A Brain Connection Classification Method Based on Multi-Passband Graph Convolution Fusion

Abstract

The present invention provides a brain connection classification method based on multi-passband graph convolution fusion, aiming to achieve brain connection classification through multi-passband feature fusion. The present invention mainly includes two main parts: whole-brain functional connectivity image feature extraction and multi-passband graph convolution fusion network. For the functional magnetic resonance data of all subjects, based on the existing graph convolution network, the present invention introduces graph scattering convolution to fuse low-pass filtering and band-pass filtering to achieve multi-passband feature learning. Furthermore, a residual filtering module and an adversarial generation module are further adopted to reduce the influence of noise, improve the robustness to interference information, extract stable and highly discriminative features, and then achieve the target task of brain connection classification.

Description

Technical Field

[0001] The present invention relates to a brain connection classification method based on multi-passband graph convolution fusion, belonging to the technical field of pattern recognition in computer image recognition. Background Art

[0002] Mental diseases have become a common health problem in the world today. They disrupt the lives of patients and also impose burdens and adverse effects on the people around them. Therefore, the detection and diagnosis of mental diseases have received more and more research attention. However, at present, the diagnosis of mental diseases mainly includes inquiry, scale evaluation and clinical manifestation diagnosis methods. This requires a large amount of energy of professional personnel, and at the same time, it is difficult to clarify the authenticity of patients' feedback, so this method is subject to relatively subjective influences. In addition, most of the research on the diagnosis of mental diseases is based on original image features and relatively simple classification methods, ignoring the noise in the data and the hidden features in different spaces. Moreover, while the incidence of mental diseases is increasing rapidly, the diagnosis rate has not increased significantly. Therefore, how to use a small number of samples to assist in diagnosis is an urgent problem to be solved.

[0003] With the in-depth research, in recent years, relevant studies have shown that there is a close connection between mental diseases and brain nerve regulation and interaction. Therefore, more and more non-invasive imaging technologies based on the brain nerves, such as functional magnetic resonance imaging (Functional Magnetic Resonance Imaging, abbreviated as fMRI), diffusion tensor imaging (Diffusion Tensor Imaging, abbreviated as DTI), electroencephalogram (Electroencephalogram, abbreviated as EEG), etc., are applied to find biological markers of mental diseases, and then assist in diagnosis. However, how to extract effective features from the imaging data and perform accurate auxiliary diagnosis is crucial. At present, the rapidly developing and increasingly mature deep learning method processes the imaging data that can help diagnose mental diseases, extracts the information that is difficult to discover behind such data, transforms the problem of disease diagnosis into a brain connection classification problem, and then realizes the diagnosis of mental diseases. For the graph topology structure information in this non-Euclidean space, the most effective current processing method is the graph convolutional neural network. However, the current most popular frequency-domain graph convolution method has the problem of over-smoothing, that is, as the network structure deepens, the node features become difficult to distinguish, often leading to the degradation of the results of node processing tasks. Summary of the Invention

[0004] To solve the above problems, the present invention provides a brain connection classification method based on multi-passband graph convolutional fusion. On the basis of the existing graph convolutional network, graph scattering convolution is introduced to fuse low-pass filtering and band-pass filtering to achieve multi-passband feature learning. Furthermore, a residual filtering module and an adversarial generation module are further adopted to reduce the influence of noise, improve the robustness to interference information, and extract stable and highly discriminative features to achieve the target task of brain connection classification.

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

[0006] A brain connection classification method based on multi-passband graph convolutional fusion, comprising the following steps:

[0007] Step 1, perform data preprocessing for the brain connection classification task, that is, process the sample original data (fMRI) to obtain significantly different features, and obtain biological markers with obvious differences between different category samples. Specifically, it includes the following sub-steps:

[0008] 1-1, extract the functional connection strength of the whole brain (gray matter, white matter) of the subject sample based on the brain gray matter and white matter partition atlas. First, use the method of brain atlas mapping to obtain the regions of interest in the brain functional connection; then, according to the BOLD change value of the blood oxygen concentration in the mapped brain regions, calculate the phase Pearson correlation coefficient between the brain regions to obtain the correlation matrix M between the brain regions; finally, take the upper triangular or lower triangular elements of the correlation matrix M, and flatten them into a one-dimensional vector as the feature of this sample;

