Brain disease prediction method and system fusing amplitude-phase information and image perception mixed experts

Abstract

This invention provides a method and system for predicting brain diseases by fusing amplitude and phase information with graph-aware hybrid expert graph neural networks. It relates to the field of neuroimaging analysis technology. The method includes: acquiring raw resting-state functional magnetic resonance imaging (fMRI) data of a subject and corresponding brain disease category labels; preprocessing the raw resting-state fMRI data; extracting the mean oxygenation level dependent signal time series of multiple brain regions of the subject and using the mean oxygenation level dependent signal time series as the signal time series; constructing a functional connectivity matrix and a phase adjacency matrix based on the signal time series; constructing a training dataset; constructing a brain disease prediction model based on a two-branch hybrid expert graph neural network; training the brain disease prediction model using the training dataset; acquiring resting-state fMRI data; inputting the resting-state fMRI data into the trained brain disease prediction model for prediction, and outputting the brain disease prediction result.

Description

Technical Field

[0001] This invention relates to the field of neuroimaging analysis technology, and in particular to a method and system for predicting brain diseases by fusing amplitude and phase information with image perception hybrid experts. Background Technology

[0002] Brain functional network analysis is a core technological direction for the diagnosis and pathological mechanism research of brain diseases. Resting-state functional magnetic resonance imaging (fMRI), with its advantages of being non-invasive and having high spatial resolution, has become an important means of capturing the correlation patterns of neural activity in brain regions. By constructing functional connectivity matrices between brain regions and exploring features such as amplitude synchronicity and phase coherence, key imaging evidence can be provided for the early screening and accurate diagnosis of brain diseases such as Alzheimer's disease and schizophrenia.

[0003] In existing technologies, brain disease prediction methods based on graph neural networks are widely used. These methods typically preprocess resting-state functional magnetic resonance imaging (fMRI) data to extract temporal features of brain region blood oxygenation level-dependent signals. Then, a single-dimensional functional connectivity matrix is ​​constructed as the graph topology input, combined with a graph convolutional network to mine brain region connectivity features. Finally, a classifier is used to achieve disease prediction. Some methods introduce a hybrid expert network architecture, which improves the model's ability to represent complex features through multiple parallel expert modules, and uses a cross-entropy loss function to optimize model parameters to improve classification accuracy.

[0004] However, existing technologies suffer from shortcomings such as one-sided mining of brain functional signal features and coarse modeling of topological structures. Feature representations suffer from single-dimensionality and fragmented correlations, resulting in low accuracy in capturing brain region interaction patterns. Furthermore, model optimization only uses classification loss as the core objective and does not incorporate key constraints such as the balance of expert module scheduling and the adaptability of branch feature fusion, which can easily lead to model overfitting and insufficient generalization ability. Summary of the Invention

[0005] To address the shortcomings of existing technologies, such as one-sided mining of brain functional signal features and coarse topological modeling, the single-dimensionality and fragmented correlation of feature representations, which result in low accuracy in capturing brain region interaction patterns; and the fact that model optimization only uses classification loss as the core objective and does not incorporate key constraints such as the balance of expert module scheduling and the adaptability of branch feature fusion, which easily leads to overfitting and insufficient generalization ability, this invention provides a brain disease prediction method and system that integrates amplitude and phase information with graph perception hybrid experts.

[0006] The technical solutions provided by the embodiments of the present invention are as follows: The first aspect of this invention provides a brain disease prediction method that integrates amplitude and phase information with graph perception hybrid experts, comprising: S1: Obtain the subject's raw resting-state functional magnetic resonance imaging data of the brain and the corresponding brain disease category labels; S2: Preprocessing of raw resting-state functional magnetic resonance imaging (fMRI) data of the brain; S3: Based on the preprocessed raw resting-state functional magnetic resonance imaging data, the mean blood oxygenation level dependent signal time series of multiple brain regions of the subjects was extracted; S4: Construct the functional connectivity matrix and phase adjacency matrix based on the signal time series; S5: Construct a training dataset based on the signal time series, functional connectivity matrix, phase adjacency matrix, and brain disease category labels; S6: Construct a brain disease prediction model based on a dual-branch hybrid expert graph neural network; S7: Train the brain disease prediction model using the training dataset; S8: Acquire resting-state functional magnetic resonance imaging (fMRI) data of the brain; S9: After processing the resting-state functional magnetic resonance imaging (fMRI) data as described above, input the data into the trained brain disease prediction model for prediction and output the brain disease prediction results.

[0007] A second aspect of this invention provides a method and system for predicting brain diseases by fusing amplitude and phase information with graph perception hybrid experts, comprising: processor; A memory storing computer-readable instructions, which, when executed by the processor, implement the brain disease prediction method as described in the first aspect, which integrates amplitude and phase information with graph-aware hybrid experts.

[0008] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the brain disease prediction method as described in the first aspect, which integrates amplitude and phase information with graph perception hybrid experts.

[0009] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this embodiment of the invention, by constructing a functional connectivity matrix and a phase adjacency matrix, the topological connectivity of brain regions is fully characterized from two dimensions: amplitude synchronicity and phase coordination. This enables comprehensive mining of brain functional signal features and refined modeling of topological structures, breaking the limitation of single feature dimensions. By constructing a brain disease prediction model based on a bi-branch hybrid expert graph neural network, the deep association and fusion of the two types of features are facilitated, strengthening the intrinsic connection between features. By training the model using a training dataset, the model optimization constraint dimensions can be enriched, improving the model's fitting effect and generalization performance. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating a brain disease prediction method that integrates amplitude and phase information with graph perception, provided in an embodiment of the present invention.

[0012] Figure 2 This is a diagram of a dual-branch hybrid expert graph neural network architecture provided in an embodiment of the present invention.

[0013] Figure 3 This is a flowchart of a hybrid expert graph isomorphic network provided in an embodiment of the present invention.

[0014] Figure 4 This is a schematic diagram of the structure of a brain disease prediction method and system that integrates amplitude and phase information with graph perception, provided in an embodiment of the present invention. Detailed Implementation

[0015] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0016] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0017] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0018] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0019] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0020] Reference manual attached Figure 1 The diagram illustrates a flowchart of a brain disease prediction method that integrates amplitude and phase information with graph perception, provided by an embodiment of the present invention.

