A nuclear magnetic resonance image reconstruction method and a terminal

Abstract

The present invention provides a nuclear magnetic resonance image reconstruction method and a terminal, which obtain a nuclear magnetic resonance image sample set, and each nuclear magnetic resonance image sample in the nuclear magnetic resonance image sample set includes a K-space image sample and a first reconstructed image; pass the K-space image sample through a preset deep neural network to obtain a predicted image and a set of predicted sensitivity mapping kernels; and obtain a second reconstructed image according to the predicted image and the set of predicted sensitivity mapping kernels; optimize the preset deep neural network according to the difference between the first reconstructed image and the second reconstructed image to obtain a target deep neural network; reconstruct a nuclear magnetic resonance image to be processed according to the target deep neural network; the present invention can ensure the stability of the image reconstruction effect under the change of different parameters, and does not depend on a determined set of sensitivity mapping kernels, realizing high-robustness nuclear magnetic resonance image reconstruction.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular, to a nuclear magnetic resonance image reconstruction method and a terminal. Background Art

[0002] Using nuclear magnetic resonance imaging to generate tissue images of the human body for auxiliary judgment of various diseases is a commonly used technical means, and because nuclear magnetic resonance imaging does not have the problem of radiation exposure, it is safer for the human body. However, during the process of nuclear magnetic resonance imaging, the image reconstruction method has a greater impact on the final imaging effect, and problems such as image artifacts often occur, affecting the quality of the image. Summary of the Invention

[0003] The technical problem to be solved by the present invention is: to provide a nuclear magnetic resonance image reconstruction method and a terminal, so as to achieve a more robust nuclear magnetic resonance image reconstruction.

[0004] To solve the above technical problem, a technical solution adopted by the present invention is:

[0005] A nuclear magnetic resonance image reconstruction method, comprising the steps of:

[0006] Obtain a set of nuclear magnetic resonance image samples, each nuclear magnetic resonance image sample in the set of nuclear magnetic resonance image samples includes a K-space image sample and a first reconstructed image;

[0007] Pass the K-space image sample through a preset deep neural network to obtain a predicted image and a set of predicted sensitivity mapping kernels; and obtain a second reconstructed image according to the predicted image and the set of predicted sensitivity mapping kernels;

[0008] Optimize the preset deep neural network according to the difference between the first reconstructed image and the second reconstructed image to obtain a target deep neural network;

[0009] Reconstruct the nuclear magnetic resonance image to be processed according to the target deep neural network.

[0010] To solve the above technical problem, another technical solution adopted by the present invention is:

[0011] A nuclear magnetic resonance image reconstruction terminal, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, the following steps are implemented:

[0012] Obtain a set of nuclear magnetic resonance image samples, each nuclear magnetic resonance image sample in the set of nuclear magnetic resonance image samples includes a K-space image sample and a first reconstructed image;

[0013] Pass the K-space image sample through a preset deep neural network to obtain a predicted image and a set of predicted sensitivity mapping kernels; and obtain a second reconstructed image according to the predicted image and the set of predicted sensitivity mapping kernels;

[0014] Optimize the preset deep neural network according to the difference between the first reconstructed image and the second reconstructed image to obtain a target deep neural network;

[0015] Reconstruct the nuclear magnetic resonance image to be processed according to the target deep neural network.

[0016] The beneficial effects of the present invention are as follows: The nuclear magnetic resonance image sample includes a K-space image sample and a first reconstructed image. Since the originally acquired nuclear magnetic resonance image is a K-space image in the frequency domain, a nuclear magnetic resonance image sample with a better reconstruction effect is obtained as a training image. During the training process, the set of sensitivity mapping kernels is used as an unknown to participate in the training optimization, and the preset deep neural network is optimized according to the difference between the predicted second reconstructed image and the already determined first reconstructed image, so that the finally obtained target deep neural network can ensure the stability of the image reconstruction effect under different parameter changes and does not depend on a determined set of sensitivity mapping kernels, realizing high-robustness nuclear magnetic resonance image reconstruction. Description of the Drawings

