A Deep Learning-Based Intelligent Magnetic Resonance Imaging Method Based on Separate Amplitude and Phase Iterations

Abstract

A deep learning-based intelligent magnetic resonance imaging method based on amplitude and phase iteration is disclosed, relating to intelligent magnetic resonance imaging. Steps: 1) Acquire data, obtaining a b value of 0 s / mm. 2 1) Use other multi-excitation diffusion magnetic resonance data with different b values; 2) Generate a training dataset and reconstruct data with a b value of 0 s / mm. 2 The process involves: 1) obtaining magnetic resonance images and estimating channel sensitivity from them; 2) reconstructing multi-excitation diffuse magnetic resonance data with other b values ​​using traditional optimization algorithms as training labels; 3) designing a deep learning network model and loss function based on alternating amplitude and phase updates; 4) training an amplitude-phase iterative reconstruction network using the training dataset; and 5) inputting the multi-excitation diffuse magnetic resonance data to be reconstructed into the trained amplitude-phase iterative network to reconstruct the magnetic resonance image. Reconstruction is performed by building iterative networks for amplitude and phase separately, which enhances the fitting degree of phase and amplitude, reduces image reconstruction errors, and improves anti-artifact capability.

Description

Technical Field

[0001] This invention relates to intelligent magnetic resonance imaging, and more particularly to imaging utilizing a b-value of 0 s / mm. 2 This paper proposes a deep learning-based intelligent magnetic resonance imaging method that uses the structural information of high signal-to-noise ratio images as features to iteratively reconstruct high-quality multi-excitation diffusion-weighted magnetic resonance images by using amplitude and phase iterations separately. Background Technology

[0002] Diffusion-weighted imaging (DWI) is a special magnetic resonance imaging method that images by detecting the diffusion motion of water molecules. Compared with ordinary magnetic resonance imaging, it has extremely high value in the diagnosis of diseases such as acute ischemic stroke and early ischemic injury. Although traditional single-excitation planar echo imaging is fast, it is very sensitive to magnetic susceptibility inhomogeneity, easily loses signals, and has obvious image distortion. Multi-excitation planar echo imaging technology reduces the echo chain length, which can effectively reduce signal loss, reduce image blurring and distortion (Anderson AW, et al., "Analysis and correction of motion artifacts indiffusion-weighted imaging," Magnetic Resonance in Medicine, 1994, 32(3):379-387.), and improve the signal-to-noise ratio. However, due to the extended sampling time caused by multiple excitations and the influence of diffusion gradient, severe phase changes are introduced between different excitations, and direct synthesis of k-space data will form strong aliasing artifacts.

[0003] Reconstruction methods without navigation echoes have better reconstruction performance than those with navigation echoes, and therefore, they have received increasing attention in recent years. Such as MUSSELS (M.Mani, et.al., "Multi-shot sensitivity-encoded diffusion data recovery using structured low-rank matrixcompletion (MUSSELS)," Magnetic Resonance in Medicine, 2017, 78: 494-507.) and PLRHM (Y. Huang, et.al., "Phase-constrained reconstruction for high-resolution multi-shot diffusion weighted image," Journal of Magnetic Resonance, 2020, 312: 106690.) and other implicit phase-based reconstruction methods as well as MUSE (Chen Nkuei, et al., "A robust multi-shot scan strategy for high-resolution diffusion weighted MRI enabled by multiplexed sensitivity-encoding (MUSE)," Neuroimage, 2013, 72: 41-47.), POCS-ICE (Guo H, et al., "POCS-enhanced inherent correction of motion-induced phaseerrors (POCS-ICE) for high-resolution Multishot diffusion MRI, “MagneticResonance in Medicine, 2016, 75(1):169-180.” and PAIR and other explicit phase-based reconstruction methods.

