Neurodegenerative disease diagnosis method for constructing dynamic brain network based on adaptive graph structure learning

Abstract

The invention discloses a neurodegenerative disease diagnosis method for constructing a dynamic brain network based on adaptive graph structure learning, and belongs to the technical field of artificial intelligence. The invention provides a self-adaptive graph structure learning multi-level nerve disease diagnosis framework, which can automatically learn and discover complex and deep connection modes among a plurality of brain areas so as to construct a dynamic brain network. According to the method, a global structure of a brain network is embedded into brain network construction based on a time window through a window mapping module, so that richness and accuracy of brain network topology feature representation are improved; in the fusion process of window features, a learnable time graph convolution module is provided to automatically capture time connectivity across time windows and effectively integrate high-order dynamic topological features extracted from different time windows. And finally, disease diagnosis and classification are realized by using a multi-layer perceptron. According to the method, the identification and classification accuracy of the neurodegenerative diseases under the fMRI data can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of dynamic brain network construction and multi-scale brain network analysis, and in particular to a neurodegenerative disease diagnosis method based on global-window dynamic graph convolution and cross-time window graph convolution based on adaptive graph structure learning. Background Art

[0002] Early diagnosis and accurate classification of neurodegenerative diseases (such as Alzheimer's disease and Parkinson's disease) are of great significance in clinical medicine, directly affecting the effectiveness of therapeutic interventions and the quality of life of patients. Brain network analysis technology based on functional magnetic resonance imaging (fMRI) can reveal the functional connectivity patterns between brain regions in a non-invasive way, providing a key basis for the discovery of disease markers. In the field of neuroscience, by tracking abnormal changes in dynamic brain networks, fMRI technology can assist in identifying key nodes in disease progression and become an important tool for monitoring neurodegenerative lesions. However, due to the high spatiotemporal dynamics and inter-individual heterogeneity of brain functional networks, the development of a diagnostic method that can accurately capture disease-specific connectivity patterns still faces severe challenges.

[0003] In general, existing methods for constructing brain networks for feature extraction can be divided into two categories. The first constructs a static functional brain network (SFBN) by directly calculating correlations from raw signal data and then transforming the raw signal data into feature vectors to represent brain information. The focus of SFBN analysis is to examine the functional connectivity of the brain in a resting state. In SFBN analysis, the brain is viewed as a whole network, however, increasing physiological evidence shows that brain functional networks change over time and contain rich temporal topological information, which the SFBN method cannot fully capture.

[0004] The second type studied dynamic functional brain networks (DFBNs), where the analysis focused on the temporal changes in brain functional connectivity, revealing the dynamic changes in network structure over time periods. These methods capture more detailed features while retaining the topological information of brain networks. However, commonly used correlation-based methods rely on linear relationships, while an increasing number of studies have shown that brain activity is complex and nonlinear. Many important neural activity patterns involve nonlinear interactions, suggesting that linear assumptions may not apply to all interactions between brain neurons in complex brain networks. In addition, correlation coefficient-based methods construct networks directly from signal correlations, ignoring multi-level structural signal correlations (e.g., different types of functional or structural connection patterns), which may exist in brain networks.

[0005] The diagnosis of traditional neurodegenerative diseases mainly relies on static fMRI analysis, which characterizes network characteristics by calculating the average functional connection strength between brain regions. Although such methods have been widely used, their neglect of dynamic interaction patterns may lead to the loss of key pathological information. For example, neurodegenerative diseases are often accompanied by dynamic attenuation or abnormal enhancement of functional connectivity in specific brain regions (such as the default mode network and hippocampus), and static network models are difficult to characterize such time-varying characteristics. In addition, single-time-scale analysis cannot distinguish the dynamic evolution of different stages of the disease, limiting the sensitivity of early diagnosis. Both types of methods are affected by the low expressiveness of predefined related computational problems and lack flexibility. Discovering more complex patterns of brain connectivity and constructing brain networks that can effectively learn and capture representative features remains a challenging task.

