A physical-based multi-site high-definition diffusion magnetic resonance imaging method

Abstract

The application discloses a physical basis model multi-position high-definition diffusion magnetic resonance imaging method, and relates to a multi-shot planar echo diffusion magnetic resonance imaging method. The method comprises the following steps: 1) acquiring k-space data of a diffusion magnetic resonance imaging to be reconstructed, which is collected by a multi-shot planar echo sequence; 2) designing a bidirectional one-dimensional low-rank reconstruction model based on a local smooth phase; 3) generating data by using a physical diffusion model; 4) training a basis model neural network by using the physical generated data; 5) solving the bidirectional one-dimensional low-rank reconstruction model; and 6) combining reconstruction parameters prompted by the basis model neural network, and generalizing reconstruction to obtain multiple-position high-definition diffusion magnetic resonance images without motion artifacts. The method realizes robust high-definition magnetic resonance diffusion imaging of multiple challenging positions without adjusting parameters by using a bidirectional one-dimensional low-rank optimization reconstruction method and a neural network to adaptively optimize reconstruction parameters.

Description

TECHNICAL FIELD

[0001] The present application relates to a multi-organ high-resolution diffusion magnetic resonance image reconstruction method, in particular to a physical-based multi-organ high-resolution diffusion magnetic resonance imaging method for multi-shot echo planar diffusion magnetic resonance images of complex phase organs such as abdomen and pelvic cavity. BACKGROUND

[0002] Diffusion magnetic resonance imaging is a non-ionizing radiation and non-invasive technique for detecting water molecule diffusion in living biological tissues, which plays an important role in the study and detection of brain function, tumors, arthritis and other diseases in multiple organs.

[0003] Conventional single-shot echo planar imaging (ss-EPI) is a commonly used diffusion imaging method in clinical diagnosis, which has fast imaging speed, but is sensitive to field inhomogeneity, leading to serious image distortion, and the image resolution is limited by the echo chain length. To reduce distortion and improve resolution, multi-shot EPI fills k-space by multiple shots to shorten the echo chain, reduce distortion, and improve resolution and signal-to-noise ratio, but motion between shots will bring motion phase errors, leading to image artifacts.

[0004] In recent years, the brain multi-shot diffusion imaging technology has been developed maturely, but the motion complex parts still need to be studied. MUSE (N. Chen et al. A robust multi-shot scan strategy for high-resolution diffusion weighted MRI enabled by multiplexed sensitivity-encoding (MUSE). Neuroimage. 2013 May 15;72:41-47) and MUSSELS (M. Mani et al. Multi-shot sensitivity-encoded diffusion data recovery using structured low-rank matrix completion (MUSSELS). Magnetic Resonance in Medicine. 2016 Aug 23;78(2):494-507) and other methods have good reconstruction effect in the brain, but it is difficult to reconstruct high-quality images in the abdomen. DONATE (C. Qian, et al. Fast and ultra-high shot diffusion MRI image reconstruction with self-adaptive Hankel subspace. Medical Image Analysis. 2025 May;102:103546) realizes the high-definition reconstruction of the brain with ultra-high shot number, but in the reconstruction of the complex phase part (such as the eye, the top of the skull, the abdomen, and the pelvic cavity), there are some layer and diffusion direction images with artifacts that are not corrected cleanly. The existing method has poor robustness when facing complex phase data, and needs to adjust the reconstruction parameters for specific layers or directions.

[0005] To improve the robustness of reconstruction, i.e. to achieve full-slice, full-direction high-quality robust reconstruction using a single reconstruction parameter. Artificial intelligence network based on generated data improves the robustness of reconstruction. PIDD (C. Qian et al. published in the conference 2023 IEEE 20th International Symposium on Biomedical Imaging (2023 IEEE International Symposium on Biomedical Imaging, ISBI) (Conference Address Cartagena, Colombia, Pages 1-5, 2023) “Physics-informed deep diffusion MRI reconstruction: Break the bottleneck of training data in artificial intelligence”) generates data with complex phase using high-order phase model, and realizes robust high-quality brain image reconstruction. However, in the part with complex phase (such as abdomen), the phase difference of different organs is large, and PIDD may cause signal loss and cannot robustly reconstruct the challenging parts such as abdomen. SUMMARY

[0006] The purpose of the present application is to solve the problems of incomplete motion artifact correction, insufficient robustness of image reconstruction, the need for manual adjustment of reconstruction parameters for different parts or slices, and the easy signal loss of existing methods in multi-shot diffusion magnetic resonance imaging with complex phase parts (such as abdomen, pelvic cavity, etc.), and to provide a physical-based multi-site high-definition diffusion magnetic resonance imaging method that realizes adaptive optimization of parameters by combining physical-based models with neural networks, robust reconstruction of multi-site high-definition artifact-free images without manual adjustment, and realizes robust, high-quality multi-site high-definition diffusion magnetic resonance imaging.