[0021] This invention provides a brain disease prediction method that integrates amplitude-phase information and image-aware hybrid expert knowledge. This method can be implemented by a brain disease prediction device based on amplitude-phase information fusion and image-aware hybrid expert knowledge, which can be a terminal or a server. The processing flow of the brain disease prediction method integrating amplitude-phase information and image-aware hybrid expert knowledge can include the following steps:

[0022] S1: Obtain the subject's raw resting-state functional magnetic resonance imaging (fMRI) data and corresponding brain disease category labels.

[0023] S2: Preprocessing of raw resting-state functional magnetic resonance imaging (fMRI) data of the brain.

[0024] It should be noted that the raw resting-state functional magnetic resonance imaging (fMRI) data undergoes a standard preprocessing procedure, including temporal correction, head motion correction, spatial normalization, denoising, smoothing, and bandpass filtering.

[0025] In this embodiment of the invention, a standard preprocessing procedure including temporal correction, head movement correction, spatial normalization, denoising, smoothing, and bandpass filtering is performed on the raw resting-state functional magnetic resonance imaging (fMRI) data. This effectively eliminates non-physiological noise and systematic biases introduced during data acquisition due to factors such as equipment hardware characteristics, unconscious head movements of subjects, and individual differences in brain structure. At the same time, smoothing improves the signal-to-noise ratio, and bandpass filtering preserves low-frequency fluctuation signals related to neural activity. Ultimately, standardized and highly reliable brain functional signal data is obtained, laying a solid data foundation for subsequent extraction of brain region amplitude and phase connectivity features and construction of accurate graph neural network analysis models.

[0026] S3: Based on the preprocessed raw resting-state functional magnetic resonance imaging (fMRI) data, the mean blood oxygenation level dependent signal time series of multiple brain regions of the subjects was extracted.

[0027] It should be noted that this is based on predefined brain atlas extraction. N The average blood oxygen level of a region of interest depends on the signal time series, thus obtaining the signal time series.

[0028] In this embodiment of the invention, brain maps are extracted from preprocessed resting-state functional magnetic resonance imaging data based on predefined brain atlases. NThe average blood oxygen level of each region of interest is dependent on the signal time series, which can focus continuous signals at the whole brain level on the core brain regions with anatomical significance, effectively reducing the redundancy and computational complexity of whole brain signals. At the same time, averaging the signals in each region of interest can further reduce local noise interference within the region, improve the stability and interpretability of the signal, and the final signal time series can accurately reflect the dynamic changes of neural activity in each brain region, providing direct and high-quality data support for subsequent calculations of amplitude synchronization and phase coordination between brain regions and the construction of adjacency matrices.

[0029] S4: Construct the functional connection matrix and phase adjacency matrix based on the signal time series.

[0030] The functional connectivity matrix is ​​a topological matrix that characterizes the connectivity of different brain regions in the brain functional network based on the synchronicity of signal amplitude fluctuations. The number of rows and columns of the matrix is ​​consistent with the total number of selected brain regions. The linear correlation between the time series of functional magnetic resonance imaging (fMRI) signals of any two brain regions is calculated (commonly using the Pearson correlation coefficient). The result is used as the element value of the corresponding position in the matrix, and then thresholded or normalized to ensure that the element values ​​are distributed between 0 and 1.

[0031] The phase adjacency matrix is ​​a topological matrix that characterizes the connection between different brain regions in the brain functional network based on signal phase coordination. The number of rows and columns of the matrix is ​​consistent with the total number of selected brain regions. First, the instantaneous phase information of the functional magnetic resonance imaging (fMRI) signals of each brain region is extracted using signal processing methods such as Hilbert transform. Then, the cosine mean of the instantaneous phase difference between any two brain regions is calculated as the phase synchronization value. The obtained result is used as the element value of the corresponding position in the matrix. After thresholding or normalization, the element values ​​are distributed between 0 and 1.

[0032] In one possible implementation, S4 specifically includes sub-steps S401 and S402: S401: Construct the functional connection matrix and phase adjacency matrix respectively based on the signal time series.

[0033] The functional connectivity matrix is ​​a brain region connectivity topology matrix constructed based on the amplitude fluctuation synchronicity of functional magnetic resonance imaging (fMRI) signals. The matrix shape is the product of the number of brain regions and the matrix itself, and its core function is to quantify the degree of synergistic correlation in signal intensity changes between different brain regions. Each element in the matrix represents the amplitude synchronization strength between two corresponding brain regions. Its value is obtained by calculating the correlation coefficient of the signal time series of each brain region and then thresholding or normalizing it. The value typically ranges from 0 to 1. A larger value indicates more synchronized amplitude fluctuations between the two brain regions and a tighter functional connection; a smaller value indicates weaker amplitude synchronicity and a sparser functional connection.

[0034] The phase adjacency matrix is ​​a brain region connectivity topology matrix constructed based on the phase synergy of functional magnetic resonance imaging (fMRI) signals. The matrix shape is the product of the number of brain regions and the matrix itself, and its core function is to quantify the degree of synchronization between signal phase changes between different brain regions. Each element in the matrix represents the phase synchronization strength between two corresponding brain regions. Its value is obtained by extracting the phase information of the signals from each brain region, calculating the cosine mean of the instantaneous phase difference, and then thresholding or normalizing it. The value typically ranges from 0 to 1. A larger value indicates more synchronized signal phase changes and a closer functional interaction between the two brain regions; a smaller value indicates weaker phase synergy and a sparser functional interaction.

[0035] In one possible implementation, constructing the functional connectivity matrix based on the signal time series in S4 specifically includes: S401A: Calculate the Pearson correlation coefficient between any two brain regions based on the signal time series.

[0036] The Pearson correlation coefficient is a statistical indicator used to measure the degree of linear correlation between two continuous variables, with a value ranging from -1 to 1. The core of its calculation is the normalization of the covariance of the two variables to eliminate the influence of differences in their dimensions. Specifically, standardization is achieved by using the product of the standard deviations of the two variables as the denominator. When the coefficient value approaches 1, it indicates a strong positive linear correlation between the two variables, meaning that an increase in one variable significantly increases the other. When the coefficient value approaches -1, it indicates a strong negative linear correlation, meaning that an increase in one variable significantly decreases the other. When the coefficient value approaches 0, it indicates that there is no significant linear correlation between the two variables.

