Brain network image analysis method for graph contrast learning based on invariance guidance

Abstract

The invention discloses a brain network image analysis method for graph contrast learning based on invariance guidance, which comprises the following steps: firstly, constructing a functional connection network by using a complete BOLD signal, storing the functional connection network in the form of a graph structure, and then preliminarily dividing the whole graph into an invariant sub-graph and a redundant sub-graph through a trainable Bernoulli edge mask; then, the I-GCL adopts an environment generator to construct a mixed graph which takes the invariant sub-graph as a core and has obvious data distribution offset as a new environment, and the ability of the model to capture the invariant sub-graph in a diversified environment is enhanced; the graph data are used as an enhanced view set in contrast learning, and the effective feature learning capability of the GCL is enhanced. Contrast loss is used for reducing inter-graph distances containing invariant features in different environments, so that a classifier can keep consistent prediction results in various environments, and invariant learning is realized. And finally, the I-GCL completes a disease classification task only by using the features of the invariant sub-graph so as to improve the performance of the model facing a brain network data set with data distribution offset.

Description

Technical Field

[0001] The present invention relates to a modeling method of a brain connection network of functional magnetic resonance imaging (fMRI) data. In order to analyze brain disease targets by computer using fMRI data, a brain disease image analysis method based on graph contrast learning guided by invariance is designed. Background Art

[0002] Brain network analysis plays a key role in understanding the biological structure and function of complex nervous systems, contributing to the early diagnosis of neurological diseases and promoting the development of neuroscience research. Functional magnetic resonance imaging provides a non-invasive solution to capture abnormal interactions between regions of interest (ROIs) in the brain by tracking the blood oxygen level-dependent (BOLD) signal that fluctuates naturally over time. Based on fMRI data, the brain network can be modeled as a network with brain regions as nodes and the relationships between regions of interest as edges. Functional connectivity networks study the statistical dependencies between remote neurophysiological events between ROIs and quantify the activity between brain regions by measuring statistical dependencies. Studies have shown that brain diseases are often associated with abnormalities in certain functional connections. Therefore, brain functional network classification has been successfully applied to computer-aided diagnosis of many brain diseases.

[0003] Brain functional network classification methods mainly include traditional machine learning methods and deep learning methods. Traditional machine learning methods include support vector machines and logistic regression. In contrast, deep learning methods show better classification performance because they can effectively extract deeper high-level features, such as brain network classification methods based on convolutional neural networks (CNN) and graph convolution neural networks (GCN). Although these deep learning methods have achieved remarkable results, they still face some problems. Existing classification models rely heavily on the same distribution assumption, that is, the training data and test data come from the same distribution. However, most of the existing public brain imaging datasets come from multiple brain imaging centers, which use different scanners, scanning settings and protocols. In addition, studies have shown that there is significant variability between individuals in different functional systems of the brain, and these differences may affect the distribution of the data and weaken the model's ability to make correct predictions. Recently, contrastive learning has achieved remarkable performance in natural image classification. Among them, graph contrastive learning (GCL) has received widespread attention as an effective method for processing fMRI data. These GCL methods focus on identifying the most discriminative features in the graph and therefore have strong generalization ability and stability. However, some of these methods construct positive samples by randomly truncating the BOLD signal, which may lead to the loss of discriminative features. In addition, the anchor map in most methods only has one corresponding front view, which may not capture the most discriminative features well, and therefore cannot effectively deal with the data distribution offset problem in the brain network.

[0004] Therefore, how to accurately and stably complete the task of brain disease classification from brain network data with data distribution deviation is still a very challenging topic in this field. Summary of the invention

