CN108010100B

CN108010100B - Single-scanning magnetic resonance quantitative T based on residual error network2Imaging reconstruction method

Info

Publication number: CN108010100B
Application number: CN201711287890.9A
Authority: CN
Inventors: 蔡淑惠; 张俊; 蔡聪波; 廖璞; 曾坤
Original assignee: Xiamen University
Current assignee: Xiamen University
Priority date: 2017-12-07
Filing date: 2017-12-07
Publication date: 2020-02-18
Anticipated expiration: 2037-12-07
Also published as: CN108010100A

Abstract

Single-scanning magnetic resonance quantitative T based on residual error network₂An imaging reconstruction method relates to a magnetic resonance imaging method. Using 4 small angles with the same deflection angleExcitation pulses, each excitation pulse being followed by an evolution time such that the transverse relaxation time T of each echo signal₂Different. A shifting gradient in both the frequency-encoding and phase-encoding dimensions is applied after each excitation pulse so that the signals generated by different excitation pulses differ in their position in k-space. A plurality of echo signals with different transverse relaxation times are obtained in one sample. Then, the sampling signal is input into a trained residual error network after normalization, zero filling and fast Fourier transform to be reconstructed to obtain quantitative T₂And (4) an image. The training data for the residual network is derived from the simulated data. Firstly, randomly generating a template, then simulating an experimental environment to sample to obtain an input image of a network, using the template as a label, and training to obtain a mapping relation between the input image and an output image.

Description

Single-scanning magnetic resonance quantitative T based on residual error network2Imaging reconstruction method

Technical Field

The invention relates to a magnetic resonance imaging method, in particular to a single-scanning magnetic resonance quantitative T based on a residual error network₂Imaging (T)₂mapping) reconstruction method.

Background

Magnetic resonance parametric imaging (e.g. T)₁Imaging, T₂Imaging and T₂ ^*Imaging) can be readily used to provide quantitative information for characterizing specific tissue characteristics^[1]. Quantitative imaging has many prominent features, the most obvious of which is the elimination of tissue property independent effects such as operator dependence, scan parameter differences, magnetic field spatial variations, and image scaling^[2]. Quantitative T₂Imaging has a wide range of clinical applications, involving early diagnosis of neurodegenerative diseases^[3]Liver iron overload measurement^[4]Myocardial infarction assessment^[5]Quantification of labeled cells^[6]Diagnosis of multiple sclerosis and epilepsy^[7]And the like. Quantitative T₂Imaging is gaining increasing attention in clinical Magnetic Resonance Imaging (MRI). Albeit upon transverse relaxationInter T₂Can be obtained by fitting spin echo MRI data of multiple different echo Times (TE), but the longer scan time of spin echo MRI allows for quantitative T in the clinic₂The image is challenging^[8]. Furthermore, the longer scan time also allows for spin-echo MRI based quantification T₂Images are susceptible to motion artifacts^[9]. Single-scan quantitative T-imaging (EPI) based on multi-echo planar imaging has been proposed₂The imaging method shows great advantages in functional magnetic resonance imaging^[10,11]. However, the resulting image is susceptible to more signal attenuation and ghost artifacts than single echo sampling^[12,13]. Other fast quantitative T₂Imaging methods, e.g. gradient spin echo (GraSE) schemes^[14]Still, a multi-echo spin-echo (MSE) strategy is employed, typically requiring several minutes^[15,16]。

Methods of planar imaging by overlapping echo separation (OLED) have been proposed^[17]Can obtain high-quality quantitative T in single scan₂Images whose temporal resolution is comparable to conventional single-scan EPI images. In addition, OLED planar imaging also shows contrast to motion artifacts and non-ideal B₁Stronger resistance of the field. However, the previous sequence contained only two excitation pulses, and thus a measurable T₂The range is very limited, which results in having a large T₂The effect is poor in areas of range, such as cerebrospinal fluid (CSF). Therefore, we improve the OLED sequence with four excitation pulses to achieve a larger T₂The measurement range. However, it is difficult to separate four overlapping echo signals by the conventional method.

