A brain network generation system and method based on a functional brain network classification task
By generating a learnable connected matrix through a time-series encoder and a graph generator, and combining it with a graph convolutional neural network for classification prediction, this approach solves the problem that existing brain network classification methods cannot capture nonlinearity and temporal relationships, thereby improving classification accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TIANJIN UNIV
- Filing Date
- 2023-04-27
- Publication Date
- 2026-05-12
AI Technical Summary
Existing brain network classification methods cannot effectively capture the nonlinearity and temporal relationships of brain network data, and graph convolutional neural networks are incompatible with handling negative weighted connections, resulting in low classification accuracy.
A system consisting of a time series encoder, a graph generator, and a graph predictor is proposed. The system extracts time series features through a transformer encoder layer, generates a learnable connected matrix, and combines it with a graph convolutional neural network for classification prediction. Optimization models for losses between classes, between different classes, and matrix sparsity are designed.
It improves the classification accuracy of functional brain networks, generates correlation matrices compatible with graph convolutional neural networks, captures nonlinearity and temporal relationships, and enhances the accuracy and stability of the model.
Smart Images
Figure CN116662850B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep learning technology, and in particular to a brain network generation system and method based on a functional brain network classification task. Background Technology
[0002] Brain network functional connectivity analysis is becoming increasingly important in brain imaging research, as it allows us to understand an individual's brain tissue. Numerous findings in neuroscience research indicate that neural circuits are key to understanding differences in brain function across populations. Functional magnetic resonance imaging (fMRI) is one of the most commonly used imaging techniques for studying brain function and tissue. The neuroimaging community is particularly interested in using brain networks to classify individuals.
[0003] In recent years, Graph Convolutional Neural Networks (GCNs) have been developed as a powerful method for processing graph structures. The advantage of GCNs stems from combining information at the local node level and the global neighborhood level. Similar to standard 2D image convolution, which pools the values of neighboring nodes, the goal of graph convolution is to aggregate neighborhood information of nodes in the graph using the graph Laplacian operator. Existing brain network analysis work typically consists of two steps: the first step is to generate a functional brain network from an individual's fMRI data. This is usually done by selecting a brain atlas with a set of regions of interest (ROIs) as nodes and extracting fMRI blood oxygen level dependent (BOLD) signal sequences from each brain region. For edge generation, measures such as Pearson correlation and partial correlation are used to compute pairwise connectivity between node pairs. Then, in the second stage, the brain connectivity measures obtained between all node pairs are used to classify or predict the individual. By using graph convolutional neural networks, higher-order information in functional brain networks can be extracted more effectively, thereby improving the classification accuracy of functional brain networks.
[0004] Most existing brain network classification methods are based on correlation scores, which have two drawbacks: First, correlation methods focus on capturing linear correlations while ignoring temporal order, meaning that shuffling the time step does not change the results. Second, there is a growing trend to apply graph convolutional neural networks (GCNs) to the brain connectivity matrix in functional magnetic resonance imaging (fMRI), but the mechanisms (i.e., message passing) of most GCNs are incompatible with existing functional brain networks that have both positive and negative weighted edges. Existing techniques fail to capture the nonlinearity and temporal relationships in brain network data and are incompatible with most graph convolutional neural networks. Therefore, these works must design additional processes within graph convolutional neural networks to handle negative weights. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a brain network generation system and method based on functional brain network classification tasks.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A brain network generation system based on a functional brain network classification task is disclosed. The system includes a time-series encoder, a graph generator, and a graph predictor based on a graph convolutional neural network, connected in sequence. The time-series encoder is used to encode Bold signal sequences. The graph generator is used to convert the encoded time-series features into a functional brain network graph. The graph predictor is used to classify and predict the functional brain network graph. The time-series encoder includes a transformer encoder layer and a multilayer perceptron A connected in sequence. The transformer encoder layer is used to extract time-series features, and the multilayer perceptron A is used to generate node features of multiple regions of interest (ROIs) after dimensionality reduction of the extracted time-series features.
[0008] Furthermore, the graph generator generates a machine learning-compatible connectivity matrix A, where the elements of the connectivity matrix A are the pairwise connectivity strengths between ROIs. Where h A =softmax(h e ), h e h is the time series feature extracted by the time series encoder. A These are normalized time series features.
