Multi-site functional magnetic resonance imaging feature extraction method

Abstract

The invention discloses a multi-site functional magnetic resonance imaging feature extraction method. The method comprises the following steps: acquiring multi-site resting state functional magnetic resonance imaging data and characterization data; performing brain region division on the preprocessed functional magnetic resonance imaging by using a plurality of templates; respectively extracting a time sequence of each brain region; calculating the correlation between the time sequences of every two brain regions to construct a brain function network; removing data heterogeneity between different sites by using a balanced homogenization algorithm; extracting identification features in the brain function network by adopting a graph attention network and a convolutional neural network; constructing a group diagram by using the multi-template features and the representation data; and dividing the group graph into a test set and a training set, and classifying by adopting cross validation and a graph wavelet neural network. According to the method, complementary characteristics of a brain function network can be deeply excavated, the utilization efficiency of functional magnetic resonance imaging is improved, the reliability and consistency of multi-site data are ensured, and a reference basis is provided for research on mental and psychological diseases and neurodegenerative diseases.

Description

Technical Field

[0001] The invention belongs to the field of biomedical information technology, relates to a brain function network, and specifically is a multi-site functional magnetic resonance imaging feature extraction method. Background Art

[0002] Brain science issues are one of the basic scientific issues facing human society and are the "ultimate frontier" for human understanding of nature and human beings themselves. In the human brain, various types of neuronal cells are connected together through synapses to form a very complex network, which is the structural basis for the brain to carry out various physiological and cognitive activities. Magnetic resonance imaging is a non-invasive imaging technology that plays an important role in the study of the brain. By using magnetic fields and radio frequency waves, magnetic resonance imaging can generate high-resolution brain structure images to help scientists observe the anatomical details of brain tissue. In addition, functional magnetic resonance imaging has further expanded the application of magnetic resonance imaging, which can detect brain activity patterns in real time and reflect changes in neural activity under specific cognitive tasks or stimulation. Magnetic resonance imaging is a high-dimensional and complex brain image, and how to extract features from brain imaging data is particularly critical. For structural magnetic resonance imaging, feature extraction usually involves anatomical features such as gray matter volume, cortical thickness, and brain region morphology; while for functional magnetic resonance imaging, feature extraction focuses more on time series, such as functional connectivity calculated based on blood oxygen level-dependent signals.

[0003] Brain functional network refers to dividing the neurons in the brain into different brain regions according to specific templates. The functions between brain regions are integrated using the correlation relationship of blood oxygen level-dependent signals, and the dynamic coordination of signals between different brain regions is recorded through functional connections. Studying functional networks can explore the functional organization pattern of the human brain and help understand the functional activity patterns of the human brain in normal or diseased states. However, most existing studies use a single template to extract features from the brain. For example, patent application CN115937581A "A classification and recognition method for multi-site fMRI data" only uses physiological brain maps for brain region division, which cannot fully represent the information in magnetic resonance imaging and has limitations.

[0004] In recent years, deep learning has been widely used in the study of brain functional networks. This research method requires a large amount of sample data. Due to the high cost of acquiring functional magnetic resonance imaging, the amount of data from a single site is small. In order to solve this problem, some institutions have pooled functional magnetic resonance imaging data from different sites to form a public data set. Public data sets have solved the problem of insufficient data to a certain extent, but they have also brought new challenges. The scanners used in different sites come from different manufacturers, the process of collecting data is not unified, and other problems make the data between different sites heterogeneous. Therefore, a feature extraction method is needed that can not only remove the data heterogeneity between different sites, but also solve the feature deviation caused by a single template. Summary of the invention

[0005] The technical problem to be solved by the present invention is that when extracting features from brain functional networks, the existing technology mostly uses a single brain region division template for extraction, which cannot effectively cope with the heterogeneity of functional magnetic resonance imaging data between different sites, and there is a certain deviation when extracting features.