[0017] Figure 1 is a flowchart of the steps of a nuclear magnetic resonance image reconstruction method according to an embodiment of the present invention;

[0018] Figure 2 is a schematic structural diagram of a nuclear magnetic resonance image reconstruction terminal according to an embodiment of the present invention;

[0019] Label Description:

[0020] 1. A nuclear magnetic resonance image reconstruction terminal; 2. A processor; 3. A memory. Detailed Embodiments

[0021] To describe the technical content, achieved objectives and effects of the present invention in detail, the following is described in conjunction with the embodiments and with reference to the accompanying drawings.

[0022] Please refer to Figure 1 , a nuclear magnetic resonance image reconstruction method, including the steps of:

[0023] Obtain a set of nuclear magnetic resonance image samples, each nuclear magnetic resonance image sample in the set of nuclear magnetic resonance image samples including a K-space image sample and a first reconstructed image;

[0024] Pass the K-space image sample through a preset deep neural network to obtain a predicted image and a set of predicted sensitivity mapping kernels; and obtain a second reconstructed image according to the predicted image and the set of predicted sensitivity mapping kernels;

[0025] Optimize the deep neural network according to the difference between the first reconstructed image and the second reconstructed image to obtain a target deep neural network;

[0026] Reconstruct the nuclear magnetic resonance image to be processed according to the target deep neural network.

[0027] As can be seen from the above description, the beneficial effects of the present invention are as follows: The nuclear magnetic resonance image sample includes a K-space image sample and a first reconstructed image. Since the originally acquired nuclear magnetic resonance image is a K-space image in the frequency domain, by obtaining a nuclear magnetic resonance image sample with better reconstruction effect as the training image, the sensitivity mapping kernel set is used as an unknown in the training process for training optimization, and the preset deep neural network is optimized according to the difference between the predicted second reconstructed image and the determined first reconstructed image, so that the finally obtained target deep neural network can ensure the stability of the image reconstruction effect under the change of different parameters, and does not depend on the determined sensitivity mapping kernel set, realizing high-robustness nuclear magnetic resonance image reconstruction.

[0028] Further, the step of passing the K-space image sample through a preset deep neural network to obtain a predicted image and a predicted sensitivity mapping kernel set includes:

[0029] Establish a total optimization objective Y = ;

[0030] where y represents the K-space image sample, A m represents a linear operator composed of fixed variables and a sampling mask A, m represents the predicted image to be calculated, λ represents a undetermined coefficient, and D(m) represents the predicted image to be calculated in the preset deep neural network D;

[0031] Obtain the predicted image m and the predicted sensitivity mapping kernel set s by alternately selecting an optimization algorithm.

[0032] As can be seen from the above description, by establishing a total optimization objective and performing learning of the preset deep neural network according to the total optimization objective, especially in the process of obtaining the predicted image to be calculated, it does not depend on the prior prediction of the sensitivity, avoiding the influence of the accuracy of the predicted sensitivity on the final result.

[0033] Further, the step of obtaining the predicted image m and the predicted sensitivity mapping kernel set s by alternately selecting an optimization algorithm includes:

[0034] Expand the total optimization objective Y according to the alternately selected optimization algorithm to obtain an expansion formula:

[0035] ;

[0036] ;

[0037] Among them, , where x represents the predicted image m or the predicted sensitivity mapping kernel set s, F represents the Fourier transform; s+ and m+ respectively represent the predicted sensitivity mapping kernel set and the predicted image generated in the next iteration after updating the preset deep neural network;

[0038] The predicted image m and the predicted sensitivity mapping kernel set s are obtained by solving the expansion according to the conjugate gradient algorithm.