[0009] Furthermore, the graph predictor includes a graph convolutional neural network, a regularization layer, and a multilayer perceptron B connected in sequence; the graph convolutional neural network is used to extract node features from the functional brain network graph; and the multilayer perceptron B is used to classify and predict the node features.
[0010] This invention also provides a brain network generation method based on a functional brain network classification task, constructing a brain network generation system based on a functional brain network classification task. This system includes a time-series encoder, a graph generator, and a graph predictor based on a graph convolutional neural network, connected sequentially. The time-series encoder is used to encode Bold signal sequences; the graph generator is used to convert the encoded time-series features into a functional brain network graph; the graph predictor is used to classify and predict the functional brain network graph. The time-series encoder includes a transformer encoder layer and a multilayer perceptron A connected sequentially. The transformer encoder layer is used to extract time-series features, and the multilayer perceptron A is used to perform dimensionality reduction processing on the extracted time-series features to generate node features of multiple ROIs.
[0011] Furthermore, the method includes the following steps:
[0012] The first step is to extract time-series features from the Bold signal sequence acquired by functional nuclear magnetic resonance imaging using the encoder layer of a transformer.
[0013] The second step is to use a multilayer perceptron A to reduce the dimensionality of the extracted time series features;
[0014] The third step is to use a graph generator to transform the dimensionality-reduced time series features into a connected matrix of brain regions.
[0015] The initial features of the brain regions are obtained by assigning initial values to each node of the brain region using the row vectors of the brain region connectivity matrix calculated based on the Pearson correlation coefficient method.
[0016] The fourth step involves multiplying the initial features of the brain region with the connected matrix generated by the graph generator to obtain a functional brain network graph, which is then input into a graph convolutional neural network.
[0017] The fifth step involves using a graph convolutional neural network to extract node features from the functional brain network graph, then concatenating all node features to generate the features of the entire graph, which are then input into a multilayer perceptron B. The multilayer perceptron B then outputs the classification results of the brain network.
[0018] Furthermore, in the third step, the node features F of node p are... p Initialized as: a vector of time series data of node p and Pearson correlation scores between all nodes in the functional brain network graph.
[0019] Furthermore, the system is constructed based on an end-to-end framework. The system is trained and optimized through a downstream functional brain network classification task in the end-to-end framework. The training loss includes inter-class loss, inter-class loss, matrix sparsity loss, and cross-entropy loss. Inter-class loss is used to minimize the difference between connectivity matrices within the same class. Inter-class loss is used to maximize the difference between connectivity matrices between different classes while maintaining the similarity of connectivity matrices within the same class. Matrix sparsity loss is used to mitigate the bias caused by large values in the functional brain network graph.
[0020] Furthermore, let the loss between similar categories be L. SCL Given a class c∈C and a set This set contains all indices of samples with label c, L SCL The calculation formula is as follows:
[0021]
[0022]
[0023] in:
[0024] C represents the set of all test samples;
[0025] c represents the test sample;
[0026] q represents the index of the test sample;
[0027] Y q,c Labels indicating the test samples;
[0028] k represents the index of the sample's functional brain network map;
[0029] A k A functional brain network diagram generated by a graph generator;
[0030] S c A functional brain network diagram representing all samples with the same label;
[0031] μ c This represents the mean of samples from different classes;
[0032] This represents the variance of samples with the same label.
[0033] Furthermore, let the loss between different classes be L. DCL L DCL The calculation formula is as follows:
[0034]
[0035] in:
[0036] e represents the index of the first type of sample;
[0037] f represents the index of the second type of sample;
[0038] 'a' represents the first type of sample;
[0039] b represents the second type of sample;
[0040] A e A functional brain network diagram representing the first type of sample;
[0041] A f Functional brain network diagram representing the second type of sample;
[0042] The variance of the functional brain network map of the first type of sample;
[0043] The variance of the functional brain network map of the second type of sample;
[0044] S a This represents the set of all functional brain network maps for the first type of samples;
[0045] S bThis represents the set of all functional brain network maps for the second type of samples;
[0046] μ a This represents the mean of the functional brain network maps of the first type of samples;
[0047] μ b This represents the mean of the functional brain network maps of the second type of samples.
[0048] Furthermore, the embedding layer at the position where the input data is added is removed from the encoder layer of the transformer, and the number of attention heads is set to 1.