[0006] The technical solution of the present invention is: a multi-site functional magnetic resonance imaging feature extraction method, which is characterized by comprising:

[0007] Step 1: Obtain multi-site resting-state functional magnetic resonance imaging data and characterization data, and pre-process the image data to eliminate the effects of brain volume differences and shaking, and obtain pre-processed images of each site;

[0008] Step 2: Use multiple brain region division templates to divide the preprocessed functional magnetic resonance imaging into brain regions, obtain multiple template division results, and extract the time series of each brain region respectively;

[0009] Step 3: Calculate the correlation between the time series of each pair of brain regions in each template division result, construct the brain function network under each template, and obtain the functional connection matrix;

[0010] Step 4: Use the balanced homogenization algorithm based on location and scale effects to process the obtained functional connectivity matrix to remove data heterogeneity between different sites;

[0011] Step 5: Use the graph attention network to process the brain function networks of different templates, aggregate the key information of each brain region for the brain function networks of each template, extract the discriminative features from them through the convolutional neural network, and splice and fuse the discriminative features of different templates into multi-template features;

[0012] Step 6: Take the subjects in multiple sites as nodes, use the multi-template features as node features, use the multi-template features and the similarity of the representation data to construct edges, and construct a group graph;

[0013] Step 7: Divide the group graph into a test set and a training set, use cross-validation and graph wavelet neural network for classification training, and the trained graph wavelet neural network is used to classify functional magnetic resonance imaging.

[0014] Furthermore, in step 1, the data are preprocessed through a configurable brain connectome analysis pipeline, including slice time correction, motion correction, skull removal, voxel intensity normalization, noise signal removal, bandpass filtering, and spatial registration.

[0015] Furthermore, in step 2, the brain region division template includes at least two of the AAL template, the HO template, and the CC200 template.

[0016] Furthermore, in step 3, the Pearson correlation is used to calculate the correlation between the time series of each pair of brain regions, and the calculation formula is:

[0017]

[0018] In the formula, r i and r j denote the time series of the ith and jth brain regions respectively, and E(·) denotes the mathematical expectation of the sequence.

[0019] Furthermore, in step 4, the balanced homogenization algorithm is used to homogenize the data between different sites to remove the heterogeneity of the data between different sites. It is assumed that the functional connectivity value y of all site data is m,n,v Obey the model:

[0020] y m,n,v =a v +X m,n β v +γ m,v +δ m,v ε m,n,v (2)

[0021] Where m is the site number; n is the subject number; v is the functional connection matrix vector number of the subject; a v represents the mean of the vth connection value of all subjects; X m,n is the design matrix for the covariates; β v Yes and X m,n The corresponding regression coefficient vector; γ m,v and δ m,v Corresponding to the additive and multiplicative effects of site connectivity, respectively; ε m,n,v represents the residual term; y m,n,v Data after balanced homogenization algorithm It is expressed as:

[0022]

[0023] In the formula, a v , β v , γ m,v , a v and δ mv They are respectively v , β v , γ m,v , a v and δ m,v The empirical Bayesian estimate of .

[0024] Furthermore, in step 5, the graph attention network input is the node feature F = [F1, F2, ..., F N ] and the adjacency matrix representing the edge features; the relationship between node p and its first-order adjacent nodes is represented by the attention coefficient:

[0025]

[0026] In the formula, a pq is the attention coefficient between node p and node q, expressed as the connection between nodes; represents the nodes that are first-order adjacent to node p in the adjacency matrix; W is a linear change matrix; [·||·] represents the features after serial transformation; σ() is a shared attention mechanism, which is implemented using a single-layer feedforward neural network; LeakyRelu is a nonlinear activation function of the feedforward neural network; after calculating the attention coefficients of all nodes, the node features are expressed as:

[0027]

[0028] In the formula, F′ p Represents the output features of node p after aggregating the features of neighboring nodes; uses a three-layer convolutional neural network to convolve the output features to obtain feature vectors; and concatenates feature vectors of different templates to obtain multi-template features.

