A Brain Functional Connectivity Classification Method Based on Multi-Stage Graph Convolution Fusion

Abstract

The present invention proposes a brain functional connectivity classification method based on multi-stage graph convolution fusion, which aims to achieve brain functional connectivity classification by learning the features of samples in different subspaces and using graph convolution fusion. First, the original resting-state functional magnetic resonance data is processed to extract the features of the significant brain functional connectivity of each sample as the original features of each sample; secondly, the deep subspace clustering theory is used to extract the features of the original features in different subspace dimensions; in addition, the topological graph structure that can better represent the relationship between samples is self-learned by using the self-expression attribute of the data in multiple stages; finally, using the learned multi-stage features and topological graph structure, the classification result is output in the way of graph convolution fusion. The present invention not only learns multi-stage features and effectively fuses them, provides more comprehensive and complementary information at different stages for the classification decision of brain functional connectivity, and can obtain good classification results with less supervised information.

Description

Technical Field

[0001] The present invention belongs to the field of pattern recognition and relates to a method for classifying brain functional connectivity based on multi-stage graph convolution fusion. Background Art

[0002] In recent years, many studies have shown that there is a close connection between many mental diseases such as depression and brain nerve regulation and functional connectivity. Therefore, more and more non-invasive imaging technologies based on the brain nerve, such as functional magnetic resonance imaging (fMRI), diffusion tensor imaging (DTI), electroencephalogram (EEG), etc., are applied to find biological markers of some mental diseases and distinguish them in terms of brain functional connectivity. However, most of these studies use the original image features and relatively simple classification methods, ignoring the noise in the data and the hidden features of brain functional connectivity in different spaces. How to extract effective features from the image data and use a small number of samples for accurate classification of brain functional connectivity is crucial.

[0003] Since the categories within different categories of brain functional connectivity often show common correlation relationships, and the human brain nervous system is a complex interaction system. Therefore, for this complex interaction relationship data, the graph convolutional neural network is the best choice as the current network model for dealing with graph data learning tasks of complex interactions. However, due to the noise interference generated during the acquisition and processing of this type of data, traditional graph convolutional neural networks are difficult to effectively handle this type of scenario, and the graph convolutional learning model with a single branch is also easily affected by over-smoothing and losing important information. In addition, due to the large amount of professional manpower and long time period required for labeling brain functional connectivity data, how to achieve good classification performance using a small amount of label information is also the focus of the present invention. Summary of the Invention

[0004] To solve the above problems, the present invention proposes a method for classifying brain functional connectivity based on multi-stage graph convolution fusion, which extracts more significant features and further improves the classification accuracy of brain functional connectivity.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for classifying brain functional connectivity based on multi-stage graph convolution fusion, comprising the following steps:

[0007] Step 1, process the resting-state functional magnetic resonance imaging data of the samples to obtain significant brain functional connectivity features with obvious differences between two groups of samples:

[0008] 1-1. For whole-brain functional connectivity calculation using fMRI data, first, preprocess the fMRI images. Register the individual to the standard space according to the gray matter and white matter brain parcellation atlases for brain region division. Then, obtain the mean blood oxygenation level-dependent signal of each brain region, which is the signal value representing the current moment of that brain region. After that, calculate the correlation of the time series of brain region partitions based on the mean time series signals of each brain region to obtain the functional connectivity strength between different brain regions, and get the correlation matrix. Finally, to save storage space and computational complexity, take the upper triangular or lower triangular elements of the correlation matrix and flatten them into a one-dimensional vector as the whole-brain functional connectivity feature of this sample.

[0009] 1-2. Obtaining significant functional connectivity features. Select the data of significant brain regions with high discriminability. First, to verify whether the selected functional connectivity features can distinguish between two groups of samples, calculate the mean of the brain region correlation matrices of all samples in each of the two categories to represent the brain region correlation metric value of that category and conduct a discriminant analysis. After that, on the basis of discriminability, use the two-sample t-test in statistics to find the brain regions with significant differences between the two sample sets. Finally, according to the calculated t-test statistic, refer to the critical value table of the t-distribution to obtain the brain region correlation difference metric value between the two categories for all brain regions, and take the brain region correlation with high discriminability as the original feature of the sample {X i |X i ∈R d , i ∈ {1, 2,..., N}}, where N is the total number of subject samples and i is the sample number. X i is the original feature of the sample numbered i, d is the dimension, and R d is the d-dimensional Euclidean space.