[0049] The advantages and positive effects of this invention are as follows: This invention proposes a brain network generation system based on a functional brain network classification task. This system includes a time-series encoder, a graph generator, and a graph predictor based on a graph convolutional neural network. It generates a correlation matrix of the brain network by directly extracting functional magnetic resonance imaging (fMRI) oxygenation level-dependent (BOLD) signal sequences from each brain region, thereby improving the classification accuracy of functional brain networks and generating the correlation matrix for subsequent analysis. This system is compatible with graph convolutional neural networks to improve the accuracy of functional brain network classification. Compared to existing work, this invention designs three modules—a time-series encoder, a graph generator, and a graph predictor—that can extract the nonlinear and temporal relationships of fMRI oxygenation level-dependent (BOLD) signal sequences from each brain region, thus improving the model's accuracy. Furthermore, this invention can also generate the correlation matrix of the brain network. Attached Figure Description
[0050] Figure 1 This is a schematic diagram of a brain network generation system based on a functional brain network classification task according to the present invention.
[0051] Figure 2 This is a block diagram of the time series encoder structure in this invention; N × This indicates the number of encoder layers in the transformer.
[0052] Figure 3 This is a diagram of the graph convolutional neural network structure in this invention. Detailed Implementation
[0053] To further understand the invention's content, features, and effects, the following embodiments are provided, and detailed descriptions are given below in conjunction with the accompanying drawings:
[0054] The Chinese definitions of the following English abbreviations and phrases in this application are as follows:
[0055] GCN: Graph Convolutional Neural Network.
[0056] Transformer: A sequence model proposed by Google.
[0057] fMRI: Functional magnetic resonance imaging data.
[0058] BOLD: Blood oxygen level dependent signal.
[0059] MLP: Multilayer Perceptron, is a type of fully connected neural network.
[0060] ReLU: The activation function in neural networks.
[0061] BatchNorm1D: Regularization.
[0062] ROI: Region of Interest in the brain.
[0063] Softmax: Normalization function.
[0064] readout: The operation of splicing node features in the graph.
[0065] like Figures 1 to 3 As shown, a brain network generation system based on a functional brain network classification task is disclosed. The system includes a time-series encoder, a graph generator, and a graph predictor based on a graph convolutional neural network, connected in sequence. The time-series encoder is used to encode Bold signal sequences. The graph generator is used to convert the encoded time-series features into a functional brain network graph. The graph predictor is used to classify and predict the functional brain network graph. The time-series encoder includes a transformer encoder layer and a multilayer perceptron A connected in sequence. The transformer encoder layer is used to extract time-series features, and the multilayer perceptron A is used to generate node features of multiple ROIs after dimensionality reduction of the extracted time-series features.
[0066] Preferably, the graph generator can generate a machine learning-compatible connectivity matrix A, where the elements of the connectivity matrix A can be the pairwise connectivity strengths between ROIs; Where h A =softmax(h e ), h e h is the time series feature extracted by the time series encoder. A These are normalized time series features.
[0067] Preferably, the graph predictor may include a graph convolutional neural network, a regularization layer, and a multilayer perceptron B connected in sequence; the graph convolutional neural network can be used to extract node features of a functional brain network graph; the multilayer perceptron B can be used to classify and predict the node features.
[0068] This invention also provides a brain network generation method based on a functional brain network classification task, constructing a brain network generation system based on a functional brain network classification task. This system includes a time-series encoder, a graph generator, and a graph predictor based on a graph convolutional neural network, connected sequentially. The time-series encoder is used to encode Bold signal sequences; the graph generator is used to convert the encoded time-series features into a functional brain network graph; the graph predictor is used to classify and predict the functional brain network graph. The time-series encoder includes a transformer encoder layer and a multilayer perceptron A connected sequentially. The transformer encoder layer is used to extract time-series features, and the multilayer perceptron A is used to perform dimensionality reduction processing on the extracted time-series features to generate node features of multiple ROIs.
[0069] Preferably, the method may include the following steps:
[0070] The first step is to extract time-series features from the Bold signal sequence acquired by functional nuclear magnetic resonance imaging using the encoder layer of a transformer.
[0071] The second step is to use a multilayer perceptron A to reduce the dimensionality of the extracted time series features;
[0072] The third step is to use a graph generator to transform the dimensionality-reduced time series features into a connected matrix of brain regions.
