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

Abstract

A kind of fMRI data analysis model construction method and device based on two-stage graph neural network, it is related to computer-aided diagnosis field;The present application first constructs brain network graph by the pre-processing of fMRI signal, captures brain functional connection information;Then, the first stage brain network feature embedding vector learning is used, and the connection features related to disease are extracted, to provide input for subsequent prediction task;Second, a group graph structure with conditional constraints is constructed to suppress the interference caused by indirect factors on feature representation;Finally, introduce phenotype data information, enhance individual feature expression, improve model prediction accuracy;The model is trained by brain network level GNN and group level GNN in stages, and the complementary advantages of the two in the data analysis of MDD are fully utilized.The model significantly improves the generalization ability and global reasoning ability of the model, and supports the accurate analysis result of brain region data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer-aided diagnosis, and in particular to a two-stage graph neural network-based fMRI data analysis model construction method and device. BACKGROUND

[0002] Major Depressive Disorder (MDD) is the second most disabling mental illness in the world. Its diagnosis has long relied on subjective evaluation of the symptom checklist in the Diagnostic and Statistical Manual of Mental Disorders (DSM), lacking objective biomarker support. Functional Magnetic Resonance Imaging (fMRI) technology can quantify brain functional connectivity (FC) through Blood Oxygenation Level Dependent (BOLD) signals, providing a potential approach for objective detection and data analysis of MDD. However, traditional machine learning methods are difficult to effectively analyze the non-linear dynamic connection characteristics between brain regions, and are limited by the analysis paradigm of single-modal data.

[0003] In recent years, Graph Neural Network (GNN) has been widely applied in the analysis of neuropsychiatric diseases due to its advantages in processing graph-structured data. In brain network modeling, GNN can effectively depict the non-Euclidean structure and high-order interaction in brain functional connectivity through node feature propagation mechanism, showing strong representation ability for brain region functional patterns. Meanwhile, researchers have extended GNN to group-level modeling paradigm, where each node represents a subject, and node features are derived from individual image, behavior or phenotype data. Edges reflect the similarity or shared statistical properties between individuals, enabling GNN to extract potential structural features from the group and improve classification accuracy and generalization ability. These two levels of GNN methods have been widely applied in the auxiliary diagnosis of neuropsychiatric diseases, achieving better performance than traditional methods.

[0004] However, in existing MDD-related research, GNN models at the brain network level are easily disturbed by indirect factors such as gender, age, and acquisition site, making it difficult to extract shared patterns from group data. While group graph models can integrate cross-individual information, they often ignore the connection structure within the brain network, lacking interpretability for key brain regions and connection pathways. The above problems mainly arise from imperfect brain network feature construction strategies, graph structure design, and heterogeneous individual feature processing methods, necessitating an analysis framework that can model brain network and group relationships, balance generalizability and interpretability. SUMMARY

[0005] To address the limitations of the aforementioned technologies in assisting the diagnosis of MDD, this invention proposes a two-stage graph neural network framework, B2P-GL (Brain-to-Population Graph Learning). This framework leverages the complementary advantages of both brain-level GNNs and population-level GNNs through phased training, significantly enhancing the model's generalization and global reasoning capabilities, and supporting accurate analysis of brain region data related to neuropsychiatric diseases such as MDD.

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

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

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

[0009] Step 2: Construct and train the first-stage learning module; First, use the graph attention network to update the node features in the brain network graph, then use the graph pooling module to aggregate and generate brain network feature embedding vectors, and then use joint loss optimization to train and obtain the first-stage brain network feature embedding vectors.

[0010] Step 3: Embed the brain network features into vectors as individual node features, and use the similarity of non-image conditional information between individual node features as connection edges to construct a group network graph based on conditional constraints.

[0011] Step 4: Construct and train the second-stage prediction module; First, use 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; next, use the fusion module to fuse the phenotypic features with the heterogeneous individual node features based on the gating mechanism to obtain the fused representation; finally, use joint loss optimization for training to 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 map was constructed through the following steps:

[0014] Brain regions were divided from functional magnetic resonance imaging (fMRI) signals, and blood oxygenation level-dependent signals were extracted from each brain region.

