CN110265148B - Dynamic functional mode learning method inspired by fMRI brain network mechanism - Google Patents

Dynamic functional mode learning method inspired by fMRI brain network mechanism Download PDF

Info

Publication number
CN110265148B
CN110265148B CN201910536898.7A CN201910536898A CN110265148B CN 110265148 B CN110265148 B CN 110265148B CN 201910536898 A CN201910536898 A CN 201910536898A CN 110265148 B CN110265148 B CN 110265148B
Authority
CN
China
Prior art keywords
brain
network
resting state
tested
dynamic
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.)
Active
Application number
CN201910536898.7A
Other languages
Chinese (zh)
Other versions
CN110265148A (en
Inventor
石玉虎
曾卫明
邓金
鲁佳
聂玮芳
李颖
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shanghai Maritime University
Original Assignee
Shanghai Maritime University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Shanghai Maritime University filed Critical Shanghai Maritime University
Priority to CN201910536898.7A priority Critical patent/CN110265148B/en
Priority to PCT/CN2019/093167 priority patent/WO2020252805A1/en
Priority to JP2021500628A priority patent/JP2022536881A/en
Publication of CN110265148A publication Critical patent/CN110265148A/en
Application granted granted Critical
Publication of CN110265148B publication Critical patent/CN110265148B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/50—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for simulation or modelling of medical disorders

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Public Health (AREA)
  • Medical Informatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Databases & Information Systems (AREA)
  • Pathology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Epidemiology (AREA)
  • Primary Health Care (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a dynamic functional mode learning method based on fMRI brain network mechanism inspiration, which comprises the following steps: collecting a plurality of tested resting state fMRI brain image data; preprocessing the resting state fMRI brain image data; according to the preprocessed resting state fMRI brain image data, a group level resting state brain function network and an individual level resting state brain function network and a corresponding time sequence are respectively obtained by a GICA-IR method; calculating a dynamic function connection matrix between the resting brain function networks corresponding to each tested object by using a sliding time window method, and opening the upper triangular elements of the dynamic function connection matrix into dynamic function connection vectors so as to obtain a dynamic function connection vector set corresponding to all tested objects; and extracting the inherent dynamic function connection mode of the brain hidden in the dynamic function connection vector set by utilizing a deep neural network model and an affine propagation clustering algorithm. The invention provides a solid foundation for disclosing the basic principle of brain cognitive activities, the damaged mechanism of cranial nerve diseases, exploring occupational brain plasticity recombination characteristics and the like.

