Neuroheterogeneity-guided dynamic brain network analysis method and system
By employing a time-series graph learning strategy guided by neural heterogeneity, dynamic brain networks are decomposed and weighted, addressing the problem of insufficient spatiotemporal heterogeneity capture in existing methods. This enables the identification of key nodes in brain network reorganization and the effective capture of spatiotemporal features, thereby improving the recognition performance of brain imaging data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-03
- Publication Date
- 2026-03-10
AI Technical Summary
Existing dynamic brain network analysis methods have failed to effectively capture the spatiotemporal heterogeneity of neural nodes in the brain, neglected the spatiotemporal coordination and sequence dependence of DFBNs, and limited in-depth exploration of spatiotemporal relationships within DFBNs.
A neural heterogeneity-guided temporal graph learning strategy is adopted. DFBNs are decomposed into topological consistency networks and temporal trend networks through a spatiotemporal pattern decoupling module. Their similarity is calculated and weighted. Combined with spatiotemporal heterogeneity weighting and time propagation graph convolutional network, the spatiotemporal features of heterogeneous DFBNs are captured.
It enables the identification of key nodes in brain network reorganization and the flexible capture of spatiotemporal features, improving the recognition performance of brain imaging data, especially in the analysis of brain disease imaging data.
Smart Images
Figure CN121639581A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of image processing, and particularly relates to an intelligent recognition technology of brain image data, and in particular to a neural heterogeneity guided dynamic brain network analysis method and system. BACKGROUND
[0002] Functional magnetic resonance imaging (fMRI) measures neural activity by detecting changes in blood oxygen level dependent signals and is often used to construct functional brain networks. In fact, the brain is constantly reorganizing even at rest. Obviously, compared with static functional brain networks, dynamic functional brain networks (DFBNs) can more comprehensively describe the topological evolution of the brain. Therefore, effective analysis of the spatiotemporal structure of DFBNs is crucial for brain image analysis.
[0003] To capture the time-varying structure of DFBNs, they are usually modeled as a series of dynamic brain graphs. In a dynamic brain graph, nodes represent brain regions and edges represent temporal connections between these regions. Existing dynamic brain graph analysis methods usually use graph convolutional networks (GCN) to extract topological features and then use temporal convolution to capture the temporal correlation between brain regions. Although these methods have made significant progress in the analysis of dynamic brain graphs, most of them ignore the key fact that the brain exhibits significant spatiotemporal heterogeneity: due to its functional characteristics, certain neural nodes in DFBNs exhibit extensive connectivity or more active temporal evolution. For example, the posterior cingulate cortex forms stable and tight connections with the frontal and parietal lobes, while the connection strength between the primary motor cortex and the supplementary motor cortex exhibits enhanced temporal variability. These neural nodes with high spatiotemporal variability can flexibly adjust the reconstruction pattern of functional networks, which is a key factor driving brain reorganization. Therefore, identifying the spatiotemporal heterogeneity of neural nodes is of great significance for elucidating the evolution mechanism of the brain.
[0004] However, accurately capturing the spatiotemporal coordination of neural nodes faces double challenges. (1) The spatiotemporal coordination of DFBNs. Although the key nodes maintain stable connections, the brain network can dynamically adjust the connections according to cognitive demands to achieve effective information integration. This spatial consistency and temporal trend together constitute the neural basis that supports complex cognitive functions. (2) The sequence dependence of DFBNs. Due to the continuity of brain activity and the hysteresis of information interaction, the current connection pattern is systematically influenced by the previous network state. Rich sequence dependence is exhibited between neural nodes.
[0005] To address these challenges, we propose a neural heterogeneity-guided temporal graph learning strategy (NeuroH-TGL) to comprehensively capture the intrinsic evolutionary mechanisms of DFBNs. Specifically, to simulate the spatiotemporal coordination of DFBNs, we design a spatiotemporal pattern decoupling module that decomposes DFBNs into topological consistency networks and temporal trend networks. Then, we calculate the similarity between the topological consistency network and the temporal trend network and use it as spatial and temporal heterogeneity weights. Subsequently, we apply spatiotemporal heterogeneity weights to the DFBNs to highlight key nodes driving network reorganization and construct a brain network that integrates heterogeneity. Finally, we develop a temporal propagation graph convolutional network to further capture the propagation mechanism of heterogeneous neural information in the current topology, thereby flexibly capturing the spatiotemporal features in heterogeneous DFBNs. Summary of the Invention
[0006] Purpose of the invention: To address the shortcomings of existing image processing technologies in exploring the spatiotemporal heterogeneity of brain imaging data, this invention provides a method and system for dynamic brain network analysis guided by neural heterogeneity.
