Method for dividing cerebral cortex super-vertex based on T1 magnetic resonance image

Abstract

The application discloses a brain cortex super-vertex division method based on T1 magnetic resonance imaging, which comprises the following steps: S1, establishing a data set and performing pretreatment; S2, calculating vertex connection in a brain region to obtain a vertex connection matrix; S3, obtaining a group original brain region super-vertex space; S4, further obtaining a group whole brain super-vertex space according to step S3; S5, completing individualized super-vertex division according to the divided group super-vertex label; S6, constructing different group characteristic super-vertex connection networks and individualized characteristic super-vertex connection networks according to the vertex characteristics of the brain atlas space and the division results of the group level super-vertex and the individualized super-vertex; and S7, verifying the effectiveness of the super-vertex division according to a machine learning algorithm. Compared with the existing brain atlas, the method greatly improves the resolution of brain division, does not need to determine the number of sub-regions, and can automatically perform brain region division to construct super-vertices.

Description

TECHNICAL FIELD

[0001] The present application relates to the biomedical technology field, and in particular to a method for dividing cerebral cortex hyper-vertices based on T1 magnetic resonance images. BACKGROUND

[0002] The brain is considered to be the most complex and mysterious organ of the human body, and the research on the function and structure of the brain has a long history. It is generally believed that the function of the brain depends largely on the structure, and the function of different regions with different structural organizations may also be different, so the fine study of the structure of the brain regions is of great significance for people to understand the function of the brain. Brain atlas is a basic method for studying the structure and function of different regions of the brain, and in recent years, many studies have constructed brain atlas by structural magnetic resonance and functional magnetic resonance.

[0003] Based on functional magnetic resonance images, functional brain atlas mainly defines the functional division of the brain through functional connection characteristics; based on structural magnetic resonance images, structural brain atlas mainly defines the structural division of the brain through white matter fiber connection characteristics. Generally speaking, due to the characteristics of functional magnetic resonance images, the spatial resolution of brain function atlas is low; due to the non-uniformity of tensor reconstruction and fiber tracking algorithm, the interpretability of brain structure atlas still needs to be explored. In addition to functional magnetic resonance imaging and diffusion tensor imaging, T1 magnetic resonance structural image has gradually become a new hotspot in brain connectivity research. In recent years, brain connectivity research based on T1 magnetic resonance structural image mainly focuses on the morphological characteristics of brain regions, and analyzes the brain region connection network through the Pearson correlation between the morphological characteristics of brain regions, which ignores the relationship between the morphological characteristics of the internal vertices of the brain regions.

[0004] As an important tool for studying brain science, brain atlas can help us understand the relationship between brain function and structure, and can also provide great help for clinical diagnosis and treatment. However, the brain atlas constructed by various existing methods has many shortcomings: the scale of brain division is not fine enough, and the relationship between the division and brain diseases is not clear; the number of brain divisions is difficult to determine, and the brain regions cannot be automatically segmented, and prior knowledge is needed to mark the brain regions. SUMMARY

[0005] In view of the shortcomings of the existing method, the purpose of the present application is to provide a method for dividing cerebral cortex hyper-vertices based on T1 magnetic resonance images, which is more detailed on the basis of the existing brain atlas and automatically divides the hyper-vertices to solve the problem of fine positioning of the brain regions.

[0006] In order to solve the above technical problems, the technical scheme of the present application is as follows:

[0007] A method for dividing cerebral cortex hyper-vertices based on T1 magnetic resonance images, comprising the following steps:

[0008] S1, obtaining T1 magnetic resonance brain image data to establish a data set, and preprocessing;

[0009] S2, calculating vertex connection in brain region according to the vertex of brain region division of brain atlas, obtaining vertex connection matrix;

[0010] Wherein the brain atlas is a brain atlas disclosed in the prior art, and the brain atlas selected in the Talairach space is: Desikan / KillianyAtlas can be divided into 68 brain regions, and DestrieuxAtlas can be divided into 148 brain regions. The brain atlas selected in the MNI space is: AAL (automated anatomical labeling) brain atlas, a brand-new human brain atlas drawn by the Institute of Automation, Chinese Academy of Sciences (http: / / atlas.brainnetome.org / ).

[0011] S3, obtaining average vertex connection matrix from vertex connection matrix of the same brain region in the whole data set, clustering the average vertex connection matrix to obtain group level super vertex, mapping the group level super vertex to the corresponding original brain region, and then obtaining group original brain region super vertex space;

[0012] Using the average vertex connection matrix of all samples in the data set as the basis for super vertex division, the difference of individual sample brain space is reduced, and the division result is the group brain cortex super vertex space of the data set.

