Brain structure network weight construction method based on microstructure quantitative characteristics

Abstract

The invention discloses a brain structure network weight construction method based on microstructure quantitative characteristics. The method comprises the steps that firstly, a brain structure fiber bundle sampling point template is obtained through processing according to an MNI space brain fiber bundle template, then a group data set is obtained through processing according to an improved multi-mode magnetic resonance imaging method and the brain structure fiber bundle sampling point template, and then an attention variation auto-encoder model is constructed. And inputting the group data set into the attention variation auto-encoder model for training, finally inputting the obtained vector value of the brain structure feature map of the individual to be detected into the trained attention variation auto-encoder model for processing, and directly taking the processing result as the weight value of the edge in the brain structure network. According to the method, information imbalance caused by length difference of the fiber bundles is overcome, information of various different microstructures of the multi-modal brain is provided, and a new direction is provided for the field of brain structure network weight research.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and particularly relates to a method for constructing the weights of a brain structure network based on quantitative features of microstructures. Background Art

[0002] The brain is the control center of the human body, and its complex structure and functional interactions determine an individual's cognition, behavior, and emotion. In order to study how the brain processes information and its functional performance in different states, the construction and analysis of brain structure networks become particularly important. A brain structure network refers to a complex network formed by different regions of the brain through neural connections (edges), where nodes usually represent anatomical or functional regions in the brain, and edges represent the connections between different regions. The weights of network edges are usually used to measure the connection strength or information transmission ability between different regions.

[0003] In the construction of brain structure networks, the calculation of weights is crucial. Currently, the weight calculation of most brain structure networks relies on macroscopic features, such as the number of streamlines and fiber density of white matter fiber bundles. These methods describe the connection strength between different regions through the analysis of white matter fiber bundles in the brain. However, these calculation methods based on macroscopic features often ignore the information of microstructures and cannot fully reflect the true connections and network structures between different regions of the brain.

[0004] In recent years, with the progress of neuroimaging techniques, more and more studies have begun to use microstructural features (such as microscopic tissue characteristics obtained by techniques like diffusion magnetic resonance imaging) to calculate the weights of brain structure networks. Microstructural features can provide more detailed brain connection information and reveal the complexity of the brain in terms of function and structure. However, most existing methods rely on the averaging process of microstructural features and ignore the rich detailed information of quantitative analysis along fiber bundles. This simplified process may lead to inaccurate capture of individual differences and structural changes, thereby affecting the accuracy and reliability of brain network analysis. Although the current methods for quantitative analysis along fiber bundles have not been applied to the construction of brain structure network weights, they also have certain defects. They adopt the same quantitative number for fiber bundles of different lengths and ignore the information imbalance caused by the length differences of fiber bundles. In addition, traditional brain structure network analysis methods usually rely on static models, cannot fully consider the comprehensive utilization of multimodal data, and ignore the complex relationships between different data sources. Summary of the Invention

[0005] The object of the present invention is to address the deficiencies of the prior art. The present invention provides a method for constructing brain structure network weights based on quantitative features of microstructures or a detection method for related neuroimaging images. The present invention can quantitatively analyze multi-modal brain microstructure features along fiber bundles, and select specific sampling numbers for fiber bundles of different lengths during the quantification process. By using the multi-modal quantitative features of the baseline population, an attention variational autoencoder model is trained, and the model is used to calculate the structural variation intensity of an individual. This structural variation intensity is ultimately defined as the weight of the brain network. This method can be widely applied to multiple fields such as brain structure network analysis and has high application value.

[0006] The technical method adopted by the present invention is as follows:

[0007] S1. Establish a brain data model in a computer. Obtain and process the center template streamline according to the MNI space brain fiber bundle template in sequence, and perform microstructural quantitative sampling method processing on the center template streamline to obtain a brain structure fiber bundle sampling point template.

[0008] The brain data model is a standard brain model obtained from MRI scan results of a large number of normal subjects. The MNI space brain fiber bundle template can be a brain white matter fiber bundle anatomical template in the Standard Neuroimaging Space. The MNI space brain fiber bundle template provides a basis for tracking and segmenting the main fiber bundles in the brain structure, and operations on the MNI space brain fiber bundle template provide a basis for subsequent implementation. The brain structure fiber bundle sampling point template is the sampling points of all template streamlines obtained from the MNI space brain fiber bundle template, which is convenient for directly applying the sampling numbers of all template streamlines to the brain structure feature map after multi-modal magnetic resonance imaging method scanning in subsequent implementation.

[0009] S2. Use an improved multi-modal magnetic resonance imaging method to scan several brain data models in a known database in a computer to obtain several brain structure imaging maps. Each brain structure imaging map is subjected to feature extraction processing to obtain a corresponding brain structure feature map. Comprehensive processing is performed on all brain structure feature maps according to the brain structure fiber bundle sampling point template obtained in step S1 to obtain the vector value of each brain structure feature map, and the vector values of all brain structure feature maps are combined to construct a population data set.

[0010] The brain data models in the database are standardized brain images in a database constructed from a large number of known neuroimaging data.

[0011] S3. Build an attention variational autoencoder model in a computer, and input the population dataset into the attention variational autoencoder model for training to obtain a trained attention variational autoencoder model.

[0012] S4. Use the same method as in step S2 in a computer to obtain the vector value of the brain structure feature map of the individual to be tested, input the vector value of the brain structure feature map of the individual to be tested into the trained attention variational autoencoder model for model processing, and then directly use the result of the model processing as the weight value of the edge in the brain structure network, and finally obtain the weight map of all edges in the brain structure network.

[0013] The weight map of all edges in the brain structure network is a new research object. The present invention provides a method for obtaining the weight map of all edges in the brain structure network for further research in the field of brain structure network weight research.

