Brain disease classification method and system

Abstract

The invention discloses a brain disease classification method and system. Precise diagnosis is realized through multi-modal data fusion and dynamic modeling. The method comprises the following steps: collecting multi-modal brain image information and cognitive behavior information of a user; performing dynamic function connection analysis on the resting state functional magnetic resonance time sequence signal to obtain a time-varying brain network feature matrix, and performing white matter fiber bundle topology reconstruction on a structure connection matrix; constructing a four-dimensional correlation tensor by using the time-varying network features, the structural connection weights and the anatomical features through a neurodynamic model; performing multi-task learning on the four-dimensional correlation tensor based on a time-varying graph neural network model, and outputting a quantitative diagnosis result; and finally generating a clinical classification report integrating the individualized brain network remodeling target, the disease progress risk layering and the treatment response prediction. By dynamically fusing the structure and functional features, comprehensive characterization of the pathological mechanism of the brain disease is realized, and decision support with both accuracy and interpretation is provided for clinical diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis, and in particular to a brain disease classification method and system. Background Art

[0002] The clinical diagnosis of brain diseases primarily relies on a combination of medical imaging and neuropsychological assessment. Currently, commonly used diagnostic methods are typically based on a single type of brain scan data, such as analyzing morphological changes in specific brain regions using structural magnetic resonance imaging or calculating correlations between different brain regions using functional imaging data. While these methods can identify some abnormal features, they struggle to fully reflect the complex pathological mechanisms of brain diseases.

[0003] In terms of functional analysis, conventional methods often treat functional connections between brain regions as fixed and unchanging, ignoring the dynamic nature of neural activity. Techniques used to analyze brain structural connectivity typically focus solely on the physical connectivity characteristics of white matter fibers, failing to fully consider their interactions with neural functional activity. Existing computer-aided diagnosis systems mostly use traditional classification algorithms, which have limited ability to integrate multi-source data and typically output simple classification labels that lack an understanding of the underlying biological mechanisms of the disease.

[0004] With the advancement of medical technology, clinical practice has placed higher demands on the diagnosis of brain diseases, requiring new analytical methods that can integrate multi-dimensional information, reflect the dynamic characteristics of the disease, and provide interpretable results. This demand is particularly prominent in the diagnosis of neurological diseases with complex pathological mechanisms, such as neurodegenerative diseases and mental disorders. Summary of the Invention

[0005] In view of the above problems, the present invention provides a brain disease classification method and system. By dynamically fusing multimodal brain images and cognitive behavioral data, a time-varying neural network model is constructed to achieve accurate classification and mechanism explanation, solving the problem of difficulty in comprehensively characterizing the dynamic pathological characteristics of brain diseases.

[0006] To achieve the above objectives, in a first aspect, the present application provides a method for classifying brain diseases, comprising: Collect multimodal brain imaging information and cognitive behavioral information from users. Multimodal brain imaging information includes resting-state functional magnetic resonance imaging time series signals, diffusion tensor imaging structural connectivity matrix, and T1-weighted anatomical features. Cognitive behavioral information is obtained through standardized neuropsychological scales. Dynamic functional connectivity analysis was performed on the resting-state functional magnetic resonance (fMRI) time series signals to obtain a time-varying brain network feature matrix. The dynamic functional connectivity analysis was configured to calculate phase synchrony based on a sliding window, and white matter fiber bundle topology was reconstructed on the diffusion tensor imaging structural connectivity matrix to obtain a structural connectivity weight matrix. The time-varying brain network feature matrix, structural connection weight matrix and T1-weighted anatomical features are used to construct a four-dimensional correlation tensor through a neurodynamic model. The dimensions of the four-dimensional correlation tensor include spatial dimension, temporal dimension, connection strength dimension and gray matter morphology dimension. A time-varying graph neural network model is used to perform multi-task joint learning on a four-dimensional correlation tensor, outputting quantitative diagnostic results including disease classification probabilities, markers of neural circuit abnormalities, and trajectories of cognitive decline. The time-varying graph neural network model is configured as a coupled oscillator embedding module with biophysical constraints. An auxiliary clinical classification report containing personalized brain network remodeling targets, disease progression risk stratification, and treatment response prediction is generated based on the degree of cross-modal alignment of quantitative diagnostic results and cognitive behavioral information.

[0007] In some embodiments, dynamic functional connectivity analysis is performed on resting-state functional magnetic resonance (fMRI) time series signals to obtain a time-varying brain network feature matrix. The dynamic functional connectivity analysis is configured to calculate phase synchrony based on a sliding window, including: Performing a sliding window analysis on resting-state functional magnetic resonance (fMRI) time series signals to obtain multiple sliding windows. The sliding window analysis is configured as a sliding window algorithm based on an adaptive window length. The window length of the sliding window is dynamically adjusted according to the signal-to-noise ratio of signals from multiple brain regions. The phase synchronization strength of the brain region corresponding to each sliding window is calculated and expressed by formula (1). Formula (1) is as follows: ; In formula (1), for Momentary brain area and brain areas The strength of phase synchronization between is the total number of time points in a single sliding window, For brain area for The instantaneous phase at a time point, For brain area for The instantaneous phase at a time point, Computational brain area and brain areas The complex exponential function of the phase difference vector between To count letters, ; Repeat the above steps until the phase synchronization strength of all brain regions is calculated, and construct a two-dimensional dynamic functional connectivity matrix based on the phase synchronization strength, which is expressed by formula (2). Formula (2) is as follows: ; In formula (2), is the dynamic functional connectivity matrix; The dynamic functional connectivity matrix is reduced in dimension by non-negative matrix decomposition, and the time-varying brain network feature matrix is extracted and expressed by formula (3). Formula (3) is as follows: ; In formula (3), is the static network component after non-negative matrix decomposition, is the time-varying brain network feature matrix.

[0008] In some embodiments, performing white matter fiber bundle topology reconstruction on the diffusion tensor imaging structural connectivity matrix to obtain a structural connectivity weight matrix includes: A probabilistic fiber tracking algorithm was used to reconstruct the whole-brain white matter fiber pathways based on the diffusion tensor imaging structural connectivity matrix, and the initial structural connectivity strength of the brain region was calculated based on the spatial density distribution of the fiber bundles. The initial structural connection matrix is normalized by graph theory measurement to obtain the structural connection matrix, which is expressed by formula (4). Formula (4) is as follows: ; In formula (4), For brain area and brain areas The initial structural connection strength between is the average anisotropy fraction of the fiber bundle, is the number of connected fibers, is the average fiber length, For brain area and brain areas The distance between the centroids, is the Gaussian kernel width; The structural connection matrix is expressed by formula (5), which is as follows: ; In formula (5), is the structural connection matrix; Perform diffusion geometric mapping on the structural connection matrix to obtain a low-dimensional manifold space; The spectral clustering algorithm is used to identify features in the low-dimensional manifold space to obtain modular features, which include the intra-module connection enhancement coefficient and the inter-module connection attenuation coefficient. The structural connection matrix is weighted according to the modular characteristics, and the intra-module connection enhancement coefficient and inter-module connection attenuation coefficient are used for dual parameter adjustment to obtain the structural connection weight matrix.

[0009] In some embodiments, a probabilistic fiber tracking algorithm is used to reconstruct the whole-brain white matter fiber pathways based on the diffusion tensor imaging structural connectivity matrix, and the initial structural connectivity strength between brain regions is calculated based on the spatial density distribution of the fiber bundles, including: In the anisotropic field defined by the diffusion tensor imaging structural connectivity matrix, the white matter fiber propagation paths were simulated by Monte Carlo random walks. Each simulated white matter fiber propagation path was probabilistically sampled based on the direction of the local diffusion tensor principal eigenvector, resulting in multiple fiber bundle paths. The fiber bundle paths generated by repeated sampling are spatially clustered, and the connected paths passing through the preset white matter skeleton area are retained and recorded as the initial fiber bundle paths; The cross-distribution density of the initial fiber bundle paths between various brain regions was counted, and geometric correction was performed based on the length of the initial fiber bundle paths to generate an initial structural connection matrix with directional weights.

[0010] In some embodiments, a spectral clustering algorithm is applied to the low-dimensional manifold space to perform feature recognition to obtain modular features, including: Calculate multiple eigenvectors based on the Laplacian matrix of the low-dimensional manifold space and construct the eigenvector space; The k-means clustering algorithm was used to partition the feature vector space and obtain the modular partitioning of the brain network; Calculate the membership of each brain region in the module of the modular partitioning of the brain network and generate a modular feature matrix; The structural connection matrix is weighted according to the modular characteristics, and the intra-module connection enhancement coefficient and inter-module connection attenuation coefficient are used for dual parameter adjustment to obtain the structural connection weight matrix, including: Apply an enhancement coefficient to the brain area connections belonging to the same module to improve the connection strength within the module, which is recorded as the intra-module connection enhancement coefficient; Apply an attenuation coefficient to the cross-module brain area connection to suppress the inter-module connection strength, which is recorded as the inter-module connection attenuation coefficient; Repeat the above steps until all brain regions are traversed; The connection enhancement coefficients within multiple modules and the connection attenuation coefficients between modules are normalized to generate a structural connection weight matrix with modular topological constraints.

[0011] In some embodiments, a four-dimensional correlation tensor is constructed by combining the time-varying brain network feature matrix, the structural connection weight matrix, and the T1-weighted anatomical features through a neurodynamic model, including: Construct a multimodal feature fusion framework to map the dynamic phase synchronization pattern of the time-varying brain network feature matrix to the energy transfer equation of the neural mass unit; A white matter conduction delay constraint is imposed on the structural connectivity weight matrix, which is coupled to the spatial propagation term of the neural mass unit via the diffusion-reaction equation; Extracting the cortical thickness and gray matter density distribution from T1-weighted anatomical features as the local excitability regulation parameters of neural mass units; By integrating the energy transfer equation, spatial propagation term and excitability regulation parameters through tensor product operation, a four-dimensional correlation tensor with four-dimensional characteristics of space-time-connectivity-morphology is generated. The spatial dimension is configured as the spatial topological arrangement of neural mass units, the temporal dimension is configured as the oscillation period of dynamic phase synchronization, the connection strength dimension is configured as the synaptic weight of white matter conduction, and the gray matter morphology dimension is configured as the gradient change of cortical thickness.