[0009] 1-2, use statistical methods to statistically analyze the whole brain connection features of different category samples, and obtain brain connection features with significant differences between different category samples. First, divide all samples into multiple categories, and calculate the average value of the brain region correlation matrix in step 1-1 for each category of samples to represent the brain region correlation measurement value of this category; then, use the classic two-sample t-test in statistics to find the brain regions with significant differences in different category sample sets. Calculate the t-test statistic, and then according to the degrees of freedom df = N A +N B -2 and the calculated t-test statistic, refer to the t-distribution table, that is, the critical value table to determine the P value. Finally, use the False Discovery Rate (FDR) to correct the P value, and then significant brain region correlation features with true significance can be found under a very low error rate.

[0010] Step 2, construct a graph of significant features between different category samples constructed in step 1. After processing all N samples of the brain region signal data, extract the significantly different brain region functional connection features X∈R N×d, construct the relationship matrix between samples in the way of K-nearest neighbors, that is, the graph topology structure G(V, E) between samples. Each sample serves as a node in the graph to form the set of V nodes in the graph, and the feature of each node is the feature extracted from this sample.

[0011] Step 3, taking the graph topology structure between samples constructed in Step 2 as the input, and adopting the method of multi-passband graph convolution fusion for feature extraction. The multi-passband graph convolution fusion network mainly consists of three parts: the Scattering Graph Convolutional Fusion Module (SGCFM), the Residual Filter Module (RFM), and the Generative Adversarial Module (GAM), and specifically includes the following sub-steps:

[0012] 3-1. First, use the Scattering Graph Convolutional Fusion Module to process the sample features through Fourier transform and geometric scattering transform respectively, and extract the low-pass and band-pass frequency signal data containing the main information. Then, splice and fuse the sample low-frequency and high-frequency information results obtained from the low-pass and band-pass branches. Finally, concatenate the information of each channel corresponding to each branch, which is the sample feature learned by the branch;

[0013] 3-2. Adopt a Residual Filter Module to reduce the captured spectrum to the defined cut-off frequency, and then filter out the high-frequency noise generated when fusing the band-pass and low-pass branches;

[0014] 3-3. Introduce the Generative Adversarial Module for the obtained brain region signal feature X to guide the scattering graph convolution network to better learn useful information. That is, based on the low-pass and band-pass information of the feature X obtained in Step 1, randomly sample from the normal distribution H hp ~N(0, 1) to act as the high-frequency noise information of the data. Then, input the low-pass, band-pass, and high-pass information of all sample brain region signal features X into the generator network, and reconstruct the original feature X of the sample brain region signal through the generator network.

[0015] As an improved scheme of the present invention, in Step 1-2, the threshold selection of the P-value with significant difference defaults to 0.05, that is, the brain region with P = 0.05 is used as the sample significant feature.

[0016] As an improved scheme of the present invention, in Step 3-1, the multi-passband information fusion is carried out in the way of using formula (2):

[0017]

[0018] Among them, And respectively represent the features learned by the low-pass and band-pass branches; C lp , C bp respectively represent the number of channels of the low-pass branch and the band-pass branch; the information of the channels corresponding to each branch is concatenated to form the sample features learned by the branch. For the k-th channel of the low-pass branch The calculation method is shown in formula (3):

[0019]

[0020] where σ(·) represents the non-linear activation function, and the |·| q method is adopted; for each channel of the band-pass branch, the calculation method is shown in formula (4):

[0021]

[0022] where, J k represents the geometric scattering operators of different orders and scales; the low-pass part of the scattering graph convolution fusion module and the band-pass part The information fusion of the two passband branches is defined as shown in formula (5):

[0023]

[0024] learnable parameters come from the convolutional layers of each branch; are the bias learning parameters of each branch respectively.

[0025] As an improved scheme of the present invention, in step 3-2, the residual filtering module is composed of a hyperparameter α, an adjacency matrix A, and a degree matrix D, and the specific definition is shown in formula (6):

[0026]

[0027] After the residual filtering module is used for the scattering graph convolution fusion module, the final output is shown in formula (7):

[0028]

[0029] where represents the learnable parameter weight matrix; when H l+1 is the output of the last layer, set d l+1 = 2, that is, equal to the number of classification categories; represents the learnable bias matrix.

