CN115730253A - Dynamic brain network state construction method based on graph core - Google Patents
Dynamic brain network state construction method based on graph core Download PDFInfo
- Publication number
- CN115730253A CN115730253A CN202211456613.7A CN202211456613A CN115730253A CN 115730253 A CN115730253 A CN 115730253A CN 202211456613 A CN202211456613 A CN 202211456613A CN 115730253 A CN115730253 A CN 115730253A
- Authority
- CN
- China
- Prior art keywords
- network
- dynamic
- brain
- matrix
- network state
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Magnetic Resonance Imaging Apparatus (AREA)
Abstract
The invention relates to a dynamic brain network state construction method based on a graph core. The invention comprises the following steps: s10, constructing a dynamic connection matrix of each sample through a sliding window by using an interested time sequence obtained from the resting state functional magnetic resonance data, and thinning the obtained dynamic connection matrix to obtain a coefficient dynamic connection network with a more simplified structure; s20, constructing a structural similarity matrix among the dynamic brain networks in the step S10 by using a shortest path graph core, selecting a network with the most similar structure for each network, merging, constructing a network state, and repeating the operation until a set number of network states are generated; and S30, each state has feature representation, the features of the network states obtained in the step S20 are extracted, all the network states of the same sample are represented by one packet, the feature set of all the states is called as a packet feature, and a multi-instance classifier is used for obtaining a final classification result based on the packet feature.
Description
Technical Field
The invention relates to the technical field of dynamic brain network state construction, in particular to a dynamic brain network state construction method based on a graph kernel.
Background
The brain diseases seriously affect the cognitive function, consciousness, mental health and the like of the patient, and bring great pain to the patient. Brain disease assessment analysis mainly depends on clinical symptoms, and objective and reliable biomarkers are urgently needed to guide brain disease assessment analysis. At present, the construction of the dynamic brain network adopts L2 distance to measure the similarity between all connection matrixes, ignores the internal graph structure information of the brain network represented by the connection matrixes, performs dynamic brain network state construction on all sample sets, ignores the huge difference of brain activity individuals, cannot accurately depict the activity information of the brain of each sample, and cannot reflect the abnormal connection information caused by diseases.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention aims to provide a dynamic brain network state construction method based on a graph core.
In order to realize the purpose of the invention, the technical scheme is as follows: a dynamic brain network state construction method based on a graph core comprises the following steps:
s10, constructing a dynamic connection matrix of each sample through a sliding window by using an interested time sequence obtained from the resting state functional magnetic resonance data, and thinning the obtained dynamic connection matrix to obtain a coefficient dynamic connection network with a more simplified structure;
s20, constructing a structural similarity matrix between the dynamic brain networks in the step S10 by using a shortest path graph core, selecting a network with the most similar structure for each network, merging the networks, constructing a network state, and repeating the operation until a set number of network states are generated;
and S30, each state has feature representation, the features of the network states obtained in the step S20 are extracted, all the network states of the same sample are represented by one packet, the feature set of all the states is called as a packet feature, and a multi-instance classifier is used for obtaining a final classification result based on the packet feature.
As a preferred technical scheme of the invention: the step S10 includes the steps of:
s11, giving time sequence
Where K = 1.. K, K is the total number of samples, P =90 is all the numbers of interest, X i (k) ＝{x i (k) (1),x i (k) (2),...,x i (k) (N)} T Is the time sequence of the brain region i, and N is the number of all time points;
s12, if the sliding window is rectangular, the connection weight between the brain network brain area i and the brain area j in the dynamic brain network connection matrix m of the dynamic brain network connection matrix m is as follows:
wherein m is equal to [0,N-L +1 ∈ [ ]]Is the starting point of the time window, L is the length of the window, x i (n) an activity value at the nth time point of the ith brain region in the brain network, x j (n) represents an activity value at an nth time point of a jth brain region in the brain network, m' = m + L-1;
s13, thinning the obtained dynamic connection matrix and reserving each connectionConnecting information of the top 25% of the weight in the matrix, and recording the sparse connection matrix as A (k) ＝{A (k) [1],A (k) [2],...,A (k) [M]And M is the number of the dynamic brain network connection matrixes.
