Image reconstruction method for magnetic resonance imaging distortion correction and magnetic resonance imaging system

Abstract

The present invention provides an image reconstruction method and a magnetic resonance imaging system for distortion correction of magnetic resonance imaging, comprising the following steps: 1) collecting magnetic resonance downsampled k-space data through an EPI fast imaging sequence; 2) obtaining a B0 field map; 3) inputting the downsampled k-space data and the information of the B0 field map into an artificial neural network for magnetic resonance image reconstruction, wherein the artificial neural network is a plurality of convolutional neural networks or Transformer neural networks connected in series, and updating intermediate values ​​through the neural network, thereby cyclically iterating by using the conjugate gradient principle until final convergence, and the training process of the artificial neural network updates the loss function by the intermediate value, thereby iteratively updating the neural network parameter θ until the training converges, and outputting the reconstructed distortion-corrected magnetic resonance image. The present invention utilizes a recursive neural network for EPI image reconstruction to improve the signal-to-noise ratio of magnetic resonance imaging and the accuracy of EPI reconstruction.

Description

Technical Field

[0001] The invention relates to the field of magnetic resonance imaging, in particular to an image reconstruction method for magnetic resonance imaging distortion correction and a magnetic resonance imaging system. Background Art

[0002] Magnetic resonance imaging (MRI) has been widely used for noninvasive imaging due to its excellent sensitivity and versatile contrast capability for soft tissues. However, the main limitation of MRI is its speed. Echo planar imaging (EPI) is a fast encoding method that has become an indispensable component of parametric mapping, functional MRI, diffusion MRI, and quantitative susceptibility mapping. Although EPI is very fast, it is seriously troubled by problems such as geometric distortion and signal loss, among which geometric distortion is the primary problem that is extremely difficult to avoid. Distortion is usually related to echo spacing, field of view, and B0 field map information. In order to fundamentally correct the distortion, it is necessary to add B0 field map information to eliminate the distortion. The existing technology has the problem that it is difficult to introduce B0 field map information to eliminate image distortion. Calculation based on experience is not only time-consuming, but also difficult to achieve good results. Summary of the invention

[0003] In view of this, the present invention aims at the problem of geometric distortion of EPI magnetic resonance imaging in the above-mentioned prior art and provides an image reconstruction method for magnetic resonance imaging distortion correction based on a model and guided by physical information.

[0004] The technical solution of the present invention is to provide a magnetic resonance image reconstruction method comprising the following steps:

[0005] 1) Acquire MRI downsampled k-space data using EPI fast imaging sequence;

[0006] 2) Obtain the B0 field map;

[0007] 3) Inputting the downsampled k-space data and the B0 field map into an artificial neural network for magnetic resonance image reconstruction, wherein the artificial neural network is a plurality of convolutional neural networks or Transformer neural networks connected in series, wherein the k-space data follows the conjugate gradient algorithm after entering the artificial neural network, and updates the intermediate value through the artificial neural network, thereby using the conjugate gradient algorithm to iterate until the final convergence, and the training process of the artificial neural network updates the loss function by the intermediate value, thereby iteratively updating the neural network parameter θ until the training converges, and outputting the reconstructed distortion-corrected magnetic resonance image .

[0008] Optionally, in step 3), the information of the B0 field image is incorporated into the coding matrix In The model solves the image to be reconstructed and maps it to the artificial neural network for magnetic resonance image reconstruction, and establishes the magnetic resonance image reconstruction equation: ;

[0009] in is the encoding matrix , represents the sampling mask, It represents the distortion correction operation, which includes the inhomogeneity information of the B0 field map, etc. This information can be obtained by pre-scanning or calculated by scanning the phase encoding gradient in the opposite direction; is the Fourier transform operation, is the SENSE encoding model, Represents a convolutional neural network or other convolution-based artificial neural network operation, is the image to be reconstructed, is the measured k-space data, θ is the neural network parameter, is the regularization weight coefficient.

[0010] As an option, the gradient of the magnetic resonance image reconstruction equation is ,in , represents the convolutional artificial neural network operation, represents the Hermitian matrix operation of E.

[0011] As an option, the corresponding loss function of the neural network is ,in I is the final reconstructed image, is the regularization term.