[0006] In summary, studying how to break through the technical barriers of dynamic brain network modeling and achieve robust capture of deep multi-brain region connection patterns is a core challenge in this field. The interweaving of spatiotemporal heterogeneity of dynamic networks and disease-specific variation makes it much more difficult than static network analysis. Therefore, developing an fMRI diagnostic method that integrates multi-scale dynamic features and is resistant to noise interference has become an urgent need to promote precision medicine for neurodegenerative diseases. Summary of the invention

[0007] Purpose of the invention: In view of the shortcomings of the neurodegenerative disease diagnosis method for building dynamic brain networks, the first purpose of the present invention is to provide a neurodegenerative disease diagnosis method based on adaptive graph structure learning and cross-time window graph convolution, which can learn to discover complex and deep connection patterns between multiple brain regions, optimize the fusion of window features through cross-time window graph convolution, and improve the richness and accuracy of feature representation. Based on the application of this method, the present invention provides a flexible and effective framework for building dynamic brain networks, which can well capture the complex spatiotemporal topological features in dynamic brain networks.

[0008] Technical solution: A neurodegenerative disease diagnosis method based on adaptive graph structure learning to build a dynamic brain network is provided. The method is characterized in that the method captures the deep and complex brain network connection mode of fMRI data through adaptive graph structure learning, models the dynamic brain network, and learns the spatiotemporal dependency across time windows, and enhances the feature fusion method between windows to improve the richness and accuracy of the brain network topological feature representation, including the following steps:

[0009] S1, acquiring fMRI neurological disease sequence data, preprocessing the data, and obtaining preprocessed data;

[0010] S2, using non-overlapping sliding windows to divide the input fMRI time series data into time windows to facilitate the subsequent construction of dynamic brain networks;

[0011] Take the preprocessed neurodegenerative disease (fMRI) data as input and define the fMRI data as X fMRI ∈R N ×M , where N is the number of brain regions, that is, the number of channels of fMRI signals, and M is the length of the fMRI time series. The window data after division is X i ∈R B×N×L , where B is the number of samples in a training batch and L is the length of each time window;

[0012] S3, based on adaptive graph structure learning to capture the complex and deep connectivity patterns of the brain network, build a learnable global brain network structure, and use Xavier normal distribution to initialize the matrix weights, initialized to A global =Xavier(N×N), where N is the number of brain regions, capturing the connectivity across brain regions;

[0013] Adjacency Matrix A Global It is a learnable parameter that is dynamically optimized through the end-to-end training process. In each operation, the matrix weights are updated based on the gradient of the model's loss function:

[0014]

[0015] In the formula, A n represents the adjacency matrix of the nth iteration, η represents the learning rate, L Graph represents the graph structure learning loss function used to constrain the brain network to construct the adjacency matrix;

[0016] S4. Through the window mapper constructed by the multi-layer perceptron, the adaptively learned overall brain network structure is mapped to the window brain network with local attributes, so as to embed the global structure of the brain network into the construction of the brain network based on the time window, thereby enhancing the richness and accuracy of the representation of the topological features of the brain network:

[0017]

[0018] In the formula, vec(·) represents flattening to a vector, f θ (·) represents the mapping function, ReLU(·) represents the activation function, is the brain network adjacency matrix of the ith window;

[0019] S5. Both the global-scale brain network and the window-scale brain network are passed through a graph convolution module with shared parameters to obtain global-scale and window-scale features, which can more effectively capture brain connection patterns at different levels and provide more accurate and detailed representations.

[0020] The final feature matrix of the global scale and window scale can be obtained:

[0021]

[0022] In the formula, f global ∈R N×T and f window ∈R N×T The window fMRI time series data are respectively passed through the global adjacency matrix and the window adjacency matrix to build brain networks of different scales. The features learned after the shared graph convolution layer, X represents the window fMRI time series data, represents the adjacency matrix with added self-connectivity, is the degree matrix, Θ is the graph convolution layer parameter, is the feature of the i-th window;

[0023] S6, built a plug-and-play window feature fusion module. Specifically, the dependencies between time windows are automatically captured through adaptive graph structure learning, a time graph is constructed from multiple window features, and these window features are effectively fused through graph convolution. Similar to S3 initializing the global brain network structure, the matrix weights are initialized using Xavier normal distribution, initialized to A Time =Xavier(T×T), where T represents the number of time windows;

[0024] Use graph convolutional networks to obtain fused window features:

[0025]

[0026] In the formula, the topological structure of the time graph comes from the adaptive graph structure learning A Time , the node feature of each node comes from the window feature obtained by S5 represents the adjacency matrix with added self-connectivity, is the degree matrix, Θ is the graph convolution layer parameter, ReLU is the activation function, and Falten means flattening the feature matrix;

[0027] S7. The obtained features are input into a classifier built using a fully connected layer to achieve diagnosis and classification of neurodegenerative diseases.

