Real-time magnetic resonance image reconstruction method and system based on deep unfolding neural network

Abstract

The application provides a real-time magnetic resonance image reconstruction method and system based on a deep unfolding neural network, and comprises the following steps: step S1: building a deep unfolding neural network based on low-rank and sparse decomposition; step S2: training the built deep unfolding neural network by using simulation data; and step S3: using the trained deep unfolding neural network to perform real-time image reconstruction. By adopting the mode of unfolding an iterative algorithm based on low-rank and sparse decomposition into a deep neural network, the application solves the problem of long reconstruction time of the iterative algorithm applied to real-time magnetic resonance imaging, and realizes real-time reconstruction of magnetic resonance interventional images.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of magnetic resonance image reconstruction, and in particular to a real-time magnetic resonance image reconstruction method and system based on a deep unfolding neural network. BACKGROUND

[0002] Magnetic resonance imaging is a medical imaging technique that was officially used in clinical practice in 1982 internationally. It uses static magnetic fields and radio frequency magnetic fields to image human tissues. In the imaging process, clear images with high contrast can be obtained without using electronic ion radiation or contrast agents. It can reflect organ abnormalities and early lesions from the inside of human molecules. It is superior to X-ray CT in many ways. Although X-CT solves the problem of image overlap, the images provided are still spatial distribution images of tissue X-ray absorption, and cannot provide physiological state information of human organs. When the absorption coefficient of the diseased tissue and the surrounding normal tissue is the same, valuable information cannot be provided. Only when the disease develops to change the shape, position and size of the organ to give an abnormal feeling can it be found. In addition to the anatomical type characteristics of X-ray CT, i.e. obtaining non-overlapping proton density body layer images, the magnetic resonance imaging device can also accurately measure the atomic nucleus relaxation time T1 and T2 by using the principle of nuclear magnetic resonance, and can reflect the information of the chemical structure in human tissues. The image reconstructed by the computer is a composition image (chemical structure image), which has the ability to display and represent different tissues with the same density and different chemical structures of the same tissue through images. This facilitates the differentiation between gray matter and white matter in the brain, and has great superiority in early diagnosis of tissue necrosis, malignant diseases and degenerative diseases, and the contrast of soft tissues is more accurate.

[0003] The Chinese invention patent document with publication number CN112881958A discloses a magnetic resonance interventional imaging method based on low-rank and sparse decomposition. It includes: during the interventional process, continuously collecting k-space data using a golden angle radial sampling method, grouping the collected k-space data, and using a method based on low-rank and sparse decomposition and framelet transform to reconstruct the magnetic resonance interventional image.

[0004] The Chinese invention patent document with publication number CN111932650A discloses a nuclear magnetic resonance image reconstruction method based on a high-throughput deep unfolding network. A high-throughput deep unfolding network is designed for training and reconstruction. The method includes the following steps: step 1: constructing a training data set; step 2: constructing a high-throughput deep unfolding network; step 3: high-throughput deep unfolding network training process; step 4: applying the trained high-throughput deep unfolding network to the compressed sensing magnetic resonance image reconstruction process.

[0005] For the related technology in the above, the inventors believe that the iterative algorithm in the above method has multiple regularizations that need to be manually tuned, which is very tedious and time-consuming, and a large number of iteration steps are required to obtain a good reconstruction effect, which takes a long time to calculate and is not suitable for real-time magnetic resonance interventional imaging. SUMMARY

[0006] In view of the defects in the prior art, the purpose of the present application is to provide a real-time magnetic resonance image reconstruction method and system based on a deep unfolding neural network.

[0007] According to the present application, a real-time magnetic resonance image reconstruction method based on a deep unfolding neural network is provided, comprising the following steps:

[0008] Step S1: building a deep unfolding neural network based on low-rank and sparse decomposition;

[0009] Step S2: training the built deep unfolding neural network with simulation data;

[0010] Step S3: using the trained deep unfolding neural network for real-time image reconstruction.

