An alzheimer disease feature extraction and classification method based on complex brain network

Abstract

The application discloses a kind of Alzheimer's disease feature extraction and classification method based on complex brain network, comprising the following steps: step S1) data pre-processing and constructing brain connectivity matrix;Step S2) complex brain network measurement;Step S3) obtain key brain area based on local measurement index;Step S4) NBS analysis is carried out based on binary brain network, and important brain area is obtained and global measurement index is calculated;Step S5) extract feature vector and carry out classification.This scheme is first according to functional magnetic resonance imaging data to carry out fine brain partition and construct brain connectivity matrix, and key brain area is extracted by local measurement index, and then global measurement index and local measurement index are calculated further for key brain area, then feature extraction is carried out and classification is carried out, to provide fast and accurate auxiliary diagnosis basis for the disease diagnosis of Alzheimer's disease.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of magnetic resonance data analysis, in particular to an Alzheimer's disease feature extraction and classification method based on a complex brain network. BACKGROUND

[0002] Alzheimer's disease (AD) is a kind of insidious and irreversible neurodegenerative disease. Brain imaging can reveal the brain changes of neurodegenerative diseases to a certain extent, including structural magnetic resonance imaging, functional magnetic resonance imaging, diffusion tensor imaging, positron emission computed tomography and the like. The research method based on neuroimaging provides a research means for exploring the change rule of key brain regions and key path abnormalities in the AD brain network, and deeply understanding the physiological mechanism of the disease and early diagnosis of the disease. With the development of imaging genetics, using imaging and biomarkers as intermediate phenotypes has higher sensitivity in identifying AD gene-related functional connections. In the prior art, most of the researches do not use fine brain partition brain imaging data to study brain region structure and functional brain imaging, brain network topology attributes, and do not directly form a functional magnetic resonance connection matrix according to the whole brain region data, and do not calculate the global and local metric indexes of the brain network, and then perform feature extraction and classification / identification. SUMMARY

[0003] The application provides an Alzheimer's disease feature extraction and classification method based on a complex brain network. First, fine brain partition is performed according to functional magnetic resonance imaging data, and a brain connectivity matrix is constructed. Key brain regions are extracted through local metric indexes. Global and local metric indexes of the key brain regions are further calculated, then feature extraction and classification are performed, and a quick and accurate auxiliary diagnosis basis is provided for disease diagnosis of Alzheimer's disease.

[0004] The application calculates the global and local metric indexes of the complex brain network of patients with different degrees of cognitive impairment and healthy people, and proposes a method for screening out key brain regions with significant differences according to the local metric indexes.

[0005] The application extracts 30 characteristic values from the data of 360 brain regions and the data of 36 key brain regions, uses the 30 characteristic values as prior conditions of machine learning for model training, finally obtains a higher level of binary classification, and the accuracy of feature extraction and classification using the local measurement index of 36 key brain regions is > the accuracy of feature extraction and classification using the local measurement index of 360 brain regions > the accuracy of feature extraction and classification using the global measurement index and the local measurement index of 360 brain regions > the accuracy of feature extraction and classification using the global measurement index of 360 brain regions, and the classification result of secondary feature extraction is obviously better than other related researches.

[0006] In order to achieve the above purpose, the application adopts the following technical solutions:

[0007] A feature extraction and classification method for Alzheimer's disease based on a complex brain network comprises the following steps: step S1) data preprocessing and construction of a brain connectivity matrix; step S2) complex brain network measurement; step S3) obtaining key brain regions based on a local measurement index; step S4) NBS analysis based on a binary brain network, obtaining important brain regions and calculating a global measurement index; and step S5) extracting a feature vector and performing classification. The application calculates the global measurement index and the local measurement index of the complex brain network of patients with different degrees of cognitive impairment and healthy people, proposes a method for screening key brain regions with significant differences according to the local measurement index, then extracts 30 characteristic values from the data of 360 brain regions and the data of 36 key brain regions, uses the 30 characteristic values as prior conditions of machine learning for model training, finally obtains a higher level of binary classification, and the accuracy of feature extraction and classification using the local measurement index of 36 key brain regions is > the accuracy of feature extraction and classification using the local measurement index of 360 brain regions > the accuracy of feature extraction and classification using the global measurement index and the local measurement index of 360 brain regions > the accuracy of feature extraction and classification using the global measurement index of 360 brain regions, and the classification result of secondary feature extraction is obviously better than other related researches.

