FMRI data analysis model construction method and device based on two-stage graph neural network

Abstract

The invention discloses an fMRI data analysis model construction method and device based on a two-stage graph neural network, and relates to the field of computer-aided diagnosis. According to the method, firstly, a brain network diagram is constructed through preprocessing of fMRI signals, and brain function connection information is captured; embedding vector learning by using the brain network features in the first stage, extracting connection features related to diseases, and providing input for subsequent prediction tasks; secondly, a group graph structure with conditional constraints is constructed, and interference of indirect factors on feature representation is inhibited; finally, phenotype data information is introduced, individual feature expression is enhanced, and the model prediction accuracy is improved; according to the model, through staged training of a brain network level GNN and a group level GNN, complementary advantages of the brain network level GNN and the group level GNN in MDD data analysis are fully exerted. According to the model, the generalization ability and the global reasoning ability of the model are remarkably improved, and an accurate analysis result of brain region data is supported.

Description

Technical Field

[0001] The present invention relates to the field of computer-aided diagnosis, and in particular to a method and device for constructing an fMRI data analysis model based on a two-stage graph neural network. Background Art

[0002] Major Depressive Disorder (MDD), the second-leading psychiatric disorder in terms of disability worldwide, has long relied on subjective assessments based on the Diagnostic and Statistical Manual of Mental Disorders (DSM) symptom scale for diagnosis, lacking objective biomarkers. Functional Magnetic Resonance Imaging (fMRI) technology, which quantifies brain functional connectivity (FC) through blood oxygenation level-dependent (BOLD) signals, offers a potential avenue for analyzing objective MDD data. However, traditional machine learning methods struggle to effectively analyze the nonlinear dynamic connectivity characteristics of brain regions and are limited by the analytical paradigm of single-modal data.

[0003] In recent years, Graph Neural Networks (GNNs) have been widely used in the analysis and research of neuropsychiatric diseases due to their advantages in processing graph-structured data. In brain network-level modeling, GNNs can effectively characterize the non-Euclidean structure and high-order interactions in the brain's functional connections through the feature propagation mechanism between nodes, demonstrating a powerful representation of the functional patterns of brain regions. At the same time, researchers have also expanded GNNs to a group-level modeling paradigm. In this framework, each node represents a subject, and the node features are derived from the individual's imaging, behavioral, or phenotypic data. The edges reflect the similarities or shared statistical properties between individuals, allowing GNNs to explore the potential structural features in the group and improve the accuracy and generalization of classification. These two levels of GNN methods have been widely used in auxiliary diagnosis tasks for neurological diseases, achieving performance that is superior to traditional methods.

[0004] However, in existing MDD research, GNN models at the brain network level are easily affected by indirect factors such as gender, age, and collection site, making it difficult to extract shared patterns from group data. While group graph models can integrate cross-individual information, they often overlook the internal connectivity structure of the brain network and lack interpretability for key brain regions and connection pathways. These issues stem primarily from the imperfect strategies for constructing brain network features, graph structure design, and methods for handling heterogeneous individual features. There is an urgent need for an analytical framework that can collaboratively model the relationship between brain networks and groups, while balancing generalization and interpretability. Summary of the Invention

[0005] To address the limitations of these technologies in assisting the diagnosis of MDD, this paper proposes a two-stage graph neural network framework, B2P-GL (Brain-to-Population Graph Learning). This framework leverages the complementary strengths of GNNs at both the brain network level and the population level in MDD data analysis by training them in phases. This framework significantly improves the model's generalization and global reasoning capabilities, enabling accurate analysis of brain region data for neuropsychiatric disorders like MDD.

[0006] The technical solution adopted in the present invention is as follows:

[0007] The fMRI data analysis model construction method based on a two-stage graph neural network includes the following steps:

[0008] Step 1: Obtain a functional magnetic resonance imaging (fMRI) signal dataset as training data and preprocess it to obtain a brain network diagram based on functional connectivity;

[0009] Step 2: Build and train the first-stage learning module. First, use the graph attention network to update the node features in the brain network graph. Then, aggregate them through the graph pooling module to generate the brain network feature embedding vector. Then, train it through joint loss optimization to obtain the first-stage brain network feature embedding vector.