[0007] Technical solution: A dynamic brain network analysis method guided by neural heterogeneity, the method comprising: S1. Acquire functional magnetic resonance imaging data as sample data, and then preprocess the sample data, including head motion correction and removal operations. S2. Use overlapping sliding windows to divide functional magnetic resonance imaging data to construct a dynamic functional brain network, and then decouple the dynamic functional brain network into a topologically consistent network and a temporal trend network that conform to brain activity. S3. Based on topological consistency networks and temporal trend networks, capture the spatial and temporal heterogeneity weights of different brain regions in the brain, thereby identifying key nodes that drive brain network reorganization. S4. Obtain a heterogeneous dynamic functional brain network by weighting the topologically consistent network and the time-trend network respectively. S5. Based on the convolution operation of the time propagation graph, the propagation of neural information in the time dimension is simulated. The heterogeneous spatiotemporal features of the brain network in the previous moment are used to guide the information aggregation of the brain network in the next moment, thereby realizing the extraction of spatiotemporal features of brain images from heterogeneous dynamic functional brain networks. S6. Input the spatiotemporal features of the obtained brain images into a multilayer perceptron to predict the category of brain image data to be identified, and analyze the influence of brain disease image data features on spatiotemporal heterogeneity of brain regions, thereby completing the dynamic brain network analysis guided by neural heterogeneity.
[0008] Furthermore, the preprocessing described in step S1 includes using SPM8, implemented in the DPARSF toolbox of MATLAB, to preprocess all functional magnetic resonance imaging data. After obtaining the original DICOM data, data that does not meet the noise requirements are removed. Then, the Realign procedure is applied to correct head motion and remove sample data that do not meet the head motion requirements. Finally, the T1-weighted structural image and functional image are registered.
[0009] Furthermore, the segmentation tools in DPARSF and the new segment+DARTEL tool were used to register T1-weighted structural images and functional images. Covariate regression was used to remove noise signals, including noise from white matter and cerebrospinal fluid in the functional magnetic resonance imaging data.
[0010] Further, in step S2, based on the preprocessed sample data, the data is divided into T subsequences using T overlapping windows. For each subsequence, the correlation coefficient between paired brain region subsequences is calculated using the Pearson correlation coefficient, thereby obtaining the dynamic functional brain network. To simulate the sparsity of brain networks, the connection strength is further reduced to less than... α The element is set to 0.
[0011] Step S2 for each brain network First, two independent graph convolutional networks are used to extract the topology-consistent network. and time trend network :
[0012]
[0013] in This represents a graph convolutional network. , I It is the identity matrix. This represents the degree matrix after adding self-loops. Represents the subsequence under the t-th window. , , and Each represents a learnable parameter matrix. It is an activation function; To enhance the discriminative power between topologically consistent networks and time-trending networks, similarity constraints between them are implemented. The structure is as follows: , in Indicates the number of brain regions. This represents the i-th brain region within the t-th window, and · represents the dot product operation. express Norm; To ensure the complementarity of the topologically consistent network and the time-trending network, the decoupled features are added together to obtain a reconstructed representation. Mean squared error is introduced as a factor in the reconstruction representation and its relationship to the brain network. Reconstruction loss between The expression is as follows:
[0014] in This represents element addition. Topological consistency refers to the high stability of certain network structures across different windows. Further, similarity constraints are imposed on topologically consistent networks in adjacent windows, and this similarity is encouraged. The similarity increases as training progresses. The expression is as follows:
[0015] The above operations ensure that topologically consistent networks and temporally trending networks are unique and complementary within the same window, while also promoting the similarity of consistent networks across different windows, thereby better aligning with the brain's intrinsic spatiotemporal coordination.
