CN110522448A - A kind of brain network class method based on figure convolutional neural networks - Google Patents

A kind of brain network class method based on figure convolutional neural networks Download PDF

Info

Publication number
CN110522448A
CN110522448A CN201910631316.3A CN201910631316A CN110522448A CN 110522448 A CN110522448 A CN 110522448A CN 201910631316 A CN201910631316 A CN 201910631316A CN 110522448 A CN110522448 A CN 110522448A
Authority
CN
China
Prior art keywords
brain
convolutional neural
neural networks
method based
node
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.)
Granted
Application number
CN201910631316.3A
Other languages
Chinese (zh)
Other versions
CN110522448B (en
Inventor
舒华忠
高舒雯
吴颖真
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.)
Southeast University
Original Assignee
Southeast 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 Southeast University filed Critical Southeast University
Priority to CN201910631316.3A priority Critical patent/CN110522448B/en
Publication of CN110522448A publication Critical patent/CN110522448A/en
Application granted granted Critical
Publication of CN110522448B publication Critical patent/CN110522448B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/05—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves 
    • A61B5/055—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves  involving electronic [EMR] or nuclear [NMR] magnetic resonance, e.g. magnetic resonance imaging
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/16—Devices for psychotechnics; Testing reaction times ; Devices for evaluating the psychological state
    • A61B5/165—Evaluating the state of mind, e.g. depression, anxiety
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/40—Detecting, measuring or recording for evaluating the nervous system
    • A61B5/4076—Diagnosing or monitoring particular conditions of the nervous system
    • A61B5/4082—Diagnosing or monitoring movement diseases, e.g. Parkinson, Huntington or Tourette
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/40—Detecting, measuring or recording for evaluating the nervous system
    • A61B5/4076—Diagnosing or monitoring particular conditions of the nervous system
    • A61B5/4088—Diagnosing of monitoring cognitive diseases, e.g. Alzheimer, prion diseases or dementia
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device

Landscapes

  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Neurology (AREA)
  • Animal Behavior & Ethology (AREA)
  • Public Health (AREA)
  • Veterinary Medicine (AREA)
  • General Health & Medical Sciences (AREA)
  • Surgery (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Pathology (AREA)
  • Biomedical Technology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Medical Informatics (AREA)
  • Psychiatry (AREA)
  • Developmental Disabilities (AREA)
  • Neurosurgery (AREA)
  • Artificial Intelligence (AREA)
  • Physiology (AREA)
  • Child & Adolescent Psychology (AREA)
  • Hospice & Palliative Care (AREA)
  • Psychology (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Mathematical Physics (AREA)
  • Fuzzy Systems (AREA)
  • Signal Processing (AREA)
  • High Energy & Nuclear Physics (AREA)
  • Evolutionary Computation (AREA)
  • Radiology & Medical Imaging (AREA)
  • Social Psychology (AREA)
  • Educational Technology (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)
  • Image Analysis (AREA)

Abstract

The brain network class method based on figure convolutional neural networks that the invention discloses a kind of, comprising the following steps: firstly, the right dependence signal of the blood oxygen for extracting each brain area from cerebral function nuclear magnetic resonance image；Secondly, building is able to reflect the mind map of functional connection topological features between brain region；Again, the mind map of building and practical diagnosis tag are input in figure convolutional neural networks and carry out feature learning and model training.The present invention is used for brain network class.

