A brain network construction method based on brain region weight correlation

Abstract

The present invention discloses a brain network construction method based on brain region weight correlation. The method collects fMRI data of the subject's brain in a resting state and performs preprocessing. A brain network connectivity matrix is ​​constructed based on the correlation of physiological signals BOLD sequence. Then, three local features of the brain region BOLD sequence are extracted. The brain region weight sequence is obtained through SVM training, and the correlation of the weights is calculated. A brain network connectivity matrix based on the brain region weight correlation is constructed. The two connectivity matrices are fused to obtain a final brain network connectivity matrix. Finally, a corresponding brain network is constructed based on the connectivity matrix. Compared with existing methods, the method of the present invention not only constructs a brain network connectivity matrix based on the BOLD sequence, but also calculates the brain region weight sequence and correlation based on the local features of the BOLD sequence. It can focus on connections that have a greater impact on the disease but are relatively weak. The constructed brain network helps to discover brain regions that are truly related to the disease.

Description

Technical Field

[0001] The present invention belongs to the technical field of brain network construction, and specifically relates to a brain network construction method based on brain region weight correlation. Background Art

[0002] Functional magnetic resonance imaging (fMRI), as a non-destructive and dynamic detection technology, has increasingly become an important method for observing brain activity and revealing the relationship between the brain and mind. In order to model the interconnected brain tissue, neuroscientists are increasingly using the concept of network to describe and explain the interactions of brain tissue. Brain network connectivity is of great significance to the study of many diseases. Studies have shown that the brain network of patients with ASD has low brain integration network circuits and local circuit network functions are over-connected. The level of brain network connectivity in patients with schizophrenia is reduced, and the pattern of functional connectivity is more random. Major depression may involve changes in anatomical connectivity between brain regions that play a key role in emotion processing.

[0003] CN201710541557.X proposes a brain network construction method based on functional magnetic resonance (fMRI) psychophysiological interaction, which includes collecting fMRI data under task status and performing preprocessing such as time layer correction, head motion correction, and motion artifact reduction. A PPI model of brain cognitive decision-making is constructed based on the preprocessed fMRI data. The active components of brain regions related to task parameters are obtained through the PPI model, thereby constructing a brain network related to task parameters.

[0004] CN201710944535.8 proposes a fusion brain network construction method based on structural connectivity and functional connectivity. By fusing the structural connectivity matrix with the resting-state functional connectivity matrix, it can more effectively discover the differences in network attribute indicators between the brain networks of patients with brain diseases and those of normal people, thereby providing certain help for the research of various brain diseases.

[0005] Existing methods for constructing brain networks mostly determine the connections between brain regions based on the subject's physiological signals and brain structure. These brain networks can differ depending on the physiological signal characteristics of different individuals. This requires identifying disease-related connections from a vast number of unrelated brain network connections or inputting them into deep learning models for iterative learning, resulting in high computational complexity. Furthermore, constructing brain networks based on physiological signals typically requires setting a threshold. Some brain connections whose physiological signal correlation is below the threshold may be disease-related, but are often overlooked by researchers, affecting the accuracy of brain network construction. Summary of the Invention

[0006] To solve the above technical problems, the present invention proposes a brain network construction method based on the correlation of brain region weights. It not only explores brain network connections from the perspective of physiological signals, but also explores brain network connections from the perspective of the correlation of feature weights of different brain regions, so as to better reveal disease-related connections and provide auxiliary decision-making for the subsequent determination of disease-related abnormal brain regions.

[0007] The technical solution adopted by the present invention is a method for constructing a brain network based on the correlation of brain region weights, and the specific steps are as follows:

[0008] S1, collect data and perform data preprocessing;

[0009] Functional magnetic resonance imaging (fMRI) data were collected from subjects in the resting state and measured using a high-resolution structural image standard brain template space. Statistical parametric mapping (SPM) software was used to perform temporal correction, spatial registration, normalization, and smoothing preprocessing on the fMRI data. Filtering was used to remove high-frequency physiological noise and low-frequency signal drift, resulting in whole-brain BOLD time series data.

[0010] A standardized brain region template is selected to divide the functional magnetic resonance imaging of the brain into several brain regions (ROIs). Each brain region contains multiple voxels. The BOLD sequence of the pth voxel in the i-th brain region is expressed as

[0011] in, represents the BOLD signal at time point t, and T represents the length of the BOLD time series.

