Functional brain network classification method fusing multi-scale hierarchical dynamic graph convolution

Abstract

The application belongs to the technical field of brain image intelligent analysis and graph neural network, and discloses a functional brain network classification method fusing multi-scale hierarchical dynamic graph convolution, which acquires data containing whole brain region time sequence matrix, demographic phenotype information and classification labels, and constructs a sample data set; converts the time sequence matrix into a multi-scale positive and negative dual-channel dynamic functional connection graph sequence, and obtains a window-level aggregated connection graph through time sequence attention weighted aggregation; realizes brain region node to topological super node hierarchical mapping by using a double-branch graph convolution, and outputs a window-level spatial feature; obtains a subject global brain function representation through single-scale time sequence modeling and cross-scale adaptive fusion; builds a group graph by combining the representation and the phenotype, outputs a prediction probability through group graph convolution, phenotype gate fusion and a classifier, and trains until the model converges; and the trained model can complete individual functional brain network classification. The application can combine brain dynamic change representation, group relationship and phenotype priori, and is suitable for auxiliary identification of brain diseases.

Description

Technical Field

[0001] This invention relates to the fields of intelligent brain image analysis and graph neural network technology, and in particular to a functional brain network classification method that integrates multi-scale hierarchical dynamic graph convolution. It can be widely applied to scenarios such as functional brain network modeling based on functional magnetic resonance imaging data, brain disease assisted identification, and medical artificial intelligence analysis. Background Technology

[0002] With the development of functional magnetic resonance imaging (fMRI) technology, brain disease identification methods based on functional brain networks have attracted widespread attention. The goal of functional brain network classification is to determine the category to which a subject belongs based on the time series or functional connectivity characteristics of whole-brain regions, thereby providing a basis for the auxiliary diagnosis of autism spectrum disorder, mild cognitive impairment, depression, and other brain diseases.

[0003] Most existing functional brain network classification methods are based on static functional connectivity matrices, assuming that brain region connectivity remains unchanged throughout the scan. However, brain functional activity exhibits significant dynamic changes, and static connectivity representations are insufficient to fully capture the fine-grained changes in brain networks over time. To address this, some methods introduce sliding time windows to construct dynamic functional connectivity maps, but existing dynamic analysis methods still have the following shortcomings: First, most methods model only on a single time scale, making it difficult to simultaneously consider both short-term local dynamic patterns and long-term stable trends; second, most methods do not differentiate between positively and negatively correlated connections, resulting in insufficient utilization of discriminative information for different types of connectivity; third, existing methods typically analyze the brain networks of individual subjects independently, ignoring the group similarity in brain connectivity patterns and phenotypic information among different subjects; fourth, although some methods incorporate deep learning models, their ability to jointly model spatial structural features, temporal dependence features, and group relationship features remains insufficient, affecting classification accuracy and generalization performance.

[0004] Therefore, proposing a functional brain network classification method that can simultaneously integrate multi-scale dynamic connectivity information, population graph relationship information, and phenotypic information to improve the accuracy and robustness of functional brain network classification has become an urgent problem to be solved. Summary of the Invention

[0005] In view of this, the present invention provides a functional brain network classification method that integrates multi-scale hierarchical dynamic graph convolution to solve the problem in the prior art that it is difficult to simultaneously characterize the multi-timescale dynamic features, positive and negative connection difference features, group similarity relationship features and phenotypic auxiliary features of functional brain networks.

[0006] This invention provides a functional brain network classification method that integrates multi-scale hierarchical dynamic graph convolution, comprising:

[0007] S1: Construct a sample dataset and divide it into a training set, a validation set, and a test set according to a preset ratio. Each sample in the sample dataset contains input data and label data. The input data is the time series matrix of the whole brain regions of an individual and demographic phenotypic information. The label data is a functional brain network classification label.

[0008] S2: Construct a multi-scale hierarchical dynamic graph convolutional model for functional brain network classification, and train the model parameters using the training set. Training the multi-scale hierarchical dynamic graph convolutional model includes the following steps:

[0009] S21: Construct a multi-scale dual-channel dynamic functional connectivity map sequence based on the time series matrix of the whole brain regions of an individual, which can simultaneously characterize the dynamic change pattern of the network at different time granularities. The dual-channel dynamic functional connectivity map sequence at any scale includes a positive connectivity map sequence and a negative connectivity map sequence.

[0010] S22: Perform temporal attention weighted aggregation on the positive connection subframe graph sequence and the negative connection subframe graph sequence corresponding to each sliding window at the same scale to obtain window-level aggregated positive connection graph and aggregated negative connection graph;

[0011] S23: A symmetrical and parameter-independent bi-branch graph convolutional network is used to perform hierarchical feature mapping from the brain region node level to the topological supernode level on the aggregated positive connection graph and the aggregated negative connection graph, and to fuse and construct a window-level spatial feature vector that takes into account both positive and negative connection modes.

[0012] S24: Perform hierarchical temporal processing on the window-level spatial feature vector, aggregate temporal dimension fluctuation information and fuse it into a single and complete individual-level global representation of brain function. The hierarchical temporal processing includes: single-scale temporal modeling and cross-scale adaptive fusion.

[0013] S25: Calculate multi-dimensional similarity based on the brain function representation vectors and demographic phenotypic information of all individuals in the training set, and construct a normalized group graph; mine individual group association features through group graph convolution, and input the group features and phenotypic embedded features into the classifier after fusion through a gating mechanism, and output the class prediction probability; calculate the loss by combining the sample's real label, backpropagate and update all learnable parameters of the model until the model converges;

[0014] S3: Use the trained multi-scale hierarchical dynamic graph convolutional model to perform functional brain network classification on the individuals to be classified.

[0015] Preferably, in S1, each element in the whole-brain time series matrix represents the blood oxygenation level-dependent signal intensity of different brain regions of an individual at different times; the demographic phenotypic information includes the individual's gender, age, and data collection site information.

[0016] Further optimization, in S1, the steps for obtaining the time series matrix of an individual's whole brain regions are as follows:

[0017] Resting-state scans of individuals are performed using functional magnetic resonance imaging (fMRI) to obtain raw image data of the whole brain's functional connectivity.

[0018] The original image data is preprocessed, and a whole-brain time series matrix is ​​extracted based on a preset AAL map.

[0019] The time series matrix of the whole brain regions is subjected to standardized preprocessing, which includes linear detrending, bandpass filtering, and Z-score normalization.

[0020] Further optimization, in S21, the method for constructing the multi-scale dual-channel dynamic functional connectivity graph sequence is as follows:

[0021] For each individual's whole-brain time series, at least two different time scales are preset;

[0022] For any time scale, the time series of an individual's whole brain regions is segmented using sliding windows, and the total number of sliding windows is calculated.

[0023] The whole-brain brain region time series within each sliding window at each scale is divided into sub-sliding windows to obtain the sub-frames corresponding to each sliding window.