[0011] Preferably, the step S1 comprises the following steps:

[0012] Step S1.1: in magnetic resonance imaging, the changing background and continuously dynamically changing interventional features are decomposed into a low-rank matrix L and a sparse matrix S from an interventional image sequence x, i.e. x=L+S, and the spatial sparse characteristics of L and S are used to build a low-rank and sparse decomposition-based model for magnetic resonance interventional image reconstruction;

[0013] Step S1.2: unfolding the low-rank and sparse decomposition-based model into a deep neural network; the deep unfolding neural network is composed of multiple iteration blocks, each iteration block corresponds to one iteration in the low-rank and sparse decomposition-based iterative algorithm, the sparse transformation ψ in the low-rank and sparse decomposition-based iterative algorithm is replaced by a structure composed of multiple convolution layers and activation functions, the weight parameters of the convolution kernel in the convolution layer are learned during the training of the deep unfolding neural network, and the regularization parameters in the low-rank and sparse decomposition-based iterative algorithm are learned during the training of the deep unfolding neural network;

[0014] Step S1.3: the first input of the first iteration block is the parameter after network initialization, the first input of the second and subsequent iteration blocks is the output of the previous iteration block, and the second input of each iteration block includes k-space data obtained by radial sampling and a coil sensitivity map, and finally a reconstructed image is output after passing through all the iteration blocks.

[0015] Preferably, in the step S1.1, the low-rank and sparse decomposition-based model for magnetic resonance interventional image reconstruction is:

[0016]

[0017] where E = ΩFC is the k-space encoding operator, C is the coil sensitivity, F is the Fourier transform, Ω is the sampling pattern; d is the acquired k-space data; is the TV transform along the S temporal direction; ψ is the sparse transform, λ L , λ s , and is the regularization parameter; k-space denotes the Fourier space.

[0018] Preferably, the step S2 comprises the following steps:

[0019] Step S2.1: obtaining a simulated intervention image sequence as a training set by simulation, performing non-uniform Fourier transform on the intervention image sequence to obtain simulated radial sampling k-space data;

[0020] Step S2.2: taking the simulated radial sampling k-space data and the coil sensitivity map as the input of the deep unfolding neural network to be trained, and taking the deep unfolding neural network as the output of the reconstructed image;

[0021] Step S2.3: calculating the loss of the reconstructed image output by the deep unfolding neural network and the intervention image sequence, updating the parameters in the deep unfolding neural network by the way of back propagation according to the loss, the parameters in the deep unfolding neural network including the weights of the convolution kernel in the convolution layer and the regularization parameter; obtaining the trained network parameters after multiple rounds of updating and training, that is, obtaining the trained deep unfolding neural network model.

[0022] Preferably, the step S3 comprises the following steps:

[0023] Step S3.1: acquiring a full sampling reference image before real-time imaging, for estimating the coil sensitivity map;

[0024] Step S3.2: acquiring radial k-space data in real time during real-time intervention imaging, taking the acquired k-space data and the estimated coil sensitivity map as the input of the trained deep unfolding neural network, and then reconstructing the magnetic resonance intervention image in real time.

[0025] According to the real-time magnetic resonance image reconstruction system based on the deep unfolding neural network provided by the application, the following modules are included:

[0026] Module M1: building a deep unfolding neural network based on low-rank and sparse decomposition;

[0027] Module M2: training the built deep unfolding neural network with simulation data;

[0028] Module M3: Real-time reconstruction of images using the trained deep unfolding neural network.

[0029] Preferably, the module M1 comprises the following modules:

[0030] Module M1.1: In magnetic resonance imaging, decompose the varying background and continuously changing interventional features from the interventional image sequence x into a low-rank matrix L and a sparse matrix S, i.e. x = L + S, by using the spatial sparsity characteristics of L and S, a low-rank and sparse decomposition-based model is used for magnetic resonance interventional image reconstruction;

[0031] Module M1.2: Unfold the low-rank and sparse decomposition-based model into a deep neural network; the deep unfolding neural network is composed of multiple iteration blocks, each iteration block corresponds to one iteration in the low-rank and sparse decomposition-based iterative algorithm, the sparse transformation ψ in the low-rank and sparse decomposition-based iterative algorithm is replaced by a structure composed of multiple convolutional layers and activation functions, the weight parameters of the convolution kernel in the convolutional layer are learned during the training of the deep unfolding neural network, and the regularization parameter in the low-rank and sparse decomposition-based iterative algorithm is learned during the training of the deep unfolding neural network;

[0032] Module M1.3: The first input of the first iteration block is the network initialized parameter, the first input of the second and subsequent iteration blocks is the output of the previous iteration block, and the second input of each iteration block includes the k-space data obtained by radial sampling and the coil sensitivity map. The reconstructed image is finally output after passing through all the iteration blocks.

