A method and system for brain disease identification based on spatiotemporal multi-view functional brain networks

Abstract

This invention belongs to the field of medical image analysis technology, specifically relating to a brain disease identification method and system based on a spatiotemporal multi-view functional brain network. The method includes dividing a preprocessed brain functional image into several brain regions based on a predefined brain region template, extracting a time series matrix, and dividing the time series matrix into several time windows. Within each time window, a correlation matrix is ​​calculated using several functional connectivity metrics to construct a spatiotemporal multi-view functional brain network. This invention uses a sliding window technique to divide the brain region time series into multiple time windows, capturing the dynamic time-varying characteristics of brain functional connectivity and avoiding static analysis from obscuring key temporal information. Simultaneously, it integrates four functional connectivity metrics—Pearson correlation, higher-order functional connectivity, mutual information, and sparse representation—to characterize linear relationships, higher-order interactions, nonlinear dependencies, and sparse dependencies between brain regions, respectively, fully utilizing complementary information from multiple views to significantly improve the comprehensiveness and richness of the functional brain network representation.

Description

Technical Field

[0001] This invention belongs to the field of medical image analysis technology, specifically relating to a brain disease identification method and system based on spatiotemporal multi-view functional brain networks. Background Technology

[0002] Resting-state functional magnetic resonance imaging (rs-fMRI), with its non-invasive nature and high spatial resolution, has become a core technology in neuroscience and brain disease research. It indirectly reflects brain neural activity by capturing blood-oxygen-level-dependent (BOLD) fluctuations. Functional brain networks (FBNs) constructed based on rs-fMRI data can characterize functional connectivity patterns between brain regions and are key biomarkers for identifying neurological and psychiatric disorders.

[0003] Among existing methods for constructing brain-free neural networks (FBNs), Pearson's correlation (PC) is widely used due to its computational simplicity, but it can only capture linear relationships and is difficult to reveal nonlinear dependencies between brain regions, and it is prone to generating spurious connections. Although there are improved methods such as mutual information (MI), high-order functional connectivity (HOFC), and sparse representation (SR) that can capture nonlinear relationships, high-order interactions, and sparse dependencies respectively, the representational ability of a single functional connectivity metric is limited. At the same time, there are significant individual differences among patients with brain diseases, resulting in insufficient generalization ability of single-view FBNs.

[0004] Furthermore, traditional methods are mostly based on static functional connectivity analysis, neglecting the dynamic and time-varying characteristics of brain functional connectivity. Although dynamic functional connectivity analysis techniques have partially addressed this issue through methods such as sliding windows, effectively integrating brain network information from multiple time windows and multiple views remains a challenge. Existing multi-view fusion methods struggle to adapt to the varying reliability of different views across different subjects and time windows, limiting the diagnostic accuracy of the model in heterogeneous populations. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, this invention provides a method, system, terminal, and storage medium for brain disease identification based on spatiotemporal multi-view functional brain networks.

[0006] In a first aspect, the present invention provides a method for identifying brain diseases based on spatiotemporal multi-view functional brain networks, comprising: S1. The acquired brain functional images are preprocessed, and the brain functional images are divided into several brain regions based on predefined brain region templates. The time series of each brain region is extracted to obtain a time series matrix; where the brain functional images correspond to brain disease category labels. S2. The time series matrix is ​​divided into several time windows using a sliding window analysis strategy to obtain a set of time windows; S3. Within each time window, the correlation matrix is ​​calculated using several functional connectivity metrics to construct a spatiotemporal multi-view functional brain network. The functional connectivity metrics include Pearson correlation, higher-order functional connectivity, mutual information, and sparse representation. S4. For each spatiotemporal multi-view functional brain network, construct a graph convolutional encoder and a multilayer perceptron classification head, perform representation learning to obtain view-level output, obtain view evidence through a non-negative activation function, and calculate the Dirichlet distribution parameters. ,in Number of brain disease categories; S5. Aggregate spatial dimension information and perform the first round of evidence fusion on the multi-view Dirichlet distribution parameters within each time window t to obtain the time window-level Dirichlet distribution parameters. ; S6. Aggregate time-dimensional information and perform a second round of evidence fusion on the Dirichlet distribution parameters of each time window to obtain the final Dirichlet distribution parameters. This enables the classification and identification of brain diseases.

[0007] Further improvements to this technical solution include step S1, which includes: S11. Preprocess the acquired resting-state functional magnetic resonance images of the brain. The brain functional images are labeled with brain disease categories. Use the DPARSF toolbox of MATLAB to remove the first p time points of the image, and perform head motion artifacts and time layer correction. Register the corrected image to the standard space and apply a time bandpass filter of 0.01Hz to 0.1Hz. S12. Based on predefined brain region templates, the preprocessed resting-state functional magnetic resonance imaging (fMRI) images are divided into N brain regions. The brain region templates are selected from the AAL116 template or the CC200 template. The AAL116 template divides the resting-state fMRI images into 116 brain regions, and the CC200 template divides the resting-state fMRI images into 200 brain regions. S13. Extract the time series data of each brain region to form a time series matrix. ,in Let T0 represent the BOLD signal vectors of N brain regions at time t, where T0 is the total number of time points and N is the number of brain regions.

[0008] Further improvements to this technical solution include step S2, which includes: S21. Configure the parameters required for sliding window analysis, including window length W and sliding step size s; S22. Based on parameters W and s, the time series matrix X is divided into T consecutive time windows. Each time window X_t extracts brain region signals from W consecutive time points in the original sequence, and the overlapping or non-overlapping relationship between adjacent time windows is controlled by the step size s. S23. Integrate all time windows to form a time window set. ,in This represents the data matrix for the t-th time window, where T is the total number of time windows. T_0 represents the original time point.

[0009] Further improvements to this technical solution include step S3, which includes: S31. Within each time window X_t, select multiple functional connectivity metrics to construct a multi-view brain network. The functional connectivity metrics include Pearson correlation, higher-order functional connectivity, mutual information, and sparse representation. S32. For linear correlation measures, calculate the Pearson correlation coefficient matrix. : ; in, The Pearson correlation coefficient between the i-th and j-th brain regions; S33. Calculating higher-order functional connectivity matrices based on Pearson correlation matrices. : ; in, To standardize the Pearson correlation matrix, , and are the mean and standard deviation of the i-th row of the Pearson correlation matrix, respectively; Represents the L2 norm; S34. For nonlinear correlation metrics, calculate the mutual information matrix. : ; in, and These are the marginal probability distributions of the i-th and j-th brain regions, respectively; for and The joint probability distribution is calculated from the time series normalized to [0,1] using histogram estimation; simultaneously, the sparse representation matrix is ​​calculated. : ; in, This represents the standardized time series vector of the i-th brain region within the time window t; The coefficients representing the sparse representation of brain region j to brain region i are denoted by . For regularization parameters; S35. Integrate all functional connection matrices to construct a spatiotemporal multi-view functional brain network. ;in, Each view corresponds to a metric; The node feature matrix; This is the adjacency matrix for the corresponding metric.

[0010] Further improvements to this technical solution include step S4, which includes: S41. For each time window t and view m, construct a graph convolutional encoder. This encoder uses a multi-layer graph convolutional neural network, similar to a spatiotemporal multi-view functional brain network. Adjacency matrix in As the adjacency matrix of the graph, the node representation is updated through graph convolution operations, using a preset node feature matrix X_m(t)∈R^{N×F} as input. The formula is as follows: ; in, This represents the updated node embedding matrix; N is the number of brain regions; F is the input node feature dimension; G is the feature dimension of the graph convolution output; The weight matrix is ​​a learnable matrix; For bias terms; The non-linear activation function is used; the graph convolutional encoder is composed of multiple stacked graph convolutional layers, each of which includes adjacency matrix normalization, feature transformation and activation function operation, thereby extracting the spatial structural features of the brain network. S42, Output of the graph convolution encoder A multilayer perceptron classification head is constructed, and the average along the brain region dimension is used to obtain a time window-level graphical representation. Its formula is: ; S43. Represent the pooled graph. The input is a multilayer perceptron (MLP), which consists of multiple fully connected layers, normalization layers, and activation functions. The view-level output is obtained through forward propagation. Its formula is: ; Where C represents the number of brain disease categories; the specific structure of the MLP includes an input layer, a hidden layer, and an output layer, with the hidden layer using LayerNorm normalization and the GELU activation function; S44, View-level output Calculate view evidence using the non-negative activation function Softplus(·). Its formula is: ; The Softplus function ensures that the evidence is non-negative. S45. Calculate the Dirichlet distribution parameters based on the evidence deep learning framework. Its formula is: ; The view-level Dirichlet distribution parameters are obtained and used to model classification uncertainty, where, The kth component The strength of evidence corresponds to the k-th class, where C is the number of classes.