[0004] Recently, the rapid development of deep learning has made significant contributions to many fields, including diffusion magnetic resonance imaging. Data-driven deep learning methods such as DL-MUSE (H. Zhang, et al., “Deep learning based multiplexed sensitivity-encoding (DL-MUSE) for high-resolution multi-shot DWI,” Neuroimage, 244:118632, 2021.), model-based methods such as M-MUSSELS (Aggarwal HK, et al., “MoDL-MUSSELS: model-based deep learning for Multi-shot sensitivity encoded diffusion MRI,” IEEE Transactions on Medical Imaging, 2020, 39(4):1268-1277.), and MORN (Wang F, et al., “Multiple b-value model-based residual network (MORN) for accelerated high-resolution diffusion-weighted Imaging,” IEEE Journal of Biomedical and Health Informatics, 2022, 26(9):4575-4586.) have achieved good results in diffusion-weighted magnetic resonance imaging.

[0005] The b-value is a key parameter in DWI imaging. It not only controls the degree of diffusion weighting in diffusion magnetic resonance images but also encodes different tissue features into the DWI signal. For free diffusion, the magnetic resonance signal attenuates as the b-value increases. Since the b-value is 0 s / mm... 2 The acquired images were not affected by the diffusion gradient, exhibiting clear structure and a high signal-to-noise ratio. Therefore, to maximize the signal-to-noise ratio and preserve edge information, this method introduces a b value of 0s / mm. 2The sample information (Jones DK, et al., "Optimal strategies for measuring diffusion in anisotropic systems by magnetic resonance imaging," Magnetic Resonance in Medicine, 1999, 42(3): 515-525.) improves the accuracy of the network. Summary of the Invention

[0006] The purpose of this invention is to address the problems of existing multi-excitation diffusion-weighted imaging, which struggles to resist motion artifacts and noise due to the influence of diffusion gradients, by providing a deep learning-based intelligent magnetic resonance imaging method based on separate amplitude and phase iterations. This method utilizes a b-value of 0 s / mm. 2 The structural information of the high signal-to-noise ratio image is used as the amplitude and phase of the feature fusion module to iteratively reconstruct a high-quality multi-excitation diffusion-weighted magnetic resonance image.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A deep learning-based intelligent magnetic resonance imaging method based on amplitude and phase iterations, the method comprising the following steps:

[0009] 1) Data collection: Obtain the b value as 0s / mm 2 Other b-valued multi-excitation diffusion magnetic resonance data;

[0010] 2) Generate training dataset: Reconstruct b value to 0s / mm 2 The magnetic resonance images were obtained and the channel sensitivity was estimated from them. The multi-excitation diffusion magnetic resonance data with other b values ​​were reconstructed using traditional optimization algorithms as training labels.

[0011] 3) Design of amplitude and phase iterative reconstruction network: Design a deep learning network model and loss function based on alternating amplitude and phase updates;

[0012] 4) Training the network: Use the training dataset obtained in step 2) to train the amplitude and phase iterative reconstruction network designed in step 3);

[0013] 5) Reconstructing the magnetic resonance image: Input the multi-excitation diffusion magnetic resonance data to be reconstructed into the amplitude and phase iterative network trained in step 4) to reconstruct the magnetic resonance image.

[0014] In step 1), data is collected, and the value of b is obtained as 0s / mm. 2 magnetic resonance data Magnetic resonance data with other b values Where N and M represent the length of the frequency and phase coding dimensions of the image, respectively, H is the total number of channels, and J is the number of excitations.

[0015] In step 2), a training dataset is generated, and the obtained b value is 0s / mm. 2 magnetic resonance data Where N and M represent the lengths of the frequency and phase coding dimensions of the image, respectively, and H is the total number of channels; from y b0 Channel sensitivity is estimated in the middle After performing a two-dimensional inverse Fourier transform, the amplitude is obtained by combining the channels using the least squares method. Acquire diffusion-weighted magnetic resonance data for other b-values ​​to be reconstructed. Where J represents the number of excitations, and the training labels are obtained by reconstruction using traditional optimization algorithms such as PAIR (Qian C, et al., “A paired phase and magnitude reconstruction for advanced diffusion-weighted imaging,” IEEE Transactions on Biomedical Engineering, 2023, 70(12): 3425-3435.). Image with phase Separate the phase and amplitude to obtain the phase. and amplitude Will I ref After two-dimensional Fourier transform, the following is obtained: y b1 and m b0 As network input, x ref m ref p ref As labels for network training, they together constitute a set of training data.

