Dynamic functional connection classification method based on convolutional bidirectional gated recurrent unit

Abstract

The application discloses a dynamic functional connectivity classification method based on a convolution bidirectional gated recurrent unit and belongs to the field of brain science research. The method comprises the following steps: multi-scale topological feature extraction, bidirectional dependent spatiotemporal feature extraction and bidirectional dependent spatiotemporal feature fusion. First, a convolution neural network is used to extract multi-scale topological features from each time point functional connectivity network. Then, the bidirectional gated recurrent unit is used to extract bidirectional dependent spatiotemporal features from the multi-scale topological feature time series. Finally, the method uses a one-dimensional convolution neural network to fuse the forward and backward spatiotemporal features to obtain joint spatiotemporal features for classification, which are input to a fully connected neural network classifier for classification. Experimental results on multiple brain disease data sets show that the method has better classification performance compared with other dynamic functional connectivity classification methods.

Description

Technical Field

[0001] This invention belongs to the field of brain science research. It designs a dynamic functional connectivity classification method based on convolutional bidirectional gated recurrent units to target the computer-aided diagnosis of brain diseases based on dynamic functional connectivity. Background Technology

[0002] Resting-state functional magnetic resonance imaging (rs-fMRI) is a non-invasive neuroimaging technique that measures the blood oxygen level dependent (BOLD) signal of each region of interest (ROI) in the brain. The correlation coefficient (e.g., Pearson correlation coefficient) between BOLD signals of different ROIs quantifies the functional similarity between ROIs and is called functional connectivity (FC). Functional connectivity plays a crucial role in cognitive processes in the brain. Some studies have shown that many brain disorders, such as Attention Deficit Hyperactivity Disorder (ADHD), Autism Spectrum Disorder (ASD), and Alzheimer's Disease (AD), are accompanied by breaks and abnormalities in the FC. Therefore, research on FC classification has attracted much attention from researchers. They typically use machine learning methods to extract various features of the FC for classification, thereby extracting biomarkers for the diagnosis of brain diseases.

[0003] Functional connectivity (FC) comprises static functional connectivity (SFC) and dynamic functional connectivity (DFC). SFC quantifies the functional similarity between different regions of interest (ROIs) throughout the entire fMRI time series, while DFC quantifies the functional similarity between different ROIs within each time interval of the fMRI time series. Because DFC reflects the evolution of FC over time and is expected to provide richer biomarkers, DFC classification is generally considered more beneficial for computer-aided diagnosis of brain diseases, revealing the pathogenic causes of brain disorders.

[0004] Currently, many DFC (Discrete Familiarity) classification methods have emerged. These can be categorized into methods based on traditional machine learning and methods based on deep learning. In the early stages, researchers typically extracted features from DFC data using traditional machine learning methods such as Principal Component Analysis (PCA), Independent Component Analysis (ICA), and k-means, and then used Support Vector Machines (SVM) for classification. Although these methods are easy to construct and have good interpretability, the shallow features they extract have poor discriminative power and do not take into account the dynamic characteristics of DFC. Therefore, the performance of DFC classification methods based on traditional machine learning needs further improvement.

[0005] In recent years, deep learning methods have been used for DFC (Distributed Full-Frame) classification due to their ability to extract deep features from data. These methods mainly include three types: those based on Recurrent Neural Networks (RNNs), those based on Deep Autoencoders (AEs), and those based on Convolutional Neural Networks (CNNs). RNN-based methods flatten the upper triangular matrix of the FC matrix at each time step into a vector, and then use a Long Short-Term Memory (LSTM) network to extract its long-short-term dependencies for classification, improving classification performance. However, this method does not consider the topological features of the FC network at each time step, and the high dimensionality of the FC network leads to severe overfitting, making it difficult to extract long-term dependencies. AE-based methods use deep autoencoders to extract deep features from the FC at each time step, and then use these features to train a fully connected neural network for classification. This method reduces the feature dimensionality during the autoencoder feature extraction process, mitigating the overfitting of the fully connected neural network to some extent. CNN-based methods use one-dimensional convolutional kernels with sparse strategies to eliminate redundant information in the DFC data, thus mitigating the overfitting problem of deep models. While these methods achieve good classification performance, they still neglect the inherent multi-scale topological features and spatiotemporal features with bidirectional long-term dependencies in the DFC (Digital Flow Framework) during feature extraction. These features are considered to contain discriminative information related to brain diseases and play an important role in computer-aided diagnosis of brain diseases. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a convolutional bidirectional gated recurrent unit (CBGRU) for DFC classification, referred to as DFC-CBGRU. This method comprises three key operations: multi-scale topological feature extraction, bidirectional spatiotemporal feature extraction, and bidirectional spatiotemporal feature fusion. Specifically, the method first uses a convolutional neural network to extract multi-scale topological features from the functionally connected network at each time step; then, it utilizes a bidirectional gated recurrent unit (GRU) to extract bidirectional spatiotemporal features from the time series of multi-scale topological features; finally, it uses a one-dimensional convolutional neural network to fuse the forward and backward spatiotemporal features to obtain joint spatiotemporal features for classification, which are then input into a fully connected neural network classifier for classification.

