Brain state analysis method and device based on fiber bundle guide graph auto-encoder

Abstract

The invention provides a brain state analysis method and device based on a fiber bundle guide graph auto-encoder, and the method comprises the steps: taking the dissection connection of white matter fiber bundles as a priori constraint, limiting an analysis range in a brain region connected with a specific fiber bundle in a functional connection dynamic graph construction stage, and employing a graph auto-encoder model to construct a functional connection dynamic graph. And the propagation path of the function information in the neural network is constrained by a binary matrix representing anatomical connection, so that deep fusion of the structure function information is realized in nonlinear dimensionality reduction, and finally, the fiber bundle related brain state with clear anatomical attribution and high specificity and the dynamic index thereof are identified. According to the method, the homogenization hypothesis of traditional whole brain state analysis is broken through, the macroscopic function is dynamically and accurately anchored on the microscopic anatomical pathway, and the anatomical specificity and biological interpretability of the recognized brain state are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of neuroscience technology, and in particular to a method and apparatus for brain state analysis based on fiber tractography autoencoders. Background Technology

[0002] As a highly complex and dynamic system, the brain's cognitive functions rely on the coordination and reorganization of large-scale neural clusters across time and space. With the development of neuroimaging techniques such as functional magnetic resonance imaging (fMRI), researchers can non-invasively observe whole-brain neural activity, leading to the concept of "brain states"—a dynamic framework describing the evolution of the overall activity patterns of the nervous system over time. Studying brain states and their dynamic transitions is crucial for understanding the neural basis of cognitive function and the mechanisms of nervous system diseases.

[0003] Currently, computational neuroscience has developed various techniques for analyzing brain states based on functional magnetic resonance imaging (fMRI) data. However, the brain states identified by existing methods rely entirely on the transient statistical characteristics derived from fMRI signals, failing to integrate the brain's inherent anatomical structural information, particularly the white matter fiber tract connectivity network that forms the physical basis of neural information transmission. This implicitly assumes a homogeneous approach to the entire brain, assigning equal weight to all possible functional connections within the state profile. This results in identified states that are often global, vague macroscopic patterns, potentially confounded by functional signals from different anatomical pathways, thus lacking clear anatomical attribution and specificity. This limitation diminishes the potential of these techniques in revealing the functional dynamics of specific neural circuits, locating abnormal circuits associated with neurological diseases, and providing clinical biomarkers with clear biological explanations.

[0004] Therefore, there is an urgent need for a new technological solution that can overcome the above limitations, in order to identify brain functional states with greater anatomical specificity, biological interpretability and clinical translational potential by relying on prior knowledge of the anatomical structure of the fibrous fiber bundles when analyzing the dynamic state of brain function. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a brain state analysis method and apparatus based on fiber tract guide map autoencoder to eliminate or improve one or more defects in the prior art, and overcome the problems of mixed state representation, insufficient anatomical specificity and poor biological interpretability caused by the lack of white matter fiber tract anatomical structure constraints in the existing methods.

[0006] One aspect of the present invention provides a brain state analysis method based on fiber tractograph autoencoders, the method comprising the following steps: Functional magnetic resonance imaging (fMRI) data of the brain of the subject to be analyzed is preprocessed and spatially registered with the individual brain structure to extract time series data of blood oxygen dependence levels in multiple regions of interest; the regions of interest are gray matter brain regions connected by specified white matter fiber bundles. The blood oxygen dependence level data time series of each region of interest are processed by a bandpass filter to extract the target low-frequency oscillation component and perform Hilbert transform. The instantaneous phase of the signal at each time point is calculated to obtain the instantaneous phase time series of each region of interest. At each time point, the instantaneous phase difference between all the regions of interest is calculated and the corresponding cosine value is calculated as the phase coherence value. A connection relationship is established between the regions of interest corresponding to a set proportion of phase coherence values ​​with higher values ​​to form a brain connectivity map. The brain connectivity map includes a binary adjacency matrix representing the connection relationship between the regions of interest and a node feature matrix composed of the phase coherence values. Based on a pre-trained graph autoencoder model, the node feature matrix at each time point is used as input, and the binary adjacency matrix is ​​used as a constraint to extract the low-dimensional latent representation corresponding to each region of interest. Cluster the low-dimensional latent representations of each region of interest at all time points, with each cluster representing a brain state; assign a unique state label to each time point to obtain a discrete sequence of brain states switching over time, and calculate the occurrence probability, average duration, and state transition probability of each brain state.

[0007] In some embodiments, before preprocessing the functional magnetic resonance imaging data of the brain of the subject to be analyzed, the method further includes: Multiple target white matter fiber tracts of the human brain are superimposed with a preset anatomical atlas to determine the gray matter brain regions connected by the target white matter fiber tracts as regions of interest; the target white matter fiber tracts include the bilateral anterior thalamic radiation, inferior fronto-occipital fasciculus, inferior longitudinal fasciculus, superior longitudinal fasciculus, superior temporal longitudinal fasciculus, uncinate fasciculus, cingulate hippocampus, cingulate gyrus and corticospinal tract. Preprocessing of the functional magnetic resonance imaging (fMRI) data of the brain of the subject to be analyzed includes: performing temporal correction, head motion correction, spatial normalization and smoothing, and filtering and denoising on the fMRI data.