Description

Dynamic functional mode learning method inspired by fMRI brain network mechanism
Technical Field
The invention belongs to the technical field of brain imaging image processing, and particularly relates to a dynamic functional mode learning method based on fMRI brain network mechanism inspiration.
Background
The brain is the most important organ of the human body, controls various cognitive behaviors of human beings such as thinking, consciousness, emotion, memory and the like, is the central nervous system for realizing high-level cognitive function activities of the human beings, and is one of the most complex and precise systems known to the human beings so far. How to know the brain and explore the cognitive mechanism of the brain nerve activity is an important field for the domestic and foreign scientific community to make breakthroughs and exploration, and has very important research value.
In recent years, with the continuous development of scientific technology, brain function imaging technology has become one of the most focused research hotspots and frontiers in the field of brain science. Among them, the functional Magnetic Resonance Imaging (fMRI) combines the information of three aspects of function, anatomy and image, not only can display the position, size and range of the brain function activation region, but also can directly display the exact anatomical position of the activation region, and has many superior characteristics of non-invasion, non-trauma, non-radiation, repeatable, accurate positioning, higher time and space resolution, etc., thus being widely applied to brain science research in various fields.
fMRI-based brain function network connectivity analysis is an important aspect of brain science research using this technology. Because the brain is a highly complex system which is always in dynamic change, the quantitative description of the dynamic change of the brain function connection mode along with the time is more in line with the essence of brain nerve activity, and richer information can be provided for exploring the basic attribute of a brain network. Therefore, the dynamic brain function network connection analysis method is perfected and developed, the dynamic characteristics of the brain are researched through the dynamic function connection mode among fMRI brain networks, the significance is achieved, the internal information hidden in fMRI signals can be effectively mined, the brain nerve activity rule can be better revealed, and the fMRI technology is promoted to play an important role. However, the current dynamic brain function network connection analysis method has several key problems in the aspects of brain network construction, dynamic brain function mode extraction and the like, and needs to be further solved systematically.
Based on the above, the present invention is directed to analyzing a global brain dynamic function connection pattern of a living human brain by fully utilizing the attribute characteristics between brain network mechanisms and adopting a group independent component analysis with intrinsic reference signal (GICA-IR) and sliding time window analysis based on a data intrinsic reference signal, and methods such as a deep neural network model and an affine propagation clustering algorithm, etc., under the condition of obtaining brain image big data by using an fMRI neural imaging technology. Thereby providing a solid foundation for revealing basic principles of brain cognitive activities (such as brain development and brain aging), impaired mechanisms of cranial nerve diseases (such as autism, attention deficit hyperactivity disorder, depression, schizophrenia, Alzheimer's disease and Parkinson's disease), exploring occupational brain plasticity reorganization characteristics (such as mariners and pilots) and the like.
Disclosure of Invention
The invention aims to provide a dynamic function connection mode learning method inspired by an fMRI brain network mechanism, which comprises the steps of extracting interesting group level and individual level resting state brain function networks in fMRI brain image data and corresponding time sequences thereof by a GICA-IR method, then calculating a dynamic function connection matrix corresponding to the brain networks in each group of data by adopting a sliding time window method, and finally performing cluster analysis on a dynamic function connection vector set by using a deep neural network model and an affine propagation clustering algorithm to extract an inherent dynamic brain function connection mode in a human brain.
In order to achieve the purpose, the invention is realized by the following technical scheme:
a dynamic functional mode learning method based on fMRI brain network mechanism inspiration comprises the following steps:
s1, collecting a plurality of tested resting state fMRI brain image data;
s2, preprocessing the resting state fMRI brain image data acquired in the step S1 to obtain preprocessed resting state fMRI brain image data;
s3, according to the preprocessed resting state fMRI brain image data in the step S2, a group-level resting state brain function network and a corresponding time sequence thereof and an individual-level resting state brain function network and a corresponding time sequence thereof are respectively obtained by a GICA-IR method;
s4, calculating a dynamic function connection matrix between the resting brain function networks corresponding to each tested object by using a sliding time window method, and expanding the upper triangular elements into dynamic function connection vectors so as to obtain a dynamic function connection vector set corresponding to all tested objects;
and S5, extracting the inherent dynamic function connection mode of the brain hidden in the dynamic function connection vector set by utilizing a deep neural network model and an affine propagation clustering algorithm.
Preferably, in step S1, the subject is asked to keep the brain awake during the data acquisition process, and lies down in the magnetic resonance apparatus.
Preferably, in the step S2, the preprocessing operation includes temporal layer correction, head movement correction, spatial normalization, spatial smoothing, filtering, linear drift elimination, and covariate regression.
Preferably, all preprocessing operations are done by DPARSF software.
Preferably, in step S3, the individual-level resting brain function networks and the corresponding time series are obtained by spatio-temporal bi-regression based on the group-level resting brain function networks and the corresponding time series.
