Abdomen multi-excitation high-definition diffusion magnetic resonance intelligent reconstruction method

Abstract

The invention discloses an abdomen multi-excitation high-definition diffusion magnetic resonance intelligent reconstruction method, and relates to intelligent reconstruction of an abdomen diffusion magnetic resonance image acquired based on a multi-excitation staggered plane echo sequence. Comprising the following steps: 1) acquiring frequency domain data of an abdominal diffusion magnetic resonance image acquired by using a multi-excitation staggered plane echo sequence; 2) designing an intelligent high-definition abdominal diffusion magnetic resonance reconstruction model; 3) constructing a large amount of paired training data, and training the phase segmentation network in the reconstruction model to obtain trained network parameters; 4, the intelligent high-definition abdominal diffusion magnetic resonance reconstruction model is solved through a convex set projection algorithm, and a final reconstruction result is obtained.Compared with a traditional reconstruction algorithm, the method has the better reconstruction signal-to-noise ratio and artifact suppression capacity, and high-definition diffusion image reconstruction can be achieved.

Description

Technical Field

[0001] The present invention belongs to an intelligent reconstruction method of abdominal multi-excitation high-definition diffusion magnetic resonance, and in particular relates to an intelligent reconstruction method of abdominal multi-excitation high-definition diffusion magnetic resonance by acquiring a multi-excitation interleaved planar echo sequence without a navigation echo. Background Art

[0002] Diffusion MRI is the only imaging method that can detect the movement of water molecules in vivo without invasiveness or radiation, and is widely used in clinical and neuroscience research (M. Bernstein, F. King, X. Zhou. Handbook of MRIPulse Sequences. Elsevier , 2004.). However, due to the limitation of the echo time and echo chain length of the rapid acquisition sequence, clinical diffusion imaging has a low signal-to-noise ratio and resolution (3-4 mm within the layer). 2 ).

[0003] Multiple-shot interleaved echo-planar sequences can be used to obtain diffusion data with submillimeter ultra-high resolution. However, due to physiological movements of the human body (such as heart and breathing), there may be phase changes caused by movement between each shot (R. Bammer, “Basic principles of diffusion-weighted imaging”, European Journal of Radiology , vol. 45, pp.169-184, 2003.)), so a reconstruction algorithm is needed to correct the motion phase (A. Anderson, and J. Gore, “Analysis and correction of motion artifactsin diffusion weighted imaging,” Magnetic Resonance in Medicine , vol. 32, pp.379-87, 1994.).

[0004] Several reconstruction algorithms for motion phase correction of multi-excitation diffusion MRI have been proposed. Among them, MUSE (N(NK Chen, A. Guidon, HC Chang, and AW Song, "A robust multi-shot scanstrategy for high-resolution diffusion weighted MRI enabled by multiplexedsensitivity-encoding (MUSE)," Neuroimage, vol. 72, pp. 41-47, 2013.) is a representative non-navigation echo reconstruction scheme. Recently, structured low-rank matrix reconstruction methods based on smooth phase characteristics, such as MUSSELS (M. Mani, M. Jacob, D. Kelley, et al. , “Multi‐shot sensitivity‐encodediffusion data recovery using structured low‐rank matrix completion (MUSSELS),” Magnetic Resonance in Medicine , vol. 78, pp. 494-507, 2017.), PLRHM (Y. Huang, X. Zhang, H. Guo, et al. , “Phase-constrained reconstruction of high-resolution multi-shot diffusion weighted image,” Journal of Magnetic Resonance , vol. 312, pp. 106690, 2020.), PAIR (C. Qian, Z Wang, B Shi, et al .,“A paired phase and magnitude reconstruction for advanced diffusion-weightedimaging,” IEEE Transactions on Biomedical Engineering , DOI: 10.1109 / TBME.2023.3288031, 2023) and physically generated data schemes (Q. Yang, Z. Wang, K. Guo, et al ., “Physics-driven synthetic data learning for biomedical magneticresonance: The imaging physics-based data synthesis paradigm for artificial intelligence,” IEEE Signal Processing Magazine , vol. 40, no. 2, pp. 129-140,2023.) Intelligent reconstruction algorithm PIDD (C. Qian, Z. Wang, X. Zhang, et al., “Physics-informeddeep diffusion MRI reconstruction: Break the bottleneck of training data inartificial intelligence,” 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI) , pp. 1-5, 2023, DOI: 10.1109 / ISBI53787.2023.10230538.) have been proposed. These algorithms can achieve robust 4-shot high-definition diffusion image reconstruction of the brain.