Description

A kind of brain network class method based on figure convolutional neural networks
Technical field
The brain network class method based on figure convolutional neural networks that the present invention relates to a kind of belongs to Digital image technology neck Domain.
Background technique
With the further development of society and science and technology, having had, which was more and more once considered incurable disease, is sent out The existing cause of disease simultaneously proposes corresponding treatment method therewith.As physical condition of the people to oneself is more paid attention to, to medical skill It is also had higher requirement in terms of art, especially at this stage, people increasingly pay close attention to the medical procedure of cerebral disease.Because human brain has Extremely complex structure and function understands the pathological characters of cerebral disease it is desirable to the operative mechanism by understanding brain and examines Disconnected method, countries in the world have put into a large amount of man power and material and have studied, such as USA and EU puts into 3,800,000,000 dollars respectively With 1,000,000,000 Euros, start brain project.On the one hand brain science research achievement will be best understood from brain, protection greatly for the mankind Brain, exploitation brain potential etc. make significant contribution, while also contributing to deepening to depression (Major Depressive Disorder, MDD), Alzheimer disease (Alzheimer's disease, AD) and its early stage, that is, mild cognitive function Obstacle (Mild cognitive impairment, MCI), the brains disease such as Parkinson's disease (Parkinson's disease, PD) The understanding of disease, and be the new method that this series nerve disease finds early diagnosis and therapy.Therefore, how research carries out Brain network class either all has very important significance for the health of clinical neurology research or people.
Functional magnetic resonance imaging (fMRI) is the doctor of a kind of common noninvasive description brain structure and connection features Learn image.FMRI can be used to explore working mechanism and rule of the brain under quiescent condition, and what is mainly reflected is brain mind Functional connection features through network.Its image-forming principle is that radiography is shaken by magnetic to capture the change of the hemodynamics under neuron activity Change, the right dependence signal of blood oxygen (BOLD signal) of each tissue points of brain is obtained and record, to reflect living body tranquillization indirectly Neuron activity situation under state.Thus it is possible to which obtaining brain function by fMRI is connected to the network matrix come observation analysis, research It whether there is significant function connects sex differernce between ordinary person and patient's intracerebral each region, auxiliary diagnosis come with this.
Influence of the diagnostic mode vulnerable to factors such as doctors experience and levels clinically at present, diagnostic result it is more subjective and It may there is a situation where mistaken diagnosis.In recent years, has the biology that disease is found in a large amount of research from the angle of brain medical image Diagnosis index.Currently used method is to directly adopt the function connects weight of Different brain region to be learnt as feature and divided Analysis, but this feature has ignored the topology information of brain network and causes diagnostic accuracy not high.
Summary of the invention
In recent years, has the biological diagnosis index that disease is found in a large amount of research from the angle of brain medical image.Mesh Preceding common method is to directly adopt the function connects weight of Different brain region to be learnt and analyzed as feature, but this is special The topology information that sign has ignored brain network causes diagnostic accuracy not high.The present invention is brain network class, provide it is a kind of from The topological structure (mind map) of brain network is extracted in fMRI as input feature vector, and the brain network based on figure convolutional neural networks divides Class method.
The present invention uses following technical scheme to solve above-mentioned technical problem:
The present invention provides a kind of brain network class method based on figure convolutional neural networks, comprising the following steps:
Step 1, it obtains cerebral function nuclear magnetic resonance image (fMRI) and it is pre-processed, therefrom extract each brain area Right dependence (BOLD) signal of blood oxygen, specifically include:
1-1, all fMRI data all use tranquillization state functional data processing auxiliary tool (DPARSF) 2.3 advanced Version kit is pre-processed, and for each subject, preceding ten frame of captured fMRI can all be dropped full to reach magnetic The stable state of sum.
1-2 carries out pretreated fMRI and big brain map to map available each brain region (L ROIs) FMRI, i.e., the blood oxygen of L brain region it is right rely on signal (BOLD signal) situation of change；
Step 2, building is able to reflect the mind map of functional connection topological features between brain region, the data structure The adjacency matrix of node label and binaryzation can be usedIt completely represents, and with adjacency matrix tensorMode stored. It specifically includes:
The BOLD signal of L brain region in step 1-2 is carried out Pearson correlation analysis two-by-two, obtains one by 2-1 L × L Pearson correlation adjacency matrix, i.e. generation brain are connected to the network matrix, the blood on matrix between each subregion of element representation Oxygen concentration relative coefficient:
Wherein, x=[x1,x2,…,xn] and y=[y1,y2,…,yn] indicate to carry out two groups of Pearson correlation analysis Signal, i.e. the BOLD signal of any two brain area.Area of the distribution of correlation coefficient that pearson correlation is analyzed in [- 1 ,+1] In, negative number representation is negatively correlated and positive number indicates to be positively correlated.It indicates more related closer to 1, indicates more not phase closer to 0 It closes；
2-2 takes square operation to related coefficient, related coefficient is mapped in [0,1] section, then carry out threshold value two-value Change processing, obtains the correlation adjacency matrix A (L × L) of binaryzation.Threshold binarization refers to the correlation in step 2-1 is adjacent Meet element r in matrixijCorresponding position a greater than threshold value TijIt is assigned a value of 1, is otherwise assigned a value of 0, binaryzation mode is expressed as follows:
Wherein, aijIndicate that the element on A, i, j=1,2 ..., L, T indicate quantization threshold.
2-3, the adjacency matrix A calculate node label according to the binaryzation acquired in step 2-2.Here node label is adopted With node in-degree information Di(i=1,2 ..., L) is characterized, i.e., is superimposed all elements of the i-th row in A:
2-4, the neighbour using the node label calculated in step 2-3 as partitioning standards, to the binaryzation acquired in step 2-2 It connects matrix A and carries out a quantization operation and obtain adjacency matrix tensor(N indicates node label Characteristic), wherein each slice AnOnly encode a certain specific node label numerical value of the mind map in an adjacency matrix Feature.