[0011] Further improvements to this technical solution include step S5, which includes: S51. For each time window t, the Dirichlet distribution parameters based on each view m. Calculate the total Dirichlet intensity for each view. ,evidence Classification beliefs and view uncertainty The specific formula is as follows: ; ; ; ; Where k represents the category index; S52. The first round of fusion is performed using an evidence fusion framework to fuse the Dirichlet distribution parameters of all views within time window t. This framework, based on Dempster-Shafer theory, sequentially fuses the parameters in the view set pairwise: for any two views... and Calculate its strength and Classification beliefs and Uncertainty and The conflict factor K is calculated to measure consistency between views, and the merged classification beliefs are updated. and uncertainty The specific formula is as follows: ; ; ; ; ; ; This fusion process is applied iteratively to all views until the parameters of the entire view set are integrated into a single representation; S53, Classification Beliefs Based on Fusion and uncertainty Derivation of Time Window-Level Dirichlet Distribution Parameters Calculate the total intensity after fusion, S_t=C / u, and calculate the evidence. Finally obtained Each component The first round of evidence fusion in the spatial dimension was completed, and time window-level parameters were obtained. .

[0012] Further improvements to this technical solution include step S6, which includes: S61. Integrate the set of time window-level Dirichlet distribution parameters for all time windows. This collection contains spatial fusion results from various local time periods in the time dimension; S62. Set of Dirichlet distribution parameters for each time window along the time dimension. The second round of evidence fusion is performed using the same evidence fusion framework as in step S5, fusing the parameters in the time window set pairwise: for any two time windows, the Dirichlet distribution parameters... and Calculate its strength and Classification beliefs and Uncertainty and And calculate the conflict factor K and update the fused classification beliefs. and uncertainty The specific formula is as follows: ; ; ; ; ; ; This fusion process is applied iteratively to all time window parameters until the parameters of the entire time window set are integrated into a single final Dirichlet distribution parameter. This completes the second round of evidence fusion in the time dimension; S63, Based on the final Dirichlet distribution parameters Achieving Classification and Recognition of Brain Diseases: Computing Final Evidence and total strength The classification result is determined by taking the expected value or the maximum probability, whereby the probability of each category is calculated. The category with the highest probability is selected as the final diagnosis result, using the following formula: ; in, To predict brain disease categories, we aim to achieve intelligent diagnosis of brain diseases based on two-layer evidence fusion.

[0013] Secondly, the present invention provides a brain disease identification system based on a spatiotemporal multi-view functional brain network, comprising: The preprocessing module is used to preprocess the acquired brain functional images, divide the brain functional images into several brain regions based on predefined brain region templates, extract the time series of each brain region, and obtain a time series matrix; among them, the brain functional images correspond to brain disease category labels; The sliding window analysis module is used to divide the time series matrix into several time windows using the sliding window analysis strategy, thereby obtaining a set of time windows. The multi-view brain network construction module is used to calculate the correlation matrix within each time window using several functional connectivity metrics to construct a spatiotemporal multi-view functional brain network; the functional connectivity metrics include Pearson correlation, higher-order functional connectivity, mutual information, and sparse representation. The graph convolution and classification module is used to construct a graph convolutional encoder and a multilayer perceptron classification head for each spatiotemporal multi-view functional brain network, perform representation learning and obtain view-level output, obtain view evidence through a non-negative activation function, and calculate the Dirichlet distribution parameters. ,in Number of brain disease categories; The spatial evidence fusion module is used to aggregate spatial dimension information. Within each time window t, it performs the first round of evidence fusion on the multi-view Dirichlet distribution parameters to obtain the time window-level Dirichlet distribution parameters. ; The temporal evidence fusion module aggregates temporal information and performs a second round of evidence fusion on the Dirichlet distribution parameters of each time window to obtain the final Dirichlet distribution parameters. This enables the classification and identification of brain diseases.

[0014] Thirdly, the present invention provides a terminal, comprising: Processor, memory, among which, This memory is used to store computer programs. The processor is used to retrieve and run the computer program from memory, causing the terminal to perform the terminal method described above.

[0015] Fourthly, the present invention provides a computer storage medium storing instructions that, when executed on a computer, cause the computer to perform the methods described in the above aspects.

[0016] The beneficial effects of this invention are as follows: This invention uses a sliding window technique to divide brain region time series into multiple time windows, capturing the dynamic time-varying characteristics of brain functional connectivity and avoiding static analysis from obscuring key temporal information. At the same time, it integrates four functional connectivity measures: Pearson correlation, higher-order functional connectivity, mutual information, and sparse representation, which respectively characterize linear relationships, higher-order interactions, nonlinear dependencies, and sparse dependencies between brain regions. By making full use of complementary information from multiple views, it significantly improves the comprehensiveness and richness of functional brain network representation and solves the technical pain point of limited representation capabilities of a single measure.

[0017] For spatiotemporal multi-view functional brain networks, a multi-layer graph convolutional encoder is used to deeply mine the spatial structural features of the brain network. Through adjacency matrix normalization, feature transformation and stacking of activation functions, the local and global correlations of brain regions are accurately captured. Combined with pooling and fully connected processing of the multilayer perceptron classification head, the high-dimensional brain network features are transformed into low-dimensional, compact view-level outputs, which effectively enhances the feature discrimination power for distinguishing brain disease categories and provides high-quality basic features for subsequent classification and fusion.

[0018] Spatially, based on Dempster-Shafer theory, the Dirichlet distribution parameters of multiple views within the same time window are fused. By calculating classification beliefs, uncertainties, and conflict factors, the reliability of different views is dynamically adapted. Temporally, evidence fusion is performed again on the fusion results of each time window to integrate complementary information from the temporal dimension. This two-layer fusion mechanism can adaptively handle view heterogeneity among different subjects and in different time windows, effectively reducing interference from spurious connections and individual differences, and significantly improving the accuracy, stability, and generalization ability of brain disease identification. It maintains stable performance in multi-site and multi-brain region template data.

[0019] The preprocessing stage utilizes MATLAB's DPARSF toolbox, combined with head movement correction and bandpass filtering to effectively remove noise interference and ensure data quality. It supports the selection of two brain region templates, AAL116 and CC200, and the sliding window parameters (window length W, step size s) can be flexibly configured to adapt to rs-fMRI data under different acquisition conditions. The classification process incorporates non-negative activation functions and Dirichlet distributions to model uncertainties, making the classification results more interpretable and providing a reliable reference for clinical auxiliary diagnosis. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic flowchart illustrating a method according to an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the architecture of the two-layer evidence fusion network of the present invention.

[0023] Figure 3 This is a schematic block diagram of a system according to an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the structure of a terminal provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the specific embodiments. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0027] The brain disease identification method based on spatiotemporal multi-view functional brain network provided in this embodiment of the invention is executed by a computer device, and correspondingly, the brain disease identification system based on spatiotemporal multi-view functional brain network runs in the computer device.

[0028] Figure 1 This is a schematic flowchart illustrating a method according to an embodiment of the present invention. Wherein, Figure 1 The implementing entity can be a brain disease recognition system based on a spatiotemporal multi-view functional brain network. Depending on different needs, the order of the steps in this flowchart can be changed, and some can be omitted.

[0029] like Figure 1 As shown, the method includes: S1. The acquired brain functional images are preprocessed, and the brain functional images are divided into several brain regions based on predefined brain region templates. The time series of each brain region is extracted to obtain a time series matrix; where the brain functional images correspond to brain disease category labels. S2. The time series matrix is ​​divided into several time windows using a sliding window analysis strategy to obtain a set of time windows; S3. Within each time window, the correlation matrix is ​​calculated using several functional connectivity metrics to construct a spatiotemporal multi-view functional brain network. The functional connectivity metrics include Pearson correlation, higher-order functional connectivity, mutual information, and sparse representation. S4. For each spatiotemporal multi-view functional brain network, construct a graph convolutional encoder and a multilayer perceptron classification head, perform representation learning to obtain view-level output, obtain view evidence through a non-negative activation function, and calculate the Dirichlet distribution parameters. ,in Number of brain disease categories; S5. Aggregate spatial dimension information and perform the first round of evidence fusion on the multi-view Dirichlet distribution parameters within each time window t to obtain the time window-level Dirichlet distribution parameters. ; S6. Aggregate time-dimensional information and perform a second round of evidence fusion on the Dirichlet distribution parameters of each time window to obtain the final Dirichlet distribution parameters. This enables the classification and identification of brain diseases.