[0005] However, it is still difficult for the above algorithms to reliably achieve the more challenging abdominal multi-excitation high-definition diffusion reconstruction. This is mainly because the displacement of the liver and surrounding organs is mainly dominated by two physiological movements: respiratory movement and cardiac movement, which causes subvoxel phase shifts in the organs, resulting in attenuation of non-diffusion signals, and then artifacts related to frequency-space trajectory errors. Summary of the invention

[0006] The purpose of the present invention is to provide an abdominal multi-excitation high-definition diffusion magnetic resonance intelligent reconstruction method with significant signal-to-noise ratio advantages and better artifact suppression capability in abdominal high-definition diffusion reconstruction, in order to address the problems existing in existing diffusion reconstruction methods, such as the difficulty in achieving abdominal multi-excitation high-definition diffusion reconstruction.

[0007] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions.

[0008] An abdominal multi-excitation high-definition diffusion magnetic resonance intelligent reconstruction method comprises the following steps:

[0009] 1) Acquire frequency domain data of abdominal diffusion magnetic resonance images acquired using a multi-shot interleaved echo-planar sequence;

[0010] 2) Design an intelligent high-definition abdominal diffusion MRI reconstruction model;

[0011] 3) Construct a large amount of paired training data and use it to train the phase segmentation network in the reconstruction model to obtain the trained network parameters;

[0012] 4) The convex set projection algorithm is used to solve the intelligent high-definition abdominal diffusion MRI reconstruction model and obtain the final reconstruction result;

[0013] In step 1), the frequency domain data of the abdominal diffusion magnetic resonance image acquired by using a multi-excitation interleaved echo-planar sequence is obtained as follows:

[0014]

[0015] in, represents the frequency domain data of the abdominal diffusion MRI image acquired by J times of H channels, also called k-space) N and M represent the frequency encoding number and phase encoding number of a single diffusion MRI image, respectively; is the diffusion magnetic resonance image of the target solution; is the channel sensitivity map of the signal receiving coil used to collect data; represents the noise during acquisition; is the sampling operator determined by the frequency domain trajectory, indicating zero filling at unsampled points; is the Fourier transform operator; and are the field of real numbers and the field of complex numbers respectively.

[0016] In step 2), the intelligent high-definition abdominal diffusion magnetic resonance reconstruction model is:

[0017]

[0018] in, represents the frequency domain data of the abdominal diffusion MRI image acquired by J excitations of H channels; N and M represent the frequency encoding number and phase encoding number of a single diffusion MRI image, respectively. is the diffusion magnetic resonance image of the target solution, is the channel sensitivity map of the signal receiving coil used to collect data; is the sampling operator determined by the frequency domain trajectory, indicating zero filling at unsampled points; is the Fourier transform operator; and are the domain of real numbers and the domain of complex numbers respectively; and are the square of the Frobenius norm and the nuclear norm, respectively; and are two regularization parameters; represents the operator for selecting the Pth region of the image domain; represents the operator for selecting smooth phase regions in the image domain, represents the operator for selecting the non-smooth phase region in the image domain; is an operator for constructing structured low-rank matrices; is an operator that constructs a local low-rank matrix, where c is the coordinate of the center of the selected image block.