[0008] In some embodiments, the formula for calculating the phase coherence value is: ; in, Indicates a point in time Time Interest Area Area of ​​Interest The phase coherence values ​​between, Indicates a point in time Time Interest Area Corresponding to the instantaneous phase of the aforementioned blood oxygen-dependent level data time series, Indicates a point in time Time Interest Area The instantaneous phase of the time series of blood oxygen-dependent level data.

[0009] In some embodiments, the frequency range of the bandpass filter is 0.01~0.1Hz; Establishing a connection between the regions of interest corresponding to a set proportion of phase coherence values ​​with higher values ​​includes setting the set proportion to 40-45% based on a proportional threshold method.

[0010] In some embodiments, the training steps of the graph autoencoder include: A training sample set is obtained, which includes a sample brain connectivity graph at multiple time points. The sample brain connectivity graph contains a sample binary adjacency matrix representing the connectivity between the regions of interest and a sample node feature matrix composed of the phase coherence values. An initial encoder and an initial decoder are obtained. The encoder is obtained by stacking multiple layers of graph convolutional networks or graph attention networks. The initial decoder includes fully connected layers, inner product operation layers, and activation function layers. The initial encoder and the initial decoder are trained using the training sample set. The initial encoder takes the sample brain connectivity graph as input and transmits and aggregates the information carried by the sample node feature matrix along the edges with existing connections marked by the sample binary adjacency matrix, while edges without connections are suppressed, outputting a low-dimensional latent representation of the sample. The initial decoder takes the low-dimensional latent representation of the sample and the sample binary adjacency matrix as input and outputs a reconstruction matrix of the sample node feature matrix. The parameters of the initial encoder and the initial decoder are updated based on the binary cross-entropy loss of the reconstruction matrix and the sample node feature matrix to obtain the graph autoencoder and graph autodecoder.

[0011] In some embodiments, clustering the low-dimensional latent representations of each region of interest at all time points further includes: clustering using the K-means algorithm, Gaussian mixture model, or spectral clustering algorithm, and determining the optimal number of clusters using the elbow rule or silhouette coefficient.

[0012] In some embodiments, the method further includes: using a hidden Markov model to simultaneously perform brain state clustering on the low-dimensional latent representation and model the state transition probability.

[0013] On the other hand, the present invention also provides a brain state analysis device based on fiber tract guide map autoencoder, including a processor, a memory, and a computer program or instructions stored in the memory. The processor is used to execute the computer program or instructions, and when the computer program or instructions are executed, the device implements the steps of the above method.

[0014] On the other hand, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0015] On the other hand, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0016] The brain state analysis method and apparatus based on fiber tract guided graph autoencoder described in this invention uses the anatomical connections of white matter fiber tracts as a priori constraints. During the functional connectivity dynamic graph construction stage, the analysis scope is limited to brain regions with specific fiber tract connections. A graph autoencoder model is used to characterize the propagation path of functional information constrained by the binary matrix of anatomical connections in the neural network, thereby achieving deep fusion of structural and functional information through nonlinear dimensionality reduction. Ultimately, it identifies fiber tract-related brain states and their dynamic indicators with clear anatomical attribution and high specificity. This invention breaks the homogeneity assumption of traditional whole-brain state analysis, accurately anchoring macroscopic functional dynamics to microscopic anatomical pathways, significantly improving the anatomical specificity and biological interpretability of the identified brain states. Simultaneously, by guiding and modulating the functional feature learning process through structural information, the robustness of the model is enhanced, providing an innovative and reliable analytical tool for revealing the functional dynamics of specific neural circuits and developing circuit-specific clinical biomarkers.

[0017] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0018] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings: Figure 1This is a schematic flowchart of a brain state analysis method based on fiber tract guide map autoencoder according to an embodiment of the present invention.

[0020] Figure 2 This is a schematic flowchart of a brain state analysis method based on a fiber tract guide map autoencoder, as described in another embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0022] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0023] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0024] Significant progress has been made in using functional magnetic resonance imaging (fMRI) to analyze the dynamic state of the brain. Mainstream methods focus on linear or nonlinear dimensionality reduction and clustering of dynamic functional connectivity matrices to identify macroscopic brain states (or metastable states) that change over time. However, these methods have a fundamental limitation: they rely primarily, or even entirely, on functional signals derived from fMRI (such as phase coherence), failing to effectively integrate inherent anatomical information of the brain as constraints and priors for analysis. This results in the identified "states" being essentially functional pattern clusters based on statistical similarity, lacking a clear anatomical basis, and exhibiting ambiguity in their biological interpretation, making it difficult to precisely map them to specific neural pathways. This problem manifests on two levels: firstly, purely function-driven analysis methods (even advanced nonlinear graph autoencoders) are susceptible to noise and individual differences in fMRI signals, and the high-dimensional features they learn may be mixed with information from different anatomical pathways, leading to insufficient state specificity. On the other hand, although the field of neuroimaging has developed mature white matter fiber tract structure analysis techniques (such as fiber tract reconstruction and segmentation using autoencoders) and complex brain dynamics analysis methods, the two have long been relatively separate. Existing research has failed to use robust anatomical connectivity maps as rigid constraints to systematically guide and modulate the extraction process of functional dynamic features. Therefore, how to deeply integrate multimodal information at the model level to resolve more specific and interpretable brain functional states within a clear anatomical framework is the main bottleneck and development direction of current technology.

