A high-order individual difference brain network construction method based on triplets

Abstract

The application discloses a kind of high-order individual difference brain network construction methods based on triplets, first acquire the brain region data of several subjects;The relative distance of two brain regions in the subject brain region data is obtained, different brain regions are taken as row and column, and reference network is constructed according to the triplet of row, column and the relative distance of two brain regions;Eliminate subject G, obtain the relative distance of two brain regions in the subject brain region data after eliminating subject G, and construct the disturbance network of subject G according to the triplet of row and column of different brain regions and the relative distance;The reference network and the disturbance network of subject G are combined by weighting, and the high-order individual difference brain network of subject G is obtained.The triplet of brain region distance obtains the high-order interaction between brain regions, better reflects the characteristics that multiple brain region interaction realizes its functional attribute in real brain, so as to better understand the working principle of brain, improve the accuracy of constructed brain network.

Description

TECHNICAL FIELD

[0001] The present application relates to brain network construction, in particular to a high-order individual difference brain network construction method based on triplets. BACKGROUND

[0002] The human brain is an extremely complex system in which neurons, clusters of neurons, or regions are connected to each other, forming a complex network. The execution of various high-level cognitive functions of the brain relies on the coordinated cooperation of different brain regions, and the pathogenesis of many mental illnesses is also related to the connection disorders of brain regions. Therefore, neuroscientists try to use brain image data of different modalities to draw brain connection maps, among which the commonly used brain image data includes structural magnetic resonance imaging sMRI which can measure the tissue structure information of gray matter (GM), white matter (WM), cerebrospinal fluid (CSF), etc., diffusion magnetic resonance imaging dMRI which describes fiber connections, and functional magnetic resonance imaging fMRI which reflects the activity of brain regions. Compared with fMRI and dMRI data, sMRI has relatively low cost and is more suitable for large-scale research, and has more consistent image quality. At the same time, previous studies have found that sMRI data can provide brain connectivity information, and the inter-regional cortical thickness correlation map obtained using sMRI is similar to fiber tractography obtained by dMRI, which indicates that morphological correlation may be related to the anatomical connectivity of the brain. Traditional group-level research can only construct a unified network for the entire subject group, which is limited to group-level statistical analysis. Although this method can reveal the statistical differences in the mean of brain networks, it cannot capture the differences between individuals, and individual differences are ubiquitous in brain development. In the field of precision medicine, constructing individual difference brain networks will help to deeply mine individual-level biological characteristics and improve the accuracy of individualized diagnosis and treatment prediction for individual subjects.

[0003] To study individual variability and identify abnormal brain morphology in patients, more and more studies have begun to focus on the construction of morphological networks at the individual level. These methods can be divided into single metric-based and multiple metrics-based individualized morphological brain networks, such as dividing three-dimensional regions to calculate similarity, constructing individual brain morphological networks based on similarity of gray matter morphology, mainly by dividing the individual brain into multiple small cubes, and calculating the similarity of gray matter density between these cubes, and finding the significant small-world property in the network. However, the nodes in this network have changed from atlas-based regions to voxel-based cubes, and these strictly extracted cubes may not correspond well to specific regions of brain anatomy or function. A set of structural covariance networks of normal controls is constructed by calculating the Pearson correlation coefficient between the gray matter volumes of each region, and then a perturbed structural covariance network is constructed by adding one patient to the control group each time; the individual structural covariance network of each patient is represented by the difference between the perturbed network and the reference network.