[0033] Preferably, in the module M1.1, the low-rank and sparse decomposition-based model for magnetic resonance interventional image reconstruction is:

[0034]

[0035] Wherein, E = ΩFC is the k-space encoding operator, C is the coil sensitivity, F is the Fourier transform, and Ω is the sampling mode; d is the acquired k-space data; is the TV transform along the S time direction; ψ is the sparse transform, λ L , λ s , and is the regularization parameter; k-space represents the Fourier space.

[0036] Preferably, the module M2 comprises the following modules:

[0037] Module M2.1: Obtain the interventional image sequence by simulation as a training set, perform non-uniform Fourier transform on the interventional image sequence to obtain simulated radial sampling k-space data;

[0038] Module M2.2: taking the simulated radial sampling k-space data and the coil sensitivity map as the input of the deep unfolding neural network to be trained, and taking the reconstructed image output by the deep unfolding neural network as the output of the module M2.2;

[0039] Module M2.3: calculating the loss between the reconstructed image output by the deep unfolding neural network and the interventional image sequence, updating the parameters in the deep unfolding neural network through back propagation according to the loss, the parameters in the deep unfolding neural network including the weights of the convolution kernels in the convolution layer and the regularization parameter, and obtaining the trained network parameters after multiple rounds of updating training, that is, obtaining the trained deep unfolding neural network model.

[0040] Preferably, the module M3 comprises the following modules:

[0041] Module M3.1: acquiring a full-sampling reference image before real-time imaging is performed, and using the reference image to estimate the coil sensitivity map;

[0042] Module M3.2: acquiring radial k-space data in real time when real-time interventional imaging is performed, and taking the acquired k-space data and the estimated coil sensitivity map as the input of the trained deep unfolding neural network, and then reconstructing the magnetic resonance interventional image in real time.

[0043] Compared with the prior art, the present application has the following beneficial effects:

[0044] 1. The present application solves the problem of long reconstruction time of iterative algorithms applied to real-time imaging of magnetic resonance by using a method of unfolding the iterative algorithm based on low-rank and sparse decomposition into a deep neural network, and realizes real-time reconstruction of magnetic resonance interventional images.

[0045] 2. The present application solves the problems of fixed sparse transformation, large amount of calculation and long time in the iterative algorithm based on low-rank and sparse decomposition by using the structure composed of convolution layer and activation function to replace the sparse transformation in the iterative algorithm based on low-rank and sparse decomposition.

[0046] 3. The present application solves the problem that the regularization parameter in the iterative algorithm based on low-rank and sparse decomposition needs to be obtained through tedious manual parameter tuning by learning the regularization parameter in the iterative algorithm based on low-rank and sparse decomposition during network training. BRIEF DESCRIPTION OF DRAWINGS

[0047] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0048] Figure 1 The structure diagram of the deep unfolding neural network;

[0049] Figure 2a training flowchart of the depth-unfolded neural network is unfolded in depth;

[0050] Figure 3 a reconstruction flowchart of the depth-unfolded neural network. DETAILED DESCRIPTION

[0051] The application will be described in detail below with specific embodiments. The following embodiments will help those skilled in the art to further understand the application, but do not limit the application in any form. It should be pointed out that those skilled in the art can make several changes and improvements without departing from the concept of the application. These all belong to the protection scope of the application.

[0052] The embodiment of the application discloses a real-time magnetic resonance image reconstruction method based on a depth-unfolded neural network, as shown in Figure 1 Figure 1 The depth-unfolded neural network based on low-rank and sparse decomposition mainly includes steps of building, training and real-time reconstruction of the depth-unfolded neural network, and specifically includes the following steps:

[0053] Step S1: building the depth-unfolded neural network based on low-rank and sparse decomposition. Figure 1 As shown in the structure diagram of the depth-unfolded neural network based on low-rank and sparse decomposition proposed by the application.