[0014] The brain structure network is the brain model data established by a computer in cognitive neuroscience, in which a network of human brain structure connections is constructed by using graph theory and network analysis methods. The edge of the brain structure network is the edge structure in the constructed network of human brain structure connections, and the weight map is a graph obtained through the connection relationship and weight values of all edges of the brain structure network.

[0015] The specific content of step S1 is as follows:

[0016] S11. Obtain a number of template fiber bundles, a number of template streamlines in each template fiber bundle, and the length of each template streamline according to the MNI space brain fiber bundle template.

[0017] The template fiber bundle is a three-dimensional model of the fiber bundle obtained from the MNI space brain fiber bundle template. The template streamline is a three-dimensional model of the streamline obtained from the three-dimensional model of the fiber bundle. The length of the template streamline is the length of the streamline in the three-dimensional model. The fiber bundle is a group of white matter fibers in the brain that form inter-cortical or cortical-subcortical connections in the MNI space brain fiber bundle template.

[0018] S12. Perform line center processing on all template streamlines in each template fiber bundle to obtain the central template streamline and the length of the central template streamline in each template fiber bundle.

[0019] The line center processing is to uniformly take m points on all template streamlines in each template fiber bundle, process the i-th taken point of all template streamlines to obtain the central point of all the t-th taken points, and process according to the same method for the central points of the m taken points to obtain m central points, and connect the m central points to obtain the central template streamline.

[0020] S13. According to the length l of the central template streamline of each template fiber bundlei , select the central template streamline with the longest length l from the central template streamlines of all template fiber bundles 0 of the central template streamline, and set the number of sampling points N of the template fiber bundle corresponding to the longest central template streamline 0 .

[0021] S14. Use the microstructural quantitative sampling method to process and obtain the number of sampling points N of other template fiber bundles except the template fiber bundle corresponding to the longest central template streamline i .

[0022] The microstructural quantitative sampling method is set according to the following formula:

[0023]

[0024] where l 0 is the length of the longest central template streamline, l i is the length of the i-th central template streamline except the longest central template streamline, N 0 is the number of samplings of the template fiber bundle where the longest central template streamline is located, N i is the number of sampling points of the i-th template fiber bundle except the template fiber bundle where the longest central template streamline is located.

[0025] S15. In each template fiber bundle, use the number of sampling points N i of the template fiber bundle as the number of sampling points N i of all template streamlines.

[0026] S16. Combine according to the number of sampling points N i of all template streamlines in each template fiber bundle to obtain a sampling point template for the brain structure fiber bundle.

[0027] The sampling point template for the brain structure fiber bundle includes the specific number of sampling points N i of all template streamlines. Subsequently, the brain structure feature map is directly sampled according to the specific number of sampling points N i of all template streamlines in the sampling point template for the brain structure fiber bundle; here, it is default that the structures (fiber bundles and streamlines, etc.) of the sampling point template for the brain structure fiber bundle and the brain structure feature map are the same, and the errors in the lengths of the fiber bundles and streamlines, etc. in both are not considered, so it is directly applied here.

[0028] The specific steps of S2 are as follows:

[0029] S201. Use an improved multi-modal magnetic resonance imaging method to scan a number of brain data models in a known database to obtain a number of brain structure imaging maps, and each brain structure imaging map is processed by feature extraction to obtain a corresponding brain structure feature map.

[0030] The improved multimodal magnetic resonance imaging method scans several brain data models in a known database using the multimodal magnetic resonance imaging method for the first time in the present invention. In the prior art, only a single-modal magnetic resonance imaging method is used for scanning, and then a single feature extraction method of the single-modal magnetic resonance imaging method is used for processing.

[0031] S202. Obtain several fiber bundles and several streamlines in each fiber bundle from each brain structure feature map.

[0032] S203. Perform tracking sampling processing on all streamlines of each fiber bundle in each brain structure feature map according to the brain structure fiber bundle sampling point template obtained in step S1, to obtain the sampling points, the number of sampling points N i and the sampling serial number x of each sampling point.

[0033] The tracking sampling processing is to sample all streamlines of each fiber bundle in each brain structure feature map according to the brain structure fiber bundle sampling point template.

[0034] S204. Divide all sampling points in each fiber bundle according to the same sampling serial number x, to obtain N i groups of serial number sampling points in each fiber bundle.

[0035] S205. Perform point center processing on each group of serial number sampling points, to obtain the central node k of each group of serial number sampling points.

[0036] The point center processing is to process N i sampling points with the same sampling serial number x in the same group to obtain the central points of N i sampling points with the same sampling serial number x in the same group, which are the central nodes k.

[0037] S206. In each group of serial number sampling points, perform Mahalanobis distance processing on N i sampling points and the central node k respectively, to obtain the Mahalanobis distance D(t) from each sampling point to the corresponding central node k.

[0038] The Mahalanobis distance is the covariance distance between two points.

[0039] S207. Perform inverse processing on the Mahalanobis distance D(t) from each sampling point to the corresponding central node k, to obtain the action weight W(t) from each sampling point to the corresponding central node k.

[0040] S208. In each group of serial number sampling points, perform normalization summation processing on the action weights W(t) of N i sampling points to the central node k, to obtain the quantitative value P(k) of each central node k.

[0041] S209. In each fiber bundle in each brain structure feature map, construct the quantitative values P(k) of all central nodes k into a vector P = [P(1), P(2), ···, P(k), ···, P(N i )] T As the vector value of each fiber bundle, combine the vector values of all fiber bundles to obtain the vector value of a brain structure feature map, that is, obtain individual data.

[0042] S210. Combine all individual data to construct a population data set.