[0010] Step 3: Embed the brain network feature vector as the individual node feature, and use the similarity of non-imaging condition information between individual node features as the connecting edge to construct a group network graph based on conditional constraints;

[0011] Step 4: Construct and train the second-stage prediction module. First, use the heterogeneous graph convolutional network to update the individual node features to obtain heterogeneous individual node features. Then, the phenotypic data fusion module outputs phenotypic features based on the phenotypic data of each individual. Then, use the fusion module to fuse the phenotypic features with the heterogeneous individual node features based on the gating mechanism to obtain a fused representation. Finally, use joint loss optimization for training and output the trained two-stage model.

[0012] Step 5: Input the functional magnetic resonance imaging signal to be predicted into the two-stage model to obtain the prediction analysis results.

[0013] The brain network diagram is constructed by the following steps:

[0014] The functional magnetic resonance imaging signals were divided into brain regions and the blood oxygen level dependent signals of each brain region were extracted;

[0015] Based on the Pearson correlation coefficient, the correlation of blood oxygen level-dependent signals between brain regions was calculated to obtain the functional connectivity coefficient, which was used as the node feature of the brain network diagram. The correlation size between brain regions was used as the edge connection of the brain network diagram.

[0016] The graph pooling module concatenates all node features processed by the graph attention network through average pooling and maximum pooling.

[0017] In step 2, the joint loss optimization specifically includes classification loss and similarity loss, and the expression is as follows:

[0018]

[0019] in represents the first stage joint loss function, Represents the classification loss, which is the ability of the supervised neural network to discriminate the category label of each node in the brain network graph. Represents the similarity loss of brain network feature embedding vectors.

[0020] The specific construction process of the group network diagram based on conditional constraints is as follows:

[0021] The non-imaging condition information includes the subject's acquisition site, gender, and age group. For each condition c in the non-imaging condition information, a group network connection under the corresponding condition constraint is constructed; the condition constraint is defined as: for any target individual node i, only the individual condition set that is different from the condition category to which the node i belongs is connected. In , nodes are selected to calculate the similarity of the brain network feature embedding vector; the similarity A of the brain network feature embedding vector is expressed as follows:

[0022]

[0023] Among them, z i is the target individual node, z j are individual nodes belonging to different condition categories, and σ is the width adjustment parameter of the kernel function;

[0024] Finally, from the individual condition set In the

[15] , the node with the highest similarity to the target individual node i-brain network feature embedding vector is selected to establish connections between individual nodes based on conditional constraints.

[0025] The phenotypic data fusion module outputs phenotypic features through the phenotypic data encoder and aligns the dimensions with the heterogeneous individual node features output by the heterogeneous graph convolutional network.

[0026] The joint loss optimization in the second stage is specifically optimized by the following formula:

[0027]

[0028] in, represents the second stage fusion loss optimization function, is the similarity loss of heterogeneous individual node features, represents the similarity loss of the fused representation.

[0029] The classification loss expression is as follows:

[0030]

[0031] Where N represents the number of individual samples, y i represents the true label of the input i-th sample, Represents the predicted label of the i-th sample based on the corresponding feature; the corresponding feature is the brain network feature embedding vector or fusion representation;

[0032] The similarity loss expression is as follows:

[0033]

[0034] in, is the similarity matrix S i,j The normalized similarity matrix, is the label similarity matrix, |·| F represents the Frobenius norm; the similarity matrix S i,j Each element S′ in i,j It represents the dot product similarity between the i-th and j-th sample brain network feature embedding vectors, and is expressed as follows:

[0035]

[0036] Among them, h i and h j are the brain network feature embedding vectors of the i-th sample and the j-th sample, respectively, and d is the embedding space dimension;

[0037] The label similarity matrix T is a symmetric binary matrix, where each element T in the label similarity matrix T i,j Represents the relationship between the i-th and j-th sample labels, and the expression is as follows:

[0038]

[0039] When two samples belong to the same category, their label similarity is set to 1; otherwise, if their categories are different, it is set to 0. The similarity loss of heterogeneous individual node features is calculated by replacing the brain network feature embedding vector with heterogeneous individual node features and fusion representation. and the similarity loss of the fusion representation

