CN115471580A - Physical intelligent high-definition magnetic resonance diffusion imaging method - Google Patents
Physical intelligent high-definition magnetic resonance diffusion imaging method Download PDFInfo
- Publication number
- CN115471580A CN115471580A CN202211040703.8A CN202211040703A CN115471580A CN 115471580 A CN115471580 A CN 115471580A CN 202211040703 A CN202211040703 A CN 202211040703A CN 115471580 A CN115471580 A CN 115471580A
- Authority
- CN
- China
- Prior art keywords
- data
- network
- module
- magnetic resonance
- diffusion
- 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
- 238000009792 diffusion process Methods 0.000 title claims abstract description 75
- 238000003384 imaging method Methods 0.000 title claims abstract description 16
- 230000033001 locomotion Effects 0.000 claims abstract description 37
- 238000012549 training Methods 0.000 claims abstract description 34
- 238000004088 simulation Methods 0.000 claims abstract description 30
- 230000035945 sensitivity Effects 0.000 claims abstract description 22
- 238000000034 method Methods 0.000 claims abstract description 21
- 230000005284 excitation Effects 0.000 claims abstract description 13
- 239000011159 matrix material Substances 0.000 claims description 16
- 239000000126 substance Substances 0.000 claims description 14
- 230000006870 function Effects 0.000 claims description 11
- 238000013461 design Methods 0.000 claims description 9
- 238000002597 diffusion-weighted imaging Methods 0.000 claims description 9
- 238000000354 decomposition reaction Methods 0.000 claims description 6
- 238000013135 deep learning Methods 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 230000015572 biosynthetic process Effects 0.000 description 2
- 230000004069 differentiation Effects 0.000 description 2
- 229940079593 drug Drugs 0.000 description 2
- 239000003814 drug Substances 0.000 description 2
- 235000020638 mussel Nutrition 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 230000000717 retained effect Effects 0.000 description 2
- 238000003786 synthesis reaction Methods 0.000 description 2
- 102100026933 Myelin-associated neurite-outgrowth inhibitor Human genes 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000002059 diagnostic imaging Methods 0.000 description 1
- 239000006185 dispersion Substances 0.000 description 1
- 230000009545 invasion Effects 0.000 description 1
- 238000002595 magnetic resonance imaging Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000009827 uniform distribution Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T11/00—2D [Two Dimensional] image generation
- G06T11/003—Reconstruction from projections, e.g. tomography
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T11/00—2D [Two Dimensional] image generation
- G06T11/003—Reconstruction from projections, e.g. tomography
- G06T11/008—Specific post-processing after tomographic reconstruction, e.g. voxelisation, metal artifact correction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/10—Image enhancement or restoration using non-spatial domain filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/80—Geometric correction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20048—Transform domain processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Radiology & Medical Imaging (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- High Energy & Nuclear Physics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Pathology (AREA)
- Heart & Thoracic Surgery (AREA)
- Medical Informatics (AREA)
- Surgery (AREA)
- Animal Behavior & Ethology (AREA)
- Public Health (AREA)
- Veterinary Medicine (AREA)
- Magnetic Resonance Imaging Apparatus (AREA)
Abstract
A physical intelligent high-definition magnetic resonance diffusion imaging method comprises the following steps: 1) Obtaining the b value of 0mm/s of single excitation or multiple excitations of multiple channels 2 The magnetic resonance image with the phase is estimated to obtain the channel sensitivity; 2) Obtaining a simulated motion phase according to a polynomial phase model of rigid motion; 3) Generation of large amounts of multi-shot diffusion-weighted image data using phased magnetic resonance images, channel sensitivities, and simulated motion phases as training for intelligent reconstruction networksAnd (5) practicing data. 4) The method comprises the steps of constructing an intelligent reconstruction network comprising a plurality of iteration blocks, wherein each iteration block comprises a low-rank module, a sparse module and a data checking module. The last iteration block also contains a denoising module. Training an intelligent reconstruction network by using simulation data; 5) Acquiring multi-excitation diffusion weighted data to be reconstructed; 6) And reconstructing multi-excitation diffusion weighted data by using the trained intelligent reconstruction network to obtain a reconstructed image.