[0012] S2. Constructing a brain network connectivity matrix based on the BOLD sequence correlation of physiological signals;

[0013] Calculate the average value of the BOLD time series corresponding to each voxel point in the brain area The calculation formula is as follows:

[0014]

[0015] in, represents the average BOLD signal of each voxel in the i-th brain region at time t, ROI i represents the i-th brain region, H represents the number of voxels in the brain region, Represents the functional properties of a brain region's time series and brain activity.

[0016] The synchronization of brain region time series is used to represent the relationship between brain region functions, and the Pearson correlation coefficient of each brain region time series is calculated.

[0017]

[0018] in, express The Pearson correlation coefficient between It represents the average of the time scale of the time series of this brain region and is calculated as follows:

[0019]

[0020] By calculating the Pearson correlation coefficient between all brain regions, the brain network connectivity matrix R based on the BOLD sequence correlation of physiological signals was obtained.

[0021] S3, acquisition of local features of brain region BOLD sequence;

[0022] Three localization indicators, namely low-frequency amplitude ALFF, proportional low-frequency amplitude fALFF and local consistency ReHo, are used to obtain the local characteristics of brain regions.

[0023] The ALFF indicator represents the average value of the square root of the signal power spectrum in the low-frequency band. First, the BOLD time series is Fourier transformed to obtain its power spectrum. The ALFF value is calculated by square rooting the power spectrum to obtain the specific low-frequency band mean value. The calculation is as follows:

[0024]

[0025] Among them, Y i It represents the square root of the value of the i-th point in the BOLD signal time series power spectrum, and N1 and N2 represent the data index positions of the discrete power spectrum corresponding to the lowest and highest frequencies of the selected frequency band, respectively.

[0026] The fALFF indicator is calculated as follows:

[0027]

[0028] The ReHo index reflects the degree of consistency of activity in local brain regions. First, the voxels in the brain region are sorted in time series (ascending order) to obtain the signal ranking values ​​of the voxels at each time point.

[0029] r t,j Represents the ascending ranking value of the signal value of voxel j at time t in the entire time series, and calculates the total ranking value of the studied voxels in the ROI at time t:

[0030]

[0031] Where H represents the number of voxels contained in the ROI.

[0032] Calculate the average of the total ranking values ​​of all study voxels in the ROI at each time:

[0033]

[0034] Where T represents the length of the BOLD time series;

[0035] The ReHo corresponding to the brain region is calculated as follows:

[0036]

[0037] S4, ROI weight sequence acquisition;

[0038] For the ReHo index, each subject obtains the ReHo index through step S3, and obtains the sample set Spl = {X1X2...X N}∈R N*d .

[0039] Among them, R represents a real number set, N represents the number of subjects, d represents the number of brain regions divided by the standard template, X N Represents the vector consisting of d ReHo indices of the Nth subject.

[0040] Each data X in the sample set Spl n (n=1,2,3,…N), with subject labels L n ; Use these sample data for SVM machine learning.

[0041] Then the sample set Sp1 is divided into M sample subsets:

[0042] The sample set Spl is divided into different groups, and the sample size of each group is K / 2. When the sample size of the last group is less than K / 2, each group is represented as Combining these groups pairwise, we get: That is, (2N / K)(2N / K-1) / 2 sample subsets of size K are represented as Spl1, Spl2, Spl3, ..., Spl M .

[0043] Wherein, M represents the number of sample subsets, and the repetition rate between different sample subsets does not exceed 50%.

[0044] For each sample subset Spl i , trained using support vector machine SVM.

[0045] After SVM training, the accuracy on this sample subset can be obtained as AUC i , and the optimized parameter W i ={w i1 ,w i2 ,w i3 ...w id}.

[0046] Among them, the parameter w id represents the weight of the dth brain region obtained by training the i-th sample subset; the weights obtained by training different sample subsets are arranged in rows to obtain the weight matrix W:

[0047]

[0048] Calculate the correlation coefficient of the weights obtained from SVM training between brain regions i and j according to the weight matrix W

[0049]

[0050] Among them, w mi represents the weight of brain region i on sample subset m, Represents the weighted average value of brain region i on each sample subset:

[0051]

[0052] The accuracy AUC obtained by training on each sample subset i The average value is used to obtain the training accuracy AUC on the ReHo index. ReHo .