[0028] Furthermore, the specific process of step S1 includes:

[0029] S1.1. All rs-fMRI data were preprocessed using SPM 9 in the DPARSF toolbox in MATLAB.

[0030] S1.2, correcting and reconstructing the initial image by dividing the serially stored data into different parts and adjusting them according to the EPI template;

[0031] S1.3, de-skew processing to reduce the effects of head motion and cerebrospinal fluid and white matter interference;

[0032] S1.4. Use AAL mapping to divide the rs-fMRI dataset into an appropriate number of ROIs.

[0033] In the method, the specific process of learning the global brain network structure described in step S3 includes:

[0034] In order to further constrain the structure of the global brain network adjacency matrix, the present invention designs a loss function for global brain network structure learning:

[0035] L Graph =λ 1 ×tr(L)+λ 2 ×||AA T || F +λ 3 ×||A|| L1

[0036] Where L is the Laplacian matrix, tr(L) represents the trace of the Laplacian matrix, and the feature smoothing loss is calculated to encourage similar nodes to be closer in the feature space. ||AA T || F The Frobenius norm is used to measure the difference between the adjacency matrix and its transpose, ensuring that the adjacency matrix remains symmetric. ||A|| L1 Use the L1 norm to reduce unnecessary connections and maintain the sparsity of the graph structure; λ 1 ,λ 2 ,λ 3 The parameters of the smoothness loss, symmetry loss, and sparsity loss constrain the importance of the three losses.

[0037] In the method, the specific process of time convolution described in step S6 includes:

[0038] Similar to the S3 learnable global brain network graph structure, the temporal convolution graph structure updates the matrix weights in each iteration based on the gradient of the model's loss function. The present invention designs a temporal graph structure loss function for temporal graph convolution:

[0039] L Time =L Decay +L smooth +γ×L fro

[0040] Where, L DecayA decay mechanism based on the distance between nodes is designed, which aims to make the connection strength of adjacent nodes (in the adjacency matrix) more similar, L Decay The specific calculation is:

[0041]

[0042] In the formula, the attenuation matrix M decay,ij =exp(-αD i,j ), α is the attenuation coefficient, which controls the attenuation speed, D i,j represents the distance measure between node i and node j;

[0043] L smooth The smoothness of the adjacency matrix is ​​realized, which is a measure of the difference in connection strength between adjacent nodes in the adjacency matrix. The specific calculation is:

[0044]

[0045] L fro The sparsity of the matrix is ​​realized, γ is a hyperparameter that controls the sparsity, and L fro The specific calculation is:

[0046]

[0047] In the method, the specific process of the classifier described in step S7 includes:

[0048] Through adaptive graph structure learning, the present invention further captures cross-window dependencies and fuses window features. After extracting high-dimensional features, the loss is obtained through the classification discriminator. CE The classification module is implemented by minimizing the following loss function:

[0049] loss CE (y,p)=-y logp-(1-y)log(1-p)

[0050] Where y is the true value label and p is the predicted value. Then, according to the global brain network structure learning loss function of claim 3 and the temporal graph structure loss function of claim 4, the overall model is optimized, and the objective function is as follows:

[0051] loss=loss CE +αL Graph +βL Time

[0052] Where α and β are the parameters for adjusting L Graph and L Time Hyperparameters of weights.

[0053] Beneficial effects: Compared with the prior art, the significant effects and substantial features of the present invention are mainly:

[0054] (1) A learnable hierarchical network structure is proposed, which can learn and discover complex and deep connection patterns between multiple brain regions through adaptive graph structure. This method provides a flexible and effective framework for building brain networks, which can well capture the complex spatiotemporal topological characteristics of dynamic brain networks.

[0055] (2) In order to maintain the learned macroscopic connectivity of the brain network and its dynamic evolution over time, a window mapping module is designed to embed the global structure of the brain network into the brain network construction based on the time window, thereby improving the richness and accuracy of the representation of brain network topological features.

