A multi-b value joint high-definition diffusion magnetic resonance intelligent imaging method

Abstract

The application relates to a multi-b-value joint high-definition diffusion magnetic resonance intelligent imaging method, and relates to image reconstruction. 1) data acquisition, containing multi-shot interleaved echo planar diffusion weighted data with b values of 0 s / mm 2 and other b values; 2) generating a training data set, reconstructing a magnetic resonance image with a b value of 0 s / mm 2 , estimating channel sensitivity from the image with a b value of 0; using PAIR to reconstruct diffusion weighted data with other b values as training labels; 3) designing a diffusion weighted intelligent imaging model and a loss function; 4) training a joint multi-b-value intelligent reconstruction network using the training data set; 5) acquiring multi-shot diffusion weighted test data to be reconstructed; 6) reconstructing multi-shot diffusion weighted data using the trained intelligent reconstruction network to obtain a reconstructed image. The time for reconstructing a single picture by using a traditional algorithm can be reduced from 1-2 min to 0.2 s, multi-shot diffusion weighted images can be quickly reconstructed, the signal-to-noise ratio and the image quality can be improved, and Nyquist ghosting can be suppressed.

Description

TECHNICAL FIELD

[0001] The present application relates to image reconstruction, in particular to a multi-b value joint high-definition diffusion magnetic resonance intelligent imaging method, which uses the structural information of high signal-to-noise ratio images with b value of 0 s / mm 2 To jointly reconstruct high-quality magnetic resonance diffusion weighted images by using an intelligent reconstruction network. BACKGROUND

[0002] Diffusion weighted imaging is a non-invasive way to detect the diffusion motion of water molecules, which has better sensitivity to tumors, edema, etc. than conventional magnetic resonance imaging (V. Baliyan, et al., “Diffusion weighted imaging: technique and applications,” World Journal of Radiology, 8:785, 2016.). Although the traditional single-shot echo planar imaging sequence has fast sampling speed and is not sensitive to motion, it often makes the single-shot echo planar resolution poor and the image distortion obvious due to the influence of the inhomogeneous field (F. Farzaneh, et al., “Analysis of T2 limitations and off-resonance effects on spatial resolution and artifacts in echo-planar imaging,” Magnetic Resonance in Medicine, 14:123-139, 1990.). The multi-shot echo planar imaging technology reduces the length of the echo chain and well reduces the influence of the inhomogeneous field, which can obtain images with high resolution and small distortion (H. An, et al., “Qualitative and quantitative comparison of image quality between single-shot echo-planar and interleaved multi-shot echo-planar diffusion-weighted imaging in female pelvis,” European Radiology, 30:1876-1884, 2020.). However, due to the influence of the diffusion gradient, there is a serious phase error between different shots, which leads to serious motion artifacts (A. W. Anderson, et al., “Analysis and correction of motion artifacts in diffusion weighted imaging,” Magnetic Resonance in Medicine, 32:379-387, 1994.).

[0003] In recent years, many reconstruction methods based on low-rank iterative models have been used to correct the motion phase between different shots to achieve motion artifact-free reconstruction. MUSSELS (M. M. Mani, et al., “Multi-shot sensitivity-encoded diffusion data recovery using structured low-rank matrix completion (MUSSELS),” Magnetic Resonance in Medicine, 78:494-507, 2017.) and PLRHM (Y. Huang, et. al., “Phase-constrained reconstruction for high-resolution multi-shot diffusion weighted image,” Journal of Magnetic Resonance, 312:106690, 2020.) and other reconstruction methods based on low-rank iterative models have a long reconstruction time due to the need for matrix decomposition, which greatly limits the application of this type of method.