[0019] In step 3), a large amount of paired training data is constructed and used to train the phase segmentation network in the reconstruction model to obtain the trained network parameters:

[0020] a) Design a phase segmentation network with multi-layer convolutional neural network

[0021] It consists of L convolutional layers, including an input layer, an output layer, and L-2 hidden layers; each hidden layer contains Z convolution kernels, and the convolution kernel size is ;

[0022] b) Construct a training data set for training the phase segmentation network

[0023] First, the frequency domain data of high-definition diffusion MRI of the abdomen with J excitations in a total of S layers are collected; a 64×64 area is selected from the center of the frequency domain of the sth layer, and a traditional optimization algorithm is used to reconstruct a multi-excitation diffusion image with low resolution but high signal-to-noise ratio. ;

[0024] The phase of the diffusion image of the jth excitation (j=1,2,…,J traverses all excitations) Perform manual segmentation and select smooth phase areas ;right Take the complement in the image domain and get the non-smooth phase region ; Change the phase and amplitude As network input, and As network training labels, a total of S×J pairs of training samples are obtained;

[0025] c) Use the constructed paired dataset for network training

[0026] Using the above dataset to train the denoiser, the loss function is defined as:

[0027]

[0028] in, is a phase division network, Represents the set of internal parameters of the entire network. The trained network parameters can be expressed as ;Σ represents the sum operation;

[0029] In step 4), the convex set projection algorithm is solved, wherein the kth iteration is described as follows:

[0030] a) Data verification items

[0031]

[0032]

[0033] in, is the image after data verification in the kth iteration; is the channel sensitivity map of the signal receiving coil used to collect data; is the image after channel merging in the kth iteration; represents the frequency domain data of the abdominal diffusion magnetic resonance image acquired by J times of excitation of H channels; are the diffusion magnetic resonance images in the kth iteration, and their initial values ​​are It can be reconstructed by selecting a 64×64 area from the center of the frequency domain using a traditional optimization algorithm; N, M, J, and H represent the frequency encoding number, phase encoding number, excitation number, and receiving coil channel number of a single diffusion magnetic resonance image, respectively; and They are the frequency domain sampling operator and its adjoint operator; and They are the Fourier transform operator and the inverse Fourier transform operator; and are the domain of real numbers and the domain of complex numbers respectively; is the regularization parameter;

[0034] b) Phase division

[0035]

[0036] in, is the image of the jth excitation after data verification and channel merging in the kth iteration; and is the operator for constructing structured low-rank matrices and its adjoint operator; is the trained phase segmentation network, and its weight parameter is ; and They are the phase smooth area and non-smooth area obtained by segmenting the image of the jth excitation. The phase segmentation results of multiple excitations are obtained by segmenting the Jth excitation number in sequence. and ;

[0037] c) Structured low-rank constraints

[0038]

[0039]

[0040] in, is the diffusion image after data verification and channel merging in the kth iteration; is the phase-segmented region, where is the smooth phase region, It is the non-smooth phase region; is the singular value hard threshold truncation operator, which means to perform singular value decomposition on the matrix Z and retain the first r psingular values, and They are the Fourier transform operator and the inverse Fourier transform operator; and represents the operator for constructing structured low-rank matrices and its inverse operator;

[0041] d) Local low-rank constraint

[0042]

[0043]

[0044] in, is the diffusion image after data verification and channel merging in the kth iteration; is the phase-segmented region, where is the smooth phase region, It is the non-smooth phase region; is the singular value soft threshold operator, which means to decompose the matrix Z into a singular value and subtract the singular value from , and retain singular values ​​greater than 0, and They are the Fourier transform operator and the inverse Fourier transform operator; and represents the operator for constructing structured low-rank matrices and its inverse operator; is the regularization parameter;

[0045] e) Update the image

[0046]

[0047] is the final reconstruction result of the kth iteration;

[0048] After K iterations, a reconstructed multi-excitation high-definition abdominal diffusion magnetic resonance image is finally obtained.

