Method, system and computer-readable medium for classifying resting-state functional magnetic resonance data of the brain by integrating multi-scale information

Abstract

The present invention discloses a brain resting-state functional magnetic resonance data classification method, system and computer-readable medium that fuse multi-scale information. The resting-state functional magnetic resonance imaging data of two groups of people in different states are obtained, and feature extraction is performed after noise is eliminated. Brain features at three scales, namely, voxel body activity, local covariance of voxel clusters, and global covariance of voxel networks, are extracted respectively, and the brain is divided into regions. Then, the average values ​​of voxel body activity, local covariance of voxel clusters, and global covariance of voxel networks of each brain region of each tested object in the two groups of people in different states are calculated. At each feature scale, a 1×N-dimensional feature matrix is ​​obtained for each tested object in the two groups of people in different states. The feature matrices at each scale are fused, and feature selection is performed on the fused feature matrix. An SVM classifier is trained using the optimal feature subset obtained by feature selection to classify the tested objects, thereby effectively improving the classification accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of magnetic resonance image processing, and relates to a brain resting-state functional magnetic resonance data classification method, system and computer-readable medium integrating multi-scale information. Background Art

[0002] Magnetic resonance imaging is a non-invasive brain observation technology that reflects brain activity by measuring magnetic resonance signals caused by changes in cerebral blood flow and cerebral blood oxygen caused by neural activity. It is widely used in clinical observation diagnosis and cognitive function research.

[0003] Neurons are the basic structural components of brain function. Comprehensively characterizing the information interaction between spatially neighboring or discrete neurons is a necessary prerequisite for accurately distinguishing brain states. Voxels are the basic information units in magnetic resonance data. A single voxel in magnetic resonance data contains a large amount of neuronal activity information. Brain decoding for magnetic resonance data uses voxels as the basic operation unit. The activity level of a single voxel represents the activity level of a cluster of neurons. The degree of activity covariation between multiple spatially neighboring voxels represents the degree of information interaction between these voxels at a local scale. The degree of activity covariation between a single voxel and other voxels in the brain represents the degree of information interaction at a global scale.

[0004] Existing resting-state functional magnetic resonance imaging data analysis methods often classify brain states based on single-scale brain state data. For example, the invention patent with publication number CN110443798A, entitled A method, device and system for detecting autism based on magnetic resonance images, trains and detects a binary classification model for autism based on extracted feature information. One of the research results presents the classification accuracy of 6 types of feature information based on the same scale. The results show that the use of single-scale information can separate the two types of samples, but the accuracy is very limited. In summary, for the problem of brain state classification, the accuracy of brain state classification is poor due to the limited representation of brain activity state by single-scale measurement parameters.

[0005] Neighboring or discrete neurons in the brain communicate with each other through common discharges and work together to support the generation of complex behaviors. Therefore, with voxels as the basic unit of brain state information, the information interaction between neurons is reflected in the activity of single voxels, the interaction between spatially adjacent voxels, and the interaction between spatially discrete voxels. Characterization parameters of different scales can provide us with relevant information about the brain from different angles. Furthermore, only by quantifying brain information at three scales, namely, single voxel scale, local scale, and global scale, can we fully characterize brain activity, which is the basic condition for accurately distinguishing brain states. Therefore, using multi-scale feature fusion methods to classify brain state data will become a technology with great application value in the fields of medical image processing and brain science. Summary of the invention

[0006] The purpose of the embodiments of the present invention is to provide a method, system and computer-readable medium for classifying resting-state functional magnetic resonance imaging data of the brain that integrates multi-scale information, so as to solve the problem of low accuracy of existing resting-state functional magnetic resonance imaging data analysis methods in classifying brain states based on single-scale brain state data.

[0007] The first technical solution adopted by the embodiment of the present invention is: a method for classifying brain resting state functional magnetic resonance imaging data by integrating multi-scale information, which is performed according to the following steps:

[0008] Step S1, obtaining resting-state functional magnetic resonance imaging data of each subject in two groups of people in different states;

[0009] Step S2, preprocessing the resting-state functional magnetic resonance imaging data of two groups of people in different states to eliminate noise;

[0010] Step S3, extracting voxel body activity, local covariation of voxel clusters, and global covariation of voxel networks from the resting-state functional magnetic resonance imaging data of the two groups of people in different states after preprocessing;

[0011] Step S4, dividing the brain into regions, and then calculating the average values ​​of the three features of voxel body activity, voxel cluster local covariation, and voxel network global covariation for each brain region of each subject in the two groups of people with different states. At each feature scale, each subject in the two groups of people with different states will obtain a 1×N dimensional feature matrix;

[0012] Step S5, fusing the 1×N-dimensional feature matrices of all the subjects in the two groups of people with different states at three different feature scales, namely, voxel body activity, voxel cluster local covariance, and voxel network global covariance, to obtain the feature fusion matrices of the two groups of people with different states;

[0013] Step S6, using the feature fusion matrix of the two groups of people in different states and the labels of the subjects of the two groups of people in different states to train the SVM classifier, and after training and testing, obtain the SVM classifier for classifying the resting-state functional magnetic resonance data of the brain;

[0014] Step S7: using the SVM classifier for classifying resting-state functional magnetic resonance data of the brain to classify the objects to be classified.

[0015] Furthermore, it also includes:

[0016] After obtaining the feature fusion matrices of the two groups of people in different states in step S5, secondary feature selection is performed on the feature fusion matrices of the two groups of people in different states, and then the feature set obtained by the secondary feature selection is used to train the SVM classifier, and the leave-one-out cross-validation method is used. After training and testing, the SVM classifier for classifying brain resting-state functional magnetic resonance data is obtained.

[0017] Furthermore, in step S7, the specific process of using the SVM classifier for classifying the resting-state functional magnetic resonance data of the brain to classify the tested object is as follows:

[0018] Acquiring resting-state functional magnetic resonance imaging data of a subject to be classified;

[0019] Using steps S2 to S5, a feature fusion matrix of the object to be classified is obtained;

[0020] Perform secondary feature selection on the feature fusion matrix of the object to be classified;

[0021] The feature set obtained by the secondary feature selection is input into the SVM classifier for classification of brain resting-state functional magnetic resonance data to classify the objects to be classified.

[0022] Furthermore, a recursive feature elimination method is used to perform secondary feature selection on the feature fusion matrix of two groups of people in different states or objects to be classified, so as to obtain the optimal feature subsets of the two groups of people in different states or objects to be classified;

[0023] After obtaining the optimal feature subsets of the two groups of people in different states, the optimal feature subsets of the two groups of people in different states are used to train the SVM classifier to obtain the SVM classifier for classifying the brain resting state functional magnetic resonance data.