Description
Technical Field
The invention relates to a physical intelligent high-definition magnetic resonance diffusion imaging method, in particular to a method for generating multi-excitation diffusion weighted imaging training data by utilizing polynomial simulation and an intelligent reconstruction network based on sparseness and low rank.
Background
Diffusion weighted imaging (Diffusion weighted imaging) is a way to assess the function and microstructure of human molecules and can detect the Diffusion motion of water molecules in tissues without invasion (V.Baliyan et al., "Diffusion weighted imaging: techniques and applications," World Journal of Radiology,8,785,2016). Multi-shot planar echo imaging techniques have the ability to improve resolution and reduce low distortion in diffusion weighted applications (h.an, x.ma, z.pan, h.guo, e.y.p. Lee, "Qualitative and quantitative compatibility of image quality between shot echo and inter-shot echo dispersion-weighted imaging in the field of specimen analysis," European radiology,30,1876-1884, 2020). However, there is a severe phase error between different shots, resulting in severe motion artifacts (A.W. Anderson, J.C. Gore, "Analysis and correction of motion artifacts in dispersion weighted imaging," Magnetic Resonance in Medicine,32,379-387, 1994).
In recent years, many reconstruction methods based on low-rank iterative models are used to correct motion phases between different excitations, and achieve reconstruction without motion artifacts. The MUSSELS constructs a structured hankel matrix by establishing phase annihilation relations among images of different excitations, and constrains the low rank of the matrix to realize reconstruction (m.mani, m.jacob, d.keley, v. Magnetic, "Multi-shot sensitivity-encoded direction data recovery using structured low-rank Matrix (MUSSELS)," Magnetic Resonance in Medicine,78,494-507, 2017). PLRHM constructs a structured low-rank matrix using smooth Phase priors of Magnetic Resonance images, and achieves reconstruction by constraining partial larger singular value sums of the low-rank matrix (y.huang et al., "Phase-constrained reconstruction for high-resolution multi-shot differentiation weighted image," Journal of Magnetic Resonance,312 106690, 2020. The PAIR uses the structural detail information of the image domain and the low rank information of the Fourier space to realize the reconstruction of multi-excitation diffusion weighting data by alternately and iteratively reconstructing the amplitude and the phase (C.Qian et al., "A-affected phase and magnetic reconstruction for advanced diffusion-weighted imaging," arXiv prediction, arXiv: 2203.14559.2022).
Recently, deep learning methods have shown great potential in multi-shot diffusion weighted Imaging (Aggarwal. H.K., M. Mani, M.Jacob, "MoDL-MUSSELS: model-based deep learning for multi-shot sensitivity-encoded differentiation MRI", IEEE Transactions on Medical Imaging,39,1268-1277, 2019). However, the multi-excitation diffusion weighted images lack high-quality training labels, and the training labels generated by the traditional iterative reconstruction method greatly limit the potential of the intelligent reconstruction methods.
In summary, the current intelligent reconstruction method for multi-excitation diffusion weighted imaging is limited in that high-quality training data is difficult to obtain, and the potential of the intelligent reconstruction method cannot be effectively exerted. The invention provides a polynomial simulation model based on rigid motion for generating multi-excitation diffusion weighted imaging training data and an intelligent reconstruction network based on sparseness and low rank (Q.Yang, Z.Wang, K.Guo, C.Cai, and X.Qu, "Physics-drive synthesis data learning for biological magnetic response: the imaging Physics based synthesis data parallel for intelligent inter-orientation," IEEE Signal processing Magazine, DOI:10.1109/MSP.2022.3183809, 2022).
Disclosure of Invention
The invention aims to provide a physical intelligent high-definition magnetic resonance diffusion imaging method.