[0053] Similarly, the weighted series correlation coefficient can be obtained based on the ALFF indicator And the training accuracy AUC on this indicator ALFF , based on the fALFF indicator, the weighted serial correlation coefficient can be obtained And the training accuracy AUC on this indicator fALFF .

[0054] S5. Construct a brain network connectivity matrix based on brain region weight correlation;

[0055] Based on the weighted sequence correlation coefficient obtained in step S4 Obtain the brain network connectivity matrix based on the ReHo index weight sequence

[0056] Similarly, the brain network connectivity matrix based on the ALFF index weight sequence can be obtained Brain network connectivity matrix based on fALFF indicator weight sequence

[0057] The above three connectivity matrices are fused to obtain the brain network connectivity matrix R based on local feature weights w :

[0058]

[0059] Where SU = AUC ReHo +AUC ALFF +AUC fALFF .

[0060] S6, brain network construction;

[0061] The connectivity matrix R obtained in step S2 and the connectivity matrix R obtained in step S5 are combined W Perform linear summation to obtain the final constructed brain network connectivity matrix:

[0062] R f =θ W +(1-θ)R (13)

[0063] Among them, R f It represents the brain network matrix that integrates the BOLD signal correlation in the human brain network and the SVM training weight correlation target. The parameter θ is determined by the actual effect.

[0064] Then determine the binary network construction threshold, by defining the network sparsity, determine the threshold, for R f In the matrix, elements greater than the threshold are retained and used as an edge in the network. If the elements are less than the threshold, the connection between the two nodes is removed to generate the corresponding brain network.

[0065] Beneficial effects of the present invention: The method of the present invention collects fMRI data of the subject's brain in a resting state and performs preprocessing, constructs a brain network connectivity matrix based on the correlation of the physiological signal BOLD sequence, then extracts three local features of the brain region BOLD sequence, obtains the brain region weight sequence through SVM training, calculates the correlation of the weights, constructs a brain network connectivity matrix based on the brain region weight correlation, fuses the two connectivity matrices to obtain the final brain network connectivity matrix, and finally constructs the corresponding brain network based on the connectivity matrix. Compared with the existing methods, the method of the present invention, on the basis of constructing a brain network connectivity matrix based on the BOLD sequence, also calculates the brain region weight sequence and correlation based on the local features of the BOLD sequence, can focus on connections that have a greater impact on the disease but are relatively weak, and the constructed brain network helps to discover brain regions that are truly related to the disease. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a flowchart of a brain network construction method based on brain region weight correlation of the present invention. DETAILED DESCRIPTION

[0067] The present invention will be further described below with reference to the accompanying drawings and examples.

[0068] like Figure 1As shown in FIG, a flow chart of a method for constructing a brain network based on brain region weight correlation of the present invention, the specific steps are as follows:

[0069] S1, collect data and perform data preprocessing;

[0070] Functional magnetic resonance imaging (fMRI) data were collected from the subjects in the resting state and measured using a high-resolution structural image standard brain template space. Statistical Parametric Mapping (SPM) software was used to perform temporal correction, spatial registration, normalization, and smoothing preprocessing on the fMRI data. Filtering (bandwidth 0.01-0.08 Hz) was used to remove high-frequency physiological noise and low-frequency signal drift, resulting in whole-brain BOLD (Blood Oxygenation Level Dependent) time series data.

[0071] A standardized brain region template is selected to divide the functional magnetic resonance imaging of the brain into several brain regions (ROIs). Each brain region contains multiple voxels. The BOLD sequence of the pth voxel in the i-th brain region is expressed as

[0072] in, represents the BOLD signal at time point t, and T represents the length of the BOLD time series.

[0073] S2. Constructing a brain network connectivity matrix based on the BOLD sequence correlation of physiological signals;

[0074] Calculate the average value of the BOLD time series corresponding to each voxel point in the brain area The calculation formula is as follows:

[0075]

[0076] in, represents the average BOLD signal of each voxel in the i-th brain region at time t, ROI i represents the i-th brain region, H represents the number of voxels in the brain region, Represents the functional properties of a brain region's time series and brain activity.