[0024] Furthermore, the specific implementation process of performing secondary feature selection on the feature fusion matrix of two groups of people in different states is as follows:

[0025] (1) The feature fusion matrix of the two groups of people with different statuses was statistically tested using the leave-one-out method to preliminarily screen the features with significant differences between the two groups of people with different statuses and obtain the feature matrix containing the features with significant differences between the two groups of people with different statuses;

[0026] (2) The feature matrix containing significant differences between the two groups of people in different states is normalized by mean variance by column to obtain the normalized feature matrix of the two groups of people in different states. The calculation formula for mean variance normalization is as follows:

[0027]

[0028] Among them, x * is the normalized data, x Yis the original data of each column in the feature matrix of two groups of people with different status after preliminary screening, containing significant differences in characteristics, μ Y According to x Y The calculated expectation, σ Y According to x Y Calculated variance;

[0029] (3) A loss function is added to the existing linear SVM classifier model to construct an SVM classifier, and the normalized feature matrices of two groups of people in different states are used, and a matrix with (the number of tested objects - 1) rows and the number of columns being the feature dimension of the normalized feature matrix is ​​used as a training sample. It is assumed that the label of one group of people in one state is 1 and the label of the other group of people in another state is -1. A corresponding label matrix of (the number of tested objects - 1) × 1 dimensions is formed according to the labels of the two groups of people in different states, and the SVM classifier is trained using the training samples and the (the number of tested objects - 1) × 1 dimension label matrix;

[0030] (4) Using the packaging method, the trained SVM classifier is used to select the optimal feature subset from the normalized feature matrices of two groups of people with different status;

[0031] Then the optimal feature subset is used to continue training the SVM classifier, and the leave-one-out cross-validation method is adopted. After training and testing, the SVM classifier for classifying the resting-state functional magnetic resonance imaging data of the brain is obtained;

[0032] When the recursive feature elimination method is used to perform secondary feature selection on the feature fusion matrix of the object to be classified, after obtaining the normalized feature matrix of the object to be classified through steps (1) and (2), step (4) is directly entered to select the optimal feature subset of the object to be classified.

[0033] Furthermore, the specific implementation process of using the trained SVM classifier to select the optimal feature subset from the normalized feature matrix of two groups of people in different states or objects to be classified is as follows:

[0034] a. Calculate the initial classification accuracy accuracy0 of the current SVM classifier;

[0035] b. Use the leave-one-out cross-validation method to subtract one column from the normalized feature matrix of two groups of people with different status or the objects to be classified, input the feature matrix composed of the remaining features into the SVM classifier, and calculate the current classification accuracy of the SVM classifier i , perform as many calculations as there are columns in the normalized feature matrix of two groups of people in different states or objects to be classified, and find all current classification accuracies that are greater than or equal to the initial classification accuracy 0 i , determine the corresponding feature to be removedi ;

[0036] c. Delete the feature determined in step b from the normalized feature matrix of two groups of people in different states or objects to be classified i To update the corresponding normalized feature matrix; and use the updated normalized feature matrices of the two groups of people in different states to obtain updated training samples to continue training the SVM classifier;

[0037] d. Repeat steps ac until deleting any feature in the normalized feature matrix of the two groups of people in different states or the objects to be classified will not improve the initial classification accuracy accuracy0 of the SVM classifier. At this time, the normalized feature matrix of the two groups of people in different states or the objects to be classified corresponding to the classification accuracy accuracy0 is the effective feature set of the two groups of people in different states or the objects to be classified; the obtained number of objects × the number of effective features N effective The feature matrix of the dimension is the optimal feature subset for two groups of people in different states or the objects to be classified.

[0038] Furthermore, a loss function is added to the existing linear SVM classifier model, and the specific calculation formula of the new objective function with the loss function added is:

[0039]

[0040] Where e represents N subject ×N effective dimensional feature matrix, N subject Indicates the number of samples, N effective represents the effective feature number, e i represents any row of e; ω df is the normal vector of the decision function, b df is the bias term of the decision function, y i Yes i The category to which it belongs is the label, whose value is -1 or +1; i (ω df e i +b df ) represents the slack variable corresponding to each sample, Represents the total classification error, and + represents the positive function, that is:

[0041]

[0042] Furthermore, in step S3, the extraction process of voxel body activity is as follows:

[0043] 1) First, convert the time series of each voxel into the frequency domain using the following formula:

[0044]

[0045] Among them, the plural a e +b e i represents a point e after the time series is transformed into the frequency domain, F(t) represents the signal expression corresponding to point e, and e starts from 1; S N Represents the number of sampling points when converting the time series from the time domain to the frequency domain, F S Indicates the sampling frequency, atan2(b e ,a e ) represents the phase of point e in the frequency domain;

[0046] 2) The information value of each voxel, i.e., the voxel body activity, is calculated by the following formula:

[0047]

[0048] Among them, ρ voxel Represents the information value of each voxel, T voxel Indicates the period of the time series corresponding to each voxel;

[0049] The local covariance of voxel clusters is calculated as follows:

[0050]

[0051] Where V represents the mutual information value of a given voxel, i.e., the local covariance of the voxel cluster, ranging from 0 to 1; is the rank sum of the time points of the ith time series, yes The average value, C is the number of time series in the tested voxel cluster, that is, the number of voxels in the tested voxel cluster, and m is the number of levels, that is, the total number of time points in the time series;

[0052] The calculation process of voxel network global covariance is as follows:

[0053] (a) Given any voxel, perform an exponential correlation analysis on its time series with the time series of other voxels in the brain. The calculation process is as follows:

[0054]

[0055] Among them, r z represents the connection strength between the time series x of a given voxel and the time series y′ of other voxels, x i Indicates the value corresponding to each index position of the time series x, y′ i represents the value corresponding to each index position of the time series y′, that is, x=[x1,x2,…,x i ,…,x nTS], y′=[y′1,y′2,…,y′ i ,…,y′ nTS ];n TS It represents the dimension of the time series of a given voxel, that is, the length of the time series. The time series of each voxel have the same dimension. represents the average value of the corresponding time series x,

[0056] (b) Calculate the connection strength between any voxel and other voxels, that is, the r between any voxel and other voxels z After the value is reached, the positive correlation r that meets the threshold requirement will be z The connection strength of a given voxel can be obtained by adding the values, and the global covariance of the corresponding voxel network can be obtained;

[0057] The specific implementation process of step S4 is as follows:

[0058] The brain is divided into regions using a predefined brain map / brain template. All voxels in each brain region will calculate three characteristic scales: voxel body activity, local covariance of voxel clusters, and global covariance of voxel networks. After the calculation is completed, the mean function of Matlab is called for each characteristic scale to return the average value of all voxels in the brain region. At each characteristic scale, each subject in the two different states will obtain a 1×N characteristic matrix.

[0059] The second technical solution adopted in the embodiment of the present invention is: a brain resting state functional magnetic resonance data classification system integrating multi-scale information, comprising:

[0060] a memory for storing instructions executable by a processor;

[0061] A processor is used to execute the instructions to implement the brain resting state functional magnetic resonance imaging data classification method integrating multi-scale information as described above.