[0010] Step 2. Using the sample significant functional connectivity features obtained in Step 1 as input, adopt deep subspace clustering to learn more comprehensive feature information at different stages and construct a topological graph structure, specifically including:

[0011] 2-1. Use an autoencoder for pre-training to learn the representations of stable sample features in different subspace dimensions, that is, the features of the sample at different stages. Use L fully connected layers to form symmetric encoder and decoder structures respectively. The encoder part takes the original features X of all subject samples as the feature representation in the first stage, and the output of each layer of the encoder corresponds to the input of the next layer. The decoder part realizes the feature reconstruction of the sample under the corresponding encoder stage through a symmetric structure. Since the encoder and decoder are symmetric structures, the output of the last layer of the encoder, that is, the Lth layer, is the input of the first layer of the decoder.

[0012] 2-2. Further utilize the self-expression property of the data, add a fully connected layer without an activation function at each stage of the autoencoder model for continued training, and then learn the affinity matrix C that can measure the affinity relationship between samples. l ;

[0013] 2-3. According to the learned affinity matrix C l , use the proportional quantization method to quantize the elements of the affinity matrix into 0 and 1 values according to the ratio, and then obtain the topological graph structure between samples; based on the edge connection relationship of the vertices in the graph topological structure, define the adjacency matrix A;

[0014] Step 3. According to the multi-stage features and graph structure learned in Step 2, use a graph convolutional neural network to fuse them, specifically including:

[0015] 3-1. Use graph convolution to learn the features of each stage and the graph topology; adopt frequency-domain graph convolution, take the sample features {X i , i ∈ {1,..., L}} learned at each stage and the adjacency matrix {A i , i ∈ {1,..., L}} representing the graph topological structure as inputs, and use graph convolution operations to achieve the combination of the feature space and the graph topology to extract the intra-sample and inter-sample feature information; on each branch of the graph convolutional neural network, the convolutional layer adopts a hierarchical information propagation rule based on the sample feature representation X i and the corresponding adjacency matrix A i realized, where h ∈ {1, 2,..., H}, H is the number of convolutional layers of the graph convolutional network, is a diagonal matrix, and the diagonal elements in the matrix are the trainable parameters of the convolutional layer, σ(·) represents the activation function, represents the output of the h-th layer; for each branch of the graph convolutional neural network, the final output is where c represents the number of categories;

[0016] 3-2. Fuse the features learned at each stage; fuse the results of the sample features convolved by the graph convolutional branch network corresponding to each stage to provide complementary information at different stages for the brain functional connectivity classification decision; for all the graph convolutional network branches corresponding to each stage, the prediction output of the entire network model is where g(·) represents the fusion function;

[0017] 3-3. Semi-supervised training and sample classification; the entire model is trained through the cross-entropy loss L fusion constraint, where Y ijIndicate whether sample i belongs to category j. If it belongs, it is 1; if not, it is 0. When the model as a whole tends to be stable, classify and discriminate the fusion result, and finally form the classification result of the sample brain functional connection.

[0018] Further, in the step 1-1, the correlation calculation uses the following calculation formula:

[0019]

[0020] where X = {X1, X2,..., X n} and Y = {Y1, Y2,..., Y n} represent two variables in the n time dimension, and represent the means of X and Y respectively.

[0021] Further, in the step 2-1, the calculation method of the output of each layer of the encoder is where l represents the number of the fully connected layer in the encoder, X l represents the output of the l-th layer of the encoder, represents the weight parameter of the (l + 1)-th stage, represents the bias parameter of the (l + 1)-th stage, and σ(·) represents the activation function adopted; the calculation method of the output of each layer of the decoder is where l represents the number of the fully connected layer in the decoder, represents the output of the l-th layer of the decoder, represents the weight parameter of the (l + 1)-th stage of the decoder, represents the bias parameter of the (l + X)-th stage of the decoder, and σ(·) represents the activation function adopted.