[0030] To facilitate understanding of the present invention, the following description further illustrates the brain disease identification method based on spatiotemporal multi-view functional brain network provided by the present invention, using the principle of the present invention and the process of brain disease identification based on spatiotemporal multi-view functional brain network in the embodiments.

[0031] First, step S1 includes: S11. Preprocess the acquired resting-state functional magnetic resonance images of the brain. The brain functional images are labeled with brain disease categories. Use the DPARSF toolbox in MATLAB to remove the first p time points of the image, and perform head motion artifacts and time layer correction to eliminate interference caused by ventricular, white matter signals and higher-order head motion effects. Register the corrected image to the standard space and apply a 0.01Hz to 0.1Hz time bandpass filter to reduce the influence of heartbeat and respiratory noise. S12. Based on predefined brain region templates, the preprocessed resting-state functional magnetic resonance imaging (fMRI) images are divided into N brain regions. The brain region templates are selected from the AAL116 template or the CC200 template. The AAL116 template divides the resting-state fMRI images into 116 brain regions, and the CC200 template divides the resting-state fMRI images into 200 brain regions to ensure the standardization and reproducibility of brain region division. S13. Extract the time series data of each brain region to form a time series matrix. ,in Let T0 represent the BOLD signal vectors of N brain regions at time t, where T0 is the total number of time points and N is the number of brain regions.

[0032] Using MATLAB R2023a software and the accompanying DPARSF 5.0 toolbox, the raw data (in NIfTI format) of resting-state functional magnetic resonance imaging (rs-fMRI) of the brain were imported. The data needed to be associated with the corresponding brain disease category label (e.g., autism spectrum disorder ASD, healthy controls HC).

[0033] S12. Brain region segmentation based on predefined templates: Template selection and import: Predefined brain region templates are selected from publicly available standardized template libraries, including AAL116 and CC200 templates. Template files (in NIfTI format) are imported through the “ROI Definition” module of the DPARSF toolbox.

[0034] Template matching and brain region segmentation: The preprocessed image, registered to the MNI152 standard space, is spatially matched with selected brain region templates. Each voxel is assigned to its corresponding brain region using a voxel overlap determination method. Specifically: When using the AAL116 template, the image is divided into 116 relatively independent brain regions (including core brain regions such as gray matter, white matter-related areas, and the limbic system). Specifically, the brain regions are divided using the DPARSF toolbox, resulting in 116 effective brain regions covering core functional areas such as the prefrontal cortex, parietal lobe, temporal lobe, occipital lobe, and limbic system. The brain region labels are consistent with the standard labels of the AAL template.

[0035] When the CC200 template is selected, the image is divided into 200 brain regions. This template provides a more refined breakdown of the default brain regions and is suitable for brain network analysis scenarios that require higher spatial resolution.

[0036] Brain region validity verification: After segmentation, extremely small brain regions with fewer than 5 voxels are automatically removed (to avoid unreliable signal extraction), and brain region label files are output to clarify the spatial range and index of each brain region, ensuring the accuracy of subsequent time series extraction.

[0037] By performing a series of preprocessing operations, such as removing initial unstable time points, head movement artifact correction, physiological noise removal, and bandpass filtering, interference from irrelevant factors such as scanning equipment, subject movement, and physiological activities was effectively eliminated, significantly improving the signal-to-noise ratio and purity of rs-fMRI data, and providing a high-quality data foundation for subsequent brain network construction.

[0038] Images of all subjects were registered to a unified MNI152 standard space and divided using AAL116 / CC200 standardized brain region templates. This eliminated individual differences in the location of brain anatomical structures among different subjects, ensuring spatial consistency of data from different subjects and different sites, and providing a guarantee for multi-center data fusion analysis and model generalization.

[0039] By extracting the time series of BOLD signals from each brain region, high-dimensional rs-fMRI image data is transformed into low-dimensional, interpretable brain functional features, which directly reflects the temporal changes in neural activity in each brain region and provides core input data for subsequent sliding window analysis and multi-view functional connectivity calculations.

[0040] It supports the selection of two brain region templates, AAL116 and CC200, and parameters such as p-value and smoothing kernel size can be adjusted according to data characteristics, adapting to the data processing needs of different scanning devices and different types of brain diseases, thus improving the universality of the technical solution. The standardized time-series matrix eliminates signal amplitude differences, ensuring the accuracy of subsequent functional connectivity measurement (such as Pearson correlation and mutual information) calculations, reducing the impact of data heterogeneity on model performance, and providing a reliable prerequisite for the construction of spatiotemporal multi-view brain networks and two-layer evidence fusion.

[0041] Secondly, step S2 includes: S21. Configure the parameters required for sliding window analysis, including window length W and sliding step size s, where W is the number of time points covered by each time window and s is the number of interval points of window sliding, which is used to control the continuity and overlap of the time windows. S22. Based on parameters W and s, the time series matrix X is divided into T consecutive time windows. Each time window X_t extracts brain region signals from W consecutive time points in the original sequence, and the overlapping or non-overlapping relationship between adjacent time windows is controlled by the step size s. S23. Integrate all time windows to form a time window set. ,in This represents the data matrix for the t-th time window, where T is the total number of time windows. T_0 represents the original time point.

[0042] In addition, step S3 includes: S31. Within each time window X_t, select multiple functional connectivity metrics to construct a multi-view brain network. The functional connectivity metrics include Pearson correlation, higher-order functional connectivity, mutual information, and sparse representation. S32. For linear correlation measures, calculate the Pearson correlation coefficient matrix. : ; in, The Pearson correlation coefficient between the i-th and j-th brain regions; S33. Calculating higher-order functional connectivity matrices based on Pearson correlation matrices. : ; in, To standardize the Pearson correlation matrix, , and are the mean and standard deviation of the i-th row of the Pearson correlation matrix, respectively; Represents the L2 norm; S34. For nonlinear correlation metrics, calculate the mutual information matrix. : ; in, and These are the marginal probability distributions of the i-th and j-th brain regions, respectively; for and The joint probability distribution is calculated from the time series normalized to [0,1] using histogram estimation; simultaneously, the sparse representation matrix is ​​calculated. : ; in, This represents the standardized time series vector of the i-th brain region within the time window t; The coefficients representing the sparse representation of brain region j to brain region i are denoted by . For regularization parameters; S35. Integrate all functional connection matrices to construct a spatiotemporal multi-view functional brain network. ;in, Each view corresponds to a metric; The node feature matrix; This is the adjacency matrix for the corresponding metric.

[0043] The selection criteria for metrics: In response to the multi-dimensional representation needs of linear relationships, higher-order interactions, non-linear dependencies and sparse dependencies in brain regions, the set of functional connectivity metrics was determined to be Pearson correlation (PC), higher-order functional connectivity (HOFC), mutual information (MI), and sparse representation (SR). These four metrics complement each other and cover the core features of brain functional connectivity.

[0044] Input data preparation: Obtain the set of time windows obtained in step S2, and for each time window... (W is the window length, N is the number of brain regions), for each Standardization preprocessing (Z-score standardization) was performed to ensure that the mean of the time series data for each brain region was 0 and the variance was 1, thus eliminating the influence of amplitude differences on connectivity metric calculation.

[0045] The metric calculation order is as follows: the calculation is performed in the order of "linear metric → higher-order metric → nonlinear metric". That is, the PC matrix is ​​calculated first, the HOFC matrix is ​​derived based on the PC matrix, and finally the MI matrix and SR matrix are calculated in parallel to improve the calculation efficiency and ensure the logical correlation between metrics.

[0046] For time windows The i-th and j-th brain regions Extract its standardized time series and Calculate the Pearson correlation coefficient. .

[0047] Matrix Construction and Verification: Arrange the correlation coefficients of all paired brain regions by row (i) and column (j) to construct the Pearson correlation coefficient matrix. The range of values ​​for matrix elements is [ [1,1]. Outliers with an absolute value greater than 0.95 (considered strong noise interference) are removed and replaced with the mean of the non-outliers in that row to ensure the rationality of the matrix.

[0048] Calculation of the Higher-Order Functional Connectivity Matrix (HOFC): PC matrix normalization: the PC matrix obtained in step S32 Standardize by row to obtain the standardized matrix. The formula is: ; in, for The mean of the i-th row. for The standard deviation of the i-th row.

[0049] View definition: The four functional connection matrices are each treated as an independent view, i.e., a set of views, where m=1 corresponds to the PC view, m=2 corresponds to the HOFC view, m=3 corresponds to the MI view, and m=4 corresponds to the SR view.

[0050] View validity verification: Calculate the sparsity (proportion of non-zero elements) of the adjacency matrix of each view. If the sparsity is lower than 5% or higher than 95%, readjust the parameters of the corresponding metrics (e.g., λ for SR, number of bins for MI) and recalculate to ensure the effective representation capability of the view.