[0040] The present invention also provides an fMRI data analysis device based on a two-stage graph neural network, characterized by comprising:

[0041] Brain network graph construction module: obtains and preprocesses the fMRI signal dataset, uses the functional connectivity coefficient as the node feature of the brain network graph, and uses the correlation size between brain regions as the edge connection of the brain network graph to construct a brain network graph based on functional connectivity;

[0042] The first-stage learning module: updates the node features in the brain network graph through the graph attention network module, and then aggregates and generates the brain network feature embedding vector through the graph pooling module to obtain the first-stage brain network feature embedding vector, which is then trained and optimized through joint loss optimization;

[0043] Group network graph construction module: This module uses brain network feature embedding vectors as individual node features and uses the similarity of non-imaging conditional information between individual node features as connecting edges to construct a group network graph based on conditional constraints.

[0044] The second stage prediction module uses a heterogeneous graph convolutional network to update individual node features to obtain heterogeneous individual node features. Then, the phenotypic data fusion module outputs phenotypic features based on the phenotypic data of each individual. The fusion module then uses a gating mechanism to fuse the phenotypic features with the heterogeneous individual node features to obtain a fused representation. Finally, joint loss optimization is used for training optimization.

[0045] Classification output module: inputs the functional magnetic resonance imaging signal to be predicted and outputs the prediction analysis results.

[0046] Beneficial effects of the present invention:

[0047] Capturing brain functional connectivity information: Through the first stage of brain network feature embedding vector learning, disease-related connectivity features are extracted to provide input for subsequent prediction tasks;

[0048] Alleviate the heterogeneity problem of medical data: Construct a group graph structure with conditional constraints to suppress the interference of indirect factors on feature representation;

[0049] Integrate phenotypic information to achieve accurate prediction: Introduce phenotypic data information to enhance individual feature expression and improve model prediction accuracy;

[0050] Supports a flexible two-stage training mechanism: the first stage can be trained independently to obtain brain network representation, or combined with the second stage to achieve group-level optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1Schematic diagram of the framework structure based on two-stage graph neural network;

[0052] Figure 2 A diagram of the training process based on a two-stage graph neural network framework;

[0053] Figure 3 Schematic diagram of the confusion matrix output from two-stage training on the REST-meta-MDD dataset. DETAILED DESCRIPTION

[0054] The following describes the embodiments of the present invention in detail with reference to the accompanying drawings and examples so that those skilled in the art can better understand the present invention.

[0055] like Figure 1 As shown, this embodiment provides an fMRI data analysis model based on a two-stage graph neural network, which adopts staged brain network level GNN and group level GNN for learning. In the first stage of brain network level learning, the brain network is constructed through functional connections, and the graph attention network and graph pooling module are used to output the first stage brain network feature embedding vector. In the second stage, the brain network feature embedding vector and phenotypic data jointly construct a group network graph, wherein the brain network feature embedding vector is used as the individual node feature, and the phenotypic data is used as a conditional constraint to construct the individual edge. The heterogeneous graph convolutional network and phenotypic data fusion module outputs the feature representation of each individual node and performs auxiliary analysis of MDD. As shown Figure 2 As shown in Figure 2, the specific construction method of the model is as follows:

[0056] Step 1: Acquire and preprocess a dataset of functional magnetic resonance imaging (fMRI) signals as training data. First, preprocess the fMRI signals of each sample and extract the BOLD signal of each brain region based on a predefined brain region division method. The Pearson correlation coefficient is used to calculate the correlation of BOLD signals between brain regions, thereby constructing the corresponding functional connectivity. A functional connectivity-based brain network diagram is constructed for each sample. The functional connectivity-based brain network consists of brain region node features and brain region edge connections. The brain region node features are the functional connectivity coefficients between each brain region and other brain regions, and the brain region edge connections are determined by the correlation between the brain regions. The functional connectivity coefficient is represented by the Pearson correlation coefficient; a larger coefficient indicates a greater correlation between brain regions.

[0057] Step 2: Build and train the first-stage learning modules, including the graph attention network, graph pooling module, and joint loss optimization module.

