A brain network analysis method based on a multi-scale kernel attention mechanism

Abstract

This invention discloses a brain network analysis method based on a multi-scale kernel attention mechanism, comprising the following steps: preprocessing functional magnetic resonance imaging (fMRI) data to obtain time series data for each brain region; constructing an initial functional brain network for each subject using Pearson correlation as the initial input features of the model; extracting and analyzing features from the initial functional brain network using an encoder based on the multi-scale kernel attention mechanism; aggregating and reading out the output of the encoder based on the multi-scale kernel attention mechanism using orthogonal clustering based on prior knowledge of brain modularity; and outputting the final recognition result from the classification layer. This method captures the complex interactions between brain regions in a flexible and interpretable manner within a deep learning model, effectively constructing high-level functional brain networks, providing valuable information for recognition tasks, and effectively exploring biomarkers while improving recognition performance, thus possessing significant theoretical and practical value.

Description

Technical Field

[0001] This invention relates to a brain network analysis method based on a multi-scale kernel attention mechanism. It constructs a functional brain network using resting-state functional magnetic resonance imaging data, and uses an encoder based on a multi-scale kernel attention mechanism to extract and analyze features from the functional brain network. This allows for the acquisition of a high-level representation of the functional brain network and valuable features in a flexible and interpretable manner, ultimately enabling the identification of brain diseases based on the brain network analysis method. Background Technology

[0002] Major depressive disorder (MDD) is a typical brain disease caused by mental disorders, characterized by depressed mood, loss of interest, and impaired cognitive function. It not only negatively impacts the patient's daily life but also places a heavy burden on their family and society. Current clinical diagnostic techniques rely primarily on questionnaires and physician experience, which are highly subjective and lead to a high rate of misdiagnosis. Therefore, an increasing number of researchers are dedicated to utilizing various neuroscience data, combined with computer technology, to provide more objective evidence for assessing patients' conditions and assisting physicians in their diagnostic work.

[0003] Resting-state functional magnetic resonance imaging (fMRI) data is favored by researchers due to its non-invasiveness, lack of radiation, and high spatial resolution. Its basic principle is to detect blood oxygen level-dependent (BOLD) signals in the brain during resting states to indirectly reflect brain activity. Constructing and analyzing functional brain networks using BOLD signals is an effective method for diagnosing major depressive disorder. Although the pathogenesis is still not fully understood, studies have shown that abnormal connectivity patterns exist in the functional brain networks of patients, containing key information that distinguishes patients from healthy individuals.

[0004] Deep learning methods based on attention mechanisms have been widely applied in the field of functional brain network analysis and have demonstrated excellent results. From a mathematical perspective, attention mechanisms obtain the interaction relationships between brain regions by calculating attention scores between features of different brain regions; the resulting attention matrix can be viewed as a high-level representation of the functional brain network. Currently, researchers mainly use dot products or Gaussian kernel functions to calculate attention. The former is insufficient in capturing complex interactions between brain regions, while the latter, although capable of capturing complex interactions, lacks flexibility and interpretability.

[0005] To enable the computational method of attention mechanisms to capture complex interactions between brain regions while maintaining flexibility and interpretability, making it more suitable for brain network analysis tasks, this invention proposes a deep learning method based on a multi-scale kernel attention mechanism. Specifically, a dual-branch encoder is used to extract features from an initial functional brain network, with each branch using a different multi-scale Gaussian kernel function to calculate attention. Each multi-scale Gaussian kernel function inherits the powerful representational capabilities of the ordinary Gaussian kernel function, effectively capturing complex interactions between brain regions. Simultaneously, multiple learnable kernel parameters enhance the flexibility of modeling and, during training, can identify discriminative functional connections and brain regions, providing an effective means for exploring biomarkers and exhibiting excellent interpretability. The readout layer aggregates and reads out brain region features, ultimately leading to the identification results of brain diseases via a classification layer. Summary of the Invention

[0006] The purpose of this invention is to develop an attention mechanism suitable for brain network analysis. It proposes a brain network analysis method based on a multi-scale kernel attention mechanism, which enables the model to capture the complex interaction relationships between brain regions in a flexible and interpretable manner, thereby obtaining effective high-level functional brain network representations and features to complete the task of brain disease identification.