Deep learning, a technique for discovering distributed features of data by combining lower level features to form more abstract higher level representation attribute classes or features, has shown explosive popularity with the availability of powerful GPUs in recent years^[18]. In particular, Convolutional Neural Networks (CNN) have caused super-resolution (supe) of imagesr-resolution, SR) reconstruction^[19]. Different network models, e.g. convolutional neural networks^[20]Residual network (ResNet)^[21]Deep-recursive convolutional network (DRCN)^[22]An efficient sub-pixel convolutional neural network (ESPCN)^[23]And generating a antagonistic network (GAN)^[24]Have been applied to obtain high resolution images. Convolutional neural networks are becoming increasingly popular in medical imaging analysis of various problems.

Reference documents:

[1]B.Zhao,F.Lam,and Z.P.Liang,"Model-based MR parameter mapping withsparsity constraints:parameter estimation and performance bounds,"IEEETransactions on Medical Imaging,vol.33,no.9,pp.1832-1844,2014.

[2]H.L.Margaret Cheng,N.Stikov,N.R.Ghugre,and G.A.Wright,"Practicalmedical applications of quantitative MR relaxometry,"Journal of MagneticResonance Imaging,vol.36,no. 4,pp.805-824,2012.

[3]S.Baudrexel,L.Nürnberger,U.Rüb,C.Seifried,J.C.Klein,T.Deller,H.Steinmetz,R. Deichmann,and R.Hilker,"Quantitative mapping of T₁and T₂*discloses nigral and brainstem pathology in early Parkinson's disease,"Neuroimage,vol.51,no.2,pp.512-520,2010.

[4]J.S.Hankins,M.B.McCarville,R.B.Loeffler,M.P.Smeltzer,M.Onciu,F.A.Hoffer,C.-S. Li,W.C.Wang,R.E.Ware,and C.M.Hillenbrand,"R₂*magneticresonance imaging of the liver in patients with iron overload,"Blood,vol.113,no.20,pp.4853-4855,2009.

[5]D.R.Messroghli,K.Walters,S.Plein,P.Sparrow,M.G.Friedrich,J.P.Ridgway,and M.U. Sivananthan,"Myocardial T₁mapping:application topatients with acute and chronic myocardial infarction,"Magnetic Resonance inMedicine,vol.58,no.1,pp.34-40,2007.

[6]W.Liu,H.Dahnke,J.Rahmer,E.K.Jordan,and J.A.Frank,"Ultrashort T₂*relaxometry for quantitation of highly concentrated superparamagnetic ironoxide(SPIO)nanoparticle labeled cells,"Magnetic Resonance in Medicine,vol.61,no.4,pp.761-766,2009.

[7]T.N.Townsend,N.Bernasconi,G.B.Pike,and A.Bernasconi,"Quantitativeanalysis of temporal lobe white matter T₂relaxation time in temporal lobeepilepsy,"Neuroimage,vol.23,no. 1,pp.318-324,2004.

[8]D.Ma,V.Gulani,N.Seiberlich,K.Liu,J.L.Sunshine,J.L.Duerk,andM.A.Griswold, "Magnetic resonance fingerprinting,"Nature,vol.495,no.7440,pp.187-193,2013.

[9]D.Le Bihan,C.Poupon,A.Amadon,and F.Lethimonnier,"Artifacts andpitfalls in diffusion MRI,"Journal of Magnetic Resonance Imaging,vol.24,no.3,pp.478-488,2006.

[10]S.Posse,S.Wiese,D.Gembris,K.Mathiak,C.Kessler,M.-L.Grosse-Ruyken,B. Elghahwagi,T.Richards,S.R.Dager,and V.G.Kiselev,"Enhancement of BOLD-contrast sensitivity by single-shot multi-echo functional MR imaging,"Magnetic Resonance in Medicine, vol.42,no.1,pp.87-97,1999.

[11]G.Hagberg,I.Indovina,J.Sanes,and S.Posse,"Real‐timequantification of T₂* changes using multiecho planar imaging and numericalmethods,"Magnetic Resonance in Medicine, vol.48,no.5,pp.877-882,2002.

[12]J.W.Evans,P.Kundu,S.G.Horovitz,and P.A.Bandettini,"Separatingslow BOLD from non-BOLD baseline drifts using multi-echo fMRI,"Neuroimage,vol.105,pp.189-197,2015.

[13]A.D.Halai,S.R.Welbourne,K.Embleton,and L.M.Parkes,"A comparisonof dual gradient‐echo and spin‐echo fMRI of the inferior temporal lobe,"HumanBrain Mapping,vol. 35,no.8,pp.4118-4128,2014.