[0016] In step 3), the amplitude-phase iterative reconstruction network uses an iterative block as its core, and forms a whole network structure by superimposing several iterative blocks. Each iterative block contains three sub-blocks; the network structure of a single iterative block is as follows:

[0017] a) The sub-block MU (Magnitude Update module) is used for magnitude and phase iteration of the network; the iterative block of the magnitude update network can be modeled as follows:

[0018]

[0019] Where j∈{1,…,J} represents the j-th excitation, J represents the total number of excitations, l∈{1,…,L} represents the l-th channel, L represents the total number of channels, and y ljFor the diffuse magnetic resonance sampling data of the l-th channel during the j-th excitation, C l This represents the sensitivity coefficient matrix of the l-th channel. Let represent the two-dimensional Fourier transform operator, u represent the undersampling operator corresponding to the k-space sampling template of the multi-excitation diffusion weighted data, P represent the phase generated by the (j-1)th excitation, m represent the amplitude image, ||·||2 represents the 2-norm, and λ1 represents the regularization parameter. This represents the U-Net deep learning denoising network with multi-level encoding and decoding, where θ represents the parameters of the deep learning denoising network.

[0020] The sum of the output of each magnitude update iteration block and the 2-norm of the label magnitude is used as the loss function of the network. Let the label be x. ref It is divided into amplitude and phase:

[0021] m ref =|x ref |

[0022] p ref =angle(x ref )

[0023] in, This represents multiple excitations of the image vector, |·| represents taking the absolute value, angle(·) represents taking the phase, and m ref Indicates amplitude label, p ref The phase label is represented; the amplitude iterative network also incorporates a feature fusion module. This module is used to embed b0 feature information to enhance amplitude reconstruction quality; it first obtains a b value of 0s / mm. 2 The amplitude m of the b0 image b0 The amplitude image of m with other b values b1 The data is concatenated by connecting 1×1 convolutional layers end-to-end to perform feature transformation, resulting in a high-dimensional feature map. This map is then input into several selective kernel convolutional blocks for feature extraction and fusion. The selective kernel convolution has two branches, using 3×3 convolutional kernels and 5×5 dilated convolutions as branches. The number of grouped convolutions is 32, with a stride of 1. Each selective kernel convolutional block has residual connections. Finally, a 1×1 convolution is used for dimensionality reduction to obtain the final amplitude data after iteration.

[0024] The loss function for the magnitude update module is defined as follows:

[0025]

[0026] Where k∈{1,…,K} represents the iteration block index, and K represents the total number of iteration blocks. Let m represent the feature fusion module network, θ represent the network parameters to be optimized, and m represent the feature fusion module network. ref , and m b0 These represent the amplitude label, the amplitude image of b1 output by the k-th iteration block, and the amplitude image of b0, respectively.

[0027] b) The sub-block PU (Phase Update module) is used for phase updates of the amplitude and phase iteration network respectively; the iterative block of the phase update network can be modeled as follows:

[0028]

[0029] Where j∈{1,…,J} represents the j-th excitation, J represents the total number of excitations, l∈{1,…,L} represents the l-th channel, and L represents the total number of channels. p represents a U-Net degree-learning denoising network for multi-level encoding and decoding. j Let represent the phase generated by the j-th excitation, M represent the amplitude generated by the (j-1)-th excitation, λ2 represent the regularization parameter, and θ represent the parameters of the deep learning denoising network. The phase image is solved using the conjugate gradient method, and the sum of the output of each phase update iteration block and the 2-norm of the label amplitude is used as the loss function of the network.

[0030]

[0031] Finally, the overall loss function is expressed as the sum of the magnitude image loss function and the phase image loss function:

[0032]

[0033] c) The DC (Data Consistency module) sub-block is used as a data verification module for the amplitude and phase iteration networks. First, the amplitude and phase outputs of the iteration block are merged in the image domain to obtain the k-space data prediction value of the j-th excitation of the l-th channel of the iteration block. The next step is to use a data verification module to ensure that the network output data and the data collected at the sampling points are in balance. The data at the unsampled points are the network output values, while the data at the sampling points are a linear combination of the sampled data and the predicted data.