[0007] To achieve the above-mentioned purpose of the application, the present application provides the following technical solutions.

[0008] The present application provides a physical-based multi-site high-definition diffusion magnetic resonance imaging method, comprising the following steps:

[0009] 1) Obtain the k-space data of the diffusion magnetic resonance imaging to be reconstructed acquired by the multi-shot echo planar sequence, for multi-site high-definition diffusion imaging reconstruction;

[0010] 2) Design a bidirectional one-dimensional low-rank reconstruction model based on local smooth phase; the model includes a two-dimensional Fourier transform operator, a channel sensitivity operator, a structured low-rank matrix extraction operator and a kernel norm constraint term, which are used to describe the low-rank characteristics of multi-shot data;

[0011] 3) Data generation using physical diffusion models, including: acquiring amplitude images using multi-organ phantoms, generating complex phase data of corresponding orders for different organs, and combining amplitude and phase to generate noisy and noise-free fully sampled multi-shot k-space data;

[0012] 4) Using physically generated data to train the base model neural network; by converting the noisy k-space data into a structured low-rank matrix and performing singular value decomposition, the neural network learns the optimal singular value cutoff parameters;

[0013] 5) Solve the bidirectional one-dimensional low-rank reconstruction model and achieve signal recovery by preserving the key singular values ​​and corresponding subspaces of the structured low-rank matrix;

[0014] 6) Combined with the reconstruction parameters suggested by the base model neural network, multi-excitation k-space data are generalized to reconstruct high-definition diffusion magnetic resonance images of various parts without motion artifacts.

[0015] In step 1), the k-space data of the diffusion magnetic resonance imaging to be reconstructed acquired by the multi-excitation echo-planar sequence is acquired, specifically: the k-space data of the diffusion magnetic resonance imaging to be reconstructed is acquired by the multi-excitation echo-planar sequence ,in represents the k-space data collected by the c-th coil during the j-th excitation, , ; represents a complex set; represents multiplication; M, N, J, and C represent the frequency encoding number, phase encoding number, total number of excitations, and total number of coil channels, respectively.

[0016] In step 2), the bidirectional one-dimensional low-rank reconstruction model is:

[0017]

[0018] in, is the k-space data of diffusion magnetic resonance imaging to be reconstructed; is the sampling template of the multi-excitation sequence; and They are the two-dimensional Fourier transform and its inverse transform; It is the channel sensitivity operator, which saves the sensitivity information of multiple channels; is the multi-shot data to be reconstructed, and Extract the frequency-coded lines and An operator that converts phase-encoded rows into a structured low-rank matrix; and are one-dimensional inverse Fourier transform operations on phase-encoding and frequency-encoding lines, respectively, denotes the Frobenius norm of a matrix; denotes the nuclear norm of a matrix; , and are weight coefficients; M denotes the number of frequency encodings, and N denotes the number of phase encodings.

[0019] In step 3), the data is generated using a physical diffusion model, specifically: using a phantom containing multiple parts such as brain gray matter, brain white matter, spine, spinal cord, adrenal gland, liver, gallbladder, stomach, pancreas, spleen, colon, kidney, ureter, artery, vein, muscle, bone, joint, fat, and skin, etc. to obtain an amplitude image:

[0020]

[0021] wherein m o represents the amplitude of the oth part, and there are O parts in total; represents a mask of the oth part; m is an amplitude image composed of each part in a layer;

[0022] Generating complex phases of all excitations of multiple organs based on a high-order phase model wherein the phase of the jth excitation is:

[0023]

[0024] wherein m, n are image coordinates; L o is the order of the generated phase of the oth part (o = 1, 2, …, O); represents a mask of the oth part; is a parameter of the generated phase of the oth part;

[0025] Further generating full-sampling multi-excitation k-space data using the generated phase, including generating noise-free multi-excitation k-space data X GT and generating noisy multi-excitation k-space data X Inp :

[0026]

[0027]

[0028] wherein, is the generated multi-excitation motion phase; is an amplitude image; and are two-dimensional Fourier transform and its inverse transform, respectively; N is noise.