[0037] S402A: Construct an initial matrix based on the Pearson correlation coefficient, and perform binarization and graph-symmetric normalization to construct the functional connectivity matrix.

[0038] Specifically, the functional connection matrix is ​​as follows: in, Represents the functional connection matrix. i and j All represent brain region indexes. r ij Indicates the first i The brain regions and the first j Pearson correlation coefficient between brain regions T Indicates the total time. t Indicates time, x i,t Indicates the first i brain regions in t The signal value at time [time] Indicates the firsti Time mean of each brain region x j,t Indicates the first j brain regions in t The signal value at time [time] Indicates the first j The average time for each brain region.

[0039] In one possible implementation, constructing the phase adjacency matrix based on the signal time series in S4 specifically includes: S401B: Performs narrowband filtering of 0.03-0.07Hz on the signal time series to obtain the narrowband filtered signal.

[0040] It should be noted that narrowband filtering (0.03-0.07Hz) of a specific frequency band is applied to the signal time series to separate specific blood oxygenation level-dependent signal information. This frequency band belongs to the core region of slow fluctuations in the original resting-state functional magnetic resonance imaging data, and applying Hilbert transform within this frequency band can better meet the signal coherence requirements. To ensure the physical meaning of phase synchronization calculation, a single narrowband must be extracted.

[0041] S402B: Based on the narrowband filtered signal, the instantaneous phase of each brain region is obtained through Hilbert transform.

[0042] The Hilbert transform is a linear integral transform widely used in signal processing. Its core function is to extract the corresponding analytic signal from a real-valued signal, thereby obtaining key features such as the instantaneous amplitude, instantaneous frequency, and instantaneous phase of the original signal. This transform generates an imaginary part signal orthogonal to the original signal by performing a specific convolution operation on the real-valued time series. Combining the two yields the analytic signal; the argument of the analytic signal corresponds to the instantaneous phase of the original signal, and the magnitude corresponds to the instantaneous amplitude of the original signal.

[0043] S403B: Based on the instantaneous phase, calculate the mean cosine of the instantaneous phase difference between any two brain regions as the phase synchronization value.

[0044] S404B: Construct a phase adjacency matrix based on the phase synchronization value and perform graph normalization.

[0045] Specifically, the phase adjacency matrix is ​​as follows: in, Represents the phase adjacency matrix. i and j All represent brain region indexes. T Indicates the total time. t Indicates time, cos Represents the cosine function. Indicates the firsti Instantaneous phase of each brain region Indicates the first j The instantaneous phase of each brain region.

[0046] S402: Perform adjacency matrix processing and normalization on the functional connection matrix and the phase adjacency matrix respectively to construct the functional connection matrix and the phase adjacency matrix.

[0047] Specifically, a threshold is applied to the functional connectivity matrix. λ ( λ =0.5), generating the final functional connection matrix; processing the phase adjacency matrix, setting the negative correlation values ​​to zero ( ), thus obtaining the phase adjacency matrix.

[0048] Furthermore, graph-symmetric normalization is performed on the functional connectivity matrix and the phase adjacency matrix respectively: in, Represents the normalized adjacency matrix. This represents an adjacency matrix with self-connections. This represents a degree matrix with self-connections. A Represents the adjacency matrix of primitive brain functions. I Represents the identity matrix.

[0049] In this embodiment of the invention, amplitude and phase adjacency matrices are first constructed based on the signal time series. The Pearson correlation coefficient is used to accurately quantify the amplitude synchronization of brain region signals. Narrowband filtering and Hilbert transform are used to accurately capture the phase coordination of the core frequency band. Then, the two types of matrices are processed and normalized. This not only fully characterizes the complex topological pattern of brain region functional connections from both amplitude and phase dimensions, making up for the lack of information in single-dimensional features, but also removes weak correlation noise and unifies the data scale through processing and normalization. This ensures that the matrix is ​​compatible with the input requirements of the subsequent dual-branch hybrid expert graph neural network, providing standardized and high-quality topological data support for the model to efficiently mine brain region amplitude-phase coupling features.

[0050] S5: Construct a training dataset based on the signal time series, functional connectivity matrix, phase adjacency matrix, and brain disease category labels.

[0051] The training dataset specifically includes a training set and a test set.

[0052] In this embodiment of the invention, by integrating signal time series, functional connectivity matrix, phase adjacency matrix, and brain disease category labels to construct a training dataset, the time-frequency dynamic information, amplitude synchronization topological features, and phase coordination topological features of brain region signals can be accurately associated with disease diagnostic labels. This ensures the comprehensiveness and complementarity of the dataset input features and provides the model with clear supervised learning objectives. At the same time, the standardized dataset construction method can effectively reduce the interference of individual differences and data noise on model training, laying a data foundation for the efficient training and parameter optimization of the dual-branch hybrid expert graph neural network, and helping the model learn more discriminative brain disease association features.

[0053] Reference manual attached Figure 2 The diagram illustrates a dual-branch hybrid expert graph neural network architecture provided by an embodiment of the present invention.

[0054] It should be noted that, firstly, based on the preprocessed functional magnetic resonance imaging (fMRI) data, three sets of core input data are constructed in parallel: firstly, the time series of mean oxygenation level dependent (BOLD) signals of each brain region are extracted; secondly, the Pearson correlation coefficient between brain regions is calculated and normalized to construct an amplitude functional connectivity matrix reflecting the synchronization intensity; and thirdly, the mean cosine of the phase difference is calculated using narrowband filtering and Hilbert transform to construct a phase adjacency matrix reflecting rhythmic coordination.