[0043] The specific step S201 is as follows:

[0044] For the brain data model of each individual, first obtain magnetic resonance imaging (MRI) maps through at least two MRI methods, and then process them through the feature extraction methods corresponding to each MRI map to obtain several brain structure feature maps; or for the brain data model of each individual, first obtain an MRI map through one MRI method, and then process it through at least two feature extraction methods corresponding to the MRI map to obtain several brain structure feature maps.

[0045] The step S201 only excludes the case where an MRI map is obtained through one MRI method and then processed only through one feature extraction method. Other cases all conform to the situation described in step S201 of the present invention.

[0046] The MRI methods include diffusion MRI method, functional MRI method or quantitative MRI method; the feature extraction methods corresponding to the MRI maps obtained by the diffusion MRI method include diffusion tensor imaging method or neurite orientation dispersion and density imaging method; the feature extraction methods corresponding to the MRI maps obtained by the functional MRI method include low-frequency amplitude analysis method; the feature extraction methods corresponding to the MRI maps obtained by the quantitative MRI method include susceptibility quantitative imaging method or magnetic resonance fingerprinting method.

[0047] The inverse processing in step S207 is set according to the following formula:

[0048]

[0049] where W(t) is the action weight of the sampling point to the corresponding central node, D(t) is the Mahalanobis distance of the sampling point to the corresponding central node, and k is a preset proportional weight value.

[0050] The attention variational autoencoder model in step S3 includes an encoder, a multi-head attention mechanism, and a decoder connected in series in sequence; both the encoder and the decoder adopt the encoder and decoder in the variational autoencoder, and the multi-head attention mechanism adopts the multi-head attention mechanism in the Transformer model.

[0051] Step S4 is specifically as follows:

[0052] S401. Use the same method as in step S2 to obtain the vector value of the brain structure feature map of the individual to be tested.

[0053] S402. Input the vector value of the brain structure feature map of the individual to be tested into the trained attention variational autoencoder model to obtain the structural variation intensity value of each fiber bundle in the brain structure feature map of the individual to be tested.

[0054] S403. Use the structural variation intensity value of each fiber bundle in the brain structure feature map of the individual to be tested as the weight of the edge between the two endpoints of each fiber bundle in the brain network, so as to construct a weight map of all edges in the brain structure network.

[0055] By combining the obtained weight map of all edges in the brain structure network with the functional connection data in the field of the brain structure network, the interaction between the functions and structures of different brain regions can be further explored, and the functional analysis of each brain region in the human brain can be carried out according to the relevant results, which is used for the efficacy evaluation of drugs and applied to the detection, analysis and judgment of CT images.

[0056] The innovation of the present invention lies in adopting reasonable sampling points for fiber bundles of different lengths, using multi-modal magnetic resonance imaging method to extract features of the brain, and for the first time in the present invention, comprehensively using the structural variation intensity value as the weight value of the edge in the brain structure network, realizing a new direction for the research field of brain structure network weights, and bringing the beneficial effect of extracting more brain network feature information.

[0057] The beneficial effects of the present invention are as follows:

[0058] 1. The present invention fully considers the information imbalance caused by the difference in fiber bundle length, and provides an adaptive sampling and quantification scheme for the physical length of the fiber bundle, so that longer fiber bundles can represent more information.

[0059] 2. The present invention adopts an improved multi-modal magnetic resonance imaging technology to overcome the deficiencies of lack and singularity of microstructural information in the brain structure feature map caused by the single-modal method adopted in the traditional method, and provides a variety of different microstructural information of the multi-modal brain.

[0060] 3. For the first time, the present invention uses the structural variation intensity value as a comprehensive method for the weight value of the edges in the brain structure network, providing a new direction for the field of brain structure network weight research and bringing the beneficial effect of extracting more brain network feature information. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of FA and NDI in Example 2 of the present invention.

[0062] Figure 2 It is a schematic diagram of the central streamline and the number of samples of the left IFOF template and the left UF template in Example 2 of the present invention.

[0063] Figure 3 It is a quantitative result diagram of FA and NDI of the left UF in the brain of an individual in Example 2 of the present invention.

[0064] Figure 4 It is a model diagram of the attention variational autoencoder in the embodiment of the present invention.

[0065] Figure 5 It is a schematic diagram of using the variational autoencoder to predict the structural variation intensity and define it as the brain structure network weight in Example 2 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0066] The following further describes the present invention in detail with reference to the drawings and embodiments. It should be noted that the following embodiments are intended to facilitate the understanding of the present invention, but do not limit it in any way.

[0067] The first embodiment of the present invention is carried out according to the following steps:

[0068] S1. Establish a brain data model in a computer and store it in a memory. Perform acquisition processing and line center processing on the brain fiber bundle template in the MNI space in sequence to obtain the central template streamline. Use the processor of the computer to process the microstructural quantitative sampling method according to the central template streamline to obtain the brain structure fiber bundle sampling point template.

[0069] The brain data model is a standard brain model obtained based on the MRI scan results of a large number of normal subjects. The MNI space brain fiber tract template can be a white matter fiber tract anatomical template in the Standard Neuroimaging Space. The MNI space brain fiber tract template provides the basis for the tracking and segmentation of the main fiber tracts in the brain structure, and operations on the MNI space brain fiber tract template provide the basis for subsequent implementation. The brain structure fiber tract sampling point template is the sampling points of all template streamlines obtained from the MNI space brain fiber tract template, which facilitates directly applying the sampling points of all template streamlines to the brain structure feature map after scanning by the multimodal magnetic resonance imaging method in subsequent implementation.

[0070] Step S1 is specifically as follows:

[0071] S11. Obtain a number of template fiber tracts, a number of template streamlines in each template fiber tract, and the length of each template streamline according to the MNI space brain fiber tract template.