[14]A.M.Sprinkart,J.A.Luetkens,F.J.Doerner,J.Gieseke,B.Schnackenburg,G. Schmitz,D.Thomas,R.Homsi,and W.Block,"Gradient Spin Echo(GraSE)imaging for fast myocardial T₂mapping,"Journal of CardiovascularMagnetic Resonance,vol.17,no.1,pp.12-20, 2015.

[15]T.J.Sumpf,A.Petrovic,M.Uecker,F.Knoll,and J.Frahm,"Fast T₂mapping with improved accuracy using undersampled spin-echo MRI and model-based reconstructions with a generating function,"IEEE Transactions onMedical Imaging,vol.33,no.12,pp.2213-2222,2014.

[16]N.Ben‐Eliezer,D.K.Sodickson,and K.T.Block,"Rapid and accurate T₂mapping from multi–spin‐echo data using Bloch‐simulation‐basedreconstruction,"Magnetic Resonance in Medicine,vol.73,no.2,pp.809-817,2015.

[17]C.B.Cai,Y.Q.Zeng,Y.C.Zhuang,S.H.Cai,L.Chen,X.H.Ding,L.J.Bao,J.H.Zhong, and Z.Chen,"Single-shot T₂mapping through Overlapping-echoDetachment(OLED)Planar Imaging,"IEEE Transactions on Biomedical Engineering,vol.64,no.10,pp.2450-2461,2017.

[18]S.S.Wang,Z.H.Su,L.Ying,X.Peng,S.Zhu,F.Liang,D.G.Feng,and D.Liang,"Accelerating magnetic resonance imaging via deep learning,"in2016 IEEE 13thInternational Symposium on Biomedical Imaging(ISBI),pp.514-517,2016.

[19]C.F.Baumgartner,O.Oktay,and D.Rueckert,"Fully ConvolutionalNetworks in Medical Imaging:Applications to Image Enhancement andRecognition,"in Deep Learning and Convolutional Neural Networks for MedicalImage Computing,pp.159-179,2017.

[20]C.Dong,C.C.Loy,K.M.He,and X.O.Tang,"Learning a deep convolutionalnetwork for image super-resolution,"in 13th European Conference on ComputerVision(ECCV),pp.184-199, 2014.

[21]K.M.He,X.Y.Zhang,S.Q.Ren,and J.Sun,"Deep residual learning forimage recognition,"in Proceedings of the IEEE Conference on Computer Visionand Pattern Recognition (CVPR),pp.770-778,2016.

[22]J.Kim,J.Kwon Lee,and K.Mu Lee,"Deeply-recursive convolutionalnetwork for image super-resolution,"in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition(CVPR),pp.1637-1645,2016.

[23]W.Z.Shi,J.Caballero,F.Huszár,J.Totz,A.P.Aitken,R.Bishop,D.Rueckert,and Z.H. Wang,"Real-time single image and video super-resolutionusing an efficient sub-pixel convolutional neural network,"in Proceedings ofthe IEEE Conference on Computer Vision and Pattern Recognition(CVPR),pp.1874-1883,2016.

[24]C.Ledig,L.Theis,F.Huszár,J.Caballero,A.Cunningham,A.Acosta,A.Aitken,A. Tejani,J.Totz,and Z.Wang,"Photo-realistic single image super-resolution using a generative adversarial network,"arXiv preprint arXiv:1609.04802,2016.

disclosure of Invention

The invention aims to obtain four echo signals in a single scanning and then reconstruct reliable quantitative T by using a deep learning method₂Single-scanning magnetic resonance quantitative T based on residual error network of image₂An imaging reconstruction method.

The invention comprises the following steps:

1) preparing an experimental sample, placing the sample to be tested on an experimental bed, fixing the sample, and conveying the experimental bed with the sample into a detection cavity of a magnetic resonance imager;

2) opening imager operation software on an operation table of a magnetic resonance imager, firstly carrying out region-of-interest positioning on an experimental sample, and then carrying out tuning, shimming, frequency correction and power correction;

3) importing a compiled OLED imaging sequence, and setting each parameter of a pulse sequence according to a specific experimental condition;

4) executing the OLED imaging sequence with the set parameters in the step 3), starting sampling, and executing the next step after data sampling is finished;

5) normalizing, zero filling and fast Fourier transform are carried out on the signals obtained in the step 4), and the signals of the k space are converted into an image domain;

6) generating a random template according to the characteristics of a sample to be tested, carrying out analog sampling on the random template by using analog software to obtain a k-space signal, and then carrying out normalization, zero filling and fast Fourier transform on the signal to obtain training data;

7) building a residual error network model by adopting a TensorFlow deep learning frame and Python, and setting relevant parameters for training;

8) adopting the data training network obtained in the step 6) until the data training network converges and reaches stability to obtain a trained network model, and then utilizing the trained network model to reconstruct the experimental data obtained in the step 5) to obtain reliable quantitative T₂And (4) an image.