[0039] As can be seen from the above description, by expanding the total optimization objective through the alternating optimization algorithm, the optimization of the total optimization objective is specifically the optimization of the predicted image and the predicted sensitivity kernel set, realizing the acquisition of the key parameters of image reconstruction, and solving through the conjugate gradient algorithm to ensure that the finally obtained solution is the optimal solution and as close as possible to the optimal result.

[0040] Furthermore, the optimizing the preset deep neural network according to the difference between the first reconstructed image and the second reconstructed image includes:

[0041] Setting a supervision loss, where the supervision loss is the structural similarity index between the first reconstructed image and the second reconstructed image;

[0042] Optimizing the preset deep neural network according to the structural similarity index.

[0043] As can be seen from the above description, the model can be optimized through continuous learning, enhancing the robustness in the image reconstruction process and being robust to the unique distributions generated by different sampling parameters.

[0044] Furthermore, obtaining the second reconstructed image according to the predicted image and the predicted sensitivity mapping kernel set includes:

[0045] ;

[0046] Among them, represents the second reconstructed image, F represents the Fourier transform, N represents the number of iterations, C represents the total number of predicted sensitivity mapping kernels, and si represents the i-th predicted sensitivity mapping kernel in the predicted sensitivity mapping kernel set s.

[0047] As can be seen from the above description, image reconstruction is realized through the predicted image and the predicted sensitivity kernel set, without the need to preset or mark the sensitivity matrix of the device in advance, saving the preparation time for image reconstruction, improving the efficiency of image reconstruction, and increasing the autonomy in the optimization process of the preset deep neural network by estimating the image training end-to-end network with square roots, ensuring the image reconstruction accuracy of the final target deep neural network.

[0048] Please refer to Figure 2 , a nuclear magnetic resonance image reconstruction terminal, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the following steps are implemented:

[0049] Obtain a set of nuclear magnetic resonance image samples, where each nuclear magnetic resonance image sample in the set of nuclear magnetic resonance image samples includes a K-space image sample and a first reconstructed image;

[0050] Pass the K-space image sample through a preset deep neural network to obtain a predicted image and a set of predicted sensitivity mapping kernels; and obtain a second reconstructed image based on the predicted image and the set of predicted sensitivity mapping kernels;

[0051] Optimize the preset deep neural network according to the difference between the first reconstructed image and the second reconstructed image to obtain a target deep neural network;

[0052] Reconstruct the nuclear magnetic resonance image to be processed according to the target deep neural network.

[0053] The beneficial effects of the present invention are as follows: The nuclear magnetic resonance image sample includes a K-space image sample and a first reconstructed image. Since the originally acquired nuclear magnetic resonance image is a K-space image in the frequency domain, by obtaining nuclear magnetic resonance image samples with better reconstruction effects as training images, the set of sensitivity mapping kernels is used as an unknown in the training process to participate in training optimization, and the preset deep neural network is optimized according to the difference between the predicted second reconstructed image and the already determined first reconstructed image, so that the finally obtained target deep neural network can ensure the stability of the image reconstruction effect under the change of different parameters and does not depend on the determined set of sensitivity mapping kernels, realizing high-robustness nuclear magnetic resonance image reconstruction.

[0054] Further, the step of passing the K-space image sample through a preset deep neural network to obtain a predicted image and a set of predicted sensitivity mapping kernels includes:

[0055] Establish a total optimization target Y = ;

[0056] where y represents the K-space image sample, A m represents a linear operator composed of fixed variables and a sampling mask A, m represents the predicted image to be calculated, λ represents a undetermined coefficient, and D(m) represents the predicted image to be calculated in the preset deep neural network D;

[0057] Obtain the predicted image m and the set of predicted sensitivity mapping kernels s by alternately selecting an optimization algorithm.

[0058] As described above, by establishing the overall optimization objective and performing learning on the deep neural network according to the overall optimization objective, especially in the process of obtaining the prediction image to be calculated, it does not rely on the prior prediction of sensitivity, avoiding the influence of the accuracy of the predicted sensitivity on the final result.