[0062] The third technical solution adopted by the embodiment of the present invention is: a computer-readable medium storing computer program code, which, when executed by a processor, implements the brain resting-state functional magnetic resonance data classification method that integrates multi-scale information as described above.

[0063] The beneficial effects of the embodiments of the present invention are: integrating voxel ontology, voxel cluster covariance and voxel network, extracting and integrating brain state data at different scales as a classification feature set, integrating valuable feature information at each scale, and effectively improving classification accuracy; using statistical test + packaging method for double feature screening, and inputting the obtained optimal feature subset, that is, the optimal feature subset obtained by secondary feature selection, into an improved support vector machine to realize classification of different brain states, effectively solving the problem of low accuracy of existing resting-state functional magnetic resonance imaging data analysis methods in classifying brain states based on single-scale brain state data. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0065] Figure 1 Schematic diagram of the structure of a brain resting-state functional magnetic resonance data classification system integrating multi-scale information in an embodiment of the present invention.

[0066] Figure 2 It is a flowchart of secondary feature selection of a method for classifying resting-state functional magnetic resonance data of the brain that integrates multi-scale information in an embodiment of the present invention.

[0067] Figure 3 It is a comparison diagram of ROC curves between the brain resting state functional magnetic resonance data classification method integrating multi-scale information in an embodiment of the present invention and other single-scale methods.

[0068] Figure 4 This is a brain region distribution map resulting from a method for classifying resting-state functional magnetic resonance data of the brain that integrates multi-scale information in an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] The adjacent or discrete neurons in the brain communicate with each other through common discharges and work together to support the generation of complex behaviors. Therefore, with voxels as the basic brain state information carrier unit, the information interaction between neurons is reflected in the activity of single voxels, the interaction between spatially adjacent voxels, and the interaction between spatially discrete voxels. The embodiment of the present invention proposes a method to characterize the above three scales of brain collaboration information, namely: voxel ontology activity, voxel cluster local covariance, and voxel network global covariance. Among them, voxel ontology activity represents the activity level of neuronal collections, which is directly related to the intensity of spontaneous neural activity in a static state and can reflect the rate of local glucose metabolism; voxel cluster local covariance represents the covariation activity of neuronal clusters contained in voxel clusters, which can be used to characterize the regional homogeneity changes of brains of different populations; voxel network global covariance represents the covariation activity between voxel clusters and other voxel clusters in the brain, which can be used to detect node characteristics of brain neural networks.

[0071] The embodiment of the present invention proposes a brain resting state functional magnetic resonance imaging data classification method based on the fusion of multi-scale information, such as Figure 1 As shown, follow these steps:

[0072] Step S1, data acquisition: obtaining resting-state functional magnetic resonance imaging data of each subject in two groups of people in different states through a nuclear magnetic resonance scanner;

[0073] Specifically, when obtaining resting-state functional magnetic resonance imaging data of each subject in two groups of people in different states through a nuclear magnetic resonance scanner, the acquisition time is set to M seconds, and the pulse sequence repetition time is TR, M is usually 360-480s. After the acquisition time reaches 5 minutes, the subsequent image analysis can have higher stability. When TR=2s, M\TR time points will be obtained. Considering that the subsequent preprocessing will remove some time points (the embodiment of the present invention is set to remove the first X time points), the remaining image data must be more than 5 minutes, and each parameter must satisfy M-TR*X>300.

[0074] Step S2, data preprocessing: preprocess the resting state functional magnetic resonance imaging data of the two groups of people in different states to solve the problem that the original data contains noise due to the influence of factors such as equipment and environment during the acquisition of resting state functional magnetic resonance imaging data, and improve the accuracy of subsequent classification. The steps of preprocessing the resting state functional magnetic resonance imaging data are as follows. After the following preprocessing, the source of false positive errors can be reduced without causing too many false negative errors:

[0075] (1) Format conversion: DICOM format is converted to NIFTI format;

[0076] (2) Removal of some time points: Remove the data of the first X time points to achieve magnetization equilibrium. When TR = 2s, X must satisfy TR*X>16s;

[0077] (3) Time layer correction to eliminate the acquisition delay between layers;

[0078] (4) Head movement correction: During the scanning process, the subject may have slight head movements, which may affect the subsequent data analysis. An algorithm is needed to fit the involuntary slight shaking of the subject's head, and then correct it to ensure that there is no significant difference in head movement between different groups of subjects.

[0079] The basic idea of ​​the algorithm for fitting the involuntary slight shaking of the head of the subject is to minimize the mismatch between the reference image (usually the first image in the time series) and the subsequent sequence images by iteratively calculating the translation and rotation parameters, so as to achieve the registration of all sequence images. The transformation process of the algorithm for fitting the involuntary slight shaking of the head of the subject is:

[0080]

[0081] Among them, M(F) represents the matching degree between the reference image and the resource image. When its value is the smallest, the matching degree is the largest. G -1 represents the inverse matrix of the reference image matrix, O TR is the desired optimal transformation space, M F The matrix represents the resource image, i.e. the image to be corrected, and F is the resource image. The head movement is decomposed into two rigid body transformations, translation and rotation, using vector Indicates the size and angle of head movement in each direction, n T1 ,n T2 ,n T3 The corresponding translation of the head of the object under test on the x, y, and z axes, n R4 ,n R5 ,n R6 The corresponding values ​​represent the rotation of the head of the object under test on the x, y, and z axes respectively. The specific calculation formula for the transformation is as follows:

[0082]

[0083] in, is the translation (n T1 ,n T2 ,n T3 ) is transformed, and its calculation formula is:

[0084]

[0085] They are rotation (n R4 ,n R5 ,n R6 ), the calculation formulas are as follows:

[0086]

[0087]

[0088]

[0089] After establishing this transformation model, the registration can be transformed into a vector The solution is generally the Gauss-Newton iterative algorithm, and the registration is performed based on the optimal rigid body transformation parameters.

[0090] (5) Remove covariates: regress the global signal, white matter signal, and cerebrospinal fluid signal. The specific method is to establish a multivariate linear regression equation with the global signal, white matter signal, and cerebrospinal fluid signal as independent variables to obtain the partial regression coefficients of the global signal, white matter signal, and cerebrospinal fluid signal. Then, calculate the partial regression sum of squares of the global signal, white matter signal, and cerebrospinal fluid signal based on the partial regression coefficients of the global signal, white matter signal, and cerebrospinal fluid signal. Finally, remove the independent variable with a small partial regression sum of squares based on the size of the partial regression sum of squares of the global signal, white matter signal, and cerebrospinal fluid signal. This can ensure that the removed covariates are not correlated with the dependent variable or have a very small correlation, thereby reducing the effects of breathing and heart. The established regression equation is as follows:

[0091] Y=m0+m global x global +m WM x WM +m CSF x CSF ;

[0092] Among them, Y represents the dependent variable mentioned above, namely the rs-MRI image, m0 is a constant term, also known as the intercept; m global is the global signal partial regression coefficient, m WM is the partial regression coefficient of white matter signal, m CSF is the partial regression coefficient of the cerebrospinal fluid signal. The partial regression coefficient indicates that when other independent variables remain unchanged, the global signal x global , white matter signal x WM , cerebrospinal fluid signal x CSF The average change in Y when any one of the two increases or decreases by one unit, x i is the global signal x global, white matter signal x WM , cerebrospinal fluid signal x CSF Any one of .