Preferably, the GICA-IR method in step S3 further comprises the following processes:
based on the preprocessed resting state fMRI brain image data, each tested object is independently subjected to ICA analysis, and meanwhile, a principal component analysis method is utilized to extract implicit intrinsic reference signals from a matrix formed by the tested independent components corresponding to the interested brain function network;
guiding the ICA to analyze fMRI brain image data on the group level by using the true reference signal, and calculating to obtain a resting state brain function network of the group level and a time sequence thereof;
and obtaining the individual level resting state brain function network and the time sequence thereof corresponding to each tested object in the group in a space-time double regression mode according to the group level resting state brain function network and the time sequence thereof obtained by calculation.
Preferably, the dynamic functional mode learning method based on the fMRI brain network mechanism inspiration,
setting and collecting K tested resting state fMRI brain image data, wherein each tested fMRI data comprises T time points and V voxels after being subjected to preprocessing operation;
the GICA-IR method in step S3 includes the following steps:
s3.1, carrying out ICA analysis on each test independently, and for the test i, ICA decomposition can be expressed as:
Xi＝MiSi,(i＝1,2,…,K) (1)
wherein, XiA fMRI observation data matrix representing T multiplied by V order;
Figure BDA0002101408170000031
represents NiX V order source signal matrix, NiRepresenting the number of independent components obtained by the tested i through group horizontal ICA decomposition, wherein each row represents an independent component obtained by the tested i in the ICA decomposition process;
Figure BDA0002101408170000032
is a column vector of size V × 1; miRepresents T × NiA mixing matrix of order, each column representing
Figure BDA0002101408170000033
The time sequence corresponding to each row in the time sequence;
order to
Figure BDA0002101408170000041
N-th of the tested i corresponding to the interested brain function networkiEach independent component, n is not less than 1i≤Ni(ii) a The component correspondence between different tested objects is measured by spatial correlation, and then a principal component analysis method is used for measuring the correspondence
Figure BDA0002101408170000042
The implicit true reference signal is extracted from the matrix R with the size of K × V, that is:
Figure BDA0002101408170000043
and according to r ═ e1' R obtains a first principal component R, which is the required implicit true reference signal, where e1Representing the eigenvector corresponding to the maximum eigenvalue;
s3.2, guiding the fMRI brain image data analysis of ICA on a group level by using the true reference signal obtained in the step S3.1, and performing group analysis by adopting a time cascade mode, wherein the group analysis comprises the following steps:
(X1；X2；…；XK)＝MS (3)
wherein S ═ S1,s2,…,sN) ' represents a source signal matrix of NxV order, each row represents a group of independent components, namely a group horizontal resting state brain function network, and N is the number of the independent components obtained by group horizontal ICA decomposition; m represents a group mixing matrix of KT × V orders, and each column represents S ═ S1,s2,…,sN) ' each row corresponds to a time sequence;
further solving the formula (3) in a constraint optimization mode
Maximization: j(s)i)＝{E[G(si)]-E[G(v)]}2 (4-1)
Constrained to: g(s)i)＝ε(si,r)-ξ≤0,h(si)＝E[si]2-1＝0 (4-2)
Wherein s isiRepresenting the found output signal; j(s)i) A comparison function representing the independence of the measurement output signals; g (-) is a non-quadratic function, v is a Gaussian random variable; r denotes the true reference signal, ε(s)iR) is a distance measure; ξ is a threshold parameter; h(s)i) The method is to ensure equality constraint of solving an optimization problem in a convex region;
s3.3, obtaining the individual horizontal resting state brain function network and the time sequence thereof corresponding to each tested object in the group in a space-time dual regression mode according to the group horizontal resting state brain function network and the time sequence thereof obtained by calculation in the step S3.2, and obtaining:
Mi＝Xi·pinv(S),Si＝pinv(Mi)·Xi,(i＝1,2,…,K) (5)
wherein S isiRepresenting a source signal corresponding to the tested i; pinv (S) refers to solving the pseudo-inverse matrix of S; miRepresenting a corresponding time series; k represents the number of test subjects.
Preferably, in step S3, according to the preprocessed resting state fMRI brain image data, nine resting state brain function networks of interest and their corresponding time sequences are calculated by using the GICA-IR, where the nine resting state brain function networks include a default network, a visual network, a bilateral visual network, an auditory network, a sensory-motor network, an execution control network, a highlight network, a working memory network, and an attention network, and the spatiotemporal information of each tested brain network in the group is obtained by a spatiotemporal dual regression manner.
Preferably, the step S4 further includes the steps of:
s4.1, for each tested object, by adopting a sliding time window method, utilizing a time window with a specific width W and taking the step size as 1 to perform time sequence T1,T2,…,TNUp-sliding, the time series of the nth brain network under the ith time window is represented as:
Figure BDA0002101408170000051
calculating the Pearson correlation coefficient between every two tested corresponding brain network time sequences under each time window to obtain T-W +1 dynamic function connection matrixes dFCMi(i is more than or equal to 1 and less than or equal to T-W +1), and the dynamic function connection matrix group formed by the dynamic function connection matrices is as follows:
dFCMG＝{dFCM1,dFCM2,…,dFCMT-W+1}； (7)
in step S4.1, the Pearson correlation coefficient between every two brain network time sequences specifically refers to the time sequence wT corresponding to the x and y brain networks under the ith time windowi xAnd wTi yThe Pearson correlation coefficient between them is expressed as follows
Figure BDA0002101408170000061
Wherein, E (wT)i x) And E (wT)i y) Respectively represent wTi xAnd wTi yExpectation of (2), D (wT)i x) And D (wT)i y) Respectively represent wTi xAnd wTi yI is more than or equal to 1 and less than or equal to T-W +1, x is more than or equal to 1 and less than or equal to N, and y is more than or equal to 1 and less than or equal to N;