[0029] In step 4), the base model neural network is trained using physically generated data; the generated noisy multi-shot k-space data X Inp Convert to a structured low-rank matrix and perform singular value decomposition:

[0030]

[0031]

[0032] in, and These are the operations of performing one-dimensional inverse Fourier transform on the phase-encoded lines and the frequency-encoded lines; and Extract the frequency-coded rows and Phase encoding rows and converting them into structured low-rank matrices; SVD is the singular value decomposition; 、 、 are the left singular matrix, singular values, and right singular matrix of the structured matrix constructed from frequency-encoded data; 、 、 are the left singular matrix, singular values, and right singular matrix of the structured matrix constructed by phase encoding data; X Inp is the generated noisy multi-shot k-space data;

[0033] The base model neural network learns the singular value truncation parameters of the structured low-rank matrix based on the generated data, that is, the number of singular values ​​optimally retained in the frequency encoding direction and phase encoding structured low-rank matrix and , so that the noisy signal has the highest peak signal-to-noise ratio after being restored by singular value truncation:

[0034]

[0035]

[0036] in, and These are the operations of performing one-dimensional inverse Fourier transform on the phase-encoded lines and the frequency-encoded lines; and Extract the frequency-coded rows and An operator that converts phase-encoded rows into a structured low-rank matrix; 、 、 are the left singular matrix, singular values, and right singular matrix of the structured matrix constructed from frequency-encoded data; 、 、 are left, singular values, and right singular matrices of structured matrix constructed by phase encoding data, respectively; M and N are the number of frequency encoding and phase encoding, respectively; X GT is noiseless multi-shot k-space data; denotes matrix multiplication; denotes the Frobenius norm of a matrix;

[0037] Physical-based neural network model is a data-driven supervised artificial intelligence neural network based on gradient backpropagation, whose loss function is:

[0038]

[0039] where the input of the neural network is the pair of one-dimensional input signals and ; the label of the network is the optimal number of reserved singular values of the structured low-rank matrix in the frequency encoding direction and the phase encoding direction and ; and denote the operation operators for selecting the mth frequency encoding row and the nth phase encoding row, respectively; and are the operations of one-dimensional inverse Fourier transform of the phase encoding row and the frequency encoding row, respectively; T is the total number of training samples; is the optimal weight parameter; M represents the number of frequency encoding, and N represents the number of phase encoding; denotes the Frobenius norm of a matrix; X Inp is the generated noisy multi-shot k-space data.

[0040] In step 5), the bidirectional one-dimensional low-rank reconstruction model uses singular value truncation to solve the kernel norm; the singular value truncation is used to constrain the low-rank matrix and are and , where denotes the singular value truncation of the low-rank matrix , and the signal recovered by retaining the first r largest singular values and the corresponding subspace; where, and are the operations of one-dimensional inverse Fourier transform of the phase encoding row and the frequency encoding row, respectively; and are the operation operators for extracting the th frequency encoding row and the th phase encoding row and converting them into a structured low-rank matrix, respectively; is the multi-shot data to be reconstructed.

[0041] In step 6), the optimal parameters are truncated using the low-rank matrix singular values ​​suggested by the base model neural network and Perform singular value truncation to achieve robust high-definition reconstruction.

[0042] Compared with the prior art, the present invention has the following outstanding technical effects and advantages:

[0043] 1. The present invention designs a bidirectional one-dimensional low-rank reconstruction model and combines it with local smooth phase constraints to perform targeted correction of phase errors and artifacts caused by inter-excitation motion in multi-excitation sequences. In particular, the reconstruction effect of complex phase areas such as the abdomen and pelvis is significantly better than that of existing low-rank algorithms (such as MUSE and MUSSELS), and can eliminate residual artifacts.

[0044] 2. The present invention generates phases of different orders for different organs and phase data for multiple organs. It constructs a physical-based model neural network based on the physical diffusion model and learns adaptive optimization of reconstruction parameters by training the neural network to achieve stable reconstruction in all levels and directions under a single parameter setting. This eliminates the need to manually adjust parameters for different parts or levels, and addresses the problem of insufficient robustness of existing methods (such as DONATE).

[0045] 3. The present invention uses a physical-based model neural network to accurately learn the signal characteristics of complex phase areas, maintaining signal integrity in multiple organ locations with large phase differences (such as the junction of the liver and intestine in the abdomen), overcoming the defect of existing methods such as PIDD that are prone to signal loss in complex phase scenarios.

[0046] 4. The present invention adopts a bidirectional one-dimensional low-rank model to strengthen local low-rank constraints, and combines the optimal singular value truncation parameters suggested by the neural network to achieve generalized reconstruction of multiple parts such as the brain, neck, and joints while improving image resolution, taking into account high-definition quality and applicability to multiple parts. Through the bidirectional one-dimensional low-rank optimization reconstruction method and the neural network adaptive optimization parameters, robust high-definition magnetic resonance diffusion imaging of multiple challenging parts is achieved without adjusting parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Fig. 1 The image shows the multi-excitation merged data of the kidney (left) and the reconstruction result (right) achieved with 2-times equally spaced undersampling based on a physical-based model multi-site high-definition diffusion magnetic resonance imaging method.

[0048] Fig. 2 This is the liver multi-excitation merged data (left) and reconstruction results (right) of 2-excitation 2-fold equally spaced undersampling based on a physical-based model multi-site high-definition diffusion magnetic resonance imaging method.

[0049] Fig. 3 Figure 2 shows the prostate multi-excitation merged data (left) and reconstruction results (right) achieved using a physical-based model multi-site high-definition diffusion MRI method.

[0050] Fig. 4 This is the four-shot sacroiliac joint multi-excitation merged data (left) and reconstruction results (right) achieved based on a physical-based model multi-site high-definition diffusion MRI method.

[0051] Fig. 5 This is the four-shot neck multi-excitation merged data (left) and the reconstruction result (right) achieved based on a physical-based model multi-site high-definition diffusion MRI method.

[0052] Fig. 6 This is the four-shot brain multi-excitation merged data (left) and reconstruction results (right) achieved based on a physical-based model multi-site high-definition diffusion magnetic resonance imaging method.

[0053] Fig. 7 This is the double-excitation knee joint multi-excitation merged data (left) and reconstruction results (right) achieved based on a physical-based model multi-site high-definition diffusion magnetic resonance imaging method. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the following embodiments will be further described in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0055] The embodiment of the present invention uses a physical-based model multi-site high-definition diffusion magnetic resonance imaging method proposed by the present invention to reconstruct the data of multiple body parts with 2 excitations (abdomen, pelvis, knee joint) and 4 excitations (brain, spinal cord, sacroiliac joint). Figs. 1-7 , the method of the present invention is described in detail.

[0056] This embodiment includes the following steps:

[0057] 1) Acquire k-space data for diffusion MRI to be reconstructed using a multi-shot echo-planar sequence ,in represents the k-space data collected by the c-th coil during the j-th excitation, , ; represents a complex set; represents multiplication; M, N, J, and C represent the frequency encoding number, phase encoding number, total number of excitations, and total number of coil channels, respectively.

[0058] 2) Bidirectional one-dimensional low-rank reconstruction model is:

[0059]

[0060] wherein is the sampling template of multi-shot sequence; and are two-dimensional Fourier transform and its inverse transform, respectively; is the channel sensitivity operator, which saves the sensitivity information of multi-channel; is the multi-shot data to be reconstructed; and are the operation operators for extracting the th frequency encoding line and the th phase encoding line and converting them into structured low-rank matrix, respectively; and are the operations of one-dimensional Fourier inverse transform of phase encoding line and frequency encoding line, respectively; denotes the Frobenius norm of matrix; denotes the nuclear norm of matrix; the weight coefficient , , .

[0061] 3) Use the generated data based on the diffusion physical model.

[0062] 3.1) Obtain the amplitude image: obtain the amplitude image using the phantom containing multiple parts such as brain gray matter, brain white matter, spine, spinal cord, adrenal gland, liver, gallbladder, stomach, pancreas, spleen, colon, kidney, ureter, artery, vein, muscle, bone, joint, fat and skin:

[0063]

[0064] wherein m o represents the amplitude of the oth part, and the number of parts O=20; represents the mask of the oth part; m is the amplitude image composed of each part in a layer.

