network. Although GCNs have to some extent solved the limitations of traditional methods, their fixed receptive field size still restricts their ability to capture the dependence relationships of distant nodes. This problem leads to difficulties for GCNs in capturing long-range dependencies between brain regions.

[0004] In addition, existing dynamic graph convolutional networks (such as deformable GCN and adaptive GCN) although introduce dynamic receptive fields to adjust convolutional operations, so as to better adapt to the requirements of different tasks, but these methods usually only focus on the feature aggregation of local neighborhood nodes and are difficult to effectively capture the long-range dependence relationships existing in the brain network. At the same time, most existing methods rely on manually setting rules to define the receptive field size and neighborhood division, which limits their flexibility and adaptability.

[0005] Although some deep learning methods attempt to overcome these problems by integrating information such as small-world attributes, hierarchical structures, and functional connectivity, most methods ignore the modular structure of the brain network and the mutual influence between the structural and functional networks.

[0006] Therefore, although existing technologies have made certain progress in brain network analysis, there are still deficiencies in being unable to effectively capture high-order topological features, long-range node dependence relationships, and in not fully combining prior knowledge in neuroscience and the interaction between structural and functional networks. These defects limit their application and effectiveness in brain disease classification.

[0007] Based on the above defects, a brain network variable graph convolution analysis method driven by brain cognition is proposed. Summary of the Invention

[0008] The object of the present invention is to provide a brain cognitive-driven variable graph convolutional analysis method for brain networks to solve the problems in the background art.

[0009] To achieve the above object, the present invention provides a brain cognitive-driven variable graph convolutional analysis method for brain networks, including the following steps:

[0010] S1. For the DTI modality data and fMRI data of the brain network, an overlapping and adaptive brain network community detection method based on deep learning is used to generate a brain region attribution matrix, and a Bernoulli-Poisson loss function is designed to optimize the parameters of the brain region attribution matrix for brain community division;

[0011] S2. Define an adaptive community detection score to evaluate the result of S1, select the most suitable brain community structure, and obtain the most suitable brain region attribution matrix;

[0012] S3. Based on the most suitable brain region attribution matrix, a one-way message passing mechanism is used to establish an adaptive variable graph convolutional network based on the brain community structure for analysis to obtain a brain disease classification result.

[0013] Preferably, in S1, the brain region attribution matrix is expressed as:

[0014]

[0015] In the formula, F is the brain region attribution matrix, ReLU(·) represents a non-linear activation function, GCN(·) represents a graph convolutional network, A is a brain region adjacency matrix constructed from DTI modality data, representing the structural information of the brain network, is the normalized brain region adjacency matrix, X is a brain region embedding matrix constructed from fMRI data, representing the functional information of the brain network, W (1) and W (2) are both weight matrices in the brain region attribution matrix.

[0016] Preferably, in S1, the Bernoulli-Poisson loss function is expressed as:

[0017]

[0018] In the formula, is the Bernoulli-Poisson loss function, E is the edge set, N is the set of node pairs without edge connections, represents the expected value of the node pair (u, v) sampled from the distribution P E of the edge set, F u is the brain region attribution vector of node u, is the transpose of the brain region attribution vector of node v, It represents the common community degree between node u and node v. exp(·) is the exponential function, and log(·) is the logarithmic function.

[0019] Preferably, in the step S2, the adaptive community detection score is expressed as:

[0020] Score = Cov + α·AvgCC - β·AvgCon;

[0021] In the formula, Score is the adaptive community detection score, Cov is the community coverage, which measures how many connected edges can be explained by the final detected community through the Bernoulli - Poisson probability distribution model. AvgCC is the weighted average community clustering coefficient, which quantifies the clustering degree between nodes and their connected neighbors in the graph. AvgCon is the weighted average community conductance, which measures the connection strength between the target community and the external community. Both α and β are parameters.

[0022] Preferably, the Bernoulli - Poisson probability distribution model is expressed as:

[0023]

[0024] In the formula, A uv is the Bernoulli - Poisson probability distribution, and Bernoulli(·) represents the Bernoulli distribution.