In the step 3), the OLED imaging sequence has the structure that an excitation pulse with a flip angle of α and a displacement gradient G are sequentially arranged₁Excitation pulse with flip angle of α, shift gradient G₂Excitation pulse with flip angle of α, shift gradient G₃Excitation pulse with flip angle of α, shift gradient G₄Refocusing pulse with flip angle of β and sampling echo chain;

4 deflection angles α excitation pulses in combination with 4 shift gradients G in the frequency encoding dimension (RO) and the phase encoding dimension (PE)₁、G₂、G₃、G₄Shifting 4 echo signals in the center of k space, and combining 4 angle excitation pulses with the layer selection gradient of a layer selection dimension (SS) to perform layer selection;

the sampling echo chain is composed of gradient chains respectively acting on a frequency encoding dimension and a phase encoding dimension, the gradient chain of the frequency encoding dimension is composed of a series of positive and negative gradients, and the gradient chain of the phase encoding dimension is composed of a series of phase encoding gradients with equal areas;

before sampling an echo chain, applying bias gradients in the directions of a frequency encoding dimension and a phase encoding dimension respectively, wherein the bias gradient area of the frequency encoding dimension is half of the area of the first frequency encoding gradient, the directions are opposite, the bias gradient area of the phase encoding dimension is half of the area of all the phase encoding gradients of the phase encoding dimension, and the directions are opposite.

In step 5), analyzing the signals obtained in step 4), theoretically deriving the nuclear spin evolution under the action of an OLED imaging sequence, wherein the observable signal intensity after the action of a refocusing pulse with a flip angle of β is in direct proportion to the following expression:

wherein the content of the first and second substances,

gamma is the spin magnetic ratio of the nuclear spin,

the position of the nuclear spins in real space; from the above formula, the sampling period has 4 echo signals modulated by different phases, and the 4 different modulation phases are θ₄、(θ₃+θ₄)、(θ₂+θ₃+θ₄) And (theta)₁+θ₂+θ₃+θ₄)。

In the step 6), the random template is randomly generated in batches by using a computer according to the feature distribution of the experimental sample, and meanwhile, the complexity of the random template is ensured to be higher and can contain all the features of the experimental sample; in the simulation sampling process, the change of a real experimental environment is considered, and an unstable factor is added, so that the robustness of the network model to an unsatisfactory experimental environment is improved; the unstable factors include excitation pulse angle deviation, shift gradient deviation, noise, and the like.

In step 7), the residual network model mainly includes: the main structure of the network and the related training parameters, the objective function of the network model is:

wherein N is the number of training images in the same batch, f (-) is a training network, W and b are network parameters, x is an input image, y is a template corresponding to the input image_changeIs a matrix with a value less than a certain threshold in y set as the threshold, y_maskIs calculated using Canny for yAnd (5) image edge information obtained by the sub-calculation.

In the step 8), the trained network model is trained by using a random template, so that the generalization is strong, and the method can be suitable for reconstruction of various samples.

The present invention increases the number of excitation pulses from 2 to 4 to expand the T in the OLED sequence₂And improve the accuracy of the CSF region. The OLED image is reconstructed from the 4 overlapping echo signals using a deep learning technique, thereby circumventing the difficulties encountered with conventional methods.

The invention uses 4 excitation pulses with the same small-angle deflection angle, a period of evolution time is left after each excitation pulse, so that 4 echoes have different transverse relaxation times, a shift gradient is added after each excitation pulse, so that 4 echo signals are deviated in the center of a signal space (k space), and then the k space image data is reconstructed by using a deep learning method to obtain a quantitative T₂And (4) an image. The method can obtain a reliable quantitative T in a single scanning₂Images while ensuring a larger T₂The measurement range.

Drawings

FIG. 1 is a sequence diagram of a single scan OLED imaging sequence employed in this patent.