[0093] Next, calculate the global signal x global , white matter signal x WM , cerebrospinal fluid signal x CSF The partial regression sum of squares is:

[0094] Take the global signal x global For example, the calculation formula is as follows:

[0095]

[0096]

[0097]

[0098] Among them, SS regression (x_global) is the global signal x global The partial regression sum of squares, SS regression (x_WM) is the white matter signal x WM The partial regression sum of squares, SS regression (x_CSF) is the cerebrospinal fluid signal x CSF The partial regression sum of squares; k is the global signal x of the independent variable global , white matter signal x WM , cerebrospinal fluid signal x CSF The amount of data, y is the estimated value of the corresponding Y, that is, in a set of independent variables x global 、x WM 、x CSF The average value of Y when taking values.

[0099] Calculate the independent variable x global 、x WM 、x CSF After calculating the partial regression sum of squares, compare their sizes and remove the independent variables with small regression contribution, that is, small partial regression sum of squares.

[0100] (6) Spatial standardization;

[0101] (7) Filtering: A specific portion of the frequency band generally reflects spontaneous neural activity. The embodiment of the present invention uses a bandpass filter for filtering to reduce the influence of low-frequency drift and high-frequency noise. The bandwidth of the bandpass filter is the difference between the higher frequency band f1 and the lower frequency band f2.

[0102] Step S3, extracting voxel ontology, voxel cluster covariance, and voxel network features from the resting-state functional magnetic resonance imaging data of the two groups of people in different states after preprocessing. The specific process is as follows:

[0103] Extraction of voxel ontology activity: Voxel ontology activity is the information value of the voxel. Each voxel in the brain contains a large number of neuron sets, and its activity level represents the intensity of the spontaneous neural activity of the voxel in the brain. The specific calculation method is:

[0104] 1) First, convert the time series of each voxel into the frequency domain using the following formula:

[0105]

[0106] Assume that at a certain point e after the time series is transformed into the frequency domain, use the complex number a e +b e i represents, F(t) represents the signal expression corresponding to point e (e starts from 1), S N Represents the number of sampling points when converting the time series from the time domain to the frequency domain, F S Indicates the sampling frequency, atan2(b e ,a e ) represents the phase of point e in the frequency domain;

[0107] 2) The information value of each voxel, i.e., the voxel body activity, is calculated by the following formula:

[0108]

[0109] Among them, ρ voxel Represents the information value of each voxel, T voxel Indicates the period of the time series corresponding to each voxel.

[0110] Extraction of voxel cluster covariance: Voxel cluster covariance is used to evaluate the mutual information of a cluster of voxels, that is, to measure the correlation between a given voxel and its nearest neighbor voxel. It is often used as an indicator of local brain activity level in research. Its calculation formula combines the time point information of the corresponding time series. The specific calculation formula is as follows:

[0111]

[0112] Where V represents the mutual information value of a given voxel, i.e., the local covariance of the voxel cluster, ranging from 0 to 1; is the rank sum of the time points of the ith time series, yes is the average value, C is the number of time series in the tested voxel cluster, that is, the number of voxels in the tested voxel cluster (a given voxel plus the number of its nearest neighboring voxels), and m is the number of levels, that is, the total number of time points in the time series.

[0113] Extraction of voxel network global covariance: Voxel network global covariance is the brain network connection strength between each voxel and all other voxels in the whole brain. The calculation process is as follows:

[0114] (a) Given any voxel, perform an exponential correlation analysis on its time series with the time series of other voxels in the brain. The calculation process is as follows:

[0115]

[0116] Among them, r z represents the connection strength between the time series x of a given voxel and the time series y′ of other voxels, x i Indicates the value corresponding to each index position of the time series x, y′ i Represents the value corresponding to each index position of the time series y′, that is, n TS It represents the dimension of the time series of a given voxel, that is, the length of the time series. The time series of each voxel have the same dimension. represents the average value of the corresponding time series x,

[0117] (b) Calculate the connection strength between any voxel and other voxels, that is, the r between any voxel and other voxels z After the value is reached, the threshold requirement (r z ≥0.25) positive correlation r z The connection strength of a given voxel can be obtained by adding the values, and the global covariance of the corresponding voxel network can be obtained, because in r z The calculation of voxel network global covariance is optimal when ≥0.25, so the threshold is set to 0.25.

[0118] Step S4, dividing the brain into regions, and then calculating the average values ​​of the three features of voxel body activity, voxel cluster local covariance, and voxel network global covariance for each brain region of each subject in the two groups of people with different states (the ratio of the total feature value of the brain region to the total prime number) to reduce the number of features;

[0119] The brain is divided into regions using predefined brain maps / brain templates (e.g., Automated Anatomical Labelling (AAL), Brainnetome Atlas). Different brain maps have different division bases and the number of brain regions N is different. For example, Automated Anatomical Labelling (AAL) divides N=116, and BrainnetomeAtlas divides N=246. All voxels in each brain region will calculate the above three scales, namely, voxel body activity, local covariance of voxel clusters, and global covariance of voxel networks. After the calculation is completed, the mean function of Matlab is called for each scale to return the average value of all voxels in the brain region. At each scale, each subject will obtain a 1×N feature matrix.

[0120] Step S5, feature fusion: the features of all the tested objects in the two groups of people with different states at three different feature scales of voxel body activity, local covariance of voxel clusters, and global covariance of voxel networks are fused to obtain a feature fusion matrix of the number of tested objects in the two groups of people with different states × 3N dimensions. First, the features of each tested object in the two groups of people with different states at three different feature scales of voxel body activity, local covariance of voxel clusters, and global covariance of voxel networks are fused to obtain 3N features of each tested object in the two groups of people with different states, and then the 3N features of all the tested objects in the two groups of people with different states are fused to obtain a feature fusion matrix of the number of tested objects in the two groups of people with different states × 3N dimensions.

[0121] Step S6, using the feature fusion matrix of the two groups of people in different states and the labels of the subjects of the two groups of people in different states to train the SVM classifier, and after training and testing, obtain the SVM classifier for classifying the resting-state functional magnetic resonance data of the brain;

[0122] Step S7: using the SVM classifier for classifying resting-state functional magnetic resonance data of the brain to classify the objects to be classified.