[0073] The initial features of the brain regions are obtained by assigning initial values to each node of the brain region using the row vectors of the brain region connectivity matrix calculated based on the Pearson correlation coefficient method.
[0074] The fourth step involves multiplying the initial features of the brain region with the connected matrix generated by the graph generator to obtain a functional brain network graph, which is then input into a graph convolutional neural network.
[0075] The fifth step involves using a graph convolutional neural network to extract node features from the functional brain network graph, then concatenating all node features to generate the features of the entire graph, which are then input into a multilayer perceptron B. The multilayer perceptron B then outputs the classification results of the brain network.
[0076] Preferably, in the third step, the node feature F of node p can be... p Initialized as: a vector of time series data of node p and Pearson correlation scores between all nodes in the functional brain network graph.
[0077] Preferably, the system can be constructed based on an end-to-end framework. The system can be trained and optimized through downstream functional brain network classification tasks in the end-to-end framework. The training loss can include inter-class loss, inter-class loss, matrix sparsity loss, and cross-entropy loss. Inter-class loss is used to minimize the difference between connectivity matrices within the same class. Inter-class loss is used to maximize the difference between connectivity matrices between different classes while maintaining the similarity of connectivity matrices within the same class. Matrix sparsity loss is used to mitigate the bias caused by large values in the functional brain network graph.
[0078] Preferably, the loss between similar classes can be set as L. SCL Given a class c∈C and a set This set contains all indices of samples with label c, L SCL The calculation formula is as follows:
[0079]
[0080]
[0081] in:
[0082] C represents the set of all test samples;
[0083] c represents the test sample;
[0084] q represents the index of the test sample;
[0085] Y q,c Labels indicating the test samples;
[0086] k represents the index of the sample's functional brain network map;
[0087] A k A functional brain network diagram generated by a graph generator;
[0088] S c A functional brain network diagram representing all samples with the same label;
[0089] μ c This represents the mean of samples from different classes;
[0090] This represents the variance of samples with the same label.
[0091] Preferably, the loss between different classes can be set as L. DCL L DCL The calculation formula is as follows:
[0092]
[0093] in:
[0094] e represents the index of the first type of sample;
[0095] f represents the index of the second type of sample;
[0096] 'a' represents the first type of sample;
[0097] b represents the second type of sample;
[0098] A e A functional brain network diagram representing the first type of sample;
[0099] A f Functional brain network diagram representing the second type of sample;
[0100] The variance of the functional brain network map of the first type of sample;
[0101] The variance of the functional brain network map of the second type of sample;
[0102] S a This represents the set of all functional brain network maps for the first type of samples;
[0103] S b This represents the set of all functional brain network maps for the second type of samples;
[0104] μ a This represents the mean of the functional brain network maps of the first type of samples;
[0105] μ b This represents the mean of the functional brain network maps of the second type of samples.
[0106] Preferably, the initial node features of the ROI can be utilized. The graph embedding is obtained by concatenating the nodes with the learnable connectivity matrix A from the graph generator.
[0107] Preferably, the embedding layer at the position where the input data is added in the encoder layer of the transformer can be removed, and the number of attention heads can be set to 1.
[0108] The working principle of the present invention will be further explained below with reference to a preferred embodiment:
[0109] A brain network generation system based on a functional brain network classification task is disclosed. The system comprises a time-series encoder, a graph generator, and a graph predictor based on a graph convolutional neural network, connected sequentially. The time-series encoder encodes Bold signal sequences; the graph generator converts the encoded time-series features into a functional brain network graph; and the graph predictor performs classification prediction on the functional brain network graph. The time-series encoder includes a transformer encoder layer and a multilayer perceptron A connected sequentially. The transformer encoder layer extracts time-series features, and the multilayer perceptron A performs dimensionality reduction on the extracted time-series features to generate node features for multiple regions of interest (ROIs). The structure and working principle of the time-series encoder, graph generator, and graph predictor are as follows:
[0110] 1. Time series encoder
[0111] BOLD signal sequences are a type of time series data. The main difference between BOLD signal sequences and ordinary time series data is that BOLD is a set of related sequences, rather than independent sequences. Traditional BOLD encoder methods, such as ICA (Independent Component Analysis) and PCA (Principal Component Analysis), ignore the temporal order. Principal Component Analysis and Independent Component Analysis can only capture linear information within or across time series. Furthermore, Pearson correlation is often used to construct functional brain network maps from ICA and PCA results. However, due to negative edge weights, functional brain network maps generated in this way are unaware of downstream tasks and are incompatible with graph convolutional neural networks. To address this issue, this invention employs a transformer encoder layer to encode the BOLD signal sequence, with the structure as follows: Figure 2 As shown. This invention makes appropriate modifications to the encoder layer of the transformer, removing the position embedding that is added to the input data, and setting the number of attention heads to 1. Because time series data only has correlation in the time dimension, the single attention head mechanism can capture the temporal correlation of the sequence.