[0056] (3) We propose a learnable temporal graph convolution module, which aims to automatically capture the temporal connectivity across time windows and effectively integrate high-order dynamic topological features extracted from different time windows.

[0057] (4) Comprehensive experiments were conducted on Alzheimer's disease and Parkinson's disease datasets. The results showed that this method is superior to existing brain network analysis methods, providing an effective method for brain network analysis and having significant advantages in the diagnosis of brain diseases. The effectiveness of this method was verified by ablation studies. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is a work flow chart of the method of the present invention;

[0059] Figure 2 It is a data preprocessing flow chart described in the present invention;

[0060] Figure 3 It is an overall framework diagram of the system to which the method described in the present invention is applied. DETAILED DESCRIPTION

[0061] In order to make the purpose, advantages and technical solutions of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be fully and clearly described below in conjunction with the accompanying drawings;

[0062] Early diagnosis and accurate classification of neurodegenerative diseases directly affect the therapeutic intervention effect and quality of life of patients, and are of great significance. The diagnosis and classification technology of neurodegenerative diseases is based on functional magnetic resonance imaging (fMRI), which can reveal the functional connection pattern between brain regions in a non-invasive way, providing a key basis for the discovery of disease markers. However, the diagnosis of traditional neurodegenerative diseases mainly relies on static fMRI analysis, which characterizes network characteristics by calculating the average functional connection strength between brain regions. Although such methods have been widely used, their neglect of dynamic interaction patterns may lead to the loss of key pathological information. In addition, the analysis of a single time scale cannot distinguish the dynamic evolution of different stages of the disease, limiting the sensitivity of early diagnosis. Both types of methods are affected by the low expressiveness of predefined related computational problems and lack flexibility. Finally, discovering more complex patterns of brain connections and constructing brain networks that can effectively learn and capture representative features are the focus of this study. Discovering complex and deep brain network connection patterns and constructing dynamic brain networks to extract spatiotemporal topological features with excellent representation capabilities need to be solved simultaneously, which is more difficult than the sum of the two.

[0063] The present invention provides a method for diagnosing neurodegenerative diseases based on adaptive graph structure learning to construct a dynamic brain network. Figure 1 The process shown and Figure 3 As shown in the method framework diagram, the embodiment adopts deep neural network technologies such as adaptive graph structure learning, graph convolutional networks, and spatiotemporal graph convolution to realize the diagnosis and classification of neurodegenerative diseases.

[0064] S1. Acquire fMRI neurological disease sequence data, preprocess the data, and obtain a preprocessed sequence matrix.

[0065] Combination Figure 2 The data preprocessing flowchart shown in the figure uses MATLAB and SPM 9 in the DPARSF toolbox to preprocess all rs-fMRI data. The initial images were corrected and reorganized by dividing the serially stored data into different parts and adjusting them according to the EPI template. The effects of head motion and interference from cerebrospinal fluid and autochromatic substances were reduced by de-skewing, and the rs-fMRI of the dataset was divided into the appropriate number of ROIs using AAL mapping.

[0066] S2, using non-overlapping sliding windows to divide the input fMRI time series data into time windows to facilitate the subsequent construction of dynamic brain networks;

[0067] Take the preprocessed neurodegenerative disease (fMRI) data as input and define the fMRI data as X fMRI ∈R N ×M, where N is the number of brain regions, that is, the number of channels of fMRI signals, and M is the length of the fMRI time series. The window data after division is X i ∈R B×N×L , where B is the number of samples in a training batch and L is the length of each time window;

[0068] S3, based on adaptive graph structure learning to capture the complex and deep connectivity patterns of brain networks and build a learnable global brain network structure;

[0069] Use Xavier normal distribution to initialize the matrix weights, initialized to A global =Xavier(N×N), where N is the number of brain regions, capturing the connectivity across brain regions;

[0070] Adjacency Matrix A Global It is a learnable parameter that is dynamically optimized through the end-to-end training process. In each operation, the matrix weights are updated based on the gradient of the model's loss function:

[0071]

[0072] In the formula, A n represents the adjacency matrix of the nth iteration, η represents the learning rate, L Graph Represents the graph structure learning loss function used to constrain the brain network to construct the adjacency matrix:

[0073] L Graph =λ 1 ×tr(L)+λ 2 ×||AA T || F +λ 3 ×||A|| L1

[0074] Where L is the Laplacian matrix, tr(L) represents the trace of the Laplacian matrix, and the feature smoothing loss is calculated to encourage similar nodes to be closer in the feature space. ||AA T || F The Frobenius norm is used to measure the difference between the adjacency matrix and its transpose, ensuring that the adjacency matrix remains symmetric. ||A|| L1 Use the L1 norm to reduce unnecessary connections and maintain the sparsity of the graph structure; λ 1 ,λ 2 ,λ 3 The parameters of the smoothness loss, symmetry loss, and sparsity loss constrain the importance of the three losses.

[0075] S4. Through the window mapper constructed by the multi-layer perceptron, the adaptively learned overall brain network structure is mapped to the window brain network with local attributes, so as to embed the global structure of the brain network into the construction of the brain network based on the time window, thereby enhancing the richness and accuracy of the representation of the topological features of the brain network:

[0076]

[0077] In the formula, vec(·) represents flattening to a vector, f θ (·) represents the mapping function, ReLU(·) represents the activation function, is the brain network adjacency matrix of the ith window;

[0078] S5. Both the global-scale brain network and the window-scale brain network are passed through a graph convolution module with shared parameters to obtain global-scale and window-scale features, which can more effectively capture brain connection patterns at different levels and provide more accurate and detailed representations.

[0079] The graph convolutional network with shared parameters can obtain the final feature matrix of the global scale and window scale:

[0080]

[0081] In the formula, f global ∈R N×T and f window ∈R N×T The window fMRI time series data are respectively passed through the global adjacency matrix and the window adjacency matrix to build brain networks of different scales. The features learned after the shared graph convolution layer, X represents the window fMRI time series data, represents the adjacency matrix with added self-connectivity, is the degree matrix, Θ is the graph convolution layer parameter, is the feature of the i-th window;

[0082] S6, built a plug-and-play temporal graph convolution feature fusion module. Specifically, the dependencies between time windows are automatically captured through adaptive graph structure learning, a temporal graph is constructed from multiple window features, and these window features are effectively fused through graph convolution. Similar to S3 initializing the global brain network structure, the matrix weights are initialized using Xavier normal distribution, initialized to A Timne =Xavier(T×T), where T represents the number of time windows;

[0083] Similar to the S3 learnable global brain network graph structure, the temporal convolution graph structure updates the matrix weights in each iteration based on the gradient of the model's loss function. The present invention designs a temporal graph structure loss function for temporal graph convolution:

[0084] L Time =L Decay +L smooth +γ×L fro

[0085] Where, L Decay A decay mechanism based on the distance between nodes is designed, which aims to make the connection strength of adjacent nodes (in the adjacency matrix) more similar, L Decay The specific calculation is:

[0086]

[0087] In the formula, the attenuation matrix M decay,ij =exp(-αD i，j ), α is the attenuation coefficient, which controls the attenuation speed, D i，j represents the distance measure between node i and node j;

[0088] L smooth The smoothness of the adjacency matrix is ​​realized, which is a measure of the difference in connection strength between adjacent nodes in the adjacency matrix. The specific calculation is:

[0089]

[0090] L fro The sparsity of the matrix is ​​realized, γ is a hyperparameter that controls the sparsity, and L fro The specific calculation is:

[0091]

[0092] Use a temporal graph convolutional network to obtain fused window features:

[0093]

[0094] In the formula, the topological structure of the time graph comes from the adaptive graph structure learning A Time , the node feature of each node comes from the window feature obtained by S5 represents the adjacency matrix with added self-connectivity, is the degree matrix, Θ is the graph convolution layer parameter, ReLU is the activation function, and Falten means flattening the feature matrix;

[0095] S7. Input the obtained features into a classifier constructed using a fully connected layer to achieve diagnosis and classification of neurodegenerative diseases;

[0096] Through adaptive graph structure learning, the present invention further captures cross-window dependencies and fuses window features. After extracting high-dimensional features, the loss is obtained through the classification discriminator.CE The classification module is implemented by minimizing the following loss function:

[0097] loss CE (y,p)=-y log p-(1-y)log(1-p)

[0098] Where y is the true value label and p is the predicted value. Then, according to the global brain network structure learning loss function of claim 3 and the temporal graph structure loss function of claim 4, the overall model is optimized, and the objective function is as follows:

[0099] loss=loss CE +αL Graph +βL Time

[0100] Where α and β are the parameters for adjusting L Graph and L Time Hyperparameters of weights.