FIG. 2 is the reconstructed quantification T₂Residual network model for image usage. In fig. 2, the residual network model contains 3 sub-networks: an input network, a residual learning network and a reconstruction network. The input network takes the real and imaginary parts of the OLED image as input and is represented by a set of feature maps. The residual error learning network is a main component for solving the task of image reconstruction. The reconstruction network is used for reconstructing a plurality of characteristic maps into a quantitative T₂And (4) an image.

FIG. 3 shows the quantification T of 3 layers of human brain₂Imaging reconstruction results; in FIG. 3, the first column represents the amplitude map of SE, wherein 16 yellow circles represent selected regions of interest (ROI) numbered with numbers 1-16, respectively; the second column represents the original amplitude diagram of the OLED, i.e. the input of the network; the third column shows the quantification T of the OLED reconstruction₂The image is the output of the network; the fourth column indicates SE fittingQuantitative of (2)₂Imaging reference images.

Fig. 4 shows the statistics of 16 ROIs. In fig. 4, a is an OLED and b is SE.

Detailed Description

The invention is further illustrated by the following figures and specific examples.

The invention provides a single-scanning magnetic resonance quantitative T based on a residual error network₂The imaging reconstruction method comprises the following steps in the specific implementation process:

(1) preparing an experimental sample, placing the sample to be detected on an experimental bed and fixing, and sending the experimental bed with the sample into a detection cavity of a magnetic resonance imager.

(2) And opening imager operation software on an operation table of the magnetic resonance imager, firstly carrying out region-of-interest positioning on the experimental sample, and then carrying out tuning, shimming, frequency correction and power correction.

(3) And importing a compiled OLED imaging sequence, and setting various parameters of the pulse sequence according to specific experimental conditions.

The OLED imaging sequence sequentially comprises an excitation pulse with a flip angle of α and a displacement gradient G₁Excitation pulse with flip angle of α, shift gradient G₂Excitation pulse with flip angle of α, shift gradient G₃Excitation pulse with flip angle of α, shift gradient G₄Refocusing pulse with flip angle β, sampling echo train.

The four small angle (α) excitation pulses incorporate four shift gradients G in the frequency encoding dimension (RO) and the phase encoding dimension (PE)₁、G₂、G₃、G₄Four echo signals are shifted in the center of k space, and four small-angle excitation pulses are combined with the layer selection gradient of a layer selection dimension (SS) to perform layer selection.

The sampling echo chain is composed of gradient chains respectively acting on a frequency encoding dimension and a phase encoding dimension, the gradient chain of the frequency encoding dimension is composed of a series of positive and negative gradients, and the gradient chain of the phase encoding dimension is composed of a series of phase encoding gradients with equal areas.

(4) And (4) executing the OLED imaging sequence with the set parameters in the step (3), starting sampling, and executing the next step after data sampling is completed.

(5) And (4) carrying out normalization, zero filling and fast Fourier transform on the signals obtained in the step (4) to convert the signals of the k space into an image domain.

The obtained signal is analyzed, and the nuclear spin evolution under the action of an OLED imaging sequence is theoretically deduced, the observable signal intensity after the action of a refocusing pulse with the flip angle of β is in direct proportion to the following expression:

gamma is the spin magnetic ratio of the nuclear spin,

the position of the nuclear spins in real space; from the above formula, the sampling period has four echo signals modulated by different phases, and the four different modulation phases are theta₄、(θ₃+θ₄)、(θ₂+θ₃+θ₄) And (theta)₁+θ₂+θ₃+θ₄)。

(6) Generating a random template according to the characteristics of an experimental sample, carrying out analog sampling on the template by using SPROM software developed by a small group of people to obtain a k-space signal, and then carrying out normalization, zero filling and fast Fourier transform on the signal to obtain training data.

The random template is generated randomly in batches by using a computer according to the characteristic distribution of the experimental sample, and meanwhile, the complexity of the template is ensured to be higher, and all the characteristics of the experimental sample can be contained. In the simulation sampling process, the change of a real experimental environment is considered, and some unstable factors such as excitation pulse angle deviation, shift gradient deviation, noise and the like are added to improve the robustness of the network model to an unsatisfactory experimental environment.

(7) And (5) building a residual error network model by using a TensorFlow deep learning frame and Python, and setting relevant parameters for training.