[0055] Further, in the feature processing stage, the signal time series is input into the time-frequency feature encoding branch. Through a one-dimensional convolutional flow, adaptive one-dimensional average pooling, and a linear mapping layer, time-domain node features are extracted. Simultaneously, Fourier operators are used to capture global asynchronous interaction information in the frequency domain. After dynamic fusion via a gating mechanism, shared node features are output. Subsequently, the dual-branch hybrid expert graph learning stage begins: the shared node features are paired with the amplitude functional connectivity matrix and the phase adjacency matrix, respectively, and input to two parameter-independent hybrid expert graph isomorphic network layers (MoE-GIN). Through its internal graph-aware gating mechanism and maximum value mask routing, nodes with different topological attributes are distributed to the optimal expert operator, extracting high-order amplitude branch graph representations and phase branch graph representations, respectively. Finally, a bidirectional cross-attention mechanism is used to achieve deep coupling and complementary enhancement of the two branch graph representations. The generated final fused features are mapped by a multilayer perceptron (MLP) to output a classification prediction score for brain diseases.

[0056] S6: Construct a brain disease prediction model based on a bi-branch hybrid expert graph neural network.

[0057] The dual-branch hybrid expert graph neural network is a modular graph learning architecture for analyzing brain functional connectivity topology. Specifically designed for the complementary brain connectivity topology information of amplitude and phase, it comprises two parallel sub-networks: a structurally symmetrical but parameter-independent amplitude branch and a phase branch. Both branches employ a hybrid expert graph neural network structure, incorporating a graph-aware gating network and multiple graph isomorphic network expert modules. The gating network adaptively assigns expert weights to each brain region node based on the shared node features and the normalized adjacency matrix of the corresponding branch. Each expert module independently extracts topological features based on node neighborhood information and then generates a branch-specific graph representation through weighted fusion. The amplitude branch uses the functional connectivity matrix as topological input, focusing on learning the synchronous correlation patterns of brain region signal intensity, while the phase branch uses the phase adjacency matrix as topological input, focusing on capturing the collaborative interaction relationships of brain region signal phases. Both branches also simultaneously output balanced and sparse losses to optimize the scheduling mechanism of the expert modules.

[0058] In this embodiment of the invention, a brain disease prediction model based on a dual-branch hybrid expert graph neural network is constructed. It can rely on the synergistic advantages of the amplitude and phase dual-branch parallel architecture and the hybrid expert mechanism. It can achieve fine mining of two types of complementary topological features by focusing on the synchronous correlation mode of brain region signal intensity and the collaborative interaction mode of phase through the structurally symmetrical but parameter-independent dual branches respectively. Furthermore, it can adaptively assign expert weights to each brain region node by using a graph-aware gating network, so that the model can flexibly call the appropriate expert module to extract core features according to the topological characteristics of brain connectivity. At the same time, the introduction of balanced loss and sparsity loss further optimizes the scheduling mechanism of expert modules, effectively improving the model's ability to represent complex patterns of brain functional connectivity and the efficiency of feature extraction. Ultimately, it provides high-performance and highly interpretable algorithmic support for the accurate prediction of brain diseases.

[0059] S7: Use the training dataset to train the brain disease prediction model.

[0060] In one possible implementation, S7 specifically includes sub-steps S701 to S707: S701: Input the signal time series from the training dataset into the time-frequency feature encoder to obtain shared node features.

[0061] The time-frequency feature encoder is a feature extraction module for functional magnetic resonance imaging (fMRI) signals of the brain. Its core function is to simultaneously mine the temporal dynamics and frequency fluctuation patterns of the signal, generating fused features that combine temporal variation patterns and frequency feature expressions. This encoder typically includes two parallel branches: time-domain feature extraction and frequency-domain feature extraction. The time-domain branch directly processes the time series of blood oxygenation level-dependent signals in the brain region, capturing dynamic features such as amplitude fluctuations and trend changes over time. The frequency-domain branch transforms the time-domain signal to the frequency domain using methods such as Fourier transform, extracting key features such as energy distribution and dominant frequency components in different frequency bands. Subsequently, a feature fusion strategy integrates the outputs of the two branches to ultimately generate a unified high-dimensional feature vector.

[0062] In one possible implementation, S701 specifically includes steps S7011 to S7019: S7011: Based on the signal time series, local temporal dynamic features are extracted through a one-dimensional convolutional neural network module.

[0063] The one-dimensional convolutional neural network module is a feature extraction unit for one-dimensional sequential data (such as time series of functional magnetic resonance imaging (fMRI) signals). It captures local temporal dependencies and dynamic changes in the sequence data through a one-dimensional convolutional kernel that slides along the time dimension. The convolution operation of this module is performed only in a single dimension, effectively extracting local patterns such as amplitude changes and trend fluctuations within different time windows. Simultaneously, pooling operations reduce the feature dimensionality while preserving key information, and activation functions introduce non-linear expressive power. Its network structure is lightweight and computationally efficient, requiring no complex dimensional transformations of the sequence data, making it highly suitable for processing one-dimensional time-series signals like fMRI.

[0064] Specifically, the local temporal dynamic features are as follows: in, H time Represents local time-domain dynamic characteristics. TimeCNN Represents a temporal convolutional neural network. This represents the signal time series of the training dataset.

[0065] S7012: Based on local temporal dynamic features, temporal node features are obtained through adaptive one-dimensional average pooling and linear mapping layers; Among them, adaptive one-dimensional average pooling automatically adjusts the division method and sampling interval of the input local temporal dynamic features according to the preset output feature dimension. By calculating the mean of each adaptively divided time segment, it realizes the standardized dimensional compression of time series features of different lengths. This can effectively reduce the feature dimension to reduce the computational load of subsequent models, while retaining the global fluctuation trend and key dynamic information of the time series signal in the time dimension. It avoids feature redundancy or loss of key time domain patterns caused by fixed pooling windows, and provides high-quality time domain features with unified dimensions and condensed information for subsequent time-frequency feature fusion.

[0066] S7013: By using the word embedding layer, the signal time series is flattened and mapped to the word embedding dimension to obtain the embedding features.

[0067] The word embedding layer is a core feature transformation module in natural language processing models. Its core function is to map discrete lexical symbols (such as words and characters) into continuous, low-dimensional, dense vectors, solving the problems of dimensionality explosion and semantic information loss in traditional one-hot encoding. Through model training or pre-training, the distance between vectors corresponding to semantically similar words is made closer, while the distance between vectors corresponding to words with large semantic differences is made farther, thus giving the vector space clear semantic association attributes. Its input is usually a sequence of word indices, and the output is a fixed-dimensional embedding vector matrix, which preserves the contextual features of words and reduces the computational complexity of subsequent models.