Step 3, the mind map constructed in step 2 and its actual classification label are sent to figure convolutional neural networks (Graph- CNN feature learning and classification diagnosis are carried out in).Figure convolutional neural networks are by picture scroll lamination, figure insertion pond layer and Quan Lian Connect layer composition.
3-1, using picture scroll product to the weighting summation of present node feature and its close node diagnostic, it is therefore an objective to preliminary poly- Class similar node, one linear filter of figure Defined, for being convex combination H ≈ h to each adjacency matrix1A1+h2A2+… +hNAN, have a corresponding filter parameter for each extraction feature (total C kind)Then there is figure Convolution is as follows:
Wherein, VinAnd VoutFor the input and output of picture scroll lamination, b is biasing；
3-2 elects the representative top for capableing of Efficient Characterization current class feature using figure insertion pond in every class vertex Point.Operating reduction number of nodes by pondization is N ', is exported as embeded matrix VembExpression formula is as follows:
3-3 extracts the characteristic that height is summarized by the insertion pondization operation of figure convolution sum figure several times.Again by complete The form that articulamentum tiles these high dimensional datas inputs and carries out the conversion of feature.By a series of full articulamentum (outputs Layer is also full articulamentum) complete final Decision Classfication.I.e. multiple figure convolution sum figure insertion pondization operations connect multiple full connections Layer constitutes network structure.
As further technical solution of the present invention, big brain map used in step 1-2 is automatic dissection label (AAL) map, the map amount to 90 brain region (separately having 26 cerebellum subregions that wouldn't be included in research range), then have L=90.
As further technical solution of the present invention, quantization threshold T value is 0.8 in step 2-2.
As further technical solution of the present invention, network structure used in step 3-3 is set as CNN (32)- CNN (32)-GEP (16)-CNN (16)-GEP (4)-FC (16)-FC (2), wherein CNN indicates that picture scroll lamination, GEP indicate figure insertion Chi Hua, FC indicate full articulamentum, and the numerical value in bracket indicates port number, the port number presentation class diagnosis of the FC layer at end Class number is 2.
The invention adopts the above technical scheme compared with prior art, has following technical effect that the invention discloses one Brain network class method of the kind based on figure convolutional neural networks realizes the two o'clock innovation on model method and brain network class, The topological structure (mind map) of brain network is innovatively extracted from fMRI herein as input feature vector, based on picture scroll product nerve net Network, which carries out feature extraction and brain network class, the present invention, can preferably be applied to brain network class, have parameter relatively fewer, learn Habit data characteristics loss is smaller, diagnostic result accuracy rate is higher, sensitivity is higher, specific higher, and the present invention can be preferable The different classes of sample properties of differentiation, indicate the biggish part of different classes of inter-sample difference, help to explain pathomechanism.
Detailed description of the invention
Fig. 1 is flow chart of the method for the present invention.
Fig. 2 is implementation process diagram of the invention.
Fig. 3 is that the adjacency matrix tensor storage mode of mind map is illustrated.
Fig. 4 is mind mapThe signal of data convolution pond process.
Fig. 5 is the present invention compared with the accuracy rate of common classification diagnostic method, sensitivity, specificity, and experiment uses 5 foldings Cross validation mode.
Specific embodiment
Technical solution of the present invention is described in further detail with reference to the accompanying drawing:
The present invention provides a kind of brain network class method based on figure convolutional neural networks and obtains first as illustrated in fig. 1 and 2 It takes fMRI data, complete brain area BOLD signal extraction；Then building is able to reflect between brain region functional connection topology knot The mind map of structure feature；Finally by the mind map of building and practical diagnosis tag be input in Graph-CNN carry out feature learning with And model training.
Firstly, the present invention provides a kind of brain network class method based on figure convolutional neural networks, comprising the following steps:
Step 1, it obtains cerebral function nuclear magnetic resonance image (fMRI) and it is pre-processed, therefrom extract each brain area Right dependence (BOLD) signal of blood oxygen, specifically include:
(1-1) all fMRI data all use tranquillization state functional data processing auxiliary tool (DPARSF) 2.3 advanced Version kit is pre-processed, and for each subject, preceding ten frame of captured fMRI can all be dropped full to reach magnetic The stable state of sum；
(1-2) carries out pretreated fMRI and big brain map to map available each brain region (L ROIs) FMRI, i.e., the right situation of change for relying on signal (BOLD signal) of the blood oxygen of L brain region, used big brain map is Automatic dissection label (AAL) map, which, which amounts to 90 brain region, (separately has 26 cerebellum subregions that wouldn't be included in research model Enclose), then there is L=90；
Step 2, building is able to reflect the mind map of functional connection topological features between brain region, the data structure The adjacency matrix of node label and binaryzation can be usedIt completely represents, and with adjacency matrix tensorMode stored. It specifically includes:
(2-1) carries out Pearson correlation analysis to the BOLD signal of L brain area in step (1-2) and generates L × L Pearson Correlation adjacency matrix, i.e. brain function are connected to the network matrix, and the blood oxygen concentration on matrix between each subregion of element representation is related Property coefficient:
Wherein, x=[x1,x2,…,xn] and y=[y1,y2,…,yn] indicate to carry out two groups of Pearson correlation analysis Signal, i.e. the BOLD signal of any two brain area.Area of the distribution of correlation coefficient that pearson correlation is analyzed in [- 1 ,+1] In, negative number representation is negatively correlated and positive number indicates to be positively correlated.It indicates more related closer to 1, indicates more not phase closer to 0 It closes；
(2-2) realizes the binary conversion treatment to brain function network connection matrix using threshold value quantizing.Related coefficient is taken Related coefficient is mapped in [0,1] section, then carries out threshold binarization treatment by square operation, obtains the correlation of binaryzation Adjacency matrixThreshold binarization refers to element r in the correlation adjacency matrix in step 2-1ijGreater than threshold value T Corresponding position aijIt is assigned a value of 1, is otherwise assigned a value of 0, binaryzation mode is expressed as follows:
Wherein, aijIndicate the element on the adjacency matrix A of binaryzation, i, j=1,2 ..., L.Quantization threshold T value is 0.8。