[0051] Multi-view set output: Integrates all validated views to form a spatiotemporal multi-view functional brain network set corresponding to time window t. .

[0052] (1) Pearson correlation coefficient matrix (PC): The Pearson correlation coefficient matrix is ​​used to capture the linear relationship between brain regions. For each time window X t First, the time series data for each brain region is processed to remove the mean, resulting in... ,in This represents the signal of the i-th brain region at time t. and These represent the mean and standard deviation of the time series data for this brain region, respectively. Then, the Pearson correlation coefficient between the i-th and j-th brain regions is calculated. as follows: ; in, and Let be the mean values ​​of brain regions i and j over time, respectively. The calculated Pearson correlation coefficient matrix is ​​denoted as: ; (2) Higher-Order Functional Connectivity Matrix (HOFC): The HOFC can capture complex higher-order interactions between brain regions. Within each time window... First, the Pearson correlation coefficient matrix P is calculated, and then standardized to obtain the standardized matrix. : ; in, and Let the mean and standard deviation of the i-th row be the values, respectively. Then, calculate the higher-order functional connectivity matrix. Recorded as: ; (3) Mutual Information Matrix (MI): The mutual information matrix is ​​used to capture nonlinear relationships between brain regions and can reflect the degree of information sharing between brain regions at different times. For each time window... First, normalization is performed to normalize the time series data of each brain region to intervals. Next, the mutual information between each pair of brain regions j and j is calculated. as follows: ; in, and These are the marginal probabilities of the i-th and j-th brain regions, respectively. This is their joint probability. Ultimately, we obtain the mutual information functional connectivity matrix. .

[0053] (4) Sparse Representation Matrix (SR): The sparse representation matrix is ​​used to reveal sparse dependencies between brain regions. For each time window... First, normalization is performed to obtain a standardized time series matrix. Subsequently, for each brain region i, a linear reconstruction is performed using the time series data of the remaining brain regions, and redundant connections are suppressed through regularization constraints. Specifically, the sparse representation is calculated as follows: ; in, This represents the standardized time series vector of the i-th brain region within the time window t. λ represents the representation coefficient of brain region j to brain region i, and λ is a regularization parameter used to control model complexity and alleviate overfitting.

[0054] This study integrates four metrics—Pearson correlation, higher-order functional connectivity, mutual information, and sparse representation—to characterize linear relationships, higher-order interactions, nonlinear dependencies, and sparse dependencies between brain regions, comprehensively covering the complex patterns of brain functional connectivity and overcoming the limitations of single-metric representation. By constructing a multi-view brain network within each time window, it preserves the temporal characteristics of dynamic functional connectivity while integrating connectivity information from different dimensions through multi-view complementarity, resulting in a more comprehensive brain network representation that better reflects the true state of brain neural activity.

[0055] Outlier removal from the PC matrix, normalization of HOFC and MI, and regularization constraints on SR effectively reduce the impact of noise interference and spurious connections, improving the quality of the functional connectivity matrix and providing high-quality input for subsequent representation learning. The combination of these four metrics can specifically capture the differences in functional connectivity between patients with brain diseases and healthy individuals (e.g., linear association strength, higher-order interaction patterns, nonlinear information transfer), providing highly discriminative structured features for subsequent evidence fusion and classification. The regularization parameters of SR and the number of histogram bins in MI can be adjusted according to data characteristics, supporting different brain region templates such as AAL116 and CC200, adapting to different acquisition conditions and different types of brain diseases in rs-fMRI data, thus improving the universality of the technical solution.

[0056] Next, step S4 includes: S41. For each time window t and view m, construct a graph convolutional encoder. This encoder uses a multi-layer graph convolutional neural network, similar to a spatiotemporal multi-view functional brain network. Adjacency matrix in As the adjacency matrix of the graph, the node representation is updated through graph convolution operations, using a preset node feature matrix X_m(t)∈R^{N×F} as input. The formula is as follows: ; in, This represents the updated node embedding matrix; N is the number of brain regions; F is the input node feature dimension; G is the feature dimension of the graph convolution output; The weight matrix is ​​a learnable matrix; For bias terms; The non-linear activation function is used; the graph convolutional encoder is composed of multiple stacked graph convolutional layers, each of which includes adjacency matrix normalization, feature transformation and activation function operation, thereby extracting the spatial structural features of the brain network. S42, Output of the graph convolution encoder A multilayer perceptron classification head is constructed, and the average along the brain region dimension is used to obtain a time window-level graphical representation. Its formula is: ; S43. Represent the pooled graph. The input is a multilayer perceptron (MLP), which consists of multiple fully connected layers, normalization layers, and activation functions. The view-level output is obtained through forward propagation. Its formula is: ; Where C is the number of brain disease categories; the specific structure of the MLP includes an input layer (dimension G), a hidden layer (dimension configurable) and an output layer (dimension C), and the hidden layer uses LayerNorm normalization and the GELU activation function; S44, View-level output Calculate view evidence using the non-negative activation function Softplus(·). Its formula is: ; The Softplus function ensures that the evidence is non-negative. S45. Calculate the Dirichlet distribution parameters based on the evidence deep learning framework. Its formula is: ; The view-level Dirichlet distribution parameters are obtained and used to model classification uncertainty, where, The kth component The strength of evidence corresponds to the k-th class, where C is the number of classes.

[0057] The multi-layer GCN encoder accurately captures local associations and global topological structures of brain regions through adjacency matrix normalization and stacked feature transformation. Compared with traditional networks, it can more fully explore the discriminative spatial features of brain functional networks, providing a high-quality foundation for classification. Global average pooling embeds and aggregates N×G-dimensional nodes into a 1×G-dimensional global graph representation, significantly reducing feature dimensionality while preserving core information, thereby reducing the computational complexity of subsequent MLP classification heads and improving model running efficiency.

[0058] The MLP classification head, combined with LayerNorm normalization, GELU activation, and Dropout techniques, effectively mitigates overfitting, enhances the non-linear expressive power of features, and improves view-level output. It can more accurately distinguish different categories of brain diseases. Through the Softplus function and Dirichlet distribution parameter mapping, the view-level output is transformed into physically meaningful evidence and uncertainty indicators, quantifying the degree of support each view provides for different categories and offering a reliable quantitative basis for subsequent adaptive evidence fusion. The regularization of GCN, the Dropout layer, and the normalization processing of MLP collectively improve the model's adaptability to data from different subjects and time windows, making the view-level Dirichlet parameters more precise. It exhibits good stability, ensuring the effectiveness of two-layer evidence fusion.

[0059] The graph convolutional encoder employs a multi-layer graph convolutional neural network (GCN) to process the feature matrix of the input nodes. and adjacency matrix Update the node representation as follows: ; in Let m be the adjacency matrix of the view m within the time window t; Let N be the node feature matrix of the inner view m within the time window t, where N is the number of brain regions and F is the node feature dimension. G is a learnable weight matrix, and G is the feature dimension of the graph convolution output. For bias terms; It is a non-linear activation function.

[0060] Encoder network structure definition: The graph convolutional encoder adopts a 3-layer stacked graph convolutional neural network (GCN) structure, with each layer having a clearly defined function: Input layer: Spatiotemporal multiview functional brain network The adjacency matrix (Functional connection matrix from step S3), node feature matrix (F is the dimension of the input feature, ranging from 20 to 50); Intermediate Layers (2 layers): Responsible for feature dimensionality enhancement and spatial structure extraction; the first layer outputs the feature dimension. The second layer outputs feature dimensions. ; Output layer: Outputs the final node embedding matrix (G=128 is the final output feature dimension).

[0061] Adjacency matrix normalization: To avoid gradient vanishing or exploding, each layer of GCN normalizes the adjacency matrix. Normalization is performed using the following formula: ; in, It is an N-order identity matrix (self-loops are added to preserve the characteristics of the nodes themselves). For degree matrix, (The diagonal elements of the degree matrix are the sum of the elements in each row of the normalized adjacency matrix).

[0062] Node representation update calculation: Each layer of GCN updates node features according to the following formula: .

[0063] The multilayer perceptron classification head processes the encoded multi-view graph representation, obtaining the class output logits for each view through several fully connected layers, normalization layers, and activation functions. Where C is the number of categories. Specifically: Output of graph convolutional neural network Averaging along the brain region dimension yields a time-window-level graphical representation:

[0064] Then, the pooled graph representation As input, the data is processed by a multilayer perceptron (MLP) to obtain the category output logits. : ; MLP consists of multiple fully connected layers, normalization layers, and activation functions.