[0012] As an option, TV (Total Variation) regularization is used, and the corresponding formula is as follows:

[0013]

[0014] in, and Represents the horizontal and vertical gradients respectively. represents the modulus of a complex number, They refer to the pixel positions respectively.

[0015] As an option, the B0 field map calculation steps are as follows:

[0016] a) Obtain two sets of low-resolution images with different echo times (TE): Select two different echo times TE1 and TE2, where TE2 is slightly longer than TE1, and obtain two sets of phase images;

[0017] b) Through the phase difference of two sets of images The local magnetic field offset is calculated,

[0018] ;

[0019] Said and are the phase values ​​at moments TE1 and TE2 respectively;

[0020] c) Convert to frequency shift graph: ,in is the time difference between TE1 and TE2;

[0021] d) Convert to B0 field map and obtain low-resolution B0 field map: ,in is the cyclotron ratio of the proton, .

[0022] Optionally, the low-resolution B0 field map obtained in step d) is converted into a high-resolution B0 field map using an Implicit Neural Representation (INR) to improve the accuracy of the B0 field map.

[0023] The present invention also provides a magnetic resonance imaging system, comprising a main magnet, a gradient system, a radio frequency system and a magnetic resonance image reconstruction system, wherein the magnetic resonance image reconstruction system performs magnetic resonance image reconstruction based on the image reconstruction method for magnetic resonance imaging distortion correction.

[0024] By using the above method and system, the present invention has the following advantages over the prior art: the present invention introduces the B0 field map to obtain distorted image information, and maps the model to an artificial neural network, and utilizes the advantage of rapid iteration of the artificial neural network to train and iterate at the same time to reconstruct a distortion-corrected magnetic resonance image. The present invention proposes a universal distortion correction framework, which uses a recursive neural network to perform EPI image reconstruction to improve the signal-to-noise ratio of magnetic resonance imaging and the accuracy of EPI reconstruction; the present invention is applicable to a variety of magnetic resonance data sets and can solve the distortion problem in EPI image reconstruction; the neural network proposed in the present invention is a self-supervised learning method, which greatly reduces the demand for training data, simplifies the reconstruction process, and improves its applicability in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flowchart of the image reconstruction method for magnetic resonance imaging distortion correction of the present invention.

[0026] Figure 2 A comparison of 2D SE-EPI imaging results using different imaging methods.

[0027] Figure 3 is the residual map of the corresponding 2D SE-EPI imaging results.

[0028] Figure 4 This is a comparison of 3D GRE-EPI imaging results under 3T MRI using different imaging methods.

[0029] Figure 5 This is a comparison of 3D GRE-EPI imaging results under 7T MRI using different imaging methods. DETAILED DESCRIPTION

[0030] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments, but the present invention is not limited to these embodiments. The present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention. In order to make the public have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, and those skilled in the art can fully understand the present invention without the description of these details.

[0031] The technical basic solution of the present invention is to provide an image reconstruction method for magnetic resonance imaging distortion correction, comprising the following steps:

[0032] 1) Acquire MRI downsampled k-space data using EPI fast imaging sequence;

[0033] 2) Obtain the B0 field map;

[0034] 3) Inputting the downsampled k-space data and the B0 field map into an artificial neural network for magnetic resonance image reconstruction, wherein the artificial neural network is a plurality of convolutional neural networks or Transformer neural networks connected in series, wherein the k-space data follows the conjugate gradient algorithm after entering the artificial neural network, and updates the intermediate value through the artificial neural network, thereby using the conjugate gradient algorithm to iterate until the final convergence, and the training process of the artificial neural network updates the loss function by the intermediate value, thereby iteratively updating the neural network parameter θ until the training converges, and outputting the reconstructed distortion-corrected magnetic resonance image .

[0035] like Figure 1 The specific process of image reconstruction in the present invention is shown in FIG. 3 ), which shows the specific reconstruction process of the magnetic resonance image in step 3), and the information of the B0 field map is incorporated into the encoding matrix In The model solves the image to be reconstructed and maps it to the artificial neural network for magnetic resonance image reconstruction, and establishes the magnetic resonance image reconstruction equation: .in is the encoding matrix , Represents the sampling mask mask, It represents a distortion correction operation, which includes information about the inhomogeneity of the B0 field map, etc. This information can be obtained by pre-scanning or calculated by scanning the phase encoding gradient in the opposite direction. is the Fourier transform operation, is the SENSE encoding model, represents the convolution operation, is the image to be reconstructed, is the measured k-space data.

[0036] After Fourier transform operation on the downsampled k-space data and B0 field map, SENSE image reconstruction is performed, and the obtained magnetic resonance image is input into the neural network for training to obtain the corresponding magnetic resonance image reconstruction equation mentioned above. The reconstruction equation mainly completes the following operations: Under the premise of ensuring data fidelity (the first term of the reconstruction equation), the initial result (the second term of the reconstruction equation) is optimized by the neural network, and it is used as the new iterative initial result to participate in the next iteration until the image reconstruction optimization problem finally converges. The convergence condition can be set to the root mean square error of two adjacent results less than 0.01%.

[0037] The gradient of the above magnetic resonance image reconstruction equation is ,in , represents the convolution operation, represents the Hermitian matrix operation, and the corresponding loss function is ,in I is the final reconstructed image, is the regularization term. Iteratively update the parameter θ until the training converges and output the reconstructed magnetic resonance image The entire self-supervised neural network flow graph is as follows Figure 1 The design idea of ​​the neural network comes from the gradient of the above image reconstruction equation. The left part is the image reconstruction process. Based on the conjugate gradient principle, the B0 field map information is introduced into the encoding matrix. In the figure, after Fourier transform, the convolutional neural network of two conv in series is used , A new neural network is combined to update and optimize the intermediate value of each iteration and enter the CG module, and the loss function is updated accordingly, thereby promoting the conjugate gradient operation until the convergence condition is finally reached, where t is the number of iterations.

[0038] In the process of training neural networks, TV (Total Variation) regularization is used, and the corresponding formula is as follows:

[0039] ；

[0040] in, and Represents the horizontal and vertical gradients respectively. represents the modulus of a complex number, They refer to the pixel positions respectively.

[0041] In step 1), the downsampling factor of the magnetic resonance k-space data is 2-32 times.

[0042] The steps for calculating the B0 field map are as follows:

[0043] a) Acquire two sets of images with different echo times (TE): Select two different echo times TE1 and TE2, where TE2 is slightly longer than TE1, and acquire two sets of phase images.

[0044] b) Through the phase difference of two sets of images The local magnetic field offset is calculated,

[0045] ;

[0046] here and are the phase values ​​at time TE1 and TE2 respectively.

[0047] c) Convert to frequency shift graph: ,in is the time difference between TE1 and TE2.

[0048] d) Convert to B0 field map: ,in is the cyclotron ratio of the proton, .

[0049] The low-resolution B0 field map obtained in step d) is converted into a high-resolution B0 field map by using the Implicit Neural Representation (INR) to improve the accuracy of the B0 field map.

[0050] In order to verify the reliability and advantages of the method of the present invention, the present invention uses three indicators for judgment and verification, namely, root mean square error (RMSE), structural similarity (SSIM) and peak signal-to-noise ratio (PSNR). When evaluating the image quality of different reconstruction methods, the root mean square error of the full sampling reference is used, as shown below:

[0051] ,in, represents the undersampling reconstruction result, Represents a fully sampled reference image. In the present invention, a fully sampled reconstructed image is used as a reference REF, such as Figure 2 (a) and (e) in the figure.

[0052] Structural similarity (SSIM) is also used to evaluate the similarity between the reference image and the reconstructed image. It is more acceptable to human visual perception and is used to compare two images as shown in the following formula: ,in represents the undersampling reconstruction result, represents the fully sampled reference image, and They are and and They are and The variance of . and The covariance of , c 1 and c 2 are two variables used to stabilize the equation when the denominator is small to avoid singular values.

[0053] When the peak signal-to-noise ratio (PSNR) is used as the evaluation basis, the larger the value, the higher the reconstruction accuracy and the less distortion.

[0054] In a specific embodiment, the image reconstruction of 2D SE-EPI of 3T brain diffusion imaging is taken as an example, and the corresponding parameters and hardware conditions are set as follows: a diffusion experiment was performed on a 3T Prisma scanner (Siemens, Erlangen, Germany) using a 32-channel head coil. The imaging parameters are as follows: 440 readout points, oversampling rate 2, slice thickness 4 mm, spatial resolution 1×1×4 mm³, b=1000s / mm², diffusion data in R inplane = 4 and Rz = 1 under acceleration conditions by 4 shot scans (2 times in AP direction, 2 times in PA direction), TR / TE=3900ms / 83ms.

[0055] Figure 2Various reconstruction results obtained from diffusion MRI are presented, including parallel imaging (SENSE) reconstruction, Hankel structured low-rank (SLR) reconstruction, and reconstruction of the proposed method. The root mean square error (RMSE: 8.8% < 12.4%, 9.1% < 12.4%) shows that the proposed method shows higher reconstruction accuracy compared with the traditional Hankel structured low-rank (SLR) reconstruction. The root mean square errors (RMSE) of these two methods are 37.2% and 31.8%, respectively, which are significantly better than the traditional SENSE reconstruction. The structural similarity index (SSIM) evaluation further confirmed these findings, among which the proposed method has the highest SSIM (0.94, 0.93), followed by the traditional SLR (0.86, 0.85) and the traditional SENSE (0.63, 0.63). The peak signal-to-noise ratio (PSNR) values ​​also show the same trend (Proposed>SLR>SENSE). In addition, as shown by the bold orange arrows, SLR and Proposed show the effectiveness of the model-based distortion correction method. When comparing SLR and the proposed method, it can be seen that the noise is further removed.

[0056] Figure 3 The residual plots in highlight the difference between the SENSE reconstruction and the distortion-corrected reference, underscoring the effectiveness of the SLR and Proposed reconstructions in correcting distortion. The SENSE reconstructed outlines deviate significantly from the reference, illustrating the limitations of traditional SENSE in dealing with geometric distortion. In contrast, the model-based SLR and Proposed, especially the Proposed, have minimal deviations from the reference, highlighting their robustness and accuracy in correcting distortion.

[0057] In addition, the quantitative evaluation results are consistent with the visual results, further confirming the conclusion that the proposed method can not only improve the geometric accuracy but also improve the overall image quality. This comprehensive evaluation highlights the potential of the proposed method as an efficient method for accurate and reliable image reconstruction in scenes susceptible to distortion.

[0058] In another embodiment, image reconstruction is performed on 3D GRE-EPI of 3T fast whole-brain imaging, and corresponding parameters and hardware conditions are set as follows:

[0059] The imaging data set was acquired on a 3T MRI scanner (Magnetom TrioSystem; Siemens Healthineers, Erlangen, Germany) equipped with a 32-channel head coil array. A three-dimensional EPI pulse sequence was used for this acquisition. Imaging parameters included a field of view of 224 × 224 × 128 mm³, TR / TE = 72 / 36 ms, a readout of 448 points, an oversampling ratio of 2, a slice thickness of 1 mm, a spatial resolution of 1 × 1 × 1 mm³, and a flip angle of 16°. The total acquisition time was 12 seconds, including a FLASH low-resolution prescan for coil sensitivity map estimation.

[0060] Figure 4 shows the performance of different reconstruction techniques at 3T, comparing SENSE, SLR, and the proposed method. The enlarged area shows the clear advantage of the proposed method, which significantly reduces artifacts, reduces noise, and enhances detail preservation. Compared with traditional methods, the proposed method significantly reduces noise and preserves fine structural details, demonstrating its effectiveness.

[0061] In yet another embodiment, image reconstruction is performed on three-dimensional GRE-EPI for 7T fast whole-brain imaging, and corresponding parameters and hardware conditions are set as follows:

[0062] To further evaluate the performance of 3D GRE-EPI, we also performed in vivo experiments on a 7T MR Magnetom scanner (Siemens, Erlangen, Germany) using a 32-channel head coil array. inplane = 5, Rz = 1. Imaging parameters included TR / TE = 50 / 25 ms, number of slices 128, number of readout points 448, oversampling ratio 2, slice thickness 1 mm, flip angle 16°, field of view 224 × 224 × 128 mm3, spatial resolution 1 × 1 × 1 mm3. The total acquisition time at Rz = 1 was 9 s, including one FLASH prescan.