[0007] To address the problems existing in the background technology, the present invention adopts the following technical solution:

[0008] (1) Data preprocessing:

[0009] The DPARSF toolbox was used to preprocess functional magnetic resonance imaging (fMRI) data acquired by the MRI instrument, including removing pre-image fragments from the MRI images. The study included time points, head movement correction, spatial smoothing and normalization, 0.01-0.1 Hz bandpass filtering, and interference signal regression. Finally, a brain atlas was used to divide the brain into several regions, and the average time series for each region was extracted. The time series for each subject is represented as follows: ,in Indicates the number of brain regions. Indicates the length of the time series. Indicates the first Time series of brain regions;

[0010] (2) Constructing an initial functional brain network:

[0011] Based on the preprocessed data, an initial functional brain network for each subject was constructed using Pearson correlations, and this network was used as the initial input feature of the model, represented as follows: The formula for calculating Pearson correlation is as follows:

[0012] ;

[0013] in, Indicates the first The brain regions and the first Pearson correlation of brain regions and They represent and The average value;

[0014] (3) Encoder based on multi-scale kernel attention mechanism for feature extraction and analysis of initial functional brain network:

[0015] The encoder based on the multi-scale kernel attention mechanism was designed as a parallel structure consisting of two branches, and the specific process is as follows: First, the initial functional brain network is... The input is fed into two different multi-scale kernel attention modules, which use different multi-scale Gaussian kernel functions 1 and 2 to calculate the correlation between brain regions, resulting in two attention matrices. and The specific calculation formula is as follows:

[0016] Multiscale Gaussian kernel function 1:

[0017] ;

[0018] Multiscale Gaussian kernel function 2:

[0019] ;

[0020] in, The first value is calculated using the multi-scale Gaussian kernel function 1. The brain regions and the first Correlation between individual brain regions The expression represents the result of the calculation using the multi-scale Gaussian kernel function 2. Correlation between individual brain regions and any other arbitrary brain regions and They represent the first The brain regions and the first Initial characteristics of each brain region Initial features representing any brain region, and These are learnable kernel parameters, because and Multiple values ​​can be taken, thus allowing for multiple learnable kernel parameters. The two different multi-scale Gaussian kernel functions inherit the powerful representational capabilities of the ordinary Gaussian kernel function, modeling complex interactions between brain regions from different perspectives to obtain a high-level representation of the functional brain network. Furthermore, the multiple learnable kernel parameters make the modeling process more flexible. In addition, each kernel parameter has a one-to-one correspondence with a unique functional connection or brain region. Specifically, the kernel parameters in multi-scale Gaussian kernel function 1 correspond one-to-one with functional connections, while the kernel parameters in multi-scale Gaussian kernel function 2 correspond one-to-one with brain regions. For example, the first... Brain regions and nuclear parameters Correspondingly, the first Brain regions and nuclear parameters Corresponding; the The brain regions and the first Functional connectivity and nuclear parameters between brain regions Correspondingly, the first The brain regions and the first Functional connectivity and nuclear parameters between brain regions Correspondingly, during model training, the kernel parameters are adaptively learned and updated, enabling the identification of discriminative functional connections and brain regions, providing clues for the discovery of biomarkers and exhibiting excellent interpretability; the softmax function is used to adjust the two attention matrices. Perform normalization operations, and the linear layers of each branch are normalized to the initial features. Project the values ​​and multiply them by the normalized attention matrix to obtain the outputs of the two multi-scale kernel attention modules. and The formula is as follows:

[0021] ;

[0022] ;

[0023] in, and Represents two normalized attention matrices. and These represent the projection matrices of the two branches respectively;

[0024] Next, and Further feature extraction is performed through layer normalization, residual connections, and a feedforward network to obtain updated features. and Finally, and The two are concatenated along the feature dimension to obtain the output of the encoder based on the multi-scale kernel attention mechanism. .

[0025] (4) Aggregation and readout of brain region features:

[0026] Based on prior knowledge of brain modularity, orthogonal clustering is used to analyze the output of an encoder based on a multi-scale kernel attention mechanism. Aggregation and reading are performed; first, the number of functional modules in the brain is set. and initialized using Xavier to obtain Cluster centers Subsequently, the Gram-Schmidt method was used to obtain the orthogonal cluster centers. The specific formula is as follows:

[0027] ;

[0028] ;

[0029] Next, the distribution matrix is ​​obtained using the projection operator. ;No. The brain regions were assigned to the first The probability of each cluster center Calculated using the following formula:

[0030] ;

[0031] Finally, the allocation matrix transpose and Multiplication yields the output features of the readout layer. ;

[0032] (5) Classification:

[0033] The output features of the readout layer After being flattened into a vector, it is input into a multilayer perceptron to obtain the model's final recognition result.