[0004] Recently, deep learning methods have shown great potential in magnetic resonance imaging (Q. Yang, et al., “Physics-driven synthetic data learning for biomedical magnetic resonance,” IEEE Signal Process Magazine, DOI:10.1109 / MSP.2022.3183809, 2022.), MODL-MUSSELS (Aggarwal. H. K., et al., “MoDL-MUSSELS: Model-based deep learning for multi-shot sensitivity-encoded diffusion MRI,” IEEE Transactions on Medical Imaging, 39:1268-1277, 2019.) and DL-MUSE (H. Zhang, et al., “Deep learning based multiplexed sensitivity-encoding (DL-MUSE) for high-resolution multi-shot DWI,” Neuroimage, 244:118632, 2021.) and other deep learning methods can quickly diffuse weighted magnetic resonance imaging. However, due to the influence of diffusion gradient, the signal-to-noise ratio of the reconstructed diffusion weighted image is significantly reduced, and the edge loss is obvious. In order to protect the edge and structure information of the diffusion weighted image and improve the signal-to-noise ratio, the method adds the image prior information of b value 0s / mm 2 . Since the image collected by b value 0s / mm 2 is not affected by the diffusion gradient, it often has clear structure and high signal-to-noise ratio, and after adding the image prior information of b value 0s / mm 2 , the edge information can be well protected, the signal-to-noise ratio can be improved, and the image quality can be enhanced. SUMMARY

[0005] The purpose of the present application is to solve the problem that the multi-shot diffusion weighted imaging is difficult to well resist motion artifacts and noise due to the influence of diffusion gradient, and provide a multi-b value joint high-definition diffusion magnetic resonance intelligent imaging method.

[0006] The present application comprises the following steps:

[0007] 1) Collect data, including b value 0s / mm 2 and other b value multi-shot interleaved echo diffusion weighted data;

[0008] 2) Generate training dataset, reconstruct b-value 0 s / mm 2 weighted data as training labels using traditional optimization algorithms like PAIR (C. Qian, et al., “A paired phase and magnitude reconstruction for advanced diffusion-weighted imaging,” arXiv preprint, arXiv:2203.14559. 2022.) etc.

[0009] 3) Design diffusion-weighted intelligent imaging model and loss function;

[0010] 4) Train joint multi-b-value intelligent reconstruction network using training dataset;

[0011] 5) Obtain multi-shot diffusion-weighted test data to be reconstructed;

[0012] 6) Reconstruct multi-shot diffusion-weighted data using trained intelligent reconstruction network to obtain reconstructed image.

[0013] In step 1), the collected data is multi-shot interleaved echo diffusion-weighted data containing b-value 0 s / mm 2 and at least one other b-value.

[0014] In step 2), the training dataset is prepared, and the obtained b-value is 0 s / mm 2 of magnetic resonance data where N and M represent the lengths of the frequency and phase encoding dimensions of the image, respectively, and H is the total number of channels; from Y b0 the channel sensitivity is estimated After two-dimensional inverse Fourier transform, the amplitudes of each channel are combined by the least squares method Obtain diffusion-weighted magnetic resonance data of other b-values to be reconstructed where J is the number of shots, and the training labels are obtained by reconstruction using traditional optimization algorithms such as PAIR (C. Qian, et al., “A paired phase and magnitude reconstruction for advanced diffusion-weighted imaging,” arXiv preprint, arXiv:2203.14559. 2022.) etc. Separate the phase and amplitude of the image with phase to obtain phase and amplitude I label is subjected to two-dimensional Fourier transform to obtain Y b1 and M b0 as network input, X label , M label , P label as network training labels, which together constitute a set of training data.