[0065] The graph convolutional encoder employs a multi-layered GCN stacked structure, combined with adjacency matrix normalization, effectively avoiding gradient vanishing or exploding problems. It can deeply capture local connections and global topological structures between brain regions. Through feature transformation of the learnable weight matrix and the action of nonlinear activation functions, it can more fully exploit the discriminative spatial features of brain functional networks compared to traditional networks, providing high-quality basic feature support for subsequent classification tasks and meeting the representation needs of structured data from brain networks.

[0066] Global average pooling aggregates an N×G dimensional node embedding matrix into a 1×G dimensional temporal window-level graph representation. While fully preserving the core spatial information of the brain network, it significantly reduces the feature dimensionality, decreases the computational complexity of subsequent MLP classification heads, and improves the overall efficiency of the model. Simultaneously, the pooling process can suppress noise interference from individual brain regions, highlight the overall functional patterns of the brain network, and enhance the robustness of features.

[0067] The MLP classification head employs a combined design of "fully connected layer + LayerNorm normalization + GELU activation + Dropout". LayerNorm effectively alleviates the feature distribution shift problem, the GELU activation function enhances the non-linear expressive power of features compared to traditional activation functions, and the Dropout layer effectively suppresses overfitting. The three work synergistically to achieve a high-level O(n) output. m (t) It can more accurately distinguish different categories of brain diseases and improve the reliability of classification results.

[0068] The MLP output is processed using the Softplus non-negative activation function to ensure that evidence E m (t) is always non-positive, strictly conforming to the mathematical requirements of evidence in the evidence-based deep learning framework. At the same time, by imposing a reasonable upper limit constraint on the evidence value, the model instability caused by excessively large values ​​is avoided, providing high-quality and compliant input data for subsequent Dirichlet distribution parameter calculation.

[0069] Mapping relationship α through Dirichlet distribution parameters m ,t=E m,t+1, transforms the view-level output into an indicator of evidence strength and uncertainty with clear physical meaning. α m Each component of t directly corresponds to the degree of evidence support for a certain category. By combining the derivation of total strength, classification belief and uncertainty, the classification reliability of each view can be quantified, providing accurate and interpretable quantitative basis for the subsequent spatial dimension evidence fusion of S5, and solving the problem that traditional fusion methods are difficult to adapt to view heterogeneity.

[0070] The L2 regularization and Dropout layer of the graph convolutional encoder, along with the normalization processing of the MLP classification head, form a multi-layered regularization system, effectively improving the model's adaptability to data from different subjects and time windows. Even when faced with individual differences and noise interference in rs-fMRI data, it can still output stable view-level Dirichlet parameters, laying a solid foundation for the effectiveness of the entire two-layer evidence fusion mechanism and ensuring the model's stable performance in multi-site and multi-brain region template data.

[0071] In addition, step S5 includes: S51. For each time window t, the Dirichlet distribution parameters based on each view m. Calculate the total Dirichlet intensity for each view. ,evidence Classification beliefs and view uncertainty The specific formula is as follows: ; ; ; ; Where k represents the category index; this step provides a quantitative basis for the reliability of each view in evidence fusion; S52. The first round of fusion is performed using an evidence fusion framework to fuse the Dirichlet distribution parameters of all views within time window t. This framework, based on Dempster-Shafer theory, sequentially fuses the parameters in the view set pairwise: for any two views... and (corresponding to) Calculate the intensity of any two views in the diagram. and Classification beliefs and Uncertainty and The conflict factor K is calculated to measure consistency between views, and the merged classification beliefs are updated. and uncertainty The specific formula is as follows: ; ; ; ; ; ; This fusion process is applied iteratively to all views until the parameters of the entire view set are integrated into a single representation; S53, Classification Beliefs Based on Fusion and uncertainty Derivation of Time Window-Level Dirichlet Distribution Parameters Calculate the total intensity after fusion, S_t=C / u, and calculate the evidence. Finally obtained Each component The first round of evidence fusion in the spatial dimension was completed, and time window-level parameters were obtained. .

[0072] By calculating classification beliefs, uncertainty, and conflict factors, the reliability of different views is dynamically adapted, avoiding the problem of traditional fixed-weight fusion ignoring view heterogeneity. Higher implicit weights are assigned to high-reliability views, improving the accuracy of the fusion results. View reliability is quantified as an uncertainty index, which can identify and reduce interference from low-reliability views during the fusion process. Simultaneously, conflict factors address inconsistencies between views, reducing the impact of false evidence on the fusion results and making time-window parameters more stable. Based on a pairwise fusion strategy using Dempster-Shafer theory, complementary features of four views (PC, HOFC, MI, and SR) are gradually aggregated, integrating multi-dimensional information including linear, higher-order, nonlinear, and sparse connections, resulting in a more comprehensive time-window representation and overcoming the limitations of single-view representation. It supports various brain disease recognition scenarios, including binary and multi-class classification. The fusion logic is independent of the specific number of views or categories, and can flexibly adapt to new functional connectivity metrics or expanded disease categories, exhibiting good scalability.

[0073] Finally, step S6 includes: S61. Integrate the set of time window-level Dirichlet distribution parameters for all time windows. This collection contains spatial fusion results from various local time periods in the time dimension; S62. Set of Dirichlet distribution parameters for each time window along the time dimension. The second round of evidence fusion is performed using the same evidence fusion framework as in step S5, fusing the parameters in the time window set pairwise: for any two time windows, the Dirichlet distribution parameters... and Calculate its strength and Classification beliefs and Uncertainty and And calculate the conflict factor K and update the fused classification beliefs. and uncertainty The specific formula is as follows: ; ; ; ; ; ; This fusion process is applied iteratively to all time window parameters until the parameters of the entire time window set are integrated into a single final Dirichlet distribution parameter. This completes the second round of evidence fusion in the time dimension; S63, Based on the final Dirichlet distribution parameters Achieving Classification and Recognition of Brain Diseases: Computing Final Evidence and total strength The classification result is determined by taking the expected value or the maximum probability, whereby the probability of each category is calculated. The category with the highest probability is selected as the final diagnosis result, using the following formula: ; in, To predict brain disease categories, we aim to achieve intelligent diagnosis of brain diseases based on two-layer evidence fusion.

[0074] The evidence fusion framework in steps S5 and S6 is as follows: Figure 2 As shown: Dirichlet parameters for the two views to be merged. and The intensity S, classification belief b, and view uncertainty u are calculated based on Dirichlet parameters. The calculation process is as follows: ; ; ; Then, calculate the conflict factor K: ; The classification belief b and the uncertainty u of the fused view are updated based on the conflict factor K: .

[0075] During training, the Dirichlet distribution parameters at the view level are... and fusion parameters , The loss function based on the Dirichlet distribution is used for optimization, including a classification term and a KL regularization term, specifically: (1) Given the one-hot vector representation of the true label y, calculate the total intensity with respect to the Dirichlet parameter α. The loss for the classification item is defined as: ; in, The true label of class k (either 1 or 0 in the one-hot vector); is the digamma function, used to calculate the expectation of the Dirichlet distribution; The total intensity of the Dirichlet distribution parameters.

[0076] Construct a parameter vector that only penalizes non-true classes. The component corresponding to the true category is fixed at 1, while the component corresponding to the non-true category remains unchanged. And calculate Dirichlet With uniform Dirichlet The KL divergence between them is used as the regularization term LKL: ; KL divergence measures the difference between the current Dirichlet distribution and the uniform distribution, thus making the model's Dirichlet distribution more focused on the true class.

[0077] Total loss is defined as:

[0078] Where λ is a weighting coefficient that adapts to the number of training steps, used to balance classification accuracy and the reasonableness of evidence distribution.

[0079] like Figure 1 As shown, this invention provides a two-layer evidence fusion model based on spatiotemporal multi-view functional brain networks for brain disease identification (i.e., a brain disease identification method based on spatiotemporal multi-view functional brain networks), comprising the following steps: S1. Preprocess the acquired brain functional images, divide the brain into several brain regions based on predefined brain region templates, and extract the time series of each brain region. ,in T0 represents the signals of N brain regions at time t, where T0 is the time point number and N is the number of brain regions. S2. Using a sliding window analysis strategy, the above time series matrix X is divided into several time windows. Let the window length be W and the step size be s, to obtain the set of time windows. ,in This represents the t-th time window; S3. Within each time window Xt, calculate the correlation matrix using multiple functional connectivity metrics, specifically selecting Pearson correlation (PC): High-order function connections (HOFC): Mutual Information (MI): Sparse representation (SR): Subsequently, a spatiotemporal multiview functional brain network was obtained at each time slice t. ; S4. For each spatiotemporal graph, construct a graph convolutional encoder and a multilayer perceptron classification head. Perform representation learning on the spatiotemporal multiview functional brain network to improve classification accuracy and prepare for subsequent evidence modeling. Specifically, for each time window t and view... Connect the corresponding functional matrices As the adjacency matrix of the graph, a pre-defined node feature matrix is ​​used to obtain node embeddings through a graph convolutional network, and then a time window-level output is obtained through a multilayer perceptron. , where C is the number of categories. For the output... Through nonnegative activation function Obtaining view evidence And calculate the Dirichlet distribution parameters. The view-level Dirichlet distribution is obtained. ; S5. Aggregate spatial dimension information and perform the first round of evidence fusion on the multi-view Dirichlet distribution within each time window. Based on the Dirichlet parameters of each view. Calculate the total Dirichlet intensity ,evidence Classification beliefs Uncertainty of view m Using an evidence fusion framework, views within each time window are fused. Obtain the time window level Dirichlet parameters. ; S6. Aggregate time-dimensional information and adjust the Dirichlet parameters for each time window along the time dimension. A second round of evidence fusion is performed, using the evidence fusion framework to calculate the relevant parameters and fuse them to obtain the final Dirichlet parameters. This enables the classification and identification of brain diseases.