[0065] 3.2) Generate complex phase: generate the complex phase of all shots of multiple organs based on the high-order phase model , wherein the phase of the jth shot is:

[0066]

[0067] wherein m, n are image coordinates; the order of the generated phase of the oth part ; represents the mask of the oth part; the parameter of the generated phase of the oth part .

[0068] 3.3) Generate k-space data: Use the generated phase to further generate full-sample multi-shot k-space data, including generating noise-free multi-shot k-space data X GT and generate the noisy multi-shot k-space data X Inp :

[0069]

[0070]

[0071] in is the generated multi-excitation motion phase; is the amplitude image; and They are two-dimensional Fourier transform and its inverse transform respectively; N is noise, and the signal-to-noise ratio is set to 0~20dB.

[0072] 4) The base model neural network is trained using physically generated data.

[0073] 4.1) Data preprocessing: The generated noisy multi-shot k-space data X Inp Convert to a structured low-rank matrix and perform singular value decomposition:

[0074]

[0075]

[0076] in, and They are respectively the one-dimensional inverse Fourier transform operations of the phase coding line and the frequency coding line; SVD is the singular value decomposition; and Extract the frequency-coded rows and An operator that converts phase-encoded rows into a structured low-rank matrix; 、 、 are the left singular matrix, singular values, and right singular matrix of the structured matrix constructed from frequency-encoded data; 、 、 They are the left singular matrix, singular values, and right singular matrix of the structured matrix constructed by phase encoding data.

[0077] 4.2) Learning singular value truncation parameters: The base model neural network learns the singular value truncation parameters of the structured low-rank matrix based on the generated data, and the number of singular values ​​optimally retained by the frequency encoding direction and phase encoding structured low-rank matrix and The noisy signal is subjected to singular value truncation recovery, and has the highest peak signal-to-noise ratio:

[0078]

[0079]

[0080] wherein and are one-dimensional Fourier inverse transform operations on phase-encoding lines and frequency-encoding lines, respectively; and are operators for extracting the th frequency-encoding line and the th phase-encoding line and converting them into structured low-rank matrices; , , are left singular matrices, singular values, and right singular matrices of structured matrices constructed from frequency-encoding data, respectively; , , are left singular matrices, singular values, and right singular matrices of structured matrices constructed from phase-encoding data, respectively; M and N are the number of frequency-encoding and phase-encoding, respectively; GT X is noise-free multi-shot k-space data; represents matrix multiplication; represents the Frobenius norm of a matrix.

[0081] Physical basis neural network model The network structure of the physical basis neural network model is ResNet18 (Deep residual learning for image recognition, K. He et al., Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770-778, 2016),

[0082] The loss function of the physical basis neural network model is:

[0083]

[0084] wherein, the input of the neural network is a pair of one-dimensional input signals and ; the label of the network is the optimal number of reserved singular values of the frequency-encoding direction and the phase-encoding structured low-rank matrix and ; and They represent the operators for selecting the mth frequency coding row and the nth phase coding row respectively; and are the one-dimensional inverse Fourier transform operations on the phase encoding line and the frequency encoding line respectively; T is the total number of training samples; is the optimal weight parameter; M represents the frequency encoding number, and N represents the phase encoding number; represents the Frobenius norm of the matrix; X Inp is the generated noisy multi-shot k-space data.

[0085] 5) The bidirectional one-dimensional low-rank reconstruction model uses singular value truncation to solve the nuclear norm. Use singular value truncation to constrain the low-rank matrix and for and ,in Represents a low-rank matrix Perform singular value truncation and retain the first r largest singular values ​​and the corresponding subspace to recover the signal; and These are the operations of performing one-dimensional inverse Fourier transform on the phase-encoded lines and the frequency-encoded lines; and Extract the frequency-coded lines and An operator that converts phase-encoded rows into a structured low-rank matrix; is the multi-shot data to be reconstructed.

[0086] 6) Use the low-rank matrix singular values ​​suggested by the base model neural network to truncate the optimal parameters and Perform singular value truncation to achieve robust high-definition reconstruction.

[0087] To verify the technical effect of the present invention, the multi-excitation diffusion data of the kidney, liver, prostate, sacroiliac joint, neck, brain, and knee joint were reconstructed. The results are as follows: Figs. 1-7 , with the kidneys, liver, prostate, and knee joints receiving two excitations, and the sacroiliac joints, neck, and brain receiving four excitations. The left side of each figure shows the artifact-bearing image after merging the original data with multiple excitations, while the right side shows the image reconstructed using this method. Experiments demonstrate that the method of the present invention effectively suppresses artifacts in multiple locations.