As a preferred technical scheme of the invention: the step S20 includes the steps of:
s21, selecting a shortest path graph core to measure the similarity of the sparse connection matrix obtained in the step S10, wherein a similarity measurement formula is as follows:
wherein s is 1 Is A [1 ]]A shortest path of(s) 2 Is A2]One shortest path of (A1)]) And S (A2)]) Are respectively A < 1 >]And A2]Set of all shortest paths in(s) 1 And s 2 Length equal delta(s) 1 ,s 2 ) Is 1, otherwise is 0, and obtains the final similarity matrix K epsilon R after solving the similarity of all the connection matrixes M×M ；
S22, according to the similarity matrix K epsilon R M×M Finding the most similar connection matrix for each dynamic connection matrix, and then performing pairwise aggregation on the matrixes in an addition aggregation mode to generate a dynamic network state
Can be expressed as:
wherein the content of the first and second substances,
is represented by A [ alpha ]]Aggregated into state with other connection matrices
And S23, repeating the steps S21 to S22 for the newly acquired network state, and continuing to acquire the similarity matrix and aggregate the similarity matrix until the set aggregation times are reached.
As a preferred technical scheme of the invention: the step S30 includes the steps of:
s31, through the steps S10 to S20, each sample is formed by a plurality of dynamic network states
The characteristics of the network states are extracted, local aggregation coefficients are selected as the characteristics of the dynamic brain networks, and the calculation formula is as follows:
wherein k is i Is a network state
Node degree of the ith brain region, and t i Is a network state
The number of triangles around the ith brain area constructs a packet feature B for each sample (k) ＝{d (k) [1],d (k) [2],...,d (k) [H]And (c) the step of (c) in which,
each packet corresponds to a label y (k) epsilon { + -1 }, wherein, -1 represents negative or not diseased, and 1 represents positive or diseased;
s32, selecting a multi-example support vector machine as a classifier to obtain a final classification result, and setting a final classification mode as a linear model: f (d) = w' phi (d), where phi is a feature map generated by an arbitrary core, then the goal of the multi-instance support vector machine is to find f that can minimize the structural risk:
where Ω can be any strictly monotonically increasing function, and l (-) is a monotonically increasing loss function, η is a regularization parameter,
the method is characterized in that the method represents the most critical example in the positive samples, namely the dynamic brain network state with the most distinguishing information, the feature vectors generated by the states and the feature vectors of the dynamic brain network state of the normal tested object are subjected to statistical analysis, namely t test, and the brain area with the P value smaller than 0.05 is taken as the distinguishing brain area, so that the diseased brain area is accurately positioned.
Compared with the prior art, the dynamic brain network state construction method based on the graph core has the following technical effects:
the dynamic brain network state construction method based on the graph core not only uses the graph core to replace the traditional matrix L2 distance, thereby being capable of better measuring the similarity between networks, but also aims at single sample design, namely constructs the dynamic brain network state for each sample, thereby being capable of better depicting the brain network dynamic activity information specific to the sample, and being capable of better guiding the learning algorithm to mine the disease related information. Finally, the brain disease identification based on the dynamic brain network is carried out under a multi-example learning framework. Under this framework, rather than simply merging the features, a packet feature can be constructed for each sample based on the constructed dynamic brain network state. The dynamic brain network state generated by the method can not only improve the identification accuracy of the brain diseases, but also effectively help to locate the lesion brain area.
Drawings
Fig. 1 is a flowchart illustrating a dynamic brain network state construction method based on a graph core according to an embodiment of the present invention;
fig. 2 is a general block diagram of a dynamic brain network state construction method based on a graph core according to an embodiment of the present invention;
FIG. 3 is a diagram illustrating a dynamic brain network state constructed according to an embodiment of the present invention;
FIG. 4 is a diagram illustrating a discriminant brain region selected according to a dynamic brain network state according to an embodiment of the present invention.