[0004] Multiple metrics-based individualized morphological brain networks can be divided into single and multi-modal networks based on multiple indicators according to imaging methods. Single modal network, for example: in the prior art, a method based on Pearson correlation is used to estimate the similarity between seven morphological measurement features of different regions, which concatenates the average and standard deviation of the number of vertices, average curvature and Gaussian curvature, surface area, gray matter volume and cortical thickness of each region into a feature vector, and uses the Pearson correlation coefficient between the feature vectors of any pair of regions to estimate the morphological connectivity between regions. The network has a reproducible small-world property. Multi-modal network, for example: structural magnetic resonance imaging and diffusion magnetic resonance imaging are combined to construct a multi-modal brain morphological network based on three morphological measurement features of cortical curvature, myelin marker and gray matter morphology, aiming to provide a multi-dimensional description of brain morphological features. However, the above individual brain network construction methods only calculate the first-order correlation between two brain regions, and cannot reflect the high-order information of brain region interaction. The brain is a highly complex network system, and the interaction between different brain regions is not just a simple two-way interaction. The above brain network construction methods are not accurate enough. SUMMARY

[0005] The purpose of the application is to provide a three-tuple-based high-order individual difference brain network construction method for improving the accuracy of brain networks.

[0006] Technical scheme: To solve the above problems, the application adopts a three-tuple-based high-order individual difference brain network construction method, which comprises the following steps:

[0007] (1) Obtain brain region data from several subjects;

[0008] (2) Obtain the relative distances between each pair of brain regions in the subject's brain region data, use different brain regions as rows and columns, and construct a reference network based on the triplets of rows, columns, and relative distances between each pair of brain regions;

[0009] (3) Eliminate subject G, obtain the relative distances between each brain region in the brain region data of the subject after excluding subject G, and construct the perturbation network of subject G based on the rows and columns of different brain regions and the triplets of relative distances;

[0010] (4) The reference network and the perturbation network of subject G are weightedly combined to obtain the high-order individual difference brain network of subject G.

[0011] Furthermore, the relative distance between each pair of brain regions in the subjects' brain region data was calculated according to the Euclidean distance formula.

[0012] Furthermore, the calculation formula for the relative distance in the reference network is:

[0013]

[0014] in, represents the distance between brain region i and brain region v relative to brain region u, x im represents the i-th brain region of the m-th subject, x vm represents the vth brain region of the mth subject, x um represents the u-th brain region of the m-th subject, N represents the number of subjects; e (ω) (x i ,x v ) is the distance between brain region i and brain region v relative to the k brain regions closest to brain region i. The values ​​of v and u are the numbers of the first k brain regions closest to brain region i. The reference network e is constructed based on the distances of all brain regions. (ω) .

[0015] Furthermore, the calculation formula for the relative distance in the perturbation network is:

[0016]

[0017] in, represents the distance between brain area i and brain area v relative to brain area u after subject G is eliminated, e (ω-G) (x i ,x v ) represents the distance between brain region i and brain region v relative to the k brain regions closest to brain region i after subject G is removed. The perturbation network e is constructed based on the distances of all brain regions. (ω-G) .

[0018] Further, the high-order individual difference brain network e of the subject G (G) is:

[0019] e (G) = β (e (ω) -e (ω-G) ) + α (e (ω-G) )

[0020] Wherein, e (ω) represents a reference network, e (ω-G) represents a perturbation network of the subject G, and β and α represent weight coefficients.

[0021] Further, the step (4) further comprises z-score standardization of the high-order individual difference brain network e G .

[0022]

[0023] Wherein, e (G) (x i ,x v ) is an edge of the high-order individual difference brain network e (G) from the brain region i to the brain region v, Z iv is the edge of e (G) (x i ,x v ) after z-score standardization, and the network of the high-order individual difference brain network e (G) after z-score standardization is Z, μ v represents the mean of the high-order individual difference brain network e (G) , and σ v represents the standard deviation of the high-order individual difference brain network e (G) .

[0024] Further, the brain region data comprises normal brain region data and abnormal brain region data, the normal brain region data and the abnormal brain region data are structural magnetic resonance image data of a public data center, are preprocessed respectively to obtain preprocessed images of the normal brain region data and the abnormal brain region data, the obtained preprocessed images are divided into brain regions, and the gray matter value feature vectors of each brain region in each data are calculated.

[0025] Further, the brain region division is performed on the preprocessed brain image data according to a brain atlas, and the brain atlas comprises an AAL atlas.

[0026] The application also adopts a computer device, which comprises a memory, a processor, and a computer program stored in the memory and capable of running on the processor, and is characterized in that the processor implements the steps of the above method when executing the computer program.