[0112] Specifically, when the feature encoder is set to the encoder layer of a u-layer transformer, it generates node features for v ROIs. The process can be broken down into the following formula:
[0113] h u =transformer_encoder u (h u-1 ),
[0114] in,
[0115] h u For the extracted time series features;
[0116] h u-1 These are the time series features extracted in the previous iteration;
[0117] d represents the dimension of the time series feature;
[0118] transformer_encoder is the proposed time-series encoder;
[0119] MLP stands for Multilayer Perceptron Model, used to reduce the dimensionality of features.
[0120] 2. Graph Generator
[0121] The graph generator is based on the encoded time series features h. e This generates a task-oriented functional brain network map for subsequent graph predictors. Let the generated connectivity matrix be A, which uses the pairwise connectivity strength between ROIs as its elements. Unlike traditional methods that typically use pairwise Pearson correlations between the raw time series of ROIs to construct functional brain network maps, this invention generates a learnable connectivity matrix, i.e. Where h A =softmax(h e A is the learnable connectivity matrix generated by the graph generator, and h is the number of nodes in the graph. e h is the feature extracted by the time series encoder. A The time-series features are normalized. The learnable connectivity matrix A can be trained and optimized through a downstream functional brain network classification task in an end-to-end framework.
[0122] The Softmax (normalization) operation highlights important ROI connections by generating larger positive edge weights, which are compatible with GCNs and valuable for interpreting brain region connectivity. In contrast, brain connectivity matrices generated by traditional (i.e., statistical methods) methods are incompatible with GCNs because they contain negative weights at the edges. To facilitate the learning of brain networks,
[0123] Existing experimental results indicate that there are marginal differences among individuals in the structural and functional brain connectivity matrix.
[0124] Inspired by this experimental result, this invention further designed three loss terms and applied them to the generator model during training, named as inter-class loss, inter-class loss, and matrix sparsity loss, respectively.
[0125] Experiments have revealed consistent resting-state functional connectivity patterns among certain individuals. To optimize the generated brain network map using these latent consistent patterns as a loss term, this invention proposes a Same Class Loss (SCL), which aims to minimize the differences in connectivity matrices between samples of the same class. Subject samples with the same predicted label are referred to as one class.
[0126] Let the loss between similar classes be L. SCL Given a class c∈C and a set This set contains all indices of samples with label c, and the average u of the learnable graph generated by the graph generator. c ,variance and L SCL Calculate using the following formula:
[0127]
[0128]
[0129] in:
[0130] C represents the set of all test samples;
[0131] c represents the test sample;
[0132] q represents the index of the test sample;
[0133] Y q,c Labels indicating the test samples;
[0134] k represents the index of the sample's functional brain network map;
[0135] A k A functional brain network diagram generated by a graph generator;
[0136] S c A functional brain network diagram representing all samples with the same label;
[0137] μ c This represents the mean of samples from different classes;
[0138] This represents the variance of samples with the same label.
[0139] Different Class Loss (DCL). Cognitive science has confirmed significant differences in the functional brain networks of different individuals. Therefore, this invention introduces a different class loss to maximize the difference in connection matrices between different classes while maintaining the similarity of connection matrices within the same class. The specific formula is shown below:
[0140]
[0141] Where A e A f Samples representing different classes; S represents the variance of different classes. a S b μ represents all samples from different classes. a μ b This represents the mean of samples from different classes.
[0142] in:
[0143] e represents the index of the first type of sample;
[0144] f represents the index of the second type of sample;
[0145] 'a' represents the first type of sample;
[0146] b represents the second type of sample;
[0147] A e A functional brain network diagram representing the first type of sample;
[0148] A f Functional brain network diagram representing the second type of sample;
[0149] The variance of the functional brain network map of the first type of sample;
[0150] The variance of the functional brain network map of the second type of sample;
[0151] S a This represents the set of all functional brain network maps for the first type of samples;
[0152] S b This represents the set of all functional brain network maps for the second type of samples;
[0153] μ a This represents the mean of the functional brain network maps of the first type of samples;
[0154] μ b This represents the mean of the functional brain network maps of the second type of samples.