Detailed Description
The present invention will be further explained with reference to the drawings so that those skilled in the art can more deeply understand the present invention and can carry out the present invention, but the present invention will be explained below by referring to examples, which are not intended to limit the present invention.
As shown in fig. 1-4, a method for constructing a dynamic brain network state based on a graph core includes the following steps:
s10, constructing a dynamic connection matrix of each sample through a sliding window by using an interested time sequence obtained from the resting state functional magnetic resonance data, and thinning the obtained dynamic connection matrix to obtain a coefficient dynamic connection network with a more simplified structure;
s20, constructing a structural similarity matrix between the dynamic brain networks in the step S10 by using a shortest path graph core, selecting a network with the most similar structure for each network, merging the networks, constructing a network state, and repeating the operation until a set number of network states are generated;
and S30, each state has feature representation, the features of the network states obtained in the step S20 are extracted, all the network states of the same sample are represented by one packet, the feature set of all the states is called as a packet feature, and a multi-instance classifier is used for obtaining a final classification result based on the packet feature.
Step S10 includes the steps of:
s11, giving time sequence
Where K = 1.., K is the total number of samples, P =90 is all the numbers of interest, X i (k) ＝{x i (k) (1),x i (k) (2),...,x i (k) (N)} T Time series for brain region iN is the number of all time points;
s12, if the sliding window is rectangular, the connection weight between the brain area i and the brain area j of the dynamic brain network connection matrix m, namely the dynamic brain network connection matrix m, is as follows:
wherein m is in the range of 0, N-L +1]Is the starting point of the time window, L is the length of the window, x i (n) an activity value at the nth time point of the ith brain region in the brain network, x j (n) represents an activity value at an nth time point of a jth brain region in the brain network, m' = m + L-1;
s13, thinning the obtained dynamic connection matrix, reserving the connection information of the top 25% of the weight in each connection matrix, and recording the connection matrix after thinning as A (k) ＝{A (k) [1],A (k) [2],...,A (k) [M]And M is the number of the dynamic brain network connection matrixes.
Step S20 includes the steps of:
s21, selecting a shortest path graph core to measure the similarity of the sparse connection matrix obtained in the step S10, wherein a similarity measurement formula is as follows:
wherein s is 1 Is A [1 ]]A shortest path of 2 Is A2]One shortest path of (A1)]) And S (A2)]) Are respectively A < 1 >]And A2]Set of all shortest paths in(s) 1 And s 2 Length equal delta(s) 1 ,s 2 ) Is 1, otherwise is 0, after the similarity of all the connection matrixes is obtained, the final similarity matrix K belongs to R M×M ；
S22, according to the similarity matrix K epsilon R M×M Finding the most similar connection matrix for each dynamic connection matrix, and then performing pairwise aggregation on the matrixes in an addition aggregation mode to generate a dynamic networkCollateral state
Can be expressed as:
wherein the content of the first and second substances,
is represented by A [ alpha ]]Aggregated into state with other connection matrices
And S23, repeating the steps S21 to S22 for the newly acquired network state, and continuing to acquire the similarity matrix and aggregate the similarity matrix until the set aggregation times are reached.
Step S30 includes the steps of:
s31, through the steps S10 to S20, each sample is composed of a plurality of dynamic network states
The characteristics of the network states are extracted, local aggregation coefficients are selected as the characteristics of the dynamic brain networks, and the calculation formula is as follows:
wherein k is i Is a network state
Node degree of the ith brain region, and t i As the state of the network
The number of triangles around the ith brain area constructs a packet feature B for each sample (k) ＝{d (k) [1],d (k) [2],...,d (k) [H]And (c) the step of (c) in which,
each packet corresponds to a label y (k) belonging to { + -1 }, wherein, -1 represents negative or not diseased, and 1 represents positive or diseased;
s32, selecting a multi-example support vector machine as a classifier to obtain a final classification result, and setting a final classification mode as a linear model: f (d) = w' φ (d), where φ is a feature map generated by an arbitrary kernel, then the goal of the multi-instance support vector machine is to find f that can minimize the structural risk:
where Ω can be any strictly monotonically increasing function, and l (-) is a monotonically increasing loss function, η is a regularization parameter,
the method is characterized in that the method represents the most critical example in the positive samples, namely the dynamic brain network state with the most distinguishing information, the feature vectors generated by the states and the feature vectors of the dynamic brain network state of the normal tested object are subjected to statistical analysis, namely t test, and the brain area with the P value smaller than 0.05 is taken as the distinguishing brain area, so that the diseased brain area is accurately positioned.