[0025] Preferably, the community coverage is expressed as:

[0026]

[0027] In the formula, is the indicator function, z v is the adjacency vector of node v, represents the connection strength between nodes u and v;

[0028] The weighted average community clustering coefficient is expressed as:

[0029]

[0030] In the formula, C i is the i - th community, CC(C i ) is the clustering coefficient of community C i ;

[0031] The weighted average community conductance is expressed as:

[0032]

[0033] In the formula, Outside(C i ) represents the nodes in community C i and the community Ci The number of edges of the external nodes, and Inside(·) represents the number of internal edges of the community.

[0034] Preferably, in step S3, the specific process of establishing an adaptive variable graph convolutional network based on the brain community structure is as follows:

[0035] S31. Define a relevance threshold to filter out the nodes with low relevance in the optimal brain region attribution matrix, and screen out the effective community attribution matrix;

[0036] S32. Based on the one-way message passing mechanism, set up a message passing channel between the nodes belonging to the same community in the effective community attribution matrix to obtain the message weight matrix;

[0037] S33. Define an adaptive variable graph convolution based on the message weight matrix.

[0038] Preferably, in step S32, the message weight matrix is expressed as:

[0039]

[0040] In the formula, M s [·] is the message passing matrix, both x and y are brain regions, is the message weight matrix, += represents the cumulative assignment operation, F′ i [·] represents the effective community attribution matrix.

[0041] Preferably, the message passing matrix is expressed as:

[0042]

[0043] Preferably, in step S33, the adaptive variable graph convolution is expressed as:

[0044]

[0045] In the formula, H (l+1) is the node feature matrix of the (l + 1)-th layer, σ is the activation function, is the degree matrix, W (l) is the weight matrix of the adaptive variable graph convolution, and ⊕ represents the tensor product.

[0046] Therefore, the brain network variable graph convolution analysis method driven by brain cognition of the present invention has the following beneficial effects:

[0047] (1) By introducing an adaptive mechanism, it can dynamically adjust the receptive field size according to the community structure of the brain regions, so as to maximize the use of modular information in the brain network, enabling the model to more accurately capture the high-order features of the brain network and improving the accuracy of analysis.

[0048] (2) By integrating the structural and functional information of brain regions, it is possible to comprehensively analyze brain networks, capture the complex interaction relationships in brain networks, avoid the limitations of the single modality of existing traditional methods, enhance the comprehensiveness of brain network analysis, and improve the accuracy in practical applications.

[0049] (3) By introducing a Bernoulli-Poisson distribution probability model in the brain community detection process, the relationship between the community structure and brain region connections is precisely constrained, significantly enhancing the adaptability and accuracy of community detection, enabling each individual to select the most suitable community structure, thus better handling the differences between individuals, and enhancing the personalization and flexibility of the analysis.

[0050] (4) A one-way message passing mechanism is adopted to simulate the directed connections between brain regions. Through this mechanism, the problem of redundant information transmission caused by undirected message passing in traditional methods can be avoided, thereby achieving a more accurate simulation of information flow and enhancing the authenticity of brain network analysis.

[0051] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Description of the Drawings

[0052] Figure 1 It is a flowchart of a brain network variable graph convolution analysis method driven by brain cognition according to the present invention. Detailed Embodiments

[0053] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention.

[0055] Embodiment

[0056] Research shows that the community structure within the brain network plays an important role in brain disease classification. When the brain is damaged, the community structure in the corresponding brain network will change significantly. Currently, clinical classification usually focuses on local modular analysis of the brain, and most brain network analysis methods also concentrate on the neighborhood of the target brain region. There are very few studies on brain diseases taking the community as a whole. Therefore, in order to make full use of brain community information, it is necessary to study brain disease classification methods based on the community structure in the brain network.