[0049] The present invention solves the problem that the existing technology is difficult to achieve high-definition diffusion reconstruction of the abdomen with multiple excitations, and has a significant signal-to-noise ratio advantage and better artifact suppression capability in high-definition diffusion reconstruction of the abdomen. In particular, the present invention can achieve better motion artifact suppression for areas such as the liver and spleen where the motion phase is large and the artifact suppression of traditional algorithms is poor, and achieve high-definition diffusion image reconstruction of the liver. The present invention has a high application prospect in high-definition diffusion imaging of moving parts, especially the abdominal liver and other areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the iterative reconstruction process of the model according to an embodiment of the present invention.

[0051] Figure 2 This is the segmentation result of the phase segmentation network provided in the embodiment of the present invention.

[0052] Figure 3 The figure is a comparison of the reconstruction results of the method proposed in the embodiment of the present invention in the liver 4-time excitation high-definition diffusion imaging and the comparison method. Among them, (a) is the comparison method 1, (b) is the comparison method 2, (c) is the comparison method 3, (d) is the comparison method 4, and (e) is the method of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in the following embodiments in conjunction with the accompanying drawings.

[0054] Example 1: Reconstruction of high-definition diffusion imaging data of abdominal liver with four shots

[0055] The frequency domain data of abdominal diffusion magnetic resonance images collected using a multi-excitation interleaved echo-planar sequence is:

[0056]

[0057] in, It represents the frequency domain data of the abdominal diffusion magnetic resonance image acquired with 32 channels and 4 excitations; is the diffusion magnetic resonance image of the target solution; is the channel sensitivity map of the signal receiving coil used to collect data; represents the noise during acquisition; is the sampling operator determined by the frequency domain trajectory, indicating zero filling at unsampled points; is the Fourier transform operator; and are the field of real numbers and the field of complex numbers respectively.

[0058] The intelligent high-definition abdominal diffusion MRI reconstruction model is:

[0059]

[0060] in, It represents the frequency domain data of the abdominal diffusion magnetic resonance image acquired with 32 channels and 4 excitations; is the diffusion image of the target solution, is the channel sensitivity map of the signal receiving coil used to collect data; is the sampling operator determined by the frequency domain trajectory, indicating zero filling at unsampled points; is the Fourier transform operator; and are the field of real numbers and the field of complex numbers respectively. and are the square of the Frobenius norm and the nuclear norm, respectively; and are two regularization parameters, taking 1.0 and 1.0 respectively; Represents different regions in the image domain. Among them, represents the operator for selecting smooth phase regions in the image domain; represents the operator for selecting the non-smooth phase region in the image domain; is an operator for constructing structured low-rank matrices; is an operator for constructing a local low-rank matrix, where c is the coordinate of the center of the selected image block;

[0061] Construct a large amount of paired training data and use it to train the phase segmentation network in the reconstruction model to obtain the trained network parameters:

[0062] a) Design a phase segmentation network with multi-layer convolutional neural network

[0063] It consists of 16 convolutional layers, including an input layer, an output layer, and 14 hidden layers; each hidden layer contains 128 convolution kernels, and the convolution kernel size is 5×5.

[0064] b) Construct a training data set for training the phase segmentation network

[0065] First, the frequency domain data of the abdominal J-shot high-definition diffusion MRI with a total of 150 layers were collected. A 64×64 area was selected from the frequency domain center of each layer and reconstructed using a traditional optimization algorithm to obtain a multi-excitation diffusion image with low resolution but high signal-to-noise ratio. .

[0066] The phase of the diffusion image of each excitation is Perform manual segmentation and select smooth phase areas .right Take the complement in the image domain and get the non-smooth phase region .

[0067] Phase and amplitude As network input, and As network training labels, a total of 6000 pairs of training samples were obtained.

[0068] c) Use the constructed paired dataset for network training

[0069] Using the above dataset to train the denoiser, the loss function is defined as:

[0070]

[0071] in, is a phase division network, Represents the set of internal parameters of the entire network. The trained network parameters can be expressed as ; represents the bi-norm term; Σ represents the summation operation.