[0025] This invention provides a brain state analysis method based on fiber tractography autoencoders, such as... Figure 1 As shown, the method includes the following steps S101~S105: Step S101: Preprocess the functional magnetic resonance imaging data of the brain of the subject to be analyzed and spatially register it with the individual brain structure to extract the time series of blood oxygen dependence level data of multiple regions of interest; the regions of interest are the gray matter brain regions connected by specified white matter fiber bundles.

[0026] Step S102: Extract the target low-frequency oscillation component from the blood oxygen-dependent level data time series of each region of interest through a bandpass filter and perform Hilbert transform. Calculate the instantaneous phase of the signal at each time point to obtain the instantaneous phase time series of each region of interest.

[0027] Step S103: Calculate the instantaneous phase difference between all regions of interest at each time point and calculate the corresponding cosine value as the phase coherence value. Establish connection relationships between regions of interest corresponding to a set proportion of phase coherence values ​​with higher values ​​to form a brain connectivity map. The brain connectivity map includes a binary adjacency matrix representing the connection relationship between regions of interest and a node feature matrix composed of phase coherence values.

[0028] Step S104: Based on the pre-trained graph autoencoder model, the low-dimensional latent representation corresponding to each region of interest is extracted using the node feature matrix at each time point as input and the binary adjacency matrix as constraint.

[0029] Step S105: Cluster the low-dimensional latent representations of each region of interest at all time points, with each cluster representing a brain state; assign a unique state label to each time point to obtain a discrete sequence of brain states switching over time, and calculate the occurrence probability, average duration, and state transition probability of each brain state.

[0030] Step S101 forms the data foundation and physical constraint starting point for the entire method. Its core objective is not traditional whole-brain data processing, but rather the construction of a structured analytical framework based on prior knowledge of white matter fiber tract anatomy. Specifically, the individual's functional magnetic resonance imaging (fMRI) data undergoes rigorous standardization preprocessing, including temporal correction, head movement correction, spatial standardization and smoothing, and filtering to eliminate non-neurophysiological noise and ensure data comparability in a standard space. After preprocessing, signals from all brain regions are not extracted; instead, targeted extraction is performed: based on the spatial superposition relationship between a pre-defined white matter fiber tract template and a standard brain region atlas (such as AAL116), gray matter brain regions with clear anatomical connections to the ends of specific fiber tracts are precisely selected and defined as regions of interest (ROIs). Target white matter fiber tracts include the bilateral anterior thalamic radii, inferior fronto-occipital fasciculus, inferior longitudinal fasciculus, superior longitudinal fasciculus, superior temporal longitudinal fasciculus, uncinate fasciculus, cingulate hippocampus, cingulate gyrus, and corticospinal tract. Subsequently, the preprocessed functional data were meticulously spatially registered with individual high-resolution structural images, and the blood oxygen dependence level (BOLD) time series of these ROIs were extracted from the registered data. This step fundamentally abandons the assumption of whole-brain homogeneity, ensuring that all subsequent analyses are constrained within specific anatomical pathways, laying the foundation for ultimately obtaining brain states with clear anatomical attribution.

[0031] Step S102 transforms the slow, indirect BOLD hemodynamic signal into a phase signal that can directly characterize the instantaneous dynamics of neural oscillations, preparing for the calculation of brain region synchronicity with millisecond-level time precision. The extracted BOLD time series contains a wide range of frequency components. First, a bandpass filter of 0.01–0.1 Hz is used to filter out high-frequency physiological noise such as heartbeat and respiration, as well as extremely low-frequency drift, retaining the low-frequency oscillation components most relevant to spontaneous neural activity. Subsequently, a Hilbert transform is applied to the filtered signal to construct an analytic signal. This analytic signal is a complex signal, with its real part being the original signal and its imaginary part being its Hilbert transform. From this analytic signal, the instantaneous phase angle of the signal at each sampling time point can be calculated. This phase angle changes continuously with time, typically ranging from [-π, π], precisely describing the instantaneous position of the neural activity in the brain region within the oscillation cycle, such as a peak, trough, or rising edge. After performing this operation for each ROI, an instantaneous phase matrix is ​​obtained, where the rows represent the brain regions corresponding to different ROIs, and the columns represent consecutive time points. This transformation elevates functional analysis from the traditional average correlation over time to instantaneous synchronization at each point in time, which is a necessary prerequisite for capturing rapid shifts in brain states.

[0032] The goal of step S103 is to quantify the instantaneous phase information obtained in the previous step into the dynamic functional connectivity strength between brain regions, and to construct a mathematical graph model that includes both connectivity topology and connectivity strength. For each time point t, the instantaneous phase difference between all pairs of regions of interest (n, p) is calculated, and the cosine value is taken to obtain the dynamic phase coherence value.