[0057] The purpose of community detection is that brain regions with high correlations will be assigned to the same community regardless of whether they are in each other's neighborhoods, while brain regions with no or low correlations will be assigned to different communities. In existing work on regional partitioning of brain networks using community detection, the mutual influence between functional and structural networks is mostly ignored, and using traditional machine learning methods will make the communities in the detection results completely independent, that is, nodes belonging to different communities are completely separated, thus blocking the message passing channels between communities, which will lead to the loss of a large amount of correlation information between nodes. On the other hand, traditional clustering methods can only utilize the structural information of the brain network during community detection and ignore the influence relationship between the brain structural network and the brain functional network.

[0058] As Figure 1 shown, a brain cognitive-driven variable graph convolutional analysis method for brain networks provided by the present invention based on the above content includes the following steps:

[0059] S1. For the DTI modal data and fMRI data of the brain network, use an overlapping and adaptive brain network community detection method based on deep learning to generate a brain region attribution matrix, design a Bernoulli-Poisson loss function to optimize the parameters of the brain region attribution matrix, and perform brain community partitioning. Specifically:

[0060] In order to make full use of the topological structure and community information of the brain network, an overlapping and adaptive brain network community detection method based on deep learning is proposed for brain community partitioning. The powerful function of the graph convolutional network in processing graph structure data is combined with community detection, and the functional attributes of the brain network are combined while using the structural information. The graph neural network is used to learn and optimize, and finally a brain region attribution matrix F is generated:

[0061]

[0062] In the formula, F is the brain region attribution matrix, ReLU(·) represents a non-linear activation function, GCN(·) represents a graph convolutional network, A is the brain region adjacency matrix constructed from DTI modal data, representing the structural information of the brain network, is the normalized brain region adjacency matrix, is the adjacency matrix after adding self-loops, is 's degree matrix, X is the brain region embedding matrix constructed from fMRI data, representing the functional information of the brain network, W (1) and W (2) are both weight matrices in the brain region attribution matrix.

[0063] Due to the significant heterogeneity among different individuals' brain networks, there is no standard community classification result as a label for training the model. Therefore, how to select the loss function for parameter optimization during the deep learning process has become an important factor affecting the community detection result. In the overlapping community detection, if nodes u and v have more common communities, then they are more likely to be connected by an edge. Therefore, the probability of an edge existing between nodes is approximately regarded as a Bernoulli-Poisson probability distribution, that is:

[0064]

[0065] In the formula, A uv is the Bernoulli-Poisson probability distribution, Bernoulli(·) represents the Bernoulli distribution, exp(·) is the exponential function, F u is the brain region membership vector of node u, is the transpose of the brain region membership vector of node v, represents the degree of common communities between nodes u and v, The larger it is, the more common communities nodes u and v have.

[0066] Based on the Bernoulli-Poisson probability distribution, the Bernoulli-Poisson loss function is obtained, which is expressed as:

[0067]

[0068] In the formula, is the Bernoulli-Poisson loss function, E is the edge set, N is the set of node pairs without edge connections, that is, the difference between the fully connected edge set and the edge set E, represents the expected value of the node pair (u, v) sampled from the distribution P E in the edge set, and log(·) is the logarithmic function.

[0069] S2. Define the adaptive community detection score to evaluate the result of S1, select the optimal brain community structure, and obtain the optimal brain region membership matrix, specifically as follows:

[0070] Due to the existence of heterogeneity among different brain networks, it is not reasonable to set a unified fixed number of communities for the brain community detection task. In order to learn the optimal community structure of different brain networks to obtain the best brain disease classification result, we comprehensively consider brain structure information, brain function information, and small-world attributes, propose a brain community detection result evaluation mechanism, and set the community detection score according to this mechanism:

[0071] Score = Cov + α·AvgCC - β·AvgCon;

[0072] Where Score is the adaptive community detection score, Cov is the community coverage. To ensure that it is not affected by the number of communities, we weighted the community conductance and clustering coefficient according to the community size. AvgCC is the weighted average community clustering coefficient, AvgCon is the weighted average community conductance, and both α and β are parameters, specifically:

[0073] 1) The community coverage measures how many connected edges can be explained by the finally detected communities through the Bernoulli-Poisson probability distribution model, expressed as:

[0074]

[0075] Where is the indicator function, z v is the adjacency vector of node v, that is, the v-th column in the membership matrix, represents the connection strength between nodes u and v, represents that there is an edge connection between nodes u and v.