[0088] The present invention generates phases of different orders for different organs, generates phase data for different organs, constructs a physical-based model neural network, and trains the neural network to adaptively obtain optimized reconstruction parameters. This method uses a bidirectional one-dimensional low-rank optimization reconstruction method and a neural network to adaptively optimize reconstruction parameters, achieving robust high-definition magnetic resonance diffusion imaging of multiple challenging areas without adjusting parameters. Compared to existing low-rank reconstruction algorithms, the present invention strengthens local low-rank constraints through a bidirectional one-dimensional low-rank model and a neural network. Phase data for multiple areas is generated, and a neural network is trained to adaptively obtain optimized reconstruction parameters. High-definition, artifact-free images can be robustly reconstructed for areas with complex phases.

[0089] The above embodiments are only preferred embodiments of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent of the present invention.

Claims

1. A physical-based model multi-site high-definition diffusion magnetic resonance imaging method, characterized in that The following steps are involved: 1) Acquire k-space data of diffusion MRI to be reconstructed using a multi-shot echo-planar sequence for high-definition diffusion imaging reconstruction of multiple sites; 2) Designing a bidirectional one-dimensional low-rank reconstruction model based on local smooth phase; the model includes a two-dimensional Fourier transform operator, a channel sensitivity operator, a structured low-rank matrix extraction operator, and a nuclear norm constraint term to characterize the low-rank characteristics of multi-shot data; The bidirectional one-dimensional low-rank reconstruction model is: Where Y is the k-space data of diffusion MRI to be reconstructed; is the sampling template of the multi-excitation sequence; and are two-dimensional Fourier transform and its inverse transform respectively; C is the channel sensitivity operator, which saves the sensitivity information of multiple channels; X is the multi-excitation data to be reconstructed, and They are operators for extracting the mth frequency-coded row and the nth phase-coded row and converting them into structured low-rank matrices; and are the one-dimensional inverse Fourier transform operations of the phase-encoded line and the frequency-encoded line, respectively. F represents the Frobenius norm of the matrix; ||·|| * represents the nuclear norm of the matrix; λ, α and β are weight coefficients; M represents the frequency encoding number, and N represents the phase encoding number; 3) Data generation using physical diffusion models, including: acquiring amplitude images using a multi-organ phantom, generating complex phase data of corresponding orders for different organs, and combining amplitude and phase to generate noisy and noise-free fully sampled multi-shot k-space data; 4) Using physically generated data to train the base model neural network; by converting the noisy k-space data into a structured low-rank matrix and performing singular value decomposition, the neural network learns the optimal singular value cutoff parameters; 5) Solve the bidirectional one-dimensional low-rank reconstruction model and achieve signal recovery by preserving the key singular values ​​and corresponding subspaces of the structured low-rank matrix; 6) Combined with the reconstruction parameters suggested by the neural network of the base model, the multi-excitation k-space data are generalized and reconstructed to obtain high-definition diffusion magnetic resonance images of various parts without motion artifacts.

2. A physical-based model multi-site high-definition diffusion magnetic resonance imaging method as claimed in claim 1, characterized in that In step 1), the k-space data of the diffusion magnetic resonance imaging to be reconstructed acquired by the multi-excitation planar echo sequence is obtained, specifically: the k-space data of the diffusion magnetic resonance imaging to be reconstructed acquired by the multi-excitation planar echo sequence Y=[Y 11 ,…,Y 1C ,Y 21 ,…,Y 2C ,…,Y jc ,…,Y J1 ,…,Y JC ],in represents the k-space data collected by the c-th coil during the j-th excitation, j = 1, 2, ..., J, c = 1, 2, ..., C; represents a set of complex numbers; × represents multiplication; M, N, J, and C represent the number of frequency encodings, the number of phase encodings, the total number of excitations, and the total number of coil channels, respectively.