[0033] In some embodiments, the formula for calculating the phase coherence value is: ; in, Indicates a point in time Time Interest Area Area of ​​Interest The phase coherence value between them Indicates a point in time Time Interest Area Corresponding to the instantaneous phase of the aforementioned blood oxygen-dependent level data time series, Indicates a point in time Time Interest Area The instantaneous phase of the time series of blood oxygen dependence level data.

[0034] This value ranges from -1 to 1; the closer it is to 1, the more synchronized the neural oscillation phases of the two brain regions are at that moment, and the stronger the functional connection. To construct a sparse and robust graph structure that reflects the most significant functional interactions, this invention employs a proportional thresholding method, setting the proportion to 40-45%, preferably retaining the top 40% of connections by strength, and converting the phase coherence value into a binary adjacency matrix. This matrix defines the graph's skeleton: a matrix element of 1 indicates that at time t, there exists a functionally significant connection between two corresponding brain regions, permissible by anatomical priors. Simultaneously, the phase coherence values ​​of each brain region with all other brain regions constitute the node feature vector for that brain region, and the feature vectors of all nodes form the node feature matrix. Thus, the brain network at each time point is formally represented as a graph structure data. , where A t It is a binary adjacency matrix containing anatomical constraints, describing whether a node is connected. X t It is a dynamic node feature matrix that describes the strength of connections. This bimodal representation provides a structurally clear and information-complementary input for subsequent graph neural networks.

[0035] Step S104 aims to achieve deep guidance and fusion of anatomical structural information with the learning process of functional dynamic features through a specially designed neural network model. The constructed graph autoencoder model typically consists of a multi-layer graph convolutional network. During training, the model uses the node feature matrix X at each time point... t As input, and crucially as the binary adjacency matrix A generated in step S103, representing the prior knowledge of fiber bundle connections. t This serves as a fixed structural constraint for graph convolution operations. During graph convolution, the aggregation and propagation of node features are strictly limited to A. t The fusion of functional information occurs along connection edges with a median of 1. This means that the fusion path of functional information fully follows the communication skeleton defined by the dissected connections, and the model compresses high-dimensional functional data into a low-dimensional latent representation Z. t In the process, it is forced to learn functional coordination patterns consistent with the underlying anatomical structure. The decoder then attempts to learn from Z... t The original phase coherence matrix is ​​reconstructed, and the encoder is driven to learn an effective compressed representation by minimizing the reconstruction loss. After pre-training, the encoder parameters are fixed and applied to dynamic graph data at all time points, outputting a series of low-dimensional feature embedding vectors. These embeddings are no longer simple functional statistical summaries, but rather encode fusion features of how functional synergy is achieved on specific anatomical pathways, providing an ideal feature space for subsequent identification of brain states with a clear anatomical basis.

[0036] In some embodiments, the training steps of the graph autoencoder include steps S1041 to S1043: Step S1041: Obtain a training sample set of sample brain connectivity graphs that includes multiple time points. The sample brain connectivity graphs contain a sample binary adjacency matrix representing the connectivity between regions of interest and a sample node feature matrix composed of phase coherence values.

[0037] Step S1042: Obtain the initial encoder and initial decoder. The encoder is obtained by stacking multiple layers of graph convolutional networks or graph attention networks. The initial decoder includes fully connected layers, inner product operation layers, and activation function layers.

[0038] Step S1043: Train the initial encoder and initial decoder using the training sample set. The initial encoder takes the sample brain connection graph as input, and transmits and aggregates the information carried by the feature matrix of the sample nodes along the edges with connection relationships marked by the binary adjacency matrix of the samples. Edges without connection relationships are suppressed, and the output is a low-dimensional latent representation of the samples. The initial decoder takes the low-dimensional latent representation of the samples and the binary adjacency matrix of the samples as input, and outputs a reconstruction matrix of the feature matrix of the sample nodes. The parameters of the initial encoder and initial decoder are updated based on the binary cross-entropy loss of the reconstruction matrix and the feature matrix of the sample nodes to obtain the graph autoencoder and graph autodecoder.

[0039] Step S105 discretizes the continuous, fused feature dynamics into a finite number of meaningful brain states and quantifies them systematically. The low-dimensional feature embeddings from all time points obtained in step S104 are used as input and partitioned using a clustering algorithm, such as K-means. The optimal number of clusters K is determined using methods such as the elbow rule. Each cluster center represents a recurring, quasi-fiber tract-related brain state; it is a prototype point in the low-dimensional feature space, corresponding to a specific structure-function coupling pattern. The algorithm assigns a state label to each time point, thereby generating a discrete sequence of brain states evolving over time. Based on this sequence, three core dynamic indicators can be calculated: 1) Occurrence probability: the proportion of each state in the total time, reflecting its dominance; 2) Average duration: the average duration of holding a state after each entry, reflecting its stability; 3) State transition probability matrix: describing the probability of switching from one state to another, revealing the reorganization patterns and transition preferences of specific fiber tract network functional modes. These quantitative indicators enable a precise and measurable characterization of the functional dynamics of specific neural pathways, providing a direct and powerful tool for comparing the dynamic characteristics of specific circuits in healthy or diseased populations.