[0054] Step S1 includes the following steps: Step S1.1: in magnetic resonance imaging, the slowly changing background and the intervention feature that is always dynamically changing are decomposed into a low-rank matrix L and a sparse matrix S from the intervention image sequence x, that is, x=L+S, by using the spatial sparse characteristics of L and S, the iterative algorithm model for magnetic resonance intervention image reconstruction based on low-rank and sparse decomposition is:

[0055]

[0056] Wherein, E=ΩFC is a k-space encoding operator, C is a coil sensitivity, F is a Fourier transform, and Ω is a sampling mode; d is the acquired k-space data; is a TV (total variation) transform along the S time direction; ψ is a sparse transform, λ L , λ s , and are coefficients of the four terms (four regularization terms) in formula (1), also called regularization parameters. After the determination of these parameters, formula (1) can be optimized and solved; k-space represents the Fourier space. represents the 2-norm, ‖·‖ * ​The kernel norm is represented by ‖·‖1, which represents the 1-norm and are common operations on matrices.

[0057] The primal dual-fixed-point algorithm (PDFP) is used to optimize the iterative algorithm model based on low-rank and sparse decomposition, which requires multiple iterative steps to solve.

[0058] Step S1.2: This invention expands the iterative algorithm based on low-rank and sparse decomposition into a deep neural network. The deep expanded neural network consists of multiple iterative blocks, each corresponding to one iteration in the iterative algorithm based on low-rank and sparse decomposition. The sparse transformation ψ in the iterative algorithm based on low-rank and sparse decomposition is replaced by a structure composed of multiple convolutional layers and activation functions. The weight parameters of the convolutional kernels in the convolutional layers can be learned during the training of the deep expanded neural network. The activation function can improve the network's expressive power. The regularization parameter λ in the iterative algorithm based on low-rank and sparse decomposition... L , λ s ,、 as well as It can also be learned during the training of deep unfolded neural networks.

[0059] Step S1.3: The first input of the first iteration block is the parameters after network initialization. The first input of the second and subsequent iteration blocks is the output of the previous iteration block. The second input of each iteration block includes the k-space data obtained by radial sampling and the coil sensitivity map. Finally, after passing through all iteration blocks, the reconstructed image is output.

[0060] Sparse transformation is based on ψ in low-rank and sparse decomposition models. Images undergoing sparse transformation can be subject to sparsity constraints. A combination of convolutional layers and activation functions replaces the sparse transformation ψ in low-rank and sparse decomposition models. Convolutional layers extract image features, while activation functions enhance the network's expressive power. The learned regularization parameter λ... L ,λ s , as well as It is used as a parameter in the network for subsequent image reconstruction.

[0061] Step S2: Train the constructed deep unfolded neural network using simulation data. Figure 2 The diagram shows the training process of the deep unfolded neural network proposed in this invention.

[0062] Step S2 includes the following steps: Step S2.1: A series of intervention image sequences are obtained through simulation as a training set. Non-uniform Fourier transform is performed on the intervention image sequences to obtain the simulated radial sampling k-space data; the coil sensitivity map is obtained by dividing each coil image by the synthesized image.

[0063] Step S2.2: Simulated radial sampling k-space data and coil sensitivityFigure 1 As an input of the deep unfolding neural network to be trained, the deep unfolding neural network outputs a reconstructed image.

[0064] Step S2.3: the deep unfolding neural network outputs a reconstructed image and calculates a loss with the interventional image sequence, and updates parameters in the deep unfolding neural network, including weights of convolution kernels in the convolution layer and a regularization parameter, through a back propagation manner according to the loss; after multiple rounds of updating training, the best network parameters are obtained, that is, a trained deep unfolding neural network model is obtained.

[0065] Step S3: using the trained deep unfolding neural network to perform real-time image reconstruction. Figure 3 The reconstruction process of the deep unfolding neural network proposed in the application is shown.

[0066] Step S3 includes the following steps: Step S3.1: before real-time imaging is performed, a full-sampling reference image is collected, which is used to estimate a coil sensitivity map. The coil sensitivity map is obtained by dividing each coil image by the synthesized image.

[0067] Step S3.2: when real-time interventional imaging is performed, radial k-space data is collected in real time, and the collected k-space data and the estimated coil sensitivity map are used as inputs of the trained deep unfolding neural network, and then a magnetic resonance interventional image is reconstructed in real time.

[0068] The embodiment of the application also discloses a real-time magnetic resonance image reconstruction system based on a deep unfolding neural network, which comprises the following modules:

[0069] Module M1: a deep unfolding neural network based on low-rank and sparse decomposition is built.

[0070] Module M1 includes the following modules: Module M1.1: in magnetic resonance imaging, a changing background and a continuously dynamically changing interventional feature are decomposed into a low-rank matrix L and a sparse matrix S from an interventional image sequence x, that is, x=L+S, and by using the spatial sparse characteristics of L and S, a model for magnetic resonance interventional image reconstruction based on low-rank and sparse decomposition is used.