[0015] In step 3), the designed diffusion-weighted intelligent imaging network contains two parts of implicit phase model and explicit phase model, wherein the implicit phase model is as follows:

[0016]

[0017] wherein is the obtained diffusion-weighted data of each channel of k-space of each excitation, wherein N and M are the lengths of the frequency and phase encoding dimensions of the image respectively, H is the number of channels, and J is the number of excitations; is the k-space of the phase diffusion-weighted image to be solved, is the channel sensitivity, is an undersampling operator corresponding to the k-space sampling template of multi-excitation diffusion-weighted data, is a two-dimensional Fourier transform operator, is a two-dimensional inverse Fourier transform operator, and is a Hadamard product, represents the square of the Frobenius norm, and λ1 is a regularization term coefficient, is a multi-layer convolutional neural network;

[0018] The implicit phase model is solved by a conjugate gradient descent algorithm to obtain:

[0019]

[0020]

[0021] wherein formula (2) is an implicit solving network, and formula (3) is a data verification term; each formula (2) and formula (3) constitutes a network iteration block, and the implicit phase model has T iteration blocks; t represents the tth iteration block, is the input of the tth implicit solving network , is the output of the tth implicit solving network ; in the data verification term of each iteration block, and are Fourier transform and inverse Fourier transform operators respectively, is an undersampling operator corresponding to the k-space sampling template of multi-excitation diffusion-weighted data, is a conjugate operator thereof, C is a channel sensitivity, and C * is a conjugate matrix of the channel sensitivity; E is an identity matrix, λ1 is a learnable regularization parameter, is a Hadamard product; Y b1 is obtained diffusion-weighted k-space data; Z t is an output variable after passing through the tth iteration network ; in the implicit solving network of each iteration block, For the multi-layer convolutional neural network, each layer of convolution contains multiple two-dimensional convolution kernels, each convolutional layer is connected by a linear rectifier function, and the input of each layer is the output of the previous layer; when t = 0, as the initialization input of the network;

[0022] The loss function of the diffusion-weighted magnetic resonance implicit network model is defined as follows:

[0023]

[0024] wherein is the output result of the tth iteration block after the nth training sample is input into the network, N is the total number of samples, and T is the total number of iteration blocks of the network, represents the square of the Frobenius norm, is the nth training label;

[0025] The second part of the designed diffusion-weighted intelligent imaging network is the explicit phase model, which separates the amplitude and phase of the output result of the last iteration block of the implicit network model to obtain the phase and the amplitude as the input of the explicit phase model;

[0026] The explicit phase model includes a phase updating network and an amplitude updating network wherein Each layer of convolution of the network contains multiple two-dimensional convolution kernels, each convolutional layer is connected by a linear rectifier function, and the input of each layer is the output of the previous layer; is a multi-scale convolutional network, and the input is obtained by splicing and the pre-reconstructed magnetic resonance image amplitude M b0 with a b value of 0 s / mm 2 ; First, a convolutional layer containing multiple filters is used for feature extraction, and then a coding and decoding operation is performed to extract information of different scales; multiple down-sampling is performed in the coding part, and the same number of up-sampling is performed in the decoding part, which can well avoid the loss of edge detail features, well protect and extract features of different b values, and guarantee the reconstruction quality.

[0027] The loss function of the explicit phase model is as follows:

[0028]

[0029] wherein is the phase of the output of the Tth iteration block of the implicit phase model after the nth sample, the amplitude of the output of the nth sample after T implicit phase model iteration blocks and the b value corresponding to the nth sample is 0 s / mm 2 the amplitude of the output of the nth sample after T implicit phase model iteration blocks is spliced in the dimension of the number of excitations; and respectively, the phase and amplitude of the nth training label; N is the total number of samples, is a phase update network, is an amplitude update network, represents the square of the Frobenius norm;

[0030] The total loss function of the multi-b-value joint intelligent reconstruction network is as follows:

[0031]

[0032] In step 4), the multi-b-value joint intelligent reconstruction network designed in step 3 is trained using the training input and training label data obtained in step 2, the learnable convolution kernel and parameters in the update network are trained through the Adam optimizer commonly used in deep learning, and finally the trained network model is obtained.

[0033] In step 5), the acquired multi-excitation diffusion-weighted test data to be reconstructed is multi-excitation interleaved planar echo diffusion-weighted data read out in the phase encoding dimension.