[0058] The Graph Attention Network (GAT) is a graph neural network architecture based on an attention mechanism, consisting of multiple layers of graph attention layers. These layers adaptively learn attention weights based on node features between neighboring nodes. These weights aggregate neighboring node features, updating the brain network feature embedding vector at the current layer and serving as input for the next graph attention layer. Different layers are connected through input and output to learn how to extract and express multi-level features within the graph structure, thereby enhancing the expressive power of the brain network feature embedding vector.

[0059] The graph pooling module obtains a brain network feature embedding vector by using average pooling and maximum pooling methods for all node features processed by the graph attention network. That is, the mean and maximum values of the features of all nodes are selected and spliced at the same time to obtain a global brain network representation, namely, a brain network feature embedding vector.

[0060] The first-stage joint loss optimization module is used to simultaneously optimize the output brain network feature embedding vector and the abnormality discrimination ability of the model. By introducing a combination of classification loss and similarity loss, the model's ability to depict brain abnormality-related patterns is improved. The first-stage joint loss function The expression is as follows:

[0061]

[0062] in, represents the classification loss, Represents feature similarity loss. The classification loss is used to supervise the neural network's ability to discriminate the category to which each node in the brain map belongs. The expression is as follows:

[0063]

[0064] Where N represents the number of individual samples, y i represents the true label of the input i-th sample, represents the predicted label of the i-th sample based on the extracted brain network representation.

[0065] The feature similarity loss compares the similarity between the two brain network feature embedding vectors in the embedding space and the difference between their label similarities, prompting the network to learn more discriminative and structural feature representations. The expression is as follows:

[0066]

[0067] in, is the similarity matrix S i,j The normalized similarity matrix, is the label similarity matrix, |·| FIt represents the Frobenius norm, which is used to measure the overall difference between two similarity matrices. i,j Each element S′ in i,j represents the dot product similarity between the i-th and j-th sample brain network feature embedding vectors. The dot product similarity between the brain network feature embedding vectors is expressed as follows:

[0068]

[0069] Among them, h i and h j are the brain network feature embedding vectors of the i-th sample and the j-th sample respectively, and d is the embedding space dimension.

[0070] The label similarity matrix T is a symmetric binary matrix used to describe the label similarity relationship between any two samples.

[0071] Among them, each element T in the label similarity matrix T i,j Represents the relationship between the i-th and j-th sample labels, and the expression is as follows:

[0072]

[0073] When two samples belong to the same category, their label similarity is set to 1; conversely, if they belong to different categories, it is set to 0. Label similarity is used as supervisory information to guide the brain network embedding structure in the feature space, making samples with the same category label closer in the brain network feature embedding vector space, while samples of different categories remain distinguishable.

[0074] When the output accuracy of the first stage stabilizes or reaches the preset training rounds, training optimization is stopped. The brain network feature embedding vector of the first stage is obtained through the brain network feature representation learning module based on the graph attention network and the first stage joint loss.

[0075] Step 3: Construct a second-stage group network diagram based on conditional constraints; the group network diagram based on conditional constraints consists of individual node features and individual connections. The individual node features are the brain network feature embedding vectors output in the first stage. The individual connections are constructed based on multiple non-imaging condition information, including the subject's acquisition site, gender, and age group. For each condition c in the non-imaging condition information, construct the group network connection under its corresponding conditional constraint. The conditional constraint is defined as: for any target node i, only the individual condition set that is different from the condition category to which node i belongs is included. In [1], nodes are selected to calculate the similarity of the brain network feature embedding vector. The similarity A of the brain network feature embedding vector is expressed as follows:

[0076]

[0077] Among them, z i is the target node, z j are individual nodes with different condition categories, and σ is the width adjustment parameter of the kernel function. In

[15] , the top-k nodes with the highest similarity to the feature embedding vector of the brain network of node i are selected to establish connections between individual nodes based on conditional constraints.

[0078] S4. Build and train the second-stage prediction module, including: heterogeneous graph convolutional network, phenotypic data fusion module, and fusion loss optimization module.