[0072] The template fiber tract is a three-dimensional model of the fiber tract obtained from the MNI space brain fiber tract template. The template streamline is a three-dimensional model of the streamline obtained from the three-dimensional model of the fiber tract. The length of the template streamline is the length of the streamline in the three-dimensional model. The fiber tract is a group of white matter fibers in the brain that form inter-cortical or cortical-subcortical connections in the MNI space brain fiber tract template.

[0073] S12. Perform line center processing on all template streamlines in each template fiber tract to obtain the central template streamline of each template fiber tract and the length of the central template streamline.

[0074] The line center processing is to uniformly take m points on all template streamlines in each template fiber tract, process the i-th taken point of all template streamlines to obtain the central point of all the t-th taken points, and process in the same way for the central points of the m taken points to obtain m central points, and connect the m central points to obtain the central template streamline.

[0075] S13. According to the length l of the central template streamline of each template fiber tract i , screen out the central template streamline with the longest length l 0 from the central template streamlines of all template fiber tracts, and set the sampling point number N 0 of the template fiber tract corresponding to the longest central template streamline.

[0076] In specific implementation, the sampling point number N 0 of the template fiber tract corresponding to the longest central template streamline is set to 100.

[0077] S14. Use the microstructural quantitative sampling method to process and obtain the number of sampling points N of other template fiber bundles except for the template fiber bundle corresponding to the longest central template streamline. i .

[0078] The microstructural quantitative sampling method is set according to the following formula:

[0079]

[0080] where l 0 is the length of the longest central template streamline, and l i is the length of the i-th central template streamline except for the longest central template streamline, N 0 is the number of samplings of the template fiber bundle where the longest central template streamline is located, and N i is the number of sampling points of the i-th template fiber bundle except for the template fiber bundle where the longest central template streamline is located.

[0081] S15. In each template fiber bundle, use the number of sampling points N i of the template fiber bundle as the number of sampling points N i for all template streamlines in the template fiber bundle;

[0082] S16. According to the number of sampling points N i of all template streamlines in each template fiber bundle, combine to obtain the sampling point template of the brain structure fiber bundle.

[0083] The sampling point template of the brain structure fiber bundle contains the specific number of sampling points N i of all template streamlines. Subsequently, the brain structure feature map is directly sampled according to the specific number of sampling points N i of all template streamlines in the sampling point template of the brain structure fiber bundle. Here, it is default that the structures (fiber bundles and streamlines, etc.) of the sampling point template of the brain structure fiber bundle and the brain structure feature map are consistent, and the errors in the lengths of the fiber bundles and streamlines, etc. in both are not considered, so direct application is carried out here.

[0084] S2. Use the improved multi-modal magnetic resonance imaging method in the computer to scan the brain data models of several individuals stored in the computer memory in the known database to obtain several brain structure imaging maps. Each brain structure imaging map is subjected to feature extraction processing to obtain the corresponding brain structure feature map. According to the sampling point template of the brain structure fiber bundle obtained in step S1, comprehensively process all brain structure feature maps to obtain the vector value of each brain structure feature map, and combine the vector values of all brain structure feature maps to construct a population data set;

[0085] The brain data models in the database are standardized brain images in the database constructed from a large number of known neuroimaging data.

[0086] Step S2 specifically includes:

[0087] S201. Use an improved multimodal magnetic resonance imaging method to scan the brain data models of several individuals stored in a computer memory's known database to obtain several brain structure imaging maps, and each brain structure imaging map is processed through feature extraction to obtain a corresponding brain structure feature map.

[0088] Step S201 specifically includes:

[0089] For the brain data model of each individual, first obtain a magnetic resonance imaging map through at least two magnetic resonance imaging methods, and then process it through at least one feature extraction method corresponding to each magnetic resonance imaging map to obtain several brain structure feature maps; or for the brain data model of each individual, first obtain a magnetic resonance imaging map through one magnetic resonance imaging method, and then process it through at least two feature extraction methods corresponding to the magnetic resonance imaging map to obtain several brain structure feature maps.

[0090] The magnetic resonance imaging methods include diffusion magnetic resonance imaging method, functional magnetic resonance imaging method, or quantitative magnetic resonance imaging method; the feature extraction methods corresponding to the magnetic resonance imaging maps obtained by the diffusion magnetic resonance imaging method include diffusion tensor imaging method or neurite orientation dispersion and density imaging method; the feature extraction methods corresponding to the magnetic resonance imaging maps obtained by the functional magnetic resonance imaging method include low-frequency amplitude analysis method; the feature extraction methods corresponding to the magnetic resonance imaging maps obtained by the quantitative magnetic resonance imaging method include susceptibility quantitative imaging method or magnetic resonance fingerprinting method.

[0091] S202. Obtain several fiber bundles and several streamlines in each fiber bundle from each brain structure feature map.

[0092] S203. According to the brain structure fiber bundle sampling point template obtained in step S1, perform tracking sampling processing on all streamlines of each fiber bundle in each brain structure feature map to obtain the sampling points, the number of sampling points N i and the sampling serial number x of each sampling point.

[0093] In specific implementation, it is defaulted that the brain structures in the MNI space brain fiber bundle template are consistent with the brain structures in the database, and the structural differences are ignored.

[0094] The tracking sampling processing is to sample all streamlines of each fiber bundle in each brain structure feature map according to the brain structure fiber bundle sampling point template. In specific implementation, when performing the tracking sampling processing, each streamline is sampled simultaneously from the same side, and all streamlines will simultaneously generate the first sampling point, the second sampling point, until the last sampling point, and the x of the xth point is the serial number of this sampling point.