[0076] 2) The community clustering coefficient quantifies the clustering degree between nodes in the graph and their connected neighbors, and the calculation formula is:

[0077]

[0078] Where R u is the number of adjacent node relationships, k u is the number of first-order neighbors of u.

[0079] The weighted average community clustering coefficient is expressed as:

[0080]

[0081] Where C i is the i-th community;

[0082] 3) The weighted average community conductance, which is used to measure the connection strength between the target community and the external community, is expressed as:

[0083]

[0084] Where Outside(C i ) represents the number of edges between the nodes in community C i and the nodes outside community C i , and Inside(·) represents the number of internal edges in the community.

[0085] The above-mentioned brain community detection and evaluation mechanism considers the structural attributes of the brain network through the community coverage Cov, the functional attributes of the brain network through the average community conductance AvgCon, and the small-world attributes of the brain network through the average community clustering coefficient AvgCC. It evaluates the results of community detection from a more comprehensive perspective, solves the problem that traditional clustering methods only divide communities based on graph structure information. At the same time, using this mechanism, the community detection results under different numbers of communities set for the same individual can be analyzed, and the best result can be selected for brain disease classification, which also solves the problem of adaptively selecting the number of communities set.

[0086] In addition, since it is not possible to determine the weight ratios of the structural, functional, and small-world attributes in different tasks, the community coverage coefficient is fixed at 1, and two parameters α and β are introduced as the coefficients of the average community conductance and the average community clustering coefficient. Grid search is used to select the most suitable parameter values for different detection tasks. The parameter range in this embodiment is 0.1 to 2.0, with a step size of 0.1.

[0087] In addition, to verify the correctness of the evaluation mechanism, the detection results generated ten times repeatedly are compared. Among them, 72.77% of the individuals obtained the same best number of communities ten times, 86.75% of the individuals obtained the same best number of communities nine times, and ablation experiments have proved that the proposed evaluation mechanism effectively improves the accuracy of disease classification.

[0088] S3. Based on the optimal brain region attribution matrix, a one-way message passing mechanism is used to establish an adaptive variable graph convolutional network based on the brain community structure for analysis, and the brain disease classification result is obtained. Specifically:

[0089] Traditional convolutional ideas have the inherent limitation of fixed receptive fields. During the convolution process, the receptive fields of convolutional kernels are almost all nodes within a certain neighborhood centered on themselves, and they cannot capture the dependencies between remote nodes. The connections between brain regions in the brain are intricate. Only calculating adjacent nodes will lead to the aggregation of redundant features and the loss of features of non-adjacent relevant nodes. Some improved graph convolution methods attempt to expand the receptive field to aggregate node information within a larger range of neighborhoods, which will cause a large amount of redundant information to be aggregated while aggregating remote node information, having an unpredictable impact on the classification result.

[0090] To address the above defects, we expand the idea of variable convolution to non-Euclidean space graph structure data, combine the receptive field with the brain community structure, and propose an adaptive variable graph convolutional network based on the brain community structure. The adaptive graph convolutional layer can adaptively adjust the receptive field according to the community structure of the brain network to aggregate the node information within the community, achieving the combination of highly relevant nodes at a distance while reducing the impact of redundant information on the nodes.

[0091] Specifically, after being selected by the brain community detection and evaluation mechanism, the optimal community structure for each individual will be selected. The optimal community structure will determine the receptive field range of the convolutional kernel during the convolution process. The connections between brain regions within a community are realized by message passing between nodes within the community, and the connections between brain regions between communities are realized by message passing between overlapping nodes between communities.