thus, the corresponding dynamic function connection matrix dFCM under the ith time windowi(1. ltoreq. i. ltoreq. T-W +1) is represented by:
Figure BDA0002101408170000062
s4.2, for each tested object, calculating a dynamic function connection vector set dFCVS, wherein the specific method is as follows:
dynamic function connection matrix dFCM in dynamic function connection matrix group dFCMG by rowiThe upper triangular elements (i is more than or equal to 1 and less than or equal to T-W +1) are stretched into a line to obtain a dynamic function connection vector dFCVi(i is more than or equal to 1 and less than or equal to T-W + 1); each vector has a size of (N × (N-1)2) × 1; T-W +1 dynamic function connection vectors are cascaded according to the time window sequence to form a dynamic function connection vector set dFCVS ═ { dFCV-1,dFCV2,…,dFCVT-W+1(ii) }, size (N × (N-1)2) × (T-W + 1); wherein, the dynamic function connection vector dFCV under the ith time windowiCan be expressed as:
Figure BDA0002101408170000063
the step S5 further includes the following steps:
s5.1, combining all tested dynamic function connection vector sets in columns to form deep neural network learning samples, and performing feature extraction by adopting a convolutional neural network model, wherein each sample represents a dynamic function connection vector corresponding to a tested object;
and S5.2, carrying out cluster analysis on the characteristic samples obtained by the deep neural network model learning in the step S5.1 by adopting an affine propagation clustering algorithm to obtain Q categories, wherein the clustering center corresponding to each category is the corresponding dynamic function connection mode hidden in the fMRI brain image data.
Compared with the prior art, the invention has the beneficial effects that:
(1) the invention introduces the brain network mechanism attribute characteristics to carry out dynamic brain function connection analysis, which is beneficial to acquiring the implicit dynamic function connection mode according to the dynamic characteristics of the brain network;
(2) according to the method, the accuracy and the efficiency of clustering are improved by combining a GICA-IR method, sliding time window analysis, a deep neural network model, an affine propagation clustering algorithm and the like;
(3) the dynamic function connection mode obtained by the invention can provide analysis basis for further research on the basic principle of brain cognitive activities, the nerve damage mechanism of brain diseases and the biological indicators of occupational brain plasticity.
Drawings
FIG. 1 is a flow chart of a dynamic functional connectivity pattern learning method inspired by the fMRI brain network mechanism of the present invention;
fig. 2 is a schematic diagram of an overall implementation of the dynamic functional connection mode learning method inspired by the fMRI brain network mechanism of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention.
As shown in fig. 1-2, the present invention provides a dynamic functional mode learning method inspired by fMRI brain network mechanism, which comprises the following steps:
s1, acquiring a plurality (for example 100) of normal healthy subject resting fMRI brain image data.
In step S1, the subject is required to keep the brain awake during data acquisition and lies in the magnetic resonance apparatus; the number of time points for each fMRI data acquisition tested was 215.
S2, preprocessing the resting state fMRI brain image data of the normal and healthy subject acquired in the step S1 to obtain preprocessed resting state fMRI brain image data of the normal and healthy subject; the preprocessing operation comprises seven steps of time layer correction, head movement correction, space standardization, space smoothing, filtering, linear drift removal and covariate regression. All preprocessing operations are performed by DPARSF software (software for brain function imaging studies).
S3, according to the preprocessed normal healthy tested resting state fMRI brain image data in the step S2, a group level resting state brain function network and a corresponding time sequence thereof, and an individual level resting state brain function network and a corresponding time sequence thereof are respectively obtained by using a GICA-IR method; wherein, the individual level information (the resting brain function network of the individual level and the corresponding time series) is obtained by a space-time dual regression mode.
S4, calculating the dynamic function connection matrix between the resting brain function networks corresponding to each tested object by using a sliding time window method, and expanding the triangular elements on the dynamic function connection matrix into dynamic function connection vectors so as to obtain a dynamic function connection vector set corresponding to all tested objects.
And S5, extracting the inherent dynamic function connection mode of the brain hidden in the dynamic function connection vector set by using a deep neural network model and an affine propagation clustering algorithm.
The GICA-IR method in step S3 includes the following steps:
and S3.1, assuming that K tested resting-state fMRI brain image data are contained, and each tested fMRI data contains T time points and V voxels after being subjected to preprocessing operation. First, Independent Component Analysis (ICA) is performed for each test individually, and for test i, the ICA can be expressed as:
Xi＝MiSi,(i＝1,2,…,K) (1)
wherein, XiA fMRI observation data matrix representing T multiplied by V order;
Figure BDA0002101408170000081
represents NiX V order source signal matrix, NiRepresenting the number of independent components obtained by the tested i through group horizontal ICA decomposition, wherein each row represents an independent component obtained by the tested i in the ICA decomposition process;
Figure BDA0002101408170000082
is a column vector of size V × 1; miRepresents T × NiA mixing matrix of order, each column representing
Figure BDA0002101408170000083
Corresponding to each row in the time series.
Then, let
Figure BDA0002101408170000084
Represents the nth of the tested i corresponding to the brain function network of interest (i.e. the brain function network of interest to be tested)i(1≤ni≤Ni) An independent component, niIndicates the nth corresponding to the tested iiAn independent component having a value in the range of 1 to NiTo (c) to (d);