[0022] Further, in the step 2, the overall constraint of the multi-stage feature and the graph learning model is shown as the following formula:

[0023]

[0024] s.t.(diag(C l ) = 0, l ∈ {1, 2,..., L}).

[0025] Further, in the step 3-1, the activation function Relu(·) = max(0, ·) is selected.

[0026] Further, in the step 3-2, the fusion function is the mean function.

[0027] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0028] The present invention uses the rapidly developing and increasingly mature deep learning method to process brain imaging data, extracts the information that is difficult to discover hidden behind such data, and thus achieves the goal of brain functional connectivity classification. The present invention not only learns multi-stage features, but also uses subspace clustering to reduce the noise of brain functional connectivity features, and learns and effectively fuses multi-stage brain functional connectivity features through multi-branch graph convolution, providing more comprehensive and complementary information at different stages for the classification decision of brain functional connectivity, and achieving good classification results with less supervised information. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flowchart of the brain functional connectivity classification method based on multi-stage graph convolution fusion of the present invention.

[0030] Figure 2 It is a schematic diagram of the implementation process of the present invention.

[0031] Figure 3 It is the comparison experiment result of the present invention on the brain functional connectivity classification task of the Zhongda-Xinxiang dataset. DETAILED DESCRIPTION OF THE INVENTION

[0032] The following will combine specific embodiments to detail the technical solutions provided by the present invention. It should be understood that the following specific embodiments are only used to illustrate the present invention and not to limit the scope of the present invention.

[0033] The present invention proposes a brain functional connectivity classification method based on multi-stage graph convolution fusion, and achieves the purpose of brain functional connectivity classification by learning the features of samples in different subspaces and using graph convolution fusion. Its process and framework are as Figure 1 and Figure 2 shown. First, process the original resting-state functional magnetic resonance (fMRI) data, calculate the whole-brain functional connectivity using the fMRI data of each sample, and extract the significant functional connectivity features therein as the original features; secondly, use the deep subspace clustering method to learn the different-stage features of the original features in different subspace dimensions, that is, the multi-stage information in the sample feature space, and use the self-expression attribute of the data in multiple stages to self-learn the topological graph structure that can better represent the relationship between samples; finally, use the GCN to learn and fuse the features and their graph structures at different stages, and give the brain functional connectivity classification result. Specifically, the present invention includes the following steps:

[0034] Step 1, process the resting-state functional magnetic resonance imaging (fMRI) data of the samples, and obtain significant brain functional connectivity features with obvious differences between two groups of samples:

[0035] (1-1) Whole-brain functional connectivity calculation uses the whole-brain functional connectivity characteristics of gray and white matter in brain diseases. First, preprocess the fMRI images, register the individual to the standard space according to the gray and white matter brain atlases, and perform brain region division. In the specific implementation of the present invention, two common gray and white matter brain atlases are selected: the Brodmann gray matter atlas and the JHU ICBM-DTI-81 white matter atlas. Using the gray matter atlas, the gray matter region can be divided into 82 gray matter brain regions, and based on the white matter atlas, the brain can be divided into 48 white matter brain regions; then, obtain the average value of the blood oxygenation level dependent (BOLD) signal of each brain region, which is the signal value representing the current moment of the brain region; after that, according to the average time series signals of each brain region, calculate the correlation of the brain region time series to obtain the functional connectivity strength between different brain regions, and obtain the correlation matrix. Here, the Pearson Correlation analysis, which classically reflects the correlation between variables, is used for correlation calculation, and the specific calculation method is formula (1).

[0036]

[0037] Among them, X = {X1, X2,..., X n} and Y = {Y1, Y2,..., Y n} represent two variables in the n time dimension. Here, they can be regarded as the BOLD signal variables of two brain regions under the sampling time series, and represent the means of X and Y respectively. cov(X, Y) represents the covariance between X and Y, and σ X and σ Y represent the standard deviations of X and Y respectively. For all brain regions, that is, 82 gray matter brain regions and 48 white matter brain regions, a total of 130 brain regions, perform Pearson correlation analysis between pairwise brain regions, and a 130×130 brain region correlation matrix M can be obtained, which is the preliminarily extracted brain functional connectivity. Finally, in order to save storage space and computational complexity, take the upper triangular or lower triangular elements of the correlation matrix and flatten them into a one-dimensional vector as the whole-brain functional connectivity feature of this sample.