[0008] As preferred, the data preprocessing in step S1 specifically comprises: based on a multi-modal brain parcellation method, performing functional region segmentation on the brain of the Alzheimer's disease patient, and subdividing the whole brain into 360 brain regions for fine brain parcellation. Step S1 mainly performs functional region segmentation on the brain of the Alzheimer's disease patient based on the multi-modal brain parcellation method of the University of Washington, and subdivides the whole brain into 360 brain regions for fine brain parcellation, and the specific operation is as follows: through the Alzheimer's Disease Neuroimaging Initiative 2 (ADNI2) data set, magnetic field distribution information, brain cortex thickness information, brain cortex myelin distribution information, task-state functional magnetic resonance imaging data and resting-state functional magnetic resonance imaging data are obtained; a Joint Human Connectome Project Multi-modal Parcellation (J-HCPMMP) method is used to combine T1 weight structural magnetic resonance imaging, resting-state functional magnetic resonance imaging and magnetic field distribution information, and automatically register the DICOM data produced by the device to the Connectivity Informatics Technology Initiative (CIFTI) grayscale space, and then the cortex vertices and subcutaneous tissues in this space are partitioned through the processing framework of HCP, so that the brain of each subject is divided into 180 sub-brain regions on the left and right sides, and a total of 360 brain regions, and this method makes up for the problem that the applicability of magnetic resonance imaging data is generally not strong due to the excessively high HCP processing protocol.

[0009] As preferred, the constructing the brain connectivity matrix in step S1 specifically comprises the following steps: step A1) performing correlation analysis on the functional magnetic resonance data between the 360 brain regions to form a 360*360 size complex brain network adjacency matrix for measuring the connection state between each functional brain region of the brain; and step A2) performing threshold processing on the complex brain network adjacency matrix to remove noise and interference information to obtain the brain connectivity matrix. The correlation analysis is performed on the functional magnetic resonance data between the 360 brain regions to form a 360*360 size complex brain network adjacency matrix for measuring the connection state between each functional brain region of the brain, and the matrix is subjected to sparse processing (threshold processing) to remove noise and interference information to obtain a sparse brain connectivity matrix.

[0010] As preferred, the complex brain network metric in step S2 specifically comprises the following steps: step B1), constructing a corresponding weighted brain network based on the brain connectivity matrix; step B2), setting the effective connections in the weighted brain network as 1 and other connections as 0, while setting the connection weights on the diagonal line as 0, to form a binary brain network; step B3), calculating the global metric index and the local metric index of the binary brain network. In the functional complex brain network, only the undirected brain region functional connection is considered, the absolute value of all correlation coefficients in the network is taken, and the negative correlation is removed, and finally two complex brain networks are constructed, including a weighted brain network and a binary brain network. The weighted brain network corresponds to the sparse brain connectivity matrix calculated in step S1; the binary brain network sets the effective connections in the weighted brain network as 1 and other connections as 0, while setting the connection weights on the diagonal line as 0.

[0011] As preferred, the global metric index comprises a complex brain network global efficiency, assortativity coefficient, small-world characteristic index, characteristic path length, hierarchy, and synchrony coefficient.

[0012] As preferred, the local metric index comprises a complex brain network node degree, local efficiency, betweenness centrality, clustering coefficient, eigenvector centrality, and shortest path.

[0013] As preferred, the key brain regions are obtained based on the local metric index in step S3, specifically comprising: performing F test on the six local metric indexes respectively, selecting the brain regions satisfying P-value < 0.01 as the key brain regions, obtaining 36 key brain regions, and further obtaining 36*36 connection matrices corresponding to the 36 key brain regions. F test is performed on the six local metric indexes of the four sets of inter-group data respectively, and the brain regions with P-value < 0.01 are selected, i.e. the brain regions with significant differences among the four groups are selected, a total of 36 brain regions are selected from 360 brain regions as key brain regions. The four sets of inter-group data are corresponding data of Alzheimer's disease patients, early mild cognitive impairment patients, late mild cognitive impairment patients, and healthy people.

