Brain network analysis method based on multi-scale kernel attention mechanism

Abstract

The invention discloses a brain network analysis method based on a multi-scale kernel attention mechanism, and the method comprises the following steps: carrying out the preprocessing of functional magnetic resonance imaging data, and obtaining a time sequence of each brain region; utilizing Pearson correlation to construct an initial functional brain network of each subject as an initial input feature of the model; performing feature extraction and analysis on the initial functional brain network based on an encoder of a multi-scale kernel attention mechanism; based on the prior knowledge of brain modularity, the output of the encoder based on the multi-scale kernel attention mechanism is aggregated and read by using an orthogonal clustering mode; and the classification layer outputs a final identification result. According to the method, a complex interaction relationship between brain regions is captured in a deep learning model in a flexible and interpretive manner, a high-level functional brain network can be effectively constructed, valuable information is provided for a recognition task, a biological marker can be effectively explored on the basis of improving the recognition performance, and the recognition efficiency is improved. The method has important theoretical significance and practical value.

Description

Technical Field

[0001] The present invention relates to a brain network analysis method based on a multi-scale kernel attention mechanism. A functional brain network is constructed using resting-state functional magnetic resonance imaging data, and an encoder based on a multi-scale kernel attention mechanism is used to extract and analyze features of the functional brain network, thereby obtaining a high-level representation of the functional brain network and acquiring valuable features in a flexible and interpretable manner, and ultimately performing brain disease identification based on the brain network analysis method. Background Art

[0002] Major depressive disorder is a typical brain disease caused by a mental disorder, characterized by low mood, loss of interest, and impaired cognitive function. This not only negatively impacts patients' daily lives but also places a heavy burden on their families and society. Existing clinical diagnostic techniques rely primarily on questionnaires and physician experience, which are highly subjective and lead to high rates of misdiagnosis. Consequently, a growing number of researchers are committed to leveraging various brain science data, combined with computer technology, to provide more objective assessments of patients' conditions and assist physicians in their diagnosis.

[0003] Resting-state functional magnetic resonance imaging (fMRI) data is favored by researchers due to its non-invasive, radiation-free nature and high spatial resolution. Its basic principle is to detect the blood oxygen level-dependent (BOLD) signal in the brain at rest to indirectly reflect brain activity. Using the BOLD signal to construct and analyze functional brain networks is an effective method for diagnosing major depressive disorder. Although the pathogenic mechanisms remain unclear, studies have shown that abnormal connectivity patterns within patients' functional brain networks contain key information that distinguishes patients from healthy controls.

[0004] Deep learning methods based on attention mechanisms have been widely used in the field of functional brain network analysis and have demonstrated excellent results. Mathematically, attention mechanisms capture the interactions between brain regions by calculating attention scores between their features. The resulting attention matrix can be viewed as a high-level representation of functional brain networks. Currently, researchers primarily calculate attention using dot products or Gaussian kernel functions. The former is insufficiently capable of capturing complex interactions between brain regions, while the latter, while capable of capturing complex interactions, lacks flexibility and interpretability.

[0005] In order to make the calculation method of the attention mechanism have both flexibility and interpretability while capturing the complex interactions between brain regions, so as to make it more suitable for brain network analysis tasks, the present invention proposes a deep learning method based on a multi-scale kernel attention mechanism. Specifically, an encoder composed of two branches is used to extract features of the initial functional brain network, and each branch uses a different multi-scale Gaussian kernel function to calculate attention. Each multi-scale Gaussian kernel function inherits the powerful representation ability of the ordinary Gaussian kernel function and can effectively capture the complex interactions between brain regions. At the same time, multiple learnable kernel parameters improve the flexibility of modeling, and can identify discriminative functional connections and brain regions during the training process, providing an effective means for the exploration of biomarkers and having excellent interpretability. The readout layer aggregates and reads the brain region features, and finally obtains the recognition results of brain diseases through the classification layer. Summary of the Invention

[0006] The purpose of this invention is to develop an attention mechanism suitable for brain network analysis and propose a brain network analysis method based on a multi-scale kernel attention mechanism, so that the model can capture the complex interactions between brain regions in a flexible and interpretable manner, thereby obtaining effective high-level functional brain network representations and features, and completing the task of identifying brain diseases.