The invention comprises the following steps:
1) Obtaining the b value of a single excitation or multiple excitations of multiple channels to be 0mm/s 2 The diffusion weighted magnetic resonance image with the phase is estimated to obtain the channel sensitivity;
2) Obtaining a plurality of groups of motion phases according to the polynomial phase model simulation;
3) Simulating the magnetic resonance image with the phase, the channel sensitivity and the motion phase to obtain diffusion weighted image data which are excited for multiple times and are used as training data of an intelligent reconstruction network;
4) The method comprises the steps of constructing an intelligent reconstruction network comprising a plurality of iteration blocks, wherein each iteration block comprises a low-rank module, a sparse module and a data checking module. The last iteration block also comprises a denoising module; training an intelligent reconstruction network by using simulation training data;
5) Acquiring multi-excitation diffusion weighted data to be reconstructed;
6) And reconstructing multi-excitation diffusion weighted data by using the trained intelligent reconstruction network to obtain a reconstructed image.
In step 1), the acquired diffusion-weighted magnetic resonance image with phase
Can be obtained by single-shot sequence, or multiple-shot sequence with b value of 0mm/s 2 N and M are the lengths of the frequency and phase encoding dimensions of the image, respectively, and the channel sensitivity is estimated using the acquired data
Wherein the total number of channels is H.
In step 2), the specific method for obtaining multiple groups of motion phases through the polynomial phase model simulation is as follows:
the polynomial motion phase model is:
wherein x, y are coordinates of the two-dimensional image, and i is an imaginary symbol;
is the phase obtained by simulation, and N and M are the frequency of the image and the length of the phase encoding dimension, respectively; l is the polynomial order, m and L-m are the x and y powers of the phases contained in the current ith order polynomial, respectively; a. The lm Is x m y l-m The coefficients of the terms are such that,
is a two-dimensional gaussian distributed noise, where μ and σ are the mean and variance, respectively; using polynomial operationThe dynamic phase model can obtain J motion phases to form a group of simulated phase data of multi-excitation diffusion weight-setting data
J is equal to the excitation times of the multi-excitation diffusion weighted data to be reconstructed by the target.
In step 3), the simulation flow formula of the diffusion weighted data excited for multiple times is as follows:
wherein the content of the first and second substances,
is the diffusion weighted fourier space data obtained by simulation,
is the complete diffusion weighted image obtained by simulation, C is the channel sensitivity, and P is the motion phase obtained by simulation; m is a diffusion weighted magnetic resonance image with phase,
is an undersampling operator corresponding to a Fourier space sampling template of multi-excitation diffusion weighted data,
is a fourier transform. By using the formula (1), a large number of simulated multi-excitation diffusion weights can be obtained as training data of the intelligent reconstruction network. Directly carrying out Fourier transform on the complete diffusion weighted image I to obtain X GT . Finally Y, X GT And C together form a pair of training data.
In the step 4), the intelligent reconstruction network comprises K iterative blocks, each iterative block comprises a low-rank module, a sparse module and a data checking module, and the last iterative block further comprises a denoising module;
the network design of the low-rank module LR is specifically as follows:
wherein the content of the first and second substances,
is a network intermediate variable, where k denotes the kth iteration block, X k-1 Is the output of the last iteration block, X 0 = Y is a network initialization input,
is a low rank module, consisting of L LR The convolution neural network comprises a plurality of two-dimensional convolution kernels, each convolution kernel comprises a plurality of two-dimensional convolution kernels, the convolution kernels are connected through a linear rectification function, and the input of each convolution kernel is the output of the previous convolution kernel.
The network design of the sparse module SP is specifically as follows:
wherein, the first and the second end of the pipe are connected with each other,
is a variable in the middle of the network,
is the output of the last low rank module,
and
respectively, fourier transform and inverse Fourier transform, soft (.;. Theta.;) k ) Is a soft threshold operation, defined as soft (x; theta k )＝max{|x|-θ k }. X/| x |, where θ k Is a trainable threshold;
from L SP The system comprises a layer convolution neural network, each layer of convolution comprises a plurality of two-dimensional convolution kernels, the convolution layers are connected through a linear rectification function, and the input of each layer is the output of the previous layer;
after the output of (2) is subjected to a soft threshold operation
The network(s) of the network(s),
and
has an antisymmetric network structure.