[0155] Matrix Sparsity Loss (SL). System models with only class loss may overemphasize differences between individuals, potentially harming the model's performance and stability. To mitigate the bias caused by large values in the generated graph and highlight the most important task-specific ROI connections, this invention further enhances the sparsity of the generated brain network, where the matrix sparsity loss is shown in the following formula:
[0156]
[0157] Where A ij Represents the weights in the generated brain connectivity network, and v represents the number of brain regions. i represents the row index in the functional brain network diagram; j represents the column index in the functional brain network diagram; L SL This represents the loss due to matrix sparsity.
[0158] Ultimately, our training objective is to minimize these three losses by optimizing the generated brain network matrix through a downstream functional brain network classification task.
[0159] 3. Graph Predictor
[0160] For graph predictors, this invention employs GCN (Graph Convolutional Neural Network) as follows: Figure 3 As shown, it learns node representations by transforming and propagating node features and structural information on the graph. In a specific implementation, this invention transforms the node features F of node p... p Initialize it as a vector of Pearson correlation scores between its time series and all nodes contained in the graph. Utilize the initial node features of the ROI. Given a learnable connectivity matrix A from the graph generator, this invention uses a k-layer graph convolutional network to predict the graph. Since the node order is fixed, permutation invariance is not a concern on the mind map. Graph-level embeddings can be obtained using a readout (node concatenation operation), where node features are concatenated with all node embeddings after the last convolutional layer. A BatchNorm1D (regularization) layer is further applied to avoid excessively large values. Finally, an MLP (Multilayer Perceptron) layer is used for classification prediction, as shown in the following formula.
[0161] h m =ReLU(Ah) m-1 W m );
[0162] h 0 =F p ;
[0163] Where A is the generated learnable connectivity matrix; ReLU is the activation function.
[0164]
[0165] in: represents the predicted sample label; v represents the number of brain regions; The brain region node features are represented by BatchNorm1D, which is a regularization layer; and MLP is a multilayer perceptron.
[0166] m represents the number of iterations;
[0167] P indicates Pearson correlation;
[0168] h m This represents the features of brain network nodes in the m-th iteration;
[0169] h m-1 This represents the features of brain network nodes in the (m-1)th iteration;
[0170] W m This represents the learnable parameters in the convolutional layer;
[0171] h 0 The initial features of the nodes are represented by the vector F of the Pearson correlation scores among all nodes. p initialization.
[0172] 4. End-to-end training
[0173] This invention combines a time series encoder, a graph generator, and a graph predictor into an end-to-end framework, such as... Figure 1 As shown, another advantage of end-to-end training is that, compared to a simple GCN prediction model, the end-to-end framework provides a larger parameter search space, thus potentially leading to performance improvements. In summary, the final training objective consists of four losses: L = L ce +L SCL +L DCL +L SL L ce It is the cross-entropy loss used for prediction; L SCL It is a loss between similar entities; L DCL It is an inter-class loss; L SL It is the loss due to matrix sparsity.
[0174] Although preferred embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these modifications are within the scope of protection of the present invention.
Claims
1. A brain network generation system based on a functional brain network classification task, characterized in that, The system comprises a time-series encoder, a graph generator, and a graph predictor based on a graph convolutional neural network, connected in sequence. The time-series encoder encodes Bold signal sequences; the graph generator transforms the encoded time-series features into a functional brain network map; and the graph predictor performs classification prediction on the functional brain network map. The time-series encoder includes a transformer encoder layer and a multilayer perceptron A connected in sequence. The transformer encoder layer extracts time-series features, and the multilayer perceptron A performs dimensionality reduction on the extracted time-series features to generate node features for multiple regions of interest (ROIs). The graph generator transforms the dimensionality-reduced time-series features into a connected matrix of brain regions.
2. The brain network generation system based on functional brain network classification task according to claim 1, characterized in that, Graph generators produce machine-readable connected matrices. , connected matrix The middle element represents the pairwise connectivity strength between ROIs; ,in , The time series features extracted by the time series encoder. These are normalized time series features.