[0012] In some embodiments, the dynamic phase synchronization pattern of the time-varying brain network feature matrix is mapped to the energy transfer equation of the neural mass unit, which is expressed by formula (6), which is as follows: ; In formula (6), is the total number of brain regions, Indicates the The energy state of each neural mass unit, represents the normalized functional connection weight of the time-varying brain network feature matrix, Indicates the The oscillation phase of the neural mass unit, Indicates the The oscillation phase of the neural mass unit, is the energy attenuation coefficient, is the external input stimulus intensity; It is coupled to the spatial propagation term of the neural mass unit through the diffusion-reaction equation and expressed by formula (7), which is as follows: ; In formula (7), is the neuronal membrane time constant, For the The action potential of a neural mass unit, represents the structural connection weight in the structural connection matrix, is the delay time variable, represents the white matter conduction delay time between brain regions, is the Sigmoid activation function, is the Gaussian noise term; The energy transfer equation, spatial propagation term, and excitability regulation parameter are integrated through tensor product operation to generate a four-dimensional correlation tensor with the four-dimensional characteristics of space-time-connectivity-morphology, which is expressed by formula (8). Formula (8) is as follows: ; In formula (8), represents a four-dimensional correlation tensor, represents a neural mass unit, Representation and Neural Mass Unit The target neural mass unit of the structural connection, represents the energy state time series, represents the discretization level of white matter conduction delay time, Indicates the discretization level The representative value of represents the tensor product operation, Indicates the The normalized value is the product of gray matter density and cortical thickness in each brain region.

[0013] In some embodiments, a time-varying graph neural network model is used to perform multi-task joint learning on a four-dimensional correlation tensor, outputting quantitative diagnostic results including disease classification probability, neural circuit abnormality markers, and cognitive function decline trajectory, including: Constructing biophysically constrained coupled oscillator modules in time-varying graphical neural network models, including: The characteristics contained in the time dimension of the four-dimensional correlation tensor are mapped into the oscillator phase dynamics equation. The natural frequency of each oscillator unit is determined by the resting-state EEG characteristic spectrum, and the coupling strength is dynamically adjusted by the connection strength dimension in the four-dimensional correlation tensor. A phase-locked detection algorithm is used to calculate the instantaneous phase synchronization matrix of the oscillator cluster, which is used as the dynamic connection feature input of the graph neural network. Spatiotemporal graph convolution based on anatomical structure constraints processes dynamic connection features. Spatiotemporal graph convolution includes spatial dimension convolution processing and temporal dimension convolution processing. Spatial dimension convolution processing includes spatial graph convolution. The adjacency matrix of spatial graph convolution is limited to the range of white matter connection paths determined by DTI fiber tracking. Temporal dimension convolution processing includes deformable convolution kernel to capture multi-scale features from event-related potentials to functional state fluctuations. Synchronously update node features and connection weights in the convolution operation of each layer of spatiotemporal graph convolution until the processing is completed; Generate quantitative diagnostic results through a multi-task attention mechanism, including: The multi-task attention mechanism includes disease classification branch, neural circuit abnormality monitoring branch, and cognitive prediction branch; The disease classification branch calculates the contribution weight of each brain region to the classification decision through dynamic graph pooling, and obtains the first quantitative result containing the first attention weight; The neural circuit anomaly detection branch uses the edge attention score to identify abnormal pathways where the functional connectivity strength and structural connectivity density deviate significantly, and obtains a second quantitative result containing a second attention weight. The cognitive prediction branch uses the temporal attention algorithm to calculate the fluctuation period and obtains the third quantitative result containing the third attention weight; A quantitative diagnosis result is generated according to the first quantitative result, the second quantitative result, and the third quantitative result.

[0014] In some embodiments, multimodal brain imaging information and cognitive behavioral information of the user are collected, where the multimodal brain imaging information includes resting-state functional magnetic resonance imaging time series signals, diffusion tensor imaging structural connectivity matrix, and T1-weighted anatomical features, including: The resting-state functional magnetic resonance timing signals were configured to be acquired using a preset functional magnetic resonance imaging sequence; The diffusion tensor imaging structural connectivity matrix was configured to acquire data using a preset diffusion-weighted imaging protocol; T1-weighted anatomical features were configured to be acquired using a high-resolution 3D structural imaging sequence, and the brain was divided into a preset number of automatically anatomically labeled brain regions using an atlas-based segmentation method.

[0015] In the second aspect, the present invention also provides a brain disease classification system, which is applicable to the brain disease classification method described in the first aspect. The system includes a data acquisition module, a data processing module, a logic operation module and a result output module. The data acquisition module is used to collect the user's multimodal brain imaging information and cognitive behavior information. The multimodal brain imaging information includes resting-state functional magnetic resonance timing signals, diffusion tensor imaging structural connection matrix and T1-weighted anatomical features. The cognitive behavior information is obtained through a standardized neuropsychological scale; the data processing module is used to perform dynamic functional connection analysis on the resting-state functional magnetic resonance timing signals to obtain a time-varying brain network feature matrix. The dynamic functional connection analysis is configured to calculate the phase synchronization based on a sliding window, and to perform white matter fiber bundle topology reconstruction on the diffusion tensor imaging structural connection matrix to obtain Structural connection weight matrix; the logical operation module is used to construct a four-dimensional correlation tensor by combining the time-varying brain network feature matrix, the structural connection weight matrix and the T1-weighted anatomical features through a neurodynamic model. The dimensions of the four-dimensional correlation tensor include spatial dimension, temporal dimension, connection strength dimension and gray matter morphology dimension; based on the time-varying graph neural network model, the four-dimensional correlation tensor is subjected to multi-task joint learning to output quantitative diagnostic results including disease classification probability, neural circuit abnormality markers and cognitive function decline trajectory. The time-varying graph neural network model is configured as a coupled oscillator embedding module with biophysical constraints; the result output module is used to generate an auxiliary clinical classification report containing personalized brain network remodeling targets, disease progression risk stratification and treatment response prediction based on the degree of cross-modal alignment between the quantitative diagnostic results and cognitive behavioral information.

[0016] Different from the existing technology, the above technical solution provides a brain disease classification method and system, which achieves accurate diagnosis through multimodal data fusion and dynamic modeling. The method includes: collecting the user's multimodal brain imaging information and cognitive behavioral information, where the multimodal brain imaging information includes resting-state functional magnetic resonance time series signals, diffusion tensor imaging structural connection matrix and T1-weighted anatomical features; performing dynamic functional connectivity analysis on the resting-state functional magnetic resonance time series signals to obtain the time-varying brain network feature matrix, and at the same time reconstructing the white matter fiber bundle topology of the structural connection matrix; constructing a four-dimensional correlation tensor by using a neurodynamic model to combine the time-varying network features, structural connection weights and anatomical features; performing multi-task learning on the four-dimensional correlation tensor based on the time-varying graph neural network model, and outputting quantitative diagnostic results including disease classification probability, neural circuit abnormality markers and cognitive function decline trajectory; and finally generating a clinical classification report that integrates individualized brain network remodeling targets, disease progression risk stratification and treatment response prediction. The present invention achieves a comprehensive characterization of the pathological mechanism of brain diseases by dynamically fusing structural and functional features, providing decision support for clinical diagnosis that is both accurate and explanatory.

[0017] The above-mentioned records related to the content of the invention are only an overview of the technical solution of this application. In order to enable ordinary technicians in this field to understand the technical solution of this application more clearly, and then implement it according to the text of the specification and the contents recorded in the drawings, and to make the above-mentioned purposes and other purposes, features and advantages of this application easier to understand, the following is an explanation in combination with the specific implementation methods and drawings of this application. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings are only used to illustrate the principles, implementation methods, applications, features and effects of the specific embodiments of the present invention and other related contents, and are not to be considered as limiting the present application.

[0019] In the drawings of the specification: Figure 1 A method step diagram showing steps S101 to S105 of the classification method described in the specific embodiment; Figure 2 A method step diagram showing steps S201 to S203 of the classification method described in the specific embodiment; Figure 3 Schematic diagram of the structure of the classification system described in the specific implementation method.

[0020] The reference numerals in the above drawings are described as follows: 1. Classification system; 11. Data acquisition module; 12. Data processing module; 13. Logical operation module; 14. Result output module. DETAILED DESCRIPTION

[0021] In order to explain in detail the possible application scenarios, technical principles, specific solutions that can be implemented, and the purpose and effects of this application, the following is a detailed description of the specific embodiments listed in conjunction with the accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of this application and are therefore only examples and are not intended to limit the scope of protection of this application.

[0022] References to "embodiments" herein mean that the specific features, structures, or characteristics described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the word "embodiment" in various places in the specification does not necessarily refer to the same embodiment, nor does it particularly limit its independence or relevance to other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the various technical features mentioned in the embodiments can be combined in any manner to form a corresponding implementable technical solution.

[0023] Unless otherwise defined, the technical terms used herein have the same meanings as those generally understood by those skilled in the art to which this application belongs; the use of relevant terms herein is only for describing specific embodiments and is not intended to limit this application.

[0024] In the description of this application, the term "and / or" is used to describe a logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and both A and B exist. In addition, the character " / " in this document generally indicates that the objects before and after are in a logical "or" relationship.

[0025] In this application, terms such as "first" and "second" are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual quantity, priority or sequence relationship between these entities or operations.