[0034] In step 6), the acquired multi-excitation planar echo diffusion-weighted data and the magnetic resonance image with a b value of 0 s / mm 2 are input into the trained intelligent reconstruction network, and a motion artifact-free image is reconstructed.

[0035] The present application can reduce the time for reconstructing a single picture by traditional algorithms such as PAIR from 1-2 min to 0.2 s, quickly reconstruct multi-excitation diffusion-weighted images, greatly improve the signal-to-noise ratio and image quality, and to some extent, suppress the Nyquist ghost. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a four-excitation diffusion-weighted image with a b value of 0 s / mm 2 and the estimated channel sensitivity.

[0037] Figure 2 is the input of the intelligent reconstruction method, i.e., the four-excitation diffusion-weighted image with a b value of 1000 s / mm 2 to be reconstructed.

[0038] Figure 3 is the output of the intelligent reconstruction method, i.e., the reconstructed b value of 1000 s / mm 2four-shot diffusion-weighted images. DETAILED DESCRIPTION

[0039] In order to make the objects, technical solutions and advantages of the present application clearer, the following embodiments will further illustrate the present application with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application. On the contrary, the present application covers any alternatives, modifications, equivalent methods and solutions defined by the claims within the spirit and scope of the present application. Further, in order to make the public better understand the present application, some specific details are described in the following detailed description of the present application. The present application can also be completely understood without these details.

[0040] The embodiment of the present application is a specific process of multi-shot k-space signal high-resolution diffusion-weighted reconstruction. In combination with the accompanying drawings, a multi-b-value joint high-definition diffusion magnetic resonance intelligent imaging method proposed by the present application is described in detail.

[0041] The specific implementation process is as follows:

[0042] First, collect data. Six volunteers are scanned using a magnetic resonance scanner with a magnetic field strength of 3.0 Tesla to obtain phase b-value 0 s / mm 2 four-shot diffusion-weighted magnetic resonance image data Y b0 and b-value 1000 s / mm 2 four-shot diffusion-weighted magnetic resonance image data Y b1 The acquisition parameters are: field of view 220*220 mm 2 , slice thickness 5 mm, coil 32 channels, and matrix size 180*180 after cropping.

[0043] Second, prepare the training set. The obtained magnetic resonance data with b-value 0 s / mm 2 is subjected to two-dimensional Fourier transform to obtain phase diffusion-weighted images , i.e., the length of the frequency dimension and the phase encoding dimension is 180, and the number of channels is 4. The channel sensitivity is estimated therefrom. The amplitudes are obtained by combining each channel by the least square method. N and M are the lengths of the frequency and phase coding dimensions of the image, respectively; H is the number of channels; and J is the number of firings. After reconstruction using the traditional optimization algorithm PAIR (C. Qian, et al., “A paired phase and magnitude reconstruction for advanced diffusion-weighted imaging,” arXiv preprint, arXiv:2203.14559.2022.), training labels are obtained. That is, the lengths of the frequency dimension and the phase encoding dimension are both 180, and the number of excitations is 4. The image with phase is... Separate the phase and amplitude to obtain the phase. and amplitude Will I label After two-dimensional Fourier transform, the following is obtained: To prepare training and validation sets, 1008 phase-coordinated magnetic resonance images were obtained from 6 volunteers. 864 training samples and 144 validation samples were selected. Each sample... Y b1 and M b0 As network input, X label M label P label These serve as labels for network training and together constitute a set of training data. The value of b is 0 s / mm. 2 Diffusion-weighted magnetic resonance image amplitude and estimated channel sensitivity, as follows Figure 1 As shown.