[0013] S4, mapping the group level super vertex of each brain region of the brain atlas obtained in step S3 to the corresponding original brain region, and then obtaining the group whole brain super vertex space;

[0014] S5, completing individualized super vertex division according to the divided group super vertex label;

[0015] S6, constructing different group feature super vertex connection network and individualized feature super vertex connection network from the division results of group level super vertex and individualized super vertex according to the vertex characteristics of brain atlas space;

[0016] The group super vertex connection network represents the standard connection of the brain space of all samples in the data set, and the connection mode is uniform; the individualized super vertex connection reflects the differential connection of each sample, and the connection mode is different.

[0017] S7, verifying the effectiveness of the hyper-vertex division according to a machine learning algorithm. Based on two hyper-vertex connection networks: a group whole-brain hyper-vertex connection network and an individualized whole-brain hyper-vertex connection network, the hyper-vertex connection of each sample in the data set is extracted as a feature vector, and a machine learning method is used to analyze the correlation between the pathological indicators of the sample and the hyper-vertex connection, to verify whether the divided hyper-vertices have a guiding effect on the evaluation of the disease.

[0018] Preferably, in S1, the preprocessing method of the data set is to segment the T1 magnetic resonance image data using an automatic segmentation tool, and then obtain the morphological features of the vertices on the T1 magnetic resonance image data.

[0019] The automatic segmentation tool can be selected from freesurfer (http: / / www.freesurfer.net / ) software, FastSurfer (https: / / fastsurfer.readthedocs.io / en / latest / ) software, and CAT12 toolkit of SPM (Statistical Parametric Mapping, https: / / www.fil.ion.ucl.ac.uk / spm / ). Different toolkits use different brain atlases and extract different morphological features after preprocessing, and different combinations of vertex morphological features calculate different vertex connections. https: / / deep-mi.org / research / fastsurfer / The preprocessing process using freesurfer software is as follows: (1) registering the individual raw image to Talairach space; (2) non-uniform field correction; (3) stripping non-brain tissue structures such as skull and scalp; (4) segmenting the brain into tissues such as gray matter, white matter and cerebrospinal fluid; (5) forming the inner and outer surfaces of the cortical image between the white matter outer surface and the gray matter outer surface; (6) using vertices as the basic unit of cortical image analysis, and completing morphological feature extraction such as cortical thickness, cortical thickness standard deviation, surface area, mean curvature, Gaussian curvature, curvature index, complexity index, gray matter volume, and sulcal depth on the inner and outer surface vertices of the cortical image.

[0020] The preprocessing process using the CAT12 toolkit of SPM software is as follows: (1) stripping non-brain tissue structures such as skull and scalp; (2) segmenting the brain into tissues such as gray matter, white matter and cerebrospinal fluid; (3) registering the subject image to the MNI (Montreal Neurological Institute) standard space; (4) modulating the influence of spatial normalization; (5) using a public brain atlas to label the partitions of each subject's brain in the standard space; (6) extracting morphological features such as cortical thickness, sulcal depth, cortical folding, and cortical complexity.

[0021] Preferably, the brain regions divided by the brain atlas are used as spatial constraints for vertex connection.

[0022] Preferably, the brain regions divided by the brain atlas are used as spatial constraints for vertex connection.

[0023] As preferred, the vertex connection is the distance between all vertices in the same brain region, and the distance calculation method can adopt: Euclidean distance, city block distance, Chebyshev distance, cosine distance, Hamming distance, Mahalanobis distance, Jaccard distance, etc.

[0024] As preferred, in S2, the calculation method of obtaining the vertex connection matrix is as follows:

[0025] Vertex: V = {feature 1, feature 2, feature 3, …, feature m};

[0026] Vertex set: {V1, V2, V3, …, Vn};

[0027] Vertex connection: dij = Distance (Vi, Vj);

[0028] Vertex connection matrix:

[0029] Where the vertex feature is the morphological feature of the preprocessed vertex, the vertex set is the initial vertex in a brain region in the brain atlas space, and Distance is the distance calculation method of the vertex. The distance calculation method can be selected from: Euclidean distance, city block distance, Chebyshev distance, cosine distance, Hamming distance, Mahalanobis distance, Jaccard distance, etc.