[0040] In some embodiments, clustering the low-dimensional latent representations of each region of interest at all time points further includes: clustering using the K-means algorithm, Gaussian mixture model, or spectral clustering algorithm, and determining the optimal number of clusters using the elbow rule or silhouette coefficient.

[0041] In some embodiments, the method further includes: using a hidden Markov model to simultaneously perform brain state clustering on the low-dimensional latent representation and model the state transition probability.

[0042] On the other hand, the present invention also provides a brain state analysis device based on fiber tract guide map autoencoder, including a processor, a memory, and a computer program or instructions stored in the memory. The processor is used to execute the computer program or instructions, and when the computer program or instructions are executed, the device implements the steps of the above method.

[0043] On the other hand, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0044] On the other hand, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0045] The present invention will now be described with reference to a specific embodiment: To overcome the two major shortcomings of existing brain state analysis methods—their detachment from anatomical foundations and the assumption of whole-brain homogeneity—this embodiment proposes a brain state analysis method based on a fiber tract guided map autoencoder. Its basic principle can be summarized as follows: using the anatomical connectivity of white matter fiber tracts as a priori constraints on functional connectivity, the guided map autoencoder performs nonlinear dimensionality reduction and feature extraction on the functional connectivity data, thereby identifying fiber tract-related brain functional states with both high biological interpretability and high dynamic specificity under the constraints of specific neural pathways.

[0046] To address the issue of existing methods that treat all brain connections equally, resulting in mixed state characteristics, this embodiment abandons the traditional whole-brain adjacency matrix. Instead, it constructs multiple fiber tract-specific subnetworks based on a defined white matter fiber tract-brain region mapping relationship. Each subnetwork contains only brain region nodes connected by a specific fiber tract and the functional connections between them. This essentially pre-defines analytical units that conform to anatomical pathways for subsequent analysis, fundamentally ensuring the structural attribution of the identified states. Through this principle, the present invention anchors the macroscopic dynamic phenomenon of brain states to the microscopic anatomical basis of specific white matter fiber tracts. What is identified is no longer a vague whole-brain state, but a clear fiber pathway state, thus providing a clear perspective for understanding the circuit mechanisms of cognitive function and diagnosing functional disorders of specific pathways.

[0047] This embodiment provides a brain state analysis method based on fiber tract guided map autoencoders. The flowchart of this method is shown below. Figure 2 As shown, this includes the following steps S1~S6: Step S1: Selection of fiber-related brain regions and extraction of BOLD signals This step aims to identify the set of brain regions involved in specific functional networks based on the anatomical connectivity map of white matter fiber tracts, and to extract their functional activity signals, thus establishing a data foundation for structural constraints in subsequent analysis. Specifically, 18 major human brain fiber tracts from the Johns Hopkins University white matter fiber tract template were first selected for analysis. These 18 fiber tracts include the bilateral anterior thalamic radii, mandibular fasciculus, inferior longitudinal fasciculus, superior longitudinal fasciculus, superior temporal longitudinal fasciculus, uncinate fasciculus, cingulate hippocampus, cingulate gyrus, and corticospinal tract. These white matter fiber tracts were then overlaid with the Automatic Anatomical Labeling 116 (AAL116) map to identify gray matter brain regions that overlap with the ends of specific white matter fiber tracts as regions of interest (ROIs). Simultaneously, standard preprocessing was performed on the functional magnetic resonance imaging (fMRI) data, including temporal correction, head motion correction, spatial normalization and smoothing, and filtering for noise reduction. Subsequently, the preprocessed functional data is registered with the individual structural space, and the BOLD time series of the above ROI set is accurately extracted from the registered data, thereby obtaining a functional data set with specific fiber bundle connections for each subject.

[0048] Step S2: Calculate the phase of the BOLD signal using Hilbert transform. The core of this step is to transform the BOLD time series of each brain region into an instantaneous phase signal that can characterize its periodic oscillatory dynamics. First, the BOLD time series extracted in the previous step is passed through a bandpass filter (typically 0.01–0.1 Hz) to retain low-frequency oscillations closely related to spontaneous brain activity. Then, a Hilbert transform is applied to the filtered time series to obtain its analytic signal. For any real-valued BOLD signal, its analytic signal is a complex signal, with the real part being the original signal itself and the imaginary part being its Hilbert transform. From this analytic signal, the instantaneous phase angle of the signal at each time point can be directly calculated. This calculation is performed throughout the entire time axis, generating a continuous instantaneous phase time series for each selected brain region.