[0030] As an improved scheme of the present invention, the specific steps for reconstructing the original features X of the sample brain region signal through the generation network in step 3-3 are shown in formula (8):

[0031]

[0032] Among them, H lp Represents the low-pass information of sample feature X, H bp Indicates that the sample feature X has all the information, H hp represents the high-pass information of the sample feature X, G(·) represents the generator network, Θ g represents the set of learnable parameters in the generator network; Represents the sample features reconstructed by the generator based on low-pass, band-pass, and high-pass information.

[0033] As an improved solution of the present invention, the adversarial generation module in step 3-3 further includes a discriminator, and the sample features reconstructed by the discriminator generator are The difference between the original feature X and the sample is used to measure the effectiveness of the low-pass and band-pass information learned by the generator and the scattering map convolution fusion module; the mean square error used to test the difference between data is used as the difference calculation function, and the specific definition is shown in formula (9):

[0034]

[0035] Here, D(·) represents the discriminator, i.e., the mean square error loss; this is used as the loss to train the generative adversarial module and guide the scattering graph convolution fusion module to learn more effective low-pass and band-pass information of the sample feature X.

[0036] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0037] This invention achieves brain connectivity classification through multi-passband feature fusion. Based on the existing graph convolutional network, graph scattering convolution is introduced to fuse low-pass and bandpass filtering to achieve multi-passband feature learning. Furthermore, a residual filtering module and an adversarial generation module are used to reduce the impact of noise, improve robustness to interference information, and extract stable, highly discriminative features to achieve the target task of brain connectivity classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the implementation process of the present invention.

[0039] Figure 2 These are comparative experimental results of the present invention on different tasks.

[0040] Figure 3 This is the ablation experiment result between modules of the present invention. DETAILED DESCRIPTION

[0041] The technical solution provided by the present invention will be described in detail below in conjunction with specific embodiments. 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.

[0042] The process and architecture of a brain connection classification method based on multi-passband graph convolution fusion proposed by the present invention are as Figure 1 shown. The specific steps include (1) extraction of whole-brain functional connectivity image features, and (2) two main parts of the multi-passband graph convolution fusion network.

[0043] 1) The main work of the whole-brain functional connectivity image feature extraction part: First, map the original brain functional magnetic resonance fMRI image data of the provided subject samples to the brain white matter atlas and gray matter atlas to obtain the brain region image data of the regions of interest (ROIs) in the brain. Then, use Pearson correlation to calculate the quantification value of the functional association degree between brain regions. Next, perform a two-sample t-test on the quantification values of the functional association degrees between brain regions of all samples, and select the brain region functional connections with significant differences for different categories of samples according to the P value. Finally, use these significantly different brain region functional connections as the input features of each sample. For specific processing details, see Steps 1 and 2.

[0044] 2) The main work of the multi-passband graph convolution fusion network part: First, the scattering graph convolution fusion module extracts the low-pass frequency band information of the sample using the graph convolution method based on Fourier transform according to the significant functional connection features of each sample brain region obtained in 1). At the same time, use the graph convolution method based on geometric scattering transform to extract the band-pass frequency band information of the sample. Then, fuse the features learned by the two passband branches to obtain the multi-passband features of the sample that fuse low-pass and band-pass information. Then, use the residual filtering module to set the cut-off frequency for the multi-passband fusion features of the sample to filter out the high-frequency noise information generated by the multi-passband fusion. At the same time, sample the normal distribution to simulate the high-pass part information of the sample, and input the low-pass, band-pass, and high-pass features of the sample into the generative adversarial module. Through the adversarial idea, guide the scattering graph convolution fusion module to better learn and extract the high-discrimination information of the sample, and at the same time improve the robustness against noise. Finally, the multi-passband fusion information of the sample passes through the residual filtering module to obtain the class prediction result, realizing the brain connection classification task. For specific details, see Step 2.

