Uncertainty-guided dual-domain magnetic resonance imaging artifact correction method and system

Abstract

The application discloses an uncertainty-guided dual-domain nuclear magnetic resonance imaging artifact correction method and system, which can reconstruct high-quality images without motion artifacts from k-space data with motion artifacts collected by a nuclear magnetic resonance imaging device. The method mainly includes five steps of motion-damaged k-space data distribution modeling, k-space uncertainty-guided dual-domain motion artifact correction model construction, dual-domain motion artifact correction network construction, dual-domain motion artifact correction network training process and motion artifact correction process. A plurality of groups of k-space data with motion artifacts, corresponding images without motion artifacts and k-space data are used to train network parameters, so that the image output and the k-space data output of the network are as close as possible to the images without motion artifacts and the k-space data. In application, the k-space data with motion artifacts is input, and high-quality nuclear magnetic resonance images without motion artifacts can be obtained from the output of the network.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of medical magnetic resonance imaging, and particularly relates to an uncertainty-guided dual-domain magnetic resonance imaging artifact correction method and system, which is used for reconstructing a high-quality motion-artifact-free image from k-space data with motion artifacts collected by a magnetic resonance imaging device. BACKGROUND

[0002] Magnetic resonance imaging is a widely used non-invasive imaging technique. Due to the long acquisition time, magnetic resonance imaging can produce motion artifacts in the reconstructed image, which reduces the quality of the magnetic resonance image and affects clinical diagnosis. Therefore, it is of great significance in clinical practice to correct motion artifacts in magnetic resonance imaging.

[0003] Existing motion artifact correction methods in magnetic resonance imaging include prospective and retrospective methods. Prospective methods use external tracking devices or navigators to measure the motion of the subject during the scanning motion correction period. However, these methods require expensive additional hardware or pulse sequence modifications to increase the accuracy of the scan. Retrospective motion correction methods either explicitly model and correct the motion in the image reconstruction algorithm or use deep learning methods to learn the mapping from the motion-artifacted magnetic resonance image to the motion-free magnetic resonance image. Specifically, some of these methods are based on a forward model of the subject's motion and use optimization algorithms to jointly estimate the motion parameters and the magnetic resonance image. Some other methods introduce convolutional neural networks into the joint optimization process to learn the magnetic resonance image prior. However, these optimization-based retrospective motion correction methods require expensive computation to explicitly estimate the motion parameters, because the optimization of the motion parameters is a non-convex optimization problem that requires many iterations to obtain a satisfactory motion parameter. Deep learning methods directly learn the mapping from the motion-affected magnetic resonance image to the motion-artifact-free magnetic resonance image by designing various deep networks. Although these deep learning methods have made some progress, the robustness of artifact removal is still a problem, and in addition, they often introduce blurring or non-physical features in the artifact correction process. Therefore, it is necessary to further explore the development of more effective motion artifact correction algorithms to further improve the quality of magnetic resonance images. SUMMARY

[0004] The present application proposes an uncertainty-guided dual-domain magnetic resonance imaging artifact correction method, which can reconstruct a high-quality motion-artifact-free image from k-space data with motion artifacts collected by a magnetic resonance imaging device.

[0005] In order to achieve the above object, the technical scheme adopted by the present application is as follows: a dual-domain motion artifact correction method based on uncertainty guidance, wherein k-space data with motion artifacts are input into a dual-domain motion artifact correction network for motion artifact correction to obtain high-quality nuclear magnetic resonance images without motion artifacts.

[0006] The dual-domain motion artifact correction network is obtained through the following steps:

[0007] The distribution of the k-space data damaged by motion is modeled through k-space uncertainty.

[0008] A k-space uncertainty guided dual-domain motion artifact correction model is constructed through the distribution of the k-space data damaged by motion and the dual-domain data prior distribution.

[0009] The calculation process of the alternating iterative algorithm for optimizing the k-space uncertainty guided dual-domain motion artifact correction model is unfolded into a dual-domain motion artifact correction network, and the dual-domain motion artifact correction network comprises a k-space uncertainty estimation block and a dual-domain reconstruction block.

[0010] The dual-domain motion artifact correction network is trained with the k-space data with motion artifacts as input, and the corresponding k-space data and nuclear magnetic resonance images without motion artifacts are output, so that the final dual-domain motion artifact correction network is obtained.

[0011] In the step 1), the distribution of the k-space data damaged by motion is modeled through k-space uncertainty; it is assumed that the k-space data damaged by motion is subject to a non-independent identically distributed Gaussian distribution at each position, wherein the reciprocal of the standard deviation of the Gaussian distribution is defined as the uncertainty of the k-space, and a smaller value is given to the k-space data severely damaged by motion, and a larger value is given to the k-space data less affected by motion.

[0012] The k-space uncertainty guided dual-domain motion artifact correction model comprises a k-space uncertainty guided data consistency term and a prior term of the nuclear magnetic resonance image in the image domain and the k-space.

[0013] Based on the distribution of the k-space data damaged by motion, the dual-domain motion artifact correction problem is modeled as a maximum log posteriori estimation model, and the solution of the maximum log posteriori estimation model is converted into a dual-domain regularization model with a k-space uncertainty guided data consistency term, which specifically comprises a log likelihood The constraint term about the k-space uncertainty derived by taking the logarithm of the corresponding k-space uncertainty guided data consistency term, the regularization term of the image domain prior, the regularization term of the k-space prior and the normalization term of the likelihood distribution.

[0014] The dual-domain motion artifact correction network is based on an alternating optimization algorithm, takes motion corrupted k-space data as input, and reconstructs a motion-free magnetic resonance image by using a plurality of operation units, each of which is composed of a k-space uncertainty estimation block and a dual-domain reconstruction block, the k-space uncertainty estimation block is used to quantify the uncertainty of the motion corrupted k-space data, and the dual-domain reconstruction block is used to reconstruct the k-space data and the magnetic resonance image under the guidance of the k-space uncertainty.

[0015] In training the dual-domain motion artifact correction network, the training data set is composed of motion corrupted undersampled k-space data, corresponding motion-free images and motion-free k-space data.

[0016] The objective function of the dual-domain motion artifact correction network training is:

[0017]

[0018] Wherein, x t and y t are the magnetic resonance image and k-space data reconstructed by the tth operation unit, and gamma is a hyperparameter used to balance the loss function between k-space and image space, in the present application, gamma is set to 0.001; the gradient of the objective function with respect to the parameters of the dual-domain motion artifact correction network is calculated by using the back propagation algorithm, and then the Adam algorithm is used to optimize the parameters of the dual-domain motion artifact correction network based on the training data set, to obtain the optimal network parameters.

[0019] Based on the above technical concept, an uncertainty guided dual-domain magnetic resonance imaging artifact correction system is provided, which comprises a correction module and a model construction module, the model construction module comprises a data modeling unit, a dual-domain motion artifact correction model construction unit, a dual-domain motion artifact correction network construction unit and a training unit;

[0020] The correction module is used for inputting the k-space data with motion artifacts into the dual-domain motion artifact correction network for motion artifact correction to obtain a high-quality motion-free magnetic resonance image without artifacts;

[0021] The data modeling unit is used for modeling the distribution of the motion corrupted k-space data by k-space uncertainty;

[0022] The dual-domain motion artifact correction model construction unit is used for constructing a k-space uncertainty guided dual-domain motion artifact correction model by using the motion corrupted k-space data distribution and a dual-domain data prior distribution;

[0023] The dual-domain motion artifact correction network construction unit is used to expand the calculation process of an alternating iterative algorithm of an optimized k-space uncertainty guided dual-domain motion artifact correction model into a dual-domain motion artifact correction network, and the dual-domain motion artifact correction network includes a k-space uncertainty estimation block and a dual-domain reconstruction block.

[0024] The training unit takes k-space data with motion artifacts as input, trains the dual-domain motion artifact correction network, and outputs approximated k-space data and magnetic resonance images without motion artifacts, so as to obtain a final dual-domain motion artifact correction network.

[0025] In addition, the present application also provides a computer device including a processor and a memory, the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes, and the processor can realize the uncertainty guided dual-domain magnetic resonance imaging artifact correction method of the present application when executing the computer executable program.

[0026] Meanwhile, a computer readable storage medium can be provided, and the computer readable storage medium stores a computer program, and the computer program can realize the uncertainty guided dual-domain magnetic resonance imaging artifact correction method of the present application when executed by a processor.

[0027] Compared with the prior art, the present application has at least the following beneficial technical effects:

[0028] The present application proposes an uncertainty guided dual-domain magnetic resonance imaging deep learning method, which can reconstruct high-quality magnetic resonance images without motion artifacts from k-space data with motion artifacts collected by a magnetic resonance imaging device. Compared with existing optimization-based motion artifact correction methods, the proposed method models motion damage based on k-space uncertainty without explicitly estimating motion parameters. Compared with existing deep learning motion correction methods, the proposed method combines a k-space uncertainty guided data consistency term into an expanded network to guide magnetic resonance imaging reconstruction. In summary, the present application can be mainly used to correct motion artifacts in magnetic resonance imaging, provide high-quality image data for doctors' later auxiliary diagnosis, and thus help to improve the diagnosis accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 is a specific implementation flowchart of the present application.

[0030] Figure 2 is a dual-domain motion artifact correction network structure diagram.

[0031] Figure 3 is a reconstructed motion artifact free instance diagram (severe motion, segmented constant motion trajectory). DETAILED DESCRIPTION

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific examples. These examples are merely illustrative and not intended to limit the invention.

[0033] refer to Figure 1 and Figure 2 The present invention provides an uncertainty-guided method for correcting motion artifacts in dual-domain nuclear magnetic resonance imaging, comprising the following steps:

[0034] I. Distribution Modeling of k-space Data for Motion-Induced Damage

[0035] This invention models motion-induced damage by measuring the uncertainty of k-space data. Specifically, this invention assumes k-space data of motion-induced damage. The distribution at each location follows a non-independent, identically distributed, and Gaussian distribution, where N is the number of k-space data points. Specifically, considering... At the i-th position, there is

[0036]

[0037] in It is an image without motion artifacts. It is the Fourier transform matrix. σ [i] ∈[1,∞) is The standard deviation at position i provides a larger value for k-space data that is severely disrupted by motion, and a smaller value for k-space data that is less affected by motion.

[0038] II. Constructing a two-domain motion artifact correction model guided by k-space uncertainty

[0039] Based on the distribution modeling of the k-space data caused by motion destruction, this invention proposes the following maximum log-posterior estimation model:

[0040]

[0041] where w∈[0,1] N This represents the uncertainty of the k-space, where each element is w. [i] =1 / σ [i] p(x) and p(y) are the prior distributions of motion artifact-free data in the image domain and k-space domain, respectively. Likelihood distribution Having been modeled by Equation (1), the solution to Equation (2) can be transformed into a bi-domain regularized model with a data consistency term guided by k-space uncertainty:

[0042]

[0043] in, and The parameter is θ I and θ K The two learnable denoisers in this invention employ a U-Net structure. The first term is the log-likelihood... The corresponding k-space uncertainty-guided data consistency term constrains the k-space data of each reconstructed MRI image and its corresponding motion-damaged k-space data under k-space uncertainty guidance; the second and third terms are regularization terms that demonstrate the image domain prior and the k-space prior; the fourth term is a constraint term on k-space uncertainty after taking the logarithm of the likelihood distribution normalization term.

[0044] III. Constructing a Two-Domain Motion Artifact Correction Network

[0045] Model optimization

[0046] The dual-domain motion artifact correction network proposed in this invention is designed based on an alternating optimization algorithm and is used to solve formula (3), such as... Figure 1 As shown, motion-destroyed k-space data is used as input to reconstruct motion-free MRI images using T operation units. Each operation unit consists of a k-space uncertainty estimation block (KU-Module) and a dual-domain reconstruction block (DDR-Module), corresponding to the subproblems of optimizing k-space uncertainty and optimizing dual-domain data containing k-space data and MRI images, respectively. The k-space uncertainty estimation block is used to quantify the uncertainty of the motion-destroyed k-space data. The dual-domain reconstruction block is responsible for reconstructing the k-space data and MRI images under the guidance of k-space uncertainty. Details of these two blocks in the t-th operation unit are as follows:

[0047] The K-space uncertainty estimation block is used to update the k-space uncertainty in formula (3). If formula (3) is optimized directly, This depends on the k-space data of the reconstructed image and the k-space data of the motion-destructive image. The difference between them. This invention extends this estimation to a learnable module, defined as:

[0048]

[0049] in The parameter is θ W The subnetwork. When t=1, this invention only... It is sent to the k-space uncertainty estimation block because there is no estimate of the reconstructed MRI image in this case.

[0050] The dual-domain reconstruction block is used to update the k-space data and the magnetic resonance image in formula (3) under the guidance of k-space uncertainty. Specifically, given the reconstructed magnetic resonance image x t-1 and the k-space uncertainty w t , the k-space data of the t-th operation unit is updated by:

[0051]

[0052] where W t = diag(w t ) ∈ [0, 1] N×N is a diagonal matrix, so the matrix inversion in formula (5) can be calculated efficiently. Define formula (5) as a k-space reconstruction module (K-Block) for solving the sub-problem of optimizing the k-space data y in formula (3). Formula (5) can be implemented by first calculating , and then calculating the k-space uncertainty guided data consistency operator UDC K in formula (5). Similarly, given the updated k-space uncertainty w t and the k-space data y t , the magnetic resonance image of the t-th operation unit is updated by:

[0053]

[0054] Formula (6) defines an image reconstruction module (I-Block) for solving the sub-problem for optimizing the magnetic resonance image x in formula (3). Formula (6) can be implemented by first calculating , and then calculating the image space uncertainty guided data consistency operator UDC I in formula (3).

[0055] The k-space reconstruction module and the image reconstruction module are combined into a dual-domain reconstruction block, which sequentially reconstructs the k-space data y t and the magnetic resonance image x t of the t-th operation unit.

[0056] In summary, by alternately connecting the k-space uncertainty estimation block and the dual-domain reconstruction block, a multi-level deep network for motion artifact correction (i.e., a dual-domain motion artifact correction network) is obtained, as shown in Figure 1 .

[0057] IV. Training process of the dual-domain motion artifact correction network

[0058] To determine the optimal parameters of the dual-domain motion artifact correction network, the application constructs a training data set for the motion artifact correction problem. The training data set is composed of a plurality of data groups, each data group being composed of k-space data damaged by motion and corresponding motion artifact-free images and motion artifact-free k-space data. In actual construction, the application first uses the acquired motion artifact-free nuclear magnetic resonance images, and then simulates the k-space data damaged by motion collected from the nuclear magnetic resonance instrument by simulating rigid body motion. The application uses the k-space data damaged by motion as the input of the dual-domain motion artifact correction network, and then trains the dual-domain motion artifact correction network using a loss function in which the k-space data and the nuclear magnetic resonance images reconstructed by different operation units are as close as possible to the motion artifact-free k-space data and the motion artifact-free images. Specifically, the application uses the loss in the image domain and the k-space domain to train the network:

[0059]

[0060] wherein x t and y t are the nuclear magnetic resonance images and the k-space data reconstructed by the tth operation unit, and γ is a hyperparameter used to balance the loss function between the k-space and the image space, and in the application, γ is set to 0.001; the gradient of the objective function with respect to the parameters of the dual-domain motion artifact correction network is calculated using the back propagation algorithm, and then the Adam algorithm is used to optimize the parameters of the dual-domain motion artifact correction network based on the training data set, to obtain the optimal network parameters.

[0061] V. Applying the trained dual-domain motion artifact correction network to ultrafast nuclear magnetic resonance imaging

[0062] Through the training process of step four, the application can determine the optimal parameters of the dual-domain motion artifact correction network. Based on the trained dual-domain motion artifact correction network, the input of the k-space data with motion artifacts is output as the reconstructed motion artifact-free nuclear magnetic resonance image by the dual-domain motion artifact correction network. Because the parameter training process of the dual-domain motion artifact correction network of step four makes the output image of the dual-domain motion artifact correction network as close as possible to the standard nuclear magnetic resonance image, the trained dual-domain motion artifact correction network can still obtain a high-quality reconstructed image in the case of motion artifacts.

[0063] Based on the technical concept of the application, an uncertainty-guided dual-domain nuclear magnetic resonance imaging artifact correction system is provided, which includes a correction module and a model construction module, and the model construction module includes a data modeling unit, a dual-domain motion artifact correction model construction unit, a dual-domain motion artifact correction network construction unit, and a training unit.

[0064] The correction module is configured to input the k-space data with motion artifacts into a dual-domain motion artifact correction network to perform motion artifact correction and obtain high-quality nuclear magnetic resonance images without motion artifacts.

[0065] The data modeling unit is configured to model a distribution of the motion-damaged k-space data through k-space uncertainty.

[0066] The dual-domain motion artifact correction model construction unit is configured to construct a k-space uncertainty guided dual-domain motion artifact correction model through the motion-damaged k-space data distribution and a dual-domain data prior distribution.

[0067] The dual-domain motion artifact correction network construction unit is configured to expand a calculation process of an alternating iterative algorithm for optimizing the k-space uncertainty guided dual-domain motion artifact correction model into a dual-domain motion artifact correction network, and the dual-domain motion artifact correction network includes a k-space uncertainty estimation block and a dual-domain reconstruction block.

[0068] The training unit is configured to input the k-space data with motion artifacts into the dual-domain motion artifact correction network for training, and output k-space data and nuclear magnetic resonance images that approximate corresponding k-space data without motion artifacts, so as to obtain a final dual-domain motion artifact correction network.

[0069] In addition, the present application can also provide a computer device including a processor and a memory, the memory being configured to store computer executable programs, the processor being configured to read part or all of the computer executable programs from the memory and execute the programs, and the processor being configured to implement the uncertainty guided dual-domain nuclear magnetic resonance imaging artifact correction method when executing part or all of the computer executable programs.

[0070] In another aspect, the present application provides a computer readable storage medium, the computer readable storage medium being configured to store a computer program, and the computer program being configured to implement the uncertainty guided dual-domain nuclear magnetic resonance imaging artifact correction method when executed by a processor.

[0071] The computer device can be a notebook computer, a desktop computer or a workstation.

[0072] The processor can be a central processing unit (CPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC) or a ready-to-program field programmable gate array (FPGA).

[0073] For the memory, it can be an internal storage unit of a notebook computer, a desktop computer or a workstation, such as a memory or a hard disk, or an external storage unit, such as a mobile hard disk or a flash card.

[0074] The computer-readable storage medium can include computer storage media and communication media. The computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. The computer-readable storage medium can include read-only memory (ROM), random access memory (RAM), solid state disk (SSD), optical disk, etc. Among them, the random access memory can include resistance random access memory (ReRAM) and dynamic random access memory (DRAM).

[0075] In the numerical experiment, the brain data of 117 subjects is used by the application, of which the data of 78 subjects is used as training data, and the data of 39 subjects is used for testing. According to the forward model Simulate in-plane and inter-plane motion, where T θ is the rigid motion matrix parameterized by translation and rotation vectors M∈{0,1} N×N is a diagonal mask matrix in k-space. And the application will reserve 7% of the k-space lines in k-space to prevent excessive distortion of the image. The motion vector is randomly selected from the Gaussian distribution N(0, 10). Regarding the motion trajectory, the application adopts piecewise constant, piecewise transient and Gaussian to simulate the motion. In addition, in order to generate various motion severities, each motion level has a series of motion corrupted k-space lines: 0-30%, 0-50% and 0-70% of the total number of k-space lines are damaged for mild, moderate and severe, respectively. In order to objectively evaluate different methods, the normalized root mean square error (NRMSE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are taken to measure the average reconstruction accuracy on the test set.

[0076] As shown in Table 1, the dual-domain motion artifact correction network (D2MC-Net) of the application is compared with four deep learning methods and one optimization-based method under different motion trajectories and motion severities. The deep learning methods include: U-Net, UPGAN, SU-Net and Alternating. The optimization-based method is Autofocuing+, and the results of the motion corrupted images without motion correction are represented as "Corrupted". The dual-domain motion artifact correction network designed by the application achieves the best reconstruction accuracy under different motion trajectories and motion severities. Figure 3is the visualization result of the reconstructed image, it can be seen that the reconstructed nuclear magnetic resonance image of the application has clear structural details and no obvious artifacts.

[0077] Table 1: Comparison results of different methods in the brain data test set under different motion trajectories and motion severity

[0078]

[0079]

Claims

1. An uncertainty-guided method for artifact correction in dual-domain nuclear magnetic resonance imaging, characterized in that, The k-space data with motion artifacts is input into a dual-domain motion artifact correction network to correct motion artifacts and obtain high-quality MRI images without motion artifacts. The dual-domain motion artifact correction network is obtained through the following steps: The distribution of k-space data with motion-induced damage is modeled using k-space uncertainty. A k-space uncertainty-guided bi-domain motion artifact correction model is constructed using the k-space data distribution of motion-damaged data and the prior distribution of bi-domain data. Based on the distribution of motion-damaged k-space data, the bi-domain motion artifact correction problem is modeled as a maximum log-posterior estimation model. The solution of the maximum log-posterior estimation model is then transformed into a bi-domain regularized model with a data consistency term guided by k-space uncertainty, specifically including log-likelihood... The constraint terms regarding k-space uncertainty are derived from the data consistency term guided by k-space uncertainty, the regularization term of the image domain prior, the regularization term of the k-space prior, and the logarithm of the normalization term of the likelihood distribution. The computation process of the alternating iterative algorithm for optimizing the k-space uncertainty-guided dual-domain motion artifact correction model is expanded into a dual-domain motion artifact correction network, which includes a k-space uncertainty estimation block and a dual-domain reconstruction block. Using k-space data with motion artifacts as input, the dual-domain motion artifact correction network is trained, and the output approximates the corresponding k-space data and MRI image without motion artifacts, thus obtaining the final dual-domain motion artifact correction network.

2. The uncertainty-guided dual-domain nuclear magnetic resonance imaging artifact correction method according to claim 1, characterized in that, Modeling the distribution of k-space data damaged by motion through k-space uncertainty; assuming that the k-space data damaged by motion follows a non-independent and identically distributed Gaussian distribution at each location, where the reciprocal of the standard deviation of the Gaussian distribution is defined as the uncertainty of k-space, a smaller value is assigned to k-space data severely damaged by motion, and a larger value is assigned to k-space data less affected by motion.

3. The uncertainty-guided dual-domain nuclear magnetic resonance imaging artifact correction method according to claim 1, characterized in that, The k-space uncertainty-guided dual-domain motion artifact correction model includes a k-space uncertainty-guided data consistency term and prior terms for MRI images in the image domain and k-space.

4. The uncertainty-guided dual-domain nuclear magnetic resonance imaging artifact correction method according to claim 1, characterized in that, The dual-domain motion artifact correction network is based on an alternating optimization algorithm. It takes motion-damped k-space data as input and uses multiple operation units to reconstruct motion-free MRI images. Each operation unit consists of a k-space uncertainty estimation block and a dual-domain reconstruction block. The k-space uncertainty estimation module is used to quantify the uncertainty of the motion-damped k-space data, and the dual-domain reconstruction module is used to reconstruct the k-space data and MRI images under the guidance of k-space uncertainty.

5. The uncertainty-guided dual-domain nuclear magnetic resonance imaging artifact correction method according to claim 1, characterized in that, When training the dual-domain motion artifact correction network, the training dataset consists of motion-damped undersampled k-space data, the corresponding motionless image, and motionless k-space data.

6. The uncertainty-guided dual-domain nuclear magnetic resonance imaging artifact correction method according to claim 1, characterized in that, Objective function for training a dual-domain motion artifact correction network for: in, For the first Reconstructed NMR images and k-space data from each operation unit It is a hyperparameter used to balance the loss function between k-space and image space, set =0.001; The gradient of the objective function with respect to the parameters of the dual-domain motion artifact correction network is calculated using the backpropagation algorithm. Then, the Adam algorithm is used to optimize the parameters of the dual-domain motion artifact correction network based on the training dataset to obtain the optimal network parameters.

7. An uncertainty-guided dual-domain nuclear magnetic resonance imaging artifact correction system, characterized in that, It includes a correction module and a model building module. The model building module includes a data modeling unit, a dual-domain motion artifact correction model construction unit, a dual-domain motion artifact correction network construction unit, and a training unit. The correction module is used to input k-space data with motion artifacts into a dual-domain motion artifact correction network to correct motion artifacts and obtain high-quality MRI images without motion artifacts. The data modeling unit is used to model the distribution of k-space data with motion damage through k-space uncertainty; The dual-domain motion artifact correction model construction unit is used to construct a dual-domain motion artifact correction model guided by k-space uncertainty through the k-space data distribution of motion-damaged data and the prior distribution of dual-domain data. Based on the distribution of k-space data with motion damage, the dual-domain motion artifact correction problem is modeled as a maximum log-posterior estimation model. The solution of the maximum log-posterior estimation model is transformed into a dual-domain regularized model with a data consistency term guided by k-space uncertainty, specifically including log-likelihood. The constraint terms regarding k-space uncertainty are derived from the data consistency term guided by k-space uncertainty, the regularization term of the image domain prior, the regularization term of the k-space prior, and the logarithm of the normalization term of the likelihood distribution. The dual-domain motion artifact correction network building unit is used to expand the computation process of the alternating iterative algorithm of the dual-domain motion artifact correction model guided by optimization k-space uncertainty into a dual-domain motion artifact correction network. The dual-domain motion artifact correction network includes a k-space uncertainty estimation block and a dual-domain reconstruction block. The training unit takes k-space data with motion artifacts as input to train the dual-domain motion artifact correction network, and outputs k-space data and MRI images that approximate the corresponding k-space data and MRI images without motion artifacts, thus obtaining the final dual-domain motion artifact correction network.

8. A computer device, characterized in that, It includes a processor and a memory, the memory being used to store a computer-executable program, the processor reading the computer-executable program from the memory and executing it, and the processor executing the computer-executable program being able to implement the uncertainty-guided dual-domain nuclear magnetic resonance imaging artifact correction method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program that, when executed by a processor, enables the uncertainty-guided dual-domain magnetic resonance imaging artifact correction method as described in any one of claims 1 to 6.

Images