[0014] As preferred, the specific process of step S4 comprises the following steps: step C1), performing NBS analysis on the binary connection matrix corresponding to the binary brain network to obtain the connectivity patterns of the Alzheimer's disease AD patients, early mild cognitive impairment EMCI patients and late mild cognitive impairment LMCI patients, selecting the brain regions satisfying P-value < 0.001 as important brain regions, obtaining 24 important brain regions, and further obtaining the 24*24 connection matrix corresponding to the 24 important brain regions; and step C2), calculating the global metric index of the 24*24 connection matrix, and comparing the global metric indexes of the 360*360 connection matrix and the 24*24 connection matrix. The binary connection matrix corresponding to the binary brain network of all subjects is used to perform NBS analysis on 10,000 permutations to find the connectivity patterns of the Alzheimer's disease AD patients, early mild cognitive impairment EMCI patients and late mild cognitive impairment LMCI patients, select the brain regions satisfying P-value < 0.001 as important brain regions, obtain 24 important brain regions, and obtain the 24*24 connection matrix; the global metric index of the 24*24 connection matrix is calculated, and the global metric indexes of the 360*360 connection matrix and the 24*24 connection matrix of the four groups of data are compared.

[0015] As preferred, the specific process of step S5 comprises the following steps: step D1), calculating the global metric index and the local metric index of the 36 key brain regions again, and extracting the feature vector through the Finser feature extraction algorithm; and step D2), adopting a machine learning algorithm to perform two-class modeling on the feature vector extracted in step D1). The global metric index and the local metric index of the key brain regions screened in step S3 are calculated again, the feature vector is extracted through the Finser feature extraction algorithm, the feature data dimension is reduced, and the model training complexity is reduced; a machine learning algorithm is adopted to perform two-class modeling on the extracted feature vector, and the feature vector is classified by using the support vector machine classification method, that is, the feature vector is identified to identify which degree of cognitive impairment patient or healthy person the feature vector corresponds to.

[0016] Therefore, the application has the advantages that global and local metric indexes of complex brain networks of patients with different degrees of cognitive impairment and healthy people are calculated, a method of screening key brain regions with significant differences according to the local metric indexes is proposed, then data of 360 brain regions and 36 key brain regions are subjected to feature extraction, 30 characteristic values are extracted, the 30 characteristic values are used as prior conditions of machine learning for model training, and finally a higher level of binary classification is obtained, and the accuracy of classification using the local metric indexes of 36 key brain regions for feature extraction is higher than that of using the local metric indexes of 360 brain regions for feature extraction, which is higher than that of using the global metric indexes and the local metric indexes of 360 brain regions for feature extraction, which is higher than that of using the global metric indexes of 360 brain regions for feature extraction, and the classification result of secondary feature extraction is obviously better than that of other related researches. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a flowchart of an embodiment of the application.

[0018] Figure 2 is a comparison chart of 360*360 connection matrix and 24*24 connection matrix global metric indexes in an embodiment of the application. DETAILED DESCRIPTION

[0019] The application will be further described below in combination with the drawings and specific embodiments.

[0020] Embodiment one:

[0021] As shown in the drawings, Figure 1 a feature extraction and classification method for Alzheimer's disease based on complex brain networks comprises the following steps:

[0022] Step 1: data preprocessing and construction of brain connectivity matrix, the specific operation is as follows:

[0023] Step 1-1: The specific parameters of structural magnetic resonance imaging are sagittal T1-weighted three-dimensional fast gradient echo imaging (T1W-3D-MPRAGE) with eight-channel sensitivity encoding (SENSE) parallel imaging algorithm, external magnetic field strength of 3 Tesla, imaging resolution of 256x256 1.0 mm, slice number of 170, slice thickness of 1.2 mm, echo time TR of 6.78 ms, and repetition time TE of 3.14 ms. The specific parameters of functional magnetic resonance imaging are 3 Tesla external magnetic field strength under the resting state of the subject, imaging resolution of 64x64 3.3125 mm, slice number of 48, slice thickness of 3.313 mm, 140 time sequences, a total of 6,720 slices, echo time TR of 3,000 ms, and repetition time TE of 30 ms. The J-HCPMMP data preprocessing method is used to automatically preprocess the T1-weighted structural data and functional magnetic resonance data, and map them to the CIFTI spatial coordinate system, then integrate the data, and divide the brain of each subject into 180 brain regions on the left and right respectively using the HCPMMP partition method, and divide the 360 brain regions into 22 groups of regions as shown in Table 1.