[0080] Example 1: The training data includes: Autism Spectrum Disorder (ASD): A total of 1,112 participants were included, of which 539 were patients and 573 were healthy controls (HC). Alzheimer's disease (AD): A total of 832 participants, including 418 patients and 414 healthy controls (HC); Schizophrenia (SZ): A total of 628 subjects were included, including 302 patients and 326 healthy controls (HC); Major depressive disorder (MDD): A total of 496 participants were included, including 245 patients and 251 healthy controls (HC).

[0081] Data partitioning: All datasets were merged and divided into training set (60.3%), validation set (19.8%), and test set (19.9%) in a 6:2:2 ratio. Stratified sampling was used to ensure a balanced distribution of data for each disease and each site, avoiding class bias.

[0082] All training images are resting-state functional magnetic resonance imaging (rs-fMRI) images, and their standardized specifications are as follows: Data format: NIfTI format (.nii.gz), compatible with cross-platform processing in MATLAB / Python; Unified scanning parameters: repetition time (TR) = 2s, echo time (TE) = 30±2ms, flip angle = 90°, matrix size = 64×64, voxel size = 3mm×3mm×4mm; Time series specifications: 140±10 original time points, 130 valid time points are retained after preprocessing; Key features of the image: captures blood oxygen level dependent (BOLD) signal fluctuations, non-invasive, in compliance with clinical ethics, and with clear differences in characteristics corresponding to different diseases (e.g., weakened default network connectivity in AD patients and abnormal prefrontal-parietal connectivity in SZ patients).

[0083] The labeling system can be divided into two categories: basic binary classification and extended multi-class classification. Basic binary classification primarily targets single brain disease identification scenarios. The core categories include healthy controls (HC) and target brain diseases (e.g., autism spectrum disorder, ASD). The healthy control label corresponds to subjects with no history of neurological or psychiatric illness and who show no abnormalities according to the DSM-5 (Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition) assessment. The target brain disease label (using ASD as an example) corresponds to subjects who meet the DSM-5 autism diagnostic criteria and exhibit typical symptoms such as social communication impairment. In this scenario, the labels use a two-dimensional one-hot encoding format, with healthy controls corresponding to [1,0] and target brain diseases corresponding to [0,1], adapting to the Dirichlet distribution loss calculation for binary classification tasks.

[0084] To adapt to the needs of recognizing multiple types of brain diseases, the labeling system can be expanded to five categories, adding three new disease labels: Alzheimer's disease (AD), schizophrenia (SZ), and major depressive disorder (MDD). The Alzheimer's disease label corresponds to subjects who meet the NINCDS-ADRDA diagnostic criteria and have cognitive decline; the schizophrenia label corresponds to subjects who meet the DSM-5 schizophrenia diagnostic criteria and have symptoms such as hallucinations or delusions; and the major depressive disorder label corresponds to subjects who meet the DSM-5 depression diagnostic criteria and have persistent low mood symptoms for ≥2 weeks. The expanded multi-classification scenario adopts a five-dimensional one-hot encoding format, with healthy controls corresponding to [1,0,0,0,0], autism spectrum disorder corresponding to [0,1,0,0,0], Alzheimer's disease corresponding to [0,0,1,0,0], schizophrenia corresponding to [0,0,0,1,0], and major depressive disorder corresponding to [0,0,0,0,1]. This ensures that each category is independent in encoding, facilitating the model's accurate learning of category differences.

[0085] Training stops when the training process reaches the preset termination condition (e.g., model performance convergence or reaching the maximum number of training rounds), and the trained model is used to infer the test data to output the final brain disease classification result based on the fusion of two-layer evidence in spatial and temporal dimensions. That is, the output shows whether the patient belongs to autism spectrum disorder (ASD), Alzheimer's disease (AD), schizophrenia (SZ), major depressive disorder (MDD), or a healthy person.

[0086] In addition, in the field of deep learning, adapting the same model architecture to different classification tasks (binary classification / multi-class classification) is a mature technology: task switching can be achieved simply by adjusting the output dimension of the classification head (number of matching categories) and the label encoding format (one-hot dimension synchronization), which will not be elaborated here.

[0087] Example 2: This technical solution can also be used for class identification of specific brain disorders, that is, for the identification of one type of brain disorder, such as for the identification of autism spectrum disorder (ASD). The training process of the corresponding model is as follows: The training data includes: Autism Spectrum Disorder (ASD, ABIDE dataset): Resting-state functional magnetic resonance imaging (rs-fMRI) data from the NYU and UM sites were selected for binary classification modeling experiments.

[0088] NYU site: A total of 184 participants, including 79 patients with acute mental illness (ASD) and 105 healthy controls (HC); UM site: A total of 145 participants, including 68 patients (ASD) and 77 healthy controls (HC).

[0089] Data partitioning: Five-fold cross-validation was used for experimental evaluation, and the two sites were run independently without cross-site data merging. Specifically, the NYU and UM sites were stratified and partitioned into five mutually exclusive subsets according to the ASD / HC categories. In each iteration, one subset was selected as the test set, and the remaining four subsets were used as the training set (the training set could be further divided into a validation set for hyperparameter selection and early stopping). The iteration was repeated five times, and the results of each subset were averaged to reduce the randomness of a single partition and avoid class bias.

[0090] All training images are resting-state functional magnetic resonance imaging (rs-fMRI) images, and their standardized specifications are as follows: Data format: NIfTI format (.nii.gz), compatible with cross-platform processing in MATLAB / Python; Scan parameters (set by site): NYU site: Scanner 3.0 T Allegra; Repetition time (TR) = 2000 ms, Echo time (TE) = 15 ms, Flip angle = 90°, Number of layers = 33, Number of time points = 180, Voxel size = 3.0 mm × 3.0 mm × 4.0 mm, Single scan duration approximately 6 min; UM site: Scanner 3.0 T Philips; Repetition time (TR) = 3000 ms, Echo time (TE) = 30 ms, Flip angle = 80°, Number of layers = 48, Number of time points = 140, Voxel size = 3.438 mm × 3.438 mm × 3.000 mm, Single scan duration approximately 7 min; Time series specifications: 180 original time points for NYU site and 140 original time points for UM site; valid time points are retained after preprocessing for feature extraction and model training.

[0091] Core image features: Captures blood oxygen level-dependent (BOLD) signal fluctuations and functional connectivity patterns. It is non-invasive, ethically sound, and can be used to learn the differences in brain functional network features between ASD and HC.

[0092] Labeling System: This invention employs a basic binary classification labeling system, with categories including healthy controls (HC) and autism spectrum disorder (ASD). The healthy control label corresponds to participants with no history of neurological or psychiatric illness and normal DSM-5 assessment results; the ASD label corresponds to participants who meet the DSM-5 diagnostic criteria for autism and exhibit typical symptoms such as social communication impairment. In this scenario, the labels use a two-dimensional one-hot encoding format, with healthy controls corresponding to [1,0] and ASD corresponding to [0,1], adapting to the Dirichlet distribution loss calculation for binary classification tasks.

[0093] In this embodiment, to ensure training stability, the learning rate is set to 0.002 and the batch size to 16, to accommodate the characteristics of FRI numerical degradation and limited sample size. During the modeling of dynamic functional connections (ffkJ / frozen function links), the sliding window width is set to 25 and the step size to 20. By introducing overlapping regions between adjacent time windows, the temporal resolution and stability of ffkJ / frozen function link estimation are balanced. During the training process of evidence deep learning and evidence fusion, this embodiment introduces a KL regularized annealing mechanism based on the Dirichlet distribution, with annealing steps set to 500. This allows the weights of the KL regularization terms to gradually increase during training, thereby avoiding the adverse effects of excessively strong regularization constraints on evidence modeling in the early stages of training and improving the stability and reliability of uncertainty modeling during multi-view evidence fusion. Training stops when the training process reaches the preset termination condition (e.g., model performance convergence or the maximum number of training rounds is reached), and the trained model is used to infer the test data, outputting the final brain disease classification result based on the fusion of two-layer evidence in spatial and temporal dimensions.