[0007] In order to solve the problems existing in the background technology, the present invention adopts the following technical solutions: (1) Data preprocessing: The DPARSF toolbox was used to preprocess the functional magnetic resonance imaging data collected by the magnetic resonance instrument, including removing the front time points, head motion correction, spatial smoothing and normalization, 0.01-0.1 Hz bandpass filtering and interference signal regression; finally, the brain was divided into several brain regions using the brain atlas, and the average time series of each brain region was extracted. The time series of each subject is expressed as ,in represents the number of brain regions, represents the length of the time series, Indicates the Time series of brain regions; (2) Constructing the initial functional brain network: Based on the preprocessed data, the Pearson correlation was used to construct the initial functional brain network of each subject and used it as the initial input feature of the model, expressed as ; The calculation formula of Pearson correlation is as follows: ; in, Indicates the brain regions and Pearson correlations among brain regions, and Respectively and The average value of (3) The encoder based on the multi-scale kernel attention mechanism extracts and analyzes features of the initial functional brain network: The encoder based on the multi-scale kernel attention mechanism is designed as a parallel structure consisting of two branches. The specific process is as follows: First, the initial functional brain network Input 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, and obtain two attention matrices and , the specific calculation formula is as follows: Multi-scale Gaussian kernel function 1: ; Multi-scale Gaussian kernel function 2: ; in, represents the first brain regions and The correlation between brain regions, represents the first The correlation between a brain region and any other brain region, and Respectively represent brain regions and The initial characteristics of each brain region, represents the initial features of any brain region, and is a learnable kernel parameter, since and It can take multiple values, so there are multiple learnable kernel parameters; the two different multi-scale Gaussian kernel functions inherit the powerful representation ability of the ordinary Gaussian kernel function, model the complex interactive relationship between brain regions from different perspectives, and obtain a high-level representation of the functional brain network, and 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; among them, the kernel parameters in the multi-scale Gaussian kernel function 1 correspond one-to-one to the functional connection, and the kernel parameters in the multi-scale Gaussian kernel function 2 correspond one-to-one to the brain region; for example, Brain region and nuclear parameters Corresponding, Brain region and nuclear parameters corresponding to; brain regions and Functional connectivity and nuclear parameters between brain regions Corresponding, brain regions and Functional connectivity and nuclear parameters between brain regions Corresponding; during model training, the kernel parameters are adaptively learned and updated, which can identify discriminative functional connections and brain regions, provide clues for the discovery of biomarkers, and have excellent interpretability; the two attention matrices are Perform normalization operation, the linear layer of each branch is normalized to the initial feature Project and multiply with the normalized attention matrix to obtain the output of the two multi-scale kernel attention modules and , the formula is as follows: ; ; in, and Represents two normalized attention matrices, and Represent the projection matrices of the two branches respectively; Next, and After further feature extraction through layer normalization, residual connection and feedforward network, the updated features are obtained. and Finally, and The two are concatenated in the feature dimension to obtain the output of the encoder based on the multi-scale kernel attention mechanism .

[0008] (4) Aggregate and read out brain region features: Based on the prior knowledge of brain modularity, the output of the encoder based on the multi-scale kernel attention mechanism is clustered using orthogonal clustering. Perform aggregation and readout; first, set the number of functional modules of the brain , and use Xavier initialization to get Cluster centers ; Then, the Gram-Schmidt method is used to obtain the orthogonal cluster centers The specific formula is as follows: ; ; Next, use the projection operator to get the distribution matrix ;No. brain regions were assigned to The probability of cluster centers Calculated by the following formula: ; Finally, the allocation matrix Transpose and Multiply to get the output features of the readout layer ; (5) Classification: The output features of the readout layer After flattening into a vector, it is input into the multi-layer perceptron to obtain the final recognition result of the model.

[0009] After adopting the above technical solution, the beneficial technical effect of the present invention is: the multi-scale kernel attention mechanism used in the present invention can not only capture the complex interactive relationship between brain regions in a flexible and interpretable manner, thereby obtaining effective high-level functional brain network representations and features, but also can discover discriminative functional connections and brain regions, which is conducive to improving recognition performance and exploring biomarkers. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is an implementation flow chart of the present invention.

[0011] Figure 2 It is a specific framework diagram of the present invention.

[0012] Figure 3 This is the computational flow chart of the multi-scale kernel attention mechanism proposed in this invention. DETAILED DESCRIPTION

[0013] The present invention will be further described below with reference to the accompanying drawings and specific implementation examples.

[0014] like Figure 1-Figure 3 As shown in Figure 2, the brain network analysis method based on the multi-scale kernel attention mechanism includes the following steps:

[0015] (1) Data preprocessing: Acquire functional magnetic resonance imaging data and preprocess the data using the DPARSF toolbox. This process uses the method described in the 2010 paper DPARSF: a MATLAB toolbox for “pipeline” data analysis of resting-state fMRI by Yan Chao-Gan and Zang Yu-Feng, published in the journal Frontiers in Systems Neuroscience, Volume 4, Issue 13. Preprocessing includes removing the front-end of the magnetic resonance image. The data were processed through the following steps: time points, head motion correction, spatial smoothing and normalization, 0.01-0.1 Hz bandpass filtering and interference signal regression; finally, the brain was divided into 200 brain regions using the CC200 brain atlas, and the average time series of each brain region was extracted. The time series of each subject is represented as ,in represents the number of brain regions, represents the length of the time series, Indicates the The DPARSF toolbox is used to preprocess functional magnetic resonance imaging (fMRI) data collected by MRI machines. This example uses data from the three largest sites in the publicly available REST-meta-MDD dataset (site 20, site 21, and site 25). This dataset includes data from 841 subjects, including 457 patients with major depressive disorder and 384 healthy subjects.