[0079] The heterogeneous graph convolutional network (HGCN) is a graph neural network structure suitable for multi-relational connections. It is used to learn conditionally constrained group network graphs. The overall network is composed of multiple layers of heterogeneous graph convolutional layers. These layers perform conditional edge-based feature aggregation on neighbor node information from different conditional categories (site, gender, and age group). By averaging the features learned under different conditions, the features of individual nodes in each layer are updated to obtain heterogeneous individual node features.

[0080] The phenotypic data fusion module is used to introduce the phenotypic data (age, gender, site and length of education) of each individual as auxiliary information, and align the phenotypic features output by the phenotypic data encoder with the heterogeneous individual node features output by the heterogeneous graph convolutional network.

[0081] The fusion module uses a gating mechanism to encode the phenotypic features and individual node characteristics Fusion is performed, where N2 is the number of individual nodes. The gating mechanism inputs the phenotypic features P and the individual node features Z into the gating network to generate a fusion coefficient matrix G, which is defined as:

[0082] G=σ(W g [Z,P]+b g )

[0083] Among them, σ(·) is the Sigmoid activation function, [·,·] represents the feature cascade operation, W g with b g is the learnable parameter in the gating network. The final fusion representation output is The expression is as follows:

[0084]

[0085] Where ⊙ represents element-wise multiplication, ReLU(·) is the activation function, and LayerNorm(·) is used to normalize the fused feature representation.

[0086] The second stage fusion loss optimization module is used to simultaneously optimize the node features and the model's discriminative ability before and after the fusion of phenotypic data. By introducing classification loss and similarity loss before and after the fusion of phenotypic data, the model's learning ability and classification performance for brain data abnormalities and phenotypic data are improved. The second stage fusion loss optimization function The expression is as follows:

[0087]

[0088] Among them, the classification loss Same as the first stage, the similarity loss of heterogeneous individual node feature Z is Fusion Representation The similarity loss is

[0089] When the model's accuracy stabilizes or reaches the preset number of training rounds, training optimization is stopped and the optimal two-stage model is saved and output. In the second stage, the MDD prediction module, based on heterogeneous graph neural networks and phenotypic data fusion optimization, achieves accurate data analysis through data analysis of fMRI signals and conditional constraints on individual samples.

[0090] This embodiment further provides an fMRI data analysis device based on a two-stage graph neural network, comprising:

[0091] The brain network map construction module acquires and preprocesses a functional magnetic resonance imaging (fMRI) signal dataset, using functional connectivity coefficients as node features and inter-brain correlations as edge connections to construct a functional connectivity-based brain network map. First, the fMRI signals of each sample are preprocessed, and the BOLD signal of each brain region is extracted based on a predefined brain region delineation method. The Pearson correlation coefficient is used to calculate the correlation between the BOLD signals of each brain region, thereby constructing the corresponding functional connectivity. For each sample, a functional connectivity-based brain network map is constructed. The functional connectivity-based brain network consists of brain region node features and brain region edge connections. The brain region node features are the functional connectivity coefficients between each brain region and other brain regions, and the brain region edge connections are determined by the correlation between the brain regions. The functional connectivity coefficient is calculated using the Pearson correlation coefficient; a larger coefficient indicates greater correlation between brain regions.

[0092] The first-stage learning module: updates the node features in the brain network graph through the graph attention network module, then aggregates and generates the brain network feature embedding vector through the graph pooling module to obtain the first-stage brain network feature embedding vector, which is then trained and optimized through the first-stage joint loss optimization module;

[0093] The Graph Attention Network (GAT) is a graph neural network architecture based on an attention mechanism, consisting of multiple layers of graph attention layers. These layers adaptively learn attention weights based on node features between neighboring nodes. These weights aggregate neighboring node features, updating the brain network feature embedding vector at the current layer and serving as input for the next graph attention layer. Different layers are connected through input and output to learn how to extract and express multi-level features within the graph structure, thereby enhancing the expressive power of the brain network feature embedding vector.

[0094] The graph pooling module obtains a brain network feature embedding vector by using average pooling and maximum pooling methods for all node features processed by the graph attention network. That is, the mean and maximum values of the features of all nodes are selected and spliced at the same time to obtain a global brain network representation, namely, a brain network feature embedding vector.