[0095] S204. Divide all the sampling points in each fiber bundle in each brain structure feature map according to the same sampling serial number x to obtain N i groups of serial number sampling points in each fiber bundle.

[0096] In specific implementation, take the first sampling point of all the streamlines in each fiber bundle as the first group of serial number sampling points, take the second sampling point of all the streamlines as the second group of serial number sampling points, and so on until the sampling ends.

[0097] S205. Perform point centering processing on each group of serial number sampling points in each fiber bundle in each brain structure feature map to obtain the central node k of each group of serial number sampling points.

[0098] Point centering processing is to process N i sampling points with the same group sampling serial number x to obtain the central point of N i sampling points with the same group sampling serial number x, which is the central node k.

[0099] S206. In each group of serial number sampling points, perform Mahalanobis distance processing on N i sampling points and the central node k respectively to obtain the Mahalanobis distance D(t) from each sampling point to the corresponding central node k.

[0100] The Mahalanobis distance is the covariance distance between two points.

[0101] S207. Perform inverse ratio processing on the Mahalanobis distance D(t) from each sampling point to the corresponding central node k to obtain the action weight W(t) from each sampling point to the corresponding central node k.

[0102] The inverse ratio processing is set according to the following formula:

[0103]

[0104] where W(t) is the action weight from the sampling point to the corresponding central node, D(t) is the Mahalanobis distance from the sampling point to the corresponding central node, and k is a preset proportional weight value.

[0105] S208. In each group of serial number sampling points, perform normalization summation processing on the action weights W(t) of N i sampling points to the central node k to obtain the quantitative value P(k) of each central node k.

[0106] S209. In each fiber bundle in each brain structure feature map, construct the quantitative values P(k) of all the central nodes k into a vector P = [P(1), P(2), ···, P(k), ···, P(N i )] TAs the vector value of each fiber bundle, the vector values of all fiber bundles are combined to obtain the vector value of a brain structure feature map, that is, individual data is obtained.

[0107] S210. Combine all individual data to construct a population data set.

[0108] S3. Construct an attention variational autoencoder model in a computer, and input the population data set into the attention variational autoencoder model for training to obtain a trained attention variational autoencoder model.

[0109] The attention variational autoencoder model in step S3 includes an encoder, a multi-head attention mechanism, and a decoder connected in series in sequence; both the encoder and the decoder adopt the encoder and decoder in the variational autoencoder, and the multi-head attention mechanism adopts the multi-head attention mechanism in the Transformer model.

[0110] S4. Use the same method as in step S2 in a computer to obtain the vector value of the brain structure feature map of the individual to be tested, input the vector value of the brain structure feature map of the individual to be tested into the trained attention variational autoencoder model for model processing, and then directly use the result of the model processing as the weight value of the edge in the brain structure network, and finally obtain the weight map of all edges in the brain structure network.

[0111] The brain structure network is the brain model data established by a computer in cognitive neuroscience, in which a network of human brain structure connections is constructed using graph theory and network analysis methods. The edges of the brain structure network are the edge structures in the constructed network of human brain structure connections, and the weight map is a graph obtained from the connection relationships and weight values of all edges in the brain structure network. By default, the fiber bundles of the brain structure feature map of the individual to be tested are regarded as an edge of the brain structure network, and so on. All fiber bundles have corresponding edges in the brain structure network, thus constructing a network graph, and the structural variation intensity value of the fiber bundle is used as the weight value of the corresponding edge, and finally the weight map of all edges in the brain structure network is obtained.

[0112] Step S4 is specifically as follows:

[0113] S401. Use the same method as in step S2 to obtain the vector value of the brain structure feature map of the individual to be tested.

[0114] S402. Input the vector value of the brain structure feature map of the individual to be tested into the trained attention variational autoencoder model to obtain the structural variation intensity value of each fiber bundle in the brain structure feature map of the individual to be tested.

[0115] S403. Use the structural variation intensity value of each fiber bundle in the brain structure feature map of the individual to be tested as the weight of the edge between the two endpoints of each fiber bundle in the brain network, thereby constructing the weight map of all edges in the brain structure network.

[0116] By default, the fiber bundle of the brain structure feature map of the individual to be measured is regarded as an edge of the brain structure network. By analogy, all fiber bundles have corresponding edges in the brain structure network, thus constructing a network graph. The structural variation intensity value of the fiber bundle is used as the weight value of the corresponding edge, and finally the weight map of all edges in the brain structure network is obtained.

[0117] Example 2

[0118] In this example, 108 subjects were selected as the baseline group, including 71 males and 52 females, with an average age of 18.33 years. In addition, 1 subject was selected as the test individual.

[0119] Select the MNI-space brain fiber bundle template proposed by Garyfallidis et al. For fiber bundle i in the template, calculate the geometric center of all streamlines within the fiber bundle, defined as the central streamline, and obtain the length l of the central streamline of each fiber bundle in the template. i 。

[0120] Taking the left inferior frontal-occipital fasciculus (IFOF) as the reference for the microstructural quantitative sampling method and the left uncinate fasciculus (UF) as the research object. For the sampling number N 1 of the microstructural quantitative sampling method is set according to the following formula:

[0121]

[0122] where l 0 is the length of the central template streamline of the left IFOF, N 0 is the sampling number of the left IFOF, l 1 is the length of the central template streamline of the left UF, and N 1 is the sampling number of the left UF.

[0123] Substitute the length l 0 = 151.37 mm of the central template streamline of the left IFOF, the sampling number N 0 = 100 of the left IFOF, and the length l 1 = 77.44 mm of the central template streamline of the left UF into the formula of the microstructural quantitative sampling method, and obtain the sampling number N 1 = 51 of the left UF. As Figure 2 shown, 1 is the left IFOF, 2 is the left UF, 3 is the central streamline of the left IFOF, with a length of 151.37 mm, and the quantitative sampling number N 0is 100, 4 is the central streamline of the left UF, its length is 77.44 mm, and the obtained number of samples N 1 is 51. The number of samples obtained by the above method corresponds to the actual length of the fiber bundle, avoiding the information imbalance caused by the fiber bundle length difference during the subsequent sampling along the fiber bundle.