Node label is arranged in (2-3).Adjacency matrix A calculate node label according to the binaryzation acquired in step 2-2.This In node label use node in-degree information Di(i=1,2 ..., L) is characterized, i.e., folds all elements of the i-th row in A Add:
(2-4) completes the storage of the adjacency matrix tensor to projection mind map topological structure according to step.In step 2-3 The node label calculated is partitioning standards, carries out a quantization operation to the adjacency matrix A of the binaryzation acquired in step 2-2 and obtains To adjacency matrix tensor(characteristic of N expression node label), wherein each is sliced AnThe only feature of a certain specific node label numerical value of the coding mind map in an adjacency matrix.By taking Fig. 3 as an example, it is assumed that work as forebrain It include 6 brain region nodes in figure.If there is connection in quantification treatment operation between node and remaining node, between node Edge label can then be assigned 1, it is connectionless, be assigned a value of 0.It is as straight in numbered the node that the node for being 2 and number are 4,5,6 Connect all exist connection, then the node label of No. 2 nodes is then assigned 3, i.e., expression present node there are three nodes therewith There are connections.Then in mind map interior joint label there are the possibility of three kinds of values { 1,2,3 }, the storage square of the C=3 mind map at this time The dimension of battle array (adjacency matrix tensor) is 3 × 6 × 6.
Step 3: finally, the mind map constructed in step 2 and its actual classification label are sent to figure convolutional neural networks (Graph-CNN) feature learning and classification diagnosis are carried out in.Figure convolutional neural networks by picture scroll lamination, figure insertion pond layer with And full articulamentum is constituted.Network structure of the invention is set as CNN (32)-CNN (32)-GEP (16)-CNN (16)-GEP (4)- FC (16)-FC (2), wherein CNN indicates that picture scroll lamination, GEP indicate that figure insertion pond, FC indicate full articulamentum, the number in bracket Value indicates port number, and the class number of the port number presentation class diagnosis of the FC layer at end is 2.
(3-1) is using picture scroll product to the weighting summation of present node feature and its close node diagnostic, it is therefore an objective to preliminary Cluster similar node.One linear filter of figure Defined, for being convex combination H ≈ h to each adjacency matrix1A1+h2A2 +…+hNAN.There is a corresponding filter parameter for each extraction feature (total C kind)Then have Picture scroll product is as follows:
Wherein, VinAnd VoutFor the input and output of picture scroll lamination, b is biasing.
(3-2) elects the representative top for capableing of Efficient Characterization current class feature using figure insertion pond in every class vertex Point.Operating reduction number of nodes by pondization is N ', is exported as embeded matrix VembExpression formula is as follows:
(3-3) extracts the characteristic that height is summarized by the insertion pondization operation of figure convolution sum figure several times.Again by complete The form that articulamentum tiles these high dimensional datas inputs and carries out the conversion of feature.By a series of full articulamentum (outputs Layer is also full articulamentum) complete final Decision Classfication.
The present invention realizes the innovation of the two o'clock on model method and brain network class.It is innovative herein on model method Ground extracts the topological structure (mind map) of brain network as input feature vector, base from cerebral function nuclear magnetic resonance image (fMRI) Feature extraction and brain network class are carried out in figure convolutional neural networks (Graph-CNN).
Application Example:
132 data instances provided below with the attached middle large hospital of Southeast China University, it is of the invention based on picture scroll to illustrate The depression classification diagnosis method of product neural network.Data set includes 50 normal controls (HC) and 82 patients with depression (MDD).Patients with depression can be subdivided into two classes again: (1) drug is effective (RD), and totally 42；(2) drug ineffective (NRD), altogether 40.All subjects receive Siemens's 3T High-resolution MRI scanning (uniform birdcage coils) in the hospital.Subject is flat It lies, reduction head movement as far as possible is closely fixed with belt and foam pad in head.During the scanning process, all subjects are It is required to close eyes, loosens to keep regaining consciousness and do not go to think anything.High-resolution three-dimension t1 weighted image is prepared by magnetization Double echo steady state obtains, and the parameter of use is specific as follows: repetition time (TR)=1900 millisecond (ms), echo time (TE)=2.48ms, flip angle (FA)=9 °, acquisition matrix=256 × 256, the visual field (FOV)=250 × 250 square millimeter (mm2), thickness=1.0 millimeter (mm), gap=0mm, 176.Eight minutes tranquillization state functional mris (rs-fMRI) Acquisition parameter be provided that TR=2000ms, TE=25ms, FA=90 °, acquisition matrix=64 × 64, FOV=240 × 240mm2；Thickness=3mm, gap=0mm, 36 axial slices, volume 240,3.75 × 3.75 mm2It is parallel to bicommissural line Flat resolution.
Experiment condition: it now chooses a computer and is tested, which is configured with Intel processors (3.4GHz) and 16GB random access memory, 64 bit manipulation systems.The present invention is based on the depression of figure convolutional neural networks point Class diagnostic model is by brain area signal extraction, the building of brain topological diagram and the big module composition of figure convolutional neural networks three.Brain area signal mentions Modulus block, to the conversion of each subregion BOLD signal, is write, MATLAB according to the full brain fMRI of big brain map Mapping implementation using m language Compiling；Brain topological diagram constructs module to construct mind map data, realizes that the correlation adjacency matrix of quantization and node label are asked Solution is write, MATLAB compiling using m language；Picture scroll product neural network module using picture scroll product, priori pond mode to mind map into Row signature analysis, and differentiated by full articulamentum implementation pattern, it is write using Python, Pycharm compiling.
Fig. 5 is the contrast and experiment of this method and currently used classification of diseases algorithm, specifically includes accuracy rate, sensitive Degree, specificity analysis.As can be seen from the table, at the cross check system of same ratio (5-fold), base proposed in this paper In the method for figure convolutional neural networks be either in terms of medical diagnosis on disease or outcome prediction all have more preferably sensitivity and The coincidence factor highest of specificity, i.e. sample predictions classification and concrete class.
The above, the only specific embodiment in the present invention, but scope of protection of the present invention is not limited thereto, appoints What is familiar with the people of the technology within the technical scope disclosed by the invention, it will be appreciated that expects transforms or replaces, and should all cover Within scope of the invention, therefore, the scope of protection of the invention shall be subject to the scope of protection specified in the patent claim.