[0034] After adopting the above technical solution, the beneficial technical effects of the present invention are: the multi-scale nuclear attention mechanism used in the present invention can not only capture the complex interaction relationship between brain regions in a flexible and interpretable manner, thereby obtaining effective high-level functional brain network representations and features, but also discover discriminative functional connections and brain regions, which is beneficial to improving recognition performance and exploring biomarkers. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the implementation of the present invention.

[0036] Figure 2 This is a specific framework diagram of the present invention.

[0037] Figure 3 This is a flowchart of the computation of the multi-scale kernel attention mechanism proposed in this invention. Detailed Implementation

[0038] The present invention will be further described below with reference to the accompanying drawings and specific implementation examples.

[0039] like Figures 1-3 As shown, the brain network analysis method based on multi-scale nucleus attention mechanisms includes the following steps:

[0040] (1) Data preprocessing:

[0041] Functional magnetic resonance imaging (fMRI) data were acquired, and the DPARSF toolbox was used to preprocess the fMRI data acquired by the MRI scanner. This process adopted the method described in the 2010 paper "DPARSF: a MATLAB toolbox for 'pipeline' data analysis of resting-state fMRI" published by Yan Chao-Gan and Zang Yu-Feng in *Frontiers in Systems Neuroscience*, Volume 4, Issue 13. Preprocessing included removing pre-existing features from the MRI images. The study included time points, head movement correction, spatial smoothing and normalization, 0.01-0.1 Hz bandpass filtering, and interference signal regression. Finally, the CC200 brain atlas was used to divide the brain into 200 regions, and the average time series for each region was extracted. The time series for each subject is represented as follows: ,in Indicates the number of brain regions. Indicates the length of the time series. Indicates the first Time series data for individual brain regions; additionally, the website http: / / rfmri.org / REST-meta-MDD publicly provides a dataset preprocessed using the DPARSF toolkit to analyze functional magnetic resonance imaging (fMRI) data acquired by an MRI scanner. In this example, the experiment used data from the three largest sites (site20, site21, and site25) in the publicly available REST-meta-MDD dataset, encompassing data from 841 participants, including 457 patients with major depressive disorder and 384 healthy individuals.

[0042] (2) Constructing an initial functional brain network:

[0043] Based on the preprocessed data, an initial functional brain network for each subject was constructed using Pearson correlations, and this network was used as the initial input feature of the model, represented as follows: The formula for calculating Pearson correlation is as follows:

[0044] ;

[0045] in, Indicates the first The brain regions and the first Pearson correlation of brain regions and They represent and The average value;

[0046] (3) Encoder based on multi-scale kernel attention mechanism for feature extraction and analysis of initial functional brain network:

[0047] The encoder based on the multi-scale kernel attention mechanism was designed as a parallel structure consisting of two branches, and the specific process is as follows: First, the initial functional brain network is... The input is fed into two different multi-scale kernel attention modules, which use different multi-scale Gaussian kernel functions 1 and 2 to calculate the correlation between brain regions, resulting in two attention matrices. and The specific calculation formula is as follows:

[0048] Multiscale Gaussian kernel function 1:

[0049] ;

[0050] Multiscale Gaussian kernel function 2:

[0051] ;

[0052] in, The first value is calculated using the multi-scale Gaussian kernel function 1. The brain regions and the first Correlation between individual brain regions The expression represents the result of the calculation using the multi-scale Gaussian kernel function 2. Correlation between individual brain regions and any other arbitrary brain regions and They represent the first The brain regions and the first Initial characteristics of each brain region Initial features representing any brain region, and These are learnable kernel parameters, because and Multiple values ​​can be taken, thus allowing for multiple learnable kernel parameters. The two different multi-scale Gaussian kernel functions inherit the powerful representational capabilities of the ordinary Gaussian kernel function, modeling complex interactions between brain regions from different perspectives to obtain a high-level representation of the functional brain network. Furthermore, the multiple learnable kernel parameters make the modeling process more flexible. In addition, each kernel parameter has a one-to-one correspondence with a unique functional connection or brain region. Specifically, the kernel parameters in multi-scale Gaussian kernel function 1 correspond one-to-one with functional connections, while the kernel parameters in multi-scale Gaussian kernel function 2 correspond one-to-one with brain regions. For example, the first... Brain regions and nuclear parameters Correspondingly, the first Brain regions and nuclear parameters Corresponding; the The brain regions and the first Functional connectivity and nuclear parameters between brain regions Correspondingly, the first The brain regions and the first Functional connectivity and nuclear parameters between brain regions Correspondingly, during model training, the kernel parameters are adaptively learned and updated, enabling the identification of discriminative functional connections and brain regions, providing clues for the discovery of biomarkers and exhibiting excellent interpretability; the softmax function is used to adjust the two attention matrices. Perform normalization operations, and the linear layers of each branch are normalized to the initial features. Project the values ​​and multiply them by the normalized attention matrix to obtain the outputs of the two multi-scale kernel attention modules. and The formula is as follows:

[0053] ;

[0054] ;

[0055] in, and Represents two normalized attention matrices. and These represent the projection matrices of the two branches respectively;

[0056] Next, and Further feature extraction is performed through layer normalization, residual connections, and a feedforward network to obtain updated features. and Finally, and The two are concatenated along the feature dimension to obtain the output of the encoder based on the multi-scale kernel attention mechanism. .

[0057] (4) Aggregation and readout of brain region features:

[0058] Based on prior knowledge of brain modularity, orthogonal clustering is used to analyze the output of an encoder based on a multi-scale kernel attention mechanism. Aggregation and reading are performed; first, the number of functional modules in the brain is set. and initialized using Xavier to obtain Cluster centers Subsequently, the Gram-Schmidt method was used to obtain the orthogonal cluster centers. The specific formula is as follows:

[0059] ;

[0060] ;

[0061] Next, the distribution matrix is ​​obtained using the projection operator. ;No. The brain regions were assigned to the first The probability of each cluster center Calculated using the following formula:

[0062] ;

[0063] Finally, the allocation matrix transpose and Multiplication yields the output features of the readout layer. ;

[0064] (5) Classification:

[0065] The output features of the readout layer After being flattened into a vector, it is input into a multilayer perceptron to obtain the model's final result for identifying severe depression.

[0066] (6) Experiment:

[0067] To verify the effectiveness of the proposed method, we conducted experiments identifying patients with major depressive disorder on three sites of the REST-meta-MDD dataset. The identification results on the test set were used to evaluate the model, and the evaluation metrics included accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC). The definitions of the metrics are as follows: , , TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative, respectively.

[0068] (7) Experimental results:

[0069] Table 1: Results of the method of the present invention and 6 other methods for identifying major depressive disorder.

[0070] ;

[0071] Table 1 presents the experimental results of the proposed method and six representative methods, including two traditional machine learning methods (CC+SVM and EC+SVM) and four deep learning methods (GAT, FBNETGEN, K-Transformer, and BNT). The experimental results show that the proposed method achieves the best performance in three metrics: Accuracy = 64.4%, AUC = 70.2%, and Specificity = 61.8%. Although it does not achieve the best results in sensitivity, the proposed method achieves a better balance between specificity and sensitivity, demonstrating the best overall performance.

[0072] Compared to two traditional machine learning methods (CC+SVM, EC+SVM), the deep learning method of this invention can adaptively extract features, which is beneficial to improving recognition performance. Compared to two deep learning methods based on graph neural networks (GAT, FBNETGEN), the method proposed in this invention benefits from the existence of a multi-scale kernel attention mechanism, which can effectively encode long-range dependencies between brain regions, overcoming the shortcomings of graph neural networks in grasping global information, thus exhibiting superior recognition performance. Compared to three deep learning methods equipped with traditional dot product attention mechanisms or Gaussian kernel attention mechanisms (GAT, K-Transformer, BNT), the multi-scale kernel attention mechanism proposed in this invention can not only capture complex interactions between brain regions, but also effectively identify important brain regions and functional connections through multiple learnable kernel parameters, possessing better flexibility and interpretability, making up for the shortcomings of previous attention mechanisms, and therefore performing optimally in experiments.

[0073] The advantages of this embodiment are as follows: its encoder based on the multi-scale kernel attention mechanism is designed as a parallel structure consisting of two branches, each containing a multi-scale kernel attention module; the initial functional brain network is input into two different multi-scale kernel attention modules of the encoder, which respectively use two different multi-scale Gaussian kernel functions with multiple learnable kernel parameters to calculate the correlation between brain regions, resulting in two attention matrices; the two different multi-scale Gaussian kernel functions inherit the powerful representational ability of ordinary Gaussian kernel functions, modeling the complex interaction relationships between brain regions from different perspectives, obtaining a high-level representation of the functional brain network, and the multiple learnable kernel parameters make the modeling process more flexible; at the same time, the kernel parameters in one multi-scale Gaussian kernel function have a one-to-one correspondence with functional connectivity, and the kernel parameters in the other multi-scale Gaussian kernel function have a one-to-one correspondence with brain regions; during model training, the kernel parameters are adaptively learned and updated, which can identify discriminative functional connectivity and brain regions, providing clues for the discovery of biomarkers, with excellent interpretability, which is conducive to improving recognition performance.