The residual error network model mainly comprises: the body structure of the network and the associated training parameters. The objective function of the network model is:

where N is the number of training images in the same batch, f (-) is the training network, W and b are the network parameters, x is the input image, y is the template corresponding to the input image, y is the number of training images in the same batch_changeIs a matrix with a value of y less than 0.06 set to 0.06_maskThe image edge information is obtained by Canny operator for y.

(8) Training a network by using the data obtained in the step (6) until the network converges and reaches a stable state to obtain a trained network model, and then reconstructing the experimental data obtained in the step (5) to obtain reliable quantitative T₂And (4) an image.

The trained network model is trained by using the random template, so that the generalization is strong, and the method can be suitable for reconstruction of various samples.

Specific examples are given below:

quantification of T with single-scan magnetic resonance based on residual error network₂The imaging reconstruction method was performed in human brain experiments to verify the feasibility of the invention. The experiment was performed under a human nmr 3T imager. On the operation table of the magnetic resonance imager, corresponding operation software in the imager is opened, the region of interest of the imaging object is firstly positioned, and then tuning, shimming, power and frequency correction are carried out. To evaluate the effectiveness of the method to obtain images, SE imaging experiments were performed as a comparison under the same environment. Then introduced intoThe compiled OLED imaging sequence (as shown in FIG. 1) sets the parameters of the pulse sequence according to the specific experimental situation, the experimental parameters of this embodiment are set as follows, the FOV of the imaging field is 22cm × 22cm, the excitation time of the 15 ° excitation pulse is 3ms, the excitation time of the 180 ° refocusing pulse is 3ms, the first echo time is 57.8ms, the second echo time is 83.9ms, the third echo time is 135.9ms, the fourth echo time is 162.1ms, the sampling points of the frequency coding dimension and the phase coding dimension are 192, after the above experimental parameters are set, the sampling is directly started, in FIG. 1, α is the excitation pulse flip angle, β is the refocusing pulse flip angle, G is the flip angle of the refocusing pulse β, and the parameters of the pulse sequence are set according to₁、G₂、G₃And G₄Four shift gradients in frequency encoding dimension and phase encoding dimension; g_crSelecting destruction gradients of 3 dimensions for the frequency encoding dimension, the phase encoding dimension and the level; TE₁、TE₂、TE₃And TE₄The time lengths of the 4 echoes respectively; the Ecoh1, Ecoh2, Ecoh3, and Ecoh4 are the center positions of 4 echo signals, respectively.

And (5) after the data sampling is finished, reconstructing the data according to the steps (5) to (8). The residual error network model used for reconstruction is shown in fig. 2, and comprises an input network, a residual error learning network and a reconstruction network. The input network takes the real part and the imaginary part of the OLED image as input and uses a group of characteristic maps for representation, the residual error learning network is a main component for solving the image reconstruction task, and the reconstruction network is used for reconstructing a plurality of characteristic maps into a quantitative T₂And (4) an image. Reconstructed quantitative T₂The image is shown in fig. 3, (a) is an anatomical image of a human brain acquired by a Spin Echo (SE) sequence, in which 16 regions marked by yellow circles are regions of interest (ROIs), which are numbered with numerals 1 to 16, respectively. (b) Is the original amplitude map acquired by the OLED sequence, in which the diagonal stripes are caused by overlapping the four echo signals. (c) Quantitative T derived from reconstruction of OLED data₂And (4) an image. (d) Is a quantitative T fit by SE₂And (4) an image. T for 16 interesting regions in the figure₂The values are subjected to quantitative statistics, i.e. T for each region of interest₂The values were averaged to obtain FIG. 4. The quantification T obtained with the OLED sequence can be seen in FIG. 3₂The image as a whole agreed with SE, and the statistical results in FIG. 4 also indicate T for 16 ROI regions₂The values are all of better accuracy.

The invention uses 4 small angle excitation pulses with the same deflection angle, and there is a period of evolution time after each excitation pulse, so that the transverse relaxation time T of each echo signal₂Different. At the same time, a shift gradient in both the frequency-encoding and phase-encoding dimensions is applied after each excitation pulse, so that the signals generated by different excitation pulses differ in their position in k-space. In this way, a plurality of echo signals with different transverse relaxation times are obtained in one sample. Then, the sampling signal is input into a trained residual error network after normalization, zero filling and fast Fourier transform to be reconstructed to obtain quantitative T₂And (4) an image. The training data for the residual network is derived from the simulated data. Firstly, a template is randomly generated, then an experiment environment is simulated for sampling to obtain an input image of a network, the template is used as a label, and a mapping relation between the input image and an output image is obtained through training. The method provided by the invention can obtain reliable quantitative T in single scanning₂And (4) an image.