[0072] In step 4), the convex set projection algorithm is solved, wherein the kth iteration is described as follows:

[0073] a) Data verification items

[0074]

[0075]

[0076] in, is the image after data verification in the kth iteration; is the channel sensitivity map of the signal receiving coil used to collect data; is the image after channel merging in the kth iteration; It represents the frequency domain data of the abdominal diffusion magnetic resonance image acquired with 32 channels and 4 excitations; are the diffusion magnetic resonance images in the kth iteration, and their initial values ​​are It can be reconstructed by selecting a 16×16 region from the center of the frequency domain and using a traditional optimization algorithm; and They are the frequency domain sampling operator and its adjoint operator; and They are the Fourier transform operator and the inverse Fourier transform operator; and are the field of real numbers and the field of complex numbers respectively. is the regularization parameter and takes the value of 1.0.

[0077] b) Phase division

[0078]

[0079] in, is the image of the jth excitation after data verification and channel merging in the kth iteration; and is the operator for constructing structured low-rank matrices and its adjoint operator; is the trained phase segmentation network, and its weight parameter is ; and are the phase smooth region and non-smooth region segmented from the image of the jth excitation, respectively. Figure 2As shown. Sequentially divide the number of excitations 4 times to obtain the phase segmentation result of multiple excitations and .

[0080] c) Structured low-rank constraints

[0081]

[0082]

[0083] in, is the diffusion image after data verification and channel merging in the kth iteration; is the phase-segmented region, where is the smooth phase region, It is the non-smooth phase region; is the singular value hard threshold truncation operator, which means to perform singular value decomposition on the matrix Z and retain the first r p singular values, r1 = 10, r2 = 100; and They are the Fourier transform operator and the inverse Fourier transform operator; and Represents the operator for constructing structured low-rank matrices and its inverse operator.

[0084] d) Local low-rank constraint

[0085]

[0086]

[0087] in, is the diffusion image after data verification and channel merging in the kth iteration; is the phase-segmented region, where is the smooth phase region, It is the non-smooth phase region; is the singular value soft threshold operator, which means to decompose the matrix Z into a singular value and subtract the singular value from , and retain the singular values ​​greater than 0, where =0.01, =0.1; and They are the Fourier transform operator and the inverse Fourier transform operator; and Represents the operator for constructing structured low-rank matrices and its inverse operator. is the regularization parameter and takes the value as 2.0.

[0088] d) Update the image

[0089]

[0090] is the final reconstruction result of the kth iteration.

[0091] After 10 iterations, a reconstructed multi-excitation high-definition abdominal diffusion magnetic resonance image is finally obtained. Figure 1 shown.

[0092] In order to verify the technical effect of the present invention, the method of the present invention is compared with four traditional reconstruction algorithms: MUSE, MUSSELS, LLR, and PAIR, and a comprehensive evaluation of these algorithms in terms of reconstruction signal-to-noise ratio and artifact suppression capabilities is performed. Figure 3 The reconstruction results of the method proposed in the embodiment of the present invention in liver 4-time excitation high-definition diffusion imaging are compared with the comparison method. Among them:

[0093] Comparative method 1: MUSE (N.-k. Chen, A. Guidon, H.-C. Chang, and AW Song, "A robust multi-shot scan strategy for highresolution diffusion weighted MRI enabled by multiplexed sensitivity-encoding (MUSE)," Neuroimage, vol. 72, pp.41-47, 2013.).

[0094] Comparative method 2: MUSSELS (M. Mani, HK Aggarwal, V. Magnotta, and M. Jacob, "Improved MUSSELS reconstruction for high-resolution multi-shot diffusionweighted imaging," Magnetic Resonance in Medicine, vol. 83, no. 6, pp. 2253-2263, 2020).

[0095] Comparative method 3: LLR (Y. Hu et al., "Motion-robust reconstruction of multishot diffusion-weighted images without phase estimation through locally low-rank regularization," Magnetic Resonance in Medicine, vol. 81, no. 2, pp.1181-1190, 2019.).

[0096] Comparative method 4: PAIR (C. Qian et al., "A Paired Phase and MagnitudeReconstruction for Advanced Diffusion-Weighted Imaging," IEEE Transactions onBiomedical Engineering, vol. 70, no. 12, pp. 3425-3435, 2023.).