[0005] In view of the low accuracy and poor stability of fMRI data in analyzing brain disease images, a brain network classification method based on invariance-guided graph contrastive learning (I-GCL) is proposed. Among them, invariance learning is a major strategy to deal with the problem of out-of-distribution data generalization. It aims to learn the invariant relationship between features and labels across different data distributions by keeping the predictor making similar predictions in different environments. Therefore, if the invariant features in brain network data can be used to guide graph contrastive learning, it is expected to alleviate the impact of data distribution shift. First, the functional connection network is constructed using the complete BOLD signal and stored in the form of a graph structure. Then, the entire graph is preliminarily divided into invariant subgraphs and redundant subgraphs through a trainable Bernoulli edge mask. Then, I-GCL uses an environment generator to construct a mixed graph with an invariant subgraph as the core and with obvious data distribution shift as a new environment, which enhances the model's ability to capture invariant subgraphs in diverse environments. After completing this step, each acquisition will have multiple graph data representing different environments. Next, the method uses these graph data as an enhanced view set in contrastive learning, so that the anchor graph has multiple front views, enhancing the ability of GCL to learn effective features. Contrastive loss is used to reduce the distance between graphs containing invariant features in different environments so that the classifier maintains consistent prediction results in multiple environments to achieve invariant learning. Finally, I-GCL only uses the features of the invariant subgraphs to complete the disease classification task to improve the performance of the model in the face of brain network datasets with data distribution shifts.

[0006] The main idea of ​​implementing the present invention is: in view of the distribution shift phenomenon of brain disease data sets from multiple brain imaging centers, the present invention integrates invariant learning into contrastive learning, and enhances the generalization ability of the model by capturing the determined invariant subgraphs in the brain network that are stable with the labels. Specifically, on the one hand, the diversified environments generated in invariant learning are used as enhanced atlases in contrastive learning. On the other hand, contrastive learning is used to shorten the distance in the feature space between the environmental features containing the invariant subgraphs so that the prediction results of the graph data containing the invariant subgraphs remain consistent to achieve the purpose of invariant learning.

[0007] A method for analyzing brain disease images based on graph contrast learning guided by invariance, comprising the following steps:

[0008] Step 1: Data acquisition: Experiments were conducted on two segmented brain maps of the ABIDE I dataset, which is commonly used for brain network image classification. The brain network image classification performance was evaluated after the time series data was converted into brain network graph data.

[0009] Step 2: Extraction of invariant subgraphs: Based on the classification data of brain network images, we first use a convolutional neural network to generate node features, and then construct edge-level features based on this. Then, we use a trainable Bernoulli mask to preliminarily divide the entire brain network image into invariant subgraphs and redundant subgraphs.

[0010] Step 3, Diversified Environment Generation: First, the redundant subgraphs with different labels are mixed with the invariant subgraph to obtain the enhanced redundant subgraph. Then, it is spliced ​​with the invariant subgraph to obtain the mixed graph, which has obvious data distribution differences from the invariant subgraph. So far, each acquisition has four types of graph data, namely the original graph, the invariant subgraph, the redundant subgraph and the mixed graph. The four graphs are used as four different environments.

[0011] Step 4: Two-layer contrastive learning: The four graphs are regarded as multi-view graph sets in contrastive learning. First, they are passed through a weight-sharing graph encoder to obtain graph-level features, and then they are mapped to the same feature space again through a projection head. Then, a two-layer contrastive learning framework is used to improve the model's ability to identify invariant subgraphs from both individual and group levels.

[0012] Step 5: Classification of invariant sub-image features: The three types of image data containing invariant sub-images are passed through a classifier consisting of two fully connected layers to obtain classification features. The entire loss function consists of a two-layer contrast loss and a cross entropy loss in the classification task. Finally, during testing, only the classification features of the invariant sub-images are used to complete the image analysis task.

[0013] Compared with the prior art, the present invention has the following obvious advantages and beneficial effects:

[0014] (1) The present invention proposes a graph contrast learning method based on invariance guidance, which completes the disease classification task by identifying invariant subgraphs to alleviate the problems of low accuracy and instability of classification methods caused by data distribution deviation in brain network data.

[0015] (2) The present invention proposes a two-layer contrastive learning structure to capture the most discriminative features at the individual level and to capture the consistency between patient groups at the group level, thereby improving the classification performance of the model.

[0016] (3) Experiments on two real disease datasets show that the proposed method outperforms several advanced existing methods in terms of accuracy and has certain interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 : Training flow chart of the model involved in this method.

[0018] Figure 2 : Important brain area map learned by I-GCL on the ABIDE I real dataset. DETAILED DESCRIPTION

[0019] The specific implementation methods and detailed steps of the present invention are described below. The specific implementation process of the present invention is as follows: Figure 1 As shown, specifically including:

[0020] (Step 1) Data acquisition.