[0015] The correlation between blood oxygen level dependent signals of each brain region is calculated based on a Pearson correlation coefficient to obtain a functional connection coefficient as a node feature of the brain network graph, and the correlation between brain regions as an edge connection of the brain network graph.

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

[0017] In step 2, the joint loss optimization specifically includes a classification loss and a similarity loss, and the expressions are as follows:

[0018]

[0019] Wherein represents the first-stage joint loss function, represents the classification loss, and the supervised neural network discriminates the category label of each node in the brain network graph, represents the similarity loss of the brain network feature embedding vector.

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

[0021] The non-image condition information includes the acquisition site, gender and age group of the subject, and for each condition c in the non-image condition information, a group network connection under the condition constraint is constructed respectively; the condition constraint is defined as: for any target individual node i, only in the individual condition set , the similarity of the node pair brain network feature embedding vector is selected; the similarity A of the brain network feature embedding vector is expressed as follows:

[0022]

[0023] Wherein, z i is the target individual node, z j is the individual node with different condition categories, and sigma is the width adjustment parameter of the kernel function.

[0024] Finally, from the individual condition set , the node with the highest similarity of the brain network feature embedding vector with the target individual node i is selected to establish the connection between the individual nodes based on the conditional constraint.

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

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

[0027]

[0028] wherein, represents a second-stage fusion loss optimization function, is a similarity loss of heterogeneous individual node features, represents a similarity loss of fusion representation.

[0029] The classification loss expression is as follows:

[0030]

[0031] wherein, N represents the number of individual samples, y i represents the real label of the input ith sample, represents the predicted label of the ith sample based on the corresponding features, wherein the corresponding features are brain network feature embedding vectors or fusion representations;

[0032] The similarity loss expression is as follows:

[0033]

[0034] wherein, is a similarity matrix S i,j a normalized similarity matrix, is a label similarity matrix, |·| F represents the Frobenius norm; the similarity matrix S i,j each element S′ i,j represents the dot product similarity between the ith and jth sample brain network feature embedding vectors, and the expression is as follows:

[0035]

[0036] wherein, h i and h j are the brain network feature embedding vectors of the ith sample and the jth sample, respectively, and d is the embedding space dimension;

[0037] The label similarity matrix T is a symmetric binary matrix, wherein each element T i,j in the label similarity matrix T represents the label relationship between the ith and jth samples, and the expression is as follows:

[0038]

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

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

[0041] A brain network graph construction module; an fMRI signal dataset is acquired and preprocessed, a functional connection coefficient is taken as a node feature of a brain network graph, and the correlation between brain regions is taken as an edge connection of the brain network graph, so that a brain network graph based on functional connection is constructed;

[0042] A first-stage learning module: the node features in the brain network graph are updated through a graph attention network module, then brain network feature embedding vectors are aggregated through a graph pooling module to obtain first-stage brain network feature embedding vectors, and training optimization is performed through joint loss optimization;

[0043] A group network graph construction module: the brain network feature embedding vectors are taken as individual node features, and the similarity of non-image condition information between individual node features is taken as a connection edge, so that a group network graph based on condition constraints is constructed;

[0044] A second-stage prediction module: the individual node features are updated by using a heterogeneous graph convolution network to obtain heterogeneous individual node features; then a phenotype data fusion module outputs phenotype features based on the phenotype data of each individual; then a fusion module is used to fuse the phenotype features and the heterogeneous individual node features based on a gating mechanism to obtain a fusion representation; and finally, training optimization is performed through joint loss optimization;

[0045] A classification output module: the fMRI signal to be predicted is input, and a prediction analysis result is output.