[0024] Based on the time series segments of the whole brain regions corresponding to the subframes, the Pearson correlation coefficients between each pair of brain regions are calculated, the subframe correlation matrix is ​​constructed, and preprocessing and Fisher Z transform are performed to obtain the subframe transformation matrix.

[0025] The subframe transformation matrix is ​​split into positive and negative connected subframe maps according to the sign of the elements. All windows at each scale are traversed to finally form a multi-scale dual-channel dynamic function connection map sequence.

[0026] Further optimization involves preprocessing the subframe correlation matrix, including: zeroing out the diagonal, replacing outliers, and cropping the values.

[0027] Further optimization, in S22, performing temporal attention weighted aggregation on the positive or negative connected subframe sequence corresponding to the sliding window at a single scale includes the following steps:

[0028] For the single scale, calculate the intensity score of each frame of the motion graph within the sliding window, wherein the motion graph is a positively connected subframe graph or a negatively connected subframe graph;

[0029] Based on the intensity score of each frame, the temporal attention weights corresponding to each frame are calculated through Softmax normalization.

[0030] The dynamic graph within the sliding window is weighted and summed according to the time-dimensional attention weights to obtain a window-level aggregated positive connection graph or aggregated negative connection graph.

[0031] Further optimization, in S23, the process of constructing the window-level spatial feature vector is as follows: The graph-level feature vectors corresponding to the aggregated positive connectivity graph and the aggregated negative connectivity graph are obtained respectively, and then concatenated by channels to obtain the window-level spatial feature vector; wherein, the method for obtaining the graph-level feature vector is as follows:

[0032] Add self-connections to aggregated positive or negative connection graphs to construct an enhanced adjacency matrix. After calculating the degree matrix based on the enhanced adjacency matrix, multi-layer symmetric normalized graph convolution is performed, and residual connections are introduced to obtain the enhanced residual node feature matrix.

[0033] Based on the enhanced residual node feature matrix, a learningable soft assignment network is used to generate the mapping relationship between brain region nodes and super nodes, thus obtaining the super node soft assignment matrix.

[0034] Based on the supernode soft assignment matrix, the original brain region nodes are aggregated into supernodes, resulting in a supernode feature matrix and a supernode adjacency matrix.

[0035] The symmetric normalized graph convolution operation is performed again at the supernode level to capture the spatial interaction patterns between higher-order functional modules and obtain the output state feature matrix.

[0036] By utilizing a graph readout mechanism that combines global average pooling and global max pooling, the supernode feature matrix is ​​aggregated along the node dimension into a graph-level feature vector of fixed dimension.

[0037] Further optimization, in S24, the single-scale temporal modeling is as follows: all window-level spatial features at the same scale are arranged temporally to form a feature sequence, which is then input into a Transformer encoder after linear projection and positional encoding to capture long-range temporal dependencies between windows; contextual smoothing features are obtained through temporal attention pooling, and transient salient features are extracted through temporal max pooling; the two types of features are concatenated to obtain the scale-specific temporal features; the cross-scale adaptive fusion is as follows: the temporal features at each scale are projected into query vectors and key vectors respectively, and the adaptive fusion weights of each scale are calculated through an additive attention mechanism; the temporal features at each scale are weighted and summed according to the adaptive fusion weights to generate an individual-level brain function representation vector.

[0038] Further optimization, S25 specifically includes:

[0039] The cosine similarity of brain function between individuals is calculated based on the brain function representation vectors of all individuals in the training set; a phenotypic association matrix is ​​constructed based on the demographic phenotypic information of individuals; the brain connectivity similarity matrix and each phenotypic similarity matrix are fused according to preset weights to obtain a comprehensive group similarity matrix.

[0040] Add self-loops to the comprehensive group similarity matrix and perform symmetric normalization to generate a group graph adjacency matrix;

[0041] The brain function representation vectors of all individuals are concatenated into an initial node feature matrix, and then multi-layer residual graph convolution operation is performed with the adjacency matrix of the population graph to obtain the population relationship feature vector.

[0042] Demographic phenotypic information is mapped into phenotypic feature vectors using a multilayer perceptron; group relationship feature vectors and phenotypic feature vectors are input into a gating fusion mechanism to obtain a fused comprehensive feature vector.

[0043] The fused integrated feature vector is then input into the classifier for probability prediction.

[0044] Further optimization involves using the cross-entropy loss function to calculate the network loss value during the training phase of the multi-scale hierarchical dynamic graph convolution model.

[0045] The functional brain network classification method provided by this invention integrates multi-scale hierarchical dynamic graph convolution. Through multi-scale dynamic mapping, it can simultaneously characterize the dynamic changes of brain networks at different temporal granularities; through positive and negative connection bi-branch spatial graph convolution and supernode pooling, it can enhance the representation ability of local connection patterns and higher-order topological relationships; through temporal coding and cross-scale fusion, it can improve the modeling ability of temporal evolution features of brain networks; and through population graph convolution and phenotypic information fusion, it can make full use of the similarity relationships and prior information among subjects, thereby improving the accuracy and robustness of functional brain network classification. Attached Figure Description

[0046] Figure 1 The overall flowchart of the functional brain network classification method that integrates multi-scale hierarchical dynamic graph convolution provided by the present invention;

[0047] Figure 2 This is a schematic diagram of the multi-scale dual-channel dynamic functional connection diagram sequence construction process in this invention;

[0048] Figure 3 A flowchart for temporal attention aggregation within a window;

[0049] Figure 4 This is a flowchart for feature extraction from a bibranch spatial graph;

[0050] Figure 5 A flowchart for hierarchical timing processing;

[0051] Figure 6 A flowchart for constructing a population graph and fusing phenotypic features for classification. Detailed Implementation

[0052] The present invention will be further described below with reference to specific embodiments.

[0053] like Figure 1 As shown, this invention provides a functional brain network classification method that integrates multi-scale hierarchical dynamic graph convolution, comprising the following steps:

[0054] S1: Construct a sample dataset and divide it into a training set, a validation set, and a test set according to a preset ratio. Each sample in the sample dataset contains input data and label data. The input data is the whole brain region time series matrix and demographic phenotypic information of the subject. The label data is a functional brain network classification label. In this embodiment, it is a disease binary classification label, namely typical developmental normal control (TD) or autism spectrum disorder (ASD). The label data serves as supervision data during the network training phase, which is used to constrain the network to learn brain network topological features, temporal evolution features, and group relationship features simultaneously.

[0055] In order to ensure the sufficiency of parameter training for deep learning networks and group graph convolution, the total number of subject samples... Ideally, the dataset should be in the hundreds or thousands (e.g., selected from the public dataset ABIDE). However, the technical solution of the present invention is not limited thereto;

[0056] Each element in the whole-brain time series matrix represents the oxygenation level-dependent (BOLD) signal intensity of different brain regions of an individual at different times.