[0016] (2) Constructing the initial functional brain network: Based on the preprocessed data, the Pearson correlation was used to construct the initial functional brain network of each subject and used it as the initial input feature of the model, expressed as ; The calculation formula of Pearson correlation is as follows: ; in, Indicates the brain regions and Pearson correlations among brain regions, and Respectively and The average value of (3) The encoder based on the multi-scale kernel attention mechanism extracts and analyzes features of the initial functional brain network: The encoder based on the multi-scale kernel attention mechanism is designed as a parallel structure consisting of two branches. The specific process is as follows: First, the initial functional brain network Input 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, and obtain two attention matrices and , the specific calculation formula is as follows: Multi-scale Gaussian kernel function 1: ; Multi-scale Gaussian kernel function 2: ; in, represents the first brain regions and The correlation between brain regions represents the first The correlation between a brain region and any other brain region, and Respectively represent brain regions and The initial characteristics of each brain region, represents the initial features of any brain region, and is a learnable kernel parameter, since and It can take multiple values, so there are multiple learnable kernel parameters; the two different multi-scale Gaussian kernel functions inherit the powerful representation ability of the ordinary Gaussian kernel function, model the complex interactive relationship between brain regions from different perspectives, and obtain a high-level representation of the functional brain network, and 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; among them, the kernel parameters in the multi-scale Gaussian kernel function 1 correspond one-to-one to the functional connection, and the kernel parameters in the multi-scale Gaussian kernel function 2 correspond one-to-one to the brain region; for example, Brain region and nuclear parameters Corresponding, Brain region and nuclear parameters Corresponding; brain regions and Functional connectivity and nuclear parameters between brain regions Corresponding, brain regions and Functional connectivity and nuclear parameters between brain regions Corresponding; during model training, the kernel parameters are adaptively learned and updated, which can identify discriminative functional connections and brain regions, provide clues for the discovery of biomarkers, and have excellent interpretability; the two attention matrices are Perform normalization operation, the linear layer of each branch is normalized to the initial feature Project and multiply with the normalized attention matrix to obtain the output of the two multi-scale kernel attention modules and , the formula is as follows: ; ; in, and Represents two normalized attention matrices, and Represent the projection matrices of the two branches respectively; Next, and After further feature extraction through layer normalization, residual connection and feedforward network, the updated features are obtained. and Finally, and The two are concatenated in the feature dimension to obtain the output of the encoder based on the multi-scale kernel attention mechanism .

[0017] (4) Aggregate and read out brain region features: Based on the prior knowledge of brain modularity, the output of the encoder based on the multi-scale kernel attention mechanism is clustered using orthogonal clustering. Perform aggregation and readout; first, set the number of functional modules of the brain , and use Xavier initialization to get Cluster centers ; Then, the Gram-Schmidt method is used to obtain the orthogonal cluster centers The specific formula is as follows: ; ;

[0018] Next, use the projection operator to get the distribution matrix ;No. brain regions were assigned to The probability of cluster centers Calculated by the following formula: ; Finally, the allocation matrix Transpose and Multiply to get the output features of the readout layer ; (5) Classification: The output features of the readout layer After flattening into a vector, it is input into the multi-layer perceptron to obtain the model's final major depression recognition result.

[0019] (6) Experiment: To validate the effectiveness of our proposed method, we conducted experiments on identifying patients with major depression on three sites in the REST-meta-MDD dataset. The recognition results on the test set were used to evaluate the model using metrics including accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC). The metrics are defined as follows: , , , where TP, TN, FP and FN represent true positive, true negative, false positive and false negative, respectively.

[0020] (7) Experimental results: Table 1: Identification results of the method of the present invention and six other methods on major depression.

[0021] ; 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 achieved the best results in terms of three metrics: accuracy = 64.4%, area under the curve (AUC) = 70.2%, and specificity = 61.8%. Although it did not achieve the best results in terms of sensitivity, the proposed method achieved a better balance between specificity and sensitivity, demonstrating the best overall performance.