[0024] Table 1 22 groups of brain regions divided in this embodiment

[0025]

[0026] Step 1-2: Correlation analysis is performed on the functional magnetic resonance data between the 360 brain regions to form a 360*360 size complex brain network adjacency matrix for measuring the connection state between each functional brain region of the brain. The matrix is processed to be sparse (threshold processing) to remove noise and interference information, and a sparse brain connectivity matrix is obtained.

[0027] Step 2: Complex brain network measurement, the specific operation is as follows:

[0028] Step 2-1: The weighted brain network corresponds to the sparse brain connectivity matrix calculated in step 1; the binary brain network sets the effective connection in the weighted brain network to 1 and other connections to 0, and sets the connection weight on the diagonal line to 0;

[0029] Step 2-2: According to the connection matrix corresponding to the binary brain network calculated in step 1, the global measurement index of each brain region is calculated using Gentna, including complex brain network global efficiency, assortativity coefficient, small-world characteristic index, characteristic path length, hierarchy, and synchronicity coefficient.

[0030] Step 2-3: According to the binary brain network corresponding to the connection matrix calculated in step 1, the local metric indicators of each brain area are calculated using Gentna, including the degree of each node of the complex brain network, local efficiency, betweenness centrality, clustering coefficient, eigenvector centrality, and shortest path.

[0031] Step 3: Obtain the key brain area based on the local metric indicators, and the specific operation is as follows:

[0032] Step 3-1: For the four sets of intergroup data, statistical analysis is performed according to the local metric indicators, and 36 key brain areas corresponding to P-value < 0.01 under six local metric indicators are obtained. The four sets of intergroup data are Alzheimer's disease patient data, early mild cognitive impairment patient data, late mild cognitive impairment patient data, and healthy person data.

[0033] Step 4: Perform NBS analysis based on the binary brain network to obtain important brain areas and calculate global metric indicators, and the specific operation is as follows:

[0034] Step 4-1: Perform NBS analysis on 10000 permutations using the connection matrix corresponding to the binary brain network of all subjects to discover the connectivity patterns of AD, EMCI, and LMCI patients, and obtain 24 important brain areas through P-value < 0.001, and obtain a 24*24 connection matrix;

[0035] Step 4-2: Calculate the global metric indicators of the 24*24 connection matrix, and compare the global metric indicators of the 360*360 connection matrix and the 24*24 connection matrix of the four sets of intergroup data, as shown in Figure 2

[0036] Step 5: Extract feature vectors and perform classification, and the specific operation is as follows:

[0037] Step 5-1: Calculate the global metric indicators and local metric indicators of the 36 key brain areas selected in step 3 again, and extract feature vectors from the local metric indicators of 360 brain areas (Local 360areas), the local metric indicators of 36 key brain areas (Local 36areas), the global metric indicators of 360 brain areas (Globe 360areas), and the global metric indicators + local metric indicators of 360 brain areas (Local + Globe 360areas) through Finser feature extraction algorithm; Step 5-2: Use machine learning algorithm to perform two-class modeling on the extracted feature vectors, and this embodiment uses support vector machine method to classify the feature vectors, i.e. to identify the feature vectors, and identify which degree of cognitive impairment patient or healthy person the feature vector corresponds to. ​

[0038] As shown in Table 2, the accuracy of feature extraction classification using the local metric index of 36 key brain regions > the accuracy of feature extraction classification using the local metric index of 360 brain regions > the accuracy of feature extraction classification using the global metric index + the local metric index of 360 brain regions > the accuracy of feature extraction classification using the global metric index of 360 brain regions.