Claims (7)

1. a kind of brain network class method based on figure convolutional neural networks, which comprises the following steps:
Step 1, it obtains cerebral function nuclear magnetic resonance image (fMRI) and it is pre-processed, therefrom extract the blood of each brain area Right dependence (BOLD) signal of oxygen,
Step 2, building is able to reflect between brain region the mind map of functional connection topological features, which can be with With the adjacency matrix of node label and binaryzationIt completely represents, and with adjacency matrix tensorMode stored；
Step 3, the mind map constructed in step 2 and its actual classification label are sent to figure convolutional neural networks (Graph-CNN) Middle progress feature learning and classification diagnosis, figure convolutional neural networks are by picture scroll lamination, figure insertion pond layer and full articulamentum It constitutes.
2. a kind of brain network class method based on figure convolutional neural networks according to claim 1, which is characterized in that institute Step 1 is stated, cerebral function nuclear magnetic resonance image (fMRI) is obtained and it is pre-processed, therefrom extracts the blood oxygen of each brain area Right dependence (BOLD) signal, specifically includes:
1-1, all fMRI data all use tranquillization state functional data to handle auxiliary tool (DPARSF) 2.3 advanced editions tools Packet is pre-processed, and for each subject, preceding ten frame of captured fMRI can be all dropped to reach magnetically saturated steady Determine state；
1-2 carries out pretreated fMRI and big brain map to map available each brain region (L ROIs) FMRI, i.e., the right situation of change for relying on signal (BOLD signal) of the blood oxygen of L brain region.
3. a kind of brain network class method based on figure convolutional neural networks according to claim 2, which is characterized in that
Step 2, building is able to reflect the mind map of functional connection topological features between brain region, specifically includes:
The BOLD signal of L brain region in step 1-2 is carried out Pearson correlation analysis two-by-two, obtains a L × L by 2-1 Pearson correlation adjacency matrix, i.e. generation brain are connected to the network matrix, and the blood oxygen on matrix between each subregion of element representation is dense Spend relative coefficient:
Wherein, x=[x1,x2,…,xn] and y=[y1,y2,…,yn] indicate to carry out two groups of letters of Pearson correlation analysis Number, i.e. the BOLD signal of any two brain area, section of the distribution of correlation coefficient that pearson correlation is analyzed in [- 1 ,+1] Interior, negative number representation is negatively correlated and positive number indicates to be positively correlated.It indicates more related closer to 1, indicates more uncorrelated closer to 0；
2-2 takes square operation to related coefficient, and related coefficient is mapped in [0,1] section, then is carried out at threshold binarization Reason, obtains the correlation adjacency matrix A (L × L) of binaryzation.Threshold binarization refers to the adjacent square of correlation in step 2-1 Element r in battle arrayijCorresponding position a greater than threshold value TijIt is assigned a value of 1, is otherwise assigned a value of 0, binaryzation mode is expressed as follows:
Wherein, aijIndicate that the element on A, i, j=1,2 ..., L, T indicate quantization threshold.
2-3, the adjacency matrix A calculate node label according to the binaryzation acquired in step 2-2.Here node label is using section Point in-degree information Di(i=1,2 ..., L) is characterized, i.e., is superimposed all elements of the i-th row in A:
2-4, the adjoining square using the node label calculated in step 2-3 as partitioning standards, to the binaryzation acquired in step 2-2 Battle array A carries out a quantization operation and obtains adjacency matrix tensor(N indicates node mark The characteristic of label), wherein each is sliced AnOnly a certain specific node label numerical value of the coding mind map in an adjacency matrix Feature.
4. a kind of brain network class method based on figure convolutional neural networks according to claim 3, which is characterized in that step Rapid 3, the mind map constructed in step 2 and its actual classification label are sent in figure convolutional neural networks (Graph-CNN) and carried out Feature learning and classification diagnosis, specific as follows:
3-1, using picture scroll product to the weighting summation of present node feature and its close node diagnostic, it is therefore an objective to preliminary clusters phase Like node, one linear filter of figure Defined, for being convex combination H ≈ h to each adjacency matrix1A1+h2A2+…+hNAN, There is a corresponding filter parameter for each extraction feature (total C kind)Then there is picture scroll product such as Under:
Wherein, VinAnd VoutFor the input and output of picture scroll lamination, b is biasing；
3-2 is elected the representative vertex for capableing of Efficient Characterization current class feature in every class vertex using figure insertion pond, led to Crossing pondization operation reduction number of nodes is N ', is exported as embeded matrix VembExpression formula is as follows:
3-3 extracts the characteristic that height is summarized, then by connecting entirely by the insertion pondization operation of figure convolution sum figure several times These high dimensional datas are inputted and are carried out the conversion of feature in the form to tile by layer.By a series of full articulamentum (output layers It is full articulamentum) complete final Decision Classfication.I.e. multiple figure convolution sum figure insertion pondization operations connect multiple full articulamentum structures At network structure.
5. a kind of brain network class method based on figure convolutional neural networks according to claim 1, which is characterized in that
Big brain map used in the step 1-2 is automatic dissection label (AAL) map, which amounts to 90 brains point Area (separately has 26 cerebellum subregions that wouldn't be included in research range), then has L=90.
6. a kind of brain network class method based on figure convolutional neural networks according to claim 1, which is characterized in that step Quantization threshold T value is 0.8 in rapid 2-2.
7. a kind of brain network class method based on figure convolutional neural networks according to claim 1, which is characterized in that step Network structure used in rapid 3-3 is set as CNN (32)-CNN (32)-GEP (16)-CNN (16)-GEP (4)-FC (16)-FC (2), wherein CNN indicates that picture scroll lamination, GEP indicate that figure insertion pond, FC indicate full articulamentum, and the numerical value in bracket indicates channel The class number of number, the port number presentation class diagnosis of the FC layer at end is 2.
CN201910631316.3A 2019-07-12 2019-07-12 Brain network classification method based on atlas neural network Active CN110522448B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910631316.3A CN110522448B (en) 2019-07-12 2019-07-12 Brain network classification method based on atlas neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910631316.3A CN110522448B (en) 2019-07-12 2019-07-12 Brain network classification method based on atlas neural network