[0045] The model proposed by the present invention is different from the traditional graph convolution method. It mainly makes decisions by filtering and extracting the low-pass and band-pass two passband data fusions through frequency domain transformation. It makes full use of the multi-passband features of the data and filters out the high-frequency noise information. In addition, use the generative adversarial method to constrain and guide the multi-passband feature learning module to extract more effective information. Specifically, the present invention includes the following steps:

[0046] Step 1: Perform data preprocessing for the brain connection classification task, that is, process the original sample data to obtain significantly different features. In the present invention, functional magnetic resonance imaging (fMRI) data is used as the research object to obtain biological markers with obvious differences between different categories of samples. The specific processing method can be divided into the following two sub-steps:

[0047] 1-1. Extract the functional connection strength of the whole brain (gray matter, white matter) of the subject samples based on the brain gray matter and white matter partition atlas. First, obtain the regions of interest in the brain functional connection by means of brain atlas mapping; then, it is necessary to calculate the BOLD values between brain regions to obtain the functional correlation analysis between brain regions. The present invention uses the classical Pearson Correlation analysis that reflects the correlation between variables. The specific calculation method is as shown in formula (1):

[0048]

[0049] where X = {X1, X2,..., X n} and Y = {Y1, Y2,..., Y n} represent two variables in the n time dimension, r X,Y represents the Pearson correlation coefficient between variables X and Y, cov(X, Y) is the covariance between variables X and Y, and σ X and σ Y represent the standard deviations of X and Y respectively. In the present invention, they can be regarded as the BOLD signal variables of two brain regions under the sampling time series. Perform Pearson correlation analysis between all brain regions, that is, 82 gray matter brain regions and 48 white matter brain regions, a total of 130 brain regions, to obtain a 130×130 brain region correlation matrix M. Finally, take the upper triangular or lower triangular elements of the correlation matrix M and flatten them into a one-dimensional vector as the features of this sample;

[0050] 1-2. Use statistical methods to perform statistical analysis on the whole brain connection features of the subject samples to obtain brain connection features with significant differences between samples belonging to different categories. First, in order to verify whether the selected functional connection features can distinguish the differences between samples belonging to different categories, the present invention first divides all samples into multiple categories, and then passes all N samples through the processing method in step 1-1 to obtain the correlation matrices {M i ∈R 130 ×130 , i ∈ {1, 2,..., N}} of 130 brain regions of N samples. Next, use the classical two-sample t-test in statistics to find the brain regions with significant differences in different categories of sample sets. Then, according to the degrees of freedom df = NA +N B -2. Compare the calculated t-test statistic with the t-distribution table, i.e., the critical value table, to determine the P-value. Finally, correct the P-value using the False Discovery Rate (FDR), and then significant brain region correlation features with true significance can be found at a very low error rate. Among them, the threshold α for the P-value 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.

[0051] Step 2: Construct a graph of significant features between different categories of samples constructed in Step 1. After processing all N samples of the brain region signal data, extract the brain region functional connection features X ∈ R with significant differences N×d , and construct the relationship matrix between samples and samples in the way of K-nearest neighbors, that is, the graph topology structure G(V, E) of the samples. Each sample forms the V node set in the graph as a node in the graph, and the feature of each node is the feature extracted from this sample. The relationship matrix constructed by K-nearest neighbors serves as the E edge set in the graph, and a small number of samples contain labels. Next, based on a small number of samples with labels, learn to deduce the categories of unlabeled samples to achieve the classification of brain connections.

[0052] Step 3: Using the graph topology structure between samples constructed in Step 2 as the input, adopt the method of multi-band graph convolution fusion for feature extraction. The multi-band graph convolution fusion network mainly consists of three parts: the Scattering Graph Convolutional Fusion Module (SGCFM), the Residual Filter Module (RFM), and the Generative Adversarial Module (GAM), and specifically includes the following sub-steps:

[0053] 3-1. First, use the Scattering Graph Convolutional Fusion Module to process the sample features through Fourier transform and geometric scattering transform respectively, and extract the low-pass and band-pass frequency signal data containing the main information. Then, fuse the sample low-frequency and high-frequency information results obtained from the low-pass and band-pass branches. In order to define the hierarchical propagation rule, the present invention uses the formula (2) for multi-band information fusion.

[0054]

[0055] Among them, and respectively represent the features learned by the low-pass and band-pass branches. C lp , C bprespectively represent the number of channels of the low-pass branch and the band-pass branch. The information of the channels corresponding to each branch is concatenated to form the sample features learned by the branch. For the k-th channel of the low-pass branch The calculation method is as shown in formula (3).