[0027] The application also adopts a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the above method.

[0028] Beneficial effects: Compared with the prior art, the application has the remarkable advantages that the high-order interaction between brain regions is obtained through the triplets of brain region distances, the characteristics of the multi-brain region interaction in the real brain to realize the functional properties are better reflected, the working principle of the brain is better understood, and the accuracy of the constructed brain network is improved. The individualized difference network is constructed by using the perturbed network, which is helpful to mine the biological characteristics at the individual level and improve the accuracy of individualized diagnosis and treatment prediction. Since the brain regions usually interact with the brain regions close to them rather than the brain regions far away from them, only the brain regions close to the brain region under discussion are considered to calculate the triplets, so that the real brain activity is simulated and the dimensionality of the features is effectively reduced. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 Fig. 1 is a flowchart of the brain network construction method in the application.

[0030] Figure 2 Fig. 2 is a specific flowchart of the brain network construction method in the application.

[0031] Figure 3 Fig. 3 is a scatter plot of the real score and the predicted score of the cognitive total score obtained after performing the partial least squares regression on the BSNIP1 and BSNIP2 data sets in the brain network construction method in the application. DETAILED DESCRIPTION

[0032] As shown in Fig. 1, the brain network construction method based on triplets in the embodiment includes the following steps: Figure 1

[0033] Step S1: Extract the gray matter values of the sMRI brain regions of the healthy subjects and the schizophrenia subjects according to the brain atlas.

[0034] The brain atlas used can be the Brainnetome Atlas, the Automated Anatomy Labeling Atlas (AAL), the Allen Brain Atlas, etc. In the embodiment, the AAL template dividing the brain into 90 brain regions (excluding the cerebellum) is used to construct the individual difference brain network, and in actual application, one or more atlases can be selected as needed.

[0035] ​The sMRI data of schizophrenia and the corresponding normal person are taken as examples for illustration in this embodiment, but the protection scope of the present application is not limited thereto. The data can be sMRI data of all mental disorders, including mental diseases such as schizophrenia and depression, and non-mental diseases such as bipolar disorder.

[0036] In this embodiment, the gray matter of the structural magnetic resonance image sMRI is extracted as its feature, and the preprocessing and feature extraction of the sMRI are specifically as follows: (1) standardized to the Montreal standard space, preferably, resampled to 3*3*3mm; (2) segmented into white matter, gray matter and cerebrospinal fluid; (3) smoothed the gray matter with an 8mm half-maximum full-width Gaussian filter kernel; (4) detected outliers of the subjects to ensure that all subjects are correctly segmented; (5) obtained the standard space coordinates of the AAL template, and aligned the preprocessed gray matter image to the AAL template space using a linear registration method; (6) divided the aligned gray matter image into 90 brain regions, ensuring that each voxel is correctly assigned to the corresponding brain region; (7) averaged the voxel values in each brain region to obtain the gray matter value feature vector of each subject.

[0037] After the above processing, the structural magnetic resonance image sMRI data is converted into the gray matter value feature, and a 90*N feature matrix A is obtained, where N is the number of subjects. The feature matrix of the healthy subjects is denoted as A1, and the number of subjects is denoted as N1; the feature matrix of the schizophrenia subjects is denoted as A2, and the number of subjects is denoted as N2.

[0038] Step S2: Calculate the relative distance between each two brain regions in the two groups of subjects according to the Euclidean distance formula, and obtain the reference network corresponding to the two groups of subjects.