when analyzing fMRI data by the ICA method, the output result from the ICA decomposition generally includes many independent components (here, how many N are generally considered to be set or estimated according to some specific method such as the minimum description length MDL), but not every component is meaningful, so that only the meaningful component is generally selected for subsequent analysis, and each independent component in the ICA output result corresponds to one brain function network, so that the corresponding independent component is generally determined by selecting the brain function network of interest;
the correspondence of the independent components among different tested subjects can be measured through spatial correlation, wherein the component correspondence refers to the correspondence among the independent components obtained by the decomposition of the ICA of different tested subjects, because the decomposition result of the ICA is unordered, for a given brain function network, the corresponding independent components are different in different tested subjects, so that the correlation among the independent components can be calculated for determining the corresponding component serial numbers of the same brain function network in different tested subjects, and therefore, as long as the component serial number of one tested corresponding brain function network is known, the independent component corresponding to the brain function network in other tested independent components (namely, the independent component with the largest corresponding correlation) can be automatically determined according to the correlation coefficient;
then, using principal component analysis method
Figure BDA0002101408170000091
The implicit true reference signal is extracted from the matrix R with the size of K × V, that is:
Figure BDA0002101408170000092
further, according to r ═ e1' R, the first principal component R (i.e., the implicit, true reference signal required) can be obtained, where e1And representing the feature vector corresponding to the maximum feature value.
S3.2, using the true reference signal obtained in step S3.1 to guide the analysis of fMRI brain image data at the group level by ICA, where the group analysis (i.e. the analysis of fMRI brain image data at the group level) is performed in a time cascade manner as follows:
(X1；X2；…；XK)＝MS (3)
wherein S ═ S1,s2,…,sN) ' a source signal matrix of order N × V, each row representing a group of independent components, i.e., a group-level resting state brain function network, N is the number of source signals obtained by group ICA analysis (i.e., formula 3), i.e., the number of independent components obtained by group-level ICA decomposition; m represents a group mixing matrix of KT × V orders, and each column represents S ═ S1,s2,…,sN) ' each row corresponds to a time series.
Further, the model (i.e. formula 3) is solved by constraint optimization
Maximization: j(s)i)＝{E[G(si)]-E[G(v)]}2 (4-1)
Constrained to: g(s)i)＝ε(si,r)-ξ≤0,h(si)＝E[si]2-1＝0 (4-2)
Wherein s isiRepresenting the output signal by a separation vector wiAnd the observed signal X is estimated, i.e. si＝wiX；J(si) A comparison function representing the independence of the measurement output signals; g (-) is a non-quadratic function, v is a Gaussian random variable; r denotes the true reference signal, ε(s)iR) is a distance measure; ξ is a threshold parameter that limits the output signal to be the only signal that satisfies the inequality constraint; constraint of equation h(s)i) Solving the optimization problem in a convex area; then, the separation vector w can be obtained by adopting an augmented Lagrange method and combining algorithms such as Newton iteration or fast fixed point iterationiTo obtain the required output signal si。
S3.3, obtaining the individual horizontal resting state brain function network and the time sequence thereof corresponding to each tested object in the group in a space-time dual regression mode according to the group horizontal resting state brain function network and the time sequence thereof obtained by calculation in the step S3.2, namely:
Mi＝Xi·pinv(S),Si＝pinv(Mi)·Xi,(i＝1,2,…,K) (5)
wherein S isiRepresenting a source signal corresponding to the tested i; pinv (S) refers to solving the pseudo-inverse matrix of S; miRepresenting a corresponding time series; k represents the number of test subjects.
The step S4 further includes the following steps:
s4.1, for each tested object, by adopting a sliding time window method, utilizing a time window with a specific width W and taking the step size as 1 to perform time sequence T1,T2,……,TNUp sliding, under the ith time windowThe time series of n brain networks is represented as:
Figure BDA0002101408170000101
calculating the Pearson correlation coefficient between every two tested corresponding brain network time sequences under each time window to obtain T-W +1 dynamic function connection matrixes dFCMi(i is more than or equal to 1 and less than or equal to T-W +1), and the dynamic function connection matrix group formed by the dynamic function connection matrices is as follows:
dFCMG＝{dFCM1,dFCM2,…,dFCMT-W+1} (7)。
in step S4.1, the pearson correlation coefficient between every two brain network time sequences herein specifically refers to the time sequence wT corresponding to the x and y brain networks under the ith time windowi xAnd wTi yThe Pearson correlation coefficient between them is expressed as follows
Figure BDA0002101408170000111
Wherein, E (wT)i x) And E (wT)i y) Respectively represent wTi xAnd wTi yExpectation of (2), D (wT)i x) And D (wT)i y) Respectively represent wTi xAnd wTi yI is more than or equal to 1 and less than or equal to T-W +1, x is more than or equal to 1 and less than or equal to N, and y is more than or equal to 1 and less than or equal to N.
Thus, the corresponding dynamic function connection matrix dFCM under the ith time windowi(1. ltoreq. i.ltoreq.T-W +1) can be represented as:
Figure BDA0002101408170000112
s4.2, for each tested object, calculating a dynamic function connection vector set dFCVS, wherein the specific method is as follows:
dynamic function connection matrix in a group of dynamic function connection matrices dFCMG by row
dFCMiThe upper triangular elements (i is more than or equal to 1 and less than or equal to T-W +1) are stretched into a line to obtain a dynamic function connection vector dFCVi(i is more than or equal to 1 and less than or equal to T-W + 1); each vector has a size of (N × (N-1)2) × 1; T-W +1 dynamic function connection vectors are cascaded according to the time window sequence to form a dynamic function connection vector set dFCVS ═ { dFCV-1,dFCV2,…,dFCVT-W+1(ii) }, size (N × (N-1)2) × (T-W + 1). Wherein, the dynamic function connection vector dFCV under the ith time windowiCan be expressed as:
Figure BDA0002101408170000121
the step S5 further includes the following steps:
and S5.1, combining all tested dynamic function connection vector sets according to columns to form a deep neural network learning sample, and extracting features by adopting a convolutional neural network model. Wherein each sample represents a dynamic function connection vector corresponding to the tested sample.
And S5.2, carrying out cluster analysis on the characteristic samples obtained by the deep neural network model learning in the step S5.1 by adopting an affine propagation clustering algorithm to obtain Q categories, wherein the clustering center corresponding to each category is the corresponding dynamic function connection mode hidden in the fMRI data.
For ease of understanding, a specific example is given for the above steps, as follows:
in step 1, 100 normal healthy subjects' resting state fMRI brain image data are collected. The number of time points for acquiring each fMRI data under test (also referred to as time series length, which is how long the time series in the analysis is corresponding to how many time points of the fMRI data are acquired) is 215.
And step 2, preprocessing the acquired resting state fMRI brain image data.
In step 3, according to the preprocessed resting state fMRI data corresponding to each tested object, a dynamic function connection vector set of each tested object is obtained by using a GICA-IR and sliding time window method; the method specifically comprises the following steps:
for each test, all interesting brain networks and their corresponding time series are calculated as follows: according to the preprocessed resting state fMRI data, calculating nine resting state brain function networks of interest in the preprocessed resting state fMRI data by using GICA-IR and corresponding time sequences T1,T2,…,T9The brain network space-time information acquisition method comprises a default network, a visual network, a bilateral visual network, an auditory network, a sensory-motor network, an execution control network, a highlight network, a working memory network and an attention network, and the brain network space-time information corresponding to each tested object in the group is obtained through a space-time dual regression mode. For fMRI data with a time series length T215, the time series
Figure BDA0002101408170000131
Wherein
Figure BDA0002101408170000132
Representing the signal strength value of the nth brain network at time i.
In step 4, for each tested object, calculating a dynamic function connection matrix set dFCMG, specifically the method is as follows:
by using a sliding time window method, a time sequence T is processed by a time window with a specific width W equal to 20 and a step size of 11,T2,……,T9Up-sliding, the time series of the nth brain network under the ith time window is represented as
Figure BDA0002101408170000133
Calculating the Pearson correlation coefficient between every two tested corresponding brain network time sequences to obtain 196 dynamic function connection matrixes dFCMi(1 ≦ i ≦ 196), and the set of dynamic function connection matrices formed by these dynamic function connection matrices is dFCMG ═ dFCM1,dFCM2,…,dFCM196}。
The Pearson correlation coefficient between every two brain network time sequences specifically refers to the time sequence wT corresponding to the x and y brain networks under the ith time windowi xAnd wTi yThe correlation coefficient of the pearson between,the formula is as follows
Figure BDA0002101408170000134
Wherein, E (wT)i x) And E (wT)i y) Respectively represent wTi xAnd wTi yExpectation of (2), D (wT)i x) And D (wT)i y) Respectively represent wTi xAnd wTi yI is more than or equal to 1 and less than or equal to 196, x is more than or equal to 1 and less than or equal to 9, and y is more than or equal to 1 and less than or equal to 9. So that the corresponding dynamic functional connection matrix dFCM is present in the ith time windowi(1. ltoreq. i. ltoreq.196) can be expressed as:
Figure BDA0002101408170000135
for each tested object, calculating a dynamic function connection vector set dFCVS by the following specific method:
dynamic function connection matrix dFCM in dynamic function connection matrix group dFCMG by rowiThe upper triangular elements (i is more than or equal to 1 and less than or equal to 196) are stretched into a line to obtain a dynamic function connection vector dFCVi(i is more than or equal to 1 and less than or equal to 196), and the size of each vector is 36 multiplied by 1; 196 dynamic function connection vectors are cascaded according to the time window sequence to form a dynamic function connection vector set dFCVS ═ { dFCV-1,dFCV2,…,dFCV196Size 36 × 196. Wherein dFCViThe dynamic function connection vector under the ith time window is specifically expressed as:
Figure BDA0002101408170000141
in step 5, a deep neural network model and an affine propagation clustering algorithm are used for carrying out clustering analysis on all tested dynamic function connection vector sets, and a dynamic function connection mode implicit in the fMRI data is obtained.
In summary, the present invention is directed to analyzing a global brain dynamic function connection pattern of a living human brain by fully utilizing the attribute characteristics between brain network mechanisms and adopting methods such as group independent component analysis (GICA-IR) and sliding time window analysis based on a data true reference signal, a deep neural network model, and an affine propagation clustering algorithm when obtaining brain image big data by using an fMRI neural imaging technique. Thereby providing a solid foundation for revealing basic principles of brain cognitive activities (such as brain development and brain aging), impaired mechanisms of cranial nerve diseases (such as autism, attention deficit hyperactivity disorder, depression, schizophrenia, Alzheimer's disease and Parkinson's disease), exploring occupational brain plasticity reorganization characteristics (such as mariners and pilots) and the like.
While the present invention has been described in detail with reference to the preferred embodiments, it should be understood that the above description should not be taken as limiting the invention. Various modifications and alterations to this invention will become apparent to those skilled in the art upon reading the foregoing description. Accordingly, the scope of the invention should be determined from the following claims.