[0056]

[0057] where A represents the adjacency matrix, D represents the degree matrix, and σ(·) represents the non-linear activation function. In this paper, |·| q is adopted. At the same time, each channel adopts a different propagation matrix, that is, different Fourier operators. Similarly, for each channel of the band-pass branch, the calculation method is as follows.

[0058]

[0059] where J k represents the geometric scattering operator of different orders and scales. Similarly, σ(·) adopts |·| q way. Finally, the low-pass part of the scattering graph convolution fusion module and the band-pass part The information fusion of the two passband branches is defined as shown in formula (5).

[0060]

[0061] Learnable parameters come from the convolutional layers of each branch. They are the bias learning parameters of each branch respectively. Finally, the multi-passband information of the sample is formed by concatenating the sample features learned by the two-branch convolution and used as the basis for subsequent brain connection classification.

[0062] 3-2. The low-pass branch and the band-pass branch are fused, and the multi-scale information at the node level is collected from the aggregation of information in different-order neighbor regions. These different-order neighbor regions may exhibit very different spectra. Therefore, the present invention adopts a residual filtering module to reduce the captured spectrum to the defined cut-off frequency, thereby filtering out high-frequency noise.

[0063] The residual filtering module is composed of a hyperparameter α, the adjacency matrix A, and the degree matrix D. The specific definition is shown in formula (6).

[0064]

[0065] After applying the residual filtering module to the scattering graph convolution fusion module, the final output can be expressed as shown in formula (7).

[0066]

[0067] where represents the learnable parameter weight matrix. When H l+1 is the output of the last layer, d l+1 can be set to 2, which is equal to the number of classification categories. represents the learnable bias matrix. Through the residual filtering module, the cut-off frequency can be set to filter out the high-frequency noise generated when the band-pass and low-pass branches are fused.

[0068] 3-3, an adversarial generation module is innovatively introduced for the obtained brain region signal features X to guide the scattering graph convolutional network to better learn useful information. That is, based on the low-pass and band-pass information of the feature X obtained in step 1, random sampling from the normal distribution H hp ~N(0,1) is used as the high-frequency noise information of the data. Then, the low-pass, band-pass, and high-pass information of all sample brain region signal features X is input into the generator network, and the original features X of the sample brain region signal are reconstructed through the generator network.

[0069] The specific steps are shown in formula (8).

[0070]

[0071] Among them, H lp represents the low-pass information of the sample feature X, H bp represents the band-pass information of the sample feature X, H hp represents the high-pass information of the sample feature X, G(·) represents the generator network, and Θ g represents the set of learnable parameters in the generator network. represents the sample feature reconstructed by the generator based on the low-pass, band-pass, and high-pass information. In the present invention, the generator network uses a simple three-layer fully connected network structure.

[0072] The adversarial generation module further includes a discriminator module. In the present invention, the difference between the sample feature reconstructed by the discriminator generator and the original sample feature X is used to measure the effectiveness of the low-pass and band-pass information learned by the generator and the scattering graph convolutional fusion module.

[0073] The present invention uses the mean square error for testing the difference between data as the difference calculation function, and the specific definition is shown in formula (9).

[0074]

[0075] Among them, D(·) represents the discriminator, that is, the mean square error loss. This is used as the loss to train the adversarial generation module and at the same time guide the scattering graph convolutional fusion module to learn more effective low-pass and band-pass information of the sample feature X.

[0076] To verify the effectiveness and superiority of the proposed multi-passband graph convolutional fusion model (MBGCFN) in brain connection classification, the accuracy of the current classical and popular machine learning semi-supervised classification methods was compared, and sensitivity and specificity indicators were used for multi-index measurement. The comparison methods selected included DeepWalk, Deep Subspace Clustering (DSC for short), Graph Convolutional Network (GCN for short), and Graph Attention Network (GAT for short).

[0077] Example:

[0078] Taking the data sets of Zhongda Hospital Affiliated to Southeast University and the Second Affiliated Hospital of Xinxiang Medical University as examples, the brain connection classification method of multi-passband graph convolution fusion of the present invention will be described below.

[0079] Experimental conditions: A computer was selected for the experiment. The configuration of this computer includes an Intel processor (3.4 GHz), 10 GB of random access memory, a 64-bit operating system, and the programming language used is Python 3.