[0007] The main idea behind this invention is that a convolutional bidirectional recurrent neural network (CBRNN) is composed of stacked CNNs and bidirectional RNNs, possessing the advantages of CNNs in extracting spatial features and bidirectional RNNs in extracting bidirectional dependency temporal features. The gating mechanisms in GRU and LSTM alleviate the gradient explosion and vanishing problems of traditional RNNs, enabling the extraction of long-term dependency information from sequential data. Therefore, CBGRU and CBLSTM have been applied in various fields in recent years, such as traffic prediction and object recognition tasks. Inspired by this, a CBGRU is proposed that can extract the spatiotemporal features of bidirectional dependencies in DFC from multiple topological scales.

[0008] A dynamic functional connectivity classification method based on convolutional bidirectional gated recurrent units mainly includes the following steps:

[0009] Step (1) Acquisition and preprocessing of rs-fMRI dataset

[0010] Step (1.1) Acquisition of rs-fMRI data: Acquire the ABIDE (Autism Brain Imaging Data Exchange) dataset, which contains fMRI data of 539 patients with autism spectrum (ASD) and 573 healthy controls;

[0011] Step (1.2) Data preprocessing: This includes steps such as slice timing correction, motion realignment, spatial normalization, spatial smoothing, nuisance signal removal, and registration.

[0012] Step (1.3) Selection of regions of interest: Using the Anatomical Automatic Labeling (ALL) template, 90 brain regions of interest were selected to obtain the average fMRI time-series signal of the corresponding brain region for each subject.

[0013] Step (1.4) Constructing Dynamic Functional Connectivity: First, the fMRI time series is divided into multiple intersecting fMRI subsets using a sliding window method. Then, the Pearson correlation coefficient between different BOLD signals within each fMRI subset is calculated as the functional connectivity strength at that time. Finally, the generated functional connectivity is stacked in chronological order to obtain dynamic functional connectivity. The obtained dynamic functional connectivity is then transformed using Fisher-r-to-z to ensure that all FC matrices follow a normal distribution.

[0014] Step (2) Data set partitioning: The resulting DFC data set is divided into training set, validation set and test set, with a data sample size ratio of 8:1:1.

[0015] Step (3) Classification of dynamic functional connections based on convolutional bidirectional gated recurrent units: such as Figure 2 As shown, the invented method takes DFC as input and outputs the predicted label of the subject. It mainly includes two processes: DFC multi-scale topological feature extraction and DFC bidirectional spatiotemporal feature extraction and fusion.

[0016] Step (3.1) Extraction of DFC multi-scale topological features: such as Figure 2 As shown, the DFC multi-scale topological feature extraction process consists of three branches, which extract node-level, module-level, and graph-level topological features from the filtered DFC, respectively. Then, the topological features from multiple scales at each time step are concatenated to obtain a multi-scale topological feature time series.

[0017] Step (3.1.1) Node-level topological feature extraction.

[0018] Step (3.1.2) Module-level topological feature extraction.

[0019] Step (3.1.3) Graph-level topological feature extraction.

[0020] Step (3.2) Extraction and fusion of spatiotemporal features of DFC bidirectional dependency: such as Figure 2 As shown, bidirectional spatiotemporal feature extraction and fusion includes two key operations: bidirectional spatiotemporal feature extraction and bidirectional spatiotemporal feature fusion.