[0029] Furthermore, in step 6, a subject is regarded as a node in the population graph, and the feature of each node is the node feature obtained in step 5; the edge is constructed by calculating the similarity between the node features and the similarity of the node representation data, and the node representation data includes gender, age and station; the definition of the population graph adjacency matrix A is:

[0030]

[0031] In the formula, Sim(N u ,N w ) represents the similarity between node features of node u and node w; γ represents the similarity between the representation data of node u and node w;

[0032]

[0033] Where r is the Pearson correlation; x(u) and x(w) are the node feature vectors of node u and node w, respectively; σ is the kernel width; for categorical representation data, γ is defined by the Kronecker function:

[0034]

[0035] For numerical representation data, the step function of γ with a threshold is defined as:

[0036]

[0037] Furthermore, in step 7, five-fold cross validation is used to divide the subjects in the group graph into five groups, and each group of subjects is used as a test set for one round of training in the order of the subjects' arrangement. In this round of training, the remaining subjects outside the test set are used as training sets and input into the graph wavelet neural network for training. The above process is repeated for five rounds of training and the classification performance indicators are calculated.

[0038] The beneficial effects of the present invention are:

[0039] 1. Through multi-template brain region division, we can deeply explore the complementary characteristics of the brain functional network, avoid the feature deviation caused by a single template, improve the accuracy of functional magnetic resonance imaging classification based on brain functional network characteristics, and improve the utilization efficiency of multi-site functional magnetic resonance imaging data.

[0040] 2. Remove data heterogeneity between different sites through balanced homogenization algorithm, ensure the reliability and consistency of multi-site data, and reduce the impact of equipment differences and inconsistent collection conditions on feature extraction results.

[0041] 3. Use graph attention network and convolutional neural network to extract discriminative features in brain functional network, use multi-template features and representation data to construct group graph, and improve the analysis ability of multi-site functional magnetic resonance imaging data.

[0042] 4. The present invention helps to deeply explore the complementary characteristics of brain functional networks, improve the utilization efficiency of functional magnetic resonance imaging, ensure the reliability and consistency of multi-site data, and provide a reference for the research of mental and psychological diseases and neurodegenerative diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0044] The present invention proposes a method for extracting features from multi-site data of a public data set, and then using the extracted features to classify different brain states. The embodiment of the present invention is described by classifying autistic patients. By extracting features from multi-site data in a public data set, a neural network classification is trained to distinguish between an autistic patient group and a normal control group. In addition, by analyzing these features, it can also be used to find medical markers for autism. Below, with reference to the accompanying drawings, through the description of the embodiments, the specific implementation methods of the present invention are further described in detail to help those skilled in the art have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.

[0045] like Figure 1 As shown, the present invention provides a multi-site functional magnetic resonance imaging feature extraction method, comprising the following steps:

[0046] Step 1: fMRI data were obtained from the ABIDE-Ⅰ dataset and preprocessed through a configurable brain connectome analysis pipeline, including slice time correction, motion correction, skull removal, voxel intensity normalization, noise signal removal, bandpass filtering [0.01-0.1 Hz], and spatial registration. Then, fMRI data that did not meet the quality requirements were removed from 1112 subjects, resulting in 505 ASD patients and 530 normal patients from 17 sites, totaling 1035 subject data.

[0047] Step 2: Use AAL and HO templates to divide the preprocessed functional magnetic resonance imaging into brain regions and extract the time series of each brain region respectively.

[0048] Step 3: Calculate the correlation between the time series of each brain region in each template division result, construct the brain function network under each template, and obtain the functional connection matrix. The correlation is calculated using Pearson correlation, and the calculation formula is:

[0049]

[0050] In the formula, r i and r j denote the time series of the ith and jth brain regions respectively, and E(·) denotes the mathematical expectation of the sequence.

[0051] Step 4: Use a balanced homogenization algorithm based on position and scale effects to process the obtained functional connectivity matrix to improve the consistency of the data and thus remove the impact of data heterogeneity between different sites.

[0052] The balanced homogenization algorithm homogenizes the data between different sites to remove the heterogeneity of data between different sites. It is assumed that all the functional connectivity values ​​y in the functional connectivity matrix m,n,v Obey the model:

[0053] y m,n,v =a v +X m,n β v +γ m,v +δ m,v ε m,n,v (2)

[0054] Where m is the site number; n is the subject number; v is the vector number in the functional connection matrix expanded by the subject; a v represents the mean of the vth connection value of all subjects; X m,n is the design matrix for the covariates; β v Yes and X m,n The corresponding regression coefficient vector; γ m,v and δ m,v Corresponding to the additive and multiplicative effects of site connectivity, respectively; ε m,n,v Represents the residual term; data after balanced homogenization algorithm It can be expressed as:

[0055]

[0056] In the formula, a v , β v , γ m,v , a v and δ mv They are respectively v , β v , γ m,v , a v and δ m,v The empirical Bayesian estimate of .