Publications (2)

Publication Number Publication Date
CN110522448A true CN110522448A (en) 2019-12-03
CN110522448B CN110522448B (en) 2023-04-11

Family

ID=68659704

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910631316.3A Active CN110522448B (en) 2019-07-12 2019-07-12 Brain network classification method based on atlas neural network

Country Status (1)

Country Link
CN (1) CN110522448B (en)

Cited By (29)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110934587A (en) * 2019-12-13 2020-03-31 深圳龙岗智能视听研究院 Alzheimer disease auxiliary diagnosis method based on atlas neural network
CN110974203A (en) * 2019-12-10 2020-04-10 电子科技大学 Heart rate imbalance analysis method based on equivalent arrangement network
CN111127441A (en) * 2019-12-25 2020-05-08 兰州大学 Multi-modal brain image depression recognition method and system based on graph node embedding
CN111339307A (en) * 2020-04-07 2020-06-26 广西师范大学 Graph residual error network text classification method based on word inverse frequency weighting
CN111419221A (en) * 2020-02-14 2020-07-17 广东司法警官职业学院 Electroencephalogram signal analysis method based on graph convolution network
CN111466876A (en) * 2020-03-24 2020-07-31 山东大学 Alzheimer's disease auxiliary diagnosis system based on fNIRS and graph neural network
CN111563533A (en) * 2020-04-08 2020-08-21 华南理工大学 Test subject classification method based on graph convolution neural network fusion of multiple human brain maps
CN112294338A (en) * 2020-09-29 2021-02-02 山东师范大学 Epilepsy detection system, equipment and medium based on graph attention network
CN112862751A (en) * 2020-12-30 2021-05-28 电子科技大学 Automatic diagnosis device for autism
CN113040715A (en) * 2021-03-09 2021-06-29 北京工业大学 Human brain function network classification method based on convolutional neural network
CN113288131A (en) * 2021-05-06 2021-08-24 广东工业大学 Non-invasive blood glucose detection method, processor and device based on graph convolution network
CN113313232A (en) * 2021-05-19 2021-08-27 华南理工大学 Functional brain network classification method based on pre-training and graph neural network
CN113555091A (en) * 2021-09-22 2021-10-26 中国科学院自动化研究所 Parkinson sMRI image-based disease classification system, method and device
CN113592836A (en) * 2021-08-05 2021-11-02 东南大学 Deep multi-modal graph convolution brain graph classification method
CN113729729A (en) * 2021-08-13 2021-12-03 北京航空航天大学 Schizophrenia early detection system based on graph neural network and brain network
CN113781385A (en) * 2021-03-19 2021-12-10 南京航空航天大学 Joint attention-seeking convolution method for brain medical image automatic classification
CN113786185A (en) * 2021-09-18 2021-12-14 安徽师范大学 Static brain network feature extraction method and system based on convolutional neural network
CN113920123A (en) * 2021-12-16 2022-01-11 中国科学院深圳先进技术研究院 Addictive brain network analysis method and device
CN114048848A (en) * 2022-01-13 2022-02-15 深圳市永达电子信息股份有限公司 Brain-like computing method and system based on memory mechanism
CN114159043A (en) * 2021-12-17 2022-03-11 天津大学 Brain function network abnormal brain node data detection method based on Qcut algorithm
CN114255228A (en) * 2021-12-23 2022-03-29 福州大学 Brain function connection network analysis method based on resting state function magnetic resonance image
CN114376558A (en) * 2022-03-24 2022-04-22 之江实验室 Brain atlas individuation method and system based on magnetic resonance and twin map neural network
CN114391826A (en) * 2022-01-24 2022-04-26 西北大学 Human characterization prediction method and device based on edge-driven graph neural network
CN114403812A (en) * 2022-03-30 2022-04-29 慧创科仪(北京)科技有限公司 Auxiliary analysis method, device and system for brain injury condition and storage medium
CN115081558A (en) * 2022-08-22 2022-09-20 北京元亦科技有限公司 Brain network data classification method and system based on graph neural network
CN115294413A (en) * 2022-10-10 2022-11-04 齐鲁工业大学 Brain disease classification method based on time sequence electroencephalogram data
CN115474939A (en) * 2022-10-21 2022-12-16 重庆邮电大学 Autism spectrum disorder recognition model based on deep expansion neural network
CN116912576A (en) * 2023-07-14 2023-10-20 太原理工大学 Self-adaptive graph convolution brain disease classification method based on brain network higher-order structure
RU2819348C1 (en) * 2023-06-16 2024-05-17 Федеральное государственное автономное образовательное учреждение высшего образования "Уральский федеральный университет имени первого Президента России Б.Н. Ельцина" Method for graphed neural network classification for absence or presence of major depressive disorder according to fmri data