[0059] Further, the obtaining of the prediction image m and the prediction sensitivity mapping kernel set s by alternately selecting the optimization algorithm includes:

[0060] Expanding the overall optimization objective Y according to the alternately selected optimization algorithm to obtain an expansion formula:

[0061] ;

[0062] ;

[0063] wherein, , where x represents the prediction image m or the prediction sensitivity mapping kernel set s, F represents the Fourier transform; s+ and m+ respectively represent the prediction sensitivity mapping kernel set and the prediction image generated in the next iteration after updating the preset deep neural network;

[0064] Solving the expansion formula according to the conjugate gradient algorithm to obtain the prediction image m and the prediction sensitivity mapping kernel set s.

[0065] As described above, by expanding the overall optimization objective through the alternating optimization algorithm, the optimization of the overall optimization objective is specifically the optimization of the prediction image and the prediction sensitivity kernel set, realizing the acquisition of the key parameters for image reconstruction, and solving through the conjugate gradient algorithm to ensure that the finally obtained solution is the optimal solution and as close as possible to the optimal result.

[0066] Further, the optimizing the deep neural network according to the difference between the first reconstructed image and the second reconstructed image includes:

[0067] Setting a supervision loss, where the supervision loss is the structural similarity index between the first reconstructed image and the second reconstructed image;

[0068] Optimizing the deep neural network according to the structural similarity index.

[0069] As described above, the model can be optimized through continuous learning, enhancing the robustness in the image reconstruction process and being robust to the unique distributions generated by different sampling parameters.

[0070] Further, obtaining the second reconstructed image according to the prediction image and the prediction sensitivity mapping kernel set includes:

[0071] ;

[0072] Among them, represents the second reconstructed image, F represents the Fourier transform, N represents the number of iterations, C represents the total number of prediction sensitivity mapping kernels, and s i represents the i-th prediction sensitivity mapping kernel in the set of prediction sensitivity mapping kernels s.

[0073] As can be seen from the above description, image reconstruction is achieved through the predicted image and the set of prediction sensitivity kernels, without the need to preset or mark the sensitivity matrix of the device in advance, saving the preparation time for image reconstruction, improving the efficiency of image reconstruction, and training the end-to-end network for image through square root estimation, increasing the autonomy of the deep neural network and ensuring the image reconstruction accuracy of the final target deep neural network.

[0074] The above-mentioned nuclear magnetic resonance image reconstruction method and terminal of the present invention can be applied to the reconstruction process of nuclear magnetic resonance images. For example, in parallel nuclear magnetic resonance image reconstruction, the following is described through specific embodiments:

[0075] Please refer to Figure 1 , the first embodiment of the present invention is:

[0076] A nuclear magnetic resonance image reconstruction method, including the steps of:

[0077] S1. Obtain a set of nuclear magnetic resonance image samples, and each nuclear magnetic resonance image sample in the set of nuclear magnetic resonance image samples includes a K-space image sample and a first reconstructed image;

[0078] Among them, the K-space image represents a frequency-domain image, such as an image in the Fourier domain;

[0079] S2. Pass the K-space image sample through a preset deep neural network to obtain a predicted image and a set of prediction sensitivity mapping kernels; and obtain a second reconstructed image according to the predicted image and the set of prediction sensitivity mapping kernels. Specifically:

[0080] S21. Establish a total optimization target Y = ;

[0081] Among them, y represents the K-space image sample, A m represents a linear operator composed of fixed variables and a sampling mask A. The sampling mask A can be represented in matrix form. m represents the predicted image to be calculated, λ represents a undetermined coefficient, and D(m) represents the predicted image to be calculated in the preset deep neural network D; where "‖‖" represents the norm;

[0082] According to the formula in S21, each step can be divided into two different sub - problems: the first sub - problem treats D(m) as a constant and uses the conjugate gradient algorithm to update m. The second sub - problem treats D() as a proximal operator and solves it by direct assignment, without relying on a pre - computed estimate of the sensitivity map, but treating it as an optimization variable;