[0044] The third step is to input the implicit phase model into the unfolded network. The specific iterations of the unfolded network are as follows:

[0045]

[0046]

[0047] Formula (2) is the implicit solution network, and formula (3) is the data validation term. Each formula (2) and formula (3) constitutes a network iteration block. The implicit phase model has a total of T iteration blocks, and T is set to 1. It is the t-th implicit solution network. Input, It is the t-th implicit solution network. The output. In the data validation items of each iteration block, and These are the Fourier transform and inverse Fourier transform operators, respectively. It is the undersampling operator corresponding to the k-space sampling template of multi-excitation diffusion-weighted data. It is its conjugate operator, C is the channel sensitivity, C * is the conjugate matrix of the channel sensitivity. E is the identity matrix, λ1 is a learnable regularization parameter set to 0.01. ⊙ is the Hadamard product. Y b1 This is the obtained diffusion-weighted k-space data. Z t It is after the t-th iteration of the network The output variables. In the implicit solution network of each iteration block, It is a multi-layer convolutional neural network, each convolutional layer contains multiple two-dimensional convolutional kernels, and the convolutional layers are connected by linear rectified functions, with the input of each layer being the output of the previous layer.

[0048] Set the number of iterations T to 1. It is the k-space of the phase-spread weighted image to be solved. When t = 0, As initial input for the network. It consists of a residual convolutional network with three residual blocks. Each residual block contains two convolutional layers and one residual connection. The number of convolutional kernels is 64. The convolutional layers are connected by a linear rectified function, and the input of each layer is the output of the previous layer.

[0049] The loss function of the diffusion-weighted magnetic resonance implicit network model is defined as follows:

[0050]

[0051] in, This is the output of the t-th iteration block after the n-th training sample is input into the network, where N is the total number of samples and T is the total number of iteration blocks in the network. This represents the square of the Frobenius norm. This is the nth training label. Formula (14) is solved using conjugate gradient descent, with the number of iterations T set to 1, and the output of the last iteration block of the expanded network is used. After performing a two-dimensional inverse Fourier transform, we obtain And Decomposed into amplitude and phase As input to the display phase model.

[0052] The explicit phase model includes a phase update network. And amplitude update network Will The reconstructed b value is 0s / mm 2 Magnetic resonance image amplitude The result is obtained by concatenating the data along the dimension of the number of excitations. Will Input amplitude update network The multi-scale convolutional network is composed of an input that is first subjected to a convolutional layer containing 64 filters for feature extraction, an encoding operation subjected to three times of down-sampling, and three times of down-sampling each subjected to a convolutional layer with the number of convolutional kernels being 128, 256 and 512, respectively. The decoding operation includes three times of up-sampling, and three times of up-sampling each subjected to a convolutional layer with the number of convolutional kernels being 512, 256 and 128, respectively. The input phase update network is The residual convolutional network is composed of three residual blocks, each residual block including two convolutional layers and a residual connection, the number of convolutional kernels being 64, each convolutional layer being connected by a linear rectifier function, and the input of each layer being the output of the previous layer.

[0053] The loss function of the explicit phase model is as follows:

[0054]

[0055] wherein, is the phase of the output of the nth sample after being subjected to the implicit phase model, is the amplitude of the output of the nth sample after being subjected to the implicit phase model the b value corresponding to the nth sample is 0s / mm 2 the amplitude of the output of the nth sample after being subjected to the implicit phase model is obtained by concatenating in the excitation number dimension; and are the phase and amplitude of the nth training label, respectively; N is the total number of samples, is the phase update network, is the amplitude update network, represents the square of the Frobenius norm.

[0056] The total loss function of the multi-b-value joint intelligent reconstruction network is as follows:

[0057]

[0058] In the fourth step, the multi-b-value joint intelligent reconstruction network designed in the third step is trained using the training input and training label data obtained in the second step, the learnable convolutional kernels and parameters in the update network are trained by the Adam optimizer commonly used in deep learning, and finally the trained network model is obtained.

[0059] In the fifth step, a magnetic resonance scanner with a magnetic field strength of 3.0 Tesla is used to scan a volunteer to obtain four excitation diffusion weighted magnetic resonance images as test data, a total of 144 samples, and the acquisition parameters are as follows: b value 1000s / mm 2 , diffusion direction 12, field of view 220*220mm 2, layer thickness 5mm, number of layers 12, coil 32 channels, and the matrix size is 180*180 after cutting.