[0057] The demographic phenotypic information includes the subject's gender, age, and collection site information, which can be collected synchronously through clinical demographic questionnaires, access to electronic medical records, or multi-center joint collection and recording.

[0058] The steps for obtaining the whole-brain time series matrix of the subjects are as follows:

[0059] S101: The subject is scanned in a resting state using a functional magnetic resonance imaging (fMRI) device to obtain raw image data of the whole brain functional connectivity of the subject;

[0060] S102: Preprocess the raw image data and extract the whole brain region time series matrix based on the preset AAL (Automated Anatomical Labeling) map;

[0061] In this embodiment, the rows of the whole-brain time-series matrix correspond to different brain regions (Region of Interest, ROI; the number of brain regions in this embodiment is...). (Set as 116), columns correspond to different sampling time points, and the time interval between two adjacent columns is the repetition time (TR), the total duration of the whole brain scan. Total number of time points compared to whole brain scan The relationship (i.e., the total number of columns in the matrix) is as follows:

[0062] ;

[0063] S103: Perform standardized preprocessing on the whole brain region time series matrix, wherein the standardized preprocessing includes linear detrending processing, bandpass filtering processing and Z-score normalization processing;

[0064] The linear detrending process is as follows: For the original BOLD signal intensity time-series vector of brain region nodes, the least squares method is used to fit and remove the trend term that shifts linearly with scan time. The calculation expression is as follows:

[0065] ;

[0066] In the formula, Indicates the sampling time point The original BOLD signal strength value at that location. and These are the linear regression coefficients and intercept term obtained by fitting using the least squares method, respectively. This represents the signal strength value after detrending processing.

[0067] The bandpass filtering process is as follows: A preset ideal bandpass filter is used to apply the detrended filter... Time-domain or frequency-domain filtering is performed to strictly constrain the signal frequency between 0.01-0.1Hz in order to filter out low-frequency physiological noise caused by breathing and heartbeat, as well as high-frequency thermal noise from the instrument.

[0068] The Z-score standardization and normalization process is as follows: the mean centering and standard deviation normalization of the filtered signal time vector are performed, and the calculation formula is as follows:

[0069] ;

[0070] In the formula, This represents the signal strength value after filtering. This represents the average signal intensity value of the brain region node during the whole-brain scan cycle. This represents the corresponding standard deviation. The final normalized BOLD signal strength value;

[0071] Through the above cascaded preprocessing, the dimensional influence and baseline noise shift between different individuals and across collection sites can be effectively eliminated, ensuring the consistency of the supervision data and the individual functional brain network representation in space and at multiple scales.

[0072] The preferred ratio of the training set, validation set, and test set is 8:1:1.

[0073] S2: Construct a multi-scale hierarchical dynamic graph convolutional model for functional brain network classification, and train the model parameters using the training set. Training the multi-scale hierarchical dynamic graph convolutional model includes the following steps:

[0074] S21: Based on the whole brain region time series matrix of the subjects, a multi-scale dual-channel dynamic functional connectivity map sequence can be constructed to simultaneously characterize the dynamic change pattern of the network at different time granularities, distinguish positive and negative correlation connections between brain regions, and provide dynamic brain network topology input for subsequent spatial feature extraction.

[0075] The method for constructing the multi-scale dual-channel dynamic function connectivity graph sequence is as follows:

[0076] S211: For the whole-brain time series of each subject, at least two different time scales are pre-set. , respectively denoted as time scale Each time scale Each with an independent window length Defined to simultaneously capture the fine-grained dynamic evolution patterns and long-term stable evolution trends of functional brain networks at different temporal granularities;

[0077] To provide a clear and concrete representation of the multi-scale mechanism, the total number of time-scale groups is used as the basis. Let's illustrate this with an example set to 2 (e.g.) Figure 2 (As shown): The time scale includes the length of the first window. and the first sliding step Defined scale 1 (short window branch) and the second window length Second sliding step Defined Scale 2 (Long Window Branch);

[0078] Scale 1 (short window branch): used to capture short-term local dynamic change patterns in the brain network; the length of the first window... The preferred value range is 30 seconds to 60 seconds, and the first sliding step size is... The preferred value range is 1 TR to 5 TRs;

[0079] In this embodiment, TR = 2 seconds, and the first window length is... Set to 30 seconds (i.e., 15 TRs), first sliding step. Set to 5 TRs;

[0080] The second scale (long window branch) is used to capture the long-term stable evolutionary trend of brain networks; the length of the second window... The preferred value range is 90 seconds to 150 seconds, and the second sliding step size is... The preferred value range is 1 TR to 5 TRs; in this embodiment, the second window length is... Set to 90 seconds (i.e., 45 TRs), second sliding step. Set to 5 TRs;

[0081] S212: For any time scale The time series data of the whole brain regions of the subjects were segmented using sliding windows, and the total number of sliding windows was calculated. ;

[0082] Assuming the total length of the whole brain region time series for each subject is For any time scale The total number of sliding windows it obtains The general calculation formula is as follows:

[0083] ;

[0084] In the formula, This represents the floor operator; These represent the corresponding time scales. The window length and sliding step size, for example, when At that time, the total number of sliding windows is recorded as the total number of short windows at scale 1. and the total number of long windows at scale 2 ;

[0085] S213: Regarding scale Next Time-series fragments of whole-brain regions within a sliding window Using a length of Step size is The sub-sliding window is further divided to subdivide the time series within the window into smaller segments. Number of subframe segments, of which the total number of subframes is The calculation formula is:

[0086] ;

[0087] In the formula, Representing scale Next The total number of subframes obtained within a window; and These are the length of the sub-sliding window and the sliding step, respectively, and their values ​​can be determined based on the window length. Set proportionally (e.g., child window length) Get window length The sliding step size of the sub-window is 1 / 3 to 1 / 2. Take 1 to 3 TRs). This represents the floor operator;

[0088] Record No. The time-series segments of the whole brain regions corresponding to each subframe are: ,in For the number of brain regions, The subframe length (i.e., the number of sampling time points within the sub-sliding window); at any given time scale Below, each sliding window After the above two partitions, all are generated The frame animation provides input for subsequent temporal attention aggregation within the S22 window;

[0089] S214: Regarding scale Next The first window Each subframe corresponds to a time sequence segment of the whole brain region. (Among them, when) hour, , ;when hour, , ), calculate the Pearson correlation coefficient between each pair of brain regions, and construct the subframe correlation matrix. The subframe transformation matrix is ​​obtained by preprocessing and Fisher Z transform. ;

[0090] Among them, the subframe correlation matrix Specifically as follows:

[0091] ;

[0092] correspond Figure 2 The illustrated embodiment shows that the subframe correlation matrices constructed at scale 1 and scale 2 are denoted as follows: and ;