[0049] Step S3: Calculation of phase coherence matrix The goal of this step is to quantify the strength of functional connectivity between any two fiber bundle-related brain regions at each instant, thereby constructing a dynamic functional connectivity network. Based on the instantaneous phase matrix obtained in the previous step, the instantaneous phase difference between all brain region pairs is calculated for each time point. Then, by calculating the cosine of this phase difference, the phase coherence of the two brain regions at that time point is obtained. The formula is as follows: in, Indicates a point in time Time brain region With brain regions The phase coherence value between them This is the brain region. The BOLD signal at the same time point The instantaneous phase. According to this definition, the phase coherence value ranges from -1 to 1. The closer the value is to 1, the more synchronized the neural oscillation phases of the two brain regions are at that moment, and the stronger the functional connection; conversely, the weaker the synchronization, the weaker the synchronicity. For each time point, the phase coherence values ​​of all brain region pairs are filled into a symmetric matrix, thus obtaining the dynamic phase coherence matrix for that moment. This matrix sequence completely characterizes the dynamic process of the rapid evolution of the internal functional connection patterns of the brain region network connected by selected fiber tracts over time, and is the core input data for characterizing the dynamics of brain function in this method.

[0050] Step S4: Construction of the brain connectivity map This embodiment abstracts the brain functional network of specific fiber connections at each time point into a graph structure. In this graph, the set of nodes Each node in Represents a specific brain region; edge set This dynamically represents the time point. At that time, the functional connectivity between all pairs of brain region nodes, where each edge Corresponding brain regions and The connections at this moment. To represent this graph structure mathematically, we construct a binary adjacency matrix. To describe its topological connections. Specifically, the elements in the matrix. The value is determined based on the dynamic phase coherence matrix obtained in the previous calculation step. Determined: We use a proportional threshold method to retain only those ranked highest in intensity. The connection, that is, these strongest connections Setting it to 1 indicates the existence of significant functional connections; while setting the remaining connections to 1 indicates the presence of significant functional connections. Set to 0. Meanwhile, each graph It is also associated with a node feature matrix. Each row of the matrix It is a node In time The feature vector. In this embodiment, this feature vector is constructed using the phase coherence weights of each brain region (node) with all other brain regions, i.e. This allows the node features to encode their global connection patterns throughout the network, providing rich and topologically consistent input information for subsequent graph neural network analysis.

[0051] Step S5: Perform phase coherence matrix dimensionality reduction using a fiber bundle-guided graph autoencoder. This step uses the anatomical connectivity information of white matter fiber bundles as a priori constraint for graph structure learning, and utilizes a graph autoencoder to perform nonlinear dimensionality reduction on the high-dimensional dynamic functional connectivity data to extract low-dimensional features implying structure-function coupling relationships. Specifically, the binary adjacency matrix generated in step S4... The study quantitatively characterized the existence of significant functional connectivity topological relationships between brain regions under anatomical prior constraints; and the corresponding node feature matrix. Each row vector is composed of continuous phase coherence weights between the corresponding brain region and all other brain regions in the network, thus dynamically reflecting the instantaneous state of each brain region's functional interaction with the whole brain. Together, these constitute the data at each time point. A complete graphical representation of the upper brain functional network In this model, structural (connection status) and functional (connection strength) information are modeled separately and complement each other. Based on this, this embodiment designs a fiber bundle-guided graph autoencoder model. The encoder of this model uses the node feature matrix at each time point. As input, and using the adjacency matrix generated in step S4 based on the anatomical prior, As a graph structure constraint, it modulates the propagation path of information between nodes. During graph convolution, the network relies on the adjacency matrix. The defined connection relationships affect node characteristics. The functional information it carries is selectively transmitted and aggregated. The functional information mainly follows... Connecting edges with a median of 1 (i.e., significant functional connections existing under anatomical constraints) are passed along, while Connections with a median of 0 are suppressed. Through this mechanism, the model can utilize high-dimensional node features... Compression into a low-dimensional latent representation In the process, not only is key functional dynamic information preserved, but anatomical connection patterns are also embedded as the skeleton of information flow, thereby achieving a deep integration of functional dynamics and anatomical structure at the feature level.

[0052] The encoder of this model outputs an embedding vector of each brain region node feature in a low-dimensional latent space. The decoder aims to reconstruct the original phase coherence matrix. Training is performed by minimizing the decoder's reconstruction loss function. The core objective is to drive the encoder to learn low-dimensional embeddings that can effectively reconstruct observed functional connectivity patterns. After training, the encoder weights are fixed and used to process the dynamic phase coherence matrix at all time points, outputting a series of low-dimensional feature embedding vectors. These vectors compactly encode the state of the brain functional network at each time point, and this state has incorporated structural constraint information from white matter fiber tracts.

[0053] Step S6: Cluster brain states using the K-means method The goal of this step is to perform cluster analysis on low-dimensional embeddings that integrate structure-function information, thereby discretizing continuous dynamic brain activity into a finite number of anatomically specific, quasi-stable brain states. First, the low-dimensional feature embeddings of all time points obtained in the previous step are used as input features for the clustering algorithm. Then, the K-means clustering algorithm is used to partition these feature vectors. The optimal number of clusters is determined using standard methods such as the elbow rule or silhouette coefficient; this number corresponds to the number of typical states that may exist in the brain on the fiber tract-related network. Each cluster center represents a fiber tract-related brain state; it is a prototype point in the low-dimensional feature space, corresponding to a specific functional connectivity pattern constrained by anatomical pathways. The algorithm assigns a state label to each time point, thereby generating a discrete sequence describing the transition of brain states over time. Based on this state sequence, the probability of occurrence of each state (i.e., the proportion of that state in the total time), the average duration (i.e., the average length of time maintained after entering that state), and the transition probability matrix between states can be further calculated. These quantitative indicators precisely characterize the dynamic reorganization patterns and temporal evolution trajectories of functional networks related to specific fiber bundles, providing measurable evidence for understanding their physiological significance and changes in diseases.