Claims

1. Single-scanning magnetic resonance quantitative T based on residual error network₂An imaging reconstruction method, characterized by comprising the steps of:

3) introducing a compiled OLED imaging sequence, and setting parameters of the pulse sequence according to specific experimental conditions, wherein the structure of the OLED imaging sequence comprises an excitation pulse with a flip angle of α and a shift gradient G₁α flip angle excitation pulseImpact and shift gradient G₂Excitation pulse with flip angle of α, shift gradient G₃Excitation pulse with flip angle of α, shift gradient G₄Refocusing pulse with flip angle of β, sampling echo chain, 4 excitation pulses with flip angle of α, and 4 shift gradients G in frequency coding dimension and phase coding dimension₁、G₂、G₃、G₄Shifting 4 echo signals in the center of k space, and performing slice selection by combining 4 excitation pulses with flip angles of α and a slice selection gradient of a slice selection dimension;

7) building a residual error network model by adopting a TensorFlow deep learning frame and Python, and setting relevant parameters for training; the residual network model comprises: the main structure of the network and the related training parameters, the objective function of the network model is:

wherein N is the number of training images in the same batch, f (-) is a training network, W and b are network parameters, x is an input image, y is a template corresponding to the input image_changeIs a matrix with a value less than a certain threshold in y set as the threshold, y_maskImage edge information obtained by Canny operator for y;

8) training the network by adopting the data obtained in the step 6) until the network is converged and stable to obtain a trained network model, and then reconstructing the experimental data obtained in the step 5) by utilizing the trained network model to obtain the network modelBy quantitative T₂And (4) an image.

2. Single-scan magnetic resonance quantitative T based on residual error network as claimed in claim 1₂The imaging reconstruction method is characterized in that in step 3), the sampling echo chain is composed of gradient chains respectively acting on a frequency encoding dimension and a phase encoding dimension, the gradient chain of the frequency encoding dimension is composed of a series of positive and negative gradients, and the gradient chain of the phase encoding dimension is composed of a series of gradients of the phase encoding dimension with equal areas.

3. Single-scan magnetic resonance quantitative T based on residual error network as claimed in claim 1₂The imaging reconstruction method is characterized in that in step 3), before sampling an echo chain, offset gradients are respectively applied to the directions of a frequency encoding dimension and a phase encoding dimension, the area of the offset gradient of the frequency encoding dimension is half of the area of the first frequency encoding gradient, the directions are opposite, the area of the offset gradient of the phase encoding dimension is half of the area of all the phase encoding gradients of the phase encoding dimension, and the directions are opposite.

4. Single-scan magnetic resonance quantitative T based on residual error network as claimed in claim 1₂The imaging reconstruction method is characterized in that in the step 5), the signals obtained in the step 4) are analyzed, the theoretical derivation is carried out on the nuclear spin evolution under the action of an OLED imaging sequence, and the observed signal intensity after the action of a refocusing pulse with the flip angle of β is in direct proportion to the following expression:

wherein the content of the first and second substances,

gamma is the spin magnetic ratio of the nuclear spin,

the position of the nuclear spins in real space; obtaining the sampling period from the above equationThere are 4 echo signals modulated with different phases, the 4 different modulation phases being theta₄、(θ₃+θ₄)、(θ₂+θ₃+θ₄) And (theta)₁+θ₂+θ₃+θ₄)。

5. Single-scan magnetic resonance quantitative T based on residual error network as claimed in claim 1₂The imaging reconstruction method is characterized in that in the step 6), the random template is randomly generated in batches by using a computer according to the characteristic distribution of the experimental sample, the random template contains all the characteristics of the experimental sample, and an unstable factor is added to improve the robustness of the network model to an unsatisfactory experimental environment; the instability factors include excitation pulse angle deviation, shift gradient deviation, and noise.

6. Single-scan magnetic resonance quantitative T based on residual error network as claimed in claim 1₂The imaging reconstruction method is characterized in that in the step 8), the trained network model is suitable for reconstruction of various samples.