[0046] The application has the following beneficial effects:

[0047] Capture brain functional connection information: the connection features related to diseases are extracted through first-stage brain network feature embedding vector learning, so as to provide input for subsequent prediction tasks;

[0048] Relieve the heterogeneity problem of medical data: a group graph structure with condition constraints is constructed to inhibit the interference of indirect factors on feature representation;

[0049] Fusion of phenotype information to realize accurate prediction: the introduction of phenotype data information enhances individual feature expression and improves model prediction accuracy;

[0050] Support flexible two-stage training mechanism: the first stage can be independently trained to obtain brain network representation, and the second stage can be combined to realize group-level optimization. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1Fig. 1 is a schematic diagram of a two-stage graph neural network framework structure;

[0052] Figure 2 Fig. 2 is a schematic diagram of a two-stage graph neural network framework training process;

[0053] Figure 3 Fig. 3 is a schematic diagram of a confusion matrix of two-stage training output under the REST-meta-MDD dataset. DETAILED DESCRIPTION

[0054] The embodiments of the present application will be described in detail below with reference to the accompanying drawings and embodiments, so that those skilled in the art can better understand the present application.

[0055] As shown in the drawings, Figure 1 The two-stage graph neural network-based fMRI data analysis model provided by the embodiment adopts a phased brain network level GNN and a group level GNN for learning. In the first stage of brain network level learning, a brain network is constructed through functional connections, and a graph attention network and a graph pooling module are used to output a first-stage brain network feature embedding vector. In the second stage, the brain network feature embedding vector and the phenotype data are used to jointly construct a group network graph, wherein the brain network feature embedding vector is used as an individual node feature, and the phenotype data is used as a condition constraint to construct individual edges. A heterogeneous graph convolution network and a phenotype data fusion module output a feature representation of each individual node and perform auxiliary analysis of MDD. As shown in the drawings, Figure 2 The specific construction method of the model is as follows:

[0056] Step 1: Obtain brain functional magnetic resonance imaging (fMRI) signal dataset as training data and pre-process; first, pre-process the fMRI signal of each sample, and extract the BOLD signal of each brain region according to the pre-defined brain region division method. The Pearson correlation coefficient is used to calculate the correlation of the BOLD signals between brain regions, thereby constructing the corresponding functional connections. For each sample, a brain network graph based on functional connections is constructed, which is composed of brain region node features and brain region edge connections. The brain region node feature is the functional connection coefficient of each brain region to other brain regions, and the brain region edge connection is determined by the correlation size of the brain region. The functional connection coefficient is represented by the Pearson correlation coefficient, and the larger the coefficient, the greater the correlation between brain regions.

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

[0058] The graph attention network (GAT) is a graph neural network structure based on an attention mechanism and is composed of multiple graph attention layers. The graph attention layer uses the node features between neighbor nodes to adaptively learn attention weights and uses the attention weights to aggregate the features of neighbor nodes, so as to update the brain network feature embedding vector in the current layer and use the brain network feature embedding vector as the input of the next layer of graph attention layer. The input and output are connected between different layers to learn the extraction and expression of multi-order features in the graph structure, thereby enhancing the expression ability of the brain network feature embedding vector.

[0059] The graph pooling module obtains the brain network feature embedding vector through average pooling and maximum pooling of 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 to obtain the global brain network representation, that is, the 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 the combination of the classification loss and the similarity loss, the ability of the model to describe the brain abnormality related pattern is improved. The first-stage joint loss function The expression is as follows:

[0061]

[0062] wherein, the classification loss, the feature similarity loss. The classification loss is used to supervise the discrimination ability of the neural network for the category to which each node in the brain graph belongs, and the expression is as follows:

[0063]

[0064] wherein, 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 promotes the network to learn a more discriminative and structural feature representation by comparing the similarity between two brain network feature embedding vectors in the embedding space and the difference in label similarity, and the expression is as follows:

[0066]

[0067] wherein, is a similarity matrix S i,j is a normalized similarity matrix, is a label similarity matrix, |·| Fdenotes the Frobenius norm, which is used to measure the overall difference between two similarity matrices. The similarity matrix S i,j each element S′ i,j denotes 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] where h i and h j are the i-th and j-th sample brain network feature embedding vectors, respectively, and d is the embedding space dimension.

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

[0071] where each element T i,j in the label similarity matrix T denotes the i-th and j-th sample label relationship, which is expressed as follows:

[0072]

[0073] When two samples belong to the same class, their label similarity is set to 1; otherwise, if their classes are different, it is set to 0. The label similarity serves as supervised information to guide the brain network embedding structure in the feature space, so that samples with the same class label are closer in the brain network feature embedding vector space, and samples of different classes remain distinguishable.