[0123] Furthermore, after obtaining the feature fusion matrix of the number of subjects × 3N dimensions for the two groups of people in different states in step S5, secondary feature selection can be performed on the feature fusion matrix of the two groups of people in different states using methods such as recursive feature elimination, and then the feature set obtained by the secondary feature selection is used to train the SVM classifier, and the leave-one-out cross-validation is used. After training and testing, the SVM classifier for classifying resting-state functional magnetic resonance imaging data of the brain is obtained.

[0124] Furthermore, the specific process of classifying the objects to be classified by using the SVM classifier for classifying the resting-state functional magnetic resonance data of the brain in step S7 is as follows:

[0125] Acquiring resting-state functional magnetic resonance imaging data of a subject to be classified;

[0126] Using steps S2 to S5, a feature fusion matrix of the object to be classified is obtained;

[0127] Use recursive feature elimination method and other methods to perform secondary feature selection on the feature fusion matrix of the object to be classified;

[0128] The feature set obtained by the secondary feature selection is input into the SVM classifier for classification of brain resting-state functional magnetic resonance data to classify the objects to be classified.

[0129] The specific implementation process of secondary feature selection for the feature fusion matrix of two groups of people in different states is as follows:

[0130] (1) The leave-one-out method was used to perform statistical tests on the characteristics of the two groups of people with different statuses, in order to preliminarily screen the characteristics with significant differences between the two groups of people with different statuses and obtain the characteristic matrix containing the characteristics with significant differences between the two groups of people with different statuses.

[0131] The purpose of the leave-one-out method is to eliminate individual differences. The specific approach is:

[0132] Each time a statistical test is performed, one of the subjects in the two groups with different statuses is removed in turn, and as many statistical tests are performed as there are subjects.

[0133] The purpose of statistical tests is to screen out features with significant differences. After each statistical test, the features with significant differences are retained, and then the feature matrix containing the features with significant differences is obtained. The specific method is:

[0134] The two-sample T test is performed using the feature fusion matrix of two groups of people with different status. Each column in the feature fusion matrix of two groups of people with different status will get a t PN Value, according to t PN The value of t determines whether each column in the feature fusion matrix of two groups of people with different status has significant differences. PN The value is calculated by the following formula:

[0135]

[0136] Among them, t PN It represents the two-sample T test statistic of two groups of people with different status. The corresponding test level p can be obtained by querying the statistical test table. It is generally believed that the test level p<0.05 is significant. It represents the average value of each column in the feature fusion matrix (number of tested objects × 3N) of the first group of tested objects (people in different states), Represents the average value of each column in the feature fusion matrix of another set of tested objects, m patient Represents the number of subjects in the first group of people, m NC Represents the number of subjects in another group of people, S patient 2 represents the variance of the first group of people in different states calculated according to their feature fusion matrix, S NC 2 Represents the variance of another group of people in different states calculated based on their feature fusion matrix.

[0137] (2) After obtaining the feature matrix containing significant differences between the two groups of people with different statuses in the initial screening, in order to ensure that the data is not dominated by a large value, the mean variance is normalized by column. The calculation formula is as follows:

[0138]

[0139] Among them, x * is the normalized data, x Y is the original data of each column in the feature matrix of two groups of people with different status after preliminary screening, containing significant differences in characteristics, μ Y According to x Y The calculated expectation, σ Y According to x Y The calculated variance.

[0140] (3) A loss function is added to the existing linear SVM classifier model to construct an SVM classifier, and the normalized feature matrices of two groups of people in different states are used, and a matrix with (the number of tested objects - 1) rows and the number of columns being the feature dimension of the normalized feature matrix is ​​used as a training sample. It is assumed that the label of one group of people in one state is 1 and the label of the other group of people in another state is -1. A corresponding label matrix of (the number of tested objects - 1)*1 dimension is formed according to the labels of the two groups of people in different states, and the SVM classifier is trained using the training samples and the label matrix of (the number of tested objects - 1)*1 dimension;

[0141] (4) In order to obtain the most representative features, the packaging method is used for secondary feature selection. The trained SVM classifier is used to select the optimal feature subset from the normalized feature matrix of two groups of people in different statuses. The packaging method is used to reduce the dimension of the features again, and the classifier itself is combined to discard irrelevant features. The features selected based on the classifier error probability should be the most effective features.

[0142] Then, the optimal feature subset is used to continue training the SVM classifier, and the leave-one-out cross-validation method is adopted. After training and testing, the SVM classifier for classifying the resting-state functional magnetic resonance imaging data of the brain is obtained. When the feature fusion matrix of the object to be classified is subjected to secondary feature selection by the recursive feature elimination method, after the normalized feature matrix of the object to be classified is obtained through steps (1) and (2), step (4) is directly entered to select the optimal feature subset of the object to be classified.

[0143] The implementation process of secondary feature selection using the packaging method is as follows Figure 2 As shown, the specific steps are as follows:

[0144] a. Calculate the initial classification accuracy accuracy0 of the SVM classifier;

[0145] b. Use the leave-one-out cross-validation method to subtract one column from the normalized feature matrix of two groups of people with different status or the objects to be classified, input the feature matrix composed of the remaining features into the SVM classifier, and calculate the current classification accuracy of the SVM classifier i , perform as many calculations as there are columns in the normalized feature matrix of two groups of people in different states or objects to be classified, and find all current classification accuracies that are greater than or equal to the initial classification accuracy 0 i , determine the corresponding feature to be removed i ;

[0146] c. Delete the feature determined in step b from the normalized feature matrix of two groups of people in different states or objects to be classified i To update the corresponding normalized feature matrix; and use the updated normalized feature matrices of the two groups of people in different states to obtain updated training samples to continue training the SVM classifier;

[0147] d. Repeat steps ac until deleting any feature in the normalized feature matrix of the two groups of people in different states or the objects to be classified will not improve the initial classification accuracy accuracy0 of the SVM classifier. At this time, the normalized feature matrix of the two groups of people in different states or the objects to be classified corresponding to the classification accuracy accuracy0 is the effective feature set of the two groups of people in different states or the objects to be classified; the obtained number of objects to be tested * the number of effective features N effective The feature matrix of the dimension is the optimal feature subset for two groups of people in different states or the objects to be classified.

[0148] Using statistical test + packaging method, an optimal feature subset is selected from the feature fusion matrix of two groups of people in different states or objects to be classified. The optimal feature subset can make the performance measurement indicators of the classification learning algorithm reach the best (such as classification accuracy). The main idea is to repeatedly construct the SVM model, select the worst feature according to the classification accuracy of the SVM model and eliminate it, and then repeat this process on the remaining feature sets until all features are traversed.

