Brain network construction and analysis method based on multivariate analysis

Abstract

The invention provides a brain network construction and analysis method based on multivariable analysis. The method is mainly used for multi-view analysis of a brain structure and a functional network. The technical problem to be solved is that information loss is caused by neglecting necessary multivariate relationships among brain region nodes when univariate analysis is carried out on a brain function network and a brain structure network. According to the method, the canonical correlation analysis method based on data driving is used for constructing the single-mode brain network, the problem that a traditional univariate analysis method neglects necessary multivariate relations is avoided, the complex relation between the brain structure and functions can be reflected more accurately, and a more reliable basis is provided for deep research of a brain working mechanism.

Description

Technical Field

[0001] The present invention relates to the interdisciplinary field of artificial intelligence and brain science based on brain networks and brain-like computing, and specifically to a brain network construction and analysis method based on multivariate analysis, which is mainly used for multi-perspective analysis of brain structure and functional networks. Background Art

[0002] With the increasing aging of society, the incidence of cognitive impairment diseases such as Alzheimer's disease and mild cognitive impairment continues to rise, placing a heavy burden on society and families. Currently, the prevention and treatment of such diseases requires a deep understanding of the brain's cognitive memory and other mechanisms. A typical method for studying brain mechanisms is to construct and analyze brain networks. However, traditional brain network modeling often uses univariate analysis methods such as Pearson correlation analysis and partial correlation analysis, which ignores the necessary multivariate relationships between brain region nodes and leads to information loss. Summary of the Invention

[0003] This paper provides a brain network construction and analysis method based on multivariate analysis, primarily for multi-perspective analysis of brain structural and functional networks. The technical problem to be solved is that univariate analysis of functional and structural brain networks ignores the necessary multivariate relationships between nodes in brain regions, resulting in information loss.

[0004] The technical solution adopted by the present invention to solve the above technical problems is:

[0005] A method for constructing and analyzing brain networks based on multivariate analysis comprises the following steps:

[0006] Step 1: Multimodal brain imaging data acquisition and preprocessing;

[0007] Multimodal brain imaging data, including structural magnetic resonance imaging (sMRI) and functional magnetic resonance imaging (fMRI), were acquired from authoritative databases such as the Alzheimer's Disease Neuroimaging Initiative (ADNI). Preprocessing of the acquired sMRI and fMRI data involved format conversion, slice time correction, motion correction, registration, segmentation, spatial normalization, smoothing, linear drift removal, and filtering to eliminate noise, improve data quality, and provide a reliable data foundation for subsequent analysis.

[0008] Step 2: Unimodal brain network construction based on canonical correlation analysis;

[0009] A data-driven canonical correlation analysis method was used to construct structural and functional brain networks. For the structural network, nodes were defined using the AAL116 template, and canonical correlations between gray matter density between brain regions were calculated as edge weights. For the functional network, nodes were also defined using the AAL116 template, and canonical correlations between time series of blood oxygenation signals between brain regions were calculated as edge weights. The constructed networks were then sparsified to obtain binary structural and functional networks.

[0010] Step 3: Unimodal brain network analysis;

[0011] The constructed unimodal brain structural and functional networks were analyzed. Using graph theory, the topological properties of the networks were calculated, including small-world properties, global network efficiency, hierarchy, and modularity. These graph theoretical properties were compared between patients with cognitive impairment and healthy controls to identify correlations between cognitive and memory impairments and brain network abnormalities. Furthermore, combining morphological analysis of brain structural images with REHO and ALFF signals extracted from functional images, a significance analysis of brain region differences was performed to identify brain regions that play a key role in cognitive and memory mechanisms.

[0012] According to the method for constructing and analyzing a brain network based on multivariate analysis described in the present application, the specific steps of preprocessing in step 1 include:

[0013] (1) Format conversion: The acquired data is in DICOM format and needs to be converted to NIFTI format for subsequent processing;

[0014] (2) Slice time correction: During the initial scan, the magnetic field in the device is unstable and the subject is in an adaptation process. Generally, the noise of the first 10 time points of the image is removed. In addition, the scanning process generally adopts the interlayer method, and the image lacks temporal consistency, so the time sequence combination needs to be redefined.