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

[0075] Step 3: Construct a second-stage conditional constraint-based group network graph; the conditional constraint-based group network graph is composed of individual node features and individual connections. The individual node features are the first-stage output brain network feature embedding vectors. The individual connections are constructed according to a plurality of non-image condition information, which includes the acquisition site, gender, and age group of the subject. For each condition c in the non-image condition information, a group network connection under the corresponding condition constraint is constructed. The condition constraint is defined as: for any target node i, only in the individual condition set where the brain network feature embedding vector similarity A is selected for the node pair. The expression of the brain network feature embedding vector similarity A is as follows:

[0076]

[0077] where z i is the target node, z j is the individual node with different condition category, and σ is the width adjustment parameter of the kernel function. Finally, the top-k nodes with the highest similarity to the brain network feature embedding vector of node i are selected from the individual condition set to establish the connection between individual nodes based on condition constraints.

[0078] S4, constructing and training the second stage prediction module, including: a heterogeneous graph convolutional network, a phenotype data fusion module, and a fusion loss optimization module.

[0079] The heterogeneous graph convolutional network is a graph neural network structure suitable for multi-relation connection, which is used for learning based on the condition-constrained group network graph. The overall network is composed of multiple layers of heterogeneous graph convolutional layers. The heterogeneous graph convolutional layers respectively perform condition-based feature aggregation on the neighbor node information from different condition categories (station, gender, age group), update the individual node features of each layer by averaging the features learned under different conditions, and obtain the heterogeneous individual node features.

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

[0081] The fusion module adopts a gating mechanism to fuse the encoded phenotype features and individual node features , where N2 is the number of individual nodes. The gating mechanism inputs the phenotype features P and individual node features Z into a gating network to generate a fusion coefficient matrix G, which is defined as:

[0082] G=σ(W g [Z,P]+b g )

[0083] where σ(·) is a Sigmoid activation function, [·,·] represents a feature concatenation operation, W g and b g are learnable parameters in the gating network. The final fusion representation output is The expression is as follows:

[0084]

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

[0086] The second stage fusion loss optimization module is used to optimize the node features and the discrimination ability of the model before and after the phenotype data fusion, and the learning ability and classification performance of the model on brain data abnormalities and phenotype data are improved by introducing the classification loss and the similarity loss before and after the phenotype data fusion. The second stage fusion loss optimization function The expression is as follows:

[0087]

[0088] The classification loss is The similarity loss of the heterogeneous individual node features Z is The fusion representation is The similarity loss of the fusion representation is

[0089] When the accuracy of the model is stable or reaches the preset training round, the training optimization is stopped, and the output best two-stage model is saved. Through the MDD prediction module based on the second stage heterogeneous graph neural network and phenotype data fusion optimization, the fMRI signal data is analyzed, the condition constraints of individual samples are realized, and the precise analysis of the data is realized.

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

[0091] A brain network graph construction module; an fMRI signal dataset is acquired and preprocessed, a functional connection coefficient is used as a node feature of a brain network graph, and a correlation size between brain regions is used as an edge connection of the brain network graph to construct a brain network graph based on the functional connection; first, the fMRI signal of each sample is preprocessed, and the BOLD signal of each brain region is extracted according to a pre-defined brain region division method. The correlation between the BOLD signals of each brain region is calculated by using the Pearson correlation coefficient, so as to construct the corresponding functional connection. The brain network graph based on the functional connection of each sample is constructed, and the brain network based on the functional connection is composed of brain region node features and brain region edge connections. The brain region node features are the functional connection coefficients of each brain region and other brain regions, and the brain region edge connections are determined by the correlation size of the brain region. The functional connection coefficient is calculated and represented by the Pearson correlation coefficient, and the greater the coefficient, the greater the correlation between the brain regions.