[0054] Corresponding to the above method, the present invention also provides an apparatus / system including a computer device, the computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the apparatus / system performs the steps of the method as described above.

[0055] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned edge computing server deployment method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0056] This embodiment does not randomly select from all brain regions of the whole brain or use all brain regions of the standard atlas. Instead, it pre-screens a set of brain regions with clear anatomical connections to the ends of specific fiber tracts based on the spatial superposition relationship between a preset white matter fiber tract template and the brain region atlas. These regions are then used as regions of interest for all subsequent analyses. This step is the fundamental premise and physical basis for all subsequent fiber tract-specific analyses, ensuring that the analyzed functional networks have a clear anatomical affiliation.

[0057] When constructing the brain network as a graph structure at each time point, this embodiment employs a dual-matrix representation strategy: a thresholded dynamic phase coherence matrix is ​​used as a binary adjacency matrix to represent the existence of significant functional connections between brain regions; simultaneously, the original dynamic phase coherence row vectors are used as node feature matrices to represent the connection strength between each brain region and the whole brain function. This separate representation method provides structured input for the subsequent graph neural network to process connection topology and node attributes separately.

[0058] Building upon the standard graph autoencoder, the binary adjacency matrix, which contains prior knowledge of fiber bundle connections, determined in step S1, is used as a rigid constraint on the information propagation path through graph convolution operations within the encoder. Specifically, during graph convolution, the aggregation and transmission of node features are strictly limited to connections with a value of 1 in the adjacency matrix. This forces the model to solidify the functional dynamics onto the communication framework defined by anatomical connections during dimensionality reduction and feature extraction, thereby guiding and modulating the learning process of functional information through structural information.

[0059] Based on the low-dimensional embeddings extracted from the aforementioned model, this embodiment identifies not globally homogeneous states through clustering, but rather discrete fiber tract-related brain states corresponding to specific fiber tract anatomical networks. Furthermore, this embodiment provides a set of quantitative index calculation methods for such states, including state occurrence probability, average duration, and intra- / inter-fiber tract state transition probability matrix. These indices enable dynamic separation and precise characterization of specific neural pathway functions.

[0060] In summary, the core effect of this invention lies in constructing a novel analytical paradigm driven by anatomical structure for dynamic functional analysis. It fundamentally breaks the implicit assumption of "whole-brain homogeneity" in traditional brain state analysis. Traditional methods treat all brain regions with equal weight, resulting in a mixed and anatomically ambiguous macroscopic statistical pattern of identified states. This invention, however, pre-limits the analysis to clusters of brain regions connected by specific white matter fiber bundles (such as the 18 major fiber bundles), precisely "anchoring" macroscopic dynamic functional activity to specific microscopic anatomical pathways. Therefore, each analyzed "brain state" is no longer a vague whole-brain pattern, but rather a "fiber bundle functional state" with clear anatomical specificity, fundamentally solving the bottleneck problem of poor state interpretability in existing technologies.

[0061] Building upon this foundation, the technical advantages of this invention are further demonstrated by the deep integration of structural and functional information within the neural network model. Existing technologies rely solely on noise-sensitive BOLD functional signals, resulting in insufficient feature robustness. This invention innovatively introduces a binary adjacency matrix representing the anatomical connections of fiber bundles as a rigid constraint into the feature learning process of the graph autoencoder. In graph convolution operations, the transmission of functional information is forcibly confined to biologically rational paths defined by the anatomical structure. This mechanism enables the model to utilize a robust anatomical framework to guide and "purify" the learning of noise-sensitive functional signals, thereby extracting more biologically meaningful and interference-resistant "structure-function" coupled features, significantly improving the robustness and reliability of the identified state features.

[0062] Ultimately, these technological innovations have given rise to novel dynamic biomarkers with high clinical application potential. Traditional whole-brain status analysis, due to its ambiguous anatomical basis, makes it difficult to pinpoint specific pathological circuits. This embodiment can precisely quantify the dynamic functional states within specific fiber tract pathways (such as pathways known to be vulnerable in schizophrenia, Alzheimer's disease, or brain injury), including their probability of occurrence, duration, and transition patterns. These indicators provide an unprecedented window into understanding the circuit mechanisms of diseases. Therefore, the fiber tract-specific dynamic state indicators generated by this invention have the potential to develop into novel biomarkers for precise disease diagnosis, efficacy evaluation, and even surgical target planning, providing a completely new technological perspective and practical tool for precision medical intervention in neurological diseases.