[0074] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and partial additions made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A brain network analysis method based on multi-scale nuclear attention mechanism, characterized in that... Includes the following steps: (1) Data preprocessing: The functional magnetic resonance imaging data acquired by the magnetic resonance instrument was preprocessed using the DPARSF toolbox; (2) Constructing an initial functional brain network: Based on the preprocessed data, an initial functional brain network for each subject was constructed using Pearson correlations, and this network was used as the initial input feature of the model, represented as follows: The formula for calculating Pearson correlation is as follows: ； in, Indicates the first The brain regions and the first Pearson correlation of brain regions and They represent and The average value; (3) Encoder based on multi-scale kernel attention mechanism for feature extraction and analysis of initial functional brain network: The encoder based on the multi-scale kernel attention mechanism was designed as a parallel structure consisting of two branches, and the specific process is as follows: First, the initial functional brain network is... The inputs are fed into two different multi-scale kernel attention modules, and different multi-scale Gaussian kernel functions 1 and 2 are used to calculate the correlation between brain regions, resulting in two attention matrices. and The specific calculation formula is as follows: Multiscale Gaussian kernel function 1: ； Multiscale Gaussian kernel function 2: ； in, The first value is calculated using the multi-scale Gaussian kernel function 1. The brain regions and the first Correlation between individual brain regions The expression represents the result of the calculation using the multi-scale Gaussian kernel function 2. Correlation between individual brain regions and any other arbitrary brain regions and They represent the first The brain regions and the first Initial characteristics of each brain region Represents the initial features of any brain region. and These are learnable kernel parameters, because and Multiple values ​​can be taken, thus allowing for multiple learnable kernel parameters. The two different multi-scale Gaussian kernel functions inherit the powerful representational capabilities of the ordinary Gaussian kernel function, modeling complex interactions between brain regions from different perspectives to obtain a high-level representation of the functional brain network. Furthermore, the multiple learnable kernel parameters make the modeling process more flexible. In addition, each kernel parameter has a one-to-one correspondence with a unique functional connection or brain region. Specifically, the kernel parameters in multi-scale Gaussian kernel function 1 correspond one-to-one with functional connections, while the kernel parameters in multi-scale Gaussian kernel function 2 correspond one-to-one with brain regions. Brain regions and nuclear parameters Correspondingly, the first Brain regions and nuclear parameters Corresponding; the The brain regions and the first Functional connectivity and nuclear parameters between brain regions Correspondingly, the first The brain regions and the first Functional connectivity and nuclear parameters between brain regions Correspondingly, during model training, the kernel parameters adaptively learn and update, enabling the identification of discriminative functional connections and brain regions, providing clues for the discovery of biomarkers; the softmax function is used to adjust the two attention matrices. Perform normalization operations, and the linear layers of each branch are normalized to the initial features. Project the values ​​and multiply them by the normalized attention matrix to obtain the outputs of the two multi-scale kernel attention modules. and The formula is as follows: ； ； in, and Represents two normalized attention matrices. and These represent the projection matrices of the two branches respectively; Next, and Further feature extraction is performed through layer normalization, residual connections, and a feedforward network to obtain updated features. and Finally, and The two are concatenated along the feature dimension to obtain the output of the encoder based on the multi-scale kernel attention mechanism. ; (4) Aggregation and readout of brain region features: Based on prior knowledge of brain modularity, orthogonal clustering is used to analyze the output of an encoder based on a multi-scale kernel attention mechanism. Aggregation and reading are performed; first, the number of functional modules in the brain is set. and initialized using Xavier to obtain Cluster centers Subsequently, the Gram-Schmidt method was used to obtain the orthogonal cluster centers. The specific formula is as follows: ； ； Next, the distribution matrix is ​​obtained using the projection operator. ;No. The brain regions were assigned to the first The probability of each cluster center Calculated using the following formula: ； Finally, the allocation matrix transpose and Multiplication yields the output features of the readout layer. ; (5) Classification: The output features of the readout layer After being flattened into a vector, it is input into a multilayer perceptron to obtain the model's final recognition result.

Images