[0149] Construction and detection of classification model: The embodiment of the present invention adds a loss function to the existing linear SVM classifier model, which is characterized by allowing some samples not to meet the constraint conditions (i.e., y i (ω df x i +b df )≥1), and at the same time, the number of samples that do not meet the constraints can be reduced as much as possible to improve the classification performance of the model, that is, some samples are allowed to make mistakes, but the samples that make mistakes should be punished.

[0150] The specific steps are: using the loss function ξ hl As a slack variable, each sample corresponds to a slack variable ξ hl1 , hl2 ,……ξ hlNsubject , and the degree to which each sample does not satisfy the constraints is different, so the penalty term C imposed on it penality is also different, corresponding to C penality1 , C penality2 , ...C penalityNsubject , multiplying and summing these two for all samples gives the total classification error C penality1 ξ hl1 , C penality2 ξ hl2 , ...C penalityNsubject ξ hlNsubject The original objective function plus the total classification error calculated above is the new objective function with the loss function added. The specific calculation formula is:

[0151]

[0152] Where e represents N subject *N effective dimensional feature matrix, N subject 、N effective Represent the number of samples and the number of effective features respectively, ω df is the normal vector of the decision function, b df is the bias term of the decision function, y i Yes i The category it belongs to, its value is -1 or +1, y i(ω dr e i +b df ) represents the slack variable corresponding to each sample, namely ξ hl1 , hl2 ,……ξ hlNsubject , Represents the total classification error, and + represents the positive function, that is:

[0153]

[0154] The above is the new classification model constructed. Next, the classification model is tested. The optimal feature subsets of the two groups of people in different states and the real labels of the two groups of people in different states are used as the input of the constructed classifier. The method selected is the leave-one-out cross-validation, that is, each time the number of tested objects obtained is * the number of effective features N. effective Remove one subject (i.e., one row) from the feature matrix, use the remaining feature matrix as the training set to train the classifier, and use the removed row as the test set to detect the classification model. The same operation is performed on the label matrix, and the corresponding row is removed each time. The output is the predicted labels for two groups of people in different states, and the classification performance is evaluated based on the true labels and the predicted labels.

[0155] Example 2

[0156] The excluded subjects are used to test the ability of the SVM classifier to reliably classify new cases. This process is repeated for each subject to obtain a relatively unbiased estimate of generalization ability. The performance of the algorithm can be described by its sensitivity, specificity, accuracy, and AUC (area under the receiver operating characteristic curve).

[0157] Sensitivity is the proportion of all outcomes that the model correctly predicts to have a true value of 1, and is calculated as follows:

[0158]

[0159] In contrast, specificity is the proportion of all outcomes for which the model correctly predicted the true value to be -1 (prediction stability), and is calculated as follows:

[0160]

[0161] Among them, TP L represents the total number of experts predicted as experts, TN L represents the total number of normal controls predicted as normal controls, FN L represents the total number of normal controls predicted as experts, FP L Represents the total number of people who predicted experts as ordinary controls.

[0162] Finally, in order to verify whether the average accuracy obtained by our proposed SVM classifier is significantly higher than the average accuracy of random guessing, we used a permutation test, the specific steps of which are to randomly permute the labels of the training data 1000 times before training, and then input them into the SVM classifier to undergo the entire classification process. The average accuracy is obtained in all permutations, and the p value is calculated as the proportion of the average accuracy obtained by our method (statistical significance is set to p<0.05), and the calculation formula is as follows:

[0163]

[0164] Among them, N random Indicates the number of times the labels of the training data are randomly permuted; n accuraty It indicates the number of times in N permutation tests that the average accuracy calculated by the sampling population is greater than the average accuracy calculated by our model.

[0165] In order to test the method of the embodiment of the present invention, the real data set of resting state functional magnetic resonance images of two different states of people, ordinary people and experts in a certain field, is used for testing. The specific process is as described above. The resting state functional magnetic resonance imaging data of ordinary people and experts are first obtained, and then preprocessed. The preprocessed resting state functional magnetic resonance imaging data of ordinary people and experts are extracted for voxel body activity, local covariation of voxel clusters, and global covariation of voxel networks, and the average values ​​of the three features of voxel body activity, local covariation of voxel clusters, and global covariation of voxel networks are calculated for each brain region of each tested object. After the above features are fused, the feature fusion matrix of two groups of people in different states is obtained. The optimal feature subset is selected using the statistical test + packaging method. The optimal feature subset and the real labels of the two groups of people in different states, ordinary people (labeled as -1) and experts (labeled as 1), are used as the input of the SVM classifier, and the output is the predicted labels of the two groups of people in different states. The classification performance is evaluated based on the real labels and the predicted labels, and the classification efficiency based on single scale and multi-scale on the real data set is obtained as shown in Table 1.

[0166] Table 1 Classification results of support vector machines based on different feature sets

[0167]

[0168] The results show that the average accuracy rate can reach about 70%. Multi-scale feature fusion performs best in distinguishing two different groups of people, with an accuracy rate of 85.36%, a sensitivity of 0.8958, and a specificity of 0.7941, which are all higher than the results of classification using a single feature. This is our additional effect.

[0169] Multi-scale feature fusion is used to show which brain regions are different between people in two different states, and then the existence of cross-scale brain regions is highlighted to locate the most likely brain regions with differences. Cross-scale brain regions refer to brain regions that have significant differences in two or more scales of voxel body activity, local covariance of voxel clusters, and global covariance of voxel networks. The most likely brain regions with differences are most likely to show differences in brain information between the two groups of people. Further analysis of the functions and networks of these brain regions can reveal their differences in brain representation and information processing mechanisms. If the two groups of people in different states are experts and normal people, the brain regions can be located and enhanced for efficient training.

[0170] In order to further compare the classification effects based on different scales, the ROC curve of each scale is plotted as follows Figure 3 As shown. The horizontal axis of the ROC curve represents the false positive rate (FPR), and the vertical axis represents the sensitivity (TPR). The closer the curve is to the upper left corner, the higher the FPR is, and the lower the FPR is, the better the classification performance is. The area under the ROC curve of the local covariance of the voxel cluster is 0.7298, the voxel body activity is 0.6415, the voxel network global covariance is 0.8088, and the multi-scale is 0.8278. It can be seen that the multi-scale feature fusion performs best in distinguishing two different groups of people (the closer to the upper left corner of the ROC curve, the better the classification performance).

[0171] This process obtains the sampling population of the small sample experiment (approximately normal distribution). Our observed value 0.6951 is near the right tail of the sampling population, indicating that this value rarely occurs. There are 6 values ​​greater than 0.6951 in the sampling population obtained in the permutation test, so the estimated P-value is 6 / 999≈0.006. The results are statistically significant, and it can be considered that the average accuracy obtained by our model is higher than the average accuracy of random guessing.

[0172] Among the 3N features of the multi-scale feature fusion of the current two groups of people in different statuses, 8 effective features were finally selected for SVM classification, as shown in Table 2.