[0038] (1-2) Obtaining significant functional connectivity features, select the significant brain region data with high discrimination. First, in order to verify whether the selected functional connectivity features can distinguish two groups of samples, the first group of sets has N1 samples, and the second group of sets has N2 samples. Calculate the brain region correlation matrices {M i ∈R 130×130, the average value of {X

[0039]

[0040] where M MDD represents the average brain region correlation matrix of the first set, and M HC represents the average brain region correlation matrix of the second set. By performing discriminant analysis, it can be seen that the brain region correlation matrix is discriminative for these two groups of samples; then, on the basis of being discriminative, a two-sample t-test in statistics is used to find the functional connectivity data with significant differences in the two types of sample sets, and the specific calculation method is as shown in (3):

[0041]

[0042] where m A , m B are the means of the two types of samples respectively, and N A , N B are the numbers of the two types of samples respectively. S 2 represents the standard deviation of the samples and can be obtained through formula (4).

[0043]

[0044] where x represents any one of the two types of samples. Through the above calculations, the t-test statistics of the two sets of samples can be obtained respectively. Then, according to the degrees of freedom df and the calculated t-test statistics, the P-value is determined by referring to the t-distribution table, that is, the critical value table. Among them, the degrees of freedom df are calculated as shown in formula (5).

[0045] df = N A + N B [[ID=~39]]- 2(5)

[0046] After obtaining the P-values of the brain region correlation differences between the two groups for all brain regions, the False Discovery Rate (FDR) is used to correct the P-values, and then the brain region correlation features with truly significant differences can be found under a very low error rate. Among them, the threshold α for the P-values with significant differences is selected as the default 0.05, that is, the brain regions with P < α = 0.05 are used as the sample significant features. That is, the brain region correlations with high discriminability are used as the sample original features {X i |X i ∈ R d , i ∈ {1, 2,..., N}}, where N is the total number of subject samples, i is the sample number, X i is the original feature of the sample numbered i, d is the dimension, and R d is the d-dimensional Euclidean space.

[0047] Step 2: Using the sample significant functional connectivity features obtained in Step 1 as input, adopt deep subspace clustering to learn more comprehensive feature information at different stages and construct a topological graph structure, specifically including:

[0048] (2-1) Perform pre-training using an autoencoder, and utilize the encoding and decoding model structures therein to learn the representations of relatively stable sample features in different subspace dimensions, that is, sample features at different stages. Use L fully connected layers respectively to form symmetric encoder and decoder structures. In the encoder part, take the original features X of all subject samples as the feature representation of the first stage, that is, X = X 1 , where X is the original feature obtained in step (1-2) of each sample. The output of each layer of the encoder corresponds to the input of the next layer, and the calculation method is:

[0049]

[0050] where l represents the number of the fully connected layer in the encoder, X l represents the output of the l-th layer of the encoder, represents the weight parameter of the (l + 1)-th stage, represents the bias parameter of the (l + 1)-th stage. σ(·) represents the activation function adopted. In the decoder part, realize the feature reconstruction of the sample under the corresponding encoder stage through a symmetric structure, and the calculation method is:

[0051]

[0052] where l represents the number of the fully connected layer in the decoder, represents the output of the l-th layer of the decoder, represents the weight parameter of the (l + 1)-th stage of the decoder, represents the bias parameter of the (l + 1)-th stage of the decoder, and σ(·) represents the activation function adopted. Since the encoder and the decoder are symmetric structures, the output of the last layer of the encoder, that is, the L-th layer, is the input of the first layer of the decoder, that is

[0053] Adopt the pre-training method to make the parameters pre-trained. The number of pre-training times is epo. Use the reconstruction loss for the autoencoder model as shown in formula (8):

[0054]

[0055] where is the reconstruction of the sample original input X 1 , and is also the output of the decoder. By expecting the reconstruction output of the decoder to approximate the original input X 1, important features of the samples are extracted. W represents the set of learnable weight parameters in the pre-trained autoencoder model. When the pre-trained reconstruction loss L pre tends to be stable, the output X l of the l-th layer of the encoder, where l ∈ {1, 2,..., L}, can be used as the relatively stable features of the samples at the l-th stage.