[0097] from Figure 3 From the comparison results, it can be seen that the method proposed in the present invention exhibits a higher reconstruction signal-to-noise ratio. While maintaining image details, the method more effectively reduces noise interference. In particular, when processing areas such as the liver and spleen where the motion phase is large and traditional algorithms are difficult to effectively suppress artifacts, the method of the present invention exhibits significant artifact suppression capabilities. The method of the present invention makes special treatment of motion artifacts in the algorithm design to achieve clearer image reconstruction. Compared with MUSE, the method of the present invention not only improves the reconstruction signal-to-noise ratio, but also significantly reduces artifacts caused by motion. Compared with MUSSELS, the present invention further improves the clarity of the image while maintaining high resolution. Although the LLR method also considers the suppression of motion artifacts, the method of the present invention performs better in this regard, especially when dealing with complex motion patterns. Although the PAIR method provides a new phase and amplitude reconstruction strategy, the method of the present invention still has obvious advantages in suppressing artifacts and maintaining image quality.

[0098] In summary, the method of the present invention is not only superior to traditional algorithms in terms of reconstruction signal-to-noise ratio and artifact suppression capabilities, but also shows higher stability and clarity when processing complex motion patterns. It performs well in high-definition diffusion imaging of moving parts, especially in the liver and other areas of the abdomen, and has broad application prospects in medical diagnosis, treatment monitoring, and biomedical research. It has significant application value and potential in high-definition diffusion imaging of moving parts.

[0099] The above embodiments are only preferred embodiments of the present invention and cannot 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 present invention.

Claims

1. A method for intelligent reconstruction of abdominal multi-excitation high-definition diffusion magnetic resonance imaging, characterized in that The following steps are involved: 1) Acquire frequency domain data of abdominal diffusion magnetic resonance images acquired using a multi-shot interleaved echo-planar sequence; 2) Design an intelligent high-definition abdominal diffusion MRI reconstruction model; 3) Construct a large amount of paired training data and use it to train the phase segmentation network in the reconstruction model to obtain the trained network parameters; 4) The convex set projection algorithm is used to solve the intelligent high-definition abdominal diffusion MRI reconstruction model and obtain the final reconstruction result.

2. The method for intelligent reconstruction of abdominal multi-excitation high-definition diffusion magnetic resonance imaging according to claim 1, characterized in that In step 1), the frequency domain data of the abdominal diffusion magnetic resonance image acquired by using a multi-excitation interleaved echo-planar sequence is obtained as follows: in, It represents the frequency domain data of the abdominal diffusion MRI image acquired by J times of H channels, also known as k-space, N and M represent the frequency encoding number and phase encoding number of a single diffusion MRI image respectively; is the diffusion magnetic resonance image of the target solution; is the channel sensitivity map of the signal receiving coil used to collect data; represents the noise during acquisition; is the sampling operator determined by the frequency domain trajectory, indicating zero filling at unsampled points; is the Fourier transform operator; and are the field of real numbers and the field of complex numbers respectively.

3. The method for intelligent reconstruction of abdominal multi-excitation high-definition diffusion magnetic resonance imaging according to claim 1, characterized in that In step 2), the intelligent high-definition abdominal diffusion magnetic resonance reconstruction model is: in, represents the frequency domain data of the abdominal diffusion magnetic resonance image acquired by J times of H channels, and N and M represent the frequency encoding number and phase encoding number of a single diffusion magnetic resonance image, respectively; is the diffusion magnetic resonance image of the target solution, is the channel sensitivity map of the signal receiving coil used to collect data; is the sampling operator determined by the frequency domain trajectory, indicating zero filling at unsampled points; is the Fourier transform operator; and are the domain of real numbers and the domain of complex numbers respectively; and are the square of the Frobenius norm and the nuclear norm, respectively; and are two regularization parameters; Indicates the first The operator of the region, represents the operator for selecting smooth phase regions in the image domain, represents the operator for selecting the non-smooth phase region in the image domain; is an operator for constructing structured low-rank matrices; is an operator that constructs a local low-rank matrix, where c is the coordinate of the center of the selected image block.

