FMRI brain detection method based on adversarial training robust graph information bottleneck

Abstract

The invention discloses an fMRI brain detection method based on adversarial training robust graph information bottleneck. The method comprises the steps that brain scanning data of diseased and healthy people are collected and preprocessed, and then a BOLD signal is obtained; an adversarial training robust graph information bottleneck model is constructed, and a BOLD signal is used for training; human body brain scanning data to be detected are collected, preprocessed and then input into the trained algorithm, whether the current human body brain is diseased or not is judged after processing, the brain area related to the current disease is determined, and human body brain detection is achieved. According to the method, the disease detection effect can be effectively improved, the model is explained by using the extracted key node connection and node importance, the method is suitable for computer-aided medical detection of human brain diseases, and the method has the advantages of improving the accuracy of medical detection and improving the accuracy of medical detection. The improvement of the deep learning model is of great significance in the aspects of interpretability and the like in the field of medical detection, and reliable and effective technical support is provided for intelligent medical detection.

Description

Technical Field

[0001] The present invention relates to a brain detection method, in particular to an fMRI brain detection method based on adversarial training robust graph information bottleneck. Background Art

[0002] As a non-invasive technique, resting-state functional magnetic resonance imaging (rs-fMRI) can obtain blood oxygen level-dependent BOLD (Blood Oxygen Level Dependent) signals that can characterize different brain regions and is widely used to study brain functional structure. Functional connectivity (FC) established through BOLD signals has been shown to be a powerful indicator for identifying abnormal brain connections. As a quantitative method, this method can achieve accurate detection and reveal disease-specific brain changes.

[0003] In the past, advances in computing performance have driven the development of computer-assisted detection using resting-state functional magnetic resonance imaging (rs-fMRI) data, including traditional machine learning-based methods and deep learning-based methods. As deep learning-based methods have shown superior performance over traditional machine learning, they have been greatly expanded. Among them, GNN graph neural network has become an attractive framework for brain network modeling due to its powerful graph embedding ability. Compared with other methods, GNN method has the following advantages: 1) better simulate the dynamic interactions between different brain regions and capture the complex relationships in the brain functional network; 2) effectively process high-dimensional data and extract useful information through feature learning of nodes and edges; 3) build personalized brain network models based on individual fMRI data, thereby achieving more accurate detection and analysis.

[0004] However, when using GNN for auxiliary detection, GNN has poor interpretability. Existing GNN models are all black box models with complex structures, making it difficult to interpret the test results. In high-risk decision-making areas such as medical testing, unreliable results will cause negative ethical and social impacts. Another more serious problem is that due to the differences in MRI scanning instruments, the magnetic field strength, internal parameters of the instrument, and the radio frequency noise of the external environment will make the collected fMRI data mixed with noise, which will cause the adjacency matrix or feature matrix input to the GNN model to have disturbances that are difficult for people to detect. These tiny disturbances will lead the GNN to output wrong results and produce explanations that go against common sense. Summary of the invention

[0005] In order to solve the problems existing in the background technology, the present invention provides an fMRI brain detection method based on adversarial training robust graph information bottleneck. The method of the present invention establishes a graph information bottleneck by extracting subgraph features and subgraph structures in a functional connection network, and inputs the graph information bottleneck into a graph neural network to obtain human brain detection results. In order to improve the robust performance of the model, a distribution-based perturbation is added to the subgraph matrix in the graph information bottleneck and then input into the graph neural network for training again to ensure that the model produces stable and reliable results.

[0006] The technical solution adopted by the present invention is:

[0007] The fMRI brain detection method based on adversarial training robust graph information bottleneck of the present invention comprises:

[0008] Step 1: Use a magnetic resonance imaging scanner to collect brain scan data of diseased and healthy people within a preset time period and obtain blood oxygen level-dependent BOLD signals after preprocessing.

[0009] Step 2: construct an adversarial training robust graph information bottleneck model, use each blood oxygen level dependent BOLD signal to train the adversarial training robust graph information bottleneck model, and obtain a trained adversarial training robust graph information bottleneck model.

