Group-level fMRI brain function network analysis method based on graph convolutional neural network
By constructing a brain functional network and extracting key subgraph structures using a graph convolutional neural network-based approach, we have solved the problems of high false positive rate and threshold dependence in the differential analysis of brain functional networks in existing technologies, and achieved more stable group-level analysis.
Patent Information
- Application Number
- CN202211330953.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-28
- Publication Date
- 2026-08-25
- Estimated Expiration
- 2042-10-28
AI Technical Summary
Existing statistical testing methods have a high false positive rate and are easily affected by threshold selection when comparing differences in brain functional networks among different subject groups. They also make it difficult to utilize node and edge features simultaneously, resulting in unstable results.
We employ a graph convolutional neural network-based approach, which involves data preprocessing, constructing a brain functional network, training a GCN classification model, and using GNNExplainer to obtain the most important subgraph structures, extract node and edge features, and perform group-level differential analysis.
It achieves more accurate and stable comparison of differences in brain functional networks, can extract node and edge features more comprehensively, reduces false positive rate, and the results do not depend on threshold selection.
Smart Images
Figure CN115670427B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of brain functional network analysis, and more particularly to an fMRI brain functional network analysis method based on GCN and GCN interpretability. Background Technology
[0002] fMRI (functional magnetic resonance imaging) measures the activity of neurons in the brain. It boasts high temporal and spatial resolution and can be used to diagnose and analyze brain diseases. Brain functional networks (BFNs) are a commonly used model for analyzing neuronal activity in the brain, and fMRI data is often analyzed using BFN models. Each node in a BFN represents a different brain region, and edges represent functional connections between these regions. Graph theory properties of BFN nodes, such as node efficiency, node degree, clustering coefficient, and betweenness, can be considered as characteristics of the nodes, while the size of the edges can be considered as characteristics of the edges.
[0003] After establishing a brain functional network, statistical tests are typically used to compare differences in nodes or edges between different subject groups to analyze the neural mechanisms of brain diseases, such as between patient and normal groups, or between pre-treatment and post-treatment patient groups. A simple two-sample t-test can be used to compare differences in edge or node characteristics between different subject groups; however, due to the significant number of nodes or edges, using only a two-sample t-test can lead to a large false positive rate. In 2010, Zalesky et al. proposed a network-based statistical (NBS) method to eliminate this error. NBS first removes highly insignificant parts based on a first threshold, then performs a permutation test. If the ratio of the size of the maximally connected subgraph to the size of the subgraph in all permutations (ranked from largest to smallest) is less than a second threshold, the permutation test is passed. NBS does not perform well in all cases; its performance largely depends on the choice of threshold, which is difficult to determine in practice.
[0004] However, statistical tests have limitations; they struggle to simultaneously utilize features from both nodes and edges in brain functional networks. Furthermore, the threshold values in statistical tests can significantly influence experimental results, making the selection of an appropriate threshold a challenging task. Summary of the Invention
[0005] The technical problem this invention aims to solve is to provide a group-level fMRI brain functional network analysis method based on graph convolutional neural networks. This method can simultaneously extract node features and edge features, providing a more comprehensive comparison of the differences between two groups of brain functional networks than statistical tests, and yielding more reliable results. Furthermore, this method is stable, unlike statistical tests whose results largely depend on the choice of threshold.
[0006] To address the aforementioned technical problems, embodiments of the present invention provide a group-level fMRI brain functional network analysis method based on graph convolutional neural networks, comprising the following steps: S1: Data acquisition, obtaining raw data, including functional magnetic resonance imaging data of the brain; S2: Data preprocessing, preprocessing the functional magnetic resonance imaging data to reduce data noise, head movement, and individual differences; S3: Construct a brain function network. Build a brain function network for each subject, perform a one-sample t-test on each group of brain function networks, remove insignificant edges, and calculate the graph theory properties of the nodes as node features. S4: Establish and train a classification model based on graph convolutional neural networks, with the input being the edges and node features of the brain functional network and the output being the category of the subject. S5: Based on the interpretability of graph convolutional neural networks, find the subgraph structure that is most important for classification. This subgraph represents the brain functional connections with the greatest differences among different groups of brain functional networks and can be used for further analysis of the neural mechanisms of brain diseases.