After the concrete implementation, the method provided by the invention is compared with the following 7 methods:
1) Static network aggregation coefficient based approach (denoted CC): in order to compare the effectiveness of dynamic networks, a method based on static network aggregation coefficients is compared, the method comprises the steps of thresholding the network, extracting features by using local aggregation coefficients and classifying by using a support vector machine.
2) Method based on node time dynamics (denoted TV): in this method, the features of the dynamic network are first extracted according to the literature (Zhang J, cheng W, liu Z, et al. Neural, electrophysiologic and anatomical basis of biological change in molecular disorders [ J ]. Brain,2016,139 (8): 2307-2321), and then classified using a support vector machine.
3) Method based on space-time node dynamics (denoted STV): in this method, features of a dynamic network are extracted and sparse feature selection is performed according to literature (Jie B, liu M, shen d. Integration of temporal and spatial properties of dynamic connectivity networks [ J ] for automatic diagnosis of sparse features, 2018,47, and finally classification is performed using a multi-core support vector machine.
4) Quadratic averaging based method (denoted RMS): in this method, a feature extraction method proposed in literature (Hutchison R M, womelsdorf T, allen E A, et al. Dynamic functional connectivity: premium, iss, and interpolations [ J ]. Neuroimage,2013, 80) is used, followed by feature selection using a sparse feature selection method, and finally classification using a support vector machine.
5) Method based on L2 matrix distance (denoted L2+ MISVM): the validity of the similarity of the connection matrix is measured for verifying the provided graph core. The conventional L2 matrix distance mode is used for constructing a dynamic network state. In order to ensure the experimental fairness, the matrix keeps the connection weight information, and a weighted local aggregation coefficient is used in the dynamic network state feature extraction, and other experimental settings are consistent with the method.
6) Multiple example classifier based approach (denoted as MISVM): in order to verify the validity of the constructed dynamic brain network state, the upper triangular elements are directly extracted as features for all original dynamic connection matrixes, and then a multi-example classifier is used for classification.
7) Graph kernel and quadratic mean combination based method (denoted GK + RMS): in order to verify the effectiveness of the multi-instance classifier, firstly, a dynamic brain network state is constructed by using a graph kernel-based method, then, the state characteristics of the dynamic brain network are extracted by using a quadratic averaging method, and finally, a support vector machine is used for classification.
Compared with other 7 methods, all the methods divide training and testing sets through cross-folding cross validation, namely, samples are divided into ten data with close numbers, wherein nine data are used as training sets, and the rest data are used as testing sets. Repeating the steps for ten times, and taking an average result as a final result. The dynamic network window size constructed herein is 20 time points, and 16 dynamic connection matrices (with overlap between time windows) are constructed in total, i.e., M =16. And we have performed dynamic brain network state aggregation twice, yielding 4 dynamic brain network states, i.e. H =4.
Comparative experimental data are as follows:
the method provided by the invention improves the accuracy, specificity and sensitivity by 10%,11% and 8% respectively.
The above-mentioned embodiments, objects, technical solutions and advantages of the present invention are further described in detail, it should be understood that the above-mentioned embodiments are only illustrative of the present invention, and are not intended to limit the scope of the present invention, and any person skilled in the art should understand that equivalent changes and modifications made without departing from the concept and principle of the present invention should fall within the protection scope of the present invention.