3. The brain network generation system based on functional brain network classification task according to claim 1, characterized in that, The graph predictor consists of a graph convolutional neural network, a regularization layer, and a multilayer perceptron B connected in sequence; the graph convolutional neural network is used to extract node features from the functional brain network graph; and the multilayer perceptron B is used to classify and predict the node features.
4. A brain network generation method based on a functional brain network classification task, characterized in that, A brain network generation system based on a functional brain network classification task is constructed. This system includes a time-series encoder, a graph generator, and a graph predictor based on a graph convolutional neural network, connected sequentially. The time-series encoder encodes Bold signal sequences; the graph generator transforms the encoded time-series features into a functional brain network graph; and the graph predictor performs classification prediction on the functional brain network graph. The time-series encoder includes a transformer encoder layer and a multilayer perceptron A connected sequentially. The transformer encoder layer extracts time-series features, and the multilayer perceptron A performs dimensionality reduction on the extracted time-series features to generate node features for multiple regions of interest (ROIs). The graph generator transforms the dimensionality-reduced time-series features into a connected matrix of brain regions.
5. The brain network generation method based on functional brain network classification task according to claim 4, characterized in that, The method includes the following steps: The first step is to extract time-series features from the Bold signal sequence acquired by functional nuclear magnetic resonance imaging using the encoder layer of a transformer. The second step is to use a multilayer perceptron A to reduce the dimensionality of the extracted time series features; The third step is to use a graph generator to transform the dimensionality-reduced time series features into a connected matrix of brain regions. The initial features of the brain regions are obtained by assigning initial values to each node of the brain region using the row vectors of the brain region connectivity matrix calculated based on the Pearson correlation coefficient method. The fourth step involves multiplying the initial features of the brain region with the connected matrix generated by the graph generator to obtain a functional brain network graph, which is then input into a graph convolutional neural network. The fifth step involves using a graph convolutional neural network to extract node features from the functional brain network graph, then concatenating all node features to generate the features of the entire graph, which are then input into a multilayer perceptron B. The multilayer perceptron B then outputs the classification results of the brain network.
6. The brain network generation method based on functional brain network classification task according to claim 5, characterized in that, In the third step, the nodes Node features Initialize as: node The vector of Pearson correlation scores between all nodes in the time series and functional brain network diagram.
7. The brain network generation method based on functional brain network classification task according to claim 4, characterized in that, The system is constructed based on an end-to-end framework. It is trained and optimized by a downstream functional brain network classification task in the end-to-end framework. The training loss includes inter-class loss, inter-class loss, matrix sparsity loss, and cross-entropy loss. The inter-class loss is used to minimize the difference between the connectivity matrices of the same class. The inter-class loss is used to maximize the difference between the connection matrices of different classes while maintaining the similarity of the connection matrices of the same class. Matrix sparsity loss is used to mitigate the bias caused by large values in functional brain network maps.
8. The brain network generation method based on functional brain network classification task according to claim 7, characterized in that, Let the loss between similar categories be... Given a certain class and set The collection contains items with tags. All indices of the samples, The calculation formula is as follows: ； ； in: C represents the set of all test samples; c represents the test sample; q represents the index of the test sample; Labels indicating the test samples; k represents the index of the sample's functional brain network map; A functional brain network diagram generated by a graph generator; A functional brain network diagram representing all samples with the same label; This represents the mean of samples from different classes; This represents the variance of samples with the same label.
9. The brain network generation method based on functional brain network classification task according to claim 7, characterized in that, ... The loss between different classes is , The calculation formula is as follows: ； in: e represents the index of the first type of sample; f represents the index of the second type of sample; 'a' represents the first type of sample; b represents the second type of sample; A functional brain network diagram representing the first type of sample; Functional brain network diagram representing the second type of sample; The variance of the functional brain network map of the first type of sample; The variance of the functional brain network map of the second type of sample; This represents the set of all functional brain network maps for the first type of samples; This represents the set of all functional brain network maps for the second type of samples; This represents the mean of the functional brain network maps of the first type of samples; This represents the mean of the functional brain network maps of the second type of samples.
10. The brain network generation method based on functional brain network classification task according to claim 4, characterized in that, Remove the embedding layer at the position where the input data is added in the encoder layer of the transformer, and set the number of attention heads to 1.