[0077] The synchronization of brain region time series is used to represent the relationship between brain region functions, and the Pearson correlation coefficient of each brain region time series is calculated.

[0078]

[0079] in, express The Pearson correlation coefficient between It represents the average of the time scale of the time series of this brain region and is calculated as follows:

[0080]

[0081] By calculating the Pearson correlation coefficient between all brain regions, the brain network connectivity matrix R based on the BOLD sequence correlation of physiological signals was obtained.

[0082] S3, acquisition of local features of brain region BOLD sequence;

[0083] The local characteristics of brain regions can be used to study abnormalities in the brain. In this embodiment, three local indicators, namely, low-frequency amplitude ALFF, proportional low-frequency amplitude fALFF, and local consistency ReHo, are used to obtain local characteristics of brain regions.

[0084] The ALFF indicator represents the average value of the square root of the signal power spectrum in the low-frequency band. First, the BOLD time series is Fourier transformed to obtain its power spectrum. The ALFF value is calculated by square rooting the power spectrum to obtain the specific low-frequency band mean value. The calculation is as follows:

[0085]

[0086] Among them, Y i It represents the square root of the value of the i-th point in the BOLD signal time series power spectrum, and N1 and N2 represent the data index positions of the discrete power spectrum corresponding to the lowest and highest frequencies of the selected frequency band, respectively.

[0087] The fALFF indicator is calculated as follows:

[0088]

[0089] The ReHo index reflects the degree of consistency of activity in local brain regions. First, the voxels in the brain region are sorted in time series (ascending order) to obtain the signal ranking values ​​of the voxels at each time point.

[0090] For example, if a given time series is: 0.1, 0.5, 0.4, 0.6, 0.7, then the ascending sort value of 0.5 in this sequence is 3, and the signal sort value of the voxel at the second time point is 3.

[0091] r t,i Represents the ascending ranking value of the signal value of voxel j at time t in the entire time series, and calculates the total ranking value of the studied voxels in the ROI at time t:

[0092]

[0093] Where H represents the number of voxels contained in the ROI;

[0094] Calculate the average of the total ranking values ​​of all study voxels in the ROI at each time:

[0095]

[0096] Where T represents the length of the BOLD time series;

[0097] The ReHo corresponding to the brain region is calculated as follows:

[0098]

[0099] S4, ROI weight sequence acquisition;

[0100] For the ReHo index, each subject obtains the ReHo index through step S3, and obtains the sample set Spl = {X1X2...X N}∈R N*d .

[0101] Among them, R represents a real number set, N represents the number of subjects, d represents the number of brain regions divided by the standard template, X N Represents the vector consisting of d ReHo indices of the Nth subject.

[0102] Each data X in the sample set Spl n (n=1,2,3,…N), with subject labels L n ; Use these sample data for SVM (Support Vector Machine) machine learning.

[0103] Then the sample set Sp1 is divided into M sample subsets:

[0104] The sample set Spl is divided into different groups, and the sample size of each group is K / 2. When the sample size of the last group is less than K / 2, samples are randomly selected from the sample set Spl to make up K / 2 samples.

[0105] Each group is represented as Combining these groups pairwise, we get: That is, (2N / K)(2N / K-1) / 2 sample subsets of size K are represented as Spl1, Spl2, Spl3, ..., Spl M .

[0106] Wherein, M represents the number of sample subsets, and the repetition rate between different sample subsets does not exceed 50%.

[0107] For each sample subset Spl i, trained using support vector machine SVM.

[0108] After SVM training, the accuracy on this sample subset can be obtained as AUC i , and the optimized parameter W i ={w i1 ,w i2 ,w i3 ...w id}.

[0109] Among them, the parameter w id represents the weight of the dth brain region obtained by training the i-th sample subset; the weights obtained by training different sample subsets are arranged in rows to obtain the weight matrix W:

[0110]

[0111] Calculate the correlation coefficient of the weights obtained from SVM training between brain regions i and j according to the weight matrix W

[0112]

[0113] Among them, w mi represents the weight of brain region i on sample subset m, Represents the weighted average value of brain region i on each sample subset:

[0114]

[0115] The accuracy AUC obtained by training on each sample subset i The average value is used to obtain the training accuracy AUC on the ReHo index. ReHo .