[0034] The data validation module can be modeled as follows:

[0035]

[0036] in, y represents the k-space data of the l-th channel excited by the j-th excitation of the k-th iteration block. lj Let represent the k-space data acquired by the j-th excitation of the l-th channel, n represent the index corresponding to the sampling point, Ω represent the set of k-space sampling points, and α is the linear combination parameter.

[0037] In step 4), the amplitude-phase iterative reconstruction network designed in step 3) is trained using the training input and training label data obtained in step 2). The learnable convolutional kernels and parameters in the network are updated by the Adam optimizer commonly used in deep learning, and finally the trained network model is obtained.

[0038] In step 5), the multi-excitation diffusion magnetic resonance data to be reconstructed and the b-value of 0 s / mm are used. 2 The magnetic resonance image is input into a trained amplitude-phase iterative network to reconstruct an image without motion artifacts.

[0039] Compared with the prior art, the present invention has the following outstanding technical effects:

[0040] This invention reconstructs images by building iterative networks for amplitude and phase data separately, enhancing the fitting degree of phase and amplitude, reducing image reconstruction errors, and improving anti-artifact capabilities. Deep learning algorithms can learn complex patterns and features from large amounts of data, helping to reduce potential errors in image reconstruction and improve imaging efficiency. Compared to traditional low-rank optimization algorithms such as PAIR, this invention's method can improve structural similarity by 3.99% and reduce reconstruction error by 58.35%, significantly improving signal-to-noise ratio and image quality, and suppressing motion artifacts to some extent. Attached Figure Description

[0041] Figure 1 These are DWI amplitude images with different b values. Among them, (a) shows images with a b value of 1000 s / mm. 2 The DWI amplitude image, (b) is a b-value of 0 s / mm 2 DWI amplitude image.

[0042] Figure 2 This is the network structure of the intelligent reconstruction method. (a) shows the overall network structure, and (b) shows the structure diagram of the k-th iteration block. FF represents the feature fusion module.

[0043] Figure 3 This is the output of the intelligent reconstruction method. Specifically, the reconstructed b-value is 1000 s / mm. 2 The four-stage excitation-diffusion weighted image. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] An embodiment of the present invention describes a specific process for multi-excitation k-space signal diffusion weighted reconstruction, as follows:

[0046] The first step was to collect data by scanning six volunteers with a magnetic resonance imaging (MRI) scanner using a magnetic resonance imaging (MRI) scanner with a magnetic field strength of 3.0 Tesla to obtain phase-dependent b-values ​​of 0 s / mm. 2 The diffusion-weighted magnetic resonance imaging data of the four excitations y b0 The corresponding b value is 1000s / mm 2 The diffusion-weighted magnetic resonance imaging data of the four excitations y b1 The scanning parameters are: field of view 220*220mm 2 The layer thickness is 5mm, the coil has 32 channels, and the matrix size is 180*180 after being trimmed.

[0047] The second step involved preparing the training dataset. A total of 1008 multi-excitation diffusion-weighted magnetic resonance images were obtained from the scans of six volunteers. Of these, 864 were selected as training samples and 144 as test samples. The obtained b-value was 0 s / mm. 2 magnetic resonance data That is, the lengths of the frequency dimension and the phase encoding dimension are both 180, and the number of channels is 4. From y b0 Channel sensitivity is estimated in the middle After performing a two-dimensional inverse Fourier transform, the amplitude is obtained by combining the channels using the least squares method. The b value to be reconstructed is obtained at 1000 s / mm. 2 diffusion-weighted magnetic resonance data The number of activations is 4, and the training labels are reconstructed using traditional optimization algorithms such as PAIR. Image with phase Separate the phase and amplitude to obtain the phase. and amplitude Will I ref After two-dimensional Fourier transform, the following is obtained: y b1 and m b0 As network input, x ref m ref p ref These serve as labels for network training and together constitute a set of training data. See DWI amplitude images with different b values ​​for details. Figure 1 This includes: a b value of 0s / mm. 2 diffusion-weighted magnetic resonance image amplitude ( Figure 1 The values ​​in Figures (a) and (b) are 1000 s / mm. 2 diffusion-weighted magnetic resonance image amplitude image ( Figure 1 Figure (b) in the middle.