[0092] The specific process of establishing the adaptive variable graph convolutional network based on the brain community structure is as follows:

[0093] S31. Define the correlation threshold to filter out the nodes with low correlation in the optimal brain region attribution matrix and select the effective community attribution matrix. Specifically:

[0094] Optimal brain region attribution matrix The values in represent the correlation between nodes and communities. For example, F s (i, j) = k, i ∈ [0, N), j ∈ [0, K u , indicating that the correlation between brain region i of sample s and community j is k. In this embodiment, a correlation threshold ρ = 0.5 is set. If k > ρ, it is considered that brain region i belongs to community j; otherwise, it is considered that the brain region has no relation with this community. Different from the traditional method that sets whether a node belongs to a community as 0 and 1, while judging whether there is a message passing channel between nodes by whether the nodes belong to the same community, we retain the correlation k for generating the community strength matrix, regarded as the connection strength between this node and the nodes within the community. Then, a self-loop is added to each node, and the community weight matrix obtained after normalizing the strength matrix is used as the weight for message passing between nodes during the convolution process.

[0095] S32. To utilize the community structure of the brain network, take the community as the structural unit of message passing. Based on the one-way message passing mechanism, set up a message passing channel between the nodes belonging to the same community in the effective community attribution matrix to obtain the message passing matrix

[0096] where F′ i represents the attribution vector of node i in the effective community attribution matrix. Although community detection provides more refined structural information for brain network analysis, simply changing the receptive field cannot fully utilize the community information. Since there are overlapping parts between communities and there is usually a clear connection direction between brain regions, such as synaptic transmission between neurons, and most previous works use an undirected adjacency matrix as the message passing matrix, that is, the weights of mutual transmission between two nodes are the same, which obviously cannot simulate the communication process between neurons in the brain. Therefore, we redefine the message passing mechanism between nodes: determine the transmission strength between nodes in this community according to the correlation of different nodes belonging to different communities, and design the message weight matrix It is used to represent the message passing intensity between nodes. The message weight matrix is jointly determined by the association degree of nodes and the membership matrix:

[0097]

[0098] In the formula, M s [·] is the message passing matrix, where both x and y are brain regions, is the message weight matrix, and += represents the cumulative assignment operation.

[0099] That is, if brain region x belongs to community j, then x will establish a message passing channel with all other brain regions in j with an intensity equal to the association degree of x to community j. Due to the existence of overlapping communities, the same brain region may establish connections with another brain region multiple times, and the sum of these values is used as the accumulation of the message passing intensity. Generally speaking, the more common communities there are between two brain regions, the greater the message passing weight between these two brain regions. In addition, since the information represented by the information flow direction between brain regions is different, we adopt a one-way message passing mechanism, that is, the message passing weight from brain region x to brain region y is different from the message passing weight from brain region y to brain region x.

[0100] S33. Based on the message weight matrix, define the adaptive variable graph convolution:

[0101]

[0102] In the formula, H (l+1) is the node feature matrix of the (l + 1)-th layer, σ is the activation function, is the degree matrix of W (l) is the weight matrix of the adaptive variable graph convolution, ⊕ represents the tensor product, represents expanding W (l) into a tensor to include additional weight information and form a new weight matrix.

[0103] By using the variable graph convolution analysis method provided by the present invention, both an adaptive variable receptive field is achieved, and the receptive field is combined with the community structure, ensuring the rationality of the change of the receptive field and solving the problem that communities are completely independent in traditional methods.

[0104] Extensive experimental verification shows that the method of the present invention performs better than existing brain network analysis methods on epilepsy and Alzheimer's disease neuroimaging datasets (ADNI), significantly improving the classification accuracy and efficiency of brain diseases, revealing effective biomarkers, demonstrating the great potential of this method in brain disease classification, providing strong support for the technological progress and application in related fields, and having broad application prospects.

[0105] Therefore, a brain cognitive-driven variable graph convolutional analysis method for brain networks of the present invention dynamically adjusts the graph convolutional receptive field by combining the community structure of the brain network, fuses brain structure and function data, optimizes community detection using a Bernoulli-Poisson probability model, and introduces a unidirectional message passing mechanism to simulate the directed information flow between brain regions, effectively capturing the long-range dependence relationships and high-order topological features between brain regions, and significantly improving the accuracy and individualization ability of brain network analysis.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A brain network variable graph convolution analysis method driven by brain cognition, characterized in that: The following steps are involved: S1. For the DTI modality data and fMRI data of the brain network, an overlapping adaptive brain network community detection method based on deep learning is used to generate the brain region attribution matrix, and a Bernoulli-Poisson loss function is designed to optimize the parameters of the brain region attribution matrix and perform brain community division. S2, define the adaptive community detection score to evaluate the results of S1, select the most suitable brain community structure, and obtain the most suitable brain region attribution matrix; S3. Based on the optimal brain region attribution matrix, a one-way message passing mechanism was used to establish an adaptive variable graph convolutional network based on the brain community structure for analysis to obtain the brain disease classification results.