[0093] in, For scale Next The whole-brain time series segment corresponding to the u-th subframe within a sliding window. Number of brain regions (in this embodiment) ), Indicates the subframe length. This indicates that the Pearson correlation coefficient is calculated for any two rows of the matrix, and any element... Indicates brain regions With brain regions The linear correlation strength of the BOLD signal within this subframe is calculated using the following formula:

[0094] ;

[0095] In the formula, Describing covariance, Indicates standard deviation, and Representing matrices respectively The row and number The range of values ​​for this element is: To avoid the standard deviation from zero due to abnormally constant signals in specific brain regions, which could cause calculation failure in the denominator, a pre-defined small positive smoothing term is preferably introduced in a cascaded manner in the denominator. (For example To ensure the absolute completeness of mathematical logic;

[0096] Among them, the subframe correlation matrix Preprocessing includes: ① Diagonal zeroing – setting the diagonal elements of the matrix to zero to eliminate the interference of brain region autocorrelation on subsequent analysis; ② Outlier replacement – ​​replacing abnormal correlation coefficient values ​​that significantly deviate from the normal distribution with the median of all connection weights in that brain region; ③ Numerical pruning – constraining the correlation coefficient values ​​to preset upper and lower limits. To suppress the negative impact of extreme noise on model training;

[0097] Specifically, the Fisher Z-transform is performed on the preprocessed subframe correlation matrix to transform the Pearson correlation coefficient from a bounded state. The formula for mapping an interval to an approximately unbounded real space is as follows:

[0098] ;

[0099] In the formula, The matrix after Fisher's Z transformation has dimensions of . ; The predefined natural logarithm operator;

[0100] S215: The matrix after Fisher Z transformation Split into positively connected subframes based on the sign of the elements. Negative connection subframe graph To learn two different types of functional connectivity patterns:

[0101] ;

[0102] correspond Figure 2 As shown, among the dual-channel features obtained after splitting, the lower window at scale 1... The corresponding positive / negative connection subframe diagrams are denoted as follows: and The window at scale 2 The corresponding positive / negative connection subframe diagrams are denoted as follows: and ;

[0103] In the formula, For positively connected subframes, retain Set all positive elements in the array, and set the rest to zero; For negative connection subframes, the absolute value of the negative elements is taken as the connection weight, and the rest are set to zero;

[0104] For scale Each window All subframes within Perform the above operations respectively to obtain the positively connected subframe graph sequence corresponding to each window. With negative connection subframe sequence ; Traversal scale All windows This constitutes a multi-scale, dual-channel dynamic functional connectivity sequence of the subjects;

[0105] S22: Perform temporal attention weighted aggregation on the positive connection subframe graph sequence and the negative connection subframe graph sequence corresponding to each sliding window at the same scale to eliminate single-window temporal noise and obtain window-level aggregated positive connection graph and aggregated negative connection graph.

[0106] like Figure 3 As shown, for a single sliding window t Frame animation graph sequence (positive connection graph branches and negative connection graph branches are processed independently and in parallel; the following will discuss the positive connection subframe graph). Negative connection subframe graph The corresponding positively connected subframe sequences and negative connection subframe sequence Unified abbreviation The specific process of weighted aggregation using the attention mechanism is as follows:

[0107] S221: Calculate the intensity score for each frame of the motion graph. This reflects the overall activity level of brain functional connectivity in that frame:

[0108] ;

[0109] in, This represents the animated graph corresponding to the u-th subframe, with dimension 1. ; This indicates taking the absolute value of all connection weights in the graph; This means calculating the mean of all non-zero elements;

[0110] S222: Based on the intensity scores of each frame, calculate the temporal attention weights corresponding to each frame using Softmax normalization. :

[0111] ;

[0112] In the formula, Indicates a specific time scale Below, the first in the current sliding window The temporal attention weights corresponding to the frame animation graph; and These represent the current sliding window's first and second elements. Frame and the The intensity score corresponding to the frame animation; This is a preset exponential function; scale The total number of subframes obtained by double partitioning within the t-th sliding window

[0113] S223: Based on the aforementioned time-dimensional attention weights, The frame dynamic graph is weighted and summed to obtain the aggregated dynamic graph representation. :

[0114] ;

[0115] The above aggregation is performed independently on the positive and negative connection subframe graph sequences, and each sliding window t ultimately yields a pair of aggregated positive connection graphs. and aggregation negative connection graph ;

[0116] S23: A structurally symmetric and parameter-independent bi-branch graph convolutional network is used to perform hierarchical feature mapping from the brain region node level to the topological supernode level on the aggregated positive connection graph and the aggregated negative connection graph, respectively, and to fuse and construct a window-level spatial feature vector that takes into account both positive and negative connection modes.

[0117] like Figure 4As shown, the positive connection graph of aggregation. and aggregation negative connection graph Each element enters its own independent graph convolution branch and sequentially completes ROI-level graph convolution and residual connection, supernode pooling, supernode graph convolution, graph readout and dual-branch feature fusion to obtain the window-level spatial feature vector corresponding to each window.

[0118] The specific processing procedure is as follows:

[0119] S231: Aggregate graph of input (using aggregated positive connection graph) Taking this as an example, the dimension is... Add self-connections to construct an enhanced adjacency matrix. After calculating the degree matrix based on the enhanced adjacency matrix, multi-layer symmetric normalized graph convolution is performed, and residual connections are introduced to obtain the enhanced residual node feature matrix.

[0120] Among them, the enhanced adjacency matrix as follows:

[0121] ;

[0122] In the formula, This represents the enhanced adjacency matrix after adding self-connections. This represents the original adjacency matrix of the input (i.e., the aggregated positive connection graph). ), This represents a predefined identity matrix, where all diagonal elements are 1 and all other elements are 0, used to preserve the initial topological features of brain region nodes during graph convolution. Number of brain regions;

[0123] Degree matrix The first in Line number The formula for calculating the diagonal element values ​​of a column is as follows:

[0124] ;

[0125] In the formula, The enhanced adjacency matrix The Summing all elements in the row yields the result. This represents the enhanced adjacency matrix. The Middle Line number The graph connects the weighted element values ​​of the columns; and These are the preset spatial topological location indices of brain region nodes, and their values ​​range from 1 to [value missing]. ; The diagonal degree matrix represents the number of brain regions. All off-diagonal elements (i.e., matrix elements whose row indices are not equal to column indices) are pre-set to zero;

[0126] In multi-layer symmetric normalized graph convolution, the first... Layer input features Calculate the first Layer output features:

[0127] ;

[0128] In the formula, For the first The input node feature matrix of the layer has a dimension of ; For the first The output node feature matrix of the layer has a dimension of The output dimension of each layer remains consistent, i.e. The weight parameter matrix is ​​a learnable matrix. For bias terms; The ReLU activation function is preferred, and residual connections are introduced on this basis to obtain the enhanced residual node feature matrix. :

[0129] ;