Claims (4)
1. A dynamic brain network state construction method based on a graph core is characterized by comprising the following steps:
s10, constructing a dynamic connection matrix of each sample through a sliding window by using an interested time sequence obtained from the resting state functional magnetic resonance data, and thinning the obtained dynamic connection matrix to obtain a coefficient dynamic connection network with a more simplified structure;
s20, constructing a structural similarity matrix among the dynamic brain networks in the step S10 by using a shortest path graph core, selecting a network with the most similar structure for each network, merging, constructing a network state, and repeating the operation until a set number of network states are generated;
and S30, each state has feature representation, the features of the network states obtained in the step S20 are extracted, all the network states of the same sample are represented by one packet, the feature set of all the states is called as a packet feature, and a multi-instance classifier is used for obtaining a final classification result based on the packet feature.
2. The method for constructing a dynamic brain network state based on a kernel of a graph according to claim 1, wherein the step S10 comprises the steps of:
s11, giving time sequence
Where K = 1.. K, K is the total number of samples, P =90 is all the numbers of interest, X i (k) ＝{x i (k) (1),x i (k) (2),...,x i (k) (N)} T Is the time sequence of the brain region i, and N is the number of all time points;
s12, if the sliding window is rectangular, the connection weight between the brain area i and the brain area j of the dynamic brain network connection matrix m, namely the dynamic brain network connection matrix m, is as follows:
wherein m is in the range of 0, N-L +1]Is the starting point of the time window, L is the length of the window, x i (n) an activity value at the nth time point of the ith brain region in the brain network, x j (n) represents an activity value at an nth time point of a jth brain region in the brain network, m' = m + L-1;
s13, thinning the obtained dynamic connection matrix, reserving the connection information of the top 25% of the weight in each connection matrix, and recording the connection matrix after thinningIs A (k) ＝{A (k) [1],A (k) [2],...,A (k) [M]And M is the number of the dynamic brain network connection matrixes.
3. The method for constructing a dynamic brain network state based on a kernel of a graph according to claim 2, wherein the step S20 comprises the steps of:
s21, selecting a shortest path graph core to measure the similarity of the sparse connection matrix obtained in the step S10, wherein a similarity measurement formula is as follows:
wherein s is 1 Is A [1 ]]A shortest path of 2 Is A2]Is a shortest path of S (A1)]) And S (A2)]) Are respectively A < 1 >]And A2]Set of all shortest paths in(s) 1 And s 2 Equal length delta(s) 1 ,s 2 ) Is 1, otherwise is 0, after the similarity of all the connection matrixes is obtained, the final similarity matrix K belongs to R M×M ；
S22, according to the similarity matrix K epsilon R M×M Finding the most similar connection matrix for each dynamic connection matrix, and then performing pairwise aggregation on the matrixes in an addition aggregation mode to generate a dynamic network state C [ l ]]It can be expressed as:
wherein, H is less than M,
is represented by A [ alpha ]]Aggregated with other connection matrices into state C [ l ]]；
And S23, repeating the steps S21 to S22 for the newly acquired network state, and continuing to obtain the similarity matrix and perform aggregation until the set aggregation times are reached.