[0021] In order to verify the effectiveness of the model of the present invention, experiments were conducted on the public dataset ABIDE I, which is commonly used to evaluate the ASD brain network classification task. The data used in the experiment can be downloaded free of charge from the ABIDE (nitrc.org) website. The dataset contains functional magnetic resonance imaging (fMRI) and phenotypic data of 1035 participants from 17 different locations around the world, including 505 ASD patients and 530 healthy controls. The original fMRI data was preprocessed by the DPARSF program. The specific steps included removing the initial five time points, performing slice time correction, processing head movements exceeding 2 degrees, and excluding collectors with horizontal head movements exceeding 2mm. Subsequent steps also included registration, smoothing and filtering of sequence images. Finally, the brain was segmented into multiple regions of interest (ROIs) with different functions using a brain segmentation template, and the average time series of each ROI was extracted as a feature. Finally, the brain was segmented into multiple regions of interest (ROIs) with different functions using a brain segmentation template, and the average time series of each ROI was extracted as a feature. After preprocessing the raw fMRI data of the ABIDE I dataset, the present invention uses the AAL brain segmentation atlas to divide the brain into smaller regions, thereby obtaining the ABIDE I dataset. The AAL atlas contains 90 brain regions and 26 cerebellar regions, and is commonly used for brain disease analysis.

[0022] Then each acquired fMRI time series data is converted into graph data, where the edge weight a ij ∈A is obtained by calculating the Pearson correlation coefficient of the time series data between the i-th brain region and the j-th brain region. Each collected image data is recorded as The node set V = {v1,v2,…,v N} represents N brain regions, and It is a sparse matrix extracted from the top 5% values ​​of positive and negative connections in the adjacency matrix A, because the setting of the sparse graph improves the performance of the model.

[0023] (Step 2) Extracting the invariant subgraph.

[0024] Before extracting the invariant subgraph and the redundant subgraph, the node features in the graph are first processed. Considering the unique topological properties of brain network data, a 3-layer node feature extraction module is designed, where L0 is the input layer and receives the adjacency matrix A represented as N×N. L1 is a cross-type filter with 16 output channels, which updates the edge-level feature a by combining the weights of all edges sharing nodes i and j. ij , the output dimension is L2 contains a 1×N convolutional filter with 32 output channels to extract node-level features from the edge-level feature map. It is then replicated N times to form The feature map is taken and input to the L3 layer. L3 is an N×1 convolution filter with 64 output channels to extract node-level features again. After the node feature extraction module, we get The feature map of each brain region is updated as follows: where X∈R N×d , d represents the dimension of the feature vector of a brain region, that is, the final number of channels, 64 dimensions.

[0025] Based on the node features in the graph, the edge-level features are Where m i and m j are the node features of the two endpoints of this edge. The edge features are then input into a multilayer perceptron (MLP) as parameters for the subsequent generation of a Bernoulli distribution, and the edges are sampled according to the distribution to divide the invariant subgraph G v ,The formula for multi-layer perceptron processing is as follows:

[0026] edge ij =σ Sigmoid (σ ReLU (e ij W3+b3)W4+b4) (1)

[0027] The MLP layer consists of two fully connected layers with trainable weights. A linear layer, a ReLU activation function, and another linear layer with weights Finally, a sigmoid function is used to convert the values ​​to the range (0,1). Given the parameter set edge ij , based on ij ~Bernoulli (edge ij ) distribution to sample the edges. Since the sampling process is not differentiable, the Gumbel Softmax reparameterization technique is used to make it differentiable. The final result ζij This means that each edge remains in G v In the present invention, the threshold α is set to 0.9. If ζ ij is greater than 0.9, it has a high probability of being selected in G v At the same time, due to the high edge density in the brain network, a maximum edge ratio r is also controlled to ensure that the number of edges in the invariant subgraph is neither too many nor too few. ij On the basis that is greater than α, the unchanged subgraph G v The number of edges in the edge mask Inv is controlled to not exceed r% of the number of edges in the original graph G. Then, the Bernoulli mask Inv is applied to the adjacency matrix Obtain the unchanged subgraph G by multiplying the elements one by one v , Similarly, redundant subgraphs The edge mask (1-Inv) is obtained by v The unselected edges in the

[0028] Step three: Generate a diversified environment.