[0130] S232: Based on the enhanced residual node feature matrix, a learnable soft allocation network adaptively generates the mapping relationship between brain region nodes and functional modules (i.e., supernodes), thus obtaining the supernode soft allocation matrix. :

[0131] ;

[0132] In the formula, The enhanced residual node feature matrix; and These are the preset learnable weight matrix and learnable bias term vector in the soft assignment network, respectively; The preset number of supernodes (set to 16 or 32 in this embodiment); This represents the normalized exponent operator performed along the row direction of the matrix, used to ensure the soft allocation matrix of the supernode. The sum of each row of elements is strictly 1;

[0133] S233: Based on the supernode soft allocation matrix Will The original brain region nodes aggregated into For each supernode, calculate the pooled supernode feature matrix. Adjacency matrix of supernodes :

[0134] ;

[0135] S234: Perform symmetric normalized graph convolution operation again at the supernode level to capture the spatial interaction patterns between higher-order functional modules and obtain the output state feature matrix. The output state feature matrix The calculation formula is as follows:

[0136] ;

[0137] In the formula, This represents the learnable weight parameter matrix of the supernode hierarchical graph convolution. This represents the corresponding learnable bias term parameter. The supernode adjacency matrix The corresponding degree matrix, The ReLU activation function is preferred.

[0138] S235: Utilizing a graph readout mechanism combining global average pooling and global max pooling, the supernode feature matrix is ​​processed. Aggregate along the node dimension to transform it into a graph-level feature vector with fixed dimensions;

[0139] The graph-level feature vectors output independently by the positive and negative connection branches are denoted as follows: and The formula is as follows:

[0140] ;

[0141] ;

[0142] In the formula, Indicates feature concatenation operation; This represents a preset global average pooling operator, specifically used to perform pooling along the node dimension on the supernode feature matrix. The arithmetic mean of the eigenvectors of all supernodes in the dataset is calculated to extract the current eigenvalue. Global smoothed contextual statistical features within a time window; This represents a preset global max-pooling operator, specifically used to perform max-pooling along the node dimension on the supernode feature matrix. The feature vectors of all supernodes are filtered dimension-wise and the maximum feature value is extracted to capture the current eigenvalue. Significant transient topological response characteristics within a time window;

[0143] S236: Use a channel splicing unit to splice the two channels to obtain a window-level spatial feature vector. :

[0144] ;

[0145] S24: Perform hierarchical temporal processing on the window-level spatial feature vectors, aggregate temporal dimension fluctuation information, and fuse them into a single and complete global representation of subject-level brain function, wherein, for example... Figure 5 As shown, hierarchical temporal processing includes: single-scale temporal modeling and cross-scale adaptive fusion;

[0146] Among them, such as Figure 5 As shown, single-scale time series modeling is specifically as follows:

[0147] First, on a specific time scale Below, at this scale, the obtained The window-level spatial feature vectors of each time window are arranged in chronological order to form a spatial feature sequence:

[0148] ;

[0149] In the formula, Indicates the scale The spatial feature sequence below; Representing scale Next Spatial feature vectors of a time window; This represents the total number of time windows at this scale. Let be the initial feature dimension of the spatial feature vector;

[0150] Subsequently, the spatial feature sequence The feature vectors in the sequence are mapped through a linear projection layer to adjust the feature dimensions, and positional encoding is introduced to assign temporal labels. The specific calculation formula is as follows:

[0151] ;

[0152] In the formula, This represents the timing input matrix after adding position encoding. The preset feature hiding dimension;

[0153] This is the learnable mapping weight matrix of the linear projection layer, used to adjust the feature dimensions from... Projection mapping to a preset hidden dimension ; This represents a preset position encoding matrix, used to assign the absolute temporal position features corresponding to the spatial feature sequence;

[0154] Next, the position-encoded sequence is input into the matrix. The data is fed into a Transformer encoder, which uses a self-attention mechanism to capture long-range temporal dependencies between windows, ultimately generating a scale-level temporal feature matrix. The mathematical mapping process is as follows:

[0155] ;

[0156] In the formula, This represents the final output scale-level temporal feature matrix, which preserves the temporal length. With hidden dimensions It remains unchanged and incorporates long-range multi-window dynamic fluctuation information across the entire time domain;

[0157] Let the scale-level temporal feature matrix output by the Transformer encoder be... The Middle The row feature vector corresponding to each time window is The importance score of each window feature is calculated using a learnable scoring function in the temporal attention pooling unit. :

[0158] ;

[0159] In the formula, Indicates the scale Next The row feature vector of each time window; and Each represents a learnable weight parameter matrix in the temporal attention pooling unit, where Hidden dimensions for pre-defined attention; and These represent the corresponding learnable bias term parameters; The preset hyperbolic tangent activation function;

[0160] Subsequently, the temporal attention weights for each window are obtained by normalization using the Softmax function. :

[0161] ;

[0162] In the formula, Indicates the first The temporal attention weights for each time window are used to characterize the importance of the current window for subject identification; This is a preset exponential function; The traversal summation operator represents the time series feature score index value, where The index is the window traversal index, which iterates from 1 to the total number of windows at that scale. ;

[0163] The temporal attention weights are used to perform a weighted summation of all window features to obtain a smooth weighted representation of global context awareness. :

[0164] ;

[0165] In the formula, This represents the average pooled feature vector that contains global context information; This represents the window weighted summation operator, where The index is used for window traversal, and its value is traversed from 1 to 1. ;

[0166] Simultaneously, time-dimensional max-pooling units are employed to capture significant transient response patterns and obtain max-pooling feature vectors. :

[0167] ;

[0168] The specific calculation methods for each dimension element are as follows:

[0169] ;

[0170] In the formula, Represents the max-pooling feature vector; This indicates the eigenvector of the eigenvector. Element values ​​in each dimension; Indicated on the time scale Next, the The first row feature vector corresponding to the first time window Each dimension element value, of which The index is the feature dimension index, and its value iterates from 1 to the preset hidden dimension. ; This represents the operator that takes the maximum value along the time window dimension;

[0171] Finally, the feature fusion unit is used to... and Perform channel stitching and fusion to obtain the current time scale. Graph-level mapping representation of the scale-level temporal feature matrix :

[0172] ;

[0173] In the formula, This represents a scale-level temporal feature vector that integrates smooth context and significant transient features; This represents the channel-dimension concatenation operation of feature vectors;

[0174] Among them, such as Figure 5 As shown, cross-scale adaptive fusion includes:

[0175] The scale-level temporal features obtained at different time scales (hereinafter referred to as...) , ,in The data is fed into a cross-scale adaptive fusion module, which projects the temporal features at each scale into query vectors using a learnable mapping matrix. With key vector :

[0176] ;