[0039] Table 2 Accuracy of cognitive impairment classification model of the present embodiment

[0040]

Claims

1. A method for feature extraction and classification of Alzheimer's disease based on complex brain networks, characterized in that, Includes the following steps: Step S1: Preprocess the specific parameters of structural magnetic resonance imaging and functional magnetic resonance imaging and construct the brain connectivity matrix; Step S2: Complex brain network measurement, constructing weighted brain networks and binary brain networks, and using the Gentna tool for global and local metrics in each brain region; Step S3: Identify key brain regions based on locality metrics; The locality metrics include the degree of each node in the complex brain network, local efficiency, betweenness centrality, clustering coefficient, eigenvector centrality, and shortest path. Step S3 specifically includes: performing F-tests on the six locality metrics respectively, selecting brain regions that satisfy P-value < 0.01 as key brain regions, obtaining 36 key brain regions, and then obtaining a 36*36 connectivity matrix corresponding to the 36 key brain regions; Step S4: Perform NBS analysis based on the binary brain network to obtain important brain regions and calculate global metrics; Step S5: Extract feature vectors and classify them. Calculate global and local metrics again for the key brain regions selected in Step S3. Extract feature vectors from the local metrics of the key brain regions. Select local metrics and use machine learning algorithms to perform binary classification modeling on the extracted feature vectors.

2. The method for feature extraction and classification of Alzheimer's disease based on complex brain networks according to claim 1, characterized in that, The preprocessing described in step S1 specifically includes: performing functional region segmentation of the brain of Alzheimer's patients based on a multimodal brain partitioning method, dividing the whole brain into 360 brain regions for fine brain partitioning.

3. The method for feature extraction and classification of Alzheimer's disease based on complex brain networks according to claim 2, characterized in that, The construction of the brain connectivity matrix in step S1 specifically includes the following steps: Step A1: Perform correlation analysis on the functional magnetic resonance imaging data between 360 brain regions to form a complex brain network adjacency matrix of size 360*360, which is used to measure the connectivity between various functional brain regions. Step A2: Perform thresholding on the adjacency matrix of the complex brain network to remove noise and interference information, and obtain the brain connectivity matrix.

4. The method for feature extraction and classification of Alzheimer's disease based on complex brain networks according to claim 1, characterized in that, The measurement of complex brain networks in step S2 specifically includes the following steps: Step B1: Construct the corresponding weighted brain network based on the brain connectivity matrix; Step B2: Set the effective connections in the weighted brain network to 1 and the other connections to 0. At the same time, set the weights of the connections located on the diagonal to 0 to form a binary brain network. Step B3: Calculate the global and local metrics of the binary brain network.

5. The method for feature extraction and classification of Alzheimer's disease based on complex brain networks according to claim 4, characterized in that, The global metrics include global efficiency of complex brain networks, isomatch coefficient, small-world property index, feature path length, hierarchical structure, and synchronization coefficient.

6. The method for feature extraction and classification of Alzheimer's disease based on complex brain networks according to claim 1, characterized in that, Step S4 includes: performing NBS analysis on the binary connection matrix corresponding to the binary brain network to obtain the connectivity patterns of Alzheimer's disease (AD) patients, early mild cognitive impairment (EMCI) patients, and late mild cognitive impairment (LMCI) patients; selecting brain regions that satisfy P-value < 0.001 as important brain regions to obtain 24 important brain regions; and then obtaining the 24*24 connection matrix corresponding to the 24 important brain regions; calculating the global metric of the 24*24 connection matrix; and comparing the global metric of the 360*360 connection matrix with that of the 24*24 connection matrix.

7. The method for feature extraction and classification of Alzheimer's disease based on complex brain networks according to claim 1, characterized in that, The specific process of step S5 includes the following steps: Step D1: Calculate the global and local metrics for the 36 key brain regions again, and extract feature vectors from the local metrics of the 360 ​​brain regions, the local metrics of the 36 key brain regions, the global metrics of the 360 ​​brain regions, and the global and local metrics of the 360 ​​brain regions respectively using the Fisher score feature extraction algorithm. Step D2: Using machine learning algorithms, perform binary classification modeling on the feature vectors extracted in step D1, and use the support vector machine method to classify the feature vectors, that is, identify the feature vectors and identify the different degrees of cognitive impairment patients or healthy people.

Images