[0021] Step (3.2.1) DFC bidirectional spatiotemporal feature extraction: such as Figure 2 As shown, a bidirectional GRU is used to extract bidirectional spatiotemporal features at each time step from a multi-scale topological feature time series. It includes forward spatiotemporal features and backward spatiotemporal features.

[0022] Step (3.2.2) DFC bidirectional spatiotemporal feature fusion: such as Figure 2 As shown, a one-dimensional convolutional neural network is used to fuse the bidirectional spatiotemporal features at each time step to obtain joint spatiotemporal features.

[0023] Step (3.2.3) Classification and Loss Function Calculation: The joint spatiotemporal features are input into the subsequent fully connected neural network for feature extraction, and the label is predicted using softmax. The cross-entropy loss between the true label and the predicted label is calculated. The weighted L2 regularization term of the model parameters is calculated. The cross-entropy loss and the weighted regularization term are added together to obtain the overall loss function of the DFC-CBGRU.

[0024] Step (3.3) Training of DFC-CBGRU: Reliable five-fold cross-validation was performed on DFC-CBGRU. In each fold cross-validation, the Adam adaptive optimization algorithm was used to minimize the loss function of DFC-CBGRU, and the structure and hyperparameters of the neural network were determined based on the classification accuracy of DFC-CBGRU on the validation set. The classification performance of DFC-CBGRU with determined parameters was tested on the test set. The results of the five tests were averaged to obtain the diagnostic performance of the proposed method for ASD patients. This helps in the detection and diagnosis of brain diseases.

[0025] Compared with existing methods, the present invention has the following significant advantages and beneficial effects:

[0026] (1) A novel dynamic functional connectivity classification method based on convolutional bidirectional GRU is proposed, which can extract bidirectional spatiotemporal features of DFC from multiple topological scales.

[0027] (2) This invention proposes a convolution process to extract multi-scale topological features composed of node-level, module-level, and graph-level topological features from the FC network at each time step. This can more comprehensively reflect the global and local spatial topological features of the FC network at each time step.

[0028] (3) This invention proposes a spatiotemporal feature extraction process to extract bidirectional dependent spatiotemporal features from the obtained multi-scale topological feature time series, which can more fully reflect the changes in the state of FC network over time.

[0029] (4) Experimental results on the ABIDE dataset show that this method can classify DFC well and can provide an auxiliary means for the diagnosis of brain diseases. Attached Figure Description

[0030] Figure 1 A diagram illustrating the DFC construction process.

[0031] Figure 2 This is a model diagram of DFC-CBGRU.

[0032] Figure 3 This is a graph for extracting node-level topological features.

[0033] Figure 4 This is a module-level topology feature extraction map.

[0034] Figure 5 This is a graph-level topological feature extraction graph.

[0035] Figure 6 This is a diagram of the GRU architecture. Detailed Implementation

[0036] The following uses the real fMRI dataset ABIDE as an example to illustrate the specific implementation steps of this invention:

[0037] Step (1) Acquisition and preprocessing of rs-fMRI data set:

[0038] Step (1.1) Acquisition of resting-state functional magnetic resonance imaging data: Download the ABIDE (Autism Brain Imaging Data Exchange) dataset from the PCP website (http: / / fcon_1000.projects.nitrc.org / indi / abide / ), which contains rs-fMRI data of 539 autistic patients and 573 healthy controls;

[0039] Step (1.2) Data Preprocessing: For each rs-fMRI data point, preprocessing was performed using the Data Processing Assistant for Resting-state fMRI (DPARSF) software. To eliminate the influence of the subjects and the environment, the data from the first four time points for each subject were first removed. Then, DPARSF software was used to process the fMRI data of each subject, mainly including six key operations: slice timing correction, motion realignment, spatial normalization, spatial smoothing, nuisance signal removal, and registration.