[0068] Specifically, the embedding features are: in, X amb Represents embedded features, Represents word embedding layer.

[0069] S7014: Apply Fast Fourier Transform to the embedded features to obtain Fast Fourier Transform features.

[0070] The Fast Fourier Transform (FFT) is an efficient implementation algorithm for the Discrete Fourier Transform (DFT). Its core function is to rapidly convert a one-dimensional discrete signal in the time domain into a signal in the frequency domain. It also supports the inverse conversion from the frequency domain to the time domain. Its key advantage lies in significantly reducing computational complexity through a divide-and-conquer strategy, thus simplifying the traditional DFT. Computational complexity optimized to ( N The number of signal sampling points can significantly improve the efficiency of signal processing. By breaking down a long sequence signal into multiple short sequences for processing, and then integrating the transformation results of each segment to obtain complete frequency domain features, key information such as the frequency components and energy distribution of each frequency band can be accurately extracted.

[0071] The specific features of the Fast Fourier Transform are as follows: in, X This represents the characteristics of the Fast Fourier Transform. F This represents the Fast Fourier Transform.

[0072] S7015: Based on the Fast Fourier Transform features, frequency domain convolution features are obtained through complex graph convolution layers.

[0073] The complex graph convolutional layer is a graph neural network module adapted for learning topological features in complex domains. Its core function is to aggregate neighborhood information and transform features in brain functional connectivity maps represented in complex form, simultaneously capturing both amplitude correlation and phase coordination information in brain region connections. This layer inputs the graph's adjacency matrix and node features as complex tensors, and the convolution kernel parameters are also defined in complex form. The operation process follows the rules of complex arithmetic, achieving weighted fusion of neighborhood features through complex multiplication and addition. This preserves the strength information of connectivity relationships while fully encoding the asynchronous interaction mode brought about by phase differences. During feature update, nonlinear transformations can be introduced by combining activation functions to further enhance the model's ability to express complex complex domain topological features.

[0074] Specifically, the frequency domain convolution features are: in, H Represents frequency domain convolution features. σ This represents the activation function. A Represents the adjacency matrix of primitive brain functions. This represents the inverse fast Fourier transform. W k Indicates the first k A complex number of weights, b k Indicates the first k One bias.

[0075] S7016: Apply inverse fast Fourier transform to the frequency domain convolution features to obtain inverse fast Fourier transform features.

[0076] The inverse fast Fourier transform (IFFT) is the inverse algorithm of the fast Fourier transform (FFT). Its core function is to convert discrete signals in the frequency domain back to signals in the time domain. It also achieves efficient computation based on a divide-and-conquer strategy, reducing the computational complexity of the traditional inverse discrete Fourier transform from... Optimized to ( N The number of signal sampling points significantly improves the processing efficiency of inverse signal transformation. By reconstructing the amplitude and phase information of the frequency domain signal through inverse operations, the fluctuation pattern and change law of the signal in the time dimension can be restored, which is a key technology for realizing bidirectional conversion between the time domain and the frequency domain.

[0077] Specifically, the inverse fast Fourier transform features are as follows: in, H ifft This represents the characteristics of the inverse fast Fourier transform.

[0078] S7017: Based on the characteristics of the inverse fast Fourier transform, the linear mapping characteristics are obtained through the linear mapping matrix.

[0079] It should be noted that the inverse fast Fourier transform features are rearranged by nodes and linearly mapped using a learnable dimensionality reduction matrix to obtain linearly mapped features.

[0080] Specifically, the linear mapping feature is as follows: in, H proj Represents the characteristics of a linear mapping. matmul Represents the matrix multiplication operator. W emb This represents a linear mapping matrix.

[0081] S7018: Based on the linear mapping characteristics, frequency domain node characteristics are obtained through a feedforward network.

[0082] Feedforward networks are neural network structures where signals propagate in one direction only. Data flows from the input layer through hidden layers to the output layer, without feedback connections or recurrent loops. Their core function is to achieve layer-by-layer mapping and abstract representation of input features through the stacking of multiple linear transformations and nonlinear activation functions. Each neuron in this network is fully connected only to the neuron in the previous layer. Backpropagation optimizes the weight parameters of each layer, minimizing the error between the model's predicted and actual values. Structurally, the number of hidden layers and the size of neurons can be flexibly adjusted to adapt to feature learning tasks of varying complexity.

[0083] It should be noted that the linear mapping features are input into a two-layer feedforward network (FFN) to obtain the frequency domain node features.

[0084] Specifically, the frequency domain node features are as follows: in, H freg Represents the characteristics of frequency domain nodes. FFN This represents a feedforward network.

[0085] S7019: Through a gated network, time-domain node features and frequency-domain node features are dynamically weighted and fused to obtain shared node features.

[0086] Among them, the gated network is a neural network module with dynamic weight allocation capabilities. Its core function is to adaptively assign weight coefficients to multiple parallel expert networks or functional branches based on the differences in input features, thereby achieving differentiated attention to and efficient fusion of different feature patterns. This module typically takes task-related core features as input, and through mapping and normalization processing using structures such as multilayer perceptrons, outputs a set of weight vectors with values ​​between 0 and 1. The magnitude of the weight value represents the degree of adaptation of the corresponding expert network or branch to the current input features. The higher the weight, the greater the contribution of the corresponding branch's output to the final fusion result; the lower the weight, the smaller the contribution.

[0087] Specifically, the characteristics of shared nodes are as follows: in, H shared Indicates shared node characteristics, It represents the Hadamah accumulation. G Represents the gating weight vector. W g This represents the weight matrix of the gated network.

[0088] Reference manual attached Figure 3 The diagram illustrates a flowchart of a hybrid expert graph isomorphic network provided by an embodiment of the present invention.

[0089] S702: Based on the shared node features and the functional connectivity matrix and phase adjacency matrix in the training dataset, the amplitude branch graph representation and the phase branch graph representation are obtained through a bi-branch hybrid expert graph learning network.