[0039] Specifically, taking the healthy subjects as an example, the process of calculating the reference network includes:

[0040] (1) Calculate the Euclidean distance between any two rows of the feature matrix, and sort them in ascending order to obtain a brain region serial number matrix N k of size 90*90. k Select the first k elements in the i-th row (i=1, 2,..., 90) of the serial number matrix N k , denoted as n i , k is 5 in this embodiment;

[0041] (2) Select the i-th row vector, u-th row vector and v-th row vector in the gray matter value feature matrix A1, denoted as x u , x v , respectively, where u, v e n k . Substitute x i , x u , x v into the following formula:

[0042]

[0043] The distance between brain region i and brain region v relative to brain region u can be obtained wherein, x *m represents the first brain region of the mth subject;

[0044] (3) The triple matrix S is calculated by the following formula (ω) Summation is performed in the dimension of u:

[0045]

[0046] The reference network e1 corresponding to the healthy subject can be obtained (ω) . Wherein, e1 (ω) (x i ,x v ) represents the distance between brain region i and brain region v relative to the distance of the nearest k brain regions from brain region i.

[0047] The same operation is performed on the schizophrenia subjects to obtain the reference network e2 corresponding to the schizophrenia subjects (ω) , wherein the brain region number matrix N k used in calculating the reference network corresponding to the schizophrenia subjects is the same as that used by the healthy subjects, and the calculation is not repeated.

[0048] Step S3: First, for the healthy subject group, the subjects G1 are sequentially removed (G1 = 1, 2,..., N1), and the operation in step S2 is repeated to obtain the disturbance network corresponding to each healthy subject. Then, for the schizophrenia group, the subjects G2 are sequentially removed (G2 = 1, 2,..., N2), and the operation in step S2 is repeated to obtain the disturbance network corresponding to each schizophrenia patient.

[0049] Taking the healthy subjects as an example, the process of calculating the disturbance network is as follows: first, the G1th column in the feature matrix A1 is deleted, wherein G1 = 1, 2,..., N1, and the matrix is obtained. Substitute it into steps S2(2)(3), and the disturbance network corresponding to each subject in the healthy subjects can be obtained The same operation is performed on the schizophrenia subjects to obtain the disturbance network corresponding to the schizophrenia subjects

[0050] Step S4: The reference network of the two groups of subjects and the disturbance network of each subject are combined by weighting, thereby obtaining the high-order individual difference brain network based on triplets.

[0051] Taking healthy subjects as an example, the specific steps for constructing a high-order individual difference brain network based on a triple include the following steps. (ω) and the perturbed network into the following formula:

[0052]

[0053] The individual difference triple high-order brain network corresponding to each healthy subject can be obtained. represents the individual difference network corresponding to the G1th subject; the β represents the weight coefficient of the reference network e1 (ω) and the perturbed network ; the α represents the weight coefficient of the perturbed network , and in the embodiment, the β is 15 and the α is 0.5.

[0054] Finally, the individual difference high-order brain network obtained is subjected to score standardization according to the following formula:

[0055]

[0056] , wherein, is the edge of the brain network from the brain region i to the brain region v, μ v is the mean of the individual difference network ; and the σ v is the standard deviation of the individual difference network . Z 1 represents the individual difference triple high-order brain network corresponding to the healthy subject after standardization, represents the edge of the individual difference triple high-order brain network Z 1 .

[0057] The same operation is performed on the schizophrenia subjects, and finally the individual difference triple high-order brain network Z 2 corresponding to each schizophrenia subject is obtained.

[0058] The embodiment of the application is tested by using the Bipolar and Schizophrenia Network for Intermediate Phenotypes (BSNIP) dataset (the BSNIP1 includes structural magnetic resonance image sMRI data of 163 schizophrenia subjects and 211 normal subjects, and the BSNIP2 includes structural magnetic resonance image sMRI data of 282 schizophrenia subjects and 421 normal subjects). The partial least squares regression is used as a prediction model, and the total score of cognition, the score of the number sequencing task, the score of the symbol coding task, the score of the labeled motor task, the score of the Tower of London task, the score of the language fluency task and the score of the language memory task are used as prior information.

[0059] like Figure 3 As shown in the figure, the scatter plot of the true score and the predicted score obtained after partial least squares regression of the total cognitive score on the BSNIP1 and BSNIP2 datasets is shown. The results show that the proposed method has good robustness in predicting the total cognitive score of different datasets.