[0048] The third step is to input the amplitude update network. The specific iterative block of the amplitude update network is as follows:

[0049]

[0050] Where j∈{1,…,J} represents the j-th excitation, J represents the total number of excitations, l∈{1,…,L} represents the l-th channel, L represents the total number of channels, and y lj For the diffuse magnetic resonance sampling data of the l-th channel during the j-th excitation, C l This represents the sensitivity coefficient matrix of the l-th channel. Let represent the two-dimensional Fourier transform operator, u represent the undersampling operator corresponding to the k-space sampling template of the multi-excitation diffusion weighted data, P represent the phase generated by the (j-1)th excitation, m represent the amplitude image, ||·||2 represents the 2-norm, and λ1 represents the regularization parameter. This represents the U-Net deep learning denoising network with multi-level encoding and decoding, where θ represents the parameters of the deep learning denoising network. The amplitude image is solved using the conjugate gradient method.

[0051] The value of b is 0s / mm 2 The b0 image is subjected to inverse Fourier transform and then channel merging to obtain the amplitude m. b0 , with the b1 image m to be reconstructed b1 The data is concatenated along the channel dimension and mapped to a high-dimensional feature map using a 1×1 convolutional kernel, serving as the input to the feature extraction network. The input first passes through a convolutional layer with 64 filters for feature extraction. The encoding operation involves three downsampling operations, each preceded by a convolutional layer with 128, 256, and 512 kernels respectively. The decoding operation includes three upsampling operations, each passing through a convolutional layer with 512, 256, and 128 kernels respectively. Subsequently, several selective kernel convolutional blocks are input for feature extraction and fusion. The selective kernel convolution has two branches, using 3×3 kernels and 5×5 dilated convolutions, with 32 grouped convolutions and a stride of 1. Each selective kernel convolutional block includes residual connections. Finally, a 1×1 convolution is used for dimensionality reduction to obtain the final amplitude data after iteration.

[0052] The loss function for the magnitude update module is defined as follows:

[0053]

[0054] Where k∈{1,…,K} represents the iterative block index. Let m represent the feature fusion module network, θ represent the network parameters to be optimized, and m represent the feature fusion module network. ref , and m b0These represent the amplitude label, the amplitude image of b1 output by the k-th iteration block, and the amplitude image of b0, respectively.

[0055] The amplitude image obtained after iteration is input into the phase update network. The specific iterative block of the phase update network is as follows:

[0056]

[0057] Where j∈{1,…,J} represents the j-th excitation, J represents the total number of excitations, l∈{1,…,L} represents the l-th channel, and L represents the total number of channels. p represents a U-Net degree-learning denoising network for multi-level encoding and decoding. j Let represent the phase generated by the j-th excitation, M represent the amplitude generated by the (j-1)-th excitation, λ2 represent the regularization parameter, and θ represent the parameters of the deep learning denoising network. The phase image is solved using the conjugate gradient method. The sum of the output of each phase update iteration block and the 2-norm of the label amplitude is used as the loss function of the network. The final loss function is:

[0058]

[0059] The overall loss function is expressed as the sum of the magnitude image loss function and the phase image loss function:

[0060]

[0061] The amplitude and phase outputs of the iterative block are combined in the image domain for data verification to obtain the k-space data prediction value of the j-th excitation of the l-th channel of the iterative block output. The data validation module can be modeled as follows:

[0062]

[0063] in, y represents the k-space data of the l-th channel excited by the j-th excitation of the k-th iteration block. lj Let represent the k-space data acquired by the j-th excitation of the l-th channel, n represent the index corresponding to the sampling point, Ω represent the set of k-space sampling points, and α is the linear combination parameter.

[0064] The fourth step involves using the training input and training label data obtained in the second step to train the amplitude and phase iterative reconstruction network designed in the third step. The Adam optimizer, commonly used in deep learning, is used to train and update the learnable convolutional kernels and parameters in the network, ultimately resulting in the trained network model.