[0083] S22. Obtain the predicted image m and the set of predicted sensitivity mapping kernels s by alternately selecting optimization algorithms:

[0084] S221. Expand the total optimization objective Y according to the alternately selected optimization algorithm to obtain an expansion formula:

[0085] (1);

[0086] (2);

[0087] Among them, , where x represents the predicted image m or the set of predicted sensitivity mapping kernels s, A s represents the second linear operator composed of fixed variables and the sampling mask A, A m represents the first linear operator composed of fixed variables and the sampling mask A, R s (s) represents the regularization term that enforces the prior on the set of sensitivity mapping kernels s; R m (m) represents the regularization term that enforces the prior on the predicted image m; F represents the Fourier transform; s+ and m+ respectively represent the set of predicted sensitivity mapping kernels and the predicted image generated in the next iteration after updating the preset deep neural network;

[0088] And, in all expansions, set =D m and =D s , j is the depth of the preset deep neural network, so that learnable weights can be effectively used; the coefficients λ s and λm are also learnable. Optimization is initialized with m (0) and s (0) , that is, the initial values without going through the iterations of the deep neural network, obtained using square - root estimation. Formulas (1) and (2) are approximately solved by the n1 and n2 steps of the conjugate gradient algorithm respectively. s + and m + are the Fourier transforms of direct assignment; these two neural networks are used as generalized denoisers and are trained in an end - to - end manner after several N external steps of alternating optimization; the conjugate gradient algorithm is based on the undersampled data y and the measurement matrices A s and A mThe loss execution given, each block matches its corresponding equation;

[0089] S222. Solve the expansion according to the conjugate gradient algorithm to obtain the predicted image m and the predicted sensitivity mapping kernel set s;

[0090] S23. Obtain the second reconstructed image according to the predicted image and the predicted sensitivity mapping kernel set ;

[0091] wherein, represents the second reconstructed image, F represents the Fourier transform, N represents the number of iterations of the preset deep neural network, C represents the total number of predicted sensitivity mapping kernels, and s i represents the i-th predicted sensitivity mapping kernel in the predicted sensitivity mapping kernel set s;

[0092] S3. Optimize the deep neural network according to the difference between the first reconstructed image and the second reconstructed image to obtain the target deep neural network, including:

[0093] S31. Set the supervision loss, and the supervision loss is the structural similarity index between the first reconstructed image and the second reconstructed image;

[0094] S32. Optimize the preset deep neural network according to the structural similarity index (SSIM), that is, L = -SSIM( , x), where x identifies the first reconstructed image;

[0095] S4. Reconstruct the to-be-processed nuclear magnetic resonance image according to the target deep neural network.

[0096] Embodiment 2 of the present invention is:

[0097] A nuclear magnetic resonance image reconstruction method, which is different from Embodiment 1 in that S4 includes:

[0098] S41. Receive the to-be-processed nuclear magnetic resonance image and preprocess the to-be-processed nuclear magnetic resonance image:

[0099] Pre-generate a sampling mask, and the sampling mask is equivalent to a mask, and the image data is processed within this range;

[0100] Obtain relevant parameters: MRI (Magnetic Resonance Imaging) sample list index, number of MRI slices nums, number of central MRI slices center_slice_idx, downsampling factor, sampling mask, direction y; and pre-estimate the sensitivity map maps;

[0101] Convert the data corresponding to the nuclear magnetic resonance images to be processed in the MRI sample list into digital tensors;

[0102] Separate the slices and samples; map always counts from zero, and the initial value of the counting parameter count = 0;

[0103] Load the nuclear magnetic resonance images to be processed;

[0104] Obtain the K-space of a specific slice, where the specific slice is the slice corresponding to the region of interest;

[0105] Store the core files, which include nuclear magnetic resonance images, sampling masks, sensitivity maps, digital tensor files, etc.;

[0106] Determine whether an external sensitivity map needs to be loaded. If so, load the external sensitivity map; in an alternative implementation, the external sensitivity map is loaded during regularization and optimization;