[0040] Step (1.3) Selection of Regions of Interest: Real rs-fMRI data contains data from hundreds of thousands of voxels. Directly using voxel data to construct the DFC results in extremely high dimensionality. This makes it difficult for machine learning algorithms to learn the complex features. Therefore, this invention uses 90 regions of interest in the brain from the Anatomical Automatic Labeling (ALL) template to obtain the average fMRI time-series signal of the corresponding brain region for each subject. On the one hand, this can greatly reduce the dimensionality of the data and reduce the computational complexity of deep learning algorithms; on the other hand, each partition of the ALL template corresponds to the structure of the human brain, which is more realistic. Specifically, the processed fMRI data contains fMRI time-series data of each voxel. The average of the fMRI time-series data of each voxel in a brain region is used to obtain the fMRI time-series data of that brain region.

[0041] Step (1.4) Constructing Dynamic Function Connections: This invention uses the sliding window method to construct DFCs. Figure 1 The process of constructing dynamic functional connectivity is demonstrated. The fMRI time series composed of BOLD signals from 90 brain regions is represented as 90 curves of length Q. (Example...) Figure 1 As shown in Figure A, this invention first determines the parameters of the sliding window, including the sliding window size w and the sliding step size s, and accordingly divides the fMRI time series into T intersecting subsets. Figure 1 As shown in B, the Pearson correlation coefficients between different BOLD signals in each subset are then calculated to generate the corresponding FC matrix. Figure 1 As shown in C, all FC matrices are finally arranged in chronological order to obtain the DFC tensor. The formula used to determine T is as follows:

[0042] T=(Qw) / s (1)

[0043] Where Q represents the length of the fMRI time series, w represents the size of the sliding window, and s represents the sliding step size.

[0044] The formula for calculating the Pearson correlation coefficient during the DFC construction process is as follows:

[0045]

[0046] Where x i This represents the BOLD signal in the i-th brain region. Let represent the average value of the BOLD signal in the i-th brain region, and τ represent the time point. This represents the strength of the functional connectivity between the i-th brain region and the j-th brain region.

[0047] After obtaining the functional connectivity matrix for each sliding window, a Fisher-r-to-z transformation is applied to make each FC matrix follow a normal distribution. The corresponding calculation formula is as follows:

[0048]

[0049] FC ij This indicates the functional connection strength used in this invention.

[0050] Step (2) Partitioning of the data set;

[0051] To fully verify the effectiveness and advancement of this invention, five-fold cross-validation was performed. First, the constructed DFC dataset was divided into five equal parts. During the five-fold cross-validation process, four of these parts were merged into a training set, and the remaining part was divided into a test set and a validation set to test the effectiveness of the proposed method. Step (3) Dynamic functional connectivity classification based on convolutional bidirectional gated recurrent units;

[0052] like Figure 2 Before the DFC (Distributed Functional Array) is input into the multi-scale topological feature extraction module, an element-wise convolutional filter is used to filter out redundant information in the DFC, aiming to alleviate model overfitting. The convolutional kernel size of the element-wise filter is 90×90, and it performs element-wise multiplication with each FC matrix in the DFC tensor, assigning a unique weight to each DFC. Then, the ReLU activation function is used to remove redundant elements in the DFC, resulting in the filtered DFC.

[0053] Step (3.1) Extraction of DFC multi-scale topological features;

[0054] like Figure 2 As shown, using the filtered DFC as input, the multi-scale topology feature extraction module extracts the multi-scale topology features of the FC network at each time step, including node-level topology features, module-level topology features, and graph-level topology features.

[0055] Step (3.1.1) Node-level topological feature extraction: such as Figure 3 As shown, the convolution process for node-level topological feature extraction includes two convolution operations: edge-to-edge (E2E) convolution and element-wise edge-to-node (E2N-EW) convolution. First, the filtered DFC (Distributed Full-Focused Array) is input into the E2E convolution kernel to extract high-level edge features, as shown in the following formula:

[0056]

[0057] Where R and C represent the row and column convolution kernels in E2E, respectively, and D... h,t D represents the t-th FC matrix in the DFC tensor. h+1,t Let represent the t-th high-level edge feature matrix in the high-level edge feature tensor, N represent the number of brain regions, and the table below represents the corresponding brain region labels i, j, and k. F() is the ReLU activation function, and d is the learnable bias.

[0058] Then, this invention uses E2N-EW convolution kernels to extract node-level topological features from the high-level edge-level feature tensor, as shown in the following formula:

[0059]

[0060] Where E is the weight matrix of the E2N-EW convolution kernel, and n represents the node-level feature vector.