[0090] The dual-branch hybrid expert graph learning network is a modular deep learning architecture for analyzing brain functional connectivity topology. Specifically designed for the complementary topological information of brain region connectivity, encompassing amplitude and phase, it comprises two parallel sub-networks: an amplitude branch and a phase branch, both structurally symmetrical but with completely independent parameters. Both branches employ a hybrid expert graph learning structure, incorporating a graph-aware gating network and multiple graph isomorphic expert modules. The gating network adaptively assigns expert weights to each brain region node based on shared node features and the corresponding branch's normalized adjacency matrix. Each expert module independently extracts topological features based on node neighborhood information and then generates a branch-specific graph representation through weighted fusion. The amplitude branch uses the functional connectivity matrix as topological input, focusing on learning synchronous correlation patterns of brain region signal intensity, while the phase branch uses the phase adjacency matrix as topological input, focusing on capturing the collaborative interaction relationships of brain region signal phases. Both branches also simultaneously output balanced and sparse losses to optimize the network's expert scheduling mechanism.

[0091] In one possible implementation, S702 specifically includes steps S7021 and S7022: S7021: Input the shared node features and functional connection matrix into the first hybrid expert graph network and output the magnitude branch graph representation.

[0092] It should be noted that shared node features are used as node features, and the functional connectivity matrix is ​​used as the adjacency matrix, both of which are input into the first hybrid expert graph network. The first hybrid expert graph network comprises an amplitude branch graph-aware gating network and an amplitude branch expert network. This process outputs an amplitude branch graph representation, as well as amplitude branch balance loss and amplitude branch sparsity loss for optimization.

[0093] S7022: Input the shared node features and phase adjacency matrix into the second hybrid expert graph network and output the phase branch graph representation.

[0094] It should be noted that shared node features are used as node features, and the phase adjacency matrix is ​​used as the adjacency matrix, both input into a parameter-independent second hybrid expert graph network. This second hybrid expert graph network comprises a phase branch graph perceptual gating network and a phase branch expert network. This process outputs a phase branch graph representation, along with the corresponding phase branch balance loss and phase branch sparsity loss.

[0095] Furthermore, the update formula for each node in the bi-branch hybrid expert graph learning layer is as follows: in, Indicates the first l Layer i Feature tensors of individual brain regions C This indicates the number of experts in a hybrid expert network. c An index representing an expert network. G c Indicates the first c Gating weights of an expert network, No. l -1st floor i Feature tensors of individual brain regions Represents the normalized adjacency matrix. E c Indicates the first c The output characteristics of an expert network N i Indicates the first i A set of domains for each brain region.

[0096] S703: Based on the amplitude branch diagram representation and the phase branch diagram representation, amplitude features and phase features are obtained through a cross-attention fusion mechanism.

[0097] The cross-attention fusion mechanism is a neural network module used to achieve deep interaction and complementary fusion of features from different sources. Its core function is to model the correlation weights between features, allowing one type of feature to focus on the key information most relevant to the task from another type of feature, thereby generating a more representative fused feature. This mechanism typically takes two sets of features to be fused as input, mapping them to query vectors, key vectors, and value vectors respectively. It calculates the similarity between the query vector and the key vector to obtain an attention weight matrix, then uses this matrix to perform a weighted sum of the value vectors, ultimately outputting a result that fuses the core information of both types of features. The magnitude of the weight value directly reflects the importance of the corresponding feature region; a higher weight indicates a greater contribution of that region to the task.

[0098] In one possible implementation, S703 specifically includes steps S7031 and S7032: S7031: Based on the amplitude branch diagram representation and the phase branch diagram representation, the amplitude feature is obtained through the first cross-attention module.

[0099] It should be noted that, using the first cross-attention module, the amplitude branch graph is represented as the query and the phase branch graph is represented as the key and value, and the amplitude feature of phase information enhancement is calculated.

[0100] Specifically, the amplitude characteristics are as follows: in, Indicates amplitude characteristics, CrossAttn This indicates a cross-attention module. Q Indicates a query. H amp This represents the amplitude branching plot. K Indicates key, H phase This represents the phase branch diagram. V Indicates the value.

[0101] S7032: Based on the amplitude branch diagram representation and the phase branch diagram representation, the phase features are obtained through the second cross-attention module.

[0102] It should be noted that, by using the second cross-attention module, with the phase branch graph as the query and the amplitude branch graph as the key and value, the phase feature with enhanced amplitude information is calculated.

[0103] Specifically, the phase characteristics are as follows: in, Indicates phase characteristics.

[0104] S704: The amplitude and phase features are spliced ​​together to obtain the final fused features.

[0105] The final fusion features are as follows: in, H final Indicates the final fusion characteristics, Concat This indicates a splicing operation. Indicates amplitude characteristics, Indicates phase characteristics.

[0106] S705: Input the final fused features into the multilayer perceptron for prediction and output the prediction result.

[0107] The multilayer perceptron (MLP) is a supervised learning classification model based on a feedforward neural network structure. Its core functionality involves using a multilayer fully connected neural network to perform layer-by-layer nonlinear transformations and abstractions on the high-dimensional input features, ultimately outputting the probability distribution of samples across different target categories to achieve classification. This classifier typically includes an input layer, several hidden layers, and an output layer. The input layer receives the feature vectors processed by the feature extraction module. The hidden layers use linear weighting operations and nonlinear activation functions to uncover complex relationships between features. The output layer uses the Softmax function to map the network output to a range of 0 to 1. The value of each output node corresponds to the probability that a sample belongs to the corresponding category, and the category corresponding to the maximum probability is the final classification result.

[0108] Multilayer perceptrons are existing technology and will not be described further in this invention.

[0109] S706: Construct a loss function based on the prediction results and the brain disease category labels of the training dataset.