[0095] The first-stage joint loss optimization module is used to simultaneously optimize the output brain network feature embedding vector and the abnormality discrimination ability of the model. By introducing a combination of classification loss and similarity loss, the model's ability to depict brain abnormality-related patterns is improved. The first-stage joint loss function The expression is as follows:

[0096]

[0097] in, represents the classification loss, Represents feature similarity loss. The classification loss is used to supervise the neural network's ability to discriminate the category to which each node in the brain map belongs. The expression is as follows:

[0098]

[0099] Where N represents the number of individual samples, y i represents the true label of the input i-th sample, represents the predicted label of the i-th sample based on the extracted brain network representation.

[0100] The feature similarity loss compares the similarity between the two brain network feature embedding vectors in the embedding space and the difference between their label similarities, prompting the network to learn more discriminative and structural feature representations. The expression is as follows:

[0101]

[0102] in, is the similarity matrix S i,j The normalized similarity matrix, is the label similarity matrix, |·| FIt represents the Frobenius norm, which is used to measure the overall difference between two similarity matrices. i,j Each element S′ in i,j represents the dot product similarity between the i-th and j-th sample brain network feature embedding vectors. The dot product similarity between the brain network feature embedding vectors is expressed as follows:

[0103]

[0104] Among them, h i and h j are the brain network feature embedding vectors of the i-th sample and the j-th sample respectively, and d is the embedding space dimension.

[0105] The label similarity matrix T is a symmetric binary matrix used to describe the label similarity relationship between any two samples.

[0106] Among them, each element T in the label similarity matrix T i,j Represents the relationship between the i-th and j-th sample labels, and the expression is as follows:

[0107]

[0108] When two samples belong to the same category, their label similarity is set to 1; conversely, if they belong to different categories, it is set to 0. Label similarity is used as supervisory information to guide the brain network embedding structure in the feature space, making samples with the same category label closer in the brain network feature embedding vector space, while samples of different categories remain distinguishable.

[0109] When the output accuracy of the first stage stabilizes or reaches the preset training rounds, training optimization is stopped. The brain network feature embedding vector of the first stage is obtained through the brain network feature representation learning module based on the graph attention network and the first stage joint loss.

[0110] Group network diagram construction module: The brain network feature embedding vector is used as the individual node feature, and the similarity of the non-imaging condition information between the individual node features is used as the connecting edge to construct a group network diagram based on conditional constraints; the non-imaging condition information includes the subject's acquisition site, gender, and age group. For each condition c in the non-imaging condition information, the group network connection under its corresponding conditional constraint is constructed. The conditional constraint is defined as: for any target node i, only the individual condition set that is different from the condition category of node i is included. In [1], nodes are selected to calculate the similarity of the brain network feature embedding vector. The similarity A of the brain network feature embedding vector is expressed as follows:

[0111]

[0112] Among them, z i is the target node, z j are individual nodes with different condition categories, and σ is the width adjustment parameter of the kernel function. In

[15] , the top-k nodes with the highest similarity to the feature embedding vector of the brain network of node i are selected to establish connections between individual nodes based on conditional constraints.

[0113] The second-stage prediction module uses a heterogeneous graph convolutional network to update individual node features to obtain heterogeneous individual node features. The phenotypic data fusion module then outputs phenotypic features based on the phenotypic data of each individual. The fusion module then fuses the phenotypic features with the heterogeneous individual node features based on a gating mechanism to obtain a fused representation. Finally, the second-stage fusion loss optimization module performs training optimization.

[0114] The heterogeneous graph convolutional network (HGCN) is a graph neural network structure suitable for multi-relational connections. It is used to learn conditionally constrained group network graphs. The overall network is composed of multiple layers of heterogeneous graph convolutional layers. These layers perform conditional edge-based feature aggregation on neighbor node information from different conditional categories (site, gender, and age group). By averaging the features learned under different conditions, the features of individual nodes in each layer are updated to obtain heterogeneous individual node features.

[0115] The phenotypic data fusion module is used to introduce the phenotypic data (age, gender, site and length of education) of each individual as auxiliary information, and align the phenotypic features output by the phenotypic data encoder with the heterogeneous individual node features output by the heterogeneous graph convolutional network.