[0061] Step (3.1.2) Module-level topological feature extraction:

[0062] like Figure 4 As shown, before extracting module-level topological features, this invention constructs a module information tensor centered on brain regions, containing modular topological information centered on each brain region at each time step. The modular topological feature extraction process includes two key convolutional operations: Element-wise Edge to Module (E2M-EW) convolution and Element-wise Module to Graph (M2G-EW) convolution. The E2M convolution extracts module-level topological features centered on each node from the filtered modular information tensor. The specific formula is as follows:

[0063]

[0064] in This represents a modular information matrix for the k-th brain region at time point t. This represents the corresponding modular features centered on brain regions, Aij This represents the weight matrix of the E2M-EW convolution kernel.

[0065] Then, this invention uses M2G-EW convolutional kernels to extract module-level features of the FC network from the modular feature matrix centered on brain regions, as shown in the following formula:

[0066]

[0067] in Let represent the module-level topology characteristics of the FC network at time t. Let P represent the k-th weight vector of the M2G-EW convolution kernel, and P is a total of M weight matrices.

[0068] Step (3.1.3) Graph-level topological feature extraction: such as Figure 5 As shown, graph-level topological feature extraction involves two convolution operations: edge-to-edge (E2E) convolution and edge-to-graph (E2G) convolution. The E2E convolution operation is the same as the node-level topological feature extraction step. After obtaining the high-level edge features, graph-level features are extracted from them using E2G convolution kernels. The corresponding formula is as follows:

[0069]

[0070] in Let be the graph-level topological features at time t. B is the k-th weight matrix of the E2G convolution kernel, and there are a total of G weight matrices.

[0071] Step (3.2) Extraction and fusion of spatiotemporal features of DFC bidirectional dependency: After obtaining the three topological features at each time point, the node-level, module-level, and graph-level topological features of DFC at each time point are concatenated to obtain a multi-scale topological feature time series. The dimension of the multi-scale topological feature time series is (T×(N+M+G)), where T represents the number of time points of DFC, N is the dimension of node-level features, M is the dimension of module-level features, and G is the dimension of graph-level features.

[0072] Step (3.2.1) DFC bidirectional spatiotemporal feature extraction: such as Figure 2 As shown, after extracting the multi-scale topological feature time series, it is fed into a bidirectional GRU to extract bidirectional spatiotemporal features. Specifically, the bidirectional GRU includes a forward GRU and a backward GRU. The forward GRU inputs data sequentially to extract forward-dependent spatiotemporal features, while the backward GRU inputs data in reverse order to extract backward-dependent spatiotemporal features. Their mechanisms for extracting time-dependent features are the same.

[0073] Figure 6 This demonstrates the process of using a forward GRU to extract forward-dependent spatiotemporal features. X represents the latent information at time t-1, containing the spatiotemporal features prior to time t. t This is the input at time t, i.e., the multi-scale topological feature information at time t. Let X... t and As input, the reset gate infers How many of them are related to the input X at this time? t Enter the relevant information and save it in the candidate hidden state. The specific formula is as follows:

[0074] Reset Door:

[0075]

[0076] Reset operation:

[0077]

[0078] Where is the sigmoid activation function and tanh is the rectified linear activation function. To reset the weight of the door.

[0079] Using and as input, the updates are used to infer candidate hidden states in the current hidden state. With the hidden information of the previous time step The specific formula is as follows:

[0080] Update Gate:

[0081]

[0082] Update operation:

[0083]

[0084] in To update the weights of the gates, To update the gate output, This is the hidden output at this point.

[0085] Similar to the forward GRU, the backward GRU uses the latent information of the next time step and the multi-scale topological features at this time as inputs to derive the spatiotemporal features of the backward dependency.

[0086] Step (3.2.2) DFC bidirectional spatiotemporal feature fusion: such as Figure 2 As shown, after obtaining the forward and backward spatiotemporal features, they are further fused to obtain joint spatiotemporal features for classification. A one-dimensional convolutional neural network is used to fuse the bidirectional spatiotemporal features at each time step. The corresponding convolution formula is as follows:

[0087]

[0088] in and V represents the i-th element of the forward and backward spatiotemporal feature vectors at time t, respectively. K The weight matrix of the one-dimensional convolution kernel is O. k This represents the spatiotemporal characteristics after fusion.