[0030] As preferred, the average vertex connection matrix is obtained by adding the vertex connection matrices of all samples in the same brain region in the data set and dividing by the number of samples. The average vertex connection matrix is the basic unit of the group level super vertex division, and the group level super vertex obtained based on the average vertex connection matrix is a standard super vertex space suitable for all samples in the data set.

[0031] As preferred, in S3, the average vertex connection matrix is clustered by a community discovery algorithm. In addition, a variety of clustering algorithms can be selected, such as traditional clustering algorithms: Kmeans clustering, DBSCAN clustering, spectral clustering, etc.; community discovery algorithms based on graph theory: Fastgreedy algorithm, Edge-betweenness algorithm, Leading eigenvector algorithm, Multilevel algorithm, etc.; deep learning clustering or graph segmentation algorithms, such as Deep Clustering Network (DCN), Deep Embedding Network (DEN), Deep Subspace Clustering Networks (DSC-Nets), Deep Multi-Manifold Clustering (DMC), etc.

[0032] As preferred, the partition method of the individualized super-vertex is: mapping the group-level super-vertex to the space of the brain atlas as the initial label of the vertex, calculating the vertex connection of the space of the brain atlas to obtain the whole brain vertex connection network, and implementing the label propagation algorithm on the basis of the whole brain vertex connection network to complete the partition of the individualized super-vertex.

[0033] The partition of the individualized super-vertex is based on the group super-vertex space, the group super-vertex division result is mapped to the individual brain space, each vertex of the individual is marked with the group super-vertex as the initial label, and then the label propagation algorithm is used to iteratively update each vertex in turn. For each vertex, the label of the neighbor vertex is counted, the label with the most label number is selected to update the vertex, if the most label number is greater than one, a label is randomly selected from the most label number to update the vertex, and the iteration is stopped until convergence. The label of all vertices in the individual brain space is updated to the new super-vertex label, and the partition of the individualized super-vertex is completed.

[0034] The partition method of the individualized super-vertex takes the group brain cortex super-vertex space as the prior starting point, calculates the vertex connection based on the magnetic resonance structure, morphology or function and multi-modal features of the individual sample, and performs label propagation on the basis of the vertex connection network of different features, so as to divide the individualized super-vertex. This method is generally applicable to the individualized brain region sub-region (super-vertex) division using the magnetic resonance structure, morphology or function and multi-modal features.

[0035] As preferred, the vertex feature is any one of the brain cortex structure, morphology, function and multi-modal feature.

[0036] Based on the group super-vertex division result, the group super-vertex division result is mapped to the individual brain space, the vertex of the individual brain space is labeled according to the group super-vertex, the super-vertex feature of each sample is extracted, and the group super-vertex connection network is constructed in the individual brain space. Based on the group super-vertex space, all samples in the data set have the same super-vertex connection network, and the super-vertex feature of the sample determines the super-vertex connection of the sample. Based on the individualized super-vertex division result, the individual super-vertex feature is extracted, and the individualized super-vertex connection network is constructed.

[0037] The present application has the following characteristics and beneficial effects:

[0038] By adopting the above technical solution, the vertices in the structural magnetic resonance image are automatically divided into different categories according to the community discovery algorithm, the super-vertex is defined as the same type of vertex in the structural magnetic resonance image, and all information of the same type of vertex is aggregated in each super-vertex, so that the structural information of the cerebral cortex can be refined, the accurate positioning of the brain partition is facilitated, the reference for the evaluation and clinical diagnosis of brain-related diseases is provided, and the present application can be applied to different brain image data sets and is helpful for the evaluation and analysis of the relationship between brain diseases and refined brain regions.

[0039] According to different vertex features in the brain atlas space, different individualized super vertex division can be realized, and the application range is wide.

[0040] Based on the super vertex connection network, the relationship between the super vertex connection and the data set sample quantization index is analyzed by using the machine learning method, and the effectiveness of the super vertex division is verified.

[0041] Compared with the existing brain atlas, the resolution of the brain division is greatly improved, and the number of sub-regions does not need to be determined, and the brain region division can be automatically performed to construct the super vertex. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0043] Figure 1 It is a flowchart of the super vertex division method of any brain region of T1 magnetic resonance image based on community discovery algorithm proposed by the present application.

[0044] Figure 2 It is a whole brain super vertex division flowchart based on T1 magnetic resonance image.

[0045] Figure 3 It is a personalized brain region sub-region division flowchart based on the whole brain super vertex space.