[0080] The experimental data for this experiment came from Zhongda Hospital Affiliated to Southeast University and Xinxiang Hospital. Magnetic resonance imaging equipment and parameters of the same model were used to collect magnetic resonance data samples from 218 subjects, including 89 healthy controls (HC) and 129 patients with major depressive disorder (MDD). All data were from outpatients and inpatients. The HC were healthy controls recruited through the same-period advertisement, whose gender, age, and education level were matched with those of MDD patients.

[0081] All functional magnetic resonance imaging (fMRI) data were obtained using a magnetic resonance machine with the same parameters. The collected parameters are as follows. The parameters for T1 imaging were: TR was 1900 ms, TE was 2.48 ms, the number of layers was 176, the slice thickness was 1 mm, and the field of view was. The parameters for resting-state functional magnetic resonance were: TR was 2000 ms, TE was 25 ms, the field of view was, the number of layers was 36, the thickness was 3 mm, and the number of time points was 240. Then, the fMRI data taken from each subject was preprocessed: motion correction, slice time correction, linear correction, registration with T1 and functional images, and regression denoising of interference signals. Finally, the original sample features of the subjects used in the present invention were obtained.

[0082] The experimental parameters are set as follows: To fairly compare the performance of different methods, all methods use the sample features after processing the brain region signal data in Step 1, that is, the 733-dimensional feature of functional connections with significantly high discriminability for white matter - white matter, gray matter - gray matter, and gray matter - white matter. At the same time, for the methods with a pre-given graph topology, the K-nearest neighbor method is used to calculate the graph topology with 10% of the total number of all samples, which is 218, as the number of neighbors. All methods are semi-supervised learning classification methods, and all use 10% of the labeled samples for learning. All comparison methods use the hyperparameters set in the original paper.

[0083] Figure 2 The results of comparative experiments on different tasks. Figure 3 This is the module ablation experiment of the method proposed by the present invention. Obviously, the method of the present invention is superior to the existing methods in terms of accuracy, specificity, and sensitivity. And the ablation experiment proves that each module of the method proposed by the present invention plays a role in brain connection classification.

[0084] The technical means disclosed in the solution of the present invention are not limited to the technical means 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 in this technical field, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A brain connection classification method based on multi-passband graph convolution fusion, characterized in that It includes the following steps: Step 1: Perform data preprocessing for the brain connection classification task, that is, process the original fMRI data of the samples to obtain features with significant differences, and obtain biological markers with significant differences between different categories of samples. Specifically, it includes the following sub-steps: 1-1: Extract the functional connection strengths of the whole-brain gray matter and white matter of the subject samples based on the gray matter and white matter partition atlas of the brain. First, obtain the regions of interest in the brain functional connection by means of brain atlas mapping; then, calculate the phase Pearson correlation coefficient between brain regions based on the BOLD change values of the brain regions obtained after mapping to obtain the correlation matrix M between brain regions; finally, take the upper triangular or lower triangular elements of the correlation matrix M and flatten them into a one-dimensional vector as the features of this sample. 1-2. Use statistical methods to perform statistical analysis on the whole-brain connection features of samples of different categories, and obtain the functional connection features of brain regions with significant differences between different categories of samples. First, divide all samples into multiple categories, and calculate the average value of the brain region correlation matrix in step 1-1 for samples of multiple categories to represent the brain region correlation metric value of this category. Then, use a two-sample t-test to find brain regions with significant differences in different categories of sample sets. Calculate the t-test statistic of the sample, and then according to the degrees of freedom df = N A + N B - 2 and the calculated t-test statistic, determine the P-value by referring to the t-distribution table, that is, the critical value table. Finally, use the false discovery rate FDR to correct the P-value, so as to find the functional connection features of brain regions with significant differences under a very low error rate. The significant difference means that the P-value meets the threshold requirements set in advance; Step 2, construct a brain region functional connectivity feature map with significant differences between different categories of samples constructed in Step 1: After processing all N samples of brain region signal data, extract the brain region functional connectivity feature X ∈ R with significant differences N ×d , construct a relationship matrix between samples through the K-nearest neighbor method, that is, the graph topology structure G(V, E) between samples; each sample serves as a node in the graph to form the V node set in the graph, and the feature of each node is the feature extracted from this sample; Step 3: Using the graph topological structure between samples constructed in Step 2 as the input, perform feature extraction by means of multi-band graph convolution fusion. The multi-band graph convolution fusion network consists of three parts: a scattering graph convolution fusion module SGCFM, a residual filtering module RFM, and a generative adversarial module GAM. Specifically, it includes the following sub-steps: 3-1: First, use the scattering graph convolution fusion module to process the sample features through Fourier transform and geometric scattering transform respectively, and extract the low-pass and band-pass frequency signal data containing information; then, splice and fuse the sample low-frequency and high-frequency information results obtained from the low-pass and band-pass branches; finally, concatenate the information of each channel corresponding to each branch to obtain the sample features learned by the branch. 3-2: Use a residual filtering module to reduce the captured spectrum to the defined cut-off frequency, thereby filtering out the high-frequency noise generated when the band-pass and low-pass branches are fused. 3-3. Introduce an adversarial generation module for the obtained brain region functional connection feature X to guide the scattering graph convolutional network to better learn useful information; that is, based on the low-pass and band-pass information of the sample features obtained in step 3-1, randomly sample from the normal distribution H hp ~N(0,1) to act as the high-frequency noise information of the data; then, input the low-pass, band-pass, and high-pass information of all sample features into the generator network, and reconstruct the original features of the sample brain region signals through the generator network.