[0089] After obtaining the joint spatiotemporal features of each time point in the DFC, they are input into a fully connected layer for dimensionality reduction and further feature extraction. Finally, the predicted label is obtained through the softmax activation function.

[0090] Step (3.2.3) Loss Function Calculation: After obtaining the predicted labels, the cross-entropy loss is calculated based on the true and predicted labels, and used as the objective function of the Adam algorithm to optimize the network connections of DFC-CBGRU. The optimization problem in the DFC-CBGRU connection optimization learning process is to minimize the loss function. The formula for the cross-entropy loss function is as follows:

[0091]

[0092] Where W and b represent the parameters of the proposed DFC-CBGRU. Let S represent the probability that the proposed method predicts the i-th sample as class u, and S represent the batch size. Furthermore, an L2 regularization term is used to further alleviate overfitting, which can limit all parameters to a very small interval; the regularization weight term is ε.

[0093] Step (3.3) Five-fold cross-validation of DFC-CBGRU: To fully verify the effectiveness and advancement of this invention, five-fold cross-validation is performed. First, the constructed DFC dataset is divided into five equal parts. During the five-fold cross-validation process, four of these parts are merged into a training set, and the remaining part is divided into a test set and a validation set to test the effectiveness of the proposed method.

[0094] In each fold cross-validation, the Adam adaptive optimization algorithm is used to minimize the loss function of DFC-CBGRU, and the structure and hyperparameters of the neural network are determined based on the classification accuracy of DFC-CBGRU on the validation set. The classification performance of DFC-CBGRU with determined parameters is then tested on the test set. The results of five tests are averaged to obtain the diagnostic performance of the proposed method for ASD patients. This contributes to the detection and diagnosis of brain diseases.

[0095] To verify the effectiveness of the method described in this invention, it was compared with different methods. The selected methods included traditional machine learning methods: Principal Component Analysis-Support Vector Machine (PCA-SVM), Independent Component Analysis-Support Vector Machine (ICA-SVM), and k-means-Support Vector Machine (k-means-SVM), and deep learning methods: Fully Connected Long Short-Term Memory (Full-BiLSTM) and Convolutional Neural Network with Sparse Strategy (SCNN). Seven classification metrics were selected for comparison: Accuracy (ACC), Sensitivity (SEN), Specificity (SPE), Precision (PRE), F1 score, Area Under the ROC Curve (AUC), and Youden index (YI).

[0096] Table 1. Performance of different DFC classification methods on the ABIDE dataset.

[0097]

[0098] Table 1 presents the comparative experimental results. The proposed method achieves the best classification performance on the ACC, SEN, PRE, F1, AUC, and YI indices. This indicates that the proposed method can extract disease-related spatiotemporal features from the DFC (Digital Functional Frame) and perform better DFC classification. Therefore, the method described in this invention can diagnose brain diseases based on the DFC and has great application prospects in computer-aided brain disease diagnosis.

Claims

1. A dynamic functional connectivity classification method based on convolutional bidirectional gated recurrent units, characterized in that, Includes the following steps: Step (1) Acquisition and preprocessing of rs-fMRI data set; Step (1.1) Acquisition of rs-fMRI data: Obtain the ABIDE dataset; Step (1.2) Data preprocessing: including inter-layer time correction, head motion correction, spatial normalization, spatial smoothing filtering, interference signal removal and registration; Step (1.3) Selection of regions of interest: Using 90 brain regions of interest from the anatomically autolabelled template, the average fMRI time-series signal of the corresponding brain region for each subject was obtained; Step (1.4) Constructing dynamic functional connections: First, the fMRI time series is divided into multiple intersecting fMRI subsets using the sliding window method; then, the Pearson correlation coefficient between different BOLD signals within each fMRI subset is calculated as the functional connection strength at that time; finally, the generated functional connections are stacked in chronological order to obtain dynamic functional connections; the obtained dynamic functional connections are transformed by Fisher-r-to-z to make all FC matrices follow a normal distribution. Step (2) Data set partitioning: The resulting DFC data set is divided into training set, validation set, and test set; Step (3) Dynamic functional connectivity classification based on convolutional bidirectional gated recurrent units: DFC is used as input and the predicted label of the subject is output; It includes two processes: DFC multi-scale topological feature extraction and DFC bidirectional spatiotemporal feature extraction and fusion; Step (3.1) Extraction of DFC multi-scale topological features: The DFC multi-scale topological feature extraction process includes three branches, which extract node-level, module-level, and graph-level topological features from the filtered DFC respectively; The topological features at multiple scales at each time point are concatenated to obtain a multi-scale topological feature time series. Step (3.2) Extraction and fusion of spatiotemporal features of DFC bidirectional dependency: Bidirectional spatiotemporal feature extraction and fusion includes two key operations, namely bidirectional spatiotemporal feature extraction and bidirectional spatiotemporal feature fusion; Step (3.3) Training DFC-CBGRU: Perform five-fold cross-validation on DFC-CBGRU.