[0029] In order to generate v A diverse environment with data distribution bias is designed, where is a mapping function that maps a label to another different label. For example, if the kth collection is a healthy control group, its label is 1, then L(Y k ) is equal to 0 for patients. First, mix all labels Y=L(Y k ) To obtain an enhanced redundant subgraph As a new environment, the environment obtained in this way and There are obvious differences in data distribution, which promotes the ability of the contrastive learning module to capture more stable invariant subgraphs in different environments. Specifically, first select all the subgraphs with labels Y = L (Y k )of Then use the (1-ζ obtained in step 2 ij ) as the input parameter of the Gumbel-Softmax reparameterization method to determine the number of edges left in each edge. Possibility of extraction Process and extraction The process is the same as in Figure 1. Get the parameter matrix (1-ζ ij ) is expressed as (1-ζ), then the process of extracting the enhanced redundant subgraph is as follows:

[0030]

[0031] Among them, through subtraction operation, select Edges with a higher probability of being dropped will lead to redundant subgraphs after enhancement. With the invariant subgraph A larger distribution difference between them helps capture the invariant subgraphs. After determining the retention probability of the middle edge, select the first M edges to form The edge mask SS of The number of edges in Compared to the same label Generate an enhanced redundant subgraph in or use its own redundant subgraph For , more diverse and complex data distribution changes can be generated. This can help improve the stability of the model when facing diverse data distribution changes.

[0032] Then, and Combined to obtain a hybrid graph As a new environment:

[0033]

[0034] So far, each collection has four types of graph data, namely, the invariant subgraph G v , redundant subgraph G s , original graph G ori That is and the mixed graph G mix We consider these graphs as four different environments and define the graph structure of each collection as G ε , where G ε ,ε∈(v,s,mix,ori).G ε contains the invariant subgraphs and other environment graphs that have deterministic and truly predictive relationships with the labels. Next, we will use G ε It serves as input data for the two-layer contrastive learning in the next summary to enhance the model's ability to learn invariant subgraphs.

[0035] (Step 4) Two-layer contrastive learning.

[0036] Four graph data are input into a graph encoder to extract graph-level features. They first aggregate neighbor node information and update their own node features through a 3-layer weight-sharing graph convolutional network (GCN) module. Then, the global structural information of the graph is obtained through global attention pooling:

[0037] Z ε =g(G ε )=Pooling(GCN(G ε )) (4)

[0038] Each layer of the GCN network is followed by a batch normalization layer and a random dropout layer, and residual connections are used to avoid gradient vanishing or gradient exploding. In order to obtain more complete graph-level features, the node hidden representations obtained from each layer of message passing are added as the final node information. Then, global attention pooling is used to obtain global graph features. Specifically, the attention weight of each node is calculated through two fully connected layers. Corresponding weight information is assigned to the node features to highlight more important features. Finally, the weighted features are aggregated to represent the features of the entire graph.

[0039] After passing through the graph encoder, a weight-sharing projection head is used to transform the vector Z ε Mapped to the same new feature representation space, we get z ε This projection head consists of two dimension-preserving linear layers with a ReLU activation function in between. From the previous summary, we can get It is the kth multi-view graph collection. I represents all the graphs in this batch. A two-layer contrastive learning method is adopted to promote the model to further learn the invariant subgraphs at the group and individual levels. At the group level, only the invariant subgraph G v . The collectors’ z with the same label Y v is considered as a positive sample pair, while the collectors with different labels have z v Considered as negative samples, a supervised comparison method is adopted:

[0040]

[0041] A(k)=I\{k} contains samples except those of the kth collector, P(k)=p∈A(k):y p =y kis the index set of all positive samples with the same label as the kth collector in this batch, |P(k)| is its cardinality, and · represents the cosine similarity. Studies have shown that participants with the same disease have similar brain network patterns, so the group supervision contrast loss calculated by formula (5) will encourage the encoder to align the feature representations of the same group by shortening their distance in the feature space. By capturing the consistency between groups with the same label, more robust representation space features can be obtained.

[0042] At the individual level, we focus on the relationship between the four types of graph data owned by a collector. Among them, é∈(v,mix,ori) is regarded as a positive sample, because they all contain the invariant relationship with the label. And they is considered as a negative sample. The individual contrast loss calculation formula is:

[0043]

[0044] By getting closer The distance between the middle graphs is used to enhance the stability of the captured invariant subgraphs in different environments. Graph data and to strengthen the independence of the invariant subgraphs, because It may also contain information associated with the label. In this way, the maximum retention of Useful information in .