[0116] Similarly, the weighted series correlation coefficient can be obtained based on the ALFF indicator And the training accuracy AUC on this indicator ALFF , based on the fALFF indicator, the weighted serial correlation coefficient can be obtained And the training accuracy AUC on this indicator fALFF .

[0117] S5. Construct a brain network connectivity matrix based on brain region weight correlation;

[0118] Based on the weighted sequence correlation coefficient obtained in step S4 Obtain the brain network connectivity matrix based on the ReHo index weight sequence

[0119] Similarly, the brain network connectivity matrix based on the ALFF index weight sequence can be obtained Brain network connectivity matrix based on fALFF indicator weight sequence

[0120] The above three connectivity matrices are fused to obtain the brain network connectivity matrix R based on local feature weights w :

[0121]

[0122] Where SU = AUC ReHo +AUC ALFF +AUC fALFF .

[0123] S6, brain network construction;

[0124] The connectivity matrix R obtained in step S2 and the connectivity matrix R obtained in step S5 are combined W Perform linear summation to obtain the final constructed brain network connectivity matrix:

[0125] R f =θ W +(1-θ)R (13)

[0126] Among them, R f It represents the brain network matrix that integrates the BOLD signal correlation in the human brain network and the SVM training weight correlation target. The parameter θ is determined by the actual effect.

[0127] Then determine the binary network construction threshold, by defining the network sparsity, determine the threshold, for R f In the matrix, elements greater than the threshold are retained and used as an edge in the network. If the elements are less than the threshold, the connection between the two nodes is removed to generate the corresponding brain network.

[0128] In summary, compared with existing methods, the method of the present invention, on the basis of constructing a brain network connectivity matrix based on the BOLD sequence, also calculates the brain region weight sequence and correlation based on the local characteristics of the BOLD sequence. It can focus on the connections that have a greater impact on the disease but are relatively weak. The constructed brain network helps to discover the brain regions that are truly related to the disease.

[0129] Those skilled in the art will appreciate that the embodiments described herein are intended to aid the reader in understanding the principles of the present invention, and it should be understood that the scope of the present invention is not limited to such specific descriptions and embodiments. Various modifications and variations are readily apparent to those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims.