[0056] (2-2) Learning the affinity matrix. First, in order to further utilize the self-expression property of the data, at each stage of the autoencoder model, a fully connected layer without an activation function is added to the output X l of the L layers of the encoder, where l ∈ {1, 2,..., L}, and continue training. The affinity matrix C l is introduced as the weight matrix of the fully connected layer. By requiring X l = X l C l , the affinity matrix C that can measure the relationship between samples is learned. l ;

[0057] (2-3) Quantifying the composition graph. First, according to the learned affinity matrix C l , using the proportional quantization method, the elements of the affinity matrix are quantized to 0 or 1 according to the ratio p, and then the topological graph structure between samples is obtained. Finally, based on the edge connection relationship of the vertices in the graph topological structure, the adjacency matrix A is defined.

[0058] (2-4) The overall constraints of the multi-stage feature and graph learning model are shown in equation (9).

[0059]

[0060] s.t. (diag(C l ) = 0, l ∈ {1, 2,..., L})

[0061] where λ1 and λ2 represent the hyperparameters for loss balance in the network. To ensure the sparsity of the affinity matrix C l obtained at each stage and avoid some overfitting problems, the l2 norm is added to the objective function.

[0062] Step 3, according to the multi-stage features and graph structure learned in Step 2, use the graph convolutional neural network to fuse them, specifically including:

[0063] (3-1) Use graph convolution to learn the features of each stage and the graph topology. Adopt frequency-domain graph convolution, and use the sample features {X i , i ∈ {1,..., L}} learned at each stage and the adjacency matrix {A i, i ∈ {1, ..., L}} as input, and use graph convolution operations to combine the feature space and graph topology to extract intra-sample and inter-sample feature information. On each branch of the graph convolutional neural network, the convolutional layer adopts the sample feature representation X i and the corresponding adjacency matrix A i to implement the hierarchical information propagation rule. The processing of the specific convolutional layer is shown in equation (10),

[0064]

[0065] where h ∈ {1, 2, ..., H}, and H is the number of convolutional layers of the graph convolutional network. is a diagonal matrix, and the diagonal elements in the matrix are the trainable parameters of the convolutional layer. σ(·) represents the activation function, and Relu(·) = max(0, ·) is selected here. represents the output of the h-th layer. For each branch of the graph convolutional neural network, the final output is:

[0066]

[0067] where c represents the number of classes.

[0068] (3-2) Fuse the features learned in each stage. Fuse the results of the convolution of the sample features learned by the graph convolutional branch network corresponding to each stage to provide complementary information at different stages for the classification decision of brain functional connectivity. For all the graph convolutional network branches corresponding to each stage, the prediction output of the entire network model is:

[0069]

[0070] where g(·) represents the fusion function, and the mean function is selected here.

[0071] (3-3) Semi-supervised training and sample classification. The entire model is trained by the cross-entropy loss L fusion constraint of the final small number of labels. The proportion of labeled samples is q, as follows:

[0072]

[0073] where Y ij indicates whether sample i belongs to class j. If it belongs, it is 1; if not, it is 0. When the overall model tends to be stable, the fusion result is classified and discriminated, and finally the classification result of the sample brain functional connectivity is formed.

[0074] In the present invention, in the brain functional connectivity classification task, multi-stage graph convolution fusion (MSGCFN) is adopted. Compared with existing classification methods, including DeepWalk, Deep Subspace Clustering (abbreviated as DSC), Graph Convolutional Network (abbreviated as GCN), Graph Attention Network (abbreviated as GAT), and multi-band graph convolution fusion model (MBGCFN). Three common classification evaluation metrics are used to evaluate the performance of the brain functional connectivity model: Accuracy, Specificity, and Sensitivity.