[0094] Both training methods only require adjusting three parameters: "number of categories C", "label dimension", and "data filtering range". There is no need to modify the core algorithm or architecture of the model, and there are no technical obstacles.

[0095] To verify the effectiveness of the model proposed in this invention, this technical solution selects data from two sites, NYU and UM, in the publicly available Autism Brain Imaging Data Exchange dataset (ABIDE) for experimental evaluation, and tests the model performance under two brain region segmentation templates, AAL116 and CC200, respectively. The relevant evaluation results are shown in Tables 1 and 2.

[0096] Table 1: Model evaluation metrics on the ABIDE dataset NYU site, where 116 represents the AAL116 template and 200 represents the CC200 template.

[0097] Table 2: Model evaluation metrics on the ABIDE dataset UM site. 116 represents the AAL116 template, and 200 represents the CC200 template.

[0098] As shown in Table 1, on the ABIDE dataset at the NYU site, the model of this invention achieved stable classification performance under both AAL116 and CC200 brain region templates. Specifically, under the AAL116 template, the model's classification accuracy (ACC) reached approximately 78%, AUC was close to 0.80, and the balance accuracy (BAC) and F1 score also remained at a high level. Under the CC200 template, the model also demonstrated good classification ability, with all indicators remaining generally stable, indicating that the proposed model has a certain degree of adaptability to different brain region segmentation schemes.

[0099] As shown in Table 2, the model of this invention also achieved relatively consistent experimental results on the ABIDE dataset UM site. Under the AAL116 template, the model's classification accuracy was close to 80%, and both AUC and BAC remained at a high level, achieving a good balance between sensitivity and specificity. Under the CC200 template, although the model performance fluctuated slightly, it still maintained good classification stability overall, indicating that the method of this invention has strong robustness under different site acquisition conditions and different brain region templates.

[0100] The experimental results from both NYU and UM sites demonstrate that the proposed two-layer evidence fusion model can stably complete brain disease classification tasks under multi-site and multi-brain region segmentation conditions. Experimental results show that by introducing spatiotemporal multi-view functional brain network modeling, evidence-based uncertainty modeling, and a two-layer fusion strategy, the model maintains good classification performance even under complex brain functional connectivity patterns, exhibiting good generalization ability and practical application potential.

[0101] In some embodiments, the brain disease recognition system 200 based on spatiotemporal multiview functional brain networks may include multiple functional modules composed of computer program segments. The computer programs for each program segment in the brain disease recognition system 200 based on spatiotemporal multiview functional brain networks may be stored in the memory of a computer device and executed by at least one processor to perform (see details). Figure 1 (Description) Functional brain disease identification based on spatiotemporal multi-view functional brain network.

[0102] In this embodiment, the brain disease recognition system 200 based on a spatiotemporal multi-view functional brain network can be divided into multiple functional modules according to the functions it performs, such as... Figure 3As shown. The functional modules may include: a preprocessing module 210, a sliding window analysis module 220, a multi-view brain network construction module 230, a graph convolution and classification module 240, a spatial evidence fusion module 250, and a temporal evidence fusion module 260. The module referred to in this invention is a series of computer program segments that can be executed by at least one processor and perform a fixed function, stored in memory. In this embodiment, the functions of each module will be described in detail in subsequent embodiments.

[0103] The preprocessing module 210 is used to preprocess the acquired brain functional images and divide them into several brain regions based on predefined brain region templates, extracting the time series of each brain region to obtain a time series matrix; wherein, the brain functional images correspond to brain disease category labels; the sliding window analysis module 220 is used to divide the time series matrix into several time windows using a sliding window analysis strategy to obtain a set of time windows; the multi-view brain network construction module 230 is used to calculate the correlation matrix using several functional connectivity metrics within each time window to construct a spatiotemporal multi-view functional brain network; the functional connectivity metrics include Pearson correlation, higher-order functional connectivity, mutual information, and sparse representation; the graph convolution and classification module 240 is used to construct a graph convolutional encoder and a multilayer perceptron classification head for each spatiotemporal multi-view functional brain network, perform representation learning and obtain view-level output, obtain view evidence through a non-negative activation function, and calculate the Dirichlet distribution parameters. ,in The number of brain disease categories; the spatial evidence fusion module 250, used to aggregate spatial dimension information, performs the first round of evidence fusion on the multi-view Dirichlet distribution parameters within each time window t, and obtains the time window-level Dirichlet distribution parameters. The time-based evidence fusion module 260 is used to aggregate time-dimensional information and perform a second round of evidence fusion on the Dirichlet distribution parameters of each time window to obtain the final Dirichlet distribution parameters. This enables the classification and identification of brain diseases.

[0104] Figure 4 This is a schematic diagram of the structure of a terminal 300 provided in an embodiment of the present invention. The terminal 300 can be used to execute the brain disease identification method based on spatiotemporal multi-view functional brain network provided in the embodiment of the present invention.

[0105] The terminal 300 may include a processor 310, a memory 320, and a communication module 330. These components communicate via one or more buses. Those skilled in the art will understand that the server structure shown in the figure does not constitute a limitation of the present invention. It may be a bus topology or a star topology, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0106] The memory 320 can be used to store the execution instructions of the processor 310. The memory 320 can be implemented by any type of volatile or non-volatile memory terminal or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. When the execution instructions in the memory 320 are executed by the processor 310, the terminal 300 is able to perform some or all of the steps in the above method embodiments.

[0107] The processor 310 serves as the control center of the storage terminal, connecting various parts of the electronic terminal via various interfaces and lines. It executes software programs and / or modules stored in the memory 320, and calls data stored in the memory to perform various functions of the electronic terminal and / or process data. The processor can be composed of integrated circuits (ICs), such as a single packaged IC or multiple packaged ICs with the same or different functions connected together. For example, the processor 310 may consist only of a central processing unit (CPU). In this embodiment of the invention, the CPU may have a single processing core or include multiple processing cores.

[0108] The communication module 330 is used to establish a communication channel, enabling the storage terminal to communicate with other terminals. It receives user data sent by other terminals or sends user data to other terminals.

[0109] The present invention also provides a computer storage medium, wherein the computer storage medium may store a program, which, when executed, may include some or all of the steps provided in the embodiments of the present invention. The storage medium may be a magnetic disk, an optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0110] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium such as a USB flash drive, mobile hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or other media capable of storing program code. It includes several instructions to cause a computer terminal (which may be a personal computer, server, or a second terminal, network terminal, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0111] The same or similar parts between the various embodiments in this specification can be referred to mutually. In particular, the terminal embodiments are basically similar to the method embodiments, so the description is relatively simple, and the relevant parts can be referred to the description in the method embodiments.

[0112] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or modules may be electrical, mechanical, or other forms.

[0113] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0114] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0115] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the present invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the present invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should also be covered within the protection scope of the present invention.

Claims

1. A method for identifying brain diseases based on spatiotemporal multi-view functional brain networks, characterized in that, include: S1. The acquired brain functional images are preprocessed, and the brain functional images are divided into several brain regions based on predefined brain region templates. The time series of each brain region is extracted to obtain a time series matrix; where the brain functional images correspond to brain disease category labels. S2. The time series matrix is ​​divided into several time windows using a sliding window analysis strategy to obtain a set of time windows; S3. Within each time window, the correlation matrix is ​​calculated using several functional connectivity metrics to construct a spatiotemporal multi-view functional brain network. The functional connectivity metrics include Pearson correlation, higher-order functional connectivity, mutual information, and sparse representation. S4. For each spatiotemporal multi-view functional brain network, construct a graph convolutional encoder and a multilayer perceptron classification head, perform representation learning to obtain view-level output, obtain view evidence through a non-negative activation function, and calculate the Dirichlet distribution parameters. ,in Number of brain disease categories; S5. Aggregate spatial dimension information and perform the first round of evidence fusion on the multi-view Dirichlet distribution parameters within each time window t to obtain the time window-level Dirichlet distribution parameters. ; S6. Aggregate time-dimensional information and perform a second round of evidence fusion on the Dirichlet distribution parameters of each time window to obtain the final Dirichlet distribution parameters. This enables the classification and identification of brain diseases.