[0046] Figure 4 It is an ADOS index prediction model based on the morphological features of the whole brain super vertex. DETAILED DESCRIPTION

[0047] In order to make the purpose, features, advantages of the present application more obvious and easy to understand, the embodiments of the present application will be described clearly and completely in the following combined with the specific embodiments involved in the drawings. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor, such as changing the use without changing the embodiments related to the basic principles of the claims, belong to the protection scope of the present application.

[0048] The present application proposes a brain cortex super vertex division method based on T1 magnetic resonance image. It should be noted that the embodiments disclosed herein are only representative, and the present application is not limited to the specific methods described herein, but can also have other embodiments or combinations of other embodiments.

[0049] The technical flow of any brain region super vertex division is as follows Figure 1As shown, the present disclosure takes the DK (Desikan / Killiany) brain atlas as an example of the division of the cerebral cortex super-vertices based on the basic brain atlas, and the specific implementation is described as follows.

[0050] The present embodiment uses the T1 magnetic resonance image data collected by the ABIDE (Autism Brain Imaging Data Exchange) database, which includes 1112 sample data.

[0051] The T1 magnetic resonance image is automatically segmented and preprocessed by the FreeSurfer software, and the processing procedure is as follows: (1) linearly registering the individual original image to the Talairach space; (2) non-uniform field correction; (3) stripping the non-brain tissue structures such as skull and scalp; (4) segmenting the brain into tissues such as gray matter, white matter and cerebrospinal fluid; (5) forming the inner surface and outer surface of the cortical image between the outer surface of the white matter and the outer surface of the gray matter; (6) using the vertex as the basic unit for analyzing the cortical image, and completing the morphological feature extraction of the vertices on the inner surface and the outer surface of the cortical image. Five morphological features are selected for each vertex, including cortical thickness, gray matter volume, surface area, average curvature and sulcal depth.

[0052] For the selection condition of the brain atlas, the present embodiment selects the 68 brain regions divided by the DK brain atlas as the basis for the division of the cerebral cortex super-vertices, divides the super-vertices on each brain region, and finally combines the division results of each brain region to form the super-vertices of the entire cerebral cortex.

[0053] For the calculation method of the vertex connection, the vertex connection in the present embodiment is calculated under the spatial constraint of the brain region i, and the city block distance is used to calculate the vertex connection between all vertices in the brain region i. The vertex connection can be represented as an N*N matrix, where N is the number of vertices in a brain region i, and the vertex connection matrix of each brain region in the brain atlas is calculated in the same way.

[0054] The average vertex connection matrix of the brain region i is obtained by adding the vertex connection matrices of all samples in the brain region i in the data set and then dividing by the number of samples. Under the DK brain atlas partition, the average vertex connection matrices of all 68 brain regions are obtained as the basic matrix for implementing the division of the cerebral cortex super-vertices.

[0055] For the division method of the average vertex connection matrix, the present embodiment uses the Multilevel algorithm to automatically cluster the average vertex connection matrix, and the vertices in the brain region i are divided into different super-vertices. The information of the super-vertices is the aggregation of the information of all vertices in the same category.

[0056] The Multilevel algorithm is applied to the average vertex connection matrix of all brain regions, each brain region is automatically divided into different numbers of super vertices, and finally the super vertices of the 68 brain regions are mapped to the standard brain region space to complete the division of the group brain cortex super vertices, and the group whole brain super vertex division process is as shown in Figure 2

[0057] The technical process of individualized brain region subregion division is as shown in Figure 3 First, the individual magnetic resonance image is obtained, the individual brain space vertex morphological features are automatically segmented and extracted, and the whole brain vertex connection network is constructed based on the morphological features. Secondly, the divided group super vertices are mapped to the individual brain space, and the individual brain space vertices take the group super vertex label as the vertex initial label. On the basis of the individual vertex connection network, the label propagation algorithm is used to update the existing super vertex label until the algorithm converges, and the vertices of the same class label define a subregion, thereby completing the division of the individualized brain region subregion.

[0058] The whole brain super vertex connection network is constructed. Each super vertex is an information aggregation of similar vertices, and the morphological features of the super vertex are the average values of the morphological features of all vertices belonging to the super vertex classification. Based on the group super vertex division result, the individual super vertex morphological features are extracted by mapping to the individual brain space, the individual brain space super vertex connection is calculated, and the group whole brain super vertex connection network is constructed. Based on the individualized super vertex division, the morphological features of the individual super vertex are directly extracted to construct the individualized whole brain super vertex connection network.