[0045] (Step 5) Classification of invariant subgraph features.

[0046] In graph invariance learning, the task is to minimize the loss of the model in multiple environments, so the supervision loss is set as:

[0047]

[0048] Among them, é∈(v,mix,ori), The graph data in all contain the invariant subgraph G v , w is a classifier with two fully connected layers, R e (w) represents the cross entropy loss on the environment é. Finally, the model completes iterative training through contrast loss and supervision loss.

[0049] In order to fully verify the superiority of the present invention, the present invention was compared with RandomForest (RF), Support Vector Machine (SVM), and Convolutional Neural Network (CNN) methods on the AAL brain atlas data on the real ABIDE I dataset, and the results were evaluated using five evaluation indicators: accuracy, precision, recall, F1score, and AUC. These six evaluation indicators are widely used in the evaluation of brain network classification. The final results are shown in Table 1 in the form of mean ± standard deviation of ten cross-validations.

[0050] Table 1: Experimental results of various methods on the ABIDE I dataset

[0051]

[0052]

[0053] As can be seen from the table above, the model is significantly higher than other comparison methods in all five evaluation indicators. Among them, based on the ABIDE I data of the AAL atlas, the proposed model has improved the accuracy indicator by about 9% and the recall indicator by 15% compared with the traditional machine learning method SVM; compared with the CNN model, it has improved the accuracy indicator by about 7% and the recall indicator by 13%. At the same time, the proposed method has a smaller standard deviation, indicating that the method has more stable performance and can cope with the distribution shift phenomenon between data from different sites by capturing the invariant subgraphs in the brain network.

[0054] In order to further verify the interpretability of the model proposed in this invention, the important brain regions are analyzed by using the obtained invariant subgraphs. The important brain regions on the two brain maps are visualized as shown in Figure 2. First, extract the important brain regions of each person, that is, the nodes in the invariant subgraph. Specifically, use the adjacency matrix of the invariant subgraph to sum the rows and columns respectively, and then take the average of the two as the importance score of each ROI (region of interest). Arrange these scores in descending order to get the top 10 important brain regions. In the AAL atlas, the top 10 brain regions are mainly concentrated in the temporal lobe and occipital lobe regions. Among them, the middle temporal gyrus (MTG.R, MTG.L) belongs to the middle temporal gyrus, which is pointed out to be one of the key areas of the "social brain" network and is more vulnerable in patients with autism spectrum disorder (ASD) than in healthy controls. At the same time, studies have shown that compared with healthy controls, ASD individuals have lower activity in the parahippocampal gyrus (PHG.R, PHG.L), which may reflect their proficiency in scene recognition and spatial navigation, and thus use fewer cognitive resources.

[0055] The above experiments show that the model I-GCL proposed in the present invention has higher accuracy and stability in the task of classifying ASD patients using fMRI data compared with other methods, and therefore has great application prospects in computer-assisted analysis of brain disease images.

Claims

1. A method for analyzing brain network images based on graph contrast learning guided by invariance, characterized by: The following steps are involved: Step 1: Data acquisition: Experiments were conducted on two segmented brain maps of the ABIDEI dataset for brain network image classification. The performance of brain network image classification was evaluated after converting the time series data into brain network graph data. Step 2: Extracting invariant subgraphs: Based on the brain network image classification data, a convolutional neural network is first used to generate node features, and then edge-level features are constructed on this basis; Then, the whole brain network graph is initially divided into invariant subgraphs and redundant subgraphs using a trainable Bernoulli mask; Step 3: Diversified environment generation: First, the redundant subgraphs with different labels are mixed with the invariant subgraph to obtain the enhanced redundant subgraph; then, the redundant subgraph is spliced ​​with the invariant subgraph to obtain the mixed graph, which has obvious data distribution differences from the invariant subgraph; so far, the collected data has four types of graph data, namely the original graph, the invariant subgraph, the redundant subgraph and the mixed graph; the four graphs are used as four different environments; Step 4: Two-layer contrastive learning: The four graphs are regarded as multi-view graph sets in contrastive learning. First, they are passed through a weight-sharing graph encoder to obtain graph-level features, and then they are mapped to the same feature space again through a projection head. Then, a two-layer contrastive learning framework is used to improve the model's ability to identify invariant subgraphs from both individual and group levels. Step 5: Classification of invariant sub-image features: The three types of image data containing invariant sub-images are passed through a classifier consisting of two fully connected layers to obtain classification features; the entire loss function is composed of a two-layer contrast loss and a cross entropy loss in the classification task; finally, during testing, only the classification features of the invariant sub-images are used to complete the image analysis task.