[0015] (3) Head motion correction: In order to eliminate the influence of the subject's head movement during the experiment, head correction is required, and samples with a movement angle greater than 2° and a horizontal head motion greater than 2 mm are eliminated;

[0016] (4) Registration: Structural images have higher spatial resolution than functional images, so the structural images are registered to the functional images to improve the accuracy of spatial standardization of the functional images;

[0017] (5) Segmentation: Segment the processed structural image into gray matter, white matter, and cerebrospinal fluid;

[0018] (6) Spatial standardization: The images of each subject are different and need to be mapped to a unified standard space. Generally, a transformation matrix is ​​used to map them to the MNI space.

[0019] (7) Smoothing: Perform Gaussian smoothing on the image to improve the image signal-to-noise ratio. By default, 4*4*4 full-width at half maximum Gaussian noise is used.

[0020] (8) Eliminate linear drift: Use regression methods to remove the impact of linear drift caused by heating of the acquisition equipment;

[0021] (9) Filtering: To reduce the impact of high-frequency noise such as heartbeat and breathing, the frequency range is controlled within 0.01 to 0.08 Hz;

[0022] According to the brain network construction and analysis method based on multivariate analysis described in the present application, it is characterized in that during the sparse processing in step 2: the threshold selection strategy is specifically as follows: the threshold is selected so that the average degree value of the binary network is greater than twice the logarithm of N and the network density is less than 0.5.

[0023] The beneficial effect of the present invention is that the present invention provides a brain network construction and analysis method based on multivariate analysis, which is mainly used for multi-perspective analysis of brain structure and function networks. The technical problem to be solved is that the univariate analysis of brain function network and brain structure network ignores the necessary multivariate relationships between brain region nodes, resulting in information loss. The present application constructs a unimodal brain network by using a data-driven canonical correlation analysis method, avoiding the problem of traditional univariate analysis methods ignoring necessary multivariate relationships, and can more accurately reflect the complex relationship between brain structure and function, providing a more reliable basis for in-depth research on the working mechanism of the brain. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Schematic diagram of brain MRI data - (a) sMRI data;

[0025] Figure 2 Schematic diagram of brain MRI data - (b) fMRI data;

[0026] Figure 3 Brain imaging data preprocessing steps;

[0027] Figure 4 The process of constructing brain structural network and brain functional network based on canonical correlation analysis. DETAILED DESCRIPTION

[0028] The following will be combined with the Figures 1-4 The present invention is described in detail, and the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0029] This invention provides a brain network construction and analysis method based on multivariate analysis, primarily for multi-perspective analysis of brain structural and functional networks. The technical problem to be solved is that univariate analysis of brain functional and structural networks ignores the necessary multivariate relationships between brain region nodes, resulting in information loss. The specific implementation is as follows:

[0030] 1. Data acquisition and preprocessing

[0031] sMRI and fMRI data were obtained from the ADNI database, where sMRI data are T1 structural images and fMRI data are resting-state functional images. The quality of brain imaging data is affected by multiple factors, such as equipment factors, subject factors, and environmental factors. Equipment factors include imaging device performance, scanning parameter settings, etc. Subject factors include head movement, physiological noise, subject status, etc. Environmental factors include environmental noise, power supply stability, etc. In order to reduce the impact of various factors, improve data quality, and reduce experimental errors, brain imaging data needs to be preprocessed to make the conclusions more accurate and credible. The steps of data preprocessing are as follows: Figure 1 As shown. The data were preprocessed using DPABI software. The specific steps include: (1) Format conversion: The acquired data is in DICOM format and needs to be converted to NIFTI format for subsequent processing. (2) Layer time correction: During the initial scan, the magnetic field in the device is unstable and the subject has an adaptation process. Generally, the noise of the first 10 time points of the image is removed. In addition, the scanning process generally adopts the interlayer method, and the image lacks time consistency, and the time sequence combination needs to be redefined. (3) Head motion correction: In order to eliminate the influence of the subject's head movement during the experiment, head correction is required, and samples with a movement angle greater than 2° and a horizontal head movement greater than 2mm are eliminated. (4) Registration: Structural images have higher spatial resolution than functional images, so the structural images are registered to the functional images to improve the accuracy of spatial standardization of the functional images. (5) Segmentation: The processed structural images are segmented into three parts: gray matter, white matter, and cerebrospinal fluid. (6) Spatial standardization: The images of each subject are different, and they need to be mapped to a unified standard space. Generally, a conversion matrix is ​​used to map them to the MNI space. (7) Smoothing: Perform Gaussian smoothing on the image to improve the signal-to-noise ratio. By default, a 4*4*4 full-width-at-half-maximum Gaussian noise is used. (8) De-linear drift: Use regression methods to remove the effects of linear drift caused by heating of the acquisition device. (9) Filtering: To reduce the effects of high-frequency noise such as heartbeat and breathing, control the frequency range to 0.01-0.08 Hz.