Claims

1. A brain network construction method based on brain region weight correlation, the specific steps are as follows: S1, collect data and perform data preprocessing; Functional magnetic resonance imaging (fMRI) data were collected from the subjects in the resting state and measured uniformly within a high-resolution structural image standard brain template space. Statistical parametric mapping (SPM) software was used to perform temporal correction, spatial registration, standardization, and smoothing preprocessing on the fMRI data. High-frequency physiological noise and low-frequency signal drift were removed through filtering to obtain whole-brain BOLD time series data. A standardized brain region template is selected to divide the functional magnetic resonance imaging of the brain into several brain region ROIs. Each brain region contains multiple voxels. The BOLD sequence of the pth voxel in the i-th brain region is expressed as in, represents the BOLD signal at time point t, and T represents the length of the BOLD time series; S2. Constructing a brain network connectivity matrix based on the BOLD sequence correlation of physiological signals; Calculate the average value of the BOLD time series corresponding to each voxel point in the brain area The calculation formula is as follows: in, represents the average BOLD signal of each voxel in the i-th brain region at time t, ROI i represents the i-th brain region, H represents the number of voxels in the brain region, Represents the functional properties of a brain region's temporal sequence and brain activity; The synchronization of brain region time series is used to represent the relationship between brain region functions, and the Pearson correlation coefficient of each brain region time series is calculated. in, express The Pearson correlation coefficient between It represents the average of the time scale of the time series of this brain region and is calculated as follows: By calculating the Pearson correlation coefficient between all brain regions, the brain network connectivity matrix R based on the BOLD sequence correlation of physiological signals was obtained; S3, acquisition of local features of brain region BOLD sequence; Three locality indices, namely low-frequency amplitude ALFF, proportional low-frequency amplitude fALFF and local consistency ReHo, were used to obtain local characteristics of brain regions. The ALFF indicator represents the average value of the square root of the signal power spectrum in the low-frequency band. First, the BOLD time series is Fourier transformed to obtain its power spectrum. The ALFF value is calculated by square rooting the power spectrum to obtain the specific low-frequency band mean value. The calculation is as follows: Among them, Y i represents the square root of the value of the i-th point in the power spectrum of the BOLD signal time series, N1 and N2 represent the data index positions of the discrete power spectrum corresponding to the lowest and highest frequencies of the selected frequency band, respectively; The fALFF indicator is calculated as follows: The ReHo index reflects the degree of consistency of activity in a local brain region. First, the voxels in the brain region are sorted in ascending order of time series to obtain the signal ranking value of the voxels at each time point. r t,j Represents the ascending ranking value of the signal value of voxel j at time t in the entire time series, and calculates the total ranking value of the studied voxels in the ROI at time t: Where H represents the number of voxels contained in the ROI; Calculate the average of the total ranking values ​​of all study voxels in the ROI at each time: Where T represents the length of the BOLD time series; The ReHo corresponding to the brain region is calculated as follows: S4, ROI weight sequence acquisition; For the ReHo index, each subject obtains the ReHo index through step S3, and obtains the sample set Spl = {X1X2...X N }∈R N*d ; Among them, R represents a real number set, N represents the number of subjects, d represents the number of brain regions divided by the standard template, X N represents the vector consisting of d ReHo indices of the Nth subject; Each data X in the sample set Spl n , n=1,2,3,…N, with subject labels L n ; Use these sample data for SVM machine learning; Then the sample set Sp1 is divided into M sample subsets: Divide the sample set Sp1 into different groups, with the sample size of each group being K / 2. If the sample size of the last group is less than K / 2, randomly select samples from the sample set Sp1 to make up for the K / 2 samples. Each group is represented as Combining these groups pairwise, we get: That is, (2N / K)(2N / K-1) / 2 sample subsets of size K are represented as Spl1, Spl2, Spl3, ..., Spl M ; Where M represents the number of sample subsets, and the repetition rate between different sample subsets does not exceed 50%; For each sample subset Spl i , trained using support vector machine SVM; After SVM training, the accuracy on this sample subset can be obtained as AUC i , and the optimized parameter W i ={w i1 ,w i2 ,w i3 ...w id }; Among them, the parameter w id represents the weight of the dth brain region obtained by training the i-th sample subset; the weights obtained by training different sample subsets are arranged in rows to obtain the weight matrix W: Calculate the correlation coefficient of the weights obtained from SVM training between brain regions i and j according to the weight matrix W Among them, w mi represents the weight of brain region i on sample subset m, Represents the weighted average value of brain region i on each sample subset: The accuracy AUC obtained by training on each sample subset i The average value is used to obtain the training accuracy AUC on the ReHo index. ReHo ; Similarly, the weighted series correlation coefficient can be obtained based on the ALFF indicator And the training accuracy AUC on this indicator ALFF , based on the fALFF indicator, the weighted serial correlation coefficient can be obtained And the training accuracy AUC on this indicator fALFF ; S5. Construct a brain network connectivity matrix based on brain region weight correlation; Based on the weighted sequence correlation coefficient obtained in step S4 Obtain the brain network connectivity matrix based on the ReHo index weight sequence Similarly, the brain network connectivity matrix based on the ALFF index weight sequence can be obtained Brain network connectivity matrix based on fALFF indicator weight sequence The above three connectivity matrices are fused to obtain the brain network connectivity matrix R based on local feature weights w : Where SUM = AUC ReHo +AUC ALFF +AUC fALFF ; S6, brain network construction; The connectivity matrix R obtained in step S2 and the connectivity matrix R obtained in step S5 are combined W Perform linear summation to obtain the final constructed brain network connectivity matrix: R f =θR W +(1-θ)R (13) Among them, R f It represents the brain network matrix that integrates the BOLD signal correlation in the human brain network and the SVM training weight correlation target. The parameter θ is determined by the actual effect. Then determine the binary network construction threshold, by defining the network sparsity, determine the threshold, for R f In the matrix, elements greater than the threshold are retained and used as an edge in the network. If the elements are less than the threshold, the connection between the two nodes is removed to generate the corresponding brain network.

Images