[0101] Through adaptive graph structure learning, the global-window brain network construction module has learned deep, complex and nonlinear brain area connection patterns. In addition, in the window feature fusion module, the model can learn cross-window dependencies and better fuse window features. The model can learn holistic brain networks with long-term characteristics and window-scale brain networks with short-term characteristics, embedding the global structure of the brain network into the construction of the brain network based on the time window, thereby enhancing the richness and accuracy of the topological feature representation of the brain network.

Claims

1. A neurodegenerative disease diagnosis method based on adaptive graph structure learning to construct a dynamic brain network, characterized in that: This method captures the deep and complex brain network connection patterns of fMRI data through adaptive graph structure learning, models dynamic brain networks, learns the spatiotemporal dependencies across time windows, and enhances the feature fusion method between windows to improve the richness and accuracy of brain network topological feature representation. It includes the following steps: S1, acquiring fMRI neurological disease sequence data, preprocessing the data, and obtaining preprocessed data; S2, using non-overlapping sliding windows to divide the input fMRI time series data into time windows to facilitate the subsequent construction of dynamic brain networks; Take the preprocessed neurodegenerative disease (fMRI) data as input and define the fMRI data as X fMRI ∈R N×M , where N is the number of brain regions, that is, the number of channels of fMRI signals, and M is the length of the fMRI time series. The window data after division is X i ∈R B×N×L , where B is the number of samples in a training batch and L is the length of each time window; S3, based on adaptive graph structure learning to capture the complex and deep connectivity patterns of the brain network, build a learnable global brain network structure, and use Xavier normal distribution to initialize the matrix weights, initialized to A global =Xavier(N×N), where N is the number of brain regions, capturing the connectivity across brain regions; Adjacency Matrix A Global It is a learnable parameter that is dynamically optimized through the end-to-end training process. In each operation, the matrix weights are updated based on the gradient of the model's loss function: In the formula, A n represents the adjacency matrix of the nth iteration, η represents the learning rate, L Graph represents the graph structure learning loss function used to constrain the brain network to construct the adjacency matrix; S4. Through the window mapper constructed by the multi-layer perceptron, the adaptively learned overall brain network structure is mapped to the window brain network with local attributes, so as to embed the global structure of the brain network into the construction of the brain network based on the time window, thereby enhancing the richness and accuracy of the representation of the topological features of the brain network: In the formula, vec(·) represents flattening to a vector, f θ (·) represents the mapping function, ReLU(·) represents the activation function, is the brain network adjacency matrix of the ith window; S5. Both the global-scale brain network and the window-scale brain network are passed through a graph convolution module with shared parameters to obtain global-scale and window-scale features, which can more effectively capture brain connection patterns at different levels and provide more accurate and detailed representations. The final feature matrix of the global scale and window scale can be obtained: In the formula, f global ∈R N×T and f window ∈R N×T The window fMRI time series data are respectively passed through the global adjacency matrix and the window adjacency matrix to build brain networks of different scales. The features learned after the shared graph convolution layer, X represents the window fMRI time series data, represents the adjacency matrix with added self-connectivity, is the degree matrix, Θ is the graph convolution layer parameter, is the feature of the i-th window; S6, built a plug-and-play window feature fusion module. Specifically, the dependencies between time windows are automatically captured through adaptive graph structure learning, a time graph is constructed from multiple window features, and these window features are effectively fused through graph convolution. Similar to S3 initializing the global brain network structure, the matrix weights are initialized using Xavier normal distribution, initialized to A Time =Xavier(T×T), where T represents the number of time windows; Use graph convolutional networks to obtain fused window features: In the formula, the topological structure of the time graph comes from the adaptive graph structure learning A Time , the node feature of each node comes from the window feature obtained by S5 represents the adjacency matrix with added self-connectivity, is the degree matrix, Θ is the graph convolution layer parameter, ReLU is the activation function, and Falten means flattening the feature matrix; S7. The obtained features are input into a classifier built using a fully connected layer to achieve diagnosis and classification of neurodegenerative diseases.