[0022] Compared to two traditional machine learning methods (CC+SVM and EC+SVM), the deep learning method of the present invention is able to adaptively perform feature extraction, which facilitates improved recognition performance. Compared to two deep learning methods based on graph neural networks (GAT and FBNETGEN), the method proposed in the present invention benefits from the use of a multi-scale kernel attention mechanism, effectively encoding long-range dependencies between brain regions. This overcomes the drawback of graph neural networks' lack of global information, resulting in superior recognition performance. Compared to three deep learning methods equipped with traditional dot-product or Gaussian kernel attention mechanisms (GAT, K-Transformer, and BNT), the multi-scale kernel attention mechanism proposed in the present invention not only captures the complex interactions between brain regions, but also incorporates multiple learnable kernel parameters that effectively identify important brain regions and functional connections. It offers greater flexibility and interpretability, overcoming the shortcomings of previous attention mechanisms and resulting in the best performance in the experiments.

[0023] The advantages of this embodiment are: its encoder based on the multi-scale kernel attention mechanism is designed as a parallel structure consisting of two branches, each branch 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, obtaining two attention matrices; the two different multi-scale Gaussian kernel functions inherit the powerful representation ability of the ordinary Gaussian kernel function, model the complex interactive relationship between brain regions from different perspectives, and obtain a high-level representation of the functional brain network, and 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 the functional connection, and the kernel parameters in the other multi-scale Gaussian kernel function have a one-to-one correspondence with the brain region; during model training, the kernel parameters are adaptively learned and updated, which can identify discriminative functional connections and brain regions, provide clues for the discovery of biomarkers, have excellent interpretability, and are conducive to improving recognition performance.

[0024] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement and partial supplement made within the spirit and principle of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A brain network analysis method based on a multi-scale kernel attention mechanism, characterized by The following steps are involved: (1) Data preprocessing: Use the DPARSF toolbox to preprocess the functional magnetic resonance imaging data collected by the magnetic resonance instrument; (2) Constructing the initial functional brain network: Based on the preprocessed data, the Pearson correlation was used to construct the initial functional brain network of each subject and used it as the initial input feature of the model, expressed as ; The calculation formula of Pearson correlation is as follows: ； in, Indicates the brain regions and Pearson correlations among brain regions, and Respectively and The average value of (3) The encoder based on the multi-scale kernel attention mechanism extracts and analyzes features of the initial functional brain network: The encoder based on the multi-scale kernel attention mechanism is designed as a parallel structure consisting of two branches. The specific process is as follows: First, the initial functional brain network Input into two different multi-scale kernel attention modules, and use different multi-scale Gaussian kernel functions 1 and 2 to calculate the correlation between brain regions, and obtain two attention matrices and , the specific calculation formula is as follows: Multi-scale Gaussian kernel function 1: ； Multi-scale Gaussian kernel function 2: ； in, represents the first brain regions and The correlation between brain regions, represents the first The correlation between a brain region and any other brain region, and Respectively represent brain regions and The initial characteristics of each brain region, represents the initial features of any brain region, and is a learnable kernel parameter, since and It can take multiple values, so there are multiple learnable kernel parameters; the two different multi-scale Gaussian kernel functions inherit the powerful representation ability of the ordinary Gaussian kernel function, model the complex interactive relationship between brain regions from different perspectives, and obtain a high-level representation of the functional brain network, and 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; among them, the kernel parameters in the multi-scale Gaussian kernel function 1 correspond one-to-one to the functional connection, and the kernel parameters in the multi-scale Gaussian kernel function 2 correspond one-to-one to the brain region; for example, Brain region and nuclear parameters Corresponding, Brain region and nuclear parameters Corresponding; brain regions and Functional connectivity and nuclear parameters between brain regions Corresponding, brain regions and Functional connectivity and nuclear parameters between brain regions Corresponding; during model training, the kernel parameters are adaptively learned and updated, which can identify discriminative functional connections and brain regions, provide clues for the discovery of biomarkers, and have excellent interpretability; the two attention matrices are Perform normalization operation, the linear layer of each branch is normalized to the initial feature Project and multiply with the normalized attention matrix to obtain the output of the two multi-scale kernel attention modules and , the formula is as follows: ； ； in, and Represents two normalized attention matrices, and Represent the projection matrices of the two branches respectively; Next, and After further feature extraction through layer normalization, residual connection and feedforward network, the updated features are obtained. and Finally, and The two are concatenated in the feature dimension to obtain the output of the encoder based on the multi-scale kernel attention mechanism ; (4) Aggregate and read out brain region features: Based on the prior knowledge of brain modularity, the output of the encoder based on the multi-scale kernel attention mechanism is clustered using orthogonal clustering. Perform aggregation and readout; first, set the number of functional modules of the brain , and use Xavier initialization to get Cluster centers ; Then, the Gram-Schmidt method is used to obtain the orthogonal cluster centers The specific formula is as follows: ； ； Next, use the projection operator to get the distribution matrix ;No. brain regions were assigned to The probability of cluster centers Calculated by the following formula: ； Finally, the allocation matrix Transpose and Multiply to get the output features of the readout layer ; (5) Classification: The output features of the readout layer After flattening into a vector, it is input into the multi-layer perceptron to obtain the final recognition result of the model.

Images