[0026] Without further limitations, in this application, the words "include", "comprise", "have" or other similar open-ended expressions used in sentences are intended to cover non-exclusive inclusion. These expressions do not exclude the presence of additional elements in the process, method or product that includes the elements, so that the process, method or product that includes a series of elements may include not only those defined elements, but also other elements that are not explicitly listed, or also include elements inherent to such process, method or product.

[0027] Consistent with the understanding in the Examination Guidelines, in this application, expressions such as "greater than," "less than," and "exceed" are understood to exclude the number itself; expressions such as "above," "below," and "within" are understood to include the number itself. Furthermore, in the description of the embodiments of this application, "multiple" means two or more (including two), and similar expressions related to "multiple," such as "multiple groups" and "multiple times," are also understood in this manner, unless otherwise specifically defined.

[0028] In the description of the embodiments of the present application, the space-related expressions used, such as "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "vertical", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc., indicate the orientation or position relationship based on the orientation or position relationship shown in the specific embodiments or drawings, and are only for the convenience of describing the specific embodiments of the present application or facilitating the reader's understanding, and do not indicate or imply that the device or component referred to must have a specific position, a specific orientation, or be constructed or operated in a specific orientation. Therefore, it should not be understood as a limitation on the embodiments of the present application.

[0029] The processor described in the embodiments of the present application can be implemented by hardware, firmware, software or a combination thereof, and can use a circuit, a single or multiple application-specific integrated circuits (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, at least one of a microprocessor, and also includes other physical, biological or chemical structures that can achieve similar or equivalent functions to the processors listed above, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some or all of the steps in the computer program or method involved in the various embodiments of the present application, or any combination of the steps mentioned therein.

[0030] The computer program involved in the embodiment can be stored in a computer device readable storage medium, which includes but is not limited to a disk, a tape, a magnetic card, a floppy disk, a flash memory, an optical disc, an optical card, a read-only memory (ROM), a random access memory (RAM), an erasable programmable ROM (EPROM) and an electrically erasable programmable ROM (EEPROM), etc., and also includes other biological, physical or chemical structures that can achieve similar or equivalent functions to the storage media listed above, such as DNA, RNA, protein and other units with information storage capabilities. In a specific embodiment, the storage medium involved can be one of the above-mentioned media types or a combination of the above-mentioned media types. In different embodiments, the computer program involved in the embodiment can be stored in a single medium in a centralized manner or in a distributed manner on multiple media. The memory containing the computer device readable storage medium can be a non-volatile memory or a random access memory. These computer device readable storage media can be built into the device or connected to the device involved in the embodiment as an external device or part of an external device. In some embodiments, a memory having a computer-readable storage medium is deployed locally; in other embodiments, a solution of deploying the memory away from the processor may also be adopted, such as a network-attached memory accessed via an RF circuit or an external port and a communication network, wherein the communication network may be the Internet, one or more intranets, a local area network (LAN), a wide area network (WLAN), a storage area network (SAN), etc., or a suitable combination thereof, as long as the computer device can access the memory. In addition, the computer program involved in the embodiment can be stored in plaintext / ciphertext form, or can be designed as training data, which can be integrated and reorganized through model training and implicitly stored in the parameter state of a deep neural network or other machine learning model.

[0031] See also Figure 1 In a first aspect, this embodiment provides a method for classifying brain diseases, comprising: S101. Collect multimodal brain imaging information and cognitive behavioral information of the user. Multimodal brain imaging information includes resting-state functional magnetic resonance imaging time series signals, diffusion tensor imaging structural connectivity matrix, and T1-weighted anatomical features. Cognitive behavioral information is obtained through standardized neuropsychological scales. S102. Perform dynamic functional connectivity analysis on the resting-state functional magnetic resonance (fMRI) time series signals to obtain a time-varying brain network feature matrix. The dynamic functional connectivity analysis is configured to calculate phase synchrony based on a sliding window, and perform white matter fiber bundle topology reconstruction on the diffusion tensor imaging structural connectivity matrix to obtain a structural connectivity weight matrix. S103, constructing a four-dimensional correlation tensor by combining the time-varying brain network feature matrix, the structural connection weight matrix, and the T1-weighted anatomical features through a neurodynamic model, wherein the dimensions of the four-dimensional correlation tensor include a spatial dimension, a temporal dimension, a connection strength dimension, and a gray matter morphology dimension; S104. Perform multi-task joint learning on a four-dimensional correlation tensor based on a time-varying graph neural network model, outputting quantitative diagnostic results including disease classification probabilities, neural circuit abnormality markers, and cognitive function decline trajectories. The time-varying graph neural network model is configured as a coupled oscillator embedding module with biophysical constraints. S105. Generate an auxiliary clinical classification report containing individualized brain network remodeling targets, disease progression risk stratification, and treatment response prediction based on the degree of cross-modal alignment between quantitative diagnostic results and cognitive behavioral information.

[0032] In step S101, the acquisition of multimodal brain imaging information is completed by medical imaging equipment, among which the resting-state functional magnetic resonance timing signal reflects the dynamic functional activity between brain regions and can be obtained using a high-temporal resolution scanning protocol to ensure that the transient characteristics of neural activity can be captured; the diffusion tensor imaging structural connection matrix represents the physical connection characteristics of white matter fiber bundles and can be reconstructed through a multi-b-value sampling scheme to improve the resolution of crossing fibers; T1-weighted anatomical features are used to quantify morphological indicators such as gray matter volume in brain regions, and isotropic voxel scanning is preferably used to provide high-spatial precision data for subsequent morphological analysis; cognitive behavioral information is obtained through standardized neuropsychological scale assessment, including cognitive domain scores such as memory and executive function. The scale selection is determined based on the core cognitive deficit domain of the target disease to ensure the biological correlation between behavioral assessment and imaging characteristics.

[0033] In step S102, the dynamic functional connectivity analysis uses an adaptive sliding window technique to process the resting-state functional magnetic resonance timing signal, and captures the time-varying characteristics of the functional connectivity by calculating the phase synchronization between brain regions within the window, and finally forms a time-varying brain network feature matrix reflecting the dynamic changes of the brain network. The window length is dynamically adjusted according to the autocorrelation characteristics of the signal, which not only ensures the fineness of the time resolution but also maintains sufficient statistical reliability; preferably, the phase synchronization calculation uses a weighted phase lag index to effectively suppress the false connections caused by the volume conduction effect.

[0034] Simultaneously, the diffusion tensor imaging structural connectivity matrix undergoes white matter fiber tract topology reconstruction. Graph theory methods are used to calculate topological metrics such as node centrality to generate a structural connectivity weight matrix. Preferably, an adaptive threshold strategy based on streamline density is introduced into the white matter fiber tract topology reconstruction process. This strategy automatically determines the significance threshold for structural connectivity by analyzing the distribution characteristics of connection strength, thus avoiding subjective bias caused by manual setting. This step extracts key features of the brain network from both functional and structural perspectives.

[0035] In step S103, the neurodynamic model performs a multimodal fusion of the time-varying brain network feature matrix, the structural connectivity weight matrix, and the T1-weighted anatomical features. The resulting four-dimensional correlation tensor has a spatial dimension corresponding to brain region location, a temporal dimension recording dynamic changes in functional connectivity, a connection strength dimension representing structural connectivity weights, and a gray matter morphology dimension storing anatomical features. This tensor structure achieves spatiotemporal alignment of features from different modalities within a unified framework.

[0036] Preferably, the neurodynamic model adopts a coupled differential equation system to mathematically model the dynamic change law of the time-varying brain network feature matrix and the constraints of the structural connection weight matrix, where the coupling strength parameter is determined by fitting the actual neuroelectrophysiological data to ensure that the model has a real biophysical basis; the construction process of the four-dimensional correlation tensor adopts tensor decomposition technology to retain the most discriminative characteristic components in each modal data through low-rank approximation, while realizing data dimensionality reduction and feature fusion.

[0037] In step S104, the time-varying graphical neural network model simulates the biophysical properties of neural activity through a coupled oscillator embedding module, performing multi-task joint learning on the four-dimensional correlation tensor. The model simultaneously outputs three types of results: disease classification probability, neural circuit abnormality markers, and cognitive function decline trajectory. The disease classification probability reflects the confidence level of the diagnosis conclusion, the neural circuit abnormality markers indicate the dysfunction of specific brain networks, and the cognitive function decline trajectory predicts the cognitive change trend of the disease progression.

[0038] Preferably, the natural frequency parameters of the coupled oscillator embedding module are personalized according to the metabolic characteristics of each brain region; the attention weight is dynamically adjusted by learning the time-varying pattern of functional connections; multi-task joint learning adopts a hierarchical sharing mechanism, in which the underlying network shares the extraction of basic neural features, and the upper network is specialized for classification tasks, anomaly labeling tasks, and trajectory prediction tasks, respectively, to improve the performance of each task while ensuring model efficiency.

[0039] In step S105, preferably, a cross-modal Bayesian inference framework is established to achieve deep alignment of quantitative diagnostic results with cognitive behavioral information. In the generated auxiliary clinical classification report, the determination of individualized brain network remodeling targets comprehensively considers the spatial distribution intensity and temporal stability of abnormal markers; the threshold for disease progression risk stratification is set based on the statistical laws of large-sample longitudinal studies; and the prediction of treatment response is evaluated by combining pharmacodynamic models and the remodeling potential of functional networks. This multi-dimensional report not only provides decision-making recommendations but also includes credibility indicators for each conclusion, assisting doctors in risk-benefit trade-offs.

[0040] This example captures the dynamic characteristics of brain networks through time-series signal analysis and leverages multimodal fusion to overcome the limitations of a single data source, ultimately enabling cross-scale correlation analysis from microscopic neural circuit abnormalities to macroscopic cognitive and behavioral changes. Data collection ensures information integrity, feature extraction preserves biological characteristics, and model construction achieves multi-dimensional integration, ultimately delivering clinical application results that balance classification accuracy and mechanistic explanations.

[0041] In some embodiments, dynamic functional connectivity analysis is performed on resting-state functional magnetic resonance (fMRI) time series signals to obtain a time-varying brain network feature matrix. The dynamic functional connectivity analysis is configured to calculate phase synchrony based on a sliding window, including: Performing a sliding window analysis on resting-state functional magnetic resonance (fMRI) time series signals to obtain multiple sliding windows. The sliding window analysis is configured as a sliding window algorithm based on an adaptive window length. The window length of the sliding window is dynamically adjusted according to the signal-to-noise ratio of signals from multiple brain regions. The phase synchronization strength of the brain region corresponding to each sliding window is calculated and expressed by formula (1). Formula (1) is as follows: ; In formula (1), for Momentary brain area and brain areas The strength of phase synchronization between is the total number of time points in a single sliding window, For brain area for The instantaneous phase at a time point, For brain area for The instantaneous phase at a time point, Computational brain area and brain areas The complex exponential function of the phase difference vector between To count letters, ; Repeat the above steps until the phase synchronization strength of all brain regions is calculated, and construct a two-dimensional dynamic functional connectivity matrix based on the phase synchronization strength, which is expressed by formula (2). Formula (2) is as follows: ; In formula (2), is the dynamic functional connectivity matrix; The dynamic functional connectivity matrix is reduced in dimension by non-negative matrix decomposition, and the time-varying brain network feature matrix is extracted, which is expressed by formula (3). Formula (3) is as follows: ; In formula (3), is the static network component after non-negative matrix decomposition, is the time-varying brain network feature matrix.

[0042] In this embodiment, when resting-state functional magnetic resonance (fMRI) time-series signals are analyzed and processed using a sliding window, an adaptive window length algorithm dynamically adjusts the window length based on the signal-to-noise ratio (SNR) of the signals in each brain region. The SNR is calculated by comparing the signal power of the target brain region with the background noise power. When high-frequency neural activity components are detected, the window length is automatically shortened to improve temporal resolution. In low-frequency bands with higher signal stability, the window length is appropriately extended to ensure statistical reliability. This adaptive mechanism overcomes the limitations of a fixed window length in capturing neural activity across different frequency bands.

[0043] Phase synchronization strength was calculated using the phase locking value (PLV) metric, which quantifies the degree of functional coupling between brain regions by statistically analyzing the mean amplitude of the complex exponential function of the instantaneous phase difference vector within a sliding window. The instantaneous phase was extracted using the Hilbert transform to ensure that the timing characteristics of neural oscillations can be accurately reflected. For a single sliding window, all brain region pairs (e.g., brain region and brain areas After the phase synchronization strength of the ) is calculated, it is arranged in rows and columns to form a dynamic functional connectivity matrix , its diagonal elements represent the stability of spontaneous activities in each brain region, and the off-diagonal elements reflect the dynamic interaction patterns between brain regions.

[0044] When the dynamic functional connectivity matrix is reduced by non-negative matrix factorization, the static network components Preserving stable connectivity patterns across time, a time-varying brain network feature matrix The network state fluctuations that evolve over time are captured. Preferably, the decomposition process is optimized using alternating least squares, and the constraints ensure that the decomposition results have clear neurobiological interpretability. The static components correspond to the inherent connectivity patterns under the constraints of the anatomical structure, and the dynamic components reflect the spontaneous transitions of the brain state that are irrelevant to the task.

[0045] This example maintains a balance between temporal resolution and statistical power through adaptive window length; phase-locked value calculation avoids interference from signal amplitude fluctuations on functional connectivity estimation; and non-negative matrix factorization effectively separates stable patterns from dynamic changes while reducing dimensionality. These methods collectively ensure that the time-varying brain network feature matrix captures rapid fluctuations in neural activity while maintaining a sufficient signal-to-noise ratio for subsequent analysis.

[0046] In some embodiments, performing white matter fiber bundle topology reconstruction on the diffusion tensor imaging structural connectivity matrix to obtain a structural connectivity weight matrix includes: A probabilistic fiber tracking algorithm was used to reconstruct the whole-brain white matter fiber pathways based on the diffusion tensor imaging structural connectivity matrix, and the initial structural connectivity strength of the brain region was calculated based on the spatial density distribution of the fiber bundles. The initial structural connection matrix is normalized by graph theory measurement to obtain the structural connection matrix, which is expressed by formula (4). Formula (4) is as follows: ; In formula (4), For brain area and brain areas The initial structural connection strength between is the average anisotropy fraction of the fiber bundle, is the number of connected fibers, is the average fiber length, For brain area and brain areas The distance between the centroids, is the Gaussian kernel width; The structural connection matrix is expressed by formula (5), which is as follows: ; In formula (5), is the structural connection matrix; Perform diffusion geometric mapping on the structural connection matrix to obtain a low-dimensional manifold space; The spectral clustering algorithm is used to identify features in the low-dimensional manifold space to obtain modular features, which include the intra-module connection enhancement coefficient and the inter-module connection attenuation coefficient. The structural connection matrix is weighted according to the modular characteristics, and the intra-module connection enhancement coefficient and inter-module connection attenuation coefficient are used for dual parameter adjustment to obtain the structural connection weight matrix.

[0047] In this embodiment, when the diffusion tensor imaging structural connectivity matrix is processed by a probabilistic fiber tracking algorithm, a multi-fiber orientation distribution model is used to overcome the analytical difficulty of crossing fibers. The calculation of the spatial density distribution of fiber bundles is achieved by counting the number of streamlines passing through each pair of brain regions, while taking into account the influence of partial volume effects. and brain areas Initial structural connection strength between The quantification of the microstructure integrity, connection density and conduction efficiency is integrated into three biophysical indicators, among which the Gaussian kernel term Gaussian kernel width is used to correct the bias of the distance between brain region centroids on the connection strength estimation Automatically calibrated based on the average connection distance of the whole brain.

[0048] Structural connectivity matrix Normalization is achieved through diffusion geometry mapping, a process that projects high-dimensional connectivity patterns onto a low-dimensional manifold space, preserving key features of the topological structure. Low-dimensional embedding employs the Laplacian eigenmap algorithm, which solves the spectral decomposition problem of the connectivity matrix and represents brain regions as low-dimensional coordinate points that preserve connectivity similarity. This effectively suppresses noise interference while highlighting the organizational principles of the network.

[0049] Preferably, when the spectral clustering algorithm is executed in a low-dimensional manifold space, the intra-module connection enhancement coefficient is determined by calculating the Z-score normalized value of the connection strength within the same cluster, reflecting the density of local connections; the inter-module connection attenuation coefficient is derived based on the edge betweenness centrality of cross-module connections, characterizing the sparse trend of long-range connections; in the dual-parameter adjustment process, the intra-module connection enhancement coefficient is nonlinearly amplified using an S-type function, and the inter-module connection attenuation coefficient is suppressed according to an inverse proportional relationship. The final generated structural connection weight matrix not only retains the original anatomical constraints, but also strengthens the modular characteristics of the network.

[0050] This embodiment improves the reconstruction accuracy of complex fiber pathways through a probabilistic fiber tracking algorithm; diffusion geometry mapping overcomes the sparsity problem of high-dimensional connectivity data; and a dual-parameter adjustment mechanism achieves adaptive enhancement of network topology, ensuring that the structural connection weight matrix reflects the true anatomical connectivity pattern while highlighting functionally relevant topological features. The reconstruction process adheres to the biological characteristics of white matter fibers, and the extraction of modular features provides a structural constraint foundation for subsequent multimodal fusion.

[0051] See also Figure 2 In some embodiments, a probabilistic fiber tracking algorithm is used to reconstruct the whole-brain white matter fiber pathways based on the diffusion tensor imaging structural connectivity matrix, and the initial structural connectivity strength between brain regions is calculated based on the spatial density distribution of the fiber bundles, including: S201, simulating white matter fiber propagation paths by Monte Carlo random walks in an anisotropic field defined by a diffusion tensor imaging structural connectivity matrix, wherein each simulated white matter fiber propagation path is probabilistically sampled based on the direction of a local diffusion tensor principal eigenvector to obtain multiple fiber bundle paths; S202, spatially clustering the fiber bundle paths generated by repeated sampling, retaining the connected paths passing through the preset white matter skeleton region, and recording them as initial fiber bundle paths; S203: Count the cross-distribution density of the initial fiber bundle paths between the various brain regions, perform geometric correction based on the length of the initial fiber bundle paths, and generate an initial structural connection matrix with directional weights.

[0052] In step S201, the anisotropy field is preferably constructed based on the fractional anisotropy map from diffusion tensor imaging. Monte Carlo random walk simulation generates candidate paths by repeatedly sampling the directions of the principal eigenvectors of the local diffusion tensor, with each simulation recording the spatial coordinate sequence of the paths. This process preferably employs ten thousand iterations to ensure coverage of intersecting fibers, while GPU parallel acceleration is used to improve computational efficiency.

[0053] In step S202, the screening of the initial fiber bundle path can be achieved through anatomical constraints: the preset white matter skeleton area is delineated with reference to the probability template of the JHU white matter atlas, and the spatial clustering uses the density peak algorithm to automatically identify the main pathways. The retained connected paths must meet the spatial angle of no more than 15 degrees with the known fiber bundle direction, which not only eliminates noise paths but also ensures that the reconstruction results conform to the laws of neuroanatomy.

[0054] In step S203, during the generation of the initial structural connection matrix, preferably, the calculation of the cross-distribution density adopts the kernel density estimation method, and the length correction factor is dynamically adjusted based on the ratio of the fiber bundle travel distance to the Euclidean distance of the brain area center of mass; the introduction of the direction weight is achieved by analyzing the angle between the initial fiber bundle path and the normal vector of the brain area surface, and the weight distribution of the forward and reverse connections follows the biological law of axon projection.

[0055] This example reconstructs whole-brain white matter fiber pathways using a probabilistic fiber tracking algorithm. It simulates fiber propagation paths based on Monte Carlo random walks and, incorporating spatial constraints from the white matter skeleton region, screens anatomically sound initial fiber bundle paths. Ultimately, an initial structural connectivity matrix with directional weights is generated through fiber bundle crossing density and length correction. This example improves reconstruction accuracy in complex fiber crossing regions through a probabilistic sampling mechanism. Spatial clustering and the anatomical constraints of a pre-set white matter skeleton ensure the biological plausibility of fiber pathways. The introduction of directional weights reflects the projection characteristics of neural conduction, providing input features with both spatial accuracy and physiological interpretability for the subsequent generation of structural connectivity weight matrices, thereby supporting the precise quantification of white matter structural connectivity in multimodal brain network analysis.

[0056] In some embodiments, a spectral clustering algorithm is applied to the low-dimensional manifold space to perform feature recognition to obtain modular features, including: Calculate multiple eigenvectors based on the Laplacian matrix of the low-dimensional manifold space and construct the eigenvector space; The k-means clustering algorithm was used to partition the feature vector space and obtain the modular partitioning of the brain network; Calculate the membership of each brain region in the module of the modular partitioning of the brain network and generate a modular feature matrix; The structural connection matrix is weighted according to the modular characteristics, and the intra-module connection enhancement coefficient and inter-module connection attenuation coefficient are used for dual parameter adjustment to obtain the structural connection weight matrix, including: Apply an enhancement coefficient to the brain area connections belonging to the same module to improve the connection strength within the module, which is recorded as the intra-module connection enhancement coefficient; Apply an attenuation coefficient to the cross-module brain area connection to suppress the inter-module connection strength, which is recorded as the inter-module connection attenuation coefficient; Repeat the above steps until all brain regions are traversed; The connection enhancement coefficients within multiple modules and the connection attenuation coefficients between modules are normalized to generate a structural connection weight matrix with modular topological constraints.

[0057] In this embodiment, the Laplacian matrix of the low-dimensional manifold space is constructed by introducing an adaptive kernel function, which dynamically adjusts the similarity measure according to the functional connectivity between brain regions, so that the feature vector space can more sensitively capture the functional reorganization characteristics of the brain network.

[0058] Preferably, the initial cluster centers of the k-means clustering algorithm are automatically determined through spectral gap analysis, which avoids the subjectivity of manually setting the number of modules. The generated modular partitioning not only reflects the static functional division, but also can capture the stable modular structure during the dynamic reorganization of the brain network through iterative optimization.

[0059] During the generation of the modular feature matrix, membership calculations utilize a probabilistic model based on connection density, comprehensively considering the strength and spatial distribution of connections between brain regions and core nodes of the module. The intra-module connection enhancement coefficient is determined by analyzing the statistical distribution of intra-module connection strengths, selectively enhancing strong connections using a Gaussian kernel function. The inter-module attenuation coefficient is dynamically adjusted based on the topological distance of cross-module connections, imposing a stronger attenuation effect on long-range cross-module connections. Normalization uses the median of inter-module connection strengths as a baseline reference value to ensure that the adjusted connection weights maintain a reasonable dynamic range.

[0060] This example combines modular feature extraction with dynamic network analysis. Through an adaptive parameter selection mechanism, the structural connection weight matrix reflects the modular reorganization patterns of brain networks under different states. This not only strengthens the topological characteristics of functional modules but also preserves network plasticity information through dynamic parameter adjustment, providing a more refined analytical tool for studying disease-related brain network reorganization mechanisms. The resulting structural connection weight matrix not only conforms to the biological organizational principles of neural networks but also adapts to individual differences and state changes, providing more explanatory connection features for subsequent multimodal fusion.

[0061] In some embodiments, a four-dimensional correlation tensor is constructed by combining the time-varying brain network feature matrix, the structural connection weight matrix, and the T1-weighted anatomical features through a neurodynamic model, including: Construct a multimodal feature fusion framework to map the dynamic phase synchronization pattern of the time-varying brain network feature matrix to the energy transfer equation of the neural mass unit; A white matter conduction delay constraint is imposed on the structural connectivity weight matrix, which is coupled to the spatial propagation term of the neural mass unit via the diffusion-reaction equation; Extracting the cortical thickness and gray matter density distribution from T1-weighted anatomical features as the local excitability regulation parameters of neural mass units; By integrating the energy transfer equation, spatial propagation term and excitability regulation parameters through tensor product operation, a four-dimensional correlation tensor with four-dimensional characteristics of space-time-connectivity-morphology is generated. The spatial dimension is configured as the spatial topological arrangement of neural mass units, the temporal dimension is configured as the oscillation period of dynamic phase synchronization, the connection strength dimension is configured as the synaptic weight of white matter conduction, and the gray matter morphology dimension is configured as the gradient change of cortical thickness.

[0062] In this example, the dynamic phase synchronization patterns of the time-varying brain network feature matrix were extracted using the Hilbert transform, and its energy transfer equation was constructed using a modified Kuramoto model. This model introduces a nonlinear phase-coupled term to characterize the cross-band interactions of neural oscillations at different frequency bands. The white matter conduction delay constraint of the structural connectivity weight matrix is estimated based on fiber bundle length and conduction velocity from DTI imaging. In the diffusion-reaction equation, this operator is expressed as a time delay operator for the spatial propagation term. This operator simulates the attenuation effect of the signal in long-range white matter pathways using an exponential decay function.

[0063] Cortical thickness and gray matter density distributions of T1-weighted anatomical features were obtained through surface-based morphological analysis. Cortical thickness gradients were calculated using the Laplacian operator to modulate the local excitability threshold of neural mass units. During the construction of a four-dimensional correlation tensor, a tensor product operation using the Kronecker product was employed to couple multidimensional features. Preferably, the spatial topology followed the brain region divisions of the Desikan-Killiany atlas. The oscillation period of dynamic phase synchronization was determined by extracting time-frequency features using wavelet transforms. The synaptic weight dimension was normalized using the modularity of the structural connectivity matrix.

[0064] This embodiment organically integrates multimodal features through a neurodynamic model, in which the energy transfer equation captures the dynamic characteristics of the functional network, the diffusion-reaction equation reflects the conduction constraints of the structural connection, and the local excitability parameter characterizes the influence of anatomical morphology. The four-dimensional correlation tensor not only retains the original characteristics of each mode, but also reveals the nonlinear interactions between them through tensor operations, providing a unified computational framework for studying the synergistic mechanism of brain function, structure, and morphology. This method elevates the separated multimodal analysis to a unified four-dimensional representation, which can more comprehensively reflect the multi-scale organizational principles of the brain.

[0065] In some embodiments, the dynamic phase synchronization pattern of the time-varying brain network feature matrix is mapped to the energy transfer equation of the neural mass unit, which is expressed by formula (6), which is as follows: ; In formula (6), is the total number of brain regions, Indicates the The energy state of each neural mass unit, represents the normalized functional connection weight of the time-varying brain network feature matrix, Indicates the The oscillation phase of the neural mass unit, Indicates the The oscillation phase of the neural mass unit, is the energy decay coefficient, which is related to the neurotransmitter metabolism rate. is the external input stimulus intensity; It is coupled to the spatial propagation term of the neural mass unit through the diffusion-reaction equation and expressed by formula (7), which is as follows: ; In formula (7), is the neuronal membrane time constant, For the The action potential of a neural mass unit, represents the structural connection weight in the structural connection matrix, is the delay time variable, Indicates the white matter conduction delay time between brain regions and the average fiber length Positive correlation, is the Sigmoid activation function, is the Gaussian noise term; The energy transfer equation, spatial propagation term, and excitability regulation parameter are integrated through tensor product operation to generate a four-dimensional correlation tensor with the four-dimensional characteristics of space-time-connectivity-morphology, which is expressed by formula (8). Formula (8) is as follows: ; In formula (8), represents a four-dimensional correlation tensor, represents a neural mass unit, Representation and Neural Mass Unit The target neural mass unit of the structural connection, represents the energy state time series, represents the discretization level of white matter conduction delay time, Indicates the discretization level The representative value of represents the tensor product operation, Indicates the The normalized value is the product of gray matter density and cortical thickness in each brain region.

[0066] In this embodiment, the dynamic phase synchronization pattern of the time-varying brain network feature matrix is extracted through Hilbert transform to extract the instantaneous phase information, wherein the sinusoidal function term of the phase difference represents the dynamic coupling relationship between brain regions; the energy state of the neural mass unit Reflects the activation level of the local neural cluster, and its rate of change is affected by the functional connection weight Modulation, the weight calculates the time-varying characteristics of dynamic functional connections through a sliding time window; energy decay coefficient It is related to the metabolic process of GABAergic neurotransmitters, and its value can be determined by fitting population-level neuroelectrophysiological data.

[0067] Neuronal membrane time constant in the diffusion-reaction equation Determined by the electrophysiological properties of cells and related to the time scale of synaptic integration of neuronal populations; structural connection weight The conduction delay time is obtained by a deterministic algorithm of white matter fiber tracking. Fiber bundle length The conversion relationship is established based on a piecewise linear model of the degree of myelination; the sigmoid activation function simulates the input-output nonlinear characteristics of the neuronal population, and its threshold parameters are calibrated according to the local field potential recording data.