[0063] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0064] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0065] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A brain state analysis method based on fiber tractography autoencoders, characterized in that, The method includes the following steps: Functional magnetic resonance imaging (fMRI) data of the brain of the subject to be analyzed is preprocessed and spatially registered with the individual brain structure to extract time series data of blood oxygen dependence levels in multiple regions of interest; the regions of interest are gray matter brain regions connected by specified white matter fiber bundles. The blood oxygen dependence level data time series of each region of interest are processed by a bandpass filter to extract the target low-frequency oscillation component and perform Hilbert transform. The instantaneous phase of the signal at each time point is calculated to obtain the instantaneous phase time series of each region of interest. At each time point, the instantaneous phase difference between all the regions of interest is calculated and the corresponding cosine value is calculated as the phase coherence value. A connection relationship is established between the regions of interest corresponding to a set proportion of phase coherence values ​​with higher values ​​to form a brain connectivity map. The brain connectivity map includes a binary adjacency matrix representing the connection relationship between the regions of interest and a node feature matrix composed of the phase coherence values. Based on a pre-trained graph autoencoder model, the node feature matrix at each time point is used as input, and the binary adjacency matrix is ​​used as a constraint to extract the low-dimensional latent representation corresponding to each region of interest. Cluster the low-dimensional latent representations of each region of interest at all time points, with each cluster representing a brain state; assign a unique state label to each time point to obtain a discrete sequence of brain states switching over time, and calculate the occurrence probability, average duration, and state transition probability of each brain state.

2. The brain state analysis method based on fiber tract guide map autoencoder according to claim 1, characterized in that, Before preprocessing the functional magnetic resonance imaging (fMRI) data of the brain of the subject to be analyzed, the method further includes: Multiple target white matter fiber tracts of the human brain are superimposed with a preset anatomical atlas to determine the gray matter brain regions connected by the target white matter fiber tracts as regions of interest; the target white matter fiber tracts include the bilateral anterior thalamic radiation, inferior fronto-occipital fasciculus, inferior longitudinal fasciculus, superior longitudinal fasciculus, superior temporal longitudinal fasciculus, uncinate fasciculus, cingulate hippocampus, cingulate gyrus and corticospinal tract. Preprocessing of the functional magnetic resonance imaging (fMRI) data of the brain of the subject to be analyzed includes: performing temporal correction, head motion correction, spatial normalization and smoothing, and filtering and denoising on the fMRI data.

3. The brain state analysis method based on fiber tractography autoencoder according to claim 1, characterized in that, The formula for calculating the phase coherence value is: ； in, Indicates a point in time Time Interest Area Area of ​​Interest The phase coherence values ​​between, Indicates a point in time Time Interest Area Corresponding to the instantaneous phase of the aforementioned blood oxygen-dependent level data time series, Indicates a point in time Time Interest Area The instantaneous phase of the time series of blood oxygen-dependent level data.

4. The brain state analysis method based on fiber tract guide map autoencoder according to claim 1, characterized in that, The frequency range of the bandpass filter is 0.01~0.1Hz; Establishing a connection between the regions of interest corresponding to a set proportion of phase coherence values ​​with higher values ​​includes setting the set proportion to 40-45% based on a proportional threshold method.

5. The brain state analysis method based on fiber tractography autoencoder according to claim 1, characterized in that, The training steps of the graph autoencoder include: A training sample set is obtained, which includes a sample brain connectivity graph at multiple time points. The sample brain connectivity graph contains a sample binary adjacency matrix representing the connectivity between the regions of interest and a sample node feature matrix composed of the phase coherence values. An initial encoder and an initial decoder are obtained. The encoder is obtained by stacking multiple layers of graph convolutional networks or graph attention networks. The initial decoder includes fully connected layers, inner product operation layers, and activation function layers. The initial encoder and the initial decoder are trained using the training sample set. The initial encoder takes the sample brain connectivity graph as input and transmits and aggregates the information carried by the sample node feature matrix along the edges with existing connections marked by the sample binary adjacency matrix, while edges without connections are suppressed, outputting a low-dimensional latent representation of the sample. The initial decoder takes the low-dimensional latent representation of the sample and the sample binary adjacency matrix as input and outputs a reconstruction matrix of the sample node feature matrix. The parameters of the initial encoder and the initial decoder are updated based on the binary cross-entropy loss of the reconstruction matrix and the sample node feature matrix to obtain the graph autoencoder and graph autodecoder.

6. The brain state analysis method based on fiber tract guide map autoencoder according to claim 1, characterized in that, Clustering the low-dimensional latent representations of the interest regions at all time points further includes: Clustering is performed using the K-means algorithm, Gaussian mixture model, or spectral clustering algorithm, and the optimal number of clusters is determined by the elbow rule or silhouette coefficient.

7. The brain state analysis method based on fiber tractography autoencoder according to claim 1, characterized in that, The method further includes: Hidden Markov Models are used to simultaneously perform brain state clustering on the low-dimensional latent representation and model the state transition probabilities.

8. A brain state analysis device based on fiber tractography autoencoder, comprising a processor, a memory, and a computer program or instructions stored in the memory, characterized in that, The processor is configured to execute the computer program or instructions, and when the computer program or instructions are executed, the device implements the steps of the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.

Images