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107909117A (en) * 2017-09-26 2018-04-13 电子科技大学 A kind of sorting technique and device based on brain function network characterization to early late period mild cognitive impairment
CN109065128A (en) * 2018-09-28 2018-12-21 郑州大学 A kind of sparse brain network establishing method of weighted graph regularization
CN109376751A (en) * 2018-08-21 2019-02-22 北京工业大学 A kind of human functional network's classification method based on convolutional neural networks

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107909117A (en) * 2017-09-26 2018-04-13 电子科技大学 A kind of sorting technique and device based on brain function network characterization to early late period mild cognitive impairment
CN109376751A (en) * 2018-08-21 2019-02-22 北京工业大学 A kind of human functional network's classification method based on convolutional neural networks
CN109065128A (en) * 2018-09-28 2018-12-21 郑州大学 A kind of sparse brain network establishing method of weighted graph regularization

Cited By (44)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110974203A (en) * 2019-12-10 2020-04-10 电子科技大学 Heart rate imbalance analysis method based on equivalent arrangement network
CN110934587A (en) * 2019-12-13 2020-03-31 深圳龙岗智能视听研究院 Alzheimer disease auxiliary diagnosis method based on atlas neural network
CN111127441A (en) * 2019-12-25 2020-05-08 兰州大学 Multi-modal brain image depression recognition method and system based on graph node embedding
CN111127441B (en) * 2019-12-25 2022-08-12 兰州大学 Multi-modal brain image depression recognition method and system based on graph node embedding
CN111419221A (en) * 2020-02-14 2020-07-17 广东司法警官职业学院 Electroencephalogram signal analysis method based on graph convolution network
CN111466876B (en) * 2020-03-24 2021-08-03 山东大学 Alzheimer's disease auxiliary diagnosis system based on fNIRS and graph neural network
CN111466876A (en) * 2020-03-24 2020-07-31 山东大学 Alzheimer's disease auxiliary diagnosis system based on fNIRS and graph neural network
CN111339307A (en) * 2020-04-07 2020-06-26 广西师范大学 Graph residual error network text classification method based on word inverse frequency weighting
CN111563533B (en) * 2020-04-08 2023-05-02 华南理工大学 Subject classification method based on graph convolution neural network fusion of various human brain maps
CN111563533A (en) * 2020-04-08 2020-08-21 华南理工大学 Test subject classification method based on graph convolution neural network fusion of multiple human brain maps
CN112294338A (en) * 2020-09-29 2021-02-02 山东师范大学 Epilepsy detection system, equipment and medium based on graph attention network
CN112862751A (en) * 2020-12-30 2021-05-28 电子科技大学 Automatic diagnosis device for autism
CN112862751B (en) * 2020-12-30 2022-05-31 电子科技大学 Automatic diagnosis device for autism
CN113040715A (en) * 2021-03-09 2021-06-29 北京工业大学 Human brain function network classification method based on convolutional neural network
CN113040715B (en) * 2021-03-09 2023-06-02 北京工业大学 Human brain function network classification method based on convolutional neural network
CN113781385B (en) * 2021-03-19 2024-03-08 南京航空航天大学 Combined attention-seeking convolution method for automatic classification of brain medical images
CN113781385A (en) * 2021-03-19 2021-12-10 南京航空航天大学 Joint attention-seeking convolution method for brain medical image automatic classification
CN113288131A (en) * 2021-05-06 2021-08-24 广东工业大学 Non-invasive blood glucose detection method, processor and device based on graph convolution network
CN113288131B (en) * 2021-05-06 2022-07-12 广东工业大学 Non-invasive blood glucose detection method, processor and device based on graph convolution network
CN113313232A (en) * 2021-05-19 2021-08-27 华南理工大学 Functional brain network classification method based on pre-training and graph neural network
CN113592836B (en) * 2021-08-05 2022-11-18 东南大学 Deep multi-modal graph convolution brain graph classification method
CN113592836A (en) * 2021-08-05 2021-11-02 东南大学 Deep multi-modal graph convolution brain graph classification method
CN113729729B (en) * 2021-08-13 2022-06-14 北京航空航天大学 Schizophrenia early detection system based on graph neural network and brain network
CN113729729A (en) * 2021-08-13 2021-12-03 北京航空航天大学 Schizophrenia early detection system based on graph neural network and brain network
CN113786185A (en) * 2021-09-18 2021-12-14 安徽师范大学 Static brain network feature extraction method and system based on convolutional neural network
CN113786185B (en) * 2021-09-18 2024-05-07 安徽师范大学 Static brain network feature extraction method and system based on convolutional neural network
CN113555091A (en) * 2021-09-22 2021-10-26 中国科学院自动化研究所 Parkinson sMRI image-based disease classification system, method and device
CN113555091B (en) * 2021-09-22 2022-02-08 中国科学院自动化研究所 Parkinson sMRI image-based disease classification system, method and device
CN113920123A (en) * 2021-12-16 2022-01-11 中国科学院深圳先进技术研究院 Addictive brain network analysis method and device
CN114159043A (en) * 2021-12-17 2022-03-11 天津大学 Brain function network abnormal brain node data detection method based on Qcut algorithm
CN114255228A (en) * 2021-12-23 2022-03-29 福州大学 Brain function connection network analysis method based on resting state function magnetic resonance image
CN114255228B (en) * 2021-12-23 2024-08-09 福州大学 Brain function connection network analysis method based on resting state function magnetic resonance image
CN114048848A (en) * 2022-01-13 2022-02-15 深圳市永达电子信息股份有限公司 Brain-like computing method and system based on memory mechanism
CN114391826A (en) * 2022-01-24 2022-04-26 西北大学 Human characterization prediction method and device based on edge-driven graph neural network
CN114376558B (en) * 2022-03-24 2022-07-19 之江实验室 Brain atlas individuation method and system based on magnetic resonance and twin map neural network
CN114376558A (en) * 2022-03-24 2022-04-22 之江实验室 Brain atlas individuation method and system based on magnetic resonance and twin map neural network
US12127826B2 (en) 2022-03-24 2024-10-29 Zhejiang Lab Brain atlas individualization method and system based on magnetic resonance and twin graph neural network
CN114403812A (en) * 2022-03-30 2022-04-29 慧创科仪(北京)科技有限公司 Auxiliary analysis method, device and system for brain injury condition and storage medium
CN115081558A (en) * 2022-08-22 2022-09-20 北京元亦科技有限公司 Brain network data classification method and system based on graph neural network
CN115294413A (en) * 2022-10-10 2022-11-04 齐鲁工业大学 Brain disease classification method based on time sequence electroencephalogram data
CN115474939A (en) * 2022-10-21 2022-12-16 重庆邮电大学 Autism spectrum disorder recognition model based on deep expansion neural network
RU2819348C1 (en) * 2023-06-16 2024-05-17 Федеральное государственное автономное образовательное учреждение высшего образования "Уральский федеральный университет имени первого Президента России Б.Н. Ельцина" Method for graphed neural network classification for absence or presence of major depressive disorder according to fmri data
CN116912576A (en) * 2023-07-14 2023-10-20 太原理工大学 Self-adaptive graph convolution brain disease classification method based on brain network higher-order structure
CN116912576B (en) * 2023-07-14 2024-10-15 太原理工大学 Self-adaptive graph convolution brain disease classification method based on brain network higher-order structure