[0068] Four-dimensional correlation tensor During the construction process, the energy state time series The white matter conduction delay time is first divided into discrete levels according to the whole brain histogram by sliding time window sampling. , each discretization level Corresponding to a representative value , Use representative delay values for each level Calculation: Normalized value of the product of gray matter density and cortical thickness The voxel data is mapped to the standard brain space through a surface registration algorithm and the parameter is calculated to reflect the structural integrity of the local microcircuit; the tensor product operation uses high-order singular value decomposition to achieve dimensionality reduction and fusion of the feature space.

[0069] This example integrates time-varying functional connectivity, white matter structural networks, and anatomical features into a unified four-dimensional representation by establishing a coupling mechanism between energy dynamics and structural conduction. Equation (6) describes the dynamic synchronization characteristics of the functional network, Equation (7) characterizes the signal propagation laws under structural constraints, and Equation (8) implements tensor fusion of multimodal features. This method, through rigorous mathematical formalization, provides a computable modeling framework for studying multiscale interactions in brain networks, achieving an organic unification of dynamic functional activity and static structural constraints.

[0070] In some embodiments, a time-varying graph neural network model is used to perform multi-task joint learning on a four-dimensional correlation tensor, outputting quantitative diagnostic results including disease classification probability, neural circuit abnormality markers, and cognitive function decline trajectory, including: Constructing biophysically constrained coupled oscillator modules in time-varying graphical neural network models, including: The characteristics contained in the time dimension of the four-dimensional correlation tensor are mapped into the oscillator phase dynamics equation. The natural frequency of each oscillator unit is determined by the resting-state EEG characteristic spectrum, and the coupling strength is dynamically adjusted by the connection strength dimension in the four-dimensional correlation tensor. The oscillator phase dynamics equation is expressed by formula (9), which is as follows: ; In formula (9), Indicates the The instantaneous phase of an oscillator unit, Indicates the The natural frequency of an oscillator unit, Representation and Node There is a set of neighboring nodes with white matter connections, represents a slice of the connection strength dimension of the four-dimensional correlation tensor, represents the phase coupling function, is the Gaussian white noise term; A phase-locked detection algorithm is used to calculate the instantaneous phase synchronization matrix of the oscillator cluster, which is used as the dynamic connection feature input of the graph neural network. Spatiotemporal graph convolution based on anatomical structure constraints processes dynamic connection features. Spatiotemporal graph convolution includes spatial dimension convolution processing and temporal dimension convolution processing. Spatial dimension convolution processing includes spatial graph convolution. The adjacency matrix of spatial graph convolution is limited to the range of white matter connection paths determined by DTI fiber tracking. Temporal dimension convolution processing includes deformable convolution kernel to capture multi-scale features from event-related potentials to functional state fluctuations. Synchronously update node features and connection weights in the convolution operation of each layer of spatiotemporal graph convolution until the processing is completed; Generate quantitative diagnostic results through a multi-task attention mechanism, including: The multi-task attention mechanism includes disease classification branch, neural circuit abnormality monitoring branch, and cognitive prediction branch; The disease classification branch calculates the contribution weight of each brain region to the classification decision through dynamic graph pooling, and obtains the first quantitative result containing the first attention weight; The neural circuit anomaly detection branch uses the edge attention score to identify abnormal pathways where the functional connectivity strength and structural connectivity density deviate significantly, and obtains a second quantitative result containing a second attention weight. The cognitive prediction branch uses the temporal attention algorithm to calculate the fluctuation period and obtains the third quantitative result containing the third attention weight; A quantitative diagnosis result is generated according to the first quantitative result, the second quantitative result, and the third quantitative result.

[0071] In this embodiment, the construction of the time-varying graphical neural network model organically combines the multi-scale dynamic characteristics of neural activity with anatomical structure constraints. The time dimension characteristics of the four-dimensional correlation tensor are mapped in a biophysically interpretable manner through the oscillator phase dynamics equation, where the natural frequency of each oscillator unit is determined by the resting-state EEG characteristic spectrum. Specifically, the power spectral density of the preprocessed EEG signal is estimated, the dominant frequency of each frequency band is extracted as the baseline oscillation frequency of the corresponding anatomical brain area, and then the macroscopic EEG characteristics are assigned to the microscopic oscillator unit through the cortical surface registration technology. The selection of the phase coupling function must meet the periodic boundary conditions, and a modified sine function containing an exponential decay term is preferably used to simulate the nonlinear coupling characteristics between neural clusters.

[0072] The implementation of the phase-lock detection algorithm must consider the non-stationary nature of neural oscillations. A sliding time window is used to calculate the instantaneous phase difference after Hilbert transformation, with the window length adaptively adjusted based on the target frequency band. When the resulting instantaneous phase synchronization matrix is used as the input for dynamic connectivity features, topological filtering of white matter connectivity pathways is performed. Specifically, only phase synchronization relationships within the anatomical connectivity range determined by DTI fiber tracking are retained. This hard constraint effectively eliminates spurious connections caused by volume conduction effects. Spatial dimension manipulation in spatiotemporal graph convolution strictly adheres to anatomical adjacency. The positions of nonzero elements in the adjacency matrix are determined by deterministic fiber tracking results, and weight initialization values are derived from the fractional anisotropy index of diffusion tensor imaging.

[0073] Deformable convolution kernels operate on a multi-scale basis in the temporal dimension, and their receptive fields can be dynamically adjusted based on the local temporal characteristics of the input signal. For transient responses such as event-related potentials, the convolution kernel automatically contracts to capture changes; for slowly changing processes such as functional state fluctuations, the convolution kernel expands the time scale. This adaptive mechanism is implemented through a learnable position offset, the adjustment range of which is limited by physiological conduction delays. The synchronous update mechanism of node features and connection weights is reflected in the following: after each layer of convolution operation, not only is the state representation of the oscillator unit updated, but the structure-function coupling coefficient is also corrected based on the gradient information output by the current layer. This two-way optimization enables the model to adapt to changes in neural plasticity.

[0074] In the multi-task attention mechanism, the dynamic graph pooling operation of the disease classification branch essentially simulates the brain region importance assessment process in clinical diagnosis. The first attention weight reflects the contribution of abnormal activity in a specific brain region to the discrimination of disease classification; the edge attention scoring mechanism adopted by the neural circuit abnormality detection branch calculates the Mahalanobis distance between functional connection strength and structural connection density to identify abnormal neural pathways whose functional activities exceed the anatomical connection constraints. Such pathways are often related to pathological compensation or loss of connectivity; the temporal attention algorithm of the cognitive prediction branch focuses on analyzing the periodic characteristics of low-frequency fluctuation signals. The third attention weight corresponds to the ranking of the importance of fluctuations at different time scales to cognitive function prediction, with special attention paid to slow oscillation components in the range of 0.01-0.1Hz.

[0075] The process of generating quantitative diagnostic results is essentially a fusion decision of multimodal biomarkers. Preferably, the disease probability output of the first quantitative result is processed using softmax normalization, the abnormal markers of the second quantitative result are represented by Z scores to represent the degree of deviation from the normal range, and the cognitive trajectory prediction of the third quantitative result is achieved through time series extrapolation. The final integration of quantitative diagnostic results requires consideration of the confidence weighting of each branch. The disease classification branch is typically weighted 1.5-2 times that of other branches to align with the classification-centric decision-making characteristics of clinical diagnosis.

[0076] This embodiment maps the dynamic characteristics of the four-dimensional correlation tensor into a phase dynamic process with neurophysiological significance by constructing a coupled oscillator module with biophysical constraints. Combined with the spatiotemporal graph convolution processing constrained by anatomical structure, it achieves an organic fusion of the multi-scale dynamic characteristics of neural activity and the topology of white matter connections. Its beneficial effects are: through the dual protection of the oscillator phase synchronization mechanism and the hard constraints of anatomical connections, the interpretability of functional connectivity features is effectively improved; the use of deformable convolution kernels to adaptively capture cross-scale features from event-related potentials to functional state fluctuations enhances the physiological rationality of temporal dynamic modeling; the multi-task attention mechanism generates quantitative diagnostic results based on a strict structure-function coupling relationship, so that disease classification, abnormal loop detection and cognitive prediction form a mutually verified closed-loop system. The final output diagnostic results have both biophysical interpretability and clinical decision support value.

[0077] In some embodiments, multimodal brain imaging information and cognitive behavioral information of the user are collected, where the multimodal brain imaging information includes resting-state functional magnetic resonance imaging time series signals, diffusion tensor imaging structural connectivity matrix, and T1-weighted anatomical features, including: The resting-state functional magnetic resonance timing signals were configured to be acquired using a preset functional magnetic resonance imaging sequence; The diffusion tensor imaging structural connectivity matrix was configured to acquire data using a preset diffusion-weighted imaging protocol; T1-weighted anatomical features were configured to be acquired using a high-resolution 3D structural imaging sequence, and the brain was divided into a preset number of automatically anatomically labeled brain regions using an atlas-based segmentation method.

[0078] In this embodiment, the acquisition of multimodal brain imaging data is collaboratively characterized through standardized processes of three types of imaging technologies. Preferably, the resting-state functional magnetic resonance time series signal is acquired using a clinical routine gradient echo planar imaging sequence, and its repetition time parameters are set according to the typical low-frequency neural oscillation characteristics. The original signal is preprocessed by head motion correction and frequency band filtering to form a time series matrix; the diffusion tensor imaging structural connectivity matrix is acquired through a multi-directional diffusion weighted imaging protocol, and the original diffusion data is processed by tensor reconstruction and deterministic fiber tracking to generate a weighted adjacency matrix reflecting the density of white matter pathways; T1-weighted anatomical features are acquired using a high-resolution three-dimensional structural imaging sequence, and the gray matter volume features are mapped to a preset anatomical template space through a standardized cortical segmentation atlas to form a structural feature vector.