4. The method for constructing a dynamic brain network state based on a kernel of a graph according to claim 3, wherein the step S30 comprises the steps of:
s31, after the steps S10 to S20, each sample is represented by a plurality of dynamic network states C [ l ], characteristics of the network states are extracted, local aggregation coefficients are selected as the characteristics of the dynamic brain networks, and the calculation formula is as follows:
wherein k is i Is a network state C [ l ]]Node degree of the ith brain region, and t i Is a network state C [ l ]]The number of triangles around the ith brain area constructs a packet feature B for each sample (k) ＝{d (k) [1],d (k) [2],...,d (k) [H]In which d is (k) [l]＝{cc (k) (1),cc (k) (2),...,cc (k) (P) }, each packet corresponds to a label y (k) belonging to { + -1 }, wherein, -1 represents negative or not diseased, and 1 represents positive or diseased;
s32, selecting a multi-example support vector machine as a classifier to obtain a final classification result, and setting a final classification mode as a linear model: f (d) = w' φ (d), where φ is a feature map generated by an arbitrary kernel, then the goal of the multi-instance support vector machine is to find f that can minimize the structural risk:
where Ω can be any strictly monotonically increasing function, and l (-) is a monotonically increasing loss function, η is a regularization parameter,
indicating the most critical example of the positive type samples, i.e.And (3) performing statistical analysis, namely t test, on the characteristic vectors generated by the states and the characteristic vectors of the normal tested dynamic brain network state, wherein the dynamic brain network state with the most distinguishing information has the characteristic vector, and taking the brain area with the P value less than 0.05 as a distinguishing brain area so as to accurately position the diseased brain area.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211456613.7A CN115730253A (en) | 2022-11-21 | 2022-11-21 | Dynamic brain network state construction method based on graph core |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211456613.7A CN115730253A (en) | 2022-11-21 | 2022-11-21 | Dynamic brain network state construction method based on graph core |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115730253A true CN115730253A (en) | 2023-03-03 |
Family
ID=85296842
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211456613.7A Pending CN115730253A (en) | 2022-11-21 | 2022-11-21 | Dynamic brain network state construction method based on graph core |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115730253A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116898401A (en) * | 2023-07-17 | 2023-10-20 | 燕山大学 | Autism spectrum disorder subtype classification method and device based on combined recursion quantification |
-
2022
- 2022-11-21 CN CN202211456613.7A patent/CN115730253A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116898401A (en) * | 2023-07-17 | 2023-10-20 | 燕山大学 | Autism spectrum disorder subtype classification method and device based on combined recursion quantification |
| CN116898401B (en) * | 2023-07-17 | 2024-02-02 | 燕山大学 | Autism spectrum disorder subtype classification method and device based on combined recursion quantification |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110444248B (en) | Cancer biomolecule marker screening method and system based on network topology parameters | |
| CN109993230B (en) | TSK fuzzy system modeling method for brain function magnetic resonance image classification | |
| CN110840468B (en) | Autism risk assessment method and device, terminal device and storage medium | |
| CN102855491B (en) | A kind of central brain functional magnetic resonance image classification Network Based | |
| Dimitriadis et al. | Improving the reliability of network metrics in structural brain networks by integrating different network weighting strategies into a single graph | |
| CN106650818B (en) | Resting state function magnetic resonance image data classification method based on high-order hyper-network | |
| WO2002019602A2 (en) | Statistical modeling to analyze large data arrays | |
| CN112418337B (en) | Multi-feature fusion data classification method based on brain function hyper-network model | |
| CN111009321A (en) | Application method of machine learning classification model in juvenile autism auxiliary diagnosis | |
| CN111863244B (en) | Functional connection mental disease classification method and system based on sparse pooling graph convolution | |
| CN102509123A (en) | Brain functional magnetic resonance image classification method based on complex network | |
| CN113160974B (en) | Mental disease biological type mining method based on hypergraph clustering | |
| CA3221066A1 (en) | Atomic-force microscopy for identification of surfaces | |
| CN115730253A (en) | Dynamic brain network state construction method based on graph core | |
| US9117121B2 (en) | Detection of disease-related retinal nerve fiber layer thinning | |
| Hill et al. | Machine learning based white matter models with permeability: An experimental study in cuprizone treated in-vivo mouse model of axonal demyelination | |
| CN114596253A (en) | Alzheimer's disease identification method based on brain imaging genome features | |
| WO2022011855A1 (en) | False positive structural variation filtering method, storage medium, and computing device | |
| CN117036793A (en) | Brain age assessment method and device based on multi-scale features of PET (positron emission tomography) images | |
| CN105787459A (en) | ERP signal classification method based on optimal score sparse determination | |
| CN114266738A (en) | Longitudinal analysis method and system for mild brain injury magnetic resonance image data | |
| CN110223786B (en) | Method and system for predicting drug-drug interaction based on nonnegative tensor decomposition | |
| Li et al. | An efficient clustering method for medical data applications | |
| CN116741384B (en) | Bedside care-based severe acute pancreatitis clinical data management method | |
| CN116364221B (en) | Brain image data processing method and system for Alzheimer disease research |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |