CN110503197A - Non-uniform magnetic field hypencephalon metabolite concentration quantization method based on neural network algorithm - Google Patents
Non-uniform magnetic field hypencephalon metabolite concentration quantization method based on neural network algorithm Download PDFInfo
- Publication number
- CN110503197A CN110503197A CN201910625103.XA CN201910625103A CN110503197A CN 110503197 A CN110503197 A CN 110503197A CN 201910625103 A CN201910625103 A CN 201910625103A CN 110503197 A CN110503197 A CN 110503197A
- Authority
- CN
- China
- Prior art keywords
- brainmetabolite
- data
- neural network
- magnetic field
- hypencephalon
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/0033—Features or image-related aspects of imaging apparatus classified in A61B5/00, e.g. for MRI, optical tomography or impedance tomography apparatus; arrangements of imaging apparatus in a room
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
-
- 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
-
- 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
Abstract
The present invention provides the non-uniform field hypencephalon metabolite concentration quantization methods based on neural network algorithm, are related to biological living brainmetabolite quantitative concentration method.Using free induction decay signal simulation softward, biological brainmetabolite nuclear magnetic resonance analog signal is generated；Establish various concentration metabolite profile diagram data collection and project training data and label data under non-uniform magnetic field；Division whole data set is training set, verifying collection, test set；It designs neural network structure and chooses neural network hyper parameter using verifying collection；Test verification is carried out to quantitative model using test data；Using the powerful capability of fitting of neural network algorithm proposed by the present invention, biological living brainmetabolite spectrogram under a kind of non-uniform magnetic field may be implemented and quickly and accurately quantify, make localization spectrum that there is wider application.
Description
Technical field
The present invention relates to biological living brainmetabolite quantitative concentration method, more particularly, to based on neural network algorithm not
Uniform magnetic field hypencephalon metabolite concentration quantization method.
Background technique
Nuclear magnetic resonance localization spectrum (or MRS) technology is currently a kind of important means in biological living detection, it can be with
Non-intrusive form obtains the metabolin information of detection biology.Under the sampling condition of short echo, metabolites kinds are up to more than 20
Class, and there are also macromolecular signals and lipid signal as its background signal.Further, since the artifact that sampling process generates is (such as
Vortex or remaining water signal) and white Gaussian noise e have an impact to the signal finally obtained.The many fittings proposed at present
Method, such as QUEST, LCModel, TARQUIN do not accomplish still stable, reliable and accurately quantify brainmetabolite.Influence spectrogram
Quality and the factor that can be used effectively, wherein be some the width of spectral line, and breadth of spectrum line depends on the equal of magnetic field
Even property.Although having some improvement and improving from the influence that the angle of hardware resists non-uniform field at present, but certain situations are also
It is that can not be solved by hardware.Such as in biological living detection, the intrinsic magnetic susceptibility of certain metabolizing tissues just not phase itself
Together, this results in the uniformity for the magnetic field environment experienced in its experiment ideal not to the utmost, and these are to be difficult to pass through hardware technology
It solves.By the nerual network technique for being successfully applied multiple fields in the near future, it is introduced into nuclear magnetic resonance localization
Spectrum field exactly utilizes its powerful fitting abstract function ability, by low resolution and lower letter under a non-uniform magnetic-field
The spectrogram of ratio of making an uproar inputs trained neural network model, the information of corresponding various metabolite concentrations is exported, thus realization amount
Change.
Summary of the invention
The main technical problem to be solved by the present invention is to provide the non-uniform magnetic field hypencephalon generations based on neural network algorithm
Thanking to object spectrogram quantization method can stablize and accurately quantify to biological living brainmetabolite concentration.
In order to solve the above technical problems, the present invention provides the non-uniform magnetic field hypencephalon generations based on neural network algorithm
Thank to object spectrogram quantization method, comprising the following steps:
1) it utilizes free induction decay signal simulation softward FID-A and selects common detectable biological living brainmetabolite mould
Quasi- basis set carries out simulating each brainmetabolite PRESS experiment, carries out band coefficient by the analog signal to each brainmetabolite and add
Power is cumulative, and fixed water signal weighting coefficient obtains final biological brainmetabolite analog signal；
2) training data is established by the analogue data of the different line widths of generation, different signal-to-noise ratio, different brainmetabolite concentration
Collection, the random fraction for taking out training data concentration is respectively as verifying collection and test set, by each analogue data of generation
Corresponding brainmetabolite weighting coefficient is as label data；
3) data prediction is carried out to above-mentioned all data sets, pretreatment includes one-dimensional Fourier transform, normalization, data
It cuts；
4) Artificial Neural Network Structures are designed, the hyper parameter for choosing neural network structure is collected using verifying, obtain quantization mould
Type；
5) test verification is carried out to quantitative model using test data: after being pre-processed to test data, input quantization
Model, i.e., its exportable corresponding each metabolite concentration information；It is described pretreatment include one-dimensional Fourier transform, normalization,
Data are cut.
In a preferred embodiment: the brainmetabolite include but is not limited to Cr, NAA, Cho, Gln, GABA, Ins, Tau,
Lac、Glu。
In a preferred embodiment: in step 1, the coefficient weighting of the analog signal of each brainmetabolite refers to: according to each
The difference of a brainmetabolite concentration in brain model solution determines that each brainmetabolite analog signal is simulated in biological brainmetabolite
Weight in signal, to obtain the coefficient weighting of the analog signal of each brainmetabolite.
In a preferred embodiment: fixing the weighting of water signal coefficient in step 1 and refer to, according to water in brain model solution
Concentration determines weight of the water signal in biological brainmetabolite analog signal, to obtain the coefficient weighting of water signal.
In a preferred embodiment: in step 5, the test data includes that data are adopted in analogue data and experiment in fact；Wherein
The real data of adopting of experiment are needed by pretreatment.
In a preferred embodiment: the data cut the dimension for referring to and reducing data.
Compared to the prior art, the present invention has more acurrate and stable advantage in quantization ability, and secondly the present invention is right
The inhomogeneities in magnetic field shows more preferably robustness.
Detailed description of the invention
Fig. 1 be the present invention relates to for biological brain metabolite concentration quantization use neural network structure.
Each metabolite concentration quantization error of Fig. 2 test set (analogue data).
Fig. 3 is to be carried out at quantization using method proposed by the present invention and Tarquin method to 10 groups of brain model analogue datas
The error of reason compares.
Fig. 4 be using method proposed by the present invention to 10 groups of brain model solution adopt in fact data carry out quantification treatment obtain it is each
Metabolite concentration quantized result.
Fig. 5 is to adopt data in fact to 10 groups of brain model solution using Tarquin method to carry out each metabolism that quantification treatment obtains
Object quantitative concentration result.
Specific embodiment
Following embodiment will the present invention is further illustrated in conjunction with attached drawing.
The embodiment of the present invention will utilize free induction decay signal simulation softward FID-A (The FID
Appliance) generate magnetic resonance signal and training network, then collected under non-uniform magnetic field magnetic resonance signal first pass through it is pre-
Processing recycling the method for the present invention exports the concentration information of each brainmetabolite.Specific implementation process is as follows:
1) it utilizes free induction decay signal simulation softward FID-A (The FID Appliance) and selects common examine
Biological living brainmetabolite simulation basis set is surveyed to carry out simulating each brainmetabolite PRESS experiment；By to each brainmetabolite
Analog signal carries out band coefficient weighted accumulation, and fixed water signal weighting coefficient obtains final brainmetabolite analog signal.
The brainmetabolite includes but is not limited to Cr, NAA, Cho, Gln, GABA, Ins, Tau, Lac, Glu.The present embodiment
Middle analogue data sampling number is 8192, and simulation home court is 7.0T by force.
The coefficient weighting of the analog signal of each brainmetabolite refers to: dense in brain model solution according to each brainmetabolite
The difference of degree determines weight of each brainmetabolite analog signal in biological brainmetabolite analog signal, to obtain each
The coefficient of the analog signal of brainmetabolite weights.
Fixed water signal coefficient weighting refers to, according to concentration of the water in brain model solution, determines water signal in biological brain
Weight in metabolin analog signal, to obtain the coefficient weighting of water signal.In the present embodiment, with reference to water in brain model solution
Concentration be 55556mM.
2) by generating different line widths (1-30Hz), different signal-to-noise ratio (about 4.7-52), different brainmetabolite concentration
The analogue data of (random number generates, and range is 0-15) establishes training dataset, the random fraction for taking out training data and concentrating
Respectively as verifying collection and test set, using brainmetabolite weighting coefficient corresponding to each analogue data of generation as label
Data.
3) data predictions are carried out to above-mentioned all data sets, including one-dimensional Fourier transform (frequency domain is transformed to by time domain,
And take real part as data input), normalization, data cut (reduce data dimension, to accelerate training process).
4) Artificial Neural Network Structures are designed, the best hyper parameter combination for choosing neural network structure are collected using verifying, such as
The unit number of hidden layer number and each layer.Final process process is as shown in Figure 1.
Optimization algorithm uses the most common Adam algorithm, and parameter initialization mode is Xavier initialization, the study of selection
Rate is 0.001.
5) test verification is carried out to quantitative model using test data, and calculates quantization error.Fig. 2 is analogue data test
Each metabolite concentration quantization error that collection is handled by the method for the present invention.Calculating whole absolutely percent error MAPE is
2.15% ± 0.95%.
Fig. 3 is (concentration to be kept to believe 10 groups of brain model analogue datas using method proposed by the present invention and Tarquin method
Cease constant, 10 groups of data line width distributions are 3.0Hz-28Hz, and by line width range, number consecutively is 1-10 from small to large) progress
The error of quantification treatment compares.It can be seen that when that is, Magnetic field inhomogeneity degree increases, comparing other methods, this hair when line width increases
Bright method shows preferable accuracy and stability.
Fig. 4 is to adopt data in fact to 10 groups of brain model solution using method proposed by the present invention (each metabolite concentration is it is known that line
Wide distribution is 1.1Hz to 25.1Hz) carry out each metabolite concentration quantized result that quantification treatment obtains.Brain model data exist
It is collected on Varian Varian 7.0T magnetic resonance tool, sampling number 8192, TE1=8ms, TE2=7ms, TR=
3s.Whole absolutely percent error MAPE is 2.39% ± 1.41%.
Fig. 5 is to adopt data in fact to 10 groups of brain model solution using Tarquin method (each metabolite concentration is it is known that line width point
Cloth range is 1.1Hz to 25.1Hz) carry out each metabolite concentration quantized result that quantification treatment obtains.Calculating global error is
5.12% ± 2.82%.Whole absolutely percent error MAPE is 5.12% ± 2.82%.
Described above, only the present invention preferably implements example, cannot limit the scope of implementation of the present invention according to this.I.e. according to this
Equivalent changes and modifications made by patent of invention range and description, should still be within the scope of the present invention.
Claims (6)
1. the non-uniform magnetic field hypencephalon metabolin spectrogram quantization method based on neural network algorithm, it is characterised in that including following step
It is rapid:
1) it utilizes free induction decay signal simulation softward FID-A and selects common detectable biological living brainmetabolite simulation base
Collection carries out simulating each brainmetabolite PRESS experiment, and it is tired to carry out band coefficient weighting by the analog signal to each brainmetabolite
Add, fixed water signal weighting coefficient obtains final biological brainmetabolite analog signal；
2) training dataset is established by the analogue data of the different line widths of generation, different signal-to-noise ratio, different brainmetabolite concentration, with
Machine takes out the fraction of training data concentration respectively as verifying collection and test set, corresponding to each analogue data by generation
Brainmetabolite weighting coefficient as label data；
3) data prediction is carried out to above-mentioned all data sets, pretreatment includes one-dimensional Fourier transform, normalization, data sanction
It cuts；
4) Artificial Neural Network Structures are designed, the hyper parameter for choosing neural network structure is collected using verifying, obtains quantitative model；
5) test verification is carried out to quantitative model using test data: after pre-processing to test data, inputs quantitative model,
Its i.e. exportable corresponding each metabolite concentration information；The pretreatment includes one-dimensional Fourier transform, normalization, data sanction
It cuts.
2. the non-uniform magnetic field hypencephalon metabolin spectrogram quantization method according to claim 1 based on neural network algorithm,
It is characterized by: the brainmetabolite includes but is not limited to Cr, NAA, Cho, Gln, GABA, Ins, Tau, Lac, Glu.
3. the non-uniform magnetic field hypencephalon metabolin spectrogram quantization method according to claim 1 based on neural network algorithm,
It is characterized by: the coefficient weighting of the analog signal of each brainmetabolite refers to: according to each brainmetabolite in brain in step 1
The difference of concentration in model solution determines weight of each brainmetabolite analog signal in biological brainmetabolite analog signal,
It is weighted to obtain the coefficient of the analog signal of each brainmetabolite.
4. the non-uniform magnetic field hypencephalon metabolin spectrogram quantization method according to claim 1 based on neural network algorithm,
Refer to it is characterized by: fixing the weighting of water signal coefficient in step 1, according to concentration of the water in brain model solution, determines that water is believed
Weight number in biological brainmetabolite analog signal, to obtain the coefficient weighting of water signal.
5. the non-uniform magnetic field hypencephalon metabolin spectrogram quantization method according to claim 1 based on neural network algorithm,
It is characterized by: the test data includes that data are adopted in analogue data and experiment in fact in step 5；Wherein the real data of adopting of experiment need
It will be by pretreatment.
6. the non-uniform magnetic field hypencephalon metabolin spectrogram quantization method according to claim 1 based on neural network algorithm,
It is characterized by: the data cut the dimension for referring to and reducing data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910625103.XA CN110503197A (en) | 2019-07-11 | 2019-07-11 | Non-uniform magnetic field hypencephalon metabolite concentration quantization method based on neural network algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910625103.XA CN110503197A (en) | 2019-07-11 | 2019-07-11 | Non-uniform magnetic field hypencephalon metabolite concentration quantization method based on neural network algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110503197A true CN110503197A (en) | 2019-11-26 |
Family
ID=68585278
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910625103.XA Pending CN110503197A (en) | 2019-07-11 | 2019-07-11 | Non-uniform magnetic field hypencephalon metabolite concentration quantization method based on neural network algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110503197A (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180096229A1 (en) * | 2016-01-26 | 2018-04-05 | Università della Svizzera italiana | System and a method for learning features on geometric domains |
| 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 |
-
2019
- 2019-07-11 CN CN201910625103.XA patent/CN110503197A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180096229A1 (en) * | 2016-01-26 | 2018-04-05 | Università della Svizzera italiana | System and a method for learning features on geometric domains |
| 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 |
Non-Patent Citations (1)
| Title |
|---|
| 段博: "非均匀磁场下核磁共振定域谱校正及量化研究", 《中国硕士学位论文全文数据库》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Benoit-Cattin et al. | The SIMRI project: a versatile and interactive MRI simulator | |
| CN105929024B (en) | Concrete defect intellectualized detection and quantitative identification method | |
| Poullet et al. | MRS signal quantitation: a review of time-and frequency-domain methods | |
| CN101088022B (en) | Magnetic resonance imaging system and method | |
| Li et al. | A subspace approach to spectral quantification for MR spectroscopic imaging | |
| US10571542B2 (en) | Time-domain MRI | |
| US20210134028A1 (en) | Method and system for simultaneous quantitative multiparametric magnetic resonance imaging (mri) | |
| Reynolds et al. | An algorithm for the automated quantitation of metabolites in in vitro NMR signals | |
| Wang et al. | Magnetic resonance fingerprinting with compressed sensing and distance metric learning | |
| Thompson et al. | Response of metabolites with coupled spins to the STEAM sequence | |
| Ginsburger et al. | Improving the realism of white matter numerical phantoms: a step toward a better understanding of the influence of structural disorders in diffusion MRI | |
| CN108132274A (en) | Echo-planar imaging is without reference scan pattern distortion antidote under non-uniform magnetic field | |
| CN109870729A (en) | Deep neural network magnetic resonance signal noise-eliminating method based on discrete cosine transform | |
| US20190353730A1 (en) | System and method for generating a magnetic resonance fingerprinting dictionary using semi-supervised learning | |
| CN110109037A (en) | A kind of inversion method of the NMR signal towards more relaxation components | |
| Shamaei et al. | Physics-informed deep learning approach to quantification of human brain metabolites from magnetic resonance spectroscopy data | |
| Ljungberg et al. | Silent T1 mapping using the variable flip angle method with B1 correction | |
| Hiltunen et al. | Quantification of human brain metabolites from in vivo1H NMR magnitude spectra using automated artificial neural network analysis | |
| US20120105060A1 (en) | Use Of Strongly Modulating Pulses In MRI For Providing Chemical Shift Selective Flip Angles | |
| Clarke et al. | Universal dynamic fitting of magnetic resonance spectroscopy | |
| CN110503197A (en) | Non-uniform magnetic field hypencephalon metabolite concentration quantization method based on neural network algorithm | |
| CN110490063A (en) | Brain model localization spectrum correction method under non-uniform field based on neural network algorithm | |
| White et al. | Optimal diffusion MRI acquisition for fiber orientation density estimation: an analytic approach | |
| Endres et al. | Phase distribution graphs for fast, differentiable, and spatially encoded Bloch simulations of arbitrary MRI sequences | |
| Shamaei et al. | A Wavelet Scattering Convolutional Network for Magnetic Resonance Spectroscopy Signal Quantitation. |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20191126 |
|
| WD01 | Invention patent application deemed withdrawn after publication |