[0079] Cognitive behavioral assessments were timed to coincide with imaging acquisition and were conducted using a combination of standardized neuropsychological scales with validated reliability and validity. These scales covered core cognitive domains, including memory, executive function, and information processing speed. Memory was assessed using a standardized delayed recall paradigm, executive function testing included conventional measures of cognitive flexibility, and information processing speed was quantified using a basic symbolic encoding task. These behavioral measures formed a multidimensional mapping relationship with neuroimaging features.

[0080] In this example, functional magnetic resonance imaging (fMRI) time-series signals are registered to individual anatomical space using a rigid-body transformation and then mapped to a standard brain template via nonlinear normalization. Diffusion tensor imaging data are aligned across subjects using white matter skeleton extraction. Cognitive and behavioral data are converted to z-scores for normalization. This processing ensures that the structural connectivity matrix constrains the scope of functional network construction, allowing the resulting fusion features to preserve both individual anatomical characteristics and functional dynamic patterns, providing neuroanatomically interpretable input features for subsequent modeling.

[0081] See also Figure 3 In the second aspect, this embodiment further provides a brain disease classification system 1, which is applicable to the brain disease classification method described in the first aspect. The system includes a data acquisition module 11, a data processing module 12, a logic operation module 13 and a result output module 14. The data acquisition module 11 is used to collect the user's multimodal brain imaging information and cognitive behavior information. The multimodal brain imaging information includes resting-state functional magnetic resonance timing signals, diffusion tensor imaging structural connection matrix and T1-weighted anatomical features. The cognitive behavior information is obtained through a standardized neuropsychological scale; the data processing module 12 is used to perform dynamic functional connection analysis on the resting-state functional magnetic resonance timing signals to obtain a time-varying brain network feature matrix. The dynamic functional connection analysis is configured to calculate the phase synchronization based on the sliding window, and perform white matter fiber bundle topology on the diffusion tensor imaging structural connection matrix. Reconstruct to obtain the structural connection weight matrix; the logical operation module 13 is used to construct a four-dimensional correlation tensor by combining the time-varying brain network feature matrix, the structural connection weight matrix and the T1-weighted anatomical features through a neurodynamic model. The dimensions of the four-dimensional correlation tensor include spatial dimension, temporal dimension, connection strength dimension and gray matter morphology dimension; based on the time-varying graph neural network model, the four-dimensional correlation tensor is subjected to multi-task joint learning to output quantitative diagnostic results including disease classification probability, neural circuit abnormality markers and cognitive function decline trajectory. The time-varying graph neural network model is configured as a coupled oscillator embedding module with biophysical constraints; the result output module 14 is used to generate an auxiliary clinical classification report including individualized brain network remodeling targets, disease progression risk stratification and treatment response prediction based on the degree of cross-modal alignment between the quantitative diagnostic results and cognitive behavioral information.

[0082] In this embodiment, the brain disease classification system 1 implements an automated analysis process for multimodal data through a modular architecture. The data acquisition module 11 directly interfaces with medical imaging equipment to ensure standardized raw data acquisition. The data processing module 12 utilizes a parallel computing architecture to simultaneously perform dynamic functional connectivity analysis and white matter topology reconstruction. The logic operation module 13 is equipped with a dedicated tensor operation unit, supporting the efficient construction of four-dimensional correlation tensors and the distributed training of time-varying graph neural networks. The result output module 14 integrates a clinical knowledge base interface to connect quantitative diagnostic results with medical decision-making systems. Communication between these modules occurs via a standardized data bus, ensuring traceability throughout the entire process, from raw data input to clinical report generation.

[0083] The system in this embodiment uses hardware acceleration to efficiently process multimodal big data. The interpretable design of the neurodynamic model ensures the reliability of clinical decision-making. The modular architecture enables the system to flexibly adapt to the imaging equipment and diagnostic needs of different medical institutions. The resulting auxiliary clinical classification report not only includes disease classification results but also provides personalized treatment recommendations based on neurobiological evidence, enabling the rapid translation of scientific discoveries into clinical applications.

[0084] By employing the above technical solutions, the present invention distinguishes itself from existing technologies and offers the following beneficial effects: Through the collaborative analysis of multimodal brain imaging and cognitive-behavioral information, it achieves high-precision and interpretable brain disease classification. Dynamic functional connectivity analysis of resting-state fMRI time-series signals captures the time-varying characteristics of brain networks, white matter fiber tract topology reconstruction from the diffusion tensor imaging structural connectivity matrix provides anatomical constraints, and T1-weighted anatomical features complement morphological information. These three elements, through a four-dimensional correlation tensor constructed using a neurodynamic model, organically integrate spatiotemporal, connectivity, and morphological features. A time-varying graph neural network model, based on a biophysically constrained coupled oscillator embedding module, transforms multimodal features into physiologically meaningful quantitative diagnostic results, including disease classification probability, neural circuit abnormality markers, and cognitive decline trajectories. The resulting auxiliary clinical classification report, through cross-modality alignment, provides personalized brain network remodeling targets, disease progression risk stratification, and treatment response prediction, providing a systematic assessment basis for clinical decision-making, from microscopic neural circuit abnormalities to macroscopic cognitive and behavioral changes. This invention overcomes the limitations of single modality analysis. Through multi-scale feature fusion and biophysical constraint modeling, it enhances the neurobiological explanatory power of diagnostic results while maintaining clinical practicality. It is particularly suitable for the auxiliary diagnosis and classification of neurodegenerative diseases such as Alzheimer's disease and Parkinson's disease.

[0085] Finally, it should be noted that although the above embodiments have been described in the specification and drawings of this application, this does not limit the scope of patent protection of this application. All technical solutions generated by replacing or modifying equivalent structures or equivalent processes based on the essential concepts of this application using the contents recorded in the specification and drawings of this application, as well as directly or indirectly implementing the technical solutions of the above embodiments in other related technical fields, are included in the scope of patent protection of this application.

Claims

1. A method for classifying brain diseases, characterized in that: include: Collecting multimodal brain imaging information and cognitive behavioral information of the user, wherein the multimodal brain imaging information includes resting-state functional magnetic resonance imaging time series signals, diffusion tensor imaging structural connectivity matrix, and T1-weighted anatomical features, and the cognitive behavioral information is obtained through standardized neuropsychological scales; performing dynamic functional connectivity analysis on the resting-state functional magnetic resonance time series signal to obtain a time-varying brain network feature matrix, wherein the dynamic functional connectivity analysis is configured to calculate phase synchrony based on a sliding window, and performing white matter fiber bundle topology reconstruction on the diffusion tensor imaging structural connectivity matrix to obtain a structural connectivity weight matrix; The time-varying brain network feature matrix, the structural connection weight matrix and the T1-weighted anatomical features are used to construct a four-dimensional correlation tensor through a neurodynamic model, wherein the dimensions of the four-dimensional correlation tensor include a spatial dimension, a temporal dimension, a connection strength dimension and a gray matter morphology dimension; Performing multi-task joint learning on the four-dimensional correlation tensor based on a time-varying graph neural network model to output a quantitative diagnostic result including disease classification probability, neural circuit abnormality markers, and cognitive function decline trajectory, wherein the time-varying graph neural network model is configured as a coupled oscillator embedding module with biophysical constraints; An auxiliary clinical classification report containing personalized brain network remodeling targets, disease progression risk stratification, and treatment response prediction is generated based on the degree of cross-modal alignment between the quantitative diagnostic results and cognitive behavioral information.

2. The brain disease classification method according to claim 1, characterized in that: Performing dynamic functional connectivity analysis on the resting-state functional magnetic resonance time series signal to obtain a time-varying brain network feature matrix, wherein the dynamic functional connectivity analysis is configured to calculate phase synchrony based on a sliding window, including: performing a sliding window analysis on the resting-state functional magnetic resonance time series signal to obtain a plurality of sliding windows, wherein the sliding window analysis is configured to be a sliding window algorithm operation based on an adaptive window length, and the window length of the sliding window is dynamically adjusted according to the signal-to-noise ratio of the signals of the plurality of brain regions; The phase synchronization strength of the brain region corresponding to each sliding window is calculated and expressed by formula (1), which is as follows: ； In formula (1), for Momentary brain area and brain areas The strength of phase synchronization between is the total number of time points in a single sliding window, For brain area for The instantaneous phase at a time point, For brain area for The instantaneous phase at a time point, Computational brain area and brain areas The complex exponential function of the phase difference vector between To count letters, ; Repeat the above steps until the phase synchronization strength of all brain regions is calculated, and construct a two-dimensional dynamic functional connectivity matrix based on the phase synchronization strength, which is expressed by formula (2). Formula (2) is as follows: ； In formula (2), is the dynamic functional connectivity matrix; The dynamic functional connectivity matrix is reduced in dimension by non-negative matrix decomposition, and the time-varying brain network feature matrix is extracted and expressed by formula (3). Formula (3) is as follows: ； In formula (3), is the static network component after non-negative matrix decomposition, is the time-varying brain network feature matrix.

3. The brain disease classification method according to claim 1, characterized in that: The diffusion tensor imaging structural connectivity matrix is subjected to white matter fiber bundle topology reconstruction to obtain a structural connectivity weight matrix, including: The diffusion tensor imaging structural connectivity matrix is used to reconstruct the whole-brain white matter fiber pathway using a probabilistic fiber tracking algorithm, and the initial structural connectivity strength of the brain region is calculated based on the spatial density distribution of the fiber bundles; The initial structural connection matrix is normalized by graph theory measurement to obtain the structural connection matrix, which is expressed by formula (4). The formula (4) is as follows: ； In formula (4), For brain area and brain areas The initial structural connection strength between is the average anisotropy fraction of the fiber bundle, is the number of connected fibers, is the average fiber length, For brain area and brain areas The distance between the centroids, is the Gaussian kernel width; The structural connection matrix is expressed by formula (5), which is as follows: ； In formula (5), is the structural connection matrix; Performing diffusion geometric mapping on the structural connection matrix to obtain a low-dimensional manifold space; Performing feature recognition on the low-dimensional manifold space using a spectral clustering algorithm to obtain modular features, wherein the modular features include an intra-module connection enhancement coefficient and an inter-module connection attenuation coefficient; The structural connection matrix is weighted according to the modular characteristics, and a dual-parameter adjustment is performed using an intra-module connection enhancement coefficient and an inter-module connection attenuation coefficient to obtain a structural connection weight matrix.