The network design of the data checking module is as follows:
wherein the content of the first and second substances,
and
respectively a fourier transform and an inverse fourier transform,
is an undersampling operator corresponding to a Fourier space sampling template of multi-excitation diffusion weighted data,
is its conjugate operator; d is an identity matrix, λ 1 Is a learnable regularization parameter.
Is the output of the sparse module; y is diffusion addition obtained by simulationThe weighted fourier space data. C is channel sensitivity, C * Is a conjugate matrix of the channel sensitivity. X k Is a network intermediate variable; if the current iteration block is not the last iteration block, X k A low rank module to be input to a next iteration block; if the current iteration block is the last iteration block, i.e. the K-th iteration block, X K Will be input to the denoising module.
The denoising module is specifically designed as follows:
wherein the content of the first and second substances,
and
respectively expressing an operator and a conjugate operator for constructing a structured Hankel matrix; SVT r Is a singular value decomposition and threshold operation operator, will
After singular value decomposition, the first r singular values are retained.
Is the final output.
The loss function for intelligent network training is:
wherein T is the total number of training samples, K is the number of network iteration blocks, and X k,t Is the output of the kth iteration block after the t-th sample is input into the network,
the training label of the t sample is generated by the simulation method; i | · | live through F Is a frobenius norm; and training and updating learnable convolution kernels and parameters in the network by a deep learning common Adam optimizer to finally obtain a trained network model.
In step 5), the acquired data to be reconstructed is multi-shot echo planar diffusion weighted data read out in segments in the phase encoding dimension.
And in step 6), inputting the acquired multiple-time excitation plane echo diffusion weighting data into a trained intelligent reconstruction network, and reconstructing to obtain an image without motion artifacts.
Compared with the existing deep learning multi-excitation diffusion weighted image reconstruction, the physical intelligent reconstruction scheme provided by the invention has the following outstanding advantages:
1. the invention generates simulation data approaching to the measured data through the rigid motion model, and does not need to train the network by the measured training data, thereby solving the problem that the multi-excitation diffusion weighted image has no high-quality training data.
2. The physical intelligent network designed by the invention comprehensively utilizes two complementary prior information of image domain sparsity and Fourier space low rank.
3. The invention can effectively remove the motion artifacts in the multi-excitation diffusion weighted image and realize rapid high-quality reconstruction.
Drawings
Fig. 1 shows a phase-shifted magnetic resonance image used for simulation and estimated channel sensitivities.
Fig. 2 is a set of simulated motion phases for four shots.
FIG. 3 shows the inputs to the intelligent reconstruction method, i.e. the b-value to be reconstructed is 1000mm/s 2 The diffusion weighted image is excited four times.
FIG. 4 shows the output of the intelligent reconstruction method, i.e. the reconstructed b value is 1000mm/s 2 The diffusion weighted image is excited four times.
Detailed Description
The embodiment of the invention is a specific process of multi-excitation Fourier space signal high-resolution diffusion weighted reconstruction, and details of a physical intelligent high-definition magnetic resonance diffusion imaging method provided by the invention are described in combination with the attached drawings.
The specific implementation process is as follows:
in the first step, a magnetic resonance scanner with the magnetic field intensity of 3.0 Tesla is used for scanning 6 volunteers to acquire a b value with a phase of 0mm/s 2 The diffusion weighted magnetic resonance image m of the four excitations is obtained with the parameters: field of view 220 x 220mm 2 The layer thickness is 5mm, the coil has 32 channels, and the matrix size is cut to 180 x 180. The 6 volunteers obtained 144 magnetic resonance images with phases and 144 channel sensitivities C. The phase-shifted magnetic resonance image and the channel sensitivity are shown in fig. 1.