3. A physical-based model multi-site high-definition diffusion magnetic resonance imaging method as claimed in claim 1, characterized in that In step 3), the data is generated using a physical diffusion model, specifically by acquiring amplitude images using a phantom comprising gray matter, white matter, spine, spinal cord, adrenal glands, liver, gallbladder, stomach, pancreas, spleen, colon, kidneys, ureters, arteries, veins, muscles, bones, joints, fat, and skin. Among them, m o They represent the amplitude of the oth part, and there are O parts in total; Represents the mask of the o-th part; m is the amplitude map of a layer composed of various parts; Generate the complex phase of all excitations of multiple organs based on the high-order phase model P = [P1, P2, ... P j …,P J ], where the phase of the j-th excitation is: Where m, n are image coordinates; L o is the order of the phase generated by the oth part; The mask representing the oth part; is the parameter of the generation phase of the oth part; The generated phase is used to further generate full-sampled multi-shot k-space data, including generating noise-free multi-shot k-space data X GT and generate the noisy multi-shot k-space data X Inp : Where P is the generated multi-excitation motion phase; m is the amplitude image; and are two-dimensional Fourier transform and its inverse transform respectively; N is noise.

4. A physical-based model multi-site high-definition diffusion magnetic resonance imaging method as claimed in claim 1, characterized in that In step 4), the base model neural network is trained using the physical generated data. Specifically, the generated noisy multi-shot k-space data X Inp Convert to a structured low-rank matrix and perform singular value decomposition: in, and These are the operations of performing one-dimensional inverse Fourier transform on the phase-encoded lines and the frequency-encoded lines; and are operators that extract the mth frequency-coded row and the nth phase-coded row and convert them into structured low-rank matrices; SVD is the singular value decomposition; are the left singular matrix, singular values, and right singular matrix of the structured matrix constructed from frequency-encoded data; are the left singular matrix, singular values, and right singular matrix of the structured matrix constructed by phase encoding data; X Inp is the generated noisy multi-shot k-space data; The base model neural network learns the singular value truncation parameters of the structured low-rank matrix based on the generated data, that is, the number of singular values ​​optimally retained in the frequency encoding direction and phase encoding structured low-rank matrix and After the noisy signal is restored through singular value truncation, it has the highest peak signal-to-noise ratio: in, and These are the operations of performing one-dimensional inverse Fourier transform on the phase-encoded lines and the frequency-encoded lines; and They are operators for extracting the mth frequency-coded row and the nth phase-coded row and converting them into structured low-rank matrices; are the left singular matrix, singular values, and right singular matrix of the structured matrix constructed from frequency-encoded data; are the left singular matrix, singular value, and right singular matrix of the structured matrix constructed by phase encoding data; M and N are the frequency encoding number and phase encoding number respectively; X GT is noise-free multi-shot k-space data; * denotes matrix multiplication; ||·|| F represents the Frobenius norm of the matrix; The physical-based neural network model PromptN(·) is a data-driven supervised artificial intelligence neural network based on gradient back propagation, and its loss function is: Among them, the input of the neural network is a paired one-dimensional input signal and The network's labels are frequency-encoded and phase-encoded. The optimal number of singular values ​​retained in the structured low-rank matrix and and They represent the operators for selecting the mth frequency coding row and the nth phase coding row respectively; and are the one-dimensional inverse Fourier transform operations on the phase encoding line and the frequency encoding line respectively; T is the total number of training samples; is the optimal weight parameter; M represents the number of frequency codes, N represents the number of phase codes; ||·|| F represents the Frobenius norm of the matrix; X Inp is the generated noisy multi-shot k-space data.

5. A physical-based model multi-site high-definition diffusion magnetic resonance imaging method as claimed in claim 1, characterized in that In step 5), the bidirectional one-dimensional low-rank reconstruction model uses singular value truncation to solve the nuclear norm; the singular value truncation is used to constrain the low-rank matrix and for and in Indicates that the low-rank matrix A is truncated for singular values, and the first r largest singular values ​​and the corresponding subspace are retained to recover the signal; where, and These are the operations of performing one-dimensional inverse Fourier transform on the phase-encoded lines and the frequency-encoded lines; and are operators that extract the mth frequency coding row and the nth phase coding row and convert them into structured low-rank matrices; X is the multi-shot data to be reconstructed.

6. A physical-based model multi-site high-definition diffusion magnetic resonance imaging method as claimed in claim 1, characterized in that In step 6), the reconstruction parameters suggested by the base model neural network adopt the low rank matrix singular value truncation optimal parameters suggested by the base model neural network and

Images