2. The multi-passband graph convolutional fusion-based brain connection classification method according to claim 1, wherein In the above Step 1-2, the threshold selection for the P value with significant differences defaults to 0.05, that is, the brain regions with P = 0.05 are used as the significant features of the samples.

3. The brain connection classification method based on multi-passband graph convolution fusion according to claim 1, characterized in that, In the above Step 3-1, the multi-band information fusion is performed in the way of formula (2): Among them, and respectively represent the features learned by the low-pass and band-pass branches; C lp , C bp respectively represent the number of channels of the low-pass branch and the band-pass branch; the information of the channels corresponding to each branch is concatenated to obtain the sample features learned by the branch. For the k-th channel of the low-pass branch The calculation method is shown in formula (3): where σ(·) represents a non-linear activation function, and |·| is adopted q For each channel of the band-pass branch, the calculation method is as shown in formula (4): Among them, J k represents geometric scattering operators of different orders and scales; the low-pass part of the scattering diagram convolution fusion module and the band-pass part The information fusion of the two passband branches is defined as shown in formula (5): Learnable parameters Convolutional layers from each branch; Bias learning parameters of each branch, respectively.

4. The multi-passband graph convolutional fusion-based brain connection classification method according to claim 3, wherein, In the above Step 3-2, the residual filtering module consists of a hyperparameter α, an adjacency matrix A, and a degree matrix D, and the specific definition is shown in formula (6): Among them, M res (α) is the residual filtering module. After using the residual filtering module in the scattering graph convolution fusion module, the final output is expressed as in formula (7): H l+1 = M res (α)H l Θ res + B res (7) where represents a learnable parameter weight matrix; when H l+1 is the output of the last layer, set d l+1 = 2, which is equal to the number of classification categories; represents a learnable bias matrix.

5. The multi-passband graph convolutional fusion-based brain connection classification method according to claim 1, characterized in that, In the above Step 3-3, the specific steps to reconstruct the original features X of the sample brain region signals through the generative network are shown in formula (8): Among them, H lp represents the low-pass information of the sample features, and H bp represents the band-pass information of the sample features, and H hp represents the high-pass information of the sample features. G(·) represents the generator network, and Θ g represents the set of learnable parameters in the generator network; represents the sample features reconstructed by the generator based on the low-pass, band-pass, and high-pass information.

6. The multi-passband graph convolutional fusion-based brain connection classification method according to claim 1, wherein, In step 3-3, the adversarial generation module further includes a discriminator, and the difference degree between the sample features reconstructed by the discriminator and the original sample features X is used to measure the effectiveness of the low-pass and band-pass information learned by the generator and the scattering graph convolution fusion module; the mean square error used to test the difference between data is used as the difference degree calculation function, and the specific definition is shown in formula (9): Among them, D(·) represents the discriminator, that is, the mean square error loss; this is used as the loss to train the generative adversarial module, and at the same time guide the scattering graph convolution fusion module to learn more effective low-pass and band-pass information of the sample features X.

Images