And secondly, simulating a motion phase. The polynomial phase model is specifically as follows:
where x, y are the coordinates of the two-dimensional image and i is an imaginary symbol.
Is the phase obtained by simulation, the frequency of the image and the length of the phase encoding dimension are both 180. The polynomial order L is 7,m and L-m are the x and y powers, respectively, of the phases that the current L-th order polynomial contains. x is the number of m y l-m Coefficient of term A lm Is from [0,0.1 l ]Is randomly obtained in the uniform distribution of the magnetic field,
is a two-dimensional gaussian distributed noise in which the mean μ and variance σ are 0 and 0.01, respectively. 4 motion phases can be obtained by utilizing a polynomial motion phase model to form simulation phase data of a group of multi-excitation diffusion weight-setting data
The co-simulation resulted in 1440 sets of motion phases.
Thirdly, simulating multi-excitation diffusion weighted data, wherein the simulation flow formula is as follows:
wherein the content of the first and second substances,
is the diffusion weighted fourier space data obtained by simulation,
is the complete diffusion weighted image obtained by simulation, C is the channel sensitivity, and P is the motion phase obtained by simulation. m is a diffusion weighted magnetic resonance image with phase,
is an undersampling operator corresponding to a Fourier space sampling template of multi-excitation diffusion weighted data,
is a fourier transform. By using the formula (1), a large number of simulated multi-shot diffusion weights can be obtained as training data of the intelligent reconstruction network. Directly carrying out Fourier transform on the complete diffusion weighted image I to obtain X GT . Finally Y, X GT And C together form a pair of training data. The amplitude data m and channel sensitivity C for each band phase can be combined with 10 sets of simulated motion phases, thus yielding 1440 sets of training data. Fig. 2 shows one set of simulated motion phases.
And fourthly, building an intelligent reconstruction network, wherein the intelligent reconstruction network comprises 5 iteration blocks, and each iteration block comprises a low-rank module, a sparse module and a data check module. The last iteration block also contains a denoising module.
a) The network design of the low-rank module LR is specifically as follows:
wherein the content of the first and second substances,
is a network intermediate variable, where k represents the kth iteration block. X k-1 Is the output of the last iteration block, X 0 Where = Y is the network initialization input, figure 3 shows the image domain of the network input, i.e. the unrereconstructed multi-shot diffusion weighted magnetic resonance data.
The low-rank module is composed of 6 layers of convolutional neural networks, each layer of convolution comprises 48 two-dimensional convolution kernels, the convolution layers are connected through a linear rectification function, and the input of each layer is the output of the previous layer.
b) The network design of the sparse module SP is specifically as follows:
wherein the content of the first and second substances,
is a variable in the middle of the network,
is the output of the last low-rank module,
and
respectively, fourier transform and inverse Fourier transform, soft (.;. Theta.;) k ) Is a soft threshold operation, defined as soft (x; theta k )＝max{|x|-θ k }. X/| x |, where θ k Is a trainable threshold, initialized to 0.01.
The convolution function is connected with each convolution layer through a linear rectification function, and each convolution layer is input into the output of the previous layer.
After the soft threshold operation, the output of (A) enters
The network(s) of the network(s),
and
has an antisymmetric network structure.
c) The network design of the data checking module is as follows:
wherein the content of the first and second substances,
and
respectively a fourier transform and an inverse fourier transform,
is an undersampling operator corresponding to a Fourier space sampling template of multi-excitation diffusion weighted data,
is its conjugate operator. D is an identity matrix, λ 1 Are learnable regularization parameters.
Is the output of the sparse module. Y is diffusion weighted fourier space data obtained from simulation. C is channel sensitivity, C * Is a conjugate matrix of channel sensitivities. X k Is a network intermediate variable. If the current iteration block is not the last iterationBlock, X k The input will be to the low rank module of the next iteration block. If the current iteration block is the last iteration block, i.e. the K-th iteration block, X K Will be input to the denoising module.
d) The denoising module is specifically designed as follows:
wherein the content of the first and second substances,
and
respectively representing an operator and a conjugate operator for constructing the structured Hankel matrix, and the size of a sliding window is [5,5 ]]。SVT r Is a singular value decomposition and threshold operation operator, will
After the singular value decomposition, the first 30 singular values are retained.