[0116] The fusion module uses a gating mechanism to encode the phenotypic features and individual node characteristics Fusion is performed, where N2 is the number of individual nodes. The gating mechanism inputs the phenotypic features P and the individual node features Z into the gating network to generate a fusion coefficient matrix G, which is defined as:

[0117] G=σ(W g [Z,P]+b g )

[0118] Among them, σ(·) is the Sigmoid activation function, [·,·] represents the feature cascade operation, W g with b g is the learnable parameter in the gating network. The final fusion representation output is The expression is as follows:

[0119]

[0120] Where ⊙ represents element-wise multiplication, ReLU(·) is the activation function, and LayerNorm(·) is used to normalize the fused feature representation.

[0121] The second stage fusion loss optimization module is used to simultaneously optimize the node features and the model's discriminative ability before and after the fusion of phenotypic data. By introducing classification loss and similarity loss before and after the fusion of phenotypic data, the model's learning ability and classification performance for brain data abnormalities and phenotypic data are improved. The second stage fusion loss optimization function The expression is as follows:

[0122]

[0123] Among them, the classification loss Same as the first stage, the similarity loss of heterogeneous individual node feature Z is Fusion Representation The similarity loss is

[0124] Classification output module: inputs the functional magnetic resonance imaging signal to be predicted and outputs the prediction analysis results.

[0125] like Figure 3 As shown in Figure 2, MDD prediction experiments were conducted on the publicly available Rest-meta-MDD data using both the brain network-level graph neural network and the two-stage graph neural network model. The experimental results are presented in a confusion matrix, where negative values indicate relative normality and positive values indicate abnormal data. The two-stage graph neural network significantly outperformed the brain network-level graph neural network in predictive performance. By optimizing BOLD signal data, the classification accuracy increased by 17.84%.

[0126] The above description of the specific implementation methods clearly demonstrates the implementation steps and processes of the present invention, providing detailed guidance for practical applications.

[0127] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. A method for constructing an fMRI data analysis model based on a two-stage graph neural network, characterized in that: The following steps are involved: Step 1: Obtain a functional magnetic resonance imaging (fMRI) signal dataset as training data and preprocess it to obtain a brain network diagram based on functional connectivity; Step 2: Build and train the first-stage learning module. First, use the graph attention network to update the node features in the brain network graph. Then, aggregate them through the graph pooling module to generate the brain network feature embedding vector. Then, train it through joint loss optimization to obtain the first-stage brain network feature embedding vector. Step 3: Embed the brain network feature vector as the individual node feature, and use the similarity of non-imaging condition information between individual node features as the connecting edge to construct a group network graph based on conditional constraints; Step 4: Construct and train the second-stage prediction module. First, use the heterogeneous graph convolutional network to update the individual node features to obtain heterogeneous individual node features. Then, the phenotypic data fusion module outputs phenotypic features based on the phenotypic data of each individual. Then, use the fusion module to fuse the phenotypic features with the heterogeneous individual node features based on the gating mechanism to obtain a fused representation. Finally, use joint loss optimization for training and output the trained two-stage model. Step 5: Input the functional magnetic resonance imaging signal to be predicted into the two-stage model to obtain the prediction analysis results.

2. The method for constructing an fMRI data analysis model based on a two-stage graph neural network according to claim 1, characterized in that: The brain network diagram is constructed by the following steps: The functional magnetic resonance imaging signals were divided into brain regions and the blood oxygen level dependent signals of each brain region were extracted; Based on the Pearson correlation coefficient, the correlation of blood oxygen level-dependent signals between brain regions was calculated to obtain the functional connectivity coefficient, which was used as the node feature of the brain network diagram. The correlation size between brain regions was used as the edge connection of the brain network diagram.

3. The method for constructing an fMRI data analysis model based on a two-stage graph neural network according to claim 2, characterized in that: The graph pooling module concatenates all node features processed by the graph attention network through average pooling and maximum pooling.

4. The method for constructing an fMRI data analysis model based on a two-stage graph neural network according to claim 3, characterized in that: In step 2, the joint loss optimization specifically includes classification loss and similarity loss, and the expression is as follows: in represents the first stage joint loss function, Represents the classification loss, which is the ability of the supervised neural network to discriminate the category label of each node in the brain network graph. Represents the similarity loss of brain network feature embedding vectors.