Claims (6)

1. A dynamic functional mode learning method based on fMRI brain network mechanism inspiration is characterized by comprising the following steps:
s1, collecting a plurality of tested resting state fMRI brain image data;
s2, preprocessing the resting state fMRI brain image data acquired in the step S1 to obtain preprocessed resting state fMRI brain image data;
s3, according to the preprocessed resting state fMRI brain image data in the step S2, a resting state brain function network at a group level and a time sequence corresponding to the resting state brain function network at an individual level are respectively obtained by adopting a group independent component analysis method based on a data true reference signal; wherein the individual level resting state brain function network and the corresponding time series thereof are obtained by a spatio-temporal bi-regression mode based on the group level resting state brain function network and the corresponding time series thereof;
the group independent component analysis method based on the true reference signal of the data book comprises the following processes:
based on the preprocessed resting state fMRI brain image data, each tested object is independently subjected to ICA analysis, and meanwhile, a principal component analysis method is utilized to extract implicit intrinsic reference signals from a matrix formed by the tested independent components corresponding to the interested brain function network;
guiding the ICA to analyze fMRI brain image data on the group level by using the true reference signal, and calculating to obtain a resting state brain function network of the group level and a time sequence thereof;
according to the calculated group level resting state brain function network and the time sequence thereof, obtaining an individual level resting state brain function network and the time sequence thereof corresponding to each tested object in the group through a space-time dual regression mode;
setting and collecting K tested resting state fMRI brain image data, wherein each tested fMRI data comprises T time points and V voxels after being subjected to preprocessing operation;
the group independent component analysis method based on the data true reference signal comprises the following steps:
s3.1, carrying out ICA analysis on each test sample independently, and for the test sample i, decomposing the ICA into the following components:
Xi＝MiSi,(i＝1,2,…,K) (1)
wherein, XiA fMRI observation data matrix representing T multiplied by V order;
Figure FDA0003019788010000021
represents NiX V order source signal matrix, NiRepresenting the number of independent components obtained by the tested i through group horizontal ICA decomposition, wherein each row represents an independent component obtained by the tested i in the ICA decomposition process;
Figure FDA0003019788010000022
is a column vector of size V × 1; miRepresents T × NiA mixing matrix of order, each column representing
Figure FDA0003019788010000023
The time sequence corresponding to each row in the time sequence;
order to
Figure FDA0003019788010000024
N-th of the tested i corresponding to the interested brain function networkiEach independent component, n is not less than 1i≤Ni(ii) a The component correspondence between different tested objects is measured by spatial correlation, and then a principal component analysis method is used for measuring the correspondence
Figure FDA0003019788010000025
The implicit true reference signal is extracted from the matrix R with the size of K × V, that is:
Figure FDA0003019788010000026
further, r ═ e'1R obtains a first principal component R, which is the required implicit true reference signal, wherein e1Representing the eigenvector corresponding to the maximum eigenvalue;
s3.2, guiding the fMRI brain image data analysis of ICA on a group level by using the true reference signal obtained in the step S3.1, and performing group analysis by adopting a time cascade mode, wherein the group analysis comprises the following steps:
(X1；X2；…；XK) MS (3) wherein S ═ S (S)1,s2,…,sN) ' represents a source signal matrix of NxV order, each row represents a group of independent components, namely a group horizontal resting state brain function network, and N is the number of the independent components obtained by group horizontal ICA decomposition; m represents a group mixing matrix of KT × V orders, and each column represents S ═ S1,s2,…,sN) ' each row corresponds to a time sequence;
solving formula (3) by means of constraint optimization
Maximization: j(s)i)＝{E[G(si)]-E[G(v)]}2 (4-1)
Constrained to: g(s)i)＝ε(si,r)-ξ≤0,h(si)＝E[si]2-1＝0 (4-2)
Wherein s isiRepresenting the found output signal; j(s)i) A comparison function representing the independence of the measurement output signals; g (-) is a non-quadratic function, v is a Gaussian random variable; r denotes the true reference signal, ε(s)iR) is a distance measure; ξ is a threshold parameter; h(s)i) The method is to ensure equality constraint of solving an optimization problem in a convex region;
s3.3, obtaining the individual horizontal resting state brain function network and the time sequence thereof corresponding to each tested object in the group in a space-time dual regression mode according to the group horizontal resting state brain function network and the time sequence thereof obtained by calculation in the step S3.2, and obtaining:
Mi＝Xi·pinv(S),Si＝pinv(Mi)·Xi,(i＝1,2,…,K) (5)
wherein S isiRepresenting a source signal corresponding to the tested i; pinv (S) refers to solving the pseudo-inverse matrix of S; miRepresenting a corresponding time series; k represents the number of the tested samples;
s4, calculating a dynamic function connection matrix between the resting brain function networks corresponding to each tested object by using a sliding time window method, and expanding the upper triangular elements into dynamic function connection vectors so as to obtain a dynamic function connection vector set corresponding to all tested objects;
step S4, including the steps of:
s4.1, for each tested object, by adopting a sliding time window method, utilizing a time window with a specific width W and taking the step size as 1 to perform time sequence T1,T2,……,TQUp-sliding, the time series of the qth brain network in the jth time window is represented as:
Figure FDA0003019788010000031
calculating the Pearson correlation coefficient between every two tested corresponding brain network time sequences under each time window to obtain T-W +1 dynamic function connection matrixes dFCMj(j is more than or equal to 1 and less than or equal to T-W +1), and the dynamic function connection matrix group formed by the dynamic function connection matrices is as follows:
dFCMG＝{dFCM1,dFCM2,…,dFCMT-W+1}； (7)
in step S4.1, the Pearson correlation coefficient between every two brain network time sequences specifically refers to the time sequence corresponding to the x and y brain networks under the jth time window
Figure FDA0003019788010000041
And
Figure FDA0003019788010000042
the Pearson correlation coefficient between them is expressed as follows
Figure FDA0003019788010000043
Wherein the content of the first and second substances,
Figure FDA0003019788010000044
and
Figure FDA0003019788010000045
respectively represent
Figure FDA0003019788010000046
And
Figure FDA0003019788010000047
in the expectation that the position of the target is not changed,
Figure FDA0003019788010000048
and
Figure FDA0003019788010000049
respectively represent
Figure FDA00030197880100000410
And
Figure FDA00030197880100000411
j is more than or equal to 1 and less than or equal to T-W +1, x is more than or equal to 1 and less than or equal to Q, and y is more than or equal to 1 and less than or equal to Q;
corresponding dynamic function connection matrix dFCM under jth time windowj(1. ltoreq. j. ltoreq. T-W +1) is represented by:
Figure FDA00030197880100000412
s4.2, for each tested object, calculating a dynamic function connection vector set dFCVS, wherein the specific method is as follows:
dynamic function connection matrix dFCM in dynamic function connection matrix group dFCMG by rowjThe upper triangular elements (j is more than or equal to 1 and less than or equal to T-W +1) are stretched into a line to obtain a dynamic function connection vector dFCVj(j is more than or equal to 1 and less than or equal to T-W + 1); each vector has a size of (Qx (Q-1)/2) x 1;
T-W +1 dynamic function connection vectors are cascaded according to the time window sequence to form a dynamic function connection vector set dFCVS ═ { dFCV-1,dFCV2,…,dFCVT-W+1(ii) }, size (Q × (Q-1)/2) × (T-W + 1);
wherein, the dynamic function connection vector dFCV under the ith time windowiExpressed as:
Figure FDA00030197880100000413
and S5, extracting the inherent dynamic function connection mode of the brain hidden in the dynamic function connection vector set by utilizing a deep neural network model and an affine propagation clustering algorithm.
2. The dynamic functional mode learning method based on fMRI brain network mechanism heuristics of claim 1,
in step S1, the subject is kept awake during data acquisition and lies down in the mri apparatus.
3. The dynamic functional mode learning method based on fMRI brain network mechanism heuristics of claim 1,
in step S2, the preprocessing operation includes temporal layer correction, head motion correction, spatial normalization, spatial smoothing, filtering, linear drift elimination, and covariate regression.
4. The method for learning a dynamic functional pattern based on the fMRI brain network mechanism heuristic according to claim 1 or 3,
all preprocessing operations are done by DPARSF software.
5. The dynamic functional mode learning method based on fMRI brain network mechanism heuristics of claim 1,
in step S3, according to the preprocessed resting state fMRI brain image data, nine resting state brain function networks of interest and time sequences corresponding thereto are calculated by using the GICA-IR, where the nine resting state brain function networks include a default network, a visual network, a bilateral visual network, an auditory network, a sensory-motor network, an execution control network, a highlight network, a working memory network, and an attention network, and the spatiotemporal information of each brain network corresponding to each test in the group is obtained in a spatiotemporal dual regression manner.
6. The dynamic functional mode learning method based on fMRI brain network mechanism heuristics of claim 1,
step S5 includes the following steps:
s5.1, combining all tested dynamic function connection vector sets in columns to form deep neural network learning samples, and performing feature extraction by adopting a convolutional neural network model, wherein each sample represents a corresponding dynamic function connection vector to be tested;
and S5.2, carrying out cluster analysis on the characteristic samples obtained by the deep neural network model learning in the step S5.1 by adopting an affine propagation clustering algorithm to obtain Q categories, wherein the corresponding clustering center of each category is a corresponding dynamic function connection mode hidden in the fMRI brain image data.
CN201910536898.7A 2019-06-20 2019-06-20 Dynamic functional mode learning method inspired by fMRI brain network mechanism Active CN110265148B (en)