[0063] Figure 5 shows images reconstructed at 7T using SENSE, SLR, and the proposed method. The increase in magnetic field strength at 7T brings more obvious artifacts and inhomogeneities, but the proposed method effectively alleviates these problems. The yellow arrows point out areas where traditional methods have difficulty resolving distortion and noise issues, while the proposed method provides a cleaner and more accurate reconstruction.

[0064] Overall, these figures highlight the significant improvements achieved by the proposed method in both 3T and 7T scanner settings, underscoring its potential for high-quality, artifact-free MRI reconstructions. Notably, the proposed method is also able to effectively resolve phase changes caused by physiological differences.

[0065] The above only describes the preferred embodiments of the present invention, but it should not be understood as limiting the claims. The present invention is not limited to the above embodiments, and its specific methods are allowed to be changed. In short, all changes made within the protection scope of the independent claims of the present invention are within the protection scope of the present invention.

Claims

1. An image reconstruction method for magnetic resonance imaging distortion correction, characterized in that: The following steps are involved: 1) Acquire MRI downsampled k-space data using EPI fast imaging sequence; 2) Obtain the B0 field map; 3) Inputting the downsampled k-space data and the information of the B0 field map into an artificial neural network for magnetic resonance image reconstruction, wherein the artificial neural network is a plurality of convolutional neural networks or Transformer neural networks connected in series, and the k-space data follows the principle of the conjugate gradient algorithm after entering the artificial neural network, and updates the intermediate value through the neural network, thereby using the conjugate gradient principle to iterate until the final convergence, and the training process of the artificial neural network uses the intermediate value to update the loss function, thereby iteratively updating the neural network parameters. θ , until the training converges, the reconstructed distortion-corrected magnetic resonance image is output ; The steps for calculating the B0 field map are as follows: a) Obtain two sets of low-resolution images with different echo times TE: Select two different echo times TE1 and TE2, where TE2 is longer than TE1, and obtain two sets of phase images; b) Through the phase difference of two sets of images The local magnetic field offset is calculated, ； Said and are the phase values ​​at moments TE1 and TE2 respectively; c) Convert to frequency shift graph: ,in is the time difference between TE1 and TE2; d) Convert to B0 field map and obtain low-resolution B0 field map: ,in is the cyclotron ratio of the proton, ; The low-resolution B0 field map obtained in step d) is converted into a high-resolution B0 field map using the neural implicit representation.

2. The image reconstruction method for magnetic resonance imaging distortion correction according to claim 1, characterized in that: In step 3), the information of the B0 field map is incorporated into the encoding matrix In The model solves the image to be reconstructed and maps it to the artificial neural network for magnetic resonance image reconstruction. The magnetic resonance image reconstruction equation is established as: ; in is the encoding matrix , represents the sampling mask, It represents the distortion correction operation, which contains the inhomogeneity information of the B0 field map, which can be obtained by pre-scanning or calculated by scanning the phase encoding gradient in the opposite direction; is the Fourier transform operation, is the SENSE encoding model, Represents a convolutional neural network or other convolution-based artificial neural network operation, is the image to be reconstructed, is the measured k-space data, θ is the neural network parameter, is the regularization weight coefficient.

3. The image reconstruction method for magnetic resonance imaging distortion correction according to claim 2, characterized in that: The gradient of the magnetic resonance image reconstruction equation is ,in , represents the convolutional artificial neural network operation, represents the Hermitian matrix operation of E.

4. The image reconstruction method for magnetic resonance imaging distortion correction according to claim 2 or 3, characterized in that: The loss function of the magnetic resonance image reconstruction equation is ,in I is the final reconstructed image, is the regularization term.

5. The image reconstruction method for magnetic resonance imaging distortion correction according to claim 4, characterized in that: Using Total Variation regularization, the corresponding formula is as follows: ； in, and Represents the horizontal and vertical gradients respectively. represents the modulus of a complex number, They refer to the pixel positions respectively.

6. A magnetic resonance imaging system, characterized in that: The invention comprises a main magnet, a gradient system, a radio frequency system and a magnetic resonance image reconstruction system, wherein the magnetic resonance image reconstruction system performs magnetic resonance image reconstruction based on the image reconstruction method for magnetic resonance imaging distortion correction according to any one of claims 1 to 5.

Images