4. The method for intelligent reconstruction of abdominal multi-excitation high-definition diffusion magnetic resonance imaging according to claim 1, characterized in that In step 3), a large amount of paired training data is constructed and used for training the phase segmentation network in the reconstruction model to obtain trained network parameters, including the following steps: a) Design a phase segmentation network with multi-layer convolutional neural network ; It consists of L convolutional layers, including an input layer, an output layer, and L-2 hidden layers; each hidden layer contains Z convolution kernels, and the convolution kernel size is ; b) constructing a training data set for training the phase segmentation network; First, the frequency domain data of the abdominal high-definition diffusion MRI with J excitations in a total of S layers are collected; a 64×64 area is selected from the center of the frequency domain of the sth layer, and a multi-excitation diffusion image with low resolution but high signal-to-noise ratio is obtained using a traditional optimization algorithm. ; The phase of the diffusion image of the j-th excitation is Perform manual segmentation, j=1,2,…,J traverse all excitations and select smooth phase regions ;right Take the complement in the image domain and get the non-smooth phase region ; Change the phase and amplitude As network input, and As network training labels, a total of S×J pairs of training samples are obtained; c) using the constructed paired dataset for network training; Using the above dataset to train the denoiser, the loss function is defined as: ， in, is a phase division network, Represents the set of internal parameters of the entire network, and the trained network parameters are expressed as ; Σ represents the summation operation.

5. The method for intelligent reconstruction of abdominal multi-excitation high-definition diffusion magnetic resonance imaging according to claim 1, characterized in that In step 4), the convex set projection algorithm is used for solving, wherein the kth iteration is described as follows: a) Data verification items ， ， in, is the image after data verification in the kth iteration; is the channel sensitivity map of the signal receiving coil used to collect data; is the image after channel merging in the kth iteration; represents the frequency domain data of the abdominal diffusion magnetic resonance image acquired by J times of excitation of H channels; are the diffusion magnetic resonance images in the kth iteration, and their initial values ​​are It is reconstructed by selecting a 64×64 region from the center of the frequency domain using a traditional optimization algorithm; N, M, J, and H represent the frequency encoding number, phase encoding number, excitation number, and receiving coil channel number of a single diffusion magnetic resonance image, respectively; and They are the frequency domain sampling operator and its adjoint operator; and They are the Fourier transform operator and the inverse Fourier transform operator; and are the domain of real numbers and the domain of complex numbers respectively; is the regularization parameter; b) Phase division in, is the image of the jth excitation after data verification and channel merging in the kth iteration; and is the operator for constructing structured low-rank matrices and its adjoint operator; is the trained phase segmentation network, and its weight parameter is ; and They are the phase smooth area and non-smooth area obtained by segmenting the image of the jth excitation. The phase segmentation results of multiple excitations are obtained by segmenting the Jth excitation number in sequence. and ; c) Structured low-rank constraints in, is the diffusion image after data verification and channel merging in the kth iteration; is the phase-segmented region, where is the smooth phase region, It is the non-smooth phase region; is the singular value hard threshold truncation operator, which means to perform singular value decomposition on the matrix Z and retain the first r p singular values, and They are the Fourier transform operator and the inverse Fourier transform operator; and represents the operator for constructing structured low-rank matrices and its inverse operator; d) Local low-rank constraint in, is the diffusion image after data verification and channel merging in the kth iteration; is the phase-segmented region, where is the smooth phase region, It is the non-smooth phase region; is the singular value soft threshold operator, which means to decompose the matrix Z into a singular value and subtract the singular value from , and retain singular values ​​greater than 0, and They are the Fourier transform operator and the inverse Fourier transform operator; and represents the operator for constructing structured low-rank matrices and its inverse operator; is the regularization parameter; e) Update the image is the final reconstruction result of the kth iteration; After K iterations, a reconstructed multi-excitation high-definition abdominal diffusion magnetic resonance image is finally obtained.

Images