4. The brain disease classification method according to claim 3, characterized in that: The diffusion tensor imaging structural connectivity matrix is used to reconstruct the whole-brain white matter fiber pathway using a probabilistic fiber tracking algorithm, and the initial structural connectivity strength between brain regions is calculated based on the spatial density distribution of the fiber bundles, including: In the anisotropic field defined by the diffusion tensor imaging structural connectivity matrix, the white matter fiber propagation paths were simulated by Monte Carlo random walks. Each simulated white matter fiber propagation path was probabilistically sampled based on the direction of the local diffusion tensor principal eigenvector, resulting in multiple fiber bundle paths. The fiber bundle paths generated by repeated sampling are spatially clustered, and the connected paths passing through the preset white matter skeleton area are retained and recorded as the initial fiber bundle paths; The cross-distribution density of the initial fiber bundle paths between various brain regions was counted, and geometric correction was performed based on the length of the initial fiber bundle paths to generate an initial structural connection matrix with directional weights.

5. The brain disease classification method according to claim 3, characterized in that: The low-dimensional manifold space is subjected to a spectral clustering algorithm for feature recognition to obtain modular features, including: Calculating a plurality of eigenvectors based on the Laplacian matrix of the low-dimensional manifold space to construct an eigenvector space; The k-means clustering algorithm is used to partition the feature vector space to obtain a modular partition of the brain network; Calculate the membership of each brain region in the module of the modular partitioning of the brain network and generate a modular feature matrix; The structural connection matrix is weighted according to the modular characteristics, and dual-parameter adjustment is performed using the intra-module connection enhancement coefficient and the inter-module connection attenuation coefficient to obtain a structural connection weight matrix, including: Apply an enhancement coefficient to the brain area connections belonging to the same module to improve the connection strength within the module, which is recorded as the intra-module connection enhancement coefficient; Apply an attenuation coefficient to the cross-module brain area connection to suppress the inter-module connection strength, which is recorded as the inter-module connection attenuation coefficient; Repeat the above steps until all brain regions are traversed; The connection enhancement coefficients within multiple modules and the connection attenuation coefficients between modules are normalized to generate a structural connection weight matrix with modular topological constraints.

6. The brain disease classification method according to claim 1, characterized in that: The time-varying brain network feature matrix, the structural connection weight matrix and the T1-weighted anatomical features are used to construct a four-dimensional correlation tensor through a neural dynamics model, including: Constructing a multimodal feature fusion framework to map the dynamic phase synchronization pattern of the time-varying brain network feature matrix to the energy transfer equation of the neural mass unit; imposing a white matter conduction delay constraint on the structural connection weight matrix, coupling it to the spatial propagation term of the neural mass unit via a diffusion-reaction equation; extracting the cortical thickness and gray matter density distribution from the T1-weighted anatomical features as local excitability regulation parameters of the neural mass unit; The energy transfer equation, spatial propagation term and excitability regulation parameter are integrated through tensor product operation to generate a four-dimensional correlation tensor with four-dimensional characteristics of space-time-connectivity-morphology, wherein the spatial dimension is configured as the spatial topological arrangement of neural mass units, the temporal dimension is configured as the oscillation period of dynamic phase synchronization, the connection strength dimension is configured as the synaptic weight of white matter conduction, and the gray matter morphology dimension is configured as the gradient change of cortical thickness.

7. The brain disease classification method according to claim 6, characterized in that: The dynamic phase synchronization pattern of the time-varying brain network characteristic matrix is mapped to the energy transfer equation of the neural mass unit, which is expressed by formula (6). The formula (6) is as follows: ； In formula (6), is the total number of brain regions, Indicates the The energy state of each neural mass unit, represents the normalized functional connection weight of the time-varying brain network feature matrix, Indicates the The oscillation phase of the neural mass unit, Indicates the The oscillation phase of the neural mass unit, is the energy attenuation coefficient, is the external input stimulus intensity; The spatial propagation term coupled to the neural mass unit through the diffusion-reaction equation is expressed by formula (7), which is as follows: ； In formula (7), is the neuronal membrane time constant, For the The action potential of a neural mass unit, represents the structural connection weight in the structural connection matrix, is the delay time variable, represents the white matter conduction delay time between brain regions, is the Sigmoid activation function, is the Gaussian noise term; The energy transfer equation, spatial propagation term, and excitability regulation parameter are integrated through tensor product operation to generate a four-dimensional correlation tensor with four-dimensional characteristics of space-time-connectivity-morphology, which is expressed by formula (8). Formula (8) is as follows: ； In formula (8), represents a four-dimensional correlation tensor, represents a neural mass unit, Representation and Neural Mass Unit The target neural mass unit of the structural connection, represents the energy state time series, represents the discretization level of white matter conduction delay time, Indicates the discretization level The representative value of represents the tensor product operation, Indicates the The normalized value is the product of gray matter density and cortical thickness in each brain region.

8. The brain disease classification method according to claim 1, characterized in that: The four-dimensional correlation tensor is subjected to multi-task joint learning based on a time-varying graph neural network model, and quantitative diagnostic results including disease classification probability, neural circuit abnormality markers, and cognitive function decline trajectory are output, including: A biophysical constrained coupled oscillator module is constructed in the time-varying graphical neural network model, including: Mapping the features contained in the time dimension of the four-dimensional correlation tensor into an oscillator phase dynamics equation, wherein the natural frequency of each oscillator unit is determined by the resting-state EEG characteristic spectrum, and the coupling strength is dynamically adjusted by the connection strength dimension in the four-dimensional correlation tensor; A phase-locked detection algorithm is used to calculate the instantaneous phase synchronization matrix of the oscillator cluster, which is used as the dynamic connection feature input of the graph neural network. The dynamic connection features are processed by spatiotemporal graph convolution based on anatomical structure constraints, the spatiotemporal graph convolution includes spatial dimension convolution processing and temporal dimension convolution processing, the spatial dimension convolution processing includes spatial graph convolution, the adjacency matrix of the spatial graph convolution is limited to the range of white matter connection paths determined by DTI fiber tracking, and the temporal dimension convolution processing includes a deformable convolution kernel to capture multi-scale features from event-related potentials to functional state fluctuations; Synchronously update node features and connection weights in the convolution operation of each layer of spatiotemporal graph convolution until the processing is completed; Generate quantitative diagnostic results through a multi-task attention mechanism, including: The multi-task attention mechanism includes a disease classification branch, a neural circuit abnormality monitoring branch, and a cognitive prediction branch; The disease classification branch calculates the contribution weight of each brain region to the classification decision through dynamic graph pooling to obtain a first quantitative result containing a first attention weight; The neural circuit anomaly detection branch uses the edge attention score to identify abnormal pathways where the functional connectivity strength and structural connectivity density deviate significantly, and obtains a second quantitative result containing a second attention weight. The cognitive prediction branch uses the temporal attention algorithm to calculate the fluctuation period and obtains the third quantitative result containing the third attention weight; The quantitative diagnosis result is generated according to the first quantification result, the second quantification result, and the third quantification result.

9. The brain disease classification method according to claim 1, characterized in that: Collect multimodal brain imaging information and cognitive behavior information of the user, wherein the multimodal brain imaging information includes resting-state functional magnetic resonance time series signals, diffusion tensor imaging structural connectivity matrix and T1-weighted anatomical features, including: The resting-state functional magnetic resonance time series signal is configured to be acquired using a preset functional magnetic resonance imaging sequence; The diffusion tensor imaging structural connectivity matrix is configured to acquire data via a preset diffusion weighted imaging protocol; The T1-weighted anatomical features are configured to be acquired using a high-resolution three-dimensional structural imaging sequence, and the brain is divided into a preset number of automatically anatomically labeled brain regions using an atlas-based segmentation method.

10. A brain disease classification system, characterized in that: The method for classifying brain diseases according to any one of claims 1 to 9, wherein the system comprises: A data acquisition module, configured to collect multimodal brain imaging information and cognitive behavioral information of the user, wherein the multimodal brain imaging information includes resting-state functional magnetic resonance imaging time series signals, diffusion tensor imaging structural connectivity matrix, and T1-weighted anatomical features, and the cognitive behavioral information is obtained through standardized neuropsychological scales; a data processing module, configured to perform dynamic functional connectivity analysis on the resting-state functional magnetic resonance time series signal to obtain a time-varying brain network feature matrix, wherein the dynamic functional connectivity analysis is configured to calculate phase synchrony based on a sliding window, and to perform white matter fiber bundle topology reconstruction on the diffusion tensor imaging structural connectivity matrix to obtain a structural connectivity weight matrix; a logic operation module for constructing a four-dimensional correlation tensor by combining the time-varying brain network feature matrix, the structural connection weight matrix, and the T1-weighted anatomical features through a neurodynamic model, wherein the dimensions of the four-dimensional correlation tensor include spatial dimension, temporal dimension, connection strength dimension, and gray matter morphology dimension; performing multi-task joint learning on the four-dimensional correlation tensor based on a time-varying graph neural network model, and outputting quantitative diagnostic results including disease classification probability, neural circuit abnormality markers, and cognitive function decline trajectory, wherein the time-varying graph neural network model is configured as a coupled oscillator embedding module with biophysical constraints; A result output module is used to generate an auxiliary clinical classification report containing individualized brain network remodeling targets, disease progression risk stratification and treatment response prediction based on the degree of cross-modal alignment between the quantitative diagnostic results and cognitive behavioral information.

Images