[0065] The fifth step is to obtain a b value of 1000 s / mm. 2 Multi-excitation planar echo diffusion weighted test data and b value of 0s / mm2 The magnetic resonance image data is input into the trained intelligent reconstruction network to reconstruct an image without motion artifacts.

[0066] Figure 2 The network structure of the intelligent reconstruction method is given. (a) shows the overall network structure, and (b) shows the structure diagram of the k-th iteration block. FF represents the feature fusion module.

[0067] Figure 3 The reconstructed b value is shown to be 1000s / mm. 2 The four-stage excitation-diffusion weighted image was obtained. Experiments show that, compared with traditional low-rank-based optimization algorithms such as PAIR, the proposed scheme in this embodiment can improve the structural similarity index by 3.99% and reduce the reconstruction error index by 58.35%, effectively improving the signal-to-noise ratio and image quality, and suppressing motion artifacts to a certain extent.

[0068] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent magnetic resonance imaging method based on separate amplitude and phase iterations, characterized by the following steps: 1) Data collection: Obtain the b value as 0s / mm 2 Other b-valued multi-excitation diffusion magnetic resonance data; 2) Generate training dataset: Reconstruct b value to 0s / mm 2 The magnetic resonance images were obtained and the channel sensitivity was estimated from them. The multi-excitation diffusion magnetic resonance data with other b values ​​were reconstructed using traditional optimization algorithms as training labels. 3) Design of amplitude and phase iterative reconstruction network: Design a deep learning network model and loss function based on alternating amplitude and phase updates; The amplitude and phase iterative reconstruction network uses an iterative block as its core, and forms a whole network structure by superimposing several iterative blocks. Each iterative block contains three sub-blocks; the network structure of a single iterative block is as follows: a) Sub-block MU is used for amplitude update of the amplitude and phase iteration network respectively; the iteration block model of the amplitude update network is as follows: Where j∈{1,…,J} represents the j-th excitation, J represents the total number of excitations, l∈{1,…,L} represents the l-th channel, L represents the total number of channels, and y lj For the diffuse magnetic resonance sampling data of the l-th channel during the j-th excitation, C l This represents the sensitivity coefficient matrix of the l-th channel. Represents the two-dimensional Fourier transform operator. Let P represent the undersampling operator corresponding to the k-space sampling template of the multi-excitation diffusion-weighted data, where P represents the phase generated by the (j-1)th excitation, m represents the amplitude image, ||·||2 represents the 2-norm, and λ1 represents the regularization parameter. This represents the U-Net deep learning denoising network with multi-level encoding and decoding, where θ represents the parameters of the deep learning denoising network. The sum of the output of each magnitude update iteration block and the 2-norm of the label magnitude is used as the loss function of the network. Let the label be x. ref It is divided into amplitude and phase: m ref =|x ref | p ref =angle(x ref ) in, This represents multiple excitations of the image vector, |·| represents taking the absolute value, angle(·) represents taking the phase, and m ref Indicates amplitude label, p ref The phase label is represented; the amplitude iterative network also incorporates a feature fusion module. This module is used to embed b0 feature information to enhance amplitude reconstruction quality; it first obtains a b value of 0s / mm. 2 The amplitude m of the b0 image b0 The amplitude image of m with other b values b1 The data is concatenated by connecting 1×1 convolutional layers end-to-end to perform feature transformation, resulting in a high-dimensional feature map. This map is then input into several selective kernel convolutional blocks for feature extraction and fusion. The selective kernel convolution has two branches, using 3×3 convolutional kernels and 5×5 dilated convolutions as branches. The number of grouped convolutions is 32, with a stride of 1. Each selective kernel convolutional block has residual connections. Finally, a 1×1 convolution is used for dimensionality reduction to obtain the final amplitude data after iteration. The loss function for the magnitude update module is defined as follows: Where k∈{1,…,K} represents the iteration block index, and K represents the total number of iteration blocks. Let m represent the feature fusion module network, θ represent the network parameters to be optimized, and m represent the feature fusion module network. ref , and m b0 These represent the amplitude label, the amplitude image of b1 output by the k-th iteration block, and the amplitude image of b0, respectively. b) The sub-block PU is used for phase update of the amplitude and phase iteration networks respectively; the iterative block model of the phase update network is as follows: Where j∈{1,…,J} represents the j-th excitation, J represents the total number of excitations, l∈{1,…,L} represents the l-th channel, and L represents the total number of channels. p represents a U-Net degree-learning denoising network for multi-level encoding and decoding. j Let represent the phase generated by the j-th excitation, M represent the amplitude generated by the (j-1)-th excitation, λ2 represent the regularization parameter, and θ represent the parameters of the deep learning denoising network. The phase image is solved using the conjugate gradient method, and the sum of the output of each phase update iteration block and the 2-norm of the label amplitude is used as the loss function of the network. Finally, the overall loss function is expressed as the sum of the magnitude image loss function and the phase image loss function: c) The sub-block DC is used as a data verification module for the amplitude and phase iteration networks respectively; firstly, the amplitude and phase outputs of the iteration block are merged in the image domain to obtain the k-space data prediction value of the j-th excitation of the l-th channel of the iteration block. Next, a data verification module is used to ensure that the network output data and the data collected at the sampling points are kept in balance. The data at the unsampled points are the network output values, while the data at the sampling points are a linear combination of the sampled data and the predicted data. The data validation module is modeled as follows: in, y represents the k-space data of the l-th channel excited by the j-th excitation of the k-th iteration block. lj Let represent the k-space data acquired by the j-th excitation of the l-th channel, n represent the index corresponding to the sampling point, Ω represent the set of k-space sampling points, and α is the linear combination parameter; 4) Training the network: Use the training dataset obtained in step 2) to train the amplitude and phase iterative reconstruction network designed in step 3); 5) Reconstructing the magnetic resonance image: Input the multi-excitation diffusion magnetic resonance data to be reconstructed into the amplitude and phase iterative network trained in step 4) to reconstruct the magnetic resonance image.