[0124] Magnetic resonance scanning was completed on a Siemens 3T Prisma machine. The acquired sequences included T1-weighted imaging (T1W) and diffusion-weighted imaging (DWI). T1W used the MPRAGE sequence, voxel size = 1×1×1 mm 3 . DWI used the SMS-EPI sequence, voxel size = 1.5×1.5×1.5 mm 3 , GRAPPA factor = 2, SMS factor = 3, b values were taken as 1000, 2000, and 3000 s / mm 3 , with 30 directions each, providing the phase-encoding direction PA and the anti-phase-encoding direction AP to reduce artifacts and improve the signal-to-noise ratio.

[0125] The DWI images were denoised by PCA, Gibbs artifacts were removed, eddy current correction and distortion correction were performed using MRtrix3 and FSL software. Two different feature extraction methods based on DWI could be selected to obtain brain structure feature maps. The diffusion tensor model was constructed using Mrtrix3 software to calculate the fractional anisotropy (FA) brain structure feature map. The neurite orientation dispersion and density imaging (NODDI) was constructed using the Matlab NODDI toolbox to calculate the axonal density fraction (NDI) brain structure feature map. Figure 1 Schematic diagrams of the FA brain structure feature map and the NDI brain structure feature map of the test individual are shown.

[0126] As Figure 1 , they are the schematic diagrams of FA and NDI. On the left is FA, and on the right is NDI. The brighter the voxel color, the higher the value at that location.

[0127] The T1W image was registered with the DWI image to obtain the fiber orientation distribution function, so as to obtain the distribution of all fiber bundles in the brain of the individual in the database. In this embodiment, the left UF in the brain of the individual was taken as an example for operation:

[0128] The number of samples N of the left UF obtained according to the previous brain fiber tract template in the MNI space 1 The left UF in the brains of 51 pairs of individuals was sampled.

[0129] As Figure 3 shown, Figure 3 a shows the relative position of the left UF in the human brain. The starting sampling point is position 1 near the lower side, and the ending sampling point is position 51 near the upper side.

[0130] All streamlines j of the left UF were evenly divided into N 1 sampling points, and the geometric center of all streamlines at each group of sampling points was calculated to obtain N 1 central nodes, which formed the central streamline; for each central node k, the Mahalanobis distance D(t) between each sampling point and the corresponding central node was calculated, and the action weight W(t) of this sampling point on the central node was inversely proportional to D(t).

[0131] For the FA brain structural feature map and the NDI brain structural feature map, the sampled values of the feature map were weighted and mapped to the corresponding central node k with the weight W(t), and the weighted sampled values were normalized and summed to obtain the quantitative value P(k) of this central node.

[0132] The vector value of the left UF in the FA brain structural feature map can be expressed as the vector P FA =[P FA (1), P FA (2), ···, P FA (k), ···, P FA (N 1 )] T , and the vector value of the left UF in the NDI brain structural feature map can be expressed as the vector P NDI =[P NDI (1), P NDI (2), ···, P NDI (k), ···, P NDI (N 1 )] T .

[0133] The quantitative results of the left UF of the individual samples are as Figure 3 shown, Figure 3 a shows the spatial position of the left UF in the brain, Figure 3As shown in Fig. b, it is the quantitative result diagram of FA and NDI. The abscissa represents the sampling point position. The 1st - 51st points are the quantitative results of FA, and the 52nd - 102nd points are the quantitative results of NDI. The ordinate is the numerical value of FA and NDI. It can be seen from Fig. a that 51 quantitative values are obtained by quantitative analysis from bottom to top along the left - hand UF. Fig. b shows that 51 features are obtained for each of the two micro - structural features, FA and NDI, and the manifestation forms of these two micro - structural features have their own characteristics, demonstrating the multi - modal micro - structural information of the brain.

[0134] Use the same method as above to process and obtain the vector values of all fiber bundles. Combine the vector values of all fiber bundles to get the vector value of a brain structure feature map, that is, obtain individual data. Combine all individual data to construct a population data set.

[0135] Construct an attention variational auto - encoder model, aiming to represent the brain structure through multi - modal micro - structural quantitative features, as Figure 4 shown. The attention variational auto - encoder model includes an encoder, a multi - head attention mechanism, and a decoder connected in series in sequence. Both the encoder and the decoder adopt the encoder and decoder in the variational auto - encoder, and the multi - head attention mechanism adopts the multi - head attention mechanism in the Transformer model. The specific steps are as follows:

[0136] First, the input data of the model comes from a population data set containing n = 108 samples. Each sample obtains R = 2 brain structure feature maps through the above steps. For the R * N i quantitative values of fiber bundle i, where N i represents the number of sampling points on fiber bundle i. The shape of the input data of the model is (n, R * N i ), that is, each input sample has R * N i input features.

[0137] Perform normalization processing on the input data, set the mean of each feature to 0 and the variance to 1, so as to eliminate the influence of different feature dimensions and scale differences on model training. The encoder contains 1 fully - connected layer and a self - attention layer. In order to balance the correlation between features, the input data is processed through the multi - head self - attention mechanism of the Transformer to generate query (Query), key (Key), and value (Value) matrices, calculate the dot product of the query and the key to obtain the attention weights, and use these weights to perform weighted averaging on the input features, thereby generating a weighted input representation, enabling the model to automatically focus on the most relevant input features.