[0107] Obtain the sensitivity map corresponding to a specific slice;

[0108] Calculate the total energy of the lines in the MRI device coil; line_energy; the range is line_energy < 1e-16;

[0109] Fill the data basis, that is, the parameters required during the calculation process;

[0110] Always remove even-numbered lines to keep the image centered originally;

[0111] Store all the K-space data corresponding to the nuclear magnetic resonance images to be processed;

[0112] Completely remove the dead lines, that is, the lines that are not used during the calculation process, and also delete them from the frequency representation of the map;

[0113] Store the K-space data without zero rows;

[0114] Obtain the central and non-central positions;

[0115] Set the fixed percentage of the center line;

[0116] (1) Given a fixed number of center lines, that is, preset a certain number, and obtain the corresponding number of center lines;

[0117] Downsample, pick up lines to ensure R = downsampling; where R represents the regularization term that enforces the prior on the sensitivity mapping kernel set s or the predicted image m

[0118] Calculate the candidate regions outside the central region, that is, the regions other than (1);

[0119] If there is no mask, immediately generate the mask mask;

[0120] Pick a random line outside the central position;

[0121] Create a sampling mask and downsampled k-space data k_sampling_mask;

[0122] Use the corresponding mask data;

[0123] Normalize the k-space;

[0124] Scale the square root estimate of the k-space data;

[0125] Obtain the initial sensitivity map;

[0126] Calculate the complex-to-real k-space;

[0127] S42. Reconstruct the preprocessed nuclear magnetic resonance image to be processed according to the target deep neural network.

[0128] Please refer to Figure 2 , and the third embodiment of the present invention is:

[0129] A nuclear magnetic resonance image reconstruction terminal 1, including a processor 2, a memory 3, and a computer program stored on the memory 3 and executable on the processor 2. When the processor 2 executes the computer program, each step in Embodiment 1 is implemented.

[0130] In summary, the present application provides a nuclear magnetic resonance image reconstruction method and terminal. An image reconstruction of nuclear magnetic resonance is realized by establishing a deep learning method based on unfolded alternating minimization, and the robustness of the reconstruction process is enhanced. A learnable model is intertwined with the optimization steps, and the entire system is end-to-end training and supervised loss. Using an uncalibrated method for the structure in a parallel nuclear magnetic resonance (MRI) model, the sensitivity map varies smoothly in space and a low-rank structure is imposed. Through alternating optimization, the image and sensitivity map kernels in the k-space are directly solved. It is robust to the distribution displacement generated by different sampling parameters.

[0131] The above are only the embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent transformation made using the specification and drawings of the present invention, or directly or indirectly applied in the related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A nuclear magnetic resonance image reconstruction method, characterized in that, it includes the steps of: Obtain a nuclear magnetic resonance image sample set, where each nuclear magnetic resonance image sample in the nuclear magnetic resonance image sample set includes a K-space image sample and a first reconstructed image; Pass the K-space image sample through a preset deep neural network to obtain a predicted image and a set of predicted sensitivity mapping kernels; and obtain a second reconstructed image according to the predicted image and the set of predicted sensitivity mapping kernels; Optimize the preset deep neural network according to the difference between the first reconstructed image and the second reconstructed image to obtain a target deep neural network; Reconstruct the nuclear magnetic resonance image to be processed according to the target deep neural network; The step of passing the K-space image sample through a preset deep neural network to obtain a predicted image and a set of predicted sensitivity mapping kernels includes: Establish the total optimization objective Y based on the K-space image samples ; Where y represents the K-space image sample, Am represents a linear operator composed of fixed variables and a sampling mask A, m represents the predicted image to be calculated, λ represents an undetermined coefficient, and D(m) represents the predicted image to be calculated in the preset deep neural network D; Obtain the predicted image m and the set of predicted sensitivity mapping kernels s through an alternating selection optimization algorithm; The step of optimizing the preset deep neural network according to the difference between the first reconstructed image and the second reconstructed image includes: Set a supervision loss, and the supervision loss is the structural similarity index between the first reconstructed image and the second reconstructed image; Optimize the preset deep neural network according to the structural similarity index; The step of obtaining a second reconstructed image according to the predicted image and the set of predicted sensitivity mapping kernels includes: ； Among them, represents the second reconstructed image, F represents the Fourier transform, N represents the number of iterations, C represents the total number of prediction sensitivity mapping kernels, and s i represents the i-th prediction sensitivity mapping kernel in the set s of prediction sensitivity mapping kernels.