[0173] Table 2 Effective features for distinguishing two brain states

[0174]

[0175] Table 2 shows the specific information of these effective features, which is our main result. In order to express the results more clearly, the distribution of brain regions is visualized, such as Figure 4This result provides a deeper understanding of the characterization of resting-state functional magnetic resonance imaging in two different groups of people. These brain regions are mainly located in the inferior frontal gyrus, inferior temporal gyrus, and insula. It can be found that the brain region with the greatest discrimination ability is the right inferior frontal gyrus, which may become a promising classification marker for the two different groups of people. Its practical significance is that we can train ordinary people to achieve the ability of experts by targeting the brain region with the greatest discrimination ability.

[0176] Example 3

[0177] An embodiment of the present invention proposes a brain resting-state functional magnetic resonance data classification system that integrates multi-scale information, including: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement a brain resting-state functional magnetic resonance data classification method that integrates multi-scale information as described in Example 1 above.

[0178] A brain resting state functional magnetic resonance data classification system integrating multi-scale information may include an internal communication bus, a processor, a read-only memory (ROM), a random access memory (RAM), a communication port, and a hard disk. The internal communication bus can realize data communication between components of a brain resting state functional magnetic resonance data classification system integrating multi-scale information. The processor can make judgments and issue prompts. In some embodiments, the processor can be composed of one or more processors. The communication port can realize data communication outside a brain resting state functional magnetic resonance data classification system integrating multi-scale information. In some embodiments, a brain resting state functional magnetic resonance data classification system integrating multi-scale information can also send and receive information and data from a network through a communication port. A brain resting state functional magnetic resonance data classification system integrating multi-scale information can also include different forms of program storage units and data storage units, such as a hard disk, a read-only memory (ROM) and a random access memory (RAM), which can store various data files used for computer processing and / or communication, and possible program instructions executed by the processor. The processor executes these instructions to implement the main part of the method. The results of the processor processing are transmitted to the subject device through the communication port and displayed on the subject interface.

[0179] The above-mentioned method for classifying resting-state functional magnetic resonance data of the brain by fusing multi-scale information can be implemented as a computer program, stored in a hard disk, and recorded in a processor for execution. Therefore, an embodiment of the present invention also provides a computer-readable medium storing a computer program code, and when the computer program code is executed by a processor, the above-mentioned method for classifying resting-state functional magnetic resonance data of the brain by fusing multi-scale information is implemented.

[0180] When a method for classifying resting-state functional magnetic resonance data of the brain that integrates multi-scale information is implemented as a computer program, it can also be stored in a computer-readable storage medium as a product. For example, a computer-readable storage medium may include, but is not limited to, a magnetic storage device (e.g., a hard disk, a floppy disk, a magnetic strip), an optical disk (e.g., a compact disk (CD), a digital versatile disk (DVD)), a smart card, and a flash memory device (e.g., an electrically erasable programmable read-only memory (EPROM), a card, a stick, a key drive). In addition, the various storage media described in the embodiments of the present invention can represent one or more devices and / or other machine-readable media for storing information. The term "machine-readable medium" may include, but is not limited to, wireless channels and various other media (and / or storage media) that can store, contain and / or carry code and / or instructions and / or data.

[0181] It should be understood that the above-mentioned embodiments are only illustrative. The embodiments described in the present invention may be implemented in hardware, software, firmware, middleware, microcode or any combination thereof. For hardware implementation, the processing unit may be implemented in one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors and / or other electronic units designed to perform the functions described in the present invention or in combination thereof.

[0182] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A method for classifying resting-state functional magnetic resonance imaging data of the brain by integrating multi-scale information, characterized in that: Follow these steps: Step S1, obtaining resting-state functional magnetic resonance imaging data of each subject in two groups of people in different states; Step S2, preprocessing the resting-state functional magnetic resonance imaging data of two groups of people in different states to eliminate noise; Step S3, extracting voxel body activity, local covariation of voxel clusters, and global covariation of voxel networks from the resting-state functional magnetic resonance imaging data of the two groups of people in different states after preprocessing; Step S4, dividing the brain into regions, and then calculating the average values ​​of the three features of voxel body activity, voxel cluster local covariation, and voxel network global covariation for each brain region of each subject in the two groups of people with different states. At each feature scale, each subject in the two groups of people with different states will obtain a 1×N dimensional feature matrix; Step S5, fusing the 1×N-dimensional feature matrices of all the subjects in the two groups of people with different states at three different feature scales, namely, voxel body activity, voxel cluster local covariance, and voxel network global covariance, to obtain the feature fusion matrices of the two groups of people with different states; Step S6, using the feature fusion matrix of the two groups of people in different states and the labels of the subjects of the two groups of people in different states to train the SVM classifier, and after training and testing, obtain the SVM classifier for classifying the resting-state functional magnetic resonance data of the brain; Step S7, using the SVM classifier for classification of resting-state functional magnetic resonance data to classify the objects to be classified; Also includes: After obtaining the feature fusion matrices of the two groups of people in different states in step S5, secondary feature selection is performed on the feature fusion matrices of the two groups of people in different states, and then the feature set obtained by the secondary feature selection is used to train the SVM classifier, and a leave-one-out cross-validation method is used. After training and testing, an SVM classifier for classifying brain resting state functional magnetic resonance data is obtained; In step S3, the extraction process of voxel body activity is as follows: 1) First, convert the time series of each voxel into the frequency domain using the following formula: Among them, the plural a e +b e i represents a point e after the time series is transformed into the frequency domain, F(t) represents the signal expression corresponding to point e, and e starts from 1; S N Represents the number of sampling points when converting the time series from the time domain to the frequency domain, F S Indicates the sampling frequency, atan2(b e ,a e ) represents the phase of point e in the frequency domain; 2) The information value of each voxel, i.e., the voxel body activity, is calculated by the following formula: Among them, ρ voxel represents the information value of each voxel, Τ voxel Indicates the period of the time series corresponding to each voxel; The local covariance of voxel clusters is calculated as follows: Where V represents the mutual information value of a given voxel, i.e., the local covariance of the voxel cluster, ranging from 0 to 1; is the rank sum of the time points of the ith time series, yes The average value, C is the number of time series in the tested voxel cluster, that is, the number of voxels in the tested voxel cluster, and m is the number of levels, that is, the total number of time points in the time series; The calculation process of voxel network global covariance is as follows: (a) Given any voxel, perform an exponential correlation analysis on its time series with the time series of other voxels in the brain. The calculation process is as follows: Among them, r z represents the connection strength between the time series x of a given voxel and the time series y′ of other voxels, x i Indicates the value corresponding to each index position of the time series x, y i ′ represents the value corresponding to each index position of the time series y′, that is, x=[x1,x2,L,x i ,L,x nTS ],y′=[y1′,y2′,L,y i ′,L,y n ' TS ];n TS It represents the dimension of the time series of a given voxel, that is, the length of the time series. The time series of each voxel have the same dimension. represents the average value of the corresponding time series x, (b) Calculate the connection strength between any voxel and other voxels, that is, the r between any voxel and other voxels z After the value is reached, the positive correlation r that meets the threshold requirement will be z The connection strength of a given voxel can be obtained by adding the values, and the global covariance of the corresponding voxel network can be obtained.