[0092] A first stage learning module: the node features in the brain network graph are updated through a graph attention network module, then the brain network feature embedding vector is aggregated and generated through a graph pooling module to obtain the brain network feature embedding vector of the first stage, and the training optimization is performed through a first stage joint loss optimization module;

[0093] The graph attention network (GAT) is a graph neural network structure based on an attention mechanism and is composed of multiple graph attention layers. The graph attention layer uses the node features between neighbor nodes to adaptively learn attention weights and uses the attention weights to aggregate the features of neighbor nodes, so as to update the brain network feature embedding vector in the current layer and use the brain network feature embedding vector as the input of the next layer of graph attention layer. The input and output are connected between different layers to learn the extraction and expression of multi-order features in the graph structure, thereby enhancing the expression ability of the brain network feature embedding vector.

[0094] The graph pooling module obtains the brain network feature embedding vector through average pooling and maximum pooling of 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 to obtain the global brain network representation, that is, the 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 the combination of the classification loss and the similarity loss, the ability of the model to describe the brain abnormality related pattern is improved. The first-stage joint loss function The expression is as follows:

[0096]

[0097] wherein, represents the classification loss, represents the feature similarity loss. The classification loss is used to supervise the discrimination ability of the neural network for the category to which each node in the brain graph belongs, and the expression is as follows:

[0098]

[0099] wherein, 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 promotes the network to learn a more discriminative and structural feature representation by comparing the similarity between two brain network feature embedding vectors in the embedding space and the difference in label similarity. The expression is as follows:

[0101]

[0102] wherein, is a similarity matrix S i,j is a normalized similarity matrix, is a label similarity matrix, |·| Fdenotes the Frobenius norm, which is used to measure the overall difference between two similarity matrices. The similarity matrix S i,j each element S' i,j denotes 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] where h i and h j are the i-th and j-th sample brain network feature embedding vectors, respectively, and d is the embedding space dimension.

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

[0106] where each element T i,j in the label similarity matrix T represents the i-th and j-th sample label relationship, expressed as follows:

[0107]

[0108] When two samples belong to the same class, their label similarity is set to 1; otherwise, if their classes are different, it is set to 0. The label similarity serves as supervised information to guide the brain network embedding structure in the feature space, so that samples with the same class label are closer in the brain network feature embedding vector space, and samples of different classes maintain distinguishability.

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

[0110] Group network graph construction module: the brain network feature embedding vector is taken as the individual node feature, and the similarity of non-image conditional information between individual node features is taken as the connection edge to construct a conditional constraint-based group network graph. The non-image conditional information includes the acquisition site, gender, and age group of the subject. For each condition c in the non-image conditional information, the corresponding group network connection under the condition constraint is constructed respectively. The condition constraint is defined as: for any target node i, only in the individual condition set where the brain network feature embedding vector similarity A between the node pair is calculated. The brain network feature embedding vector similarity A is expressed as follows:

[0111]

[0112] where z i is the target node, z j is the individual node with different condition category, and σ is the width adjustment parameter of the kernel function. Finally, the top-k nodes with the highest similarity to the brain network feature embedding vector of node i are selected from the individual condition set to establish the connection between individual nodes based on condition constraints.

[0113] The second stage prediction module: the individual node features are updated using the heterogeneous graph convolution network to obtain the heterogeneous individual node features; then the phenotype data fusion module outputs the phenotype features based on the phenotype data of each individual; then the fusion module is used to fuse the phenotype features and the heterogeneous individual node features based on the gating mechanism to obtain the fused representation; finally, the second stage fusion loss optimization module is used for training and optimization;

[0114] The heterogeneous graph convolution network is a graph neural network structure suitable for multi-relation connection, which is used for learning based on the condition-constrained group network graph. The overall network is composed of multiple layers of heterogeneous graph convolution layers. The heterogeneous graph convolution layer respectively aggregates the features based on the condition edges from the neighbor nodes under different condition categories (station, gender, age group), and updates the individual node features of each layer by averaging the features learned under different conditions to obtain the heterogeneous individual node features.

[0115] The phenotype data fusion module is used to introduce the phenotype data (age, gender, station, and education duration) of each individual as auxiliary information, and the phenotype features output by the phenotype data encoder are dimensionally aligned with the heterogeneous individual node features output by the heterogeneous graph convolution network.