[0075] Embodiment:

[0076] Taking the data sets of Zhongda Hospital Affiliated to Southeast University and the Second Affiliated Hospital of Xinxiang Medical University as an example, the multi-stage graph convolution-based brain functional connectivity classification method of the present invention will be described below.

[0077] Experimental conditions: A computer is selected for the experiment. The configuration of this computer is Nvidia 1080Ti GPU, 11G video memory, 32G memory, 2T hard disk, and 64-bit operating system. Implementing whole-brain functional connectivity calculation and significant functional connectivity features are all realized by writing code in Matlab R2018b under the Windows10 environment. The multi-stage graph convolution fusion network model is built and implemented under Windows10 using the Python language and the Pycharm2017.3 development integration tool. The visualization and graph drawing of the later experimental results part are drawn using the Python third-party drawing library Matlibplot and the software GraphPadPrism 8.

[0078] The experimental data used in this chapter are provided by Zhongda Hospital Affiliated to Southeast University and the Second Affiliated Hospital of Xinxiang Medical University, which are abbreviated as the Zhongda-Xinxiang data set. The sample collection of this data set uses the same type of magnetic resonance imaging equipment and parameters, and a total of 218 samples are included, of which 89 are healthy controls and 129 are patients, which are two different types of brain functional connectivity samples.

[0079] The experimental parameters are set as follows: the quantization ratio p is 0.08, the encoder layer number L is 3, the pre-training threshold epo is 5000, and the training label supervision threshold q is 0.1.

[0080] Figure 3 For the sample classification comparison experimental results based on the Zhongda-Xinxiang data set.

[0081] The technical means disclosed in the solution of the present invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications are also regarded as the protection scope of the present invention.