[0016] Further, step S3 includes calculating the cross-window similarity of the topologically consistent network and the temporally trending network to measure the connection density and temporal variability of the brain, respectively, thereby exploring the spatiotemporal heterogeneity of the brain; For spatial heterogeneity SH Calculate the average correlation between topologically consistent networks in all paired windows: , in Let T represent the cosine similarity, the i-th and j-th windows, and T represent the number of sliding windows. Let i and j represent the topologically consistent networks under the i-th and j-th windows, respectively.
[0017] Lower cross-window similarity in time-trend networks indicates more pronounced dynamic evolution, calculated using the following formula. TH :
[0018] in, These represent the time trend networks under the i-th and j-th windows, respectively; Through the above process, spatiotemporally heterogeneous brain regions that drive brain network reorganization are obtained.
[0019] Further, in step S4, after obtaining the spatiotemporal heterogeneity weights, the spatiotemporal heterogeneity weights are weighted onto the topological consistency network and the temporal trend network respectively to obtain the heterogeneous dynamic functional brain network; wherein, the heterogeneous dynamic functional brain network includes the heterogeneous topological consistency network. and heterogeneous time trend networks .
[0020] Further, the time-propagation graph convolution described in step S5 includes two graph convolutional networks and one two-dimensional convolutional network, and extracts features from the heterogeneous topological consistency network and the heterogeneous temporal trend network respectively through two spatiotemporal convolutional blocks with shared parameters; heterogeneous topological consistency features and heterogeneous time trend characteristics Learn in the following way: , in ; The first t The two types of features under each window are combined to obtain a fused spatiotemporal representation. After performing the same operation on the two networks for each window, the features of all windows are concatenated to obtain a global spatiotemporal representation. E : .
[0021] Furthermore, step S6 includes using cross-entropy loss. To supervise the update of model parameters, the overall training objective is expressed as: , in and It is a hyperparameter that controls the relative importance of different loss terms. and These represent the similarity loss and reconstruction loss of the spatiotemporal pattern decoupling module across all windows, respectively.
[0022] On the other hand, the present invention also provides a dynamic brain network analysis system guided by neural heterogeneity, the system being used to perform the above-described method, the system comprising: The spatiotemporal pattern decoupling module is used to simulate the spatiotemporal coordination of the brain, decoupling the dynamic functional brain network into a topologically consistent network and a time-trend network that conform to brain activity; The spatiotemporal heterogeneity mining module captures the spatial and temporal heterogeneity weights of different brain regions based on topological consistency networks and temporal trend networks, thereby identifying key nodes that drive brain network reorganization. The spatiotemporal heterogeneity weighting module weights the topologically consistent network and the temporally trending network respectively to obtain a heterogeneous dynamic functional brain network; The temporal propagation graph convolution module is used to simulate the propagation of neural information in the time dimension; The classification module uses a multilayer perceptron to map the heterogeneous spatiotemporal features of the brain image data to be identified to labels, and uses cross-entropy loss to train the classifier, thereby improving the recognition performance of the brain image data to be identified by narrowing the distance between the predicted labels and the true labels.
[0023] Functional magnetic resonance imaging (fMRI) measures neural activity by detecting changes in blood oxygen level-dependent signals and is often used to construct functional brain networks. In fact, even at rest, the brain is constantly reorganizing. Clearly, dynamic functional brain networks (DFBNs) can more comprehensively describe the topological evolution of the brain compared to static functional brain networks. Therefore, effectively analyzing the spatiotemporal structure of DFBNs is crucial for brain imaging analysis.
[0024] To capture the temporal variation of DFBNs, they are typically modeled as a series of dynamic brain maps. In these dynamic brain maps, nodes represent brain regions, and edges represent temporal connections between these regions. Existing dynamic brain map analysis methods typically use graph convolutional networks (GCNs) to extract topological features, followed by temporal convolution to capture the temporal correlations between brain regions. While these methods have made significant progress in the analysis of dynamic brain maps, most overlook the crucial fact that the brain exhibits significant spatiotemporal heterogeneity: due to their functional characteristics, certain neural nodes in DFBNs exhibit extensive connectivity or more active temporal evolution. For example, the posterior cingulate cortex forms stable and tight connections with the frontal and parietal lobes, while the connection strength between the primary motor cortex and the supplementary motor cortex shows enhanced temporal variability. These neural nodes with high spatiotemporal variability can flexibly adjust the reconstruction patterns of functional networks, which is a key factor driving brain reorganization. Therefore, identifying the spatiotemporal heterogeneity of neural nodes is crucial for elucidating the mechanisms of brain evolution.