[0057] Step 5: Use the graph attention network to process the brain function networks of different templates, aggregate the key information of each brain region, extract the discriminative features from them through the convolutional neural network and fuse them into multi-template features.

[0058] The input of the graph attention network is a set of node features F = [F1, F2, ..., F N ] and the adjacency matrix representing the edge features; the relationship between node p and its first-order adjacent nodes can be represented by the attention coefficient:

[0059]

[0060] In the formula, a pq is the attention coefficient between node p and node q, expressed as the connection between nodes;

[0061] represents the nodes that are first-order adjacent to node p in the adjacency matrix; W is a linear change matrix; [·||·] represents the features after serial transformation; σ() is a shared attention mechanism, which is implemented using a single-layer feedforward neural network; LeakyRelu is a nonlinear activation function of the feedforward neural network, and its input negative slope is 0.2; after calculating the attention coefficients of all nodes, the node features are expressed as:

[0062]

[0063] In the formula, F′ p Represents the output features of node p after aggregating the features of neighboring nodes; uses a three-layer convolutional neural network to convolve the output features to obtain feature vectors; and concatenates feature vectors of different templates to obtain multi-template features.

[0064] Step 6: Take the subjects in multiple sites as nodes, use the multi-template features as node features, use the multi-template features and the similarity of the representation data to construct edges, and construct a group graph.

[0065] In the population graph, a subject is regarded as a node, and the feature of each node is the multi-template feature obtained in step 5; the edge is constructed by calculating the similarity between the features of the nodes and the similarity of the representation data, such as gender, age, and site; the definition of the population graph adjacency matrix A is:

[0066]

[0067] Where Sim(N u ,N w ) represents the similarity between node features of node u and node w; γ represents the similarity between the representation data of node u and node w;

[0068]

[0069] Where r is the Pearson correlation; x(u) and x(w) are the node feature vectors of node u and node w respectively; σ is the kernel width;

[0070] For categorical representation information such as gender and site, the representation data similarity γ is defined by the Kronecker function:

[0071]

[0072] For numerical representation information such as age, the representation data similarity γ is defined by a step function with a threshold:

[0073]

[0074] Step 7: Divide the group graph into a test set and a training set, divide the subjects in the group graph into five parts using five-fold cross validation, and use each subject as a test set for a round of training in the order of the subjects. In this round of training, the remaining subjects outside the test set are used as training sets and input into the graph wavelet neural network for training. Repeat the above process for five rounds of training and calculate the classification performance index. After the graph wavelet neural network training is completed, this embodiment realizes the classification of functional magnetic resonance imaging of autistic patients, classifies autistic patients and normal control groups, and can be used to predict the state of functional magnetic resonance imaging of new subjects.

[0075] The present invention is described above by way of example in conjunction with the accompanying drawings. It is obvious that the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the inventive concept and technical solution of the present invention, or the inventive concept and technical solution are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A multi-site functional magnetic resonance imaging feature extraction method, characterized in that include: Step 1: Obtain multi-site resting-state functional magnetic resonance imaging data and characterization data, and pre-process the image data to eliminate the effects of brain volume differences and shaking, and obtain pre-processed images of each site; Step 2: Use multiple brain region division templates to divide the preprocessed functional magnetic resonance imaging into brain regions, obtain multiple template division results, and extract the time series of each brain region respectively; Step 3: Calculate the correlation between the time series of each pair of brain regions in each template division result, construct the brain function network under each template, and obtain the functional connection matrix; Step 4: Use the balanced homogenization algorithm based on location and scale effects to process the obtained functional connectivity matrix to remove data heterogeneity between different sites; Step 5: Use the graph attention network to process the brain function networks of different templates, aggregate the key information of each brain region for the brain function networks of each template, extract the discriminative features from them through the convolutional neural network, and splice and fuse the discriminative features of different templates into multi-template features; Step 6: Take the subjects in multiple sites as nodes, use the multi-template features as node features, use the multi-template features and the similarity of the representation data to construct edges, and construct a group graph; Step 7: Divide the group graph into a test set and a training set, use cross-validation and graph wavelet neural network for classification training, and the trained graph wavelet neural network is used to classify functional magnetic resonance imaging.