[0138] Map the weighted input representation to the latent space, and output the mean μ and variance σ 2, and then generate the latent variable z; introduce the reparameterization trick, and represent the sampling process in the latent space as:

[0139] z = μ + σ·ε

[0140] where ε is the noise sampled from the standard normal distribution N(0, 1).

[0141] During backpropagation, gradient updates are performed on the latent space, so that the VAE can generate higher-quality reconstructed data by optimizing the distribution of the latent space. The latent variable z is input into the decoder, and after passing through 1 fully connected layer, the reconstructed representation of the input data is generated. The loss function consists of two parts: the reconstruction error and the KL divergence (Kullback-Leibler Divergence). The reconstruction error is used to measure the difference between the input data and its reconstructed data, and the KL divergence is used to measure the difference between the latent space distribution and the standard normal distribution. During training, He initialization and the Adam optimizer are used, and the learning rate α = 10 -3 , set the maximum number of training epochs Epoch = 200, and stop training when the value of the loss function is less than 0.01 or the maximum number of training epochs is reached.

[0142] For a single test individual, according to the obtained R = 2 brain structure feature maps, for the R*N of the left UF of the fiber bundle i microstructure quantitative values, denote the input feature as x; input the individual's feature data into the encoder part of the pre-trained variational autoencoder. The encoder processes the input data through multiple fully connected layers and self-attention layers to generate the mean and variance of the latent space, which is used to represent the low-dimensional features of the input data; based on the mean and variance, the reparameterization trick is used to sample the latent variable z from the Gaussian distribution in the latent space; the latent variable z is input into the decoder part of the variational autoencoder, and the decoder maps the latent variable back to the original input data space through multiple fully connected layers to generate the reconstruction result of the feature data of the individual to be predicted, as the prediction output x', as Figure 5 shown.

[0143] The structural variation intensity of the individual is the mean square error (MSE) between the true value and the predicted value, and the calculation formula is as follows:

[0144]

[0145] where, x j is the i-th eigenvalue of the input data, x j ′ is the predicted value generated by the decoder, and m is the number of input features. A higher MSE value indicates that the individual deviates more from the training data distribution, indicating that the individual may have more significant structural variations, which in turn affect the weights of the edges in the brain structure network.

[0146] The brain regions connected by the left UF are the orbital part of the prefrontal lobe and the inferior temporal gyrus. As Figure 5 shown, in the construction of the brain structure network, the two brain regions are abstracted as node 5 and node 7, and the structural variation intensity is defined as the weight 6 of the connection edge between the two nodes.

[0147] By combining the weight map of all edges in the obtained brain structure network with the functional connection data in the field of brain structure network, the interaction between the functions and structures of different brain regions can be further explored. The functional analysis of each brain region in the human brain can be carried out according to the relevant results, which can be used for the efficacy evaluation of drugs and applied to the detection, analysis and judgment of CT images.

[0148] The above results verify the feasibility of the present invention and can innovatively provide a brain structure network method.

[0149] The weight map of all edges in the brain structure network is a new research object. The present invention finally provides a method for obtaining the weight map of all edges in the brain structure network for further research in the field of brain structure network weight research.

[0150] The innovation of the present invention lies in adopting reasonable sampling points for fiber bundles of different lengths, using multimodal magnetic resonance imaging method to extract features of the brain, and for the first time in the present invention, taking the structural variation intensity value as the comprehensive method of the weight value of the edge in the brain structure network, realizing providing a new direction for the field of brain structure network weight research and bringing the beneficial effect of extracting more brain network feature information.

[0151] The above embodiments are only two embodiments of the present invention and do not limit the scope of implementation of the present invention. Therefore, all changes made according to the method of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for constructing brain structure network weights based on microstructure quantitative features, characterized in that: The following steps are involved: S1. Establish a brain data model in a computer, perform acquisition processing and line center processing in sequence according to the MNI spatial brain fiber bundle template to obtain the central template streamline, perform microstructure quantitative sampling method processing according to the central template streamline, and obtain the brain structure fiber bundle sampling point template; S2. Using an improved multimodal magnetic resonance imaging method in a computer to scan several brain data models in a known database to obtain several brain structure imaging images, each brain structure imaging image is subjected to feature extraction processing to obtain a corresponding brain structure feature map, all brain structure feature maps are comprehensively processed according to the brain structure fiber bundle sampling point template obtained in step S1 to obtain a vector value of each brain structure feature map, and the vector values ​​of all brain structure feature maps are combined to construct a group data set; S3. constructing an attention variational autoencoder model in a computer, and inputting the group data set into the attention variational autoencoder model for training to obtain a trained attention variational autoencoder model; S4. Use the same method as step S2 in the computer to obtain the vector value of the brain structure feature map of the individual to be tested, input the vector value of the brain structure feature map of the individual to be tested into the trained attention variational autoencoder model for model processing, and then directly use the result of the model processing as the weight value of the edge in the brain structure network, and finally obtain the weight map of all edges in the brain structure network.