[0007] Specifically, the data preprocessing in S2 includes: removing unstable time points, time-level correction, head motion correction, spatial standardization, delinear drift, regression covariates, filtering, and removing time points with excessive head motion.
[0008] The construction of the brain functional network in S3 includes the following steps: S31: Define the nodes of the brain functional network; S32: Extract the fMRI time series of each node: According to the definition of the node, calculate the average fMRI of all voxels of each node as the fMRI of the node; S33: Calculate the correlation coefficient to obtain the brain functional network: Calculate the Pearson correlation coefficient between all nodes based on the fMRI time series of the nodes, and use it as the value of the edge connection between the nodes to obtain the brain functional connectivity matrix.
[0009] In S3, the one-sample t-test includes performing a one-sample t-test on each of the two groups of brain functional networks to be compared, and removing insignificant connections between the two groups for each brain functional network; the graph theory properties of the computation nodes include graph theory properties such as computation node efficiency, clustering coefficient, betweenness number, and node degree as node features.
[0010] The classification model of the convolutional neural network in S4 includes an input layer, two convolutional layers, an average pooling layer, a fully connected layer, and an output layer.
[0011] The training of the graph convolutional neural network classification model in S4 includes the following steps: S41: Construct a dataset using each subject's brain functional connectivity matrix, node features, and category labels; S42: Divide the dataset into training and validation sets, and use k-fold cross-validation for training; S43: Adjust the neural network hyperparameters, batch size, and number of epochs to achieve an average training accuracy and validation accuracy of over 85%. S44: Calculate the average brain functional connectivity matrix for each class, calculate node features, construct a test set, and test the GCN classification model. If the accuracy reaches 85% or higher, save the neural network parameters, the average brain functional connectivity matrix for each class, the node features of the average brain functional connectivity matrix, and the label category of the average brain functional connectivity matrix as checkpoints.
[0012] In S5, the interpretability model of the convolutional neural network is a deep neural network. It obtains the feature weights of the GCN classification model by learning the mask of the brain functional connection matrix. The optimized mask represents the subgraph structure of the brain functional network that is most important for classification. This subgraph represents the brain functional connections with the greatest differences among different groups of brain functional networks.
[0013] Implementing the embodiments of the present invention has the following beneficial effects: The GCN-based classification model provided by the present invention can simultaneously extract edge and node features. Utilizing the interpretability of GCN, it can more comprehensively compare differences in brain functional connectivity among different categories at the group level than statistical testing methods, thus obtaining more accurate results. Furthermore, this method is stable, unlike statistical testing methods whose performance is highly dependent on the choice of threshold. Attached Figure Description
[0014] Figure 1 This is a flowchart illustrating the method of the present invention; Figure 2 This is a schematic diagram illustrating the general method of the present invention; Figure 3 This is the GCN framework diagram; Figure 4 This is a diagram of the GCN interpretation model framework. Detailed Implementation
[0015] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.
[0016] like Figure 1 , Figure 2 As shown in the figure, this invention acquires raw data, preprocesses fMRI data, constructs a brain functional network, establishes and trains a GCN-based classification model, and obtains functional connectivity differences between different groups based on GCN interpretability.
[0017] First, the fMRI data is preprocessed, including: removing unstable time points, time-layer correction, head motion correction, spatial standardization, delinear drift, regression covariates, filtering, and removing time points with excessive head motion.
[0018] Remove unstable time points: fMRI scanners require warm-up time, and the signals are unstable at the beginning of the scan, so unstable time points at the beginning are removed.
[0019] Time layer correction: When a machine collects data at a given time point, it scans layer by layer. There are time differences between different layers. The purpose of time layer correction is to make the time of different layers consistent.
[0020] Head movement correction: During fMRI scanning, head movements of the subject can cause signal changes and produce artifacts. Head movement correction involves image registration of each image in the image sequence with the target image to eliminate the effects of head movements.