2. The method for analyzing brain network images based on graph contrast learning guided by invariance according to claim 1, characterized in that: In step 1, the original fMRI data were preprocessed by the DPARSF program, which included removing the initial five time points, performing slice time correction, handling head movements exceeding 2 degrees, and excluding the acquisition of brain network images with horizontal head movements exceeding 2 mm; subsequent steps also included registration, smoothing, and filtering of sequence images; finally, the brain was segmented into multiple regions of interest (ROIs) with different functions using a brain segmentation template, and the average time series of each ROI was extracted as a feature; finally, the brain was segmented into multiple regions of interest ROIs with different functions using a brain segmentation template, and the average time series of each region of interest ROI was extracted; after preprocessing the original fMRI data of the ABIDEI dataset, the AAL brain segmentation atlas was used to divide the brain into smaller regions, and the ABIDEI dataset was obtained; Then each acquired fMRI time series data is converted into graph data, where the edge weight a ij ∈A is obtained by calculating the Pearson correlation coefficient of the time series data between the i-th brain region and the j-th brain region; each collected brain network map data is recorded as The node set V = {v1,v2,…,v N } represents N brain regions, and It is a sparse matrix extracted from the first 5% of the positive and negative connections in the adjacency matrix A.

3. The method for analyzing brain network images based on graph contrast learning guided by invariance according to claim 1, characterized in that: In step 2, before extracting the invariant subgraph and the redundant subgraph, the node features in the graph are first processed; Considering the unique topological properties of brain network data, a 3-layer node feature extraction module is designed, where L0 is the input layer, receiving the adjacency matrix A represented as N×N; L1 is a cross-type filter with 16 output channels, which updates the edge-level feature a by combining the weights of all edges sharing nodes i and j. ij , the output dimension is L2 contains a 1×N convolutional filter with 32 output channels to extract node-level features from the edge-level feature map; it is then replicated N times to form The feature map is taken and input into the L3 layer; L3 is an N×1 convolution filter with 64 output channels to extract node-level features again; After the node feature extraction module, we will get The feature map of each brain region is updated as follows: where X∈R N×d , d represents the dimension of the feature vector of a brain region, i.e., the final number of channels, 64 dimensions; Based on the node features in the graph, the edge-level features are Where m i and m j are the node features of the two endpoints of this edge; the features of the edge are then input into the Multilayer Perceptron (MLP) as the parameters for the subsequent generation of the Bernoulli distribution, and the edges are sampled according to the distribution to divide the invariant subgraph G v ,The formula for multi-layer perceptron processing is as follows: edge ij =s Sigmoid (s ReLU (e ij W3+b3)W4+b4) (1) The MLP layer consists of two fully connected layers with trainable weights. A linear layer, a ReLU activation function, and another linear layer with weights Finally, a sigmoid function is used to convert the values ​​to the range (0,1); given the parameter set edge ij , based on ij ~Bernoulli (edge ij ) distribution to sample the edges. Since the sampling process is not differentiable, the GumbelSoftmax reparameterization technique is used to make it differentiable. The final result is ζ ij This means that each edge remains in G v The probability of; set the threshold α to 0.9; if ζ ij is greater than 0.9, it has a high probability of being selected in G v In the brain network, due to the high edge density, a maximum edge ratio r is also controlled to ensure that the number of edges in the invariant subgraph is neither too many nor too few; in ζ ij On the basis that is greater than α, the unchanged subgraph G v The number of edges in the edge mask Inv is controlled to not exceed r% of the number of edges in the original graph G; then, the Bernoulli mask Inv is applied to the adjacency matrix Obtain the unchanged subgraph G by multiplying the elements one by one v , Similarly, redundant subgraphs The edge mask (1-Inv) is obtained by v The unselected edges in the 4. The method for analyzing brain network images based on graph contrast learning guided by invariance according to claim 1, characterized in that: In step 3, in order to generate v A diverse environment with data distribution bias is designed, where is a mapping function that maps a label to another different label; if the kth acquisition is a healthy control group brain image, its label is 1, then L(Y k ) is equal to 0, indicating the patient's brain image; for each First, mix all labels Y=L(Y k ) To obtain an enhanced redundant subgraph As a new environment, the environment obtained in this way and There are obvious differences in data distribution, which promotes the ability of the contrastive learning module to capture more stable invariant subgraphs in different environments; specifically, first select all the subgraphs with labels Y = L (Y k )of Then use the (1-ζ obtained in step 2 ij ) as the input parameter of the Gumbel-Softmax reparameterization method to determine the number of edges left in each edge. Possibility of extraction Process and extraction The process is the same as in Get the parameter matrix (1-ζ ij ) is expressed as (1-ζ), then the process of extracting the enhanced redundant subgraph is as follows: Among them, through subtraction operation, select Edges with a higher probability of being dropped will lead to redundant subgraphs after enhancement. With the invariant subgraph A larger distribution difference between After determining the retention probability of the middle edge, select the first M edges to form The edge mask SS of The number of edges in the middle; obtained through the above operation Compared to the same label Generate an enhanced redundant subgraph in or use its own redundant subgraph In terms of data distribution, more diverse and complex changes are generated; Then, and Combined to obtain a hybrid graph As a new environment: So far, each collection has four types of graph data, namely, the invariant subgraph G v , redundant subgraph G s , original graph G ori That is and the mixed graph G mix ; We consider these graphs as four different environments and define the graph structure of each collection as G ε , where G ε ,ε∈(v,s,mix,ori);G ε contains the invariant subgraphs and other environment graphs that have deterministic and truly predictive relationships with the labels; next, we will use G ε It serves as input data for the two-layer contrastive learning in the next summary to enhance the model's ability to learn invariant subgraphs.