[0116] The fusion module adopts a gating mechanism to fuse the encoded phenotype features and the individual node features , where N2 is the number of individual nodes. The gating mechanism inputs the phenotype 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] where σ(·) is a Sigmoid activation function, [·,·] represents a feature concatenation operation, W g and b g are learnable parameters in the gating network. The final fused 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 simultaneously optimizes the node features and the model's discriminative ability before and after phenotypic data fusion. By introducing classification loss and similarity loss before and after phenotypic data fusion, it improves the model's learning ability and classification performance for brain data anomalies and phenotypic data. The second-stage fusion loss optimization function... The expression is as follows:

[0122]

[0123] Among them, classification loss Similar to the first stage, the similarity loss of the heterogeneous individual node feature Z is: Fusion representation The similarity loss is

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

[0125] like Figure 3 As shown, prediction experiments for MDD were conducted on publicly available data from Rest-meta-MDD, comparing the brain network-level graph neural network and the two-stage graph neural network model. The experimental results are presented by the confusion matrix, where negative indicates relatively normal data and positive indicates abnormal data. The two-stage graph neural network significantly outperformed the brain network-level graph neural network in prediction performance, improving classification accuracy by 17.84% through optimized processing of the BOLD signal data.

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

[0127] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the 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 within the protection scope of 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 method comprises the following steps: Step 1: obtaining a brain functional magnetic resonance imaging (fMRI) signal dataset as training data and preprocessing to obtain a brain network graph based on functional connectivity; Step 2: constructing and training a first-stage learning module; first, updating the node features in the brain network graph using a graph attention network, then aggregating to generate a brain network feature embedding vector through a graph pooling module, and then training through joint loss optimization to obtain the brain network feature embedding vector of the first stage; Step 3: constructing a group network graph based on conditional constraints by taking the brain network feature embedding vector as the individual node feature and the similarity of non-image condition information between individual node features as the connection edge; Step 4: constructing and training a second-stage prediction module; first, updating the individual node features using a heterogeneous graph convolution network to obtain heterogeneous individual node features; then, a phenotype data fusion module outputs a phenotype feature based on the phenotype data of each individual; then, using a fusion module, the phenotype feature and the heterogeneous individual node feature are fused based on a gating mechanism to obtain a fusion representation; finally, training through joint loss optimization to output the trained two-stage model; Step 5: inputting the functional magnetic resonance imaging signal to be predicted into the two-stage model to obtain a prediction analysis result. The non-imaging condition information includes a collection site, a gender, and an age group of the subject, and for each condition in the non-imaging condition information , a corresponding population network connection under a condition constraint is constructed respectively; the condition constraint is defined as: for any target individual node , only in an individual condition set different from the condition category of the node , a similarity between the node and another node is calculated by using a brain network feature embedding vector; the similarity between the brain network feature embedding vectors is expressed as follows: ; wherein, is a target individual node, is an individual node of a different condition class, is a width adjustment parameter for the kernel function; Finally, from the individual condition set the node with the highest similarity to the target individual node brain network feature embedding vector is selected to establish the connection between individual nodes based on the condition constraint.

2. The two-stage graph neural network-based fMRI data analysis model construction method according to claim 1, characterized in that, The brain network graph is constructed by the following steps: Dividing the brain regions of the functional magnetic resonance imaging signal and extracting the blood oxygen level dependent signal of each brain region; Based on the Pearson correlation coefficient, the correlation of the blood oxygen level dependent signals between brain regions is calculated to obtain a 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.

3. The method of claim 2, wherein the two-stage graph neural network-based fMRI data analysis model is constructed by: The graph pooling module obtains the node features processed by the graph attention network through average pooling and maximum pooling and then splices them.

4. The method of claim 3, wherein the two-stage graph neural network-based fMRI data analysis model is constructed by, In step 2, the joint loss optimization specifically includes a classification loss and a similarity loss, and the expression is as follows: ; wherein represents the first-stage joint loss function, represents the classification loss, supervising the discriminative ability of the neural network on the class label of each node in the brain network graph, represents the similarity loss of the brain network feature embedding vector.