2. The deep learning-based magnetic resonance intelligent imaging method based on amplitude and phase iteration as described in claim 1, characterized in that... In step 1), the collected data is used to obtain a b value of 0s / mm. 2 magnetic resonance data Magnetic resonance data with other b values Where N and M represent the length of the frequency and phase coding dimensions of the image, respectively, H is the total number of channels, and J is the number of excitations.

3. The deep learning-based magnetic resonance intelligent imaging method based on amplitude and phase iteration as described in claim 1, characterized in that... In step 2), the training dataset is generated, specifically: the obtained b value is 0s / mm. 2 magnetic resonance data Where N and M represent the lengths of the frequency and phase coding dimensions of the image, respectively, and H is the total number of channels; from y b0 Channel sensitivity is estimated in the middle After performing a two-dimensional inverse Fourier transform, the amplitude is obtained by combining the channels using the least squares method. Acquire diffusion-weighted magnetic resonance data for other b-values ​​to be reconstructed. Where J represents the number of activations, and the training labels are reconstructed using an optimization algorithm. Image with phase Separate the phase and amplitude to obtain the phase. and amplitude Will I ref After two-dimensional Fourier transform, the following is obtained: y b1 and m b0 As network input, x ref m ref p ref As labels for network training, they together constitute a set of training data.

4. The deep learning-based intelligent magnetic resonance imaging method based on amplitude and phase iteration as described in claim 1, characterized in that... In step 4), the training network uses the training input and training label data obtained in step 2) to train the amplitude-phase iterative reconstruction network designed in step 3). The learnable convolutional kernels and parameters in the network are updated by training the Adam optimizer commonly used in deep learning, and the trained network model is obtained.

5. The deep learning-based intelligent magnetic resonance imaging method based on amplitude and phase iteration as described in claim 1, characterized in that... In step 5), the reconstructed magnetic resonance image is obtained by combining the multi-excitation diffusion magnetic resonance data to be reconstructed with a b-value of 0 s / mm. 2 The magnetic resonance image is input into a trained amplitude-phase iterative network to reconstruct an image without motion artifacts.

Images