[0032] 2. Unimodal Brain Network Construction

[0033] The construction of a unimodal structural network is based on sMRI structural images. After preprocessing, the sMRI structural image data is segmented into gray matter, white matter, and cerebrospinal fluid. Based on the segmented gray matter images, the brain structural network is constructed. Nodes in the structural network are defined using a brain template. After preprocessing, the sMRI data are converted to MNI space. The AAL template is created based on MNI space. That is, the anatomical regions in the AAL template are defined using MNI coordinates. Therefore, the AAL template can be used to define structural network nodes. AAL templates include AAL90 and AAL116. AAL116 includes all cerebral regions defined by AAL90 as well as 26 cerebellar regions. Research has shown that the cerebellum is also an important participant in brain-related mechanisms. Therefore, AAL116 was used to define structural network nodes. Univariate analysis methods such as the Pearson correlation coefficient or partial correlation coefficient are often used to define structural edges. However, necessary multivariate relationships may be overlooked by univariate analysis methods. Therefore, multivariate analysis methods are used to assess correlations between nodes. Canonical correlation analysis is used to define structural network edges. After spatial normalization, each subject image has the same spatial resolution and coordinate system, and the same number of voxels in the same brain region, as delineated according to AAL116. A subject sequence is determined for the experimental group, and the gray matter density value of a voxel in this subject sequence forms a random variable. A brain region contains several voxels, forming a random variable group. Using canonical correlation analysis, the maximum correlation coefficient of the canonical variables between two brain regions is calculated as the structural connectivity strength between the two regions. After defining the connecting edges, a threshold is selected to sparsify the network.

[0034] For functional networks, the AAL116 template is also used to define nodes. Each brain region defined by the AAL116 template is considered a region of interest (ROI). The time series of a voxel signal within the ROI constitutes a random variable, and the time series of all voxels that make up the ROI form a random variable group. A linear combination of the random variable group, namely the canonical variable, is found to best represent the random variable group. By calculating the correlation between the canonical variables of the two ROIs, the connecting edges between the nodes of the two ROIs are defined. The connecting edges of all ROIs in the network are calculated and sparsely distributed to construct a functional connectivity network.

[0035] Both structural and functional networks require selecting an appropriate threshold for sparseness, transforming a fully connected weighted network into a partially connected binary network. The threshold selection strategy is as follows: select the threshold such that the average degree of the binary network is greater than twice the logarithm of N and the network density is less than 0.5.

[0036] 3. Single-modality brain network analysis

[0037] Graph-theoretic topological properties of unimodal brain networks were calculated, including small-world properties, global network efficiency, hierarchy, and modularity. By comparing these properties between patients with cognitive impairment and healthy controls, abnormal correlations between cognitive and memory impairments and brain network abnormalities were revealed. For example, patients with cognitive impairment showed reduced small-world properties, decreased global network efficiency, and abnormalities in hierarchy and modularity.

[0038] In addition, morphological analysis of brain structural images and activity signal analysis of functional images were performed to determine the important role of abnormal brain regions in brain mechanisms such as cognition and memory. The specific methods are as follows.

[0039] 1. Morphological Analysis of Structural Images

[0040] Voxel-Based Morphometry (VBM) is a neuroimaging technique used to analyze changes in brain structure. By analyzing voxel intensities across different subject groups, changes in brain tissue, such as the distribution of gray matter, white matter, and cerebrospinal fluid, can be assessed. The VBM analysis process includes image preprocessing and statistical analysis.