Claims

1. A brain functional connection classification method based on multi-stage graph convolution fusion, characterized in that Including the following steps: Step 1, process the resting-state functional magnetic resonance imaging data of the samples to obtain the significant brain functional connection features between two groups of samples: 1-1. Use fMRI data to calculate the whole-brain functional connection. First, preprocess the fMRI images, register individuals to the standard space according to the gray matter and white matter brain parcellation atlases, and perform brain region division; then, obtain the average blood oxygenation level-dependent signal of each brain region, which is the signal value representing the current moment of that brain region; after that, calculate the correlation of the time series of the brain regions to obtain the functional connection strength between different brain regions and get the correlation matrix; finally, to save storage space and computational complexity, take the upper triangular or lower triangular elements of the correlation matrix and flatten them into a one-dimensional vector as the whole-brain functional connection feature of this sample; 1-2. Obtaining significant functional connection features, selecting significant brain region data with high discrimination. First, to verify whether the selected functional connection features can distinguish between two groups of samples, calculate the average of the brain region correlation matrices of all samples in the two categories respectively to represent the brain region correlation measurement value of this category and conduct a discriminant analysis. Then, on the basis of having discrimination, use the two-sample t-test in statistics to find the brain regions with significant differences between the two sample sets. The significant difference is that the brain region correlation difference measurement value meets the pre-set requirements. Finally, according to the calculated t-test statistic, refer to the critical value table of the t-distribution to obtain the brain region correlation difference measurement value between the two categories for all brain regions, and use the brain region correlation with high discrimination as the original feature of the sample {X i |X i ∈R d , i ∈ {1, 2,..., N}}, where N is the total number of subject samples, i is the sample number, X i is the original feature of the sample numbered i, d is the dimension, and R d is the d-dimensional Euclidean space; Step 2, take the significant functional connection features of the samples obtained in Step 1 as the input, and use deep subspace clustering to learn comprehensive feature information at different stages and construct a topological graph structure, specifically including: 2-1. Use an autoencoder for pre-training to learn the representation of stable sample features in different subspace dimensions, that is, the features of the samples at different stages; use L fully connected layers to form symmetric encoder and decoder structures respectively; the encoder part takes the original features X of all subject samples as the feature representation of the first stage, and the output of each layer of the encoder corresponds to the input of the next layer; the decoder part realizes the feature reconstruction of the sample under the corresponding encoder stage through the symmetric structure. Because the encoder and decoder are symmetric structures, the output of the last layer of the encoder, that is, the output of the Lth layer, is the input of the first layer of the decoder; 2-2. Further utilize the self-expression property of the data, add a fully connected layer without an activation function at each stage of the autoencoder model for continued training, and then learn the affinity matrix C that can measure the affinity relationship between samples. l ; 2-3, according to the learned affinity matrix C l , using the proportional quantization method, the elements of the affinity matrix are quantized to 0 and 1 values according to the ratio, and then the topological graph structure between samples is obtained; based on the edge connection relationship of the vertices in the topological graph structure, the adjacency matrix A is defined; Step 3, fuse the multi-stage features and the graph structure learned in Step 2 by using a graph convolutional neural network, specifically including: 3-1. Learn the characteristics of each stage and the graph topology using graph convolution; adopt frequency-domain graph convolution, and use the sample features {X i , i ∈ {1,..., L}} learned in each stage and the adjacency matrix {A i , i ∈ {1,..., L}} representing the graph topology structure as inputs, and use graph convolution operations to realize the combination of the feature space and the topological graph to extract the intra-sample and inter-sample feature information; on each branch of the graph convolutional neural network, the convolutional layer adopts the sample feature representation X i and the corresponding adjacency matrix A i to implement the hierarchical information propagation rule, where h ∈ {1, 2,..., H}, and H is the number of convolutional layers of the graph convolutional network, is a diagonal matrix, and the diagonal elements in the matrix are the trainable parameters of the convolutional layer, σ(·) represents the activation function, represents the output of the h-th layer; for each branch of the graph convolutional neural network, the final output is where c represents the number of categories; 3-2, fuse the features learned in each stage; fuse the results of convolving the sample features learned by the graph convolutional branch network corresponding to each stage, providing complementary information at different stages for the classification decision of brain functional connectivity; for all graph convolutional network branches corresponding to each stage, the predicted output of the entire network model is where g(·) represents the fusion function; 3-3, Semi-supervised training and sample classification; The entire model is trained by the cross-entropy loss L of the final small number of labels fusion constraint, where Y ij represents whether sample i belongs to class j. If it belongs, it is 1; if it does not belong, it is 0. When the overall model tends to be stable, the fusion result is classified and discriminated, and finally the classification result of the sample brain functional connection is formed.

2. The brain functional connection classification method based on multi-stage graph convolution fusion according to claim 1, wherein, In the above Step 1-1, the following calculation formula is used for the correlation calculation: Among them, X = {X1, X2,..., X n} and Y = {Y1, Y2,..., Y n} represent two variables in the n time dimensions, and represent the means of X and Y respectively.

3. The brain functional connectivity classification method based on multi-stage graph convolution fusion according to claim 1, characterized in that In the said step 2-1, the calculation method of the output of each layer of the encoder is as follows where l represents the number of the fully connected layer in the encoder, and X l represents the output of the l-th layer of the encoder, represents the weight parameter of the (l + 1)-th stage, represents the bias parameter of the (l + 1)-th stage, and σ(·) represents the activation function adopted; the calculation method of the output of each layer of the decoder is as follows where l represents the number of the fully connected layer in the decoder, represents the output of the l-th layer of the decoder, represents the weight parameter of the (l + 1)-th stage of the decoder, represents the bias parameter of the (l + 1)-th stage of the decoder, and σ(·) represents the activation function adopted.

4. The brain functional connectivity classification method based on multi-stage graph convolution fusion according to claim 3, wherein In the above Step 2, the overall constraints of the multi-stage features and the graph learning model are shown as follows: Among them, λ1 and λ2 represent the hyperparameters for loss balance in the network, and C l is the affinity matrix.

5. The brain functional connectivity classification method based on multi-stage graph convolution fusion according to claim 1, characterized in that In the above Step 3-1, the activation function is selected as Relu(·) = max(0,·).

6. The brain functional connectivity classification method based on multi-stage graph convolution fusion according to claim 1, wherein, In the above Step 3-2, the fusion function is the mean function.

Images