[0060] As shown in Table 1, the method proposed in this embodiment is compared with the whole-brain voxel dimensionality reduction feature, brain region average gray matter value feature, and KL divergence feature in predicting the accuracy of seven cognitive scores: total cognitive score, digital sorting task, symbol encoding task, marking movement task, Tower of London task, language fluency task, and language memory based on the BSNIP1 dataset.

[0061] Table 1 Comparison of the accuracy of prediction of seven cognitive scores by four methods

[0062]

[0063] The results show that this method has the highest prediction accuracy among the four methods, indicating that the high-order individual difference brain network features proposed by this method can better capture the characteristic information related to cognitive scores.

[0064] In summary, this method has high prediction accuracy on different data sets, and in comparison with different brain network construction methods, this method has the highest prediction accuracy for different cognitive scores, indicating that the high-order individual difference brain network features proposed by this method have good robustness on different data sets and can more accurately reveal feature information related to cognitive scores.

Claims

1. A method for constructing a high-order individual difference brain network based on triplets, characterized in that, The method comprises the following steps: (1) obtaining brain region data of a plurality of subjects; (2) obtaining relative distances between two brain regions in the brain region data of the subjects, taking different brain regions as rows and columns, and constructing a reference network according to the triplets of rows, columns and relative distances between two brain regions; The relative distances between two brain regions in the brain region data of the subjects are calculated according to the Euclidean distance formula; the calculation formula of the relative distances in the reference network is: wherein, denotes the distance of brain region i to brain region v relative to brain region u, x im denotes the i-th brain region of the m-th subject, x vm denotes the v-th brain region of the m-th subject, x um denotes the u-th brain region of the m-th subject, N denotes the number of subjects; e (ω) (x i ,x v ) is the distance of brain region i to brain region v relative to the k brain regions closest to brain region i, v, u are the indices of the k brain regions closest to brain region i, a reference network e (ω) is constructed based on the distances of all brain regions; (3) removing the subject G, obtaining the relative distances between two brain regions in the brain region data of the subjects after removing the subject G, and constructing a perturbation network of the subject G according to the triplets of rows, columns and relative distances of different brain regions; the calculation formula of the relative distances in the perturbation network is: wherein, represents the distance between brain region i and brain region v relative to brain region u after removing subject G, e (ω-G) (x i , x v ) represents the distance between brain region i and brain region v relative to the k brain regions closest to brain region i after removing subject G, and the perturbation network e is constructed based on the distance of all brain regions (ω-G) ; (4) combining the reference network and the perturbation network of the subject G by weighting to obtain a high-order individual difference brain network of the subject G; The high-order individual difference brain network e of subject G (G) Is: e (G) = β(e (ω) - e (ω-G) ) + α(e (ω-G) ) where e (ω) represents the reference network, e (ω-G) represents the perturbed network of the subject G, and β, α represent weight coefficients. 2.The high-order individual difference brain network construction method according to claim 1, characterized in that, Step (4) further comprises performing z-score normalization on the high-order individual difference brain networks e (G) z-score normalization: wherein e (G) (x i ,x v ) is a high-order individual difference brain network e (G) , Z iv is an edge from brain region i to brain region v of e (G) (x i ,x v ) is an edge after z-score standardization of e (G) , Z is a network after z-score standardization of e v , μ (G) represents a mean of the vth brain region of e v , and σ (G) represents a standard deviation of the vth brain region of e 3.The high-order individual difference brain network construction method of claim 1, wherein, The brain region data includes normal brain region data and abnormal brain region data, and the normal brain region data and the abnormal brain region data are structural magnetic resonance imaging data of a public data center, which are preprocessed to obtain preprocessed images of the normal brain region data and the abnormal brain region data; the obtained preprocessed images are divided into N brain regions, and the gray matter value feature vectors of each brain region in each data are calculated. 4.The high-order individual difference brain network construction method of claim 1, wherein, The brain region division is performed on the preprocessed brain image data according to a brain atlas, and the brain atlas includes an AAL atlas.

5. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the method in any one of claims 1 to 4.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method in any one of claims 1 to 4.

Images