[0060] In the sixth step, the acquired multi-shot echo planar diffusion weighted test data with b value of 1000s / mm 2 and the magnetic resonance image data with b value of 0s / mm 2 are input into the trained intelligent reconstruction network to obtain the image without motion artifacts. Figure 2 is the input of the intelligent reconstruction method, i.e. 2 the four-shot diffusion weighted image with b value of 1000s / mm Figure 3 is the output of the intelligent reconstruction method, i.e. 2 the four-shot diffusion weighted image with b value of 1000s / mm Compared with the traditional low-rank based optimization algorithm, the embodiment of the present application can reduce the time of reconstructing a single picture by PAIR and other traditional algorithms from 1-2min to 0.2s, greatly reduce the reconstruction time, greatly improve the signal-to-noise ratio and image quality, to a certain extent, suppress the Nyquist ghost, and solve the problem that the prior art cannot well resist motion artifacts and noise.

Claims

1. A multi-b value joint high-definition diffusion magnetic resonance intelligent imaging method, characterized by The method comprises the following steps: 1) Acquire data, including multi-shot echo-planar diffusion weighted data with b values of 0 s / mm 2 and other b values; 2) Generate training dataset, reconstruct b-value 0 s / mm 2 MR images with b-value 0 s / mm2, and estimate channel sensitivity from them; use optimization algorithm to reconstruct other b-value diffusion weighted data as training labels, the specific steps are: Prepare training data set, the obtained b value is 0 s / mm 2 Magnetic resonance data Wherein, N and M represent the length of frequency and phase encoding dimension of image respectively, H is the total number of channels; From Y b0 Estimate the channel sensitivity After two-dimensional inverse Fourier transform, the amplitude of each channel is obtained by least square method Obtain the diffusion weighted magnetic resonance data of other b values to be reconstructed Wherein J is the number of excitations, and the training label is obtained by optimization algorithm Separate the phase and amplitude of the image with phase Get phase And amplitude I label After two-dimensional Fourier transform, Y Y b1 And M b0 As network input, X label , M label , P label As network training label, together constitute a set of training data; 3) designing a diffusion-weighted intelligent imaging model and a loss function; 4) training a joint multi-b-value intelligent reconstruction network using a training data set; 5) obtaining multi-shot diffusion-weighted test data to be reconstructed; 6) reconstructing the multi-shot diffusion-weighted data using the trained intelligent reconstruction network to obtain a reconstructed image.

2. The multi-b value joint high-definition diffusion magnetic resonance intelligent imaging method of claim 1, wherein In step 1), the data is acquired as multi-shot interleaved echo planar diffusion weighted data containing a b-value of 0 s / mm 2 and at least one other b-value.