[0025] However, accurately capturing the spatiotemporal coordination of neural nodes faces a dual challenge. (1) Spatiotemporal coordination of DFBNs. Although key nodes maintain stable connections, brain networks can dynamically adjust connections according to cognitive needs to achieve effective information integration. This spatial consistency and temporal trend together constitute the neural basis supporting complex cognitive functions. (2) Sequence dependence of DFBNs. Due to the continuity of brain activity and the lag in information interaction, current connection patterns are systematically influenced by previous network states. Neural nodes exhibit rich sequence dependencies.
[0026] To address these challenges, we propose a neural heterogeneity-guided temporal graph learning strategy (NeuroH-TGL) to comprehensively capture the intrinsic evolutionary mechanisms of DFBNs. Specifically, to simulate the spatiotemporal coordination of DFBNs, we design a spatiotemporal pattern decoupling module that decomposes DFBNs into topological consistency networks and temporal trend networks. Then, we calculate the similarity between the topological consistency network and the temporal trend network and use it as spatial and temporal heterogeneity weights. Subsequently, we apply spatiotemporal heterogeneity weights to the DFBNs to highlight key nodes driving network reorganization and construct a brain network that integrates heterogeneity. Finally, we develop a temporal propagation graph convolutional network to further capture the propagation mechanism of heterogeneous neural information in the current topology, thereby flexibly capturing the spatiotemporal features in heterogeneous DFBNs. Attached Figure Description
[0027] Figure 1 This is a flowchart of the implementation steps of the present invention; Figure 2 This is a framework diagram of the method described in this invention; Figure 3 It is a framework for mining spatiotemporal heterogeneity. Detailed Implementation
[0028] Dynamic Functional Brain Networks (DFBNs) are powerful tools in neuroscience research. Recent studies have shown that DFBNs contain heterogeneous neural nodes with broader connectivity and more dramatic temporal variations, which play a crucial role in coordinating brain reorganization. Furthermore, the spatiotemporal patterns of these nodes are modulated by historical brain states. However, existing methods not only neglect the spatiotemporal heterogeneity of neural nodes but also fail to effectively encode the temporal propagation mechanisms of heterogeneous activity. These limitations hinder in-depth exploration of the spatiotemporal relationships within DFBNs. To address these issues, we propose a neural heterogeneity-guided temporal map learning method for intelligent recognition processing of brain imaging data.
[0029] Figure 1 The implementation process of the method described in this invention is illustrated, including the following implementation steps: Step 1: Acquire functional magnetic resonance imaging (fMRI) data and perform the following preprocessing procedure on the data to obtain the preprocessed fMRI signal.
[0030] fMRI data preprocessing: All fMRI data were preprocessed using SPM8, implemented in the DPARSF toolbox in MATLAB.
[0031] After obtaining the raw DICOM data, we chose to delete the data from the first 10 time points because the data was quite noisy when it was first collected.
[0032] Subsequently, we applied the Realign procedure to correct head motion and removed samples with significant head motion.
[0033] Using the Bet function in the toolbox to remove scalp structures helps reduce the influence of non-brain tissues during registration, thereby improving registration accuracy.
[0034] We registered the T1-weighted structural image with the functional image.
[0035] Using the Segment and New segment+DARTEL tools in the toolbox, we registered the T1-weighted structural image with the functional image.
[0036] Nuisance covariates regression is used to further remove noise signals, such as noise from white matter and cerebrospinal fluid.
[0037] After preprocessing, the fMRI signals of each subject were divided into 90 brain regions using an automated anatomical atlas, with each brain region containing 197 time points.
[0038] Step 2: Based on each preprocessed fMRI signal, we divide it into T subsequences using T overlapping windows. For each subsequence, we calculate the correlation coefficient between paired brain region subsequences using the Pearson correlation coefficient, thereby obtaining the dynamic functional brain network. To simulate the sparsity of brain networks, we further increased the connection strength to less than... α The element is set to 0.