2. The method for constructing brain structure network weights based on microstructure quantitative features according to claim 1, characterized in that: The step S1 is specifically as follows: S11, obtaining a number of template fiber bundles, a number of template streamlines in each template fiber bundle, and the length of each template streamline according to the MNI spatial brain fiber bundle template; S12, performing line center processing on all template streamlines in each template fiber bundle to obtain a center template streamline and a length of the center template streamline; The line center processing is to uniformly select m points for all template streamlines in each template fiber bundle, process the i-th point of all template streamlines to obtain the center point of all t-th points, process the center point of the t-th point in the same way to obtain m center points, and connect the m center points to obtain the center template streamline; S13, according to the length l of the center template streamline of each template fiber bundle i , select the central template streamline with the longest length of l0 from the central template streamlines of all template fiber bundles, and set the number of sampling points N0 of the template fiber bundle corresponding to the longest central template streamline; S14. The number of sampling points N of other template fiber bundles except the template fiber bundle corresponding to the longest central template streamline is obtained by using the microstructure quantitative sampling method. i ; The microstructure quantitative sampling method is set according to the following formula: Among them, l0 is the length of the longest center template streamline, l i is the length of the i-th center template streamline except the longest center template streamline, N0 is the sampling number of the template fiber bundle where the longest center template streamline is located, N i is the number of sampling points of the i-th template fiber bundle except the template fiber bundle where the longest central template streamline is located; S15. In each template fiber bundle, the number of sampling points N of the template fiber bundle is i As the number of sampling points N for all template streamlines i ; S16, according to the number of sampling points N of all template streamlines in each template fiber bundle i The brain structure fiber bundle sampling point template is obtained by combination.

3. The method for constructing brain structure network weights based on microstructure quantitative features according to claim 1, characterized in that: The step S2 is specifically as follows: S201, using an improved multimodal magnetic resonance imaging method to scan a number of brain data models in a known database to obtain a number of brain structure imaging images, and performing feature extraction processing on each brain structure imaging image to obtain a corresponding brain structure feature image; S202, acquiring a plurality of fiber bundles and a plurality of streamlines in each fiber bundle from each brain structure feature map; S203, according to the brain structure fiber bundle sampling point template obtained in step S1, all streamlines of each fiber bundle in each brain structure feature map are traced and sampled to obtain the sampling points of each streamline and the number of sampling points N. i And the sampling number x of each sampling point; S204, divide all sampling points in each fiber bundle according to the same sampling sequence number x to obtain N sampling points in each fiber bundle. i Group number sampling point; S205, performing point center processing on each group of serial number sampling points to obtain the central node k of each group of serial number sampling points; The point center is processed as N for the same group of sampling sequence number x i The sampling points are processed to obtain N i The center point of the sampling points is the central node k; S206. In each group of serial numbered sampling points, i Each sampling point is processed with the central node k for Mahalanobis distance, and the Mahalanobis distance D(t) from each sampling point to the corresponding central node k is obtained; S207, performing inverse processing on the Mahalanobis distance D(t) from each sampling point to the corresponding central node k to obtain the action weight W(t) from each sampling point to the corresponding central node k; S208. In each group of serial numbered sampling points, i The weights W(t) of the actions of the sampling points to the central node k are normalized and summed to obtain the quantitative value P(k) of each central node k; S209. In each fiber bundle, construct the quantitative values ​​P(k) of all central nodes k into a vector P = [P(1), P(2), . . . , P(k), . . . , P(N i )] T As the vector value of each fiber bundle, the vector values ​​of all fiber bundles are combined to obtain a vector value of a brain structure feature map, that is, to obtain individual data; S210, combining all individual data to construct a group data set.

4. The method for constructing brain structure network weights based on microstructure quantitative features according to claim 3, characterized in that: The step S201 is specifically as follows: The brain data model of each individual is first subjected to at least two magnetic resonance imaging methods to obtain a magnetic resonance imaging image, and then processed by the feature extraction methods corresponding to each magnetic resonance imaging image to obtain a number of brain structure feature maps; or the brain data model of each individual is first subjected to a magnetic resonance imaging method to obtain a magnetic resonance imaging image, and then processed by at least two feature extraction methods corresponding to the magnetic resonance imaging image to obtain a number of brain structure feature maps.

5. The method for constructing brain structure network weights based on microstructure quantitative features according to claim 4, characterized in that: The magnetic resonance imaging method includes a diffusion magnetic resonance imaging method, a functional magnetic resonance imaging method or a quantitative magnetic resonance imaging method; the feature extraction method corresponding to the magnetic resonance imaging image obtained by the diffusion magnetic resonance imaging method includes a diffusion tensor imaging method or a neurite direction dispersion and density imaging method; the feature extraction method corresponding to the magnetic resonance imaging image obtained by the functional magnetic resonance imaging method includes a low-frequency amplitude analysis method; the feature extraction method corresponding to the magnetic resonance imaging image obtained by the quantitative magnetic resonance imaging method includes a magnetic susceptibility quantitative imaging method or a magnetic resonance fingerprint imaging method.

6. The method for constructing brain structure network weights based on microstructure quantitative features according to claim 3, characterized in that: The inverse ratio processing in step S207 is set according to the following formula: Wherein, W(t) is the weight of the action from the sampling point to the corresponding central node, D(t) is the Mahalanobis distance from the sampling point to the corresponding central node, and k is the preset proportional weight value.

7. The method for constructing brain structure network weights based on microstructure quantitative features according to claim 1, characterized in that: The attention variational autoencoder model in step S3 includes an encoder, a multi-head attention mechanism and a decoder connected in series in sequence; the encoder and the decoder both adopt the encoder and the decoder in the variational autoencoder, and the multi-head attention mechanism adopts the multi-head attention mechanism in the Transformer model.

8. The method for constructing brain structure network weights based on microstructure quantitative features according to claim 1, characterized in that: The step S4 is specifically as follows: S401, using the same method as step S2 to obtain the vector value of the brain structure characteristic map of the individual to be tested; S402, inputting the vector value of the brain structure feature map of the individual to be tested into the trained attention variational autoencoder model to obtain the structural variation intensity value of each fiber bundle in the brain structure feature map of the individual to be tested; S403, using the structural variation intensity value of each fiber bundle in the brain structure characteristic map of the individual to be tested as the weight of the edge between the two endpoints of each fiber bundle in the brain network, thereby constructing a weight map of all edges in the brain structure network.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images