[0177] In the formula, and These represent the learnable mapping matrices used to project the scale-level temporal features into the query space and the key space, respectively. Representing the Scale-level temporal features at group time scales, the and correspond Figure 5 The feature vector output by the query / key mapping layer corresponds to Figure 5 and the current In the illustrative embodiments, when and At that time, the corresponding scale-level temporal features are abbreviated as follows: and The query vector and key vector obtained by its projection correspond to the following respectively: as well as ;

[0178] Subsequently, attention scores at each scale were calculated using an attention score calculation unit via an additive attention mechanism. :

[0179] ;

[0180] In the formula, This represents the learnable mapping matrix used for score mapping in the additive attention mechanism; The preset hyperbolic tangent activation function; This indicates that the query vector at the corresponding scale will be... With the key vector Performing element-wise addition along the feature dimension, in the current specific implementation, when and Attention scores at each scale are calculated as follows: and ;

[0181] Next, normalization is performed using a Softmax normalization unit to obtain the adaptive fusion weights for each scale. :

[0182] ;

[0183] In the formula, Indicates the first Adaptive fusion weights corresponding to group time scales and These represent the attention scores at the corresponding time scales. This is a preset exponential function; The ergodic summation operator represents the cross-scale attention score index value, where The index is used for scale traversal, and its value is traversed from 1 to 1. For example, when When the denominator corresponds to the traversal summation operator, the concrete representation is as follows: The corresponding adaptive fusion weights are calculated as follows: and ;

[0184] Finally, the cross-scale weighted summation unit performs weighted summation of the temporal features at each scale level based on the adaptive fusion weights, ultimately generating a subject-level brain function representation vector. :

[0185] ;

[0186] In the formula, This represents the final generated subject-level brain function representation vector. Indicates the first Scale-level temporal features at group time scales; Denotes the cross-scale weighted summation operator, where The index is used for scale traversal, and its value is traversed from 1 to 1. This is used to adaptively apply normalized weights to the corresponding time-series features at each scale and perform vector weighted summation, so as to achieve the effect of normalizing weights for any... Adaptive and seamless fusion of temporal features at multiple scales, for example, when At that time, the subject-level brain function representation vector The corresponding concrete expression of the calculation formula is as follows: ;

[0187] S25: Calculate multi-dimensional similarity based on the brain function representation vectors and demographic phenotypic information of all subjects in the training set, and construct a normalized population graph; mine the population association features of the subjects through population graph convolution, and input the population features and phenotypic embedded features into the classifier after fusion through a gating mechanism, and output the category prediction probability; calculate the loss by combining the sample's real label, backpropagate and update all learnable parameters of the model until the model converges;

[0188] like Figure 6 As shown, the brain function representation vector corresponding to each subject in the training set is denoted as... in (total number of subjects in the training set), and the specific processing procedure is as follows:

[0189] S251: Calculate the cosine similarity of brain function among subjects based on the brain function representation vectors of all subjects in the training set. The calculation formula is as follows:

[0190] ;

[0191] In the formula, Indicates the first The subject and the first Cosine similarity feature value of brain functional connectivity among subjects; and They represent the first The subject and the first Brain function representation vectors of each subject; Represents the vector dot product operator; This represents finding the 2-norm of a vector;

[0192] Based on the participants' demographic phenotypic information (collection site, gender, and age), a phenotypic association matrix was constructed. The calculation method for each phenotypic similarity element is as follows:

[0193] ;

[0194] In the formula, and They represent the first The and the first The image data acquisition station identifier corresponding to each subject and These represent their corresponding gender category labels. and These represent their corresponding physical age values ​​(in years); The specific mathematical output rule for the preset indicator function is as follows: the output value is 1 when the conditional logic inside the parentheses is true, and the output value is 0 when the conditional logic inside the parentheses is false. This represents the absolute value operator, where the value 3 indicates that the age difference between the two subjects is within 3 years.

[0195] The brain connectivity similarity matrix and the phenotypic similarity matrices are fused according to preset weights to obtain a comprehensive group similarity matrix. :

[0196] ;

[0197] S252: Add self-loops to the comprehensive group similarity matrix and perform symmetric normalization to generate the group graph adjacency matrix. :

[0198] ;

[0199] In the formula, This represents the adjacency matrix of the generated group graph; This is the preset identity matrix; For matrix The corresponding diagonal degree matrix;

[0200] S253: Concatenate the brain function representation vectors of all subjects into an initial node feature matrix. And perform multi-level graph convolution operation with the group graph adjacency matrix to obtain a group relationship feature vector that integrates group similarity relationships;

[0201] Among them, the The formula for updating the node state features of a layer is:

[0202] ;

[0203] In the formula, and They represent the first Layer and first The group node state feature matrix of the layer; For the first The learnable weight parameter matrix of layer graph convolution. These are the corresponding learnable bias term parameters; Using a pre-defined activation function, residual connections are introduced to preserve the original individual brain network features, resulting in the final group relationship feature matrix. In this embodiment, the total number of convolutional layers in the group graph is preset. The row feature vector corresponding to each subject is denoted as ;

[0204] S254: Using a phenotypic encoder, the raw demographic phenotypic information of the subjects is nonlinearly embedded and mapped through a multilayer perceptron (MLP) to extract high-order phenotypic feature vectors. ; group relationship feature vector With phenotypic feature vectors In the input gating fusion mechanism, the fused comprehensive feature vector is obtained through the following steps:

[0205] First, calculate the adaptive gating weights. :

[0206] ;

[0207] In the formula, This represents the dynamically generated adaptive gating weight value for the current subject; The preset learnable weight vector in the gated fusion mechanism; This is a preset activation function used to constrain the weight values ​​to a certain value. Within the interval:

[0208] Subsequently, the two are adaptively weighted and summed based on the gate weights to obtain the fused comprehensive feature vector:

[0209] ;

[0210] S255: The adaptively fused comprehensive feature vector Input is fed into the classifier for probability prediction:

[0211] ;

[0212] The output is the predicted probability distribution vector after processing by the softmax function. The category corresponding to the one with the highest probability value is the subject's final binary disease classification result, thereby dynamically outputting TD or ASD discrimination labels (TD is a typical developmentally normal control, and ASD is autism spectrum disorder).

[0213] Specifically, during the training phase of the multi-scale hierarchical dynamic graph convolution model, the predicted probability distribution vector of the training set samples is... The network loss value is calculated by inputting the binary labels of the actual disease state into the cross-entropy loss function, and the formula for calculating the cross-entropy loss function is as follows:

[0214] ;

[0215] In the formula, This represents the calculated overall network loss value; The total number of subjects in the training set; Indicates the first The true disease status binary label of each subject (when the subject is ASD) , for TD time ); Represents the predicted probability distribution vector The probability component that predicts the subject's autism spectrum disorder;

[0216] Preferably, the initial learning rate of the network training is set to 0.001, and an exponential decay learning rate adjustment strategy based on LambdaLR is adopted to gradually decay the learning rate from the initial value within a preset number of training rounds. The total number of training rounds of the network is set to 100 rounds. During the backpropagation process, the AdamW optimizer is used to update the gradient of all learnable parameters in the model until the model converges. The network parameters of the trained model are then saved.