[0041] (1) Image preprocessing: MRI data were preprocessed to segment brain tissue into gray matter, white matter, and cerebrospinal fluid, and then standardized using the MNI template.

[0042] (2) Statistical analysis: Use statistical methods to test the differences in data distribution.

[0043] In morphological analysis, the T-test is a commonly used statistical method. The T-test is often used for inter-group comparisons and is suitable for normally distributed data with small sample sizes and unknown population standard deviations. In brain network analysis, the most commonly used form of the T-test is the two-sample T-test. The two-sample T-test is used to compare the mean differences between two independent groups. In VBM, the gray matter density of each voxel is treated as a variable, and the mean values ​​of different subject groups at that voxel are tested to see if there are significant differences, and the T-value of the difference is calculated.

[0044]

[0045] in, and is the variance of the two groups of samples, and is the mean of the two groups of samples, and n1 and n2 are the sample sizes of the two groups. The larger the T value, the more obvious the difference in the means of the two groups of data.

[0046] 2. Activity Signal Analysis

[0047] REHO signal and ALFF signal are two commonly used signal indicators in rs-fMRI to analyze the spontaneous neural activity of the brain.

[0048] The REHO (Regional Homogeneity) signal measures the temporal synchronization of BOLD signals within a local brain region, reflecting the signal consistency of adjacent voxels. Higher REHO values ​​indicate more synchronized local activity within a region. Kendall's Coefficient of Concordance (KCC) is often used to calculate the REHO signal and assess the temporal similarity of a voxel with its surrounding voxels. Typically, 27 neighboring voxels are used. A voxel and its neighboring voxels are denoted by the matrix X(t,n).

[0049]

[0050] Where t is the time point and n is the number of voxels. Calculate the rank matrix of X(t,n) to obtain R(t,n).

[0051]

[0052] Among them, r i,j Denotes the rank of the j-th voxel at the i-th time point. Calculate the Kendall coefficient KCC.

[0053]

[0054] The ALFF signal (Amplitude of Low-Frequency Fluctuations) measures the intensity of spontaneous neural activity in the brain's low-frequency range, typically in the 0.01-0.1 Hz frequency range. Compared to REHO, which describes local network activity based on voxel-by-voxel calculations, ALFF analyzes signal activity across the entire network. Its calculation process involves bandpass filtering each voxel's time series, converting the time series to the frequency domain using a fast Fourier transform, and calculating the square root of the sum of the amplitude spectra within the low-frequency range as the ALFF value. A normalized version of ALFF is called fALFF (Fractional ALFF). fALFF divides the ALFF value by the total energy across the entire frequency range to reduce the influence of high-frequency noise. After calculating ALFF, a Fisher z-transform can be performed to make the resulting data distribution more normal. This yields zALFF.

[0055]

[0056] Here, mALFF is the ALFF value of each voxel divided by the mean of the whole-brain signal, and SD is the standard deviation of the ALFF value of the whole-brain voxels.

[0057] After calculating the voxel signal, intergroup analysis requires testing the significance of signal differences between groups. A commonly used statistical method is the two-sample T-test. The definition of the two-sample T-test is given in Formula 3-1. When testing for signal differences, the signal value of each voxel is treated as a variable. The mean signal value at that voxel is tested for significant differences between groups, and the T-score of the difference is calculated.

[0058] The brain network construction and analysis method based on multivariate analysis described in this application has important application value in the following fields:

[0059] 1. Precision diagnosis and treatment of neurological diseases:

[0060] By integrating multimodal brain network information (e.g. Figure 2 It can comprehensively analyze the cross-modal network degradation mechanism of neurodegenerative diseases such as Alzheimer's disease and Parkinson's disease, accurately locate the structural-functional coupling abnormalities of key systems such as the default mode network, and provide objective imaging biomarkers for early diagnosis.

[0061] 2. Research areas on the mechanisms of mental illness:

[0062] It is suitable for the neural circuit analysis of complex mental illnesses such as depression and schizophrenia. Through multimodal core-periphery joint analysis, it reveals cross-level organizational abnormalities in the limbic system-prefrontal cortex circuit, and assists in formulating personalized neuromodulatory treatment plans.