2. According to the brain cognition driven brain network variable graph convolution analysis method according to claim 1, it is characterized in that: In S1, the brain region attribution matrix is ​​expressed as: Where F is the brain region attribution matrix, ReLU(·) represents the nonlinear activation function, GCN(·) represents the graph convolutional network, and A is the brain region adjacency matrix constructed from DTI modality data, representing the structural information of the brain network. is the brain region adjacency matrix after normalization operation, X is the brain region embedding matrix constructed by fMRI data, representing the functional information of brain network, and W( 1 and W( 2 ) are all weight matrices in the brain region attribution matrix.

3. According to the brain cognition driven brain network variable graph convolution analysis method of claim 1, it is characterized in that: In S1, the Bernoulli-Poisson loss function is expressed as: In the formula, is the Bernoulli-Poisson loss function, E is the edge set, N is the set of node pairs that do not have edge connections, Denotes the distribution P from the edge set E The expected value of the node pair (u,v) sampled in u is the brain region belonging vector of node u, is the transpose of the brain region belonging vector of node v, represents the degree of common community between node u and node v, exp(·) is the exponential function, and log(·) is the logarithmic function.

4. The method of brain network variable graph convolution analysis driven by brain cognition according to claim 1, characterized in that: In S2, the adaptive community detection score is expressed as: Score=Cov+α·AvgCC-β·AvgCon; Where Score is the adaptive community detection score, Cov is the community coverage, AvgCC is the weighted average community clustering coefficient, AvgCon is the weighted average community conductivity, and α and β are parameters.

5. According to claim 4, a brain cognition-driven brain network variable graph convolution analysis method is characterized in that: The community coverage is expressed as: In the formula, is the indicator function, z v is the adjacency vector of node v, Represents the connection strength between nodes u and v; The weighted average community clustering coefficient is expressed as: In the formula, C i is the i-th community, CC(C i ) is community C i The clustering coefficient of The weighted average community conductivity is expressed as: In the formula, Outside(C i ) represents community C i Intermediate nodes and community C i The number of edges of the external nodes, Inside(·) represents the number of edges inside the community.

6. The method of brain network variable graph convolution analysis driven by brain cognition according to claim 1, characterized in that: In S3, the specific process of establishing an adaptive variable graph convolutional network based on the brain community structure is: S31, defining a correlation threshold to filter out nodes with low correlation in the optimal brain region attribution matrix, and screening out a valid community attribution matrix; S32, based on a one-way message transmission mechanism, setting a message transmission channel between nodes belonging to the same community in the effective community belonging matrix to obtain a message weight matrix; S33. Based on the message weight matrix, define an adaptive variable graph convolution.

7. The method of brain network variable graph convolution analysis driven by brain cognition according to claim 6, characterized in that: In S32, the message weight matrix is ​​expressed as: Where M s [·] is the message passing matrix, x and y are brain regions, is the message weight matrix, += represents the cumulative assignment operation, F′ i [·] represents the effective community belonging matrix.

8. The method of brain network variable graph convolution analysis driven by brain cognition according to claim 7, characterized in that: The message passing matrix is ​​expressed as:

9. The method of brain network variable graph convolution analysis driven by brain cognition according to claim 6, characterized in that: In S33, the adaptive variable graph convolution is expressed as: In the formula, H (l+1) is the node feature matrix of the l+1th layer, σ is the activation function, is the degree matrix, W (l) is the weight matrix of adaptive variable graph convolution, Represents a tensor product.

Images