Priority Applications (3)

Application Number Priority Date Filing Date Title
CN201910536898.7A CN110265148B (en) 2019-06-20 2019-06-20 Dynamic functional mode learning method inspired by fMRI brain network mechanism
PCT/CN2019/093167 WO2020252805A1 (en) 2019-06-20 2019-06-27 Dynamic function mode learning method inspired by fmri brain network mechanism
JP2021500628A JP2022536881A (en) 2019-06-20 2019-06-27 A Dynamic Functional Mode Learning Method Inspired by fMRI Brain Network Mechanisms

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910536898.7A CN110265148B (en) 2019-06-20 2019-06-20 Dynamic functional mode learning method inspired by fMRI brain network mechanism

Publications (2)

Publication Number Publication Date
CN110265148A CN110265148A (en) 2019-09-20
CN110265148B true CN110265148B (en) 2021-06-01

Family

ID=67919794

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910536898.7A Active CN110265148B (en) 2019-06-20 2019-06-20 Dynamic functional mode learning method inspired by fMRI brain network mechanism

Country Status (3)

Country Link
JP (1) JP2022536881A (en)
CN (1) CN110265148B (en)
WO (1) WO2020252805A1 (en)

Families Citing this family (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110797123B (en) * 2019-10-28 2023-05-26 大连海事大学 Graph convolution neural network evolution method of dynamic brain structure
CN111259849B (en) * 2020-01-22 2023-05-12 深圳大学 Functional near infrared spectrum imaging resting state brain network detection method and device
CN111402212B (en) * 2020-03-04 2023-11-14 上海海事大学 Extraction method of dynamic connection activity mode of sea person brain function network
CN112382385A (en) * 2020-11-12 2021-02-19 山东中医药大学 Aura-free migraine auxiliary diagnosis algorithm based on 3D convolutional neural network
CN113345572B (en) * 2021-05-13 2024-03-15 东南大学 Brain disease judging system and judging device based on meta-analysis
CN113314216B (en) * 2021-06-01 2022-11-22 南方科技大学 Functional brain network construction method and device, electronic equipment and readable storage medium
CN114129147A (en) * 2021-10-14 2022-03-04 北京理工大学 System and method for predicting effects of Parkinson patients after DBS operation based on brain function network
CN116363404B (en) * 2023-03-30 2023-11-21 哈尔滨工业大学 Method for identifying and quantifying predictive pattern characterization of spontaneous brain activities
CN116597994B (en) * 2023-05-16 2024-05-14 天津大学 Mental disease brain function activity assessment device based on brain activation clustering algorithm
CN116313131B (en) * 2023-05-24 2023-09-15 山东大学 Brain network difference recognition system, equipment and storage medium based on imitation variables

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106709244A (en) * 2016-12-12 2017-05-24 西北工业大学 Brain function network modeling method for resting state synchronization EEG-fMRI
CN109316188A (en) * 2018-09-30 2019-02-12 上海海事大学 A kind of extracting method of migraine brain function connection mode

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101563153B1 (en) * 2013-07-24 2015-10-26 삼성전자주식회사 Method and apparatus for processing medical image signal
CN104715241B (en) * 2015-03-23 2018-11-13 南京工业大学 A kind of fMRI Feature extraction and recognition methods based on tensor resolution
CN104715261A (en) * 2015-03-23 2015-06-17 南京工业大学 FMRI dynamic brain function sub-network construction and parallel connection SVM weighted recognition method
WO2017172487A1 (en) * 2016-03-28 2017-10-05 The Board Of Trustees Of The Leland Stanford Junior University Detecting or treating post-traumatic stress syndrome
CN106204581B (en) * 2016-07-08 2019-10-18 西安交通大学 Dynamic brain function connection mode decomposition method based on PCA and K mean cluster
US10327637B2 (en) * 2017-06-08 2019-06-25 Fdna Inc. Systems, methods, and computer-readable media for patient image analysis to identify new diseases

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106709244A (en) * 2016-12-12 2017-05-24 西北工业大学 Brain function network modeling method for resting state synchronization EEG-fMRI
CN109316188A (en) * 2018-09-30 2019-02-12 上海海事大学 A kind of extracting method of migraine brain function connection mode

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
"A novel fMRI group data analysis method based on data-driven reference extracting from group subjects";Yuhu Shi等;《Computer methods and programs in biomedicine》;20150901;全文 *

Also Published As

Publication number Publication date
WO2020252805A1 (en) 2020-12-24
CN110265148A (en) 2019-09-20
JP2022536881A (en) 2022-08-22

Similar Documents

Publication Publication Date Title
CN110265148B (en) Dynamic functional mode learning method inspired by fMRI brain network mechanism
Wu et al. Bayesian Machine Learning: EEG\/MEG signal processing measurements
Thirion et al. Detection of signal synchronizations in resting-state fMRI datasets
CN111417342B (en) Brain network activity estimation system, method and storage medium
Carlson et al. An introduction to time-resolved decoding analysis for M/EEG
Miller et al. Higher dimensional analysis shows reduced dynamism of time-varying network connectivity in schizophrenia patients
CN112002428B (en) Whole brain individualized brain function map construction method taking independent component network as reference
CN109199376B (en) Decoding method of motor imagery electroencephalogram signal based on OA-WMNE brain source imaging
CN109745046B (en) Electrical impedance imaging electrode and system suitable for motion state
CN103325119A (en) Default state brain network center node detecting method based on modality fusion
Shi et al. A novel fMRI group data analysis method based on data-driven reference extracting from group subjects
CN107967686B (en) Epilepsy recognition device combining dynamic brain network and long-time and short-time memory network
CN108937968A (en) lead selection method of emotion electroencephalogram signal based on independent component analysis
Hu et al. Nonparametric matrix response regression with application to brain imaging data analysis
CN107832656B (en) Brain function state information processing method based on dynamic function brain network
CN113995422B (en) Transient brain power source positioning method and system based on nonnegative block sparse Bayesian learning
CN111402212A (en) Extraction method of dynamic connection activity mode of maritime brain function network
CN112861629B (en) Multi-window distinguishing typical pattern matching method and brain-computer interface application
CN115813409A (en) Ultra-low-delay moving image electroencephalogram decoding method
Dinh et al. Contextual minimum-norm estimates (CMNE): a deep learning method for source estimation in neuronal networks
Wu et al. A probabilistic framework for learning robust common spatial patterns
Shi et al. An improved multi-objective optimization-based CICA method with data-driver temporal reference for group fMRI data analysis
Khalid et al. Adaptive 2DCCA based approach for improving spatial specificity of activation detection in functional MRI
Paszkiel et al. Using the LORETA Method for Localization of the EEG Signal Sources in BCI Technology
Niranjan et al. Dynamic functional connectivity in the musical brain

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
GR01 Patent grant
GR01 Patent grant