[0021] Spatial normalization: Spatial normalization involves registering fMRI images to a standard MNI space. Because brain structure varies from person to person, spatial normalization can reduce individual differences and facilitate subsequent group-level analysis.
[0022] Delinear drift: Delinear drift is used to remove linear drift in a signal, which is usually caused by thermal noise such as machine heat.
[0023] Regression covariates: Regression covariates can eliminate the influence of harmful variables. Noise generated by white matter signals and cerebrospinal fluid signals can be removed by regression covariates.
[0024] Filtering: The frequency of fMRI signals is typically 0.01–0.08 Hz. High-frequency components in the signal can be eliminated or suppressed by filtering, thereby reducing image noise.
[0025] Remove time points with excessive head movement: The head movement correction step can obtain the head movement parameters at each time point, and time points with excessive head movement need to be removed.
[0026] After preprocessing, a brain functional network is constructed for each subject, resulting in a brain functional connectivity matrix. The steps are as follows: (1) Define the nodes of the brain functional network: Each node contains multiple voxels. The range of each node needs to be defined. The nodes of the brain functional network can be defined using the AAL90 template, AAL116 template, Dosenbach160 template, and Power264 template.
[0027] (2) Extract the fMRI time series of each node: According to the definition of the node, calculate the average fMRI of all voxels of each node as the fMRI of the node.
[0028] (3) Calculate the correlation coefficient to obtain the brain functional network: Calculate the Pearson correlation coefficient between all nodes based on the fMRI time series of the nodes, and use it as the value of the connection between the nodes to obtain the brain functional connection matrix, which is the brain functional network.
[0029] The preprocessing and brain functional network construction operations can be completed in one go using the Matlab plugin gretna.
[0030] Then, a one-sample t-test was performed on each of the two sets of brain functional networks being compared, removing insignificant connections between the two sets for each network. Because the Pearson correlation coefficient is rarely zero, the brain functional connectivity matrix is a dense matrix, and removing insignificant connections can reduce dimensionality.
[0031] Then, node efficiency, clustering coefficient, betweenness, and node degree are calculated as node graph theory properties as node features. The calculation of graph theory properties can also be accomplished using Gretna.
[0032] Then, train the GCN classification model, following these steps: (1) Construct a dataset using the brain functional connectivity matrix, node features, and category labels of each subject; (2) Divide the training set and validation set, and use k-fold cross-validation for training; (3) Adjust the neural network hyperparameters, batch size, and number of epochs to achieve an average training accuracy and validation accuracy of over 85%. (4) Calculate the average brain functional connectivity matrix for each class, calculate the node features, construct a test set, and test the GCN classification model. If the accuracy reaches 85% or more, save the parameters of the neural network, the average brain functional connectivity matrix for each class, the node features of the average brain functional connectivity matrix, and the label category of the average brain functional connectivity matrix as checkpoints.
[0033] The GCN classification model consists of one input layer, two convolutional layers, one average pooling layer, one fully connected layer, and one output layer, as follows: Figure 3 As shown. The function of the convolutional layer is to aggregate the features of adjacent nodes and edges, while the average pooling layer is used to represent the entire graph using the average of all nodes.
[0034] Next, GNNExplainer is chosen as the explanatory model for GCN. GNNExplainer obtains the subgraph structure that has the greatest impact on classification by maximizing mutual information. The parameters to be learned are the mask matrix of the brain functional connectivity matrix. The loss function is implemented based on the information entropy of the predicted probability, the information entropy of the mask, and the size of the mask. The framework diagram of GNNExplainer is shown below. Figure 4As shown, the specific operation steps are as follows: (1) Load the optimal checkpoint saved from the training of the GCN classification model to obtain the parameters of the neural network of the GCN classification model, the average brain functional connectivity matrix of each class, the node features of the average brain functional connectivity matrix, and the label category of the average brain functional connectivity matrix.
[0035] (2) Load the parameters of the neural network into the GCN classification model, initialize the mask, and use the average functional connectivity matrix and the node features of the average functional connectivity matrix after the mask of each class as the input of the trained GCN classification model. (3) Adjust the hyperparameters and training epochs of GNNExplainer to minimize the loss function; (4) The average brain functional connectivity matrix of each class is trained once. After training, the mask matrix is averaged. The average mask matrix represents the subgraph structure that is most important to the classification and can represent the differences in brain functional connectivity among different classes. It can be used to further analyze the neural mechanisms of brain diseases.
[0036] The above outlines the specific steps of the invention. By simultaneously extracting edge and node features through graph convolution, it provides a more comprehensive comparison of functional connectivity differences among different groups of brain functional networks compared to traditional statistical testing methods, yielding more accurate results. Furthermore, this method demonstrates stability, unlike statistical testing methods whose results largely depend on the choice of threshold.
[0037] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.
Claims
1. A group-level fMRI brain functional network analysis method based on graph convolutional neural networks, characterized in that, Includes the following steps: S1: Data acquisition, obtaining raw data, including functional magnetic resonance imaging data of the brain; S2: Data preprocessing, preprocessing the functional magnetic resonance imaging data to reduce data noise, head movement, and individual differences; S3: Construct a brain function network. Build a brain function network for each subject, perform a one-sample t-test on each group of brain function networks, remove insignificant edges, and calculate the graph theory properties of the nodes as node features. The construction of the brain functional network includes the following steps: S31: Define the nodes of the brain functional network; S32: Extract the fMRI time series of each node: According to the definition of the node, calculate the average fMRI of all voxels of each node as the fMRI of the node; S33: Calculate the correlation coefficient to obtain the brain functional network: Calculate the Pearson correlation coefficient between all nodes based on the fMRI time series of the nodes, and use it as the value of the edge connection between the nodes to obtain the brain functional connectivity matrix; The one-sample t-test includes performing a one-sample t-test on each of the two groups of brain functional networks to be compared, and removing insignificant edges between the two groups for each brain functional network; the graph theory properties of the computation nodes include graph theory properties such as computation node efficiency, clustering coefficient, betweenness, and node degree as node features. S4: Establish and train a classification model based on graph convolutional neural networks, with the input being the edges and node features of the brain functional network and the output being the category of the subject. S5: Based on the interpretability of graph convolutional neural networks, find the subgraph structure that is most important for classification. This subgraph represents the brain functional connections with the greatest differences among different groups of brain functional networks, which can be used for further analysis of the neural mechanisms of brain diseases. The interpretability model of graph convolutional neural networks is a deep neural network. The feature weights of the GCN classification model are obtained by learning the mask of the brain functional connection matrix. The optimized mask represents the subgraph structure of the brain functional network that is most important for classification. This subgraph represents the brain functional connections with the greatest differences among different groups of brain functional networks.
2. The group-level fMRI brain functional network analysis method based on graph convolutional neural networks according to claim 1, characterized in that, The data preprocessing in S2 specifically includes: removing unstable time points, time-level correction, head motion correction, spatial standardization, delinear drift, regression covariates, filtering, and removing time points with excessive head motion.
3. The group-level fMRI brain functional network analysis method based on graph convolutional neural networks according to claim 1, characterized in that, The classification model of the S4 convolutional neural network includes an input layer, two convolutional layers, an average pooling layer, a fully connected layer, and an output layer.
4. The group-level fMRI brain functional network analysis method based on graph convolutional neural networks according to claim 1, characterized in that, The training of the graph convolutional neural network classification model includes the following steps: S41: Construct a dataset using each subject's brain functional connectivity matrix, node features, and category labels; S42: Divide the dataset into training and validation sets, and use k-fold cross-validation for training; S43: Adjust the neural network hyperparameters, batch size, and number of epochs to achieve an average training accuracy and validation accuracy of over 85%. S44: Calculate the average brain functional connectivity matrix for each class, calculate node features, construct a test set, and test the GCN classification model. If the accuracy reaches 85% or higher, save the neural network parameters, the average brain functional connectivity matrix for each class, the node features of the average brain functional connectivity matrix, and the label category of the average brain functional connectivity matrix as checkpoints.