2. A multi-site functional magnetic resonance imaging feature extraction method according to claim 1, characterized in that in step 1, the data is preprocessed through a configurable brain connection group analysis process, including slice time correction, motion correction, skull removal, voxel intensity normalization, noise signal removal, bandpass filtering and spatial alignment.

3. A multi-site functional magnetic resonance imaging feature extraction method according to claim 1, characterized in that In step 2, the brain region segmentation template includes at least two of the AAL template, the HO template, and the CC200 template.

4. The multi-site functional magnetic resonance imaging feature extraction method according to claim 1, characterized in that in step 3, the correlation between the time series of two brain regions is calculated using Pearson correlation, and the calculation formula is: In the formula, r i and r j denote the time series of the ith and jth brain regions respectively, and E(·) denotes the mathematical expectation of the sequence.

5. The multi-site functional magnetic resonance imaging feature extraction method according to claim 1, characterized in that In step 4, the balanced homogenization algorithm is used to homogenize the data between different sites to remove the heterogeneity of the data between different sites. It is assumed that the functional connectivity value y of all site data m,n,v Obey the model: y m,n,v =a v +X m,n b v +g m,v +d m,v e m,n,v (2) Where m is the site number; n is the subject number; v is the functional connection matrix vector number of the subject; a v represents the mean of the vth connection value of all subjects; X m,n is the design matrix for the covariates; β v Yes and X m,n The corresponding regression coefficient vector; γ m,v and δ m,v Corresponding to the additive and multiplicative effects of site connectivity, respectively; ε m,n,v represents the residual term; y m,n,v Data after balanced homogenization algorithm It is expressed as: In the formula, a v , β v , γ m,v , a v and δ mv They are respectively v , β v , γ m,v , a v and δ m,v The empirical Bayesian estimate of .

6. A multi-site functional magnetic resonance imaging feature extraction method according to claim 1, characterized in that in step 5, the graph attention network input is the node feature F=[F1, F2, ..., F N ] and the adjacency matrix representing the edge features; the relationship between node i and its first-order adjacent nodes can be represented by the attention coefficient: In the formula, a pq is the attention coefficient between node p and node q, expressed as the connection between nodes; represents the nodes that are first-order adjacent to node p in the adjacency matrix; W is a linear change matrix; [·||·] represents the features after serial transformation; σ() is a shared attention mechanism, which is implemented using a single-layer feedforward neural network; LeakyRelu is a nonlinear activation function of the feedforward neural network; after calculating the attention coefficients of all nodes, the node features are expressed as: In the formula, F p ′ represents the output feature of node p after aggregating the features of neighboring nodes; the output feature is convolved using a three-layer convolutional neural network to obtain a feature vector; the feature vectors of different templates are concatenated to obtain multi-template features.

7. The multi-site functional magnetic resonance imaging feature extraction method according to claim 1, characterized in that In step 6, a subject is treated as a node in the population graph, and the feature of each node is the node feature obtained in step 5; The edges are constructed by calculating the similarity between node features and the similarity of node representation data, which includes gender, age and station. The definition of the group graph adjacency matrix A is: In the formula, Sim(N u ,N w ) represents the similarity between node features of node u and node w; γ represents the similarity between the data represented by node u and node w; Where r is the Pearson correlation; x(u) and x(w) are the node feature vectors of node u and node w, respectively; σ is the kernel width; for categorical representation data, γ is defined by the Kronecker function: For numerical representation data, the step function of γ with a threshold is defined as:

8. The multi-site functional magnetic resonance imaging feature extraction method according to claim 1, characterized in that In step 7, five-fold cross validation is used to divide the subjects in the population graph into five groups. Each group of subjects is used as a test set for one round of training in the order of the subjects' arrangement. In this round of training, the remaining subjects outside the test set are used as training sets and input into the graph wavelet neural network for training. The above process is repeated for five rounds of training and the classification performance indicators are calculated.

Images