3. The multi-b value joint high-definition diffusion magnetic resonance intelligent imaging method of claim 1, wherein In step 3), the diffusion-weighted intelligent imaging model comprises two parts of an implicit phase model and an explicit phase model, wherein the implicit phase model is as follows: wherein, is the diffusion-weighted data of each channel of each excitation of the k-space obtained, wherein N and M are the lengths of the frequency and phase encoding dimensions of the image respectively, H is the number of channels, and J is the number of excitations; is the k-space of the phase diffusion-weighted image to be solved, is the channel sensitivity, is the k-space sampling template corresponding to the undersampling operator of the multi-excitation diffusion-weighted data, is a two-dimensional Fourier transform operator, is a two-dimensional inverse Fourier transform operator, and is a Hadamard product, represents the square of the Frobenius norm, and λ1 is a regularization term coefficient, is a multi-layer convolutional neural network; The implicit phase model is solved by a conjugate gradient descent algorithm to obtain: wherein formula (2) is an implicit solving network, and formula (3) is a data check term; each formula (2) and formula (3) constitutes a network iteration block, and the implicit phase model has T iteration blocks in total; t represents the tth iteration block, is the input of the tth implicit solving network , is the output of the tth implicit solving network ; in the data check term of each iteration block, and are Fourier transform and inverse Fourier transform operators respectively, is an undersampling operator corresponding to a k-space sampling template of multi-shot diffusion weighted data, is a conjugate operator thereof, C is a channel sensitivity, and C * is a conjugate matrix of the channel sensitivity; E is a unit matrix, λ1 is a learnable regularization parameter, and is a Hadamard product; Y b1 is obtained diffusion weighted k-space data; Z t is an output variable after the tth iteration network ; in the implicit solving network of each iteration block, is a multi-layer convolutional neural network, each layer of convolution contains multiple two-dimensional convolution kernels, the convolutional layers are connected by a linear rectifier function, and the input of each layer is the output of the previous layer; when t=0, serves as an initialization input of the network; The loss function of the diffusion-weighted magnetic resonance implicit network model is defined as follows: wherein, is the output result of the tthiteration block after the nthtraining sample input network, N is the total number of samples, T is the total number of iteration blocks of the network, denotes the square of the Frobenius norm, is the nthtraining label; The second part of the diffusion-weighted intelligent imaging network is an explicit phase model that takes the output of the last iteration block of the implicit network model and separates the magnitude and phase to obtain the phase and the magnitude as inputs to the explicit phase model; Explicit phase model includes phase update network and amplitude update network wherein Each layer of convolution of the network includes multiple two-dimensional convolution kernels, each convolution layer is connected by a linear rectifier function, and each layer input is the output of the previous layer; is a multi-scale convolution network, the input is to and the pre-reconstructed b value is 0s / mm 2 The magnetic resonance image amplitude M b0 is obtained by splicing in the dimension of the number of excitations; First, a convolution layer containing multiple filters is used for feature extraction, and then a coding and decoding operation is performed to extract information of different scales; multiple down-sampling is performed in the coding part, and the same number of up-sampling is performed in the decoding part, so as to avoid losing edge detail features, protect and extract features of different b values, and guarantee the reconstruction quality; The loss function of the explicit phase model is as follows: wherein is the phase of the output of the nth sample after T implicit phase model iteration blocks, is the amplitude of the output of the nth sample after T implicit phase model iteration blocks and the b value corresponding to the nth sample is 0 s / mm 2 is the amplitude of the output of the nth sample after T implicit phase model iteration blocks is obtained by concatenating in the dimension of the number of excitations; and are the phase and amplitude of the nth training label, respectively; N is the total number of samples, is a phase update network, is an amplitude update network, denotes the square of the Frobenius norm. The total loss function of the multi-b-value joint intelligent reconstruction network is as follows:

4. The multi-b value joint high-definition diffusion magnetic resonance imaging method of claim 1, wherein In step 4), the specific steps of training the joint multi-b-value intelligent reconstruction network using the training data set are as follows: the multi-b-value joint intelligent reconstruction network designed in step 3) is trained using the training input and training label data obtained in step 2), the learnable convolution kernel and parameters in the network are updated by using the Adam optimizer commonly used in deep learning, and finally the trained network model is obtained.

5. The multi-b value joint high-definition diffusion magnetic resonance imaging method of claim 1, wherein In step 5), the obtained multi-shot diffusion-weighted test data to be reconstructed is multi-shot interleaved planar echo diffusion-weighted data read out in a phase encoding dimension.

6. The multi-b value joint high-definition diffusion magnetic resonance imaging method of claim 1, wherein In step 6), the multi-shot diffusion weighted data is reconstructed by using the trained intelligent reconstruction network. The acquired multi-shot planar echo diffusion weighted data and the magnetic resonance image with b value of 0 s / mm 2 are input into the trained intelligent reconstruction network to obtain an image without motion artifacts.

Images