Also Published As

Publication number Publication date
CN110522448B (en) 2023-04-11

Similar Documents

Publication Publication Date Title
CN110522448A (en) A kind of brain network class method based on figure convolutional neural networks
Dinsdale et al. Learning patterns of the ageing brain in MRI using deep convolutional networks
Liu et al. Multi-modality cascaded convolutional neural networks for Alzheimer’s disease diagnosis
CN113571195B (en) Early Alzheimer disease prediction model based on cerebellar function connection characteristics
Gull et al. Artificial intelligence in brain tumor detection through MRI scans: advancements and challenges
CN108366752B (en) Brain activity analysis device and method, storage medium, and biomarker device
Chaki et al. Deep learning for neurodegenerative disorder (2016 to 2022): A systematic review
CN106023194A (en) Amygdaloid nucleus spectral clustering segmentation method based on resting state function connection
CN113284126B (en) Method for predicting hydrocephalus shunt operation curative effect by artificial neural network image analysis
Wang et al. Sensorineural hearing loss identification via nine-layer convolutional neural network with batch normalization and dropout
Liu et al. Attention deficit/hyperactivity disorder classification based on deep spatio-temporal features of functional magnetic resonance imaging
CN112488976A (en) Multi-modal medical image fusion method based on DARTS network
CN109222902A (en) Parkinson's prediction technique, system, storage medium and equipment based on nuclear magnetic resonance
CN115424067A (en) System, method, processor and storage medium for realizing classification processing of depression subtypes based on multiple fusion brain network diagram technology
CN114463493A (en) Transcranial magnetic stimulation electric field rapid imaging method and model based on coding and decoding structure
Savaashe et al. A review on cardiac image segmentation
Alsiddiky et al. Vertebral tumor detection and segmentation using analytical transform assisted statistical characteristic decomposition model
Wang et al. Deep transfer learning-based multi-modal digital twins for enhancement and diagnostic analysis of brain mri image
Javeed et al. Brain tumor segmentation and classification with hybrid clustering, probabilistic neural networks
CN107731283A (en) A kind of image radio system based on more subspace modelings
Li et al. NVTrans‐UNet: Neighborhood vision transformer based U‐Net for multi‐modal cardiac MR image segmentation
CN115147417B (en) Function connection matrix processing system and device based on filtering method feature selection
Muthiah et al. Fusion of MRI and PET images using deep learning neural networks
Pugazhenthi et al. Identification of autism in MR brain images using deep learning networks
CN109223001A (en) Hyperactivity prediction technique, system, storage medium and equipment based on nuclear magnetic resonance

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