2. The method for classifying resting-state functional magnetic resonance data of the brain by integrating multi-scale information according to claim 1, characterized in that: In step S7, the specific process of using the SVM classifier for classifying the resting-state functional magnetic resonance data of the brain to classify the object to be classified is as follows: Acquiring resting-state functional magnetic resonance imaging data of a subject to be classified; Using steps S2 to S5, a feature fusion matrix of the object to be classified is obtained; Perform secondary feature selection on the feature fusion matrix of the object to be classified; The feature set obtained by the secondary feature selection is input into the SVM classifier for classification of brain resting-state functional magnetic resonance data to classify the objects to be classified.

3. The brain resting state functional magnetic resonance imaging data classification method integrating multi-scale information according to claim 1 or 2, characterized in that: Using recursive feature elimination method, secondary feature selection is performed on the feature fusion matrix of two groups of people in different states or objects to be classified, so as to obtain the optimal feature subsets of the two groups of people in different states or objects to be classified; After obtaining the optimal feature subsets of the two groups of people in different states, the optimal feature subsets of the two groups of people in different states are used to train the SVM classifier to obtain the SVM classifier for classifying the brain resting state functional magnetic resonance data.

4. The brain resting state functional magnetic resonance imaging data classification method integrating multi-scale information according to claim 1 or 2, characterized in that: The specific implementation process of secondary feature selection for the feature fusion matrix of two groups of people in different states is as follows: (1) The feature fusion matrix of the two groups of people with different statuses was statistically tested using the leave-one-out method to preliminarily screen the features with significant differences between the two groups of people with different statuses and obtain the feature matrix containing the features with significant differences between the two groups of people with different statuses; (2) The feature matrix containing significant differences between the two groups of people in different states is normalized by mean variance by column to obtain the normalized feature matrix of the two groups of people in different states. The calculation formula for mean variance normalization is as follows: Among them, x * is the normalized data, x Y is the original data of each column in the feature matrix of two groups of people with different status after preliminary screening, containing significant differences in characteristics, μ Y According to x Y The calculated expectation, σ Y According to x Y Calculated variance; (3) A loss function is added to the existing linear SVM classifier model to construct an SVM classifier, and the normalized feature matrices of two groups of people in different states are used, and a matrix with (the number of tested objects - 1) rows and the number of columns being the feature dimension of the normalized feature matrix is ​​used as a training sample. It is assumed that the label of one group of people in one state is 1 and the label of the other group of people in another state is -1. A corresponding label matrix of (the number of tested objects - 1) × 1 dimensions is formed according to the labels of the two groups of people in different states, and the SVM classifier is trained using the training samples and the (the number of tested objects - 1) × 1 dimension label matrix; (4) Using the packaging method, the trained SVM classifier is used to select the optimal feature subset from the normalized feature matrices of two groups of people with different status; Then the optimal feature subset is used to continue training the SVM classifier, and the leave-one-out cross-validation method is adopted. After training and testing, the SVM classifier for classifying the resting-state functional magnetic resonance imaging data of the brain is obtained; When the recursive feature elimination method is used to perform secondary feature selection on the feature fusion matrix of the object to be classified, after obtaining the normalized feature matrix of the object to be classified through steps (1) and (2), step (4) is directly entered to select the optimal feature subset of the object to be classified.

5. The method for classifying resting-state functional magnetic resonance data of the brain by integrating multi-scale information according to claim 4, characterized in that: The specific implementation process of using the trained SVM classifier to select the optimal feature subset from the normalized feature matrix of two groups of people in different states or objects to be classified is as follows: a. Calculate the initial classification accuracy accuracy0 of the current SVM classifier; b. Use the leave-one-out cross-validation method to subtract one column from the normalized feature matrix of two groups of people with different status or the objects to be classified, input the feature matrix composed of the remaining features into the SVM classifier, and calculate the current classification accuracy of the SVM classifier i , perform as many calculations as there are columns in the normalized feature matrix of two groups of people in different states or objects to be classified, and find all current classification accuracies that are greater than or equal to the initial classification accuracy 0 i , determine the corresponding feature to be removed i ; c. Delete the feature determined in step b from the normalized feature matrix of two groups of people in different states or objects to be classified i To update the corresponding normalized feature matrix; and use the updated normalized feature matrices of the two groups of people in different states to obtain updated training samples to continue training the SVM classifier; d. Repeat steps ac until deleting any feature in the normalized feature matrix of the two groups of people in different states or the objects to be classified will not improve the initial classification accuracy accuracy0 of the SVM classifier. At this time, the normalized feature matrix of the two groups of people in different states or the objects to be classified corresponding to the classification accuracy accuracy0 is the effective feature set of the two groups of people in different states or the objects to be classified; the obtained number of objects × the number of effective features N effective The feature matrix of the dimension is the optimal feature subset for two groups of people in different states or the objects to be classified.

6. The method for classifying resting-state functional magnetic resonance data of the brain by integrating multi-scale information according to claim 4, characterized in that: A loss function is added to the existing linear SVM classifier model. The specific calculation formula of the new objective function with the loss function is: Where e represents N subject ×N effective dimensional feature matrix, N subject Indicates the number of samples, N effective represents the effective feature number, e i represents any row of e; ω df is the normal vector of the decision function, b df is the bias term of the decision function, y i Yes i The category to which it belongs is the label, whose value is -1 or +1; i (ω df e i +b df ) represents the slack variable corresponding to each sample, Represents the total classification error, and + represents the positive function, that is:

7. The method for classifying resting-state functional magnetic resonance data of the brain by integrating multi-scale information according to claim 1, 2, 5 or 6, characterized in that: The specific implementation process of step S4 is as follows: The brain is divided into regions using a predefined brain map / brain template. All voxels in each brain region will calculate three characteristic scales: voxel body activity, local covariance of voxel clusters, and global covariance of voxel networks. After the calculation is completed, the mean function of Matlab is called for each characteristic scale to return the average value of all voxels in the brain region. At each characteristic scale, each subject in the two different states will obtain a 1×N characteristic matrix.

8. A brain resting-state functional magnetic resonance data classification system integrating multi-scale information, characterized in that: include: a memory for storing instructions executable by a processor; A processor, configured to execute the instructions to implement the method according to claim 1, 2, 5 or 6.

9. A computer-readable medium, characterized in that A computer program code is stored, which, when executed by a processor, implements the method according to claim 1, 2, 5 or 6.

Images