Is the final output.
The loss function for intelligent network training is:
wherein the total number of training samples is 1200, the number of verification samples is 240, and the number of network iteration blocks is 5,X k,t Is the output of the kth iteration block after the t-th sample is input into the network,
the training label of the t-th sample is generated by the simulation method. I | · | purple wind F Is the frobenius norm. Training learnable convolution kernels and parameters in an update network through a deep learning common Adam optimizer, and finallyAnd obtaining the trained network model.
Fourthly, scanning 1 volunteer by using a magnetic resonance scanner with the magnetic field intensity of 3.0 tesla to acquire four-time excited diffusion weighted magnetic resonance images, wherein the acquisition parameters are as follows: TR/TE =3000/60ms, b value 1000mm/s 2 3 in the diffusion direction, 220 × 220mm in the field of view 2 Layer thickness 5mm, layer number 12, coil 32 passageway, matrix size is 180 x 180 after cutting. And calculating to obtain the channel sensitivity.
And fifthly, inputting the collected four times of excitation diffusion weighting data into a trained intelligent reconstruction network, and reconstructing to obtain an image without motion artifacts. FIG. 4 shows that the reconstructed b value is 1000mm/s 2 The diffusion weighted image is excited four times.
Claims (7)
1. A physical intelligent high-definition magnetic resonance diffusion imaging method is characterized by comprising the following steps:
1) Obtaining the b value of 0mm/s of single excitation or multiple excitations of multiple channels 2 The diffusion weighted magnetic resonance image with the phase is estimated to obtain the channel sensitivity;
2) Simulating to obtain a plurality of groups of motion phases according to a polynomial phase model of rigid motion;
3) Simulating the magnetic resonance image with the phase, the channel sensitivity and the motion phase to obtain diffusion weighted image data which are excited for multiple times and are used as training data of an intelligent reconstruction network;
4) Constructing an intelligent reconstruction network comprising a plurality of iteration blocks, wherein each iteration block comprises a low-rank module, a sparse module and a data check module, and the last iteration block further comprises a denoising module; training an intelligent reconstruction network by using simulation training data;
5) Acquiring multi-excitation diffusion weighted data to be reconstructed;
6) And reconstructing multi-excitation diffusion weighted data by using the trained intelligent reconstruction network to obtain a reconstructed image.
2. The physically intelligent high-definition magnetic resonance diffusion imaging method as claimed in claim 1, wherein in the step 1), the method is characterized in thatAcquired diffusion weighted magnetic resonance image with phase
B-value of 0mm/s for the singly-excited or multiply-excited sequence 2 N and M are the lengths of the frequency and phase encoding dimensions of the image, respectively, and the channel sensitivity is estimated using the acquired data
Wherein the total number of channels is H.
3. The physical intelligent high-definition magnetic resonance diffusion imaging method as claimed in claim 1, wherein in step 2), the polynomial phase model is simulated to obtain a plurality of groups of motion phases as follows:
the polynomial motion phase model based on rigid body motion is:
wherein x, y are coordinates of the two-dimensional image, and i is an imaginary symbol;
is the phase obtained by simulation, and N and M are the frequency of the image and the length of the phase encoding dimension, respectively; l is the polynomial order, m and L-m are the x and y powers of the phases contained in the current ith order polynomial, respectively; a. The lm Is x m y l-m The coefficient of the term(s) is,
is a two-dimensional gaussian distributed noise, where μ and σ are the mean and variance, respectively; j motion phases can be obtained by utilizing a polynomial motion phase model to form simulation phase data of a group of multi-excitation diffusion weight-setting data
J equals the number of shots of the multi-shot diffusion weighted data of the target to be reconstructed.
4. An intelligent reconstruction diffusion weighted magnetic resonance imaging method as claimed in claim 1, wherein in step 3), the simulation flow formula of the multi-excitation diffusion weighted data is:
wherein the content of the first and second substances,
is diffusion weighted Fourier space data of each channel of each excitation obtained by simulation,
is a simulated complete diffusion weighted image, C is the channel sensitivity, P is the simulated motion phase, m is the diffusion weighted magnetic resonance image with phase,
is an undersampling operator corresponding to a Fourier space sampling template of multi-excitation diffusion weighted data,
the method is characterized in that Fourier transform is adopted, a large number of simulated multi-excitation diffusion weights are obtained by using a formula (1) and are used as training data of an intelligent reconstruction network, and an X is obtained by directly carrying out Fourier transform on a complete diffusion weighted image I GT Finally Y, X GT And C together form a pair of training data.
5. The physically intelligent high-definition magnetic resonance diffusion imaging method as claimed in claim 1, wherein in step 4), the intelligent reconstruction network comprises K iterative blocks, each iterative block comprises a low rank module, a sparse module, a data check module, and the last iterative block further comprises a denoising module;
the network design of the low-rank module LR is specifically as follows:
wherein the content of the first and second substances,
is a network intermediate variable, where k denotes the kth iteration block, X k-1 Is the output of the last iteration block, X 0 = Y is a network initialization input,
is a low rank module, consisting of L LR The system comprises a layer convolution neural network, each layer of convolution comprises a plurality of two-dimensional convolution kernels, the convolution layers are connected through a linear rectification function, and the input of each layer is the output of the previous layer;
the network design of the sparse module SP is specifically as follows:
wherein the content of the first and second substances,
is a variable in the middle of the network,
is the output of the last low rank module,
and
respectively Fourier transform and inverse Fourier transform, soft(·；θ k ) Is a soft threshold operation, defined as soft (x; theta k )＝max{|x|-θ k } · x/| x |, where θ k Is a trainable threshold;
from L SP A layer convolution neural network, each layer of convolution comprises a plurality of two-dimensional convolution kernels, the convolution layers are connected by a linear rectification function, and the input of each layer is the output of the previous layer,
after the soft threshold operation, the output of (A) enters
The network(s) of the network(s),
and
having an antisymmetric network structure;
the network design of the data checking module is as follows:
wherein the content of the first and second substances,
andrespectively a fourier transform and an inverse fourier transform,
is an undersampling operator corresponding to a Fourier space sampling template of multi-excitation diffusion weighted data,
is its conjugate operator; d is an identity matrix, λ 1 Is a learnable regularization parameter;
is the output of the sparse module, Y is the diffusion weighted Fourier space data obtained by simulation, C is the channel sensitivity * Is a conjugate matrix of channel sensitivities, X k Is a network intermediate variable; if the current iteration block is not the last iteration block, X k The low rank module is input to the next iteration block, if the current iteration block is the last iteration block, i.e. the Kth iteration block, X K Inputting the input to a denoising module;
the denoising module is specifically designed as follows:
wherein the content of the first and second substances,
and
respectively representing an operator and a conjugate operator for constructing a structured Hankel matrix, SVT r Is a singular value decomposition and threshold operation operator, will
After singular value decomposition, the first r singular values are reserved;
is the final output;
the loss function for intelligent network training is:
wherein T is the total number of training samples, K is the number of network iteration blocks, X k,t Is the output of the kth iteration block after the t-th sample is input into the network,
is the training label of the tth sample, is generated by the simulation method, | | · u F Is a frobenius norm; and training and updating learnable convolution kernels and parameters in the network by a deep learning common Adam optimizer to finally obtain a trained network model.
6. A physically intelligent high definition magnetic resonance diffusion imaging method as claimed in claim 1, characterized in that in step 5), the acquired data to be reconstructed is multi-shot planar echo diffusion weighted data read out in segments in phase encoding dimension.
7. The physical intelligent high-definition magnetic resonance diffusion imaging method according to claim 1, wherein in step 6), the acquired multi-shot echo planar diffusion weighted data are input into a trained intelligent reconstruction network, and an image without motion artifacts is reconstructed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211040703.8A CN115471580A (en) | 2022-08-29 | 2022-08-29 | Physical intelligent high-definition magnetic resonance diffusion imaging method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211040703.8A CN115471580A (en) | 2022-08-29 | 2022-08-29 | Physical intelligent high-definition magnetic resonance diffusion imaging method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115471580A true CN115471580A (en) | 2022-12-13 |
Family
ID=84368901
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211040703.8A Pending CN115471580A (en) | 2022-08-29 | 2022-08-29 | Physical intelligent high-definition magnetic resonance diffusion imaging method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115471580A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117011409A (en) * | 2023-08-10 | 2023-11-07 | 厦门大学 | Multi-position physical intelligent high-definition diffusion magnetic resonance data generation method |
-
2022
- 2022-08-29 CN CN202211040703.8A patent/CN115471580A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117011409A (en) * | 2023-08-10 | 2023-11-07 | 厦门大学 | Multi-position physical intelligent high-definition diffusion magnetic resonance data generation method |
| CN117011409B (en) * | 2023-08-10 | 2024-05-10 | 厦门大学 | Multi-position physical intelligent high-definition diffusion magnetic resonance data generation method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Liu et al. | RARE: Image reconstruction using deep priors learned without groundtruth | |
| CN107610194B (en) | Magnetic resonance image super-resolution reconstruction method based on multi-scale fusion CNN | |
| CN110378980B (en) | Multichannel magnetic resonance image reconstruction method based on deep learning | |
| Wen et al. | Transform learning for magnetic resonance image reconstruction: From model-based learning to building neural networks | |
| Lam et al. | Constrained magnetic resonance spectroscopic imaging by learning nonlinear low-dimensional models | |
| CN112150568A (en) | Magnetic resonance fingerprint imaging reconstruction method based on Transformer model | |
| CN111870245B (en) | Cross-contrast-guided ultra-fast nuclear magnetic resonance imaging deep learning method | |
| CN109410289A (en) | A kind of high lack sampling hyperpolarized gas lung MRI method for reconstructing of deep learning | |
| Luo et al. | Bayesian MRI reconstruction with joint uncertainty estimation using diffusion models | |
| CN111783631B (en) | Sparse representation-based deep learning magnetic resonance spectrum reconstruction method | |
| Fan et al. | A segmentation-aware deep fusion network for compressed sensing mri | |
| Benou et al. | De-noising of contrast-enhanced MRI sequences by an ensemble of expert deep neural networks | |
| CN112991483B (en) | Non-local low-rank constraint self-calibration parallel magnetic resonance imaging reconstruction method | |
| CN114820849A (en) | Magnetic resonance CEST image reconstruction method, device and equipment based on deep learning | |
| CN114299185A (en) | Magnetic resonance image generation method, magnetic resonance image generation device, computer equipment and storage medium | |
| Manimala et al. | Convolutional neural network for sparse reconstruction of MR images interposed with gaussian noise | |
| CN115471580A (en) | Physical intelligent high-definition magnetic resonance diffusion imaging method | |
| Gan et al. | SS-JIRCS: Self-supervised joint image reconstruction and coil sensitivity calibration in parallel MRI without ground truth | |
| Klug et al. | Analyzing the sample complexity of self-supervised image reconstruction methods | |
| CN114820636A (en) | Three-dimensional medical image segmentation model and training method and application thereof | |
| CN113487507A (en) | Dual-domain recursive network MR reconstruction method based on multi-module feature aggregation | |
| Bian et al. | CSDL‐Net: An iterative network based on compressed sensing and deep learning | |
| CN112634385A (en) | Rapid magnetic resonance imaging method based on deep Laplace network | |
| CN113034639B (en) | Magnetic resonance imaging image reconstruction method based on separable Henkel matrix | |
| CN117011409B (en) | Multi-position physical intelligent high-definition diffusion magnetic resonance data generation method |
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 |