5. The method of claim 4, wherein the two-stage graph neural network-based fMRI data analysis model is constructed by, The phenotype data fusion module outputs a phenotype feature through a phenotype data encoder, which is dimensionally aligned with the heterogeneous individual node feature output by the heterogeneous graph convolution network.

6. The method of claim 5, wherein the two-stage graph neural network-based fMRI data analysis model is constructed by: The joint loss optimization of the second stage is specifically optimized by the following formula: ; wherein, represents a second stage fusion loss optimization function, is a similarity loss for heterogeneous individual node features, represents a similarity loss for the fused representation.

7. The method of claim 6, wherein the two-stage graph neural network-based fMRI data analysis model is constructed by, The classification loss expression is as follows: ; wherein, denotes the number of individual samples, denotes the true label of the input sample, denotes the predicted label based on the corresponding features for the sample; the corresponding features are brain network feature embedding vectors or fusion representations; the similarity loss The expression is as follows: ; wherein, is a similarity matrix is a normalized similarity matrix, is a label similarity matrix, denotes the Frobenius norm; the similarity matrix each element in denotes the dot product similarity between the th sample brain network feature embedding vector and the th sample brain network feature embedding vector, expressed as follows: ; wherein, and are brain network feature embedding vectors for the th sample and the th sample, respectively, is an embedding space dimension; The tag similarity matrix Let be a symmetric binary matrix, where the label similarity matrix is... Each element in Indicates the first The and the first The relationship between the sample labels is expressed as follows: , When two samples belong to the same category, the label similarity is set to 1; otherwise, if the categories are different, it is set to 0; the similarity loss of the heterogeneous individual node features is calculated by replacing the brain network feature embedding vector with the heterogeneous individual node features and the fusion representation and the similarity loss of the fusion representation .

8. The fMRI data analysis apparatus based on two-stage graph neural network, characterized in that, It comprises: a brain network graph construction module; obtaining a brain functional magnetic resonance imaging (fMRI) signal dataset and preprocessing, 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; a first-stage learning module: updating the node features in the brain network graph through a graph attention network module, then aggregating to generate a brain network feature embedding vector through a graph pooling module, to obtain the brain network feature embedding vector of the first stage, and training and optimizing through joint loss optimization; a group network graph construction module: taking the brain network feature embedding vector as the individual node feature and the similarity of non-image condition information between individual node features as the connection edge, to construct a group network graph based on conditional constraints; a second-stage prediction module: first, updating the individual node features using a heterogeneous graph convolution network to obtain heterogeneous individual node features; then, a phenotype data fusion module outputs a phenotype feature based on the phenotype data of each individual; then, using a fusion module, the phenotype feature and the heterogeneous individual node feature are fused based on a gating mechanism to obtain a fusion representation; finally, training through joint loss optimization to output the trained two-stage model. The second stage prediction module: the individual node features are updated by using a heterogeneous graph convolution network to obtain heterogeneous individual node features; then a phenotype data fusion module outputs phenotype features based on the phenotype data of each individual; then a fusion module is used to fuse the phenotype features and the heterogeneous individual node features based on a gating mechanism to obtain a fusion representation; finally, training optimization is performed by using joint loss optimization; The classification output module: inputs the functional magnetic resonance imaging signal to be predicted, and outputs a prediction analysis result. The non-imaging condition information includes a collection site, a gender, and an age group of the subject, and for each condition in the non-imaging condition information , a corresponding population network connection under a condition constraint is constructed respectively; the condition constraint is defined as: for any target individual node , only in an individual condition set different from the condition category to which the node belongs, a similarity between the node is calculated by using a brain network feature embedding vector; the similarity between the brain network feature embedding vectors is expressed as follows: ; wherein, is a target individual node, is an individual node of a different condition class, is a width adjustment parameter for the kernel function; Finally, from the individual condition set the node with the highest similarity to the target individual node brain network feature embedding vector is selected to establish the connection between individual nodes based on the condition constraint.

Images