[0039] Step 3: Neuroscience research indicates that the brain's cognitive functions are supported by its inherent topological consistency and temporal trends. Therefore, effectively decoupling the spatiotemporal organization within dynamic functional brain networks helps reveal network dysregulation caused by brain diseases. For each brain network... First, two independent graph convolutional networks (GCNs) are used to extract the topology-consistent network. and time trend network : , , in I is the identity matrix. This represents the degree matrix after adding self-loops. , , and Each represents a learnable parameter matrix. To enhance the discriminative power between topologically consistent networks and time-trending networks, we encourage similarity constraints between them. Gradually decrease during training: , Where · represents the dot product operation. express Norm. Furthermore, to ensure the complementarity of topologically consistent and temporally trending networks, we sum the decoupled features to obtain a reconstructed representation. Then, we introduce mean squared error (MSE) as the reconstructed representation and the brain network. Reconstruction loss between:
[0040] in This represents element addition. It's worth noting that topological consistency refers to the high stability of certain network structures across different windows. Therefore, we further impose similarity constraints on topologically consistent networks in adjacent windows and encourage this similarity. Increases as training progresses:
[0041] The above operations ensure that topologically consistent networks and temporally trending networks are unique and complementary within the same window, while also promoting the similarity of consistent networks across different windows, thereby better aligning with the brain's intrinsic spatiotemporal coordination.
[0042] Step 4: In dynamic functional brain networks, some neural nodes exhibit high spatiotemporal variability, playing a crucial role in coordinating brain evolution. Therefore, we compute cross-window similarity for topologically consistent networks and temporally trending networks separately to measure brain connection density and temporal variability, thereby exploring spatiotemporal heterogeneity. For spatial heterogeneity (SH), we first calculate the average correlation between topologically consistent networks across all paired windows:
[0043] in This represents cosine similarity. In contrast, for temporal heterogeneity (TH), lower cross-window similarity in time-trending networks indicates more pronounced dynamic evolution. Therefore, we calculate TH using the following formula: ; Through the above steps, we can obtain the spatiotemporally heterogeneous brain regions that drive brain network reorganization.
[0044] Step 5: After obtaining the spatiotemporal heterogeneity weights, the spatiotemporal heterogeneity weights are applied to the topological consistency network and the temporal trend network respectively to obtain the heterogeneous dynamic functional brain network. The heterogeneous dynamic functional brain network includes the heterogeneous topological consistency network. and heterogeneous time trend networks .
[0045] Step 6: Heterogeneous information in the brain propagates continuously along the time dimension, meaning that each brain network gradually influences the state of adjacent brain networks. To flexibly capture cross-temporal interactions in heterogeneous brain networks, we further designed a temporally propagating graph convolutional network. This framework utilizes the heterogeneous spatiotemporal features of the brain network in the previous time step to guide the aggregation of information in the brain network in the next time step. In this framework, each spatiotemporal convolutional block consists of two graph convolutional networks (GCNs) and one two-dimensional convolutional network (CNN) to effectively aggregate dynamic structural information. To reduce the number of parameters, we use two parameter-shared spatiotemporal convolutional blocks to extract features from the heterogeneous topological consistency network and the heterogeneous temporal trend network, respectively. Heterogeneous topological consistency features and heterogeneous time trend characteristics You can learn in the following way: ; in Then, we sum the features of these two types under the t-th window to obtain the fused spatiotemporal representation. After performing the same operation on the two networks for each window, we concatenate the features of all windows to obtain a global spatiotemporal representation. E : ; Step 7: Finally, represent the global spacetime. E The data is fed into a multilayer perceptron to predict diagnostic results, and cross-entropy loss is used. To supervise the updating of model parameters. The overall training objective can be expressed as: ; in and It is a hyperparameter that controls the relative importance of different loss terms. and These represent the similarity loss and reconstruction loss of the spatiotemporal pattern decoupling module across all windows, respectively.