[0110] The loss function is specifically as follows: in, L total Represents the loss function. L task This represents the task loss function. α This represents the auxiliary loss weighting coefficient. L balance Represents the balanced loss function. L sparse Represents the sparse loss function.B Represents the total number of samples. m Indicates the sample index. y m Indicates the first m The true diagnostic label of each sample Indicates the first m Predicted label for each sample, This represents the magnitude branch balance loss function. This represents the phase-branch balance loss function. This represents the single-branch balanced loss function. C This indicates the number of experts in a hybrid expert network. c An index representing an expert network. p c Indicates the first c The global average participation weight of an expert network N Indicates the number of brain regions. G c Indicates the first c Gating weights of an expert network, H m Indicates the first m The global node feature tensor of each sample This represents the amplitude branch sparse loss function. This represents the phase-branch sparse loss function. This represents a single-branch sparse loss function. h m Indicates the first m Feature vector of a single brain region node in a sample.

[0111] It should be noted that the binary classification cross-entropy task loss is used to guide the model in learning amplitude-phase topological feature association patterns related to brain diseases, ensuring the model's accurate ability to distinguish brain diseases. At the same time, by weighted fusion of balance loss and sparsity loss, the expert module calling mechanism of the dual-branch hybrid expert network achieves a dual constraint of "global equilibrium and local sparsity". This ensures that all expert modules in both amplitude and phase branches are evenly called to avoid resource idleness, and forces the model to activate only a few suitable expert modules for each brain region node to improve the targeting of feature extraction. Ultimately, while strengthening the model's ability to represent complex brain functional connectivity patterns, it effectively reduces the risk of overfitting and improves the model's generalization performance and training stability.

[0112] S707: With the goal of minimizing the function value of the loss function, the model parameters of the brain disease prediction model are optimized by gradient descent to complete the training of the brain disease prediction model.

[0113] Gradient descent is a classic iterative optimization algorithm that iteratively updates model parameters along the negative direction of the gradient of the objective function (such as the model's loss function) until it finds the optimal solution that minimizes the objective function. The gradient represents the rate of change and steepest ascent direction of the objective function at the current parameter point; therefore, adjusting parameters along the negative gradient direction allows the objective function value to continuously decrease. The step size of each parameter update is controlled by the learning rate. An excessively large learning rate can cause the parameters to oscillate around the optimal solution or even diverge, while an excessively small learning rate will result in slow parameter convergence. In the training process of brain functional connectivity analysis models, gradient descent (and its variants such as stochastic gradient descent, mini-batch gradient descent, and the Adam optimizer) is used to iteratively optimize network weights. By continuously calculating the gradient of the loss function with respect to the parameters of each layer and backpropagating, the error between the model's predicted values ​​and the true labels is gradually reduced, ultimately improving the model's classification accuracy for brain diseases.

[0114] In this embodiment of the invention, a brain disease prediction model is systematically trained using a training dataset. A time-frequency feature encoder is used to deeply fuse time-domain and frequency-domain information of brain region signals to generate shared node features. A dual-branch hybrid expert graph learning network is used to accurately mine differential features of amplitude and phase topology. A cross-attention mechanism is then used to strengthen the complementary correlation between the two types of features. Combined with a multi-loss collaborative optimization function and gradient descent method, parameters are iteratively adjusted. This achieves comprehensive mining and efficient fusion of multi-dimensional features of brain functional connectivity. Furthermore, by optimizing the expert module scheduling mechanism through balanced loss and sparse loss, the model's ability to represent brain disease-related features is effectively improved. Simultaneously, the risk of overfitting is reduced, and training stability is ensured. Ultimately, a model with strong generalization performance and high prediction accuracy is obtained, providing reliable algorithmic support for the accurate diagnosis of brain diseases.

[0115] Furthermore, after training the brain disease prediction model, test set data containing signal time series, functional connectivity matrix, phase adjacency matrix, and real disease labels are input into the trained model. The forward propagation process of the model training phase is executed to obtain the predicted values ​​of the test set, which are then converted into the final predicted labels through the argmax function. Based on the real and predicted labels of the test set, the model's performance is evaluated on two independent binary classification tasks: Alzheimer's disease vs. normal control and early mild cognitive impairment vs. late mild cognitive impairment. Indicators such as accuracy, sensitivity, specificity, F1-score, and AUC are calculated. Finally, the brain disease category classification results for the corresponding subjects and the model's performance evaluation report on the two tasks are output.

[0116] S8: Acquire resting-state functional magnetic resonance imaging (fMRI) data of the brain.

[0117] S9: After processing the resting-state functional magnetic resonance imaging (fMRI) data as described above, input the data into the trained brain disease prediction model for prediction and output the brain disease prediction results.

[0118] Reference manual attached Figure 4 The diagram shows a schematic representation of the structure of a brain disease prediction method and system that integrates amplitude and phase information with graph perception, provided by the present invention.

[0119] This invention also provides a brain disease prediction method and system 20 that integrates amplitude and phase information with graph perception hybrid experts, applied to the aforementioned brain disease prediction method integrating amplitude and phase information with graph perception hybrid experts, comprising: Processor 201.

[0120] The memory 202 stores computer-readable instructions that, when executed by the processor 201, implement the brain disease prediction method that integrates amplitude and phase information with graph perception hybrid experts as described in the method embodiment.

[0121] The brain disease prediction method and system 20 that integrates amplitude and phase information with image perception and hybrid expert provided by the present invention can perform the brain disease prediction method that integrates amplitude and phase information with image perception and hybrid expert described above, and achieve the same or similar technical effects. To avoid duplication, the present invention will not elaborate further.

[0122] It should be understood that the processor in the embodiments of the present invention can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0123] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0124] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0125] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0126] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0127] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0128] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are 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 implementations should not be considered beyond the scope of this invention.

[0129] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0130] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0131] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0132] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0133] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0134] This invention provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the brain disease prediction method that integrates amplitude and phase information with graph perception hybrid experts as described in the method embodiment.

[0135] The present invention provides a computer-readable storage medium that can implement the steps and effects of the brain disease prediction method of the above-described method embodiment, which integrates amplitude and phase information with graph perception hybrid experts. To avoid repetition, the present invention will not repeat the steps.

[0136] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0137] The following points need to be explained: (1) The accompanying drawings of the embodiments of the present invention only involve the structures involved in the embodiments of the present invention. Other structures can refer to the general design.

[0138] (2) For clarity, the thickness of layers or regions is enlarged or reduced in the drawings used to describe embodiments of the invention, i.e., these drawings are not drawn to scale. It is understood that when an element such as a layer, film, region or substrate is referred to as being “above” or “below” another element, the element may be “directly” located “above” or “below” the other element or there may be intermediate elements.