[0217] S3: Use the trained multi-scale hierarchical dynamic graph convolutional model to perform functional brain network classification on the individuals to be classified.

[0218] In this step, only forward propagation calculation is performed in one direction throughout the entire process, without involving loss function calculation, gradient backpropagation or parameter update operations.

[0219] Specifically, the process involves using whole-brain time-series data and demographic phenotypic information of a single individual as initial input. Through individual multi-scale feature extraction, dynamic updating and inference of the population graph based on individual features, and feature fusion and result output, the final classification result (a clinical auxiliary diagnostic result for binary disease classification) is obtained. The classification steps are briefly described below:

[0220] (1) Single-person multi-scale feature extraction of the individual to be diagnosed

[0221] Obtain the whole brain region time series matrix and personal phenotypic information (including gender, age, and data collection site) of the individual to be classified.

[0222] First, the whole brain region time series matrix of the subject was divided into sliding windows using preset short windows (scale 1) and long windows (scale 2). The correlation matrix within each window was calculated and Fisher Z transform and positive and negative channel splitting were performed.

[0223] Next, the window-based temporal attention aggregation module is used to perform weighted aggregation to obtain a pair of aggregated positive connection graphs for each window. and aggregation negative connection graph ;

[0224] Subsequently, the aggregated graph is input into the pre-trained dual-branch spatial graph feature extraction module. Through independent graph convolution branches, ROI-level graph convolution, residual mapping, supernode pooling, supernode graph convolution, and graph readout operations are performed. The output spatial feature vectors are then concatenated to obtain the window-level spatial feature vectors for each subject within each time window. ;

[0225] Finally, the window-level spatial features are arranged into a temporal sequence in chronological order. Long-range temporal dependencies are captured by a Transformer encoder (503), and temporal attention pooling and temporal max pooling are used to extract scale-level temporal features at different scales. These scale features are then fed into a cross-scale adaptive fusion submodule, where scale weights are adaptively adjusted and weighted summation is performed. This process ultimately generates the individual brain function representation vector for the single individual to be classified, denoted as... ;

[0226] (2) Dynamic updating and reasoning of group graphs based on individual characteristics

[0227] Since the disease discrimination classification and network optimization update of this invention utilizes a population graph convolutional network to capture macroscopic similarity relationships among subjects, the following dynamic population inference strategy is adopted in order to perform population graph convolution even when only a single individual to be classified is input:

[0228] The brain function representation vector of the individual to be classified The global feature matrix of the subjects in the training set (containing) saved during the model training phase in step S2. Feature vectors of subjects The sample dimensions are concatenated to form a sample containing... Temporary group feature matrix of nodes , of which The row corresponds to the individual currently being classified;

[0229] The cosine similarity of brain connectivity between the individual to be classified and each known subject in the training set is recalculated. Combined with phenotypic information, a phenotypic association indicator function value is calculated to generate a dimensionless dataset containing the individual to be classified. The comprehensive group similarity matrix is ​​obtained, a self-loop is added to the comprehensive group similarity matrix, and symmetric normalization is performed. The group graph adjacency matrix is ​​then dynamically updated, denoted as . ;

[0230] (3) Feature fusion and autism auxiliary diagnosis result output

[0231] The updated adjacency matrix of the group graph With temporary group feature matrix The inputs are fed into a pre-trained group graph convolutional module. During forward propagation, the module utilizes pre-defined graph convolutional weight parameters for multi-layer feature extraction and residual connections. This adaptively fuses the feature vector of the individual to be classified with global contextual information from neighboring subjects with similar phenotypes and brain connectivity, outputting an updated group relationship feature vector. The row vectors belonging to the current individual to be classified are then extracted, denoted as... ;

[0232] Simultaneously, the original demographic phenotypic information of the individual to be classified is input into the trained phenotypic encoder for nonlinear embedding mapping, outputting its high-order phenotypic feature vector. ;

[0233] Will and The input gating fusion mechanism calculates adaptive gating weights using fixed gating weight parameters. The comprehensive feature vector of the subject is obtained by adaptive weighted summation. ;

[0234] Finally, The raw score is obtained by inputting the data into a multilayer perceptron classifier and calculating it through forward propagation. Then, the probability is normalized using the softmax function, outputting the predicted probability values ​​of the individual to be classified as belonging to the typical developmental normal control (TD) and the autism spectrum disorder (ASD):

[0235] ;

[0236] By comparing the probability values, the disease category corresponding to the highest probability value is the final clinical auxiliary diagnostic result for that individual subject. If... If the system outputs that the subject has a tendency to be diagnosed with autism spectrum disorder, then the system will output that the subject has a tendency to be diagnosed with autism spectrum disorder.

[0237] Validated on multiple public datasets such as ABIDE, the method proposed in this invention achieves superior classification accuracy and generalization performance compared to existing methods in the autism assisted diagnosis task.

[0238] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The apparatus of the present invention can employ a processor of a general-purpose computer, a dedicated medical testing computer, an embedded processor, or other programmable data processing device to execute instructions to implement the functional means specified in the various sub-steps of the present invention.

[0239] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A functional brain network classification method integrating multi-scale hierarchical dynamic graph convolution, characterized in that, include: S1: Construct a sample dataset and divide it into a training set, a validation set, and a test set according to a preset ratio. Each sample in the sample dataset contains input data and label data. The input data is the time series matrix of the whole brain regions of an individual and demographic phenotypic information. The label data is a functional brain network classification label. S2: Construct a multi-scale hierarchical dynamic graph convolutional model for functional brain network classification, and train the model parameters using the training set. Training the multi-scale hierarchical dynamic graph convolutional model includes the following steps: S21: Construct a multi-scale dual-channel dynamic functional connectivity map sequence based on the time series matrix of the whole brain regions of an individual, which can simultaneously characterize the dynamic change pattern of the network at different time granularities. The dual-channel dynamic functional connectivity map sequence at any scale includes a positive connectivity map sequence and a negative connectivity map sequence. S22: Perform temporal attention weighted aggregation on the positive connection subframe graph sequence and the negative connection subframe graph sequence corresponding to each sliding window at the same scale to obtain window-level aggregated positive connection graph and aggregated negative connection graph; S23: A symmetrical and parameter-independent bi-branch graph convolutional network is used to perform hierarchical feature mapping from the brain region node level to the topological supernode level on the aggregated positive connection graph and the aggregated negative connection graph, and to fuse and construct a window-level spatial feature vector that takes into account both positive and negative connection modes. S24: Perform hierarchical temporal processing on the window-level spatial feature vector, aggregate temporal dimension fluctuation information and fuse it into a single and complete individual-level global representation of brain function. The hierarchical temporal processing includes: single-scale temporal modeling and cross-scale adaptive fusion. S25: Calculate multi-dimensional similarity based on the brain function representation vectors and demographic phenotypic information of all individuals in the training set, and construct a normalized group graph; mine individual group association features through group graph convolution, and input the group features and phenotypic embedded features into the classifier after fusion through a gating mechanism, and output the class prediction probability; calculate the loss by combining the sample's real label, backpropagate and update all learnable parameters of the model until the model converges; S3: Use the trained multi-scale hierarchical dynamic graph convolutional model to perform functional brain network classification on the individuals to be classified.