[0046] From a system architecture perspective, this invention includes a spatiotemporal pattern decoupling module, a spatiotemporal heterogeneity mining module, a spatiotemporal heterogeneity weighting module, a time propagation graph convolution module, and a classification module. The overall framework structure of the proposed method is as follows: Figure 2 As shown, the spatiotemporal pattern decoupling module is used to simulate the spatiotemporal coordination of the brain. Brain coordination refers to the brain's ability to dynamically adjust connections according to cognitive needs while maintaining stable connections at key nodes to achieve effective information integration. This spatial consistency and temporal trend together constitute the neural basis supporting complex cognitive functions. The spatiotemporal pattern decoupling module decouples the dynamic functional brain network into a topologically consistent network and a temporally trend-based network that conforms to brain activity. The spatiotemporal heterogeneity mining module, as shown... Figure 3 As shown, it captures the spatial and temporal heterogeneity weights of different brain regions based on topological consistency networks and temporal trend networks, thereby identifying key nodes driving brain network reorganization. Next, we weight the spatiotemporal heterogeneity weights onto the topological consistency network and the temporal trend network respectively to obtain a heterogeneous dynamic functional brain network. Due to the continuity of brain activity and the lag in information interaction, the current brain connectivity pattern is systematically influenced by the previous network state. Neural nodes exhibit rich sequence dependencies; therefore, we designed a temporal propagation graph convolution module to simulate the propagation of neural information in the temporal dimension, thereby effectively extracting spatiotemporal features from the heterogeneous dynamic functional brain network. The classification module uses a multilayer perceptron to map the heterogeneous spatiotemporal features of each brain image data to a label, and uses cross-entropy loss to train the classifier, improving the intelligent recognition performance of brain disease image data by narrowing the distance between the predicted label and the true label.
[0047] This embodiment uses only one layer of time-propagation graph convolution to extract spatiotemporal features, where the output dimension of the graph convolution is 16 and the temporal convolution kernel size is 3. 3. ReLU is used as the activation function in all spatiotemporal graph convolutional blocks. In the classifier, the number of neurons in the two fully connected layers are 32 and 2, respectively, and softmax is used as the activation function in the last layer.
Claims
1. A method of neural heterogeneity guided dynamic brain network analysis, characterized in that, The method comprises: S1, acquiring functional magnetic resonance imaging data as sample data, and then preprocessing the sample data, including head motion correction and rejection operation on the sample data; S2, using an overlapping sliding window to divide the functional magnetic resonance imaging data to construct a dynamic functional brain network, and then decoupling the dynamic functional brain network into a topological consistency network and a time trend network consistent with brain activity; S3, based on the topological consistency network and the time trend network, the spatial and temporal heterogeneity weights of different brain regions in the brain are captured, so as to identify the key nodes driving the brain network reorganization; S4, the topological consistency network and the time trend network are weighted respectively to obtain a heterogeneous dynamic functional brain network; S5, based on a time propagation graph convolution operation, the propagation of neural information in the time dimension is simulated, and the heterogeneous spatiotemporal features of the brain network at the previous moment are used to guide the information aggregation of the brain network at the next moment, so as to extract the spatiotemporal features of the brain image from the heterogeneous dynamic functional brain network; S6, the obtained spatiotemporal features of the brain image are input into a multilayer perception machine to predict the category of the brain image data to be identified, and the influence of the brain disease image data features on the spatiotemporal heterogeneity of the brain region is analyzed, so as to complete the neural heterogeneity guided dynamic brain network analysis.
2. The dynamic brain network analysis method of claim 1, wherein, The preprocessing of step S1 includes SPM8 realized by DPARSF toolbox in MATLAB for preprocessing all functional magnetic resonance imaging data. After obtaining the original DICOM data, the data not meeting the requirements of noise is rejected, then the Realign process is applied for head motion correction, and the sample data with relatively poor head motion is rejected, and finally the T1 weighted structure image is registered with the functional image.
3. The dynamic brain network analysis method of claim 2, wherein, The registration of the T1 weighted structure image and the functional image is realized by using the segmentation tool in DPARSF and the new segment+DARTEL tool, and the noise signal is removed by using the covariate regression, wherein the noise signal includes the noise of white matter and cerebrospinal fluid in the functional magnetic resonance imaging data.
4. The dynamic brain network analysis method of claim 1, wherein, Step S2 divides the pre-processed sample data into T subsequences using T overlapping windows of length S wherein V denotes the number of brain regions; For each subsequence , the correlation coefficient between each pair of brain region subsequence was calculated using Pearson correlation coefficient, and the dynamic functional brain network was obtained α To simulate the sparsity of brain network, the elements with connection strength lower than α were set to 0.