2. The dynamic functional connectivity classification method based on convolutional bidirectional gated recurrent units according to claim 1, characterized in that, The implementation steps of step (3.1) are as follows: Step (3.1.1) Node-level topological feature extraction; Step (3.1.2) Module-level topological feature extraction; Step (3.1.3) Graph-level topological feature extraction.

3. The dynamic functional connectivity classification method based on convolutional bidirectional gated recurrent units according to claim 1, characterized in that, The implementation steps of step (3.2) are as follows: Step (3.2.1) DFC bidirectional spatiotemporal feature extraction: Use bidirectional GRU to extract the bidirectional spatiotemporal features of each time step from the multi-scale topological feature time series; including forward spatiotemporal features and backward spatiotemporal features; Step (3.2.2) DFC bidirectional spatiotemporal feature fusion: Use a one-dimensional convolutional neural network to fuse the bidirectional spatiotemporal features at each time step to obtain joint spatiotemporal features; Step (3.2.3) Classification and loss function calculation: The joint spatiotemporal features are input into the subsequent fully connected neural network for feature extraction and label prediction is performed through softmax; calculate the cross-entropy loss between the true label and the predicted label; calculate the weighted L2 regularization term of the model parameters; add the cross-entropy loss and the weighted regularization term as the loss function of the entire DFC-CBGRU.

4. The dynamic functional connectivity classification method based on convolutional bidirectional gated recurrent units according to claim 1, characterized in that, In step (3.3), in each fold cross-validation, the Adam adaptive optimization algorithm is used to minimize the loss function of DFC-CBGRU, and the structure and hyperparameters of the neural network are determined based on the classification accuracy of DFC-CBGRU on the validation set; the classification performance of DFC-CBGRU with determined parameters is tested on the test set, and the results of the five tests are averaged.

5. The dynamic functional connectivity classification method based on convolutional bidirectional gated recurrent units according to claim 1, characterized in that, In step (1.4), the parameters of the sliding window are first determined, including the sliding window size w and the sliding step size s. Based on this, the fMRI time series is divided into T intersecting subsets. The Pearson correlation coefficient between different BOLD signals in each subset is calculated to generate the corresponding FC matrix. All FC matrices are arranged in chronological order to obtain the DFC tensor. The formula used to determine T is as follows: T=(Qw)÷s (1) Where Q represents the length of the fMRI time series, w represents the size of the sliding window, and s represents the sliding step size; The formula for calculating the Pearson correlation coefficient during the DFC construction process is as follows: Where x i This represents the BOLD signal in the i-th brain region. Let represent the average value of the BOLD signal in the i-th brain region, and τ represent the time point. This represents the strength of the functional connectivity between the i-th brain region and the j-th brain region; After obtaining the functional connectivity matrix for each sliding window, the Fisher-r-to-z transformation is used to make each FC matrix follow a normal distribution. The calculation formula is as follows: FC ij Indicates the strength of the functional connection used.

6. The dynamic functional connectivity classification method based on convolutional bidirectional gated recurrent units according to claim 2, characterized in that, Step (3.1.1) Node-level topological feature extraction: The convolution process for node-level topological feature extraction includes two convolution operations: edge-to-edge convolution and element-wise edge-to-node convolution. First, the filtered DFC is input into the E2E convolution kernel to extract high-level edge features, as shown in the following formula: Where R and C represent the row and column convolution kernels in E2E, respectively, and D... h,t D represents the t-th FC matrix in the DFC tensor. h+1,t Let represent the t-th high-level edge feature matrix in the high-level edge feature tensor, N represent the number of brain regions, i, j, and k represent the corresponding brain region labels, F() is the ReLU activation function, and d is the learnable bias; node-level topological features are extracted from the high-level edge feature tensor using an E2N-EW convolution kernel, as shown in the following formula: Where E is the weight matrix of the E2N-EW convolution kernel, and n represents the node-level feature vector.