[0010] Step 3, use a magnetic resonance imaging scanner to collect the human brain scan data to be tested and perform the same preprocessing as in step 1, and then input the data into the trained adversarial training robust graph information bottleneck model. After processing, it is determined whether the current human brain is diseased and the brain area related to the current disease is determined and displayed on the display, thereby realizing human brain detection.

[0011] In the step 1, for each brain scan data, the brain scan data is first rearranged, and then timing correction is performed using head motion correction and time slice correction methods, and then spatial normalization, smoothing and linear detrending processing are performed in sequence to obtain functional magnetic resonance imaging fMRI data, and finally the functional magnetic resonance imaging fMRI data is segmented according to the anatomical automatic labeling AAL (Anatomical Automatic Labeling) template to obtain 116 regions of interest (ROIs) and their blood oxygen level-dependent BOLD signals.

[0012] In step 2, the adversarial training robust graph information bottleneck model is specifically as follows:

[0013] Step 2.1, for each human body's blood oxygen level-dependent BOLD signal, establish a graph feature matrix and a graph adjacency matrix of the blood oxygen level-dependent BOLD signal.

[0014] Step 2.2, mask the graph feature matrix to obtain the subgraph feature matrix, and use the subgraph feature matrix and the graph adjacency matrix to generate the subgraph adjacency matrix.

[0015] In step 2.3, a distribution-based adversarial sample generator is established. The subgraph adjacency matrix is ​​normalized and then eigendecomposed. The perturbation is then input into the adversarial sample generator to synthesize the adversarial sample of the normalized adjacency matrix.

[0016] In step 2.4, the adversarial sample of the normalized adjacency matrix and the subgraph feature matrix in step 2.2 are input into the graph neural network for training, and the loss function of the adversarial training robust graph information bottleneck model is constructed until the loss function converges and the training of the adversarial training robust graph information bottleneck model is completed.

[0017] In step 2.1, for each human body's blood oxygen level-dependent BOLD signal, the Pearson correlation coefficient matrix of the blood oxygen level-dependent BOLD signal is first calculated as the graph feature matrix X. d=116, then binarize the graph feature matrix X to obtain the graph adjacency matrix A.

[0018] In step 2.2, for each graph feature matrix X and graph adjacency matrix A, the graph feature matrix X is first masked using the mask matrix B to obtain the subgraph feature matrix X IB , Then the subgraph feature matrix X IB Input a fully connected layer to obtain the hidden variable Z, Z = FCN (X IB ), FCN represents the fully connected layer, and the subgraph adjacency matrix A is obtained by performing a dot product between the inner product of the latent variable Z and the graph adjacency matrix A IB ,

[0019] The mask matrix B is a binary matrix, B∈{0,1} d×d ,In order to make the mask matrix B differentiable in deep neural networks, the elements in the mask matrix B are set to Bernoulli distribution.

[0020] In step 2.3, the subgraph adjacency matrix A is IB After normalizing the graph constraints, perform eigendecomposition to obtain the eigenvalue matrix and eigenvector matrix. and Represent the subgraph adjacency matrix A respectively IB The normalized matrix and self-connection matrix of for degree matrix; U is the eigenvector matrix, Λ is the eigenvalue matrix; I is the identity matrix, and the distribution-based adversarial sample generator is a variational inference method. The eigenvalue matrix and eigenvector matrix are respectively input into the adversarial sample generator to obtain the eigenvalue perturbation δ Λ and the eigenvector perturbation δ U , then add the eigenvalue perturbation and eigenvector perturbation to the eigenvalue matrix and eigenvector matrix respectively, and finally synthesize the adversarial sample of the normalized adjacency matrix, as follows:

[0021]

[0022] in, Represents an adversarial example of a normalized adjacency matrix; Indicates adding eigenvector perturbation δ U The eigenvector matrix after Indicates adding eigenvalue perturbation δ Λ The eigenvalue matrix after is, U and Λ represent the eigenvector matrix and eigenvalue matrix respectively.

[0023] In step 2.4, the graph neural network uses GNN (Graph Neural Networks) to transform the adversarial sample of the normalized adjacency matrix and the subgraph feature matrix X IB Input into the GNN graph neural network for training, and the GNN graph neural network outputs the mean μ of the latent variable Z′ IB and variance Σ IB , and then perform Gaussian distribution and sampling to obtain the sub-latent variable Z IB , Sample represents the sampling process. Represents Gaussian distribution; finally, according to the sub-latent variable Z IB Establish a fully connected matrix from latent variables to disease detection results, as follows:

[0024]

[0025] in, Represents the disease detection result; FCN() represents the fully connected layer.

[0026] In step 2.4, the loss function L of the adversarial training robust graph information bottleneck model is as follows:

[0027]

[0028] Among them, L CE () represents the cross entropy loss function, y represents the true disease result; β represents the Lagrange multiplier, which is used to balance the cross entropy loss function L CE() and KL divergence KL(); KL() represents relative entropy or information KL divergence; represents sampling from a normal distribution.

[0029] Repeat step 2 until the loss function L converges to obtain the trained adversarial training robust graph information bottleneck model.

[0030] In the step 3, the pre-processed human brain scan data to be detected is input into the trained adversarial training robust graph information bottleneck model for processing. The trained adversarial training robust graph information bottleneck model confirms whether the human brain to be detected is diseased. When the human body is diseased, the subgraph feature matrix X is obtained through the trained adversarial training robust graph information bottleneck model. IB and the subgraph adjacency matrix A IB , the subgraph feature matrix X IB As the key nodes of the human brain area, they are connected to each other to establish a feature matrix. Each row of the feature matrix corresponds to a region of interest ROI in the brain area. The importance of each key node is obtained according to the feature matrix, as follows:

[0031]

[0032] Among them, s i and j Represents the subgraph feature matrix X IB The i-th row X IB,i and the jth row X IB,j The importance value of ||·||2 represents the bi-norm; A IB,i,j Represents the subgraph feature matrix X IB The i-th row X IB,i and the jth row X IB,j The corresponding subgraph adjacency matrix.

[0033] The higher the importance value, the more relevant the current region of interest ROI is to the current disease.

[0034] Subgraph feature matrix X IB , subgraph adjacency matrix A IB and the latent variable Z IB As the graph information bottleneck, the graph information bottleneck is used to perform interpretable analysis on the model and improve the credibility of the model.

[0035] The method of the present invention can obtain the blood oxygen level-dependent BOLD signal by dividing the collected human brain magnetic resonance images into regions and averaging the time series signals of each region of interest ROI. The functional connection matrix is ​​established using the BOLD signal, and the matrix is ​​used as the feature matrix and its binarized matrix is ​​used as the graph adjacency matrix. The two structural matrices are used as the input of the graph neural network to detect brain diseases. In order to improve the anti-interference ability and interpretability of the deep learning model, the present invention uses the graph information bottleneck method to extract subgraphs and features that are highly sensitive to disease information from the graph structure. These subgraphs provide rich explanatory information for the graph neural network model, thereby reducing redundancy and increasing the interpretability of the model. In addition, in order to improve the adaptability of the graph information bottleneck method to noise and disturbance, a robust training framework based on distributed perturbation is introduced. The framework performs eigenvalue decomposition on the normalized graph adjacency matrix of the functional brain network, and adds distribution-based perturbations to the eigenvalues ​​and eigenvectors, using an adversarial approach to ensure that the model produces stable and reliable results.

[0036] The beneficial effects of the present invention are:

[0037] 1. The present invention uses the graph information bottleneck method to detect human brain fMRI data, achieving innovation in intelligent medical detection.

[0038] 2. The present invention adopts the adversarial training method to improve the robustness of the graph information bottleneck and achieves innovation in the robustness of the model.

[0039] 3. The present invention utilizes the key subgraph features and key subgraph adjacency matrix extracted from the graph information bottleneck to effectively explain the deep learning model, making the black box model transparent, and facilitating the common sense understanding of doctors and patients for high-risk decision-making tasks such as medical testing based on deep learning models.

[0040] In summary, the method of the present invention can effectively improve the detection effect of the disease, and use the extracted key node connections and node importance to interpret the model. The method is suitable for computer-assisted medical detection of human brain diseases, and is of great significance for improving the accuracy of medical detection and improving the interpretability of deep learning models in the field of medical detection. It provides reliable and effective technical support for smart medical detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flowchart for a specific implementation of the present invention;

[0042] Figure 2 This is a schematic diagram of key node connections for patients with autism spectrum disorder (ASD) of the present invention;

[0043] Figure 3This is a schematic diagram of key node connections of a healthy control population of the present invention;

[0044] Figure 4 This is a schematic diagram of the importance weights of key nodes for patients with autism spectrum disorder (ASD) according to the present invention;

[0045] Figure 5 Schematic diagram of the importance weights of key nodes of the healthy control population of the present invention. DETAILED DESCRIPTION

[0046] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] The present invention is specifically implemented using the Autism Brain Imaging Data Exchange I (ABIDE I) dataset. The ABIDE I dataset includes 17 international datasets with a total of 1,035 subjects. Figure 1 As shown, in order to apply the human brain detection method based on the adversarial training robust graph information bottleneck model in fMRI analysis, the following steps were performed:

[0048] Step 1: Use a magnetic resonance imaging scanner to collect brain scan data of diseased and healthy people within a preset time period and obtain blood oxygen level-dependent BOLD signals after preprocessing; for each brain scan data, the brain scan data is first rearranged, and then the timing correction is performed using the head motion correction and time slice correction methods, and then spatial normalization, smoothing and linear detrending processing are performed in sequence to obtain magnetic resonance functional imaging fMRI data. In the specific implementation, the GRETNA toolbox is used for data preprocessing. The main preprocessing steps include skull stripping, slice time correction, motion correction, global average intensity normalization, interference signal regression and bandpass filtering at 0.01-0.1Hz. Finally, the magnetic resonance functional imaging fMRI data are segmented according to the anatomical automatic annotation AAL template to obtain 116 regions of interest ROIs and their blood oxygen level-dependent BOLD signals.

[0049] Step 2: Construct an adversarial training robust graph information bottleneck model, and use each blood oxygen level dependent BOLD signal to train the adversarial training robust graph information bottleneck model. The adversarial training robust graph information bottleneck model is as follows:

[0050] Step 2.1, for each human body's blood oxygen level-dependent BOLD signal, first calculate the Pearson correlation coefficient matrix of the blood oxygen level-dependent BOLD signal as the graph feature matrix X, d=116, then binarize the graph feature matrix X to obtain the graph adjacency matrix A. In specific implementation, the threshold of the binarization matrix of the binarization process is 0.4, that is, the matrix elements greater than 0.4 are set to 1, and the matrix elements less than 0.4 are set to 0, to obtain a binarized correlation coefficient matrix, which is used as the graph adjacency matrix A.

[0051] Step 2.2: For each graph feature matrix X and graph adjacency matrix A, first use the mask matrix B to mask the graph feature matrix X to obtain the subgraph feature matrix X. IB , Then the subgraph feature matrix X IB Input a fully connected layer to obtain the hidden variable Z, Z = FCN (X IB ), FCN represents the fully connected layer, and the subgraph adjacency matrix A is obtained by performing a dot product between the inner product of the latent variable Z and the graph adjacency matrix A IB ,

[0052]

[0053] The mask matrix B is a binary matrix, B∈{0,1} d×d , in order to make the mask matrix B differentiable in the deep neural network, the elements in the mask matrix B are set to Bernoulli distribution, as follows:

[0054]

[0055] Among them, B k,j represents the element in the kth row and jth column of the mask matrix B, τ represents the initial constant of the Bernoulli distribution, π k,j represents the Bernoulli distribution parameter of the element in the kth row and jth column in the mask matrix B, η represents the relaxation parameter that obeys the 0-1 distribution, η~Uniform(0,1).

[0056] Step 2.3: Establish a distribution-based adversarial sample generator and transform the subgraph adjacency matrix A IB After normalizing the graph constraints, perform eigendecomposition to obtain the eigenvalue matrix and eigenvector matrix. and Represent the subgraph adjacency matrix A respectively IB The normalized matrix and self-connection matrix of for degree matrix; U is the eigenvector matrix, Λ is the eigenvalue matrix; I is the identity matrix, and the distribution-based adversarial sample generator is a variational inference method. The eigenvalue matrix and eigenvector matrix are respectively input into the adversarial sample generator to obtain the eigenvalue perturbation δ Λ and the eigenvector perturbation δ U, then add the eigenvalue perturbation and eigenvector perturbation to the eigenvalue matrix and eigenvector matrix respectively, and finally synthesize the adversarial sample of the normalized adjacency matrix, as follows:

[0057]

[0058] in, Represents an adversarial example of a normalized adjacency matrix; Indicates adding eigenvector perturbation δ U The eigenvector matrix after Indicates adding eigenvalue perturbation δ Λ The eigenvalue matrix after is, U and Λ represent the eigenvector matrix and eigenvalue matrix respectively.

[0059] The variational inference method is used to calculate the distribution-based perturbations, including the perturbations of the eigenvector matrix and the eigenvalue matrix. The variational inference method first obtains the optimal perturbation distribution by parameterizing the posterior distribution of the perturbation and maximizing the evidence lower bound ELBO (Evidence Lower Bound). The calculation method of the evidence lower bound is as follows:

[0060]

[0061] in, Represents expectations, Denote the parameter as θ U The eigenvector perturbation distribution of Denote the parameter as θ Λ The eigenvalue perturbation distribution of U and θ Λ It can be obtained through ELBO optimization; p() represents the probability density distribution; represents the disease detection result output by the GNN graph neural network, δ U represents the perturbation of the eigenvector, δ Λ Represents the perturbation of the eigenvalue; KL() represents the relative entropy or information KL (Kullback-Leibler Divergence) divergence, which measures the difference between two distributions; and are two prior distributions, which are set as standard Gaussian distributions; by optimizing the parameters θ U and θ Λ Maximize ELBO and then use the posterior distribution and Sampling in the middle, we get the feature vector perturbation δ U and the eigenvalue perturbation δ Λ .

[0062] Step 2.4: Input the adversarial sample of the normalized adjacency matrix and the subgraph feature matrix in step 2.2 into the graph neural network for training, and construct the loss function of the adversarial training robust graph information bottleneck model until the loss function converges, and the adversarial training robust graph information bottleneck model training is completed. The graph neural network uses the GNN graph neural network to input the adversarial sample of the normalized adjacency matrix and the subgraph feature matrix X IB Input into the GNN graph neural network for training, and the GNN graph neural network outputs the mean μ of the latent variable Z′ IB and variance Σ IB , and then perform Gaussian distribution and sampling to obtain the sub-latent variable Z IB , Sample represents the sampling process. Represents Gaussian distribution; finally, according to the sub-latent variable Z IB Establish a fully connected matrix from latent variables to disease detection results, as follows:

[0063]

[0064] in, Represents the disease detection result; FCN() represents the fully connected layer.

[0065] The loss function L of the adversarial training robust graph information bottleneck model is as follows:

[0066]

[0067] Among them, L CE () represents the cross entropy loss function, y represents the true disease result; β represents the Lagrange multiplier, which is used to balance the cross entropy loss function L CE () and KL divergence KL(); KL() represents relative entropy or information KL divergence; represents sampling from a normal distribution.

[0068] Repeat step 2 until the loss function L converges to obtain the trained adversarial training robust graph information bottleneck model.

[0069] Step 3, using a magnetic resonance scanner to collect human brain scan data to be detected and perform the same preprocessing as in step 1, inputting the preprocessed human brain scan data to be detected into the trained adversarial training robust graph information bottleneck model for processing, and the trained adversarial training robust graph information bottleneck model confirms whether the human brain to be detected is diseased. When the human body is diseased, the subgraph feature matrix X is obtained through the trained adversarial training robust graph information bottleneck model. IB and the subgraph adjacency matrix A IB , the subgraph feature matrix X IBAs key nodes in the human brain, they are interconnected to establish a feature matrix, such as Figure 2 As shown in Figure 1, the key node connections of ASD patients extracted using this method are shown. The circles in the figure represent key nodes, and the horizontal lines represent the connections between nodes. Figure 3 As shown in the figure, the key node connections of the healthy control population extracted by this method are shown; each row of the feature matrix corresponds to a region of interest ROI in each brain region, and the importance of each key node is obtained according to the feature matrix, as follows:

[0070]

[0071] Among them, s i and j Represents the subgraph feature matrix X IB The i-th row X IB,i and the jth row X IB,j The importance value of ||·||2 represents the bi-norm; A IB,i,j Represents the subgraph feature matrix X IB The i-th row X IB,i and the jth row X IB,j The corresponding subgraph adjacency matrix.

[0072] Subgraph feature matrix X IB , subgraph adjacency matrix A IB and the latent variable Z IB As the graph information bottleneck, the graph information bottleneck is used to perform interpretable analysis on the model to improve the credibility of the model. It determines whether the current human brain is diseased and determines the brain area related to the current disease and displays it on the display. The higher the importance value, the more relevant the current region of interest ROI is to the current disease, and finally realizes human brain detection.

[0073] like Figure 4 As shown, it is the normalized contribution degree of ASD key nodes, which specifically shows Figure 2 The names of the key nodes in the graph and their contribution to the prediction results of the graph neural network. The height of the bar graph represents the contribution of each node. The higher the bar graph, the greater the contribution. For example Figure 4 The "left middle temporal gyrus" has a significant impact on the prediction results of ASD patients. Figure 5 As shown, it is the normalized contribution degree of key nodes in the healthy control population, which specifically shows Figure 3 The names of the key nodes and their contribution to the prediction results of the graph neural network are finally used to detect autism patients with an accuracy rate of up to 74.4%.

[0074] In summary, the human brain detection method based on the adversarial training robust graph information bottleneck model in fMRI analysis of the present invention can effectively improve the detection effect of diseases, and use the extracted key node connections and node importance to explain the model.

Claims

1. An fMRI brain detection method based on adversarial training robust graph information bottleneck, characterized in that: include: Step 1, using a magnetic resonance imaging scanner to collect brain scan data of diseased and healthy people within a preset time period and obtain blood oxygen level-dependent BOLD signals after preprocessing; Step 2, construct an adversarial training robust graph information bottleneck model, use each blood oxygen level dependent BOLD signal to train the adversarial training robust graph information bottleneck model, and obtain a trained adversarial training robust graph information bottleneck model; Step 3, use a magnetic resonance imaging scanner to collect the human brain scan data to be tested and perform the same preprocessing as in step 1, and then input the data into the trained adversarial training robust graph information bottleneck model. After processing, it is determined whether the current human brain is diseased and the brain area related to the current disease is determined and displayed on the display, thereby realizing human brain detection.

2. The fMRI brain detection method based on adversarial training robust graph information bottleneck according to claim 1, characterized in that: In the step 1, for each brain scan data, the brain scan data is first rearranged, and then timing correction is performed using head motion correction and time slice correction methods, and then spatial normalization, smoothing and linear detrending processing are performed in sequence to obtain functional magnetic resonance imaging fMRI data, and finally the functional magnetic resonance imaging fMRI data is segmented according to the anatomical automatic annotation AAL template to obtain 116 regions of interest ROIs and their blood oxygen level-dependent BOLD signals.

3. The fMRI brain detection method based on adversarial training robust graph information bottleneck according to claim 1, characterized in that: In step 2, the adversarial training robust graph information bottleneck model is specifically as follows: Step 2.1, for each human body's blood oxygen level-dependent BOLD signal, establish a graph feature matrix and a graph adjacency matrix of the blood oxygen level-dependent BOLD signal; Step 2.2, mask the graph feature matrix to obtain the subgraph feature matrix, and use the subgraph feature matrix and the graph adjacency matrix to generate the subgraph adjacency matrix; Step 2.3: Establish a distribution-based adversarial sample generator, normalize the subgraph adjacency matrix, perform eigendecomposition, and then input it into the adversarial sample generator to obtain perturbations, and then synthesize adversarial samples of the normalized adjacency matrix; In step 2.4, the adversarial sample of the normalized adjacency matrix and the subgraph feature matrix in step 2.2 are input into the graph neural network for training, and the loss function of the adversarial training robust graph information bottleneck model is constructed until the loss function converges and the training of the adversarial training robust graph information bottleneck model is completed.

4. The fMRI brain detection method based on adversarial training robust graph information bottleneck according to claim 3, characterized in that: In the step 2.1, for each human body's blood oxygen level-dependent BOLD signal, the Pearson correlation coefficient matrix of the blood oxygen level-dependent BOLD signal is first calculated as the graph feature matrix X, and then the graph feature matrix X is binarized to obtain the graph adjacency matrix A.

5. The fMRI brain detection method based on adversarial training robust graph information bottleneck according to claim 3, characterized in that: In step 2.2, for each graph feature matrix X and graph adjacency matrix A, the graph feature matrix X is first masked using the mask matrix B to obtain the subgraph feature matrix X IB , then the subgraph feature matrix X IB Input a fully connected layer to obtain the hidden variable Z, and perform dot product between the inner product of the hidden variable Z and the graph adjacency matrix A to obtain the subgraph adjacency matrix A. IB .

6. The fMRI brain detection method based on adversarial training robust graph information bottleneck according to claim 5, characterized in that: The mask matrix B is a binarized matrix, and the elements in the mask matrix B are set to Bernoulli distribution.

7. The fMRI brain detection method based on adversarial training robust graph information bottleneck according to claim 3, characterized in that: In step 2.3, the subgraph adjacency matrix A is IB After normalization, the eigenvalue matrix and eigenvector matrix are obtained by eigendecomposition. The distribution-based adversarial sample generator is a variational inference method. The eigenvalue matrix and eigenvector matrix are respectively input into the adversarial sample generator to obtain the eigenvalue perturbation δ Λ and the eigenvector perturbation δ U , then add the eigenvalue perturbation and eigenvector perturbation to the eigenvalue matrix and eigenvector matrix respectively, and finally synthesize the adversarial sample of the normalized adjacency matrix, as follows: in, Represents an adversarial example of a normalized adjacency matrix; Indicates adding eigenvector perturbation δ U The eigenvector matrix after Indicates adding eigenvalue perturbation δ Λ The eigenvalue matrix after is, U and Λ represent the eigenvector matrix and eigenvalue matrix respectively.

8. The fMRI brain detection method based on adversarial training robust graph information bottleneck according to claim 3, characterized in that: In step 2.4, the graph neural network uses the GNN graph neural network to normalize the adjacency matrix of the adversarial sample and the subgraph feature matrix X IB Input into the GNN graph neural network for training, and the GNN graph neural network outputs the mean μ of the latent variable Z′ IB and variance Σ IB , and then perform Gaussian distribution and sampling to obtain the sub-latent variable Z IB , and finally according to the sub-hidden variable Z IB Establish a fully connected matrix from latent variables to disease detection results, as follows: in, Represents disease detection results; FCN() represents a fully connected layer.

9. The fMRI brain detection method based on adversarial training robust graph information bottleneck according to claim 8, characterized in that: In step 2.4, the loss function L of the adversarial training robust graph information bottleneck model is as follows: Among them, L CE () represents the cross entropy loss function, y represents the true disease result; β represents the Lagrange multiplier; KL() represents the relative entropy or information KL divergence; represents normal distribution sampling; Repeat step 2 until the loss function L converges to obtain the trained adversarial training robust graph information bottleneck model.

10. The fMRI brain detection method based on adversarial training robust graph information bottleneck according to claim 2, characterized in that: In the step 3, the pre-processed human brain scan data to be detected is input into the trained adversarial training robust graph information bottleneck model for processing. The trained adversarial training robust graph information bottleneck model confirms whether the human brain to be detected is diseased. When the human body is diseased, the subgraph feature matrix X is obtained through the trained adversarial training robust graph information bottleneck model. IB and the subgraph adjacency matrix A IB , the subgraph feature matrix X IB As the key nodes of the human brain area, they are connected to each other to establish a feature matrix. Each row of the feature matrix corresponds to a region of interest ROI in the brain area. The importance of each key node is obtained according to the feature matrix, as follows: Among them, s i and j Represents the subgraph feature matrix X IB The i-th row X IB,i and the jth row X IB,j The importance value of ||·||2 represents the bi-norm; A IB,i,j Represents the subgraph feature matrix X IB The i-th row X IB,i and the jth row X IB,j The corresponding subgraph adjacency matrix; The higher the importance value, the more relevant the current region of interest ROI is to the current disease.

Images