5. The method for constructing an fMRI data analysis model based on a two-stage graph neural network according to claim 4, characterized in that: The specific construction process of the group network diagram based on conditional constraints is as follows: The non-imaging condition information includes the subject's acquisition site, gender, and age group. For each condition c in the non-imaging condition information, a group network connection under the corresponding condition constraint is constructed; the condition constraint is defined as: for any target individual node i, only the individual condition set that is different from the condition category to which the node i belongs is connected. In , nodes are selected to calculate the similarity of the brain network feature embedding vector; the similarity A of the brain network feature embedding vector is expressed as follows: Among them, z i is the target individual node, z j are individual nodes belonging to different condition categories, and σ is the width adjustment parameter of the kernel function; Finally, from the individual condition set In the [15], the node with the highest similarity to the target individual node i-brain network feature embedding vector is selected to establish connections between individual nodes based on conditional constraints.

6. The method for constructing an fMRI data analysis model based on a two-stage graph neural network according to claim 5, characterized in that: The phenotypic data fusion module outputs phenotypic features through the phenotypic data encoder and aligns the dimensions with the heterogeneous individual node features output by the heterogeneous graph convolutional network.

7. The method for constructing an fMRI data analysis model based on a two-stage graph neural network according to claim 6, characterized in that: The joint loss optimization in the second stage is specifically optimized by the following formula: in, represents the second stage fusion loss optimization function, is the similarity loss of heterogeneous individual node features, represents the similarity loss of the fused representation.

8. The method for constructing an fMRI data analysis model based on a two-stage graph neural network according to claim 7, characterized in that: The classification loss expression is as follows: Where N represents the number of individual samples, y i represents the true label of the input i-th sample, Represents the predicted label of the i-th sample based on the corresponding feature; the corresponding feature is the brain network feature embedding vector or fusion representation; The similarity loss expression is as follows: in, is the similarity matrix S i,j The normalized similarity matrix, is the label similarity matrix, |·| F represents the Frobenius norm; the similarity matrix S i,j Each element S′ in i,j It represents the dot product similarity between the i-th and j-th sample brain network feature embedding vectors, and is expressed as follows: Among them, h i and h j are the brain network feature embedding vectors of the i-th sample and the j-th sample, respectively, and d is the embedding space dimension; The label similarity matrix T is a symmetric binary matrix, where each element T in the label similarity matrix T i,j Represents the relationship between the i-th and j-th sample labels, and the expression is as follows: When two samples belong to the same category, their label similarity is set to 1; otherwise, if their categories are different, it is set to 0. The similarity loss of heterogeneous individual node features is calculated by replacing the brain network feature embedding vector with heterogeneous individual node features and fusion representation. and the similarity loss of the fusion representation 9. An fMRI data analysis device based on a two-stage graph neural network, characterized in that: include: Brain network graph building blocks; Acquire and preprocess the fMRI signal dataset, use the functional connectivity coefficient as the node feature of the brain network graph, and the correlation between brain regions as the edge connection of the brain network graph to construct a brain network graph based on functional connectivity. The first-stage learning module: updates the node features in the brain network graph through the graph attention network module, and then aggregates and generates the brain network feature embedding vector through the graph pooling module to obtain the first-stage brain network feature embedding vector, which is then trained and optimized through joint loss optimization; Group network graph construction module: This module uses brain network feature embedding vectors as individual node features and the similarity of non-imaging conditional information between individual node features as connecting edges to construct a group network graph based on conditional constraints. The second stage prediction module uses a heterogeneous graph convolutional network to update individual node features to obtain heterogeneous individual node features. Then, the phenotypic data fusion module outputs phenotypic features based on the phenotypic data of each individual. The fusion module then uses a gating mechanism to fuse the phenotypic features with the heterogeneous individual node features to obtain a fused representation. Finally, joint loss optimization is used for training optimization. Classification output module: inputs the functional magnetic resonance imaging signal to be predicted and outputs the prediction analysis results.

Images