[0059] The divided brain cortex super vertex is verified. ADOS (Autism Diagnostic Observation Schedule) is an important indicator for evaluating brain diseases, and samples with ADOS Total, ADOS Social, ADOS Communication, and ADOS Behavior indicators in the data set are selected as verification samples. Based on the group super vertex connection network, the super vertex connection of the sample is extracted as a feature vector to verify the prediction effect of the super vertex connection on the ADOS value.

[0060] The leave-one-out method is used for regression prediction, and the specific operation is as follows: each time, one sample is used as test data, and all other samples are used as training data. The ADOS value of the test sample is predicted by the support vector regression algorithm. The same operation is performed on all samples to obtain the predicted ADOS value of each sample. The predicted ADOS value of the sample is compared with the real ADOS value for Pearson correlation analysis.

[0061] ​The embodiments of the present application are described in detail above with reference to the accompanying drawings, but the present application is not limited to the described embodiments. Various changes, modifications, replacements, and variations of the embodiments including components can be made by those skilled in the art without departing from the principles and spirit of the present application, and still fall within the scope of the present application.

Claims

1. A method for segmenting cerebral cortex supervertices based on T1 magnetic resonance imaging, characterized in that: The steps include: S1. Acquire T1 magnetic resonance imaging data to establish a data set and perform preprocessing. S2. Calculate the vertex connections in the brain regions according to the vertices of the brain regions divided by the brain map to obtain a vertex connection matrix; The calculation method to obtain the vertex connection matrix is ​​as follows: Vertex: V = {feature 1, feature 2, feature 3, ..., feature m}; Vertex set: {V1, V2, V3, …, Vn}; Vertex connection: dij = Distance(Vi,Vj); Vertex connectivity matrix: The vertex feature is the morphological feature of the vertex after preprocessing, the vertex set is the initial vertex in a brain region in the brain atlas space, and Distance is the distance calculation method of the vertex; S3. Obtain an average vertex connection matrix from the vertex connection matrix of the same brain region in the entire dataset, cluster the average vertex connection matrix to obtain group-level supervertices, and map the group-level supervertices to the corresponding original brain region, thereby obtaining the group original brain region supervertices space; S4. According to step S3, the group-level hypervertices of each brain region of the brain atlas are obtained and mapped to the corresponding original brain regions, thereby obtaining the group-level hypervertices space of the whole brain; S5, completing individualized super-vertex division according to the super-vertex labels of the divided groups; The method for dividing individualized supervertices is as follows: mapping group-level supervertices to the space of the brain map as initial labels of vertices, calculating the vertex connections in the space of the brain map to obtain a whole-brain vertex connection network, and implementing a label propagation algorithm on the basis of the whole-brain vertex connection network to complete the division of individualized supervertices; S6. Based on the vertex features of the brain map space, different group feature super vertex connection networks and individual feature super vertex connection networks are constructed based on the division results of group-level super vertices and individual super vertexes; S7. Verify the effectiveness of super-vertex partitioning based on machine learning algorithm.

2. The method for segmenting cerebral cortex super-vertex based on T1 magnetic resonance imaging according to claim 1, characterized in that: In S1, the data set preprocessing method is to segment the T1 magnetic resonance image data using an automatic segmentation tool, and then obtain the morphological features of the vertices on the T1 magnetic resonance image data.

3. The method for cerebral cortex super-vertex segmentation based on T1 magnetic resonance imaging according to claim 1, characterized in that: The brain regions divided by the brain atlas serve as spatial constraints for vertex connections.

4. The method for segmenting cerebral cortex super-vertex based on T1 magnetic resonance imaging according to claim 1, characterized in that: The vertex connectivity is the distance between all vertices in the same brain region.

5. The method for segmenting cerebral cortex super-vertex based on T1 magnetic resonance imaging according to claim 1, characterized in that: The method for obtaining the average vertex connection matrix is ​​as follows: the average vertex connection matrix of the same brain region of all samples in the data set is added and then divided by the number of samples.

6. The method for cerebral cortex super-vertex segmentation based on T1 magnetic resonance imaging according to claim 1, characterized in that: In S3, the average vertex connection matrix is ​​clustered using a community discovery algorithm.

7. The method for segmenting cerebral cortex super-vertex based on T1 magnetic resonance imaging according to claim 1, characterized in that: The vertex feature is any one of the cerebral cortex structure, morphology, function and multimodal features.

Images