2. The brain disease identification method based on spatiotemporal multi-view functional brain network according to claim 1, characterized in that, Step S1 includes: S11. Preprocess the acquired resting-state functional magnetic resonance images of the brain. The brain functional images are labeled with brain disease categories. Use the DPARSF toolbox of MATLAB to remove the first p time points of the image, and perform head motion artifacts and time layer correction. Register the corrected image to the standard space and apply a time bandpass filter of 0.01Hz to 0.1Hz. S12. Based on predefined brain region templates, the preprocessed resting-state functional magnetic resonance imaging (fMRI) images are divided into N brain regions. The brain region templates are selected from the AAL116 template or the CC200 template. The AAL116 template divides the resting-state fMRI images into 116 brain regions, and the CC200 template divides the resting-state fMRI images into 200 brain regions. S13. Extract the time series data of each brain region to form a time series matrix. ,in Let T0 represent the BOLD signal vectors of N brain regions at time t, where T0 is the total number of time points and N is the number of brain regions.

3. The brain disease identification method based on spatiotemporal multi-view functional brain network according to claim 2, characterized in that, Step S2 includes: S21. Configure the parameters required for sliding window analysis, including window length W and sliding step size s; S22. Based on parameters W and s, the time series matrix X is divided into T consecutive time windows. Each time window X_t extracts brain region signals from W consecutive time points in the original sequence, and the overlapping or non-overlapping relationship between adjacent time windows is controlled by the step size s. S23. Integrate all time windows to form a time window set. ,in This represents the data matrix for the t-th time window, where T is the total number of time windows. T_0 is the original time point number.

4. The brain disease identification method based on spatiotemporal multi-view functional brain network according to claim 1, characterized in that, Step S3 includes: S31. Within each time window X_t, select multiple functional connectivity metrics to construct a multi-view brain network. The functional connectivity metrics include Pearson correlation, higher-order functional connectivity, mutual information, and sparse representation. S32. For linear correlation measures, calculate the Pearson correlation coefficient matrix. : ； in, The Pearson correlation coefficient between the i-th and j-th brain regions; S33. Calculating higher-order functional connectivity matrices based on Pearson correlation matrices. : ； in, To standardize the Pearson correlation matrix, , and are the mean and standard deviation of the i-th row of the Pearson correlation matrix, respectively; Represents the L2 norm; S34. For nonlinear correlation metrics, calculate the mutual information matrix. : ； in, and These are the marginal probability distributions of the i-th and j-th brain regions, respectively; for and The joint probability distribution is calculated from the time series normalized to [0,1] using histogram estimation; simultaneously, the sparse representation matrix is ​​calculated. : ； in, This represents the standardized time series vector of the i-th brain region within the time window t; The coefficients representing the sparse representation of brain region j to brain region i are denoted by . For regularization parameters; S35. Integrate all functional connection matrices to construct a spatiotemporal multi-view functional brain network. ;in, Each view corresponds to a metric; The node feature matrix; This is the adjacency matrix for the corresponding metric.

5. The brain disease identification method based on spatiotemporal multi-view functional brain network according to claim 4, characterized in that, Step S4 includes: S41. For each time window t and view m, construct a graph convolutional encoder. This encoder uses a multi-layer graph convolutional neural network, similar to a spatiotemporal multi-view functional brain network. Adjacency matrix in As the adjacency matrix of the graph, the node representation is updated through graph convolution operations, using a preset node feature matrix X_m(t)∈R^{N×F} as input. The formula is as follows: ； in, This represents the updated node embedding matrix; N is the number of brain regions; F is the input node feature dimension; G is the feature dimension of the graph convolution output; The weight matrix is ​​a learnable matrix; For bias terms; The non-linear activation function is used; the graph convolutional encoder is composed of multiple stacked graph convolutional layers, each of which includes adjacency matrix normalization, feature transformation and activation function operation, thereby extracting the spatial structural features of the brain network. S42, Output of the graph convolution encoder A multilayer perceptron classification head is constructed, and the average along the brain region dimension is used to obtain a time window-level graphical representation. Its formula is: ； S43. Represent the pooled graph. The input is a multilayer perceptron (MLP), which consists of multiple fully connected layers, normalization layers, and activation functions. The view-level output is obtained through forward propagation. Its formula is: ； Where C represents the number of brain disease categories; the specific structure of the MLP includes an input layer, a hidden layer, and an output layer, with the hidden layer using LayerNorm normalization and the GELU activation function; S44, View-level output Calculate view evidence using the non-negative activation function Softplus(·). Its formula is: ； The Softplus function ensures that the evidence is non-negative. S45. Calculate the Dirichlet distribution parameters based on the evidence deep learning framework. Its formula is: ； The view-level Dirichlet distribution parameters are obtained and used to model classification uncertainty, where, The kth component The strength of evidence corresponds to the k-th class, where C is the number of classes.

6. The brain disease identification method based on spatiotemporal multi-view functional brain network according to claim 5, characterized in that, Step S5 includes: S51. For each time window t, the Dirichlet distribution parameters based on each view m. Calculate the total Dirichlet intensity for each view. ,evidence Classification beliefs and view uncertainty The specific formula is as follows: ； ； ； ； Where k represents the category index; S52. The first round of fusion is performed using an evidence fusion framework to fuse the Dirichlet distribution parameters of all views within time window t. This framework, based on Dempster-Shafer theory, sequentially fuses the parameters in the view set pairwise: for any two views... and Calculate its strength and Classification beliefs and Uncertainty and The conflict factor K is calculated to measure consistency between views, and the merged classification beliefs are updated. and uncertainty The specific formula is as follows: ； ； ； ； ； ； This fusion process is applied iteratively to all views until the parameters of the entire view set are integrated into a single representation; S53, Classification Beliefs Based on Fusion and uncertainty Derivation of Time Window-Level Dirichlet Distribution Parameters Calculate the total intensity after fusion, S_t=C / u, and calculate the evidence. Finally obtained Each component The first round of evidence fusion in the spatial dimension was completed, and time window-level parameters were obtained. .

7. The brain disease identification method based on spatiotemporal multi-view functional brain network according to claim 6, characterized in that, Step S6 includes: S61. Integrate the set of time window-level Dirichlet distribution parameters for all time windows. This collection contains spatial fusion results from various local time periods in the time dimension; S62. Set of Dirichlet distribution parameters for each time window along the time dimension. The second round of evidence fusion is performed using the same evidence fusion framework as in step S5, fusing the parameters in the time window set pairwise: for any two time windows, the Dirichlet distribution parameters... and Calculate its strength and Classification beliefs and Uncertainty and And calculate the conflict factor K and update the fused classification beliefs. and uncertainty The specific formula is as follows: ； ； ； ； ； ； This fusion process is applied iteratively to all time window parameters until the parameters of the entire time window set are integrated into a single final Dirichlet distribution parameter. This completes the second round of evidence fusion in the time dimension; S63, Based on the final Dirichlet distribution parameters Achieving Classification and Recognition of Brain Diseases: Computing Final Evidence and total strength The classification result is determined by taking the expected value or the maximum probability, whereby the probability of each category is calculated. The category with the highest probability is selected as the final diagnosis result, using the following formula: ； in, To predict brain disease categories, we aim to achieve intelligent diagnosis of brain diseases based on two-layer evidence fusion.

8. A brain disease recognition system based on spatiotemporal multi-view functional brain networks, characterized in that, include: The preprocessing module is used to preprocess the acquired brain functional images, divide the brain functional images into several brain regions based on predefined brain region templates, extract the time series of each brain region, and obtain a time series matrix; among them, the brain functional images correspond to brain disease category labels; The sliding window analysis module is used to divide the time series matrix into several time windows using the sliding window analysis strategy, thereby obtaining a set of time windows. The multi-view brain network construction module is used to calculate the correlation matrix within each time window using several functional connectivity metrics to construct a spatiotemporal multi-view functional brain network; the functional connectivity metrics include Pearson correlation, higher-order functional connectivity, mutual information, and sparse representation. The graph convolution and classification module is used to construct a graph convolutional encoder and a multilayer perceptron classification head for each spatiotemporal multi-view functional brain network, perform representation learning and obtain view-level output, obtain view evidence through a non-negative activation function, and calculate the Dirichlet distribution parameters. ,in Number of brain disease categories; The spatial evidence fusion module is used to aggregate spatial dimension information. Within each time window t, it performs the first round of evidence fusion on the multi-view Dirichlet distribution parameters to obtain the time window-level Dirichlet distribution parameters. ; The temporal evidence fusion module aggregates temporal information and performs a second round of evidence fusion on the Dirichlet distribution parameters of each time window to obtain the final Dirichlet distribution parameters. This enables the classification and identification of brain diseases.

9. A terminal, characterized in that, include: processor; Memory used to store the processor's execution instructions; The processor is configured to perform the method of any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

Images