5. The dynamic brain network analysis method of claim 1, wherein, Step S2 extracts, for each brain network , a topological consistency network and a temporal trend network using two independent graph convolutional networks: ， ， wherein denotes a graph convolutional network, , I is an identity matrix, denotes a degree matrix after adding a self-loop, denotes a sub-sequence under the t-th window, , , and each denote a learnable parameter matrix, is an activation function; To enhance the discriminativeness between topologically consistent networks and temporally trending networks, similarity constraints between the two The construction is as follows: ， wherein denotes the number of brain regions, denotes the ith brain region under the tth window, • denotes the dot product operation, denotes norm; To ensure the complementarity of the topological consistency network and the temporal tendency network, the decoupled features are added to obtain the reconstructed representation, and the mean square error is introduced as the reconstruction loss between the reconstructed representation and the brain network , which is expressed as follows: ， wherein represents element-wise addition; Topological consistency refers to high stability of certain network structures in different windows, further imposing similarity constraints on topologically consistent networks of adjacent windows and encouraging such similarity The similarity increases as training progresses The expression is as follows: ， The above operation ensures that the topological consistency network and the time trend network are unique and complementary in the same window, and also promotes the similarity of the consistency network between different windows, so as to better coordinate with the inherent spatiotemporal consistency of the brain.
6. The dynamic brain network analysis method of claim 1, wherein, Step S3 includes calculating the cross-window similarity of the topological consistency network and the time trend network to measure the connection density and the time variability of the brain respectively, so as to explore the spatiotemporal heterogeneity of the brain; For spatial heterogeneity SH The average correlation between topologically consistent networks in all pairs of windows is computed: ， wherein denotes the cosine similarity, i and j denote the ith and jth window, respectively, and T denotes the number of sliding windows, denotes the topological consistency network under the ith and jth window, respectively; The lower cross-window similarity in the temporal trend network indicates more apparent dynamic evolution, which is calculated using the following formula TH : wherein, respectively represent the temporal trend network under the i-th and j-th window; Through the above process, the spatiotemporal heterogeneity brain region driving the brain network reorganization is obtained.
7. The dynamic brain network analysis method of claim 1, wherein, Step S4, after obtaining the spatiotemporal heterogeneity weight, weighting the spatiotemporal heterogeneity weight on the topological consistency network and the temporal trend network respectively to obtain a heterogeneity dynamic functional brain network; wherein the heterogeneity dynamic functional brain network comprises a heterogeneity topological consistency network and a heterogeneity temporal trend network .
8. The dynamic brain network analysis method of claim 1, wherein, The time propagation graph convolution of step S5 includes two graph convolution networks and one two-dimensional convolution network, and the features are extracted from the heterogeneous topological consistency network and the heterogeneous time trend network respectively through the two parameter-shared spatiotemporal convolution blocks; Heterogeneous topological consistency features and heterogeneous temporal trend features Learned in the following way: ， ； Adding the two types of features under the first t window to obtain a fused spatio-temporal representation After performing the same operation on the two networks for each window, concatenating the features of all windows to obtain a global spatio-temporal representation E : .
9. The dynamic brain network analysis method of claim 1, wherein, Step S6 comprises using a cross-entropy loss to supervise the update of the model parameters, the overall training objective is formulated as: ， wherein and are hyperparameters that control the relative importance of different loss terms, and denote the similarity loss and reconstruction loss, respectively, that the spatio-temporal pattern decoupling module decouples over all windows.
10. A neural heterogeneity guided dynamic brain network analysis system, comprising: The system is used for executing the method as claimed in any one of claims 1-9, and the system comprises: A spatiotemporal pattern decoupling module is used for simulating the spatiotemporal coordination of the brain, and decoupling the dynamic functional brain network into a topological consistency network and a time trend network consistent with brain activity; The spatio-temporal heterogeneity mining module captures spatial and temporal heterogeneity weights of different brain regions in the brain based on a topological consistency network and a time trend network respectively, so as to identify key nodes driving brain network reorganization; The spatio-temporal heterogeneity weighting module weights the topological consistency network and the time trend network respectively to obtain a heterogeneity dynamic functional brain network; The time propagation graph convolution module is used to simulate the propagation of neural information in the time dimension. The classification module uses a multilayer perceptron to map the heterogeneity spatio-temporal features of the brain image data to be identified to labels, and uses a cross-entropy loss function to train the classifier, thereby narrowing the distance between the predicted labels and the real labels to improve the identification performance of the brain image data to be identified.