2. The functional brain network classification method incorporating multi-scale hierarchical dynamic graph convolution as described in claim 1, characterized in that, In S1, each element in the whole-brain time series matrix represents the blood oxygenation level-dependent signal intensity of different brain regions of an individual at different times; the demographic phenotypic information includes the individual's gender, age, and data collection site information.

3. The functional brain network classification method incorporating multi-scale hierarchical dynamic graph convolution as described in claim 1, characterized in that, In S1, the steps for obtaining the time series matrix of an individual's whole brain regions are as follows: Resting-state scans of individuals are performed using functional magnetic resonance imaging (fMRI) to obtain raw image data of the whole brain's functional connectivity. The original image data is preprocessed, and a whole-brain time series matrix is ​​extracted based on a preset AAL map. The time series matrix of the whole brain regions is subjected to standardized preprocessing, which includes linear detrending, bandpass filtering, and Z-score normalization.

4. The functional brain network classification method incorporating multi-scale hierarchical dynamic graph convolution as described in claim 1, characterized in that, In S21, the method for constructing the multi-scale dual-channel dynamic function connectivity graph sequence is as follows: For each individual's whole-brain time series, at least two different time scales are preset; For any time scale, the time series of an individual's whole brain regions is segmented using sliding windows, and the total number of sliding windows is calculated. The whole-brain brain region time series within each sliding window at each scale is divided into sub-sliding windows to obtain the sub-frames corresponding to each sliding window. Based on the time series segments of the whole brain regions corresponding to the subframes, the Pearson correlation coefficients between each pair of brain regions are calculated, the subframe correlation matrix is ​​constructed, and preprocessing and Fisher Z transform are performed to obtain the subframe transformation matrix. The subframe transformation matrix is ​​split into positive and negative connected subframe maps according to the sign of the elements. All windows at each scale are traversed to finally form a multi-scale dual-channel dynamic function connection map sequence.

5. The functional brain network classification method fusion of multi-scale hierarchical dynamic graph convolution as described in claim 4, characterized in that, The preprocessing of the subframe correlation matrix includes: diagonal zeroing, outlier replacement, and numerical clipping.

6. The functional brain network classification method incorporating multi-scale hierarchical dynamic graph convolution as described in claim 1, characterized in that, In S22, the temporal attention weighted aggregation of the positive or negative connected subframe sequence corresponding to the sliding window at a single scale includes the following steps: For the single scale, calculate the intensity score of each frame of the motion graph within the sliding window, wherein the motion graph is a positively connected subframe graph or a negatively connected subframe graph; Based on the intensity score of each frame, the temporal attention weights corresponding to each frame are calculated through Softmax normalization. The dynamic graph within the sliding window is weighted and summed according to the time-dimensional attention weights to obtain a window-level aggregated positive connection graph or aggregated negative connection graph.

7. The functional brain network classification method incorporating multi-scale hierarchical dynamic graph convolution as described in claim 1, characterized in that, In S23, the process of constructing the window-level spatial feature vector is as follows: The graph-level feature vectors corresponding to the aggregated positive connectivity graph and the aggregated negative connectivity graph are obtained respectively, and then concatenated by channels to obtain the window-level spatial feature vector; the method for obtaining the graph-level feature vector is as follows: Add self-connections to aggregated positive or negative connection graphs to construct an enhanced adjacency matrix. After calculating the degree matrix based on the enhanced adjacency matrix, multi-layer symmetric normalized graph convolution is performed, and residual connections are introduced to obtain the enhanced residual node feature matrix. Based on the enhanced residual node feature matrix, a learningable soft assignment network is used to generate the mapping relationship between brain region nodes and super nodes, thus obtaining the super node soft assignment matrix. Based on the supernode soft assignment matrix, the original brain region nodes are aggregated into supernodes, resulting in a supernode feature matrix and a supernode adjacency matrix. The symmetric normalized graph convolution operation is performed again at the supernode level to capture the spatial interaction patterns between higher-order functional modules and obtain the output state feature matrix. By utilizing a graph readout mechanism that combines global average pooling and global max pooling, the supernode feature matrix is ​​aggregated along the node dimension into a graph-level feature vector of fixed dimension.

8. The functional brain network classification method incorporating multi-scale hierarchical dynamic graph convolution as described in claim 1, characterized in that, In S24, the single-scale temporal modeling is as follows: all window-level spatial features at the same scale are arranged temporally to form a feature sequence, which is then input into a Transformer encoder after linear projection and positional encoding to capture long-range temporal dependencies between windows; contextual smoothing features are obtained through temporal attention pooling, and transient salient features are extracted through temporal max pooling; the two types of features are concatenated to obtain the scale-specific temporal features; the cross-scale adaptive fusion is as follows: the temporal features at each scale are projected into query vectors and key vectors respectively, and the adaptive fusion weights of each scale are calculated through an additive attention mechanism; the temporal features at each scale are weighted and summed according to the adaptive fusion weights to generate an individual-level brain function representation vector.

9. The functional brain network classification method incorporating multi-scale hierarchical dynamic graph convolution as described in claim 1, characterized in that, S25 specifically includes: The cosine similarity of brain function between individuals is calculated based on the brain function representation vectors of all individuals in the training set; a phenotypic association matrix is ​​constructed based on the demographic phenotypic information of individuals; the brain connectivity similarity matrix and each phenotypic similarity matrix are fused according to preset weights to obtain a comprehensive group similarity matrix. Add self-loops to the comprehensive group similarity matrix and perform symmetric normalization to generate a group graph adjacency matrix; The brain function representation vectors of all individuals are concatenated into an initial node feature matrix, and then multi-layer residual graph convolution operation is performed with the adjacency matrix of the population graph to obtain the population relationship feature vector. Demographic phenotypic information is mapped into phenotypic feature vectors using a multilayer perceptron; group relationship feature vectors and phenotypic feature vectors are input into a gating fusion mechanism to obtain a fused comprehensive feature vector. The fused integrated feature vector is then input into the classifier for probability prediction.

10. The functional brain network classification method incorporating multi-scale hierarchical dynamic graph convolution as described in claim 1, characterized in that, During the training phase of the multi-scale hierarchical dynamic graph convolutional model, the cross-entropy loss function is used to calculate the network loss value.

Images