7. The dynamic functional connectivity classification method based on convolutional bidirectional gated recurrent units according to claim 2, characterized in that, Step (3.1.2) Module-level topological feature extraction: Before extracting module-level topological features, a module information tensor centered on brain regions is constructed, containing modular topological information centered on each brain region at each time step. The modular topological feature extraction process includes two key convolution operations: element-wise edge-to-module convolution and element-wise module-to-graph convolution. Among them, E2M convolution extracts module-level topological features centered on each node from the filtered modular information tensor. The specific formula is as follows: in This represents a modular information matrix for the k-th brain region at time point t. This represents the corresponding modular features centered on brain regions, A ij This represents the weight matrix of the E2M-EW convolution kernel; The M2G-EW convolutional kernel is used to extract module-level features of the FC network from the modular feature matrix centered on brain regions. The specific formula is as follows: in Let represent the module-level topology characteristics of the FC network at time t. Let P represent the k-th weight vector of the M2G-EW convolution kernel, and P is a total of M weight matrices.

8. The dynamic functional connectivity classification method based on convolutional bidirectional gated recurrent units according to claim 2, characterized in that, Step (3.1.3) Graph-level topological feature extraction: Graph-level topological feature extraction includes two convolution operations: edge-to-edge convolution and edge-to-graph convolution. The E2E convolution operation is the same as the node-level topological feature extraction step. After obtaining the high-level edge features, graph-level features are extracted from them using E2G convolution kernels, as shown in the following formula: in Let be the graph-level topological features at time t. B is the k-th weight matrix of the E2G convolution kernel, and there are a total of G weight matrices.

9. The dynamic functional connectivity classification method based on convolutional bidirectional gated recurrent units according to claim 1, characterized in that, Step (3.2) Extraction and fusion of spatiotemporal features of DFC bidirectional dependency: After obtaining the three topological features at each time point, the node-level, module-level, and graph-level topological features of DFC at each time point are spliced ​​together to obtain a multi-scale topological feature time series; the dimension of the multi-scale topological feature time series is (T×(N+M+G)), where T represents the number of time points of DFC, N is the dimension of node-level features, M is the dimension of module-level features, and G is the dimension of graph-level features.

10. The dynamic functional connectivity classification method based on convolutional bidirectional gated recurrent units according to claim 9, characterized in that, During the process of extracting forward-dependent spatiotemporal features using a forward-dependent GRU, The latent information at time t-1 contains the spatiotemporal features prior to time t; X t The input at time t is the multi-scale topological feature information at time t; with X... t and As input, the reset gate infers How many of them are related to the input X at this time? t Enter the relevant information and save it in the candidate hidden state. The specific formula is as follows: Reset Door: Reset operation: Where is the sigmoid activation function and tanh is the rectified linear activation function. To reset the weight of the door; Using sums as input, the update gate is used to infer candidate hidden states in the current hidden state. With the hidden information of the previous time step The specific formula is as follows: Update Gate: Update operation: in To update the weights of the gates, To update the gate output, This is the hidden output at this time; The backward GRU uses the latent information of the next time step and the multi-scale topological features at this time as input to derive the spatiotemporal features of the backward dependency. After obtaining the forward and backward spatiotemporal features, they are fused to obtain joint spatiotemporal features for classification; a one-dimensional convolutional neural network is used to fuse the bidirectional spatiotemporal features at each time step; the corresponding convolution formula is as follows: in and V represents the i-th element of the forward and backward spatiotemporal feature vectors at time t, respectively. K The weight matrix of the one-dimensional convolution kernel is O. k This represents the spatiotemporal characteristics after fusion; After obtaining the joint spatiotemporal features of each time point in the DFC, they are input into a fully connected layer for dimensionality reduction and further feature extraction. Finally, the predicted label is obtained through the softmax activation function.

Images