2. The nuclear magnetic resonance image reconstruction method according to claim 1, characterized in that, The step of obtaining the predicted image m and the set of predicted sensitivity mapping kernels s through an alternating selection optimization algorithm includes: Expand the total optimization target Y according to the alternating selection optimization algorithm to obtain an expansion formula: ； ； Among them, , where x represents the predicted image m or the predicted sensitivity mapping kernel set s, F represents the Fourier transform; s+ and m+ respectively represent the predicted sensitivity mapping kernel set and the predicted image generated in the next iteration after updating the preset deep neural network. Solve the expansion formula according to the conjugate gradient algorithm to obtain the predicted image m and the set of predicted sensitivity mapping kernels s.

3. A nuclear magnetic resonance image reconstruction terminal, including a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, when the processor executes the computer program, the following steps are implemented: Obtain a nuclear magnetic resonance image sample set, where each nuclear magnetic resonance image sample in the nuclear magnetic resonance image sample set includes a K-space image sample and a first reconstructed image; Pass the K-space image sample through a preset deep neural network to obtain a predicted image and a set of predicted sensitivity mapping kernels; and obtain a second reconstructed image according to the predicted image and the set of predicted sensitivity mapping kernels; Optimize the preset deep neural network according to the difference between the first reconstructed image and the second reconstructed image to obtain a target deep neural network; Reconstruct the nuclear magnetic resonance image to be processed according to the target deep neural network; The step of passing the K-space image sample through a preset deep neural network to obtain a predicted image and a set of predicted sensitivity mapping kernels includes: Establish the total optimization objective Y based on the K-space image samples ; Wherein, y represents the K-space image sample, Am represents a linear operator composed of fixed variables and a sampling mask A, m represents the predicted image to be calculated, λ represents an undetermined coefficient, and D(m) represents the predicted image to be calculated in the preset deep neural network D; The predicted image m and the set of predicted sensitivity mapping kernels s are obtained by an alternating selection optimization algorithm; The optimizing the deep neural network according to the difference between the first reconstructed image and the second reconstructed image includes: Setting a supervision loss, where the supervision loss is the structural similarity index between the first reconstructed image and the second reconstructed image; Optimizing the deep neural network according to the structural similarity index; Obtaining the second reconstructed image according to the predicted image and the set of predicted sensitivity mapping kernels includes: ； Among them, represents the second reconstructed image, F represents the Fourier transform, N represents the number of iterations, C represents the total number of prediction sensitivity mapping kernels, and s i represents the i-th prediction sensitivity mapping kernel in the prediction sensitivity mapping kernel set s.

4. A nuclear magnetic resonance image reconstruction terminal according to claim 3, characterized in that The obtaining the predicted image m and the set of predicted sensitivity mapping kernels s by the alternating selection optimization algorithm includes: Expanding the total optimization objective Y according to the alternating selection optimization algorithm to obtain an expansion formula: ； ； Among them, , where x represents the predicted image m or the predicted sensitivity mapping kernel set s, F represents the Fourier transform; s+ and m+ respectively represent the predicted sensitivity mapping kernel set and the predicted image generated in the next iteration after updating the preset deep neural network; Solving the expansion formula according to the conjugate gradient algorithm to obtain the predicted image m and the set of predicted sensitivity mapping kernels s.

Images