5. The method for analyzing brain network images based on graph contrast learning guided by invariance according to claim 1, characterized in that: In step 4, the four graph data are input into a graph encoder to extract graph-level features. First, a GCN module with a 3-layer weight-sharing graph convolutional network (GCN) is used to aggregate neighbor node information and update the own node features. Then, the global structural information of the graph is obtained through global attention pooling: Z ε =g(G ε )=Pooling(GCN(G ε )) (4) Each layer of the GCN network is followed by a batch normalization layer and a random dropout layer, and residual connections are used to avoid gradient vanishing or gradient exploding. To obtain more complete graph-level features, the node hidden representations obtained from each layer of message passing are added as the final node information. Then, global attention pooling is used to obtain global graph features. Specifically, the attention weight of each node is calculated through two fully connected layers. Corresponding weight information is assigned to node features to highlight more important features. Finally, the weighted features are aggregated to represent the features of the entire graph. After passing through the graph encoder, a weight-sharing projection head is used to transform the vector Z ε Mapped to the same new feature representation space, we get z ε The projection head consists of two dimension-preserving linear layers with a ReLU activation function in the middle. A two-layer contrastive learning method is adopted, which only focuses on the unchanged subgraph G at the group level. v ; The z of the collectors with the same label Y v Considered as positive sample pairs, z collected with different labels v Considered as negative samples, a supervised comparison method is adopted: A(k)=I\{k} contains samples except those of the kth collector, P(k)=p∈A(k):y p =y k is the index set of all positive samples in this batch with the same label as the k-th collection, |P(k)| is its cardinality, and · represents the cosine similarity; At the individual level, we focus on the relationship between the four types of graph data owned by a collector; Among them, é∈(v,mix,ori) is a positive sample, which contains the invariant relationship with the label. And they is regarded as a negative sample; the individual contrast loss calculation formula is: By getting closer The distance between the middle graphs can enhance the stability of the captured invariant subgraphs in different environments; Graph data and to strengthen the independence of the invariant subgraphs. contains the information associated with the label; in this way, the maximum retention of Useful information in .

6. The method for analyzing brain network images based on graph contrast learning guided by invariance according to claim 1, characterized in that: In step 5, the graph invariance learning task is to minimize the loss of the model in multiple environments, and the supervision loss is set as: Among them, é∈(v,mix,ori), The graph data in all contain the invariant subgraph G v , w is a classifier with two fully connected layers, R e (w) represents the cross entropy loss on the environment é; finally, iterative training is completed through contrast loss and supervision loss.

Images