2. The neurodegenerative disease diagnosis method based on adaptive graph structure learning to construct a dynamic brain network according to claim 1, characterized in that: The specific process of step S1 includes: S1.

1. All rs-fMRI data were preprocessed using SPM 9 in the DPARSF toolbox in MATLAB. S1.2, correcting and reconstructing the initial image by dividing the serially stored data into different parts and adjusting them according to the EPI template; S1.3, de-skew processing to reduce the effects of head motion and cerebrospinal fluid and white matter interference; S1.

4. Use AAL mapping to divide the rs-fMRI dataset into an appropriate number of ROIs.

3. The neurodegenerative disease diagnosis method based on adaptive graph structure learning to construct dynamic brain networks according to claim 1, characterized in that: The specific process of step S3 includes: In order to further constrain the structure of the global brain network adjacency matrix, the present invention designs a loss function for global brain network structure learning: L Graph =λ1×tr(L)+λ2×||A-A T || F +λ3×||A|| L1 Where L is the Laplacian matrix, tr(L) represents the trace of the Laplacian matrix, and the feature smoothing loss is calculated to encourage similar nodes to be closer in the feature space. ||AA T || F The Frobenius norm is used to measure the difference between the adjacency matrix and its transpose, ensuring that the adjacency matrix remains symmetric. ||A|| L1 The L1 norm is used to reduce unnecessary connections and maintain the sparsity of the graph structure; λ1, λ2, λ3 are parameters of smoothness loss, symmetry loss, and sparsity loss, which constrain the importance of the three losses.

4. The neurodegenerative disease diagnosis method based on adaptive graph structure learning to construct dynamic brain networks according to claim 1, characterized in that: The specific process of step S6 includes: Similar to the S3 learnable global brain network graph structure, the temporal convolution graph structure updates the matrix weights in each iteration based on the gradient of the model's loss function. The present invention designs a temporal graph structure loss function for temporal graph convolution: L Time =L Decay +L smooth +γ×L fro Where, L Decay A decay mechanism based on the distance between nodes is designed, which aims to make the connection strength of adjacent nodes (in the adjacency matrix) more similar, L Decay The specific calculation is: In the formula, the attenuation matrix M decay,ij =exp(-αD i，j ), α is the attenuation coefficient, which controls the attenuation speed, D i，j represents the distance measure between node i and node j; L smooth The smoothness of the adjacency matrix is ​​realized, which is a measure of the difference in connection strength between adjacent nodes in the adjacency matrix. The specific calculation is: L fro The sparsity of the matrix is ​​realized, γ is a hyperparameter that controls the sparsity, and L fro The specific calculation is:

5. The neurodegenerative disease diagnosis method based on adaptive graph structure learning to construct dynamic brain networks according to claim 1, characterized in that: The specific process of step S7 includes: Through adaptive graph structure learning, the present invention further captures cross-window dependencies and fuses window features. After extracting high-dimensional features, the loss is obtained through the classification discriminator. CE The classification module is implemented by minimizing the following loss function: loss CE (y,p)=-y log p-(1-y)log(1-p) Where y is the true value label and p is the predicted value. Then, according to the global brain network structure learning loss function of claim 3 and the temporal graph structure loss function of claim 4, the overall model is optimized, and the objective function is as follows: loss=loss CE +αL Graph +βL Time Where α and β are the parameters for adjusting L Graph and L Time Hyperparameters of weights.

6. The neurodegenerative disease diagnosis method based on adaptive graph structure learning to construct dynamic brain networks according to claim 1, characterized in that: Through adaptive graph structure learning, the global-window brain network construction module has learned deep, complex and nonlinear brain area connection patterns. In addition, in the window feature fusion module, the model can learn cross-window dependencies and better fuse window features. The model can learn holistic brain networks with long-term characteristics and window-scale brain networks with short-term characteristics, embedding the global structure of the brain network into the construction of the brain network based on the time window, thereby enhancing the richness and accuracy of the topological feature representation of the brain network.

Images