[0063] 3. Brain-computer interface performance optimization field:

[0064] In brain-computer interaction scenarios such as motor imagery, a multimodal network fusion model is used to decode the coordination mechanism between the motor cortex and cerebellum, greatly improving the accuracy of command recognition and significantly reducing system response delay.

[0065] 4. Neurosurgery surgery planning area:

[0066] Accurately locate the hub nodes of the epileptic focus network before surgery, protect key structure-function coupling brain areas (such as the core area of ​​the language network) during surgery, avoid postoperative cognitive function damage, and improve surgical safety.

[0067] Advantages of technology universality:

[0068] This method breaks through the limitations of single-modal analysis and is applicable to all neuroscience fields that need to explore the brain structure-function synergy mechanism. It is especially irreplaceable in scenarios such as drug efficacy evaluation and cognitive training monitoring.

[0069] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing and analyzing brain networks based on multivariate analysis, characterized in that: The following steps are involved: Step 1: Multimodal brain imaging data acquisition and preprocessing; Multimodal brain imaging data, including structural magnetic resonance imaging (sMRI) and functional magnetic resonance imaging (fMRI), were acquired from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. The acquired sMRI and fMRI data were preprocessed, including format conversion, slice time correction, head motion correction, registration, segmentation, spatial normalization, smoothing, linear drift removal, and filtering operations to eliminate noise, improve data quality, and provide a reliable data foundation for subsequent analysis. Step 2: Unimodal brain network construction based on canonical correlation analysis; A data-driven canonical correlation analysis method was used to construct brain structural and functional networks. For the structural network, nodes were defined using the AAL116 template, and the canonical correlation of gray matter density between brain regions was calculated as the edge weight. For the functional network, nodes were also defined using the AAL116 template, and the canonical correlation of blood oxygen signal time series between brain regions was calculated as the edge weight. The constructed networks were then sparsified to obtain binary structural and functional networks. Step 3: Unimodal brain network analysis; The constructed unimodal brain structural network and functional network were analyzed, and the topological properties of the network were calculated from a graph theory perspective, including small-world properties, global network efficiency, hierarchy, modularity and other indicators. These graph theory properties were compared between patients with cognitive impairment and healthy controls to find the correlation between cognitive memory impairment and brain network abnormalities. At the same time, combined with the morphological analysis of brain structural images and the REHO and ALFF signals extracted from functional images, a significance analysis of brain region differences was performed to identify brain regions that play an important role in brain mechanisms such as cognition and memory.

2. A brain network construction and analysis method based on multivariate analysis according to claim 1, characterized in that: The specific steps of preprocessing in step 1 include: (1) Format conversion: The acquired data is in DICOM format and needs to be converted to NIFTI format for subsequent processing; (2) Slice time correction: During the initial scan, the magnetic field in the device is unstable and the subject is in an adaptation process. Generally, the noise of the first 10 time points of the image is removed. In addition, the scanning process generally adopts the interlayer method, and the image lacks temporal consistency, so the time sequence combination needs to be redefined. (3) Head motion correction: In order to eliminate the influence of the subject's head movement during the experiment, head correction is required, and samples with a movement angle greater than 2° and a horizontal head motion greater than 2 mm are eliminated; (4) Registration: Structural images have higher spatial resolution than functional images, so the structural images are registered to the functional images to improve the accuracy of spatial standardization of the functional images; (5) Segmentation: Segment the processed structural image into gray matter, white matter, and cerebrospinal fluid; (6) Spatial standardization: The images of each subject are different and need to be mapped to a unified standard space. Generally, a transformation matrix is ​​used to map them to the MNI space. (7) Smoothing: Perform Gaussian smoothing on the image to improve the image signal-to-noise ratio. By default, 4*4*4 full-width at half maximum Gaussian noise is used. (8) Eliminate linear drift: Use regression methods to remove the impact of linear drift caused by heating of the acquisition equipment; (9) Filtering: To reduce the impact of high-frequency noise such as heartbeat and breathing, the frequency range is controlled within 0.01 to 0.08 Hz.

3. The method for constructing and analyzing a brain network based on multivariate analysis according to claim 1, characterized in that: During the sparse processing in step 2: the threshold selection strategy is as follows: select the threshold so that the average degree value of the binary network is greater than twice the logarithm of N and the network density is less than 0.5.

Images