[0139] (3) Where there is no conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other to obtain new embodiments.

[0140] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A brain disease prediction method fusing a mixed expert of amplitude-phase information and graph perception, characterized in that, include: S1: Obtain the subject's raw resting-state functional magnetic resonance imaging data of the brain and the corresponding brain disease category labels; S2: Preprocess the raw resting-state functional magnetic resonance imaging (fMRI) data of the brain; S3: Based on the preprocessed raw resting-state functional magnetic resonance imaging data, extract the average blood oxygenation level dependent signal time series of multiple brain regions of the subject; S4: Construct the functional connection matrix and the phase adjacency matrix based on the signal time series; S5: Construct a training dataset based on the signal time series, the functional connectivity matrix, the phase adjacency matrix, and the brain disease category labels; S6: Construct a brain disease prediction model based on a dual-branch hybrid expert graph neural network; S7: Train the brain disease prediction model using the training dataset; S8: Acquire resting-state functional magnetic resonance imaging (fMRI) data of the brain; S9: After processing the resting-state functional magnetic resonance imaging data as described above, input the data into the trained brain disease prediction model for prediction, and output the brain disease prediction results.

2. The brain disease prediction method integrating amplitude and phase information with graph perception hybrid expert as described in claim 1, characterized in that, S4 specifically includes: S401: Construct the functional connection matrix and the phase adjacency matrix respectively based on the signal time series; S402: Perform adjacency matrix processing and normalization on the functional connection matrix and the phase adjacency matrix respectively to construct the functional connection matrix and the phase adjacency matrix.

3. The brain disease prediction method integrating amplitude and phase information with graph perception hybrid expert as described in claim 1, characterized in that, The construction of the functional connection matrix based on the signal time series in step S4 specifically includes: S401A: Based on the signal time series, calculate the Pearson correlation coefficient between any two brain regions over the time series. S402A: Construct an initial matrix based on the Pearson correlation coefficient, and perform binarization and graph symmetry normalization to construct the functional connectivity matrix.

4. The brain disease prediction method integrating amplitude and phase information with graph perception hybrid expert as described in claim 1, characterized in that, The construction of the phase adjacency matrix based on the signal time series in step S4 specifically includes: S401B: Perform narrowband filtering of 0.03-0.07Hz on the time series of the signal to obtain a narrowband filtered signal; S402B: Based on the narrowband filtered signal, the instantaneous phase of each brain region is obtained through Hilbert transform; S403B: Based on the instantaneous phase, calculate the mean cosine of the instantaneous phase difference between any two brain regions as the phase synchronization value; S404B: Based on the phase synchronization value, construct the phase adjacency matrix and perform graph normalization processing.

5. The brain disease prediction method integrating amplitude and phase information with graph perception hybrid expert as described in claim 1, characterized in that, Specifically, S7 includes: S701: Input the signal time series in the training dataset into the time-frequency feature encoder to obtain shared node features; S702: Based on the shared node features and the functional connectivity matrix and phase adjacency matrix in the training dataset, the amplitude branch graph representation and the phase branch graph representation are obtained through a dual-branch hybrid expert graph learning network; S703: Based on the amplitude branch diagram representation and the phase branch diagram representation, amplitude features and phase features are obtained through a cross-attention fusion mechanism; S704: The amplitude feature and the phase feature are concatenated to obtain the final fused feature; S705: Input the final fused features into a multilayer perceptron for prediction and output the prediction result; S706: Construct a loss function based on the prediction results and the brain disease category labels of the training dataset; S707: With the goal of minimizing the function value of the loss function, the model parameters of the brain disease prediction model are optimized by gradient descent to complete the training of the brain disease prediction model.

6. The brain disease prediction method based on the fusion of amplitude and phase information and graph perception hybrid expert as described in claim 5, characterized in that, Specifically, S701 includes: S7011: Based on the signal time series, extract local temporal dynamic features through a one-dimensional convolutional neural network module; S7012: Based on the local temporal dynamic features, temporal node features are obtained through adaptive one-dimensional average pooling and linear mapping layers; S7013: The signal time series is flattened and mapped to the word embedding dimension through the word embedding layer to obtain the embedding features; S7014: Apply Fast Fourier Transform to the embedded features to obtain Fast Fourier Transform features; S7015: Based on the Fast Fourier Transform features, frequency domain convolution features are obtained through a complex graph convolution layer; S7016: Apply inverse fast Fourier transform to the frequency domain convolution features to obtain inverse fast Fourier transform features; S7017: Based on the inverse fast Fourier transform characteristics, obtain the linear mapping characteristics through the linear mapping matrix; S7018: Based on the linear mapping characteristics, frequency domain node characteristics are obtained through a feedforward network; S7019: The shared node features are obtained by dynamically weighting and fusing the time-domain node features and the frequency-domain node features through a gating network.

7. The brain disease prediction method based on the fusion of amplitude and phase information and graph perception hybrid expert as described in claim 5, characterized in that, Specifically, S702 includes: S7021: Input the shared node features and the functional connection matrix into the first hybrid expert graph network, and output the magnitude branch graph representation; S7022: Input the shared node features and the phase adjacency matrix into the second hybrid expert graph network, and output the phase branch graph representation.

8. The brain disease prediction method based on the fusion of amplitude and phase information and graph perception hybrid expert as described in claim 5, characterized in that, The S703 specifically includes: S7031: Based on the amplitude branch diagram representation and the phase branch diagram representation, the amplitude feature is obtained through the first cross-attention module; S7032: Based on the amplitude branch diagram representation and the phase branch diagram representation, the phase feature is obtained through the second cross-attention module.

9. A method and system for predicting brain diseases by integrating amplitude and phase information with graph perception hybrid experts, characterized in that, include: processor; A memory storing computer-readable instructions, which, when executed by the processor, implement the brain disease prediction method that integrates amplitude and phase information with graph perception hybrid experts as described in any one of claims 1 to 8.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the brain disease prediction method according to any one of claims 1 to 8, which integrates amplitude and phase information with graph perception hybrid experts.

Images