[0071] The model for magnetic resonance interventional image reconstruction based on low-rank and sparse decomposition is as follows:

[0072]

[0073] Wherein, E=ΩFC is a k-space encoding operator, C is a coil sensitivity, F is a Fourier transform, and Ω is a sampling mode; d is acquired k-space data; is a TV transform along the S time direction; ψ is a sparse transform, λ L , λ s , and is a regularization parameter; k-space represents Fourier space.

[0074] Module M1.2: Expanding the model based on low-rank and sparse decomposition into a deep neural network; the deep expanding neural network is composed of multiple iterative blocks, each of which corresponds to an iteration in the iterative algorithm based on low-rank and sparse decomposition, the sparse transformation ψ in the iterative algorithm based on low-rank and sparse decomposition is replaced by a structure composed of multiple convolutional layers and activation functions, the weight parameters of the convolution kernel in the convolutional layer are learned during the training of the deep expanding neural network, and the regularization parameter in the iterative algorithm based on low-rank and sparse decomposition is learned during the training of the deep expanding neural network.

[0075] Module M1.3: The first input of the first iterative block is the network initialized parameter, the first input of the second and subsequent iterative blocks is the output of the previous iterative block, and the second input of each iterative block includes the k-space data obtained by radial sampling and the coil sensitivity map, and finally the output of all iterative blocks is the reconstructed image.

[0076] Module M2: Training the deep expanding neural network built by using simulation data.

[0077] Module M2 includes the following modules: Module M2.1: Obtain the intervention image sequence by simulation as the training set, and perform non-uniform Fourier transform on the intervention image sequence to obtain the simulated radial sampling k-space data.

[0078] Module M2.2: Take the simulated radial sampling k-space data and the coil sensitivity map as the input of the deep expanding neural network to be trained, and the deep expanding neural network outputs the reconstructed image.

[0079] Module M2.3: Calculate the loss between the reconstructed image output by the deep expanding neural network and the intervention image sequence, update the parameters in the deep expanding neural network by back propagation according to the loss, the parameters in the deep expanding neural network include the weight of the convolution kernel in the convolutional layer and the regularization parameter; after multiple rounds of training, the trained network parameters are obtained, that is, the trained deep expanding neural network model is obtained.

[0080] Module M3: Real-time image reconstruction using the trained deep expanding neural network.

[0081] Module M3 includes the following modules: Module M3.1: Before real-time imaging, a full-sampled reference image is collected for estimating the coil sensitivity map.

[0082] Module M3.2: During real-time interventional imaging, radial k-space data is collected in real time, and the collected k-space data and the estimated coil sensitivity map are taken as the input of the trained deep expanding neural network, and then the magnetic resonance interventional image is reconstructed in real time.

[0083] To realize real-time magnetic resonance interventional imaging, an iterative algorithm based on low rank and sparse decomposition is expanded into a deep neural network to realize real-time reconstruction of magnetic resonance interventional images.

[0084] Those skilled in the art know that, in addition to implementing the system and each device, module and unit thereof provided by the present application in the form of pure computer readable program code, the same functions can also be realized by logically programming the method steps in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers and embedded microcontrollers. Therefore, the system and each device, module and unit thereof provided by the present application can be considered as a hardware component, and the devices, modules and units included therein for realizing various functions can also be considered as structures within the hardware component; the devices, modules and units for realizing various functions can also be considered as both software modules realizing methods and structures within the hardware component.

[0085] The specific embodiments of the present application are described above. It needs to be understood that the present application is not limited to the above specific embodiments, and various changes or modifications can be made by those skilled in the art within the scope of the claims, which does not affect the essential content of the present application. The embodiments of the present application and the features in the embodiments can be arbitrarily combined with each other without conflict.

Claims

1. A real-time magnetic resonance image reconstruction method based on a deep unfolded neural network, characterized in that, Includes the following steps: Step S1: Construct a deep unfolded neural network based on low-rank and sparse decomposition; Step S2: Train the constructed deep unfolded neural network using simulation data; Step S3: Real-time image reconstruction using a trained deep unfolded neural network; Step S1 includes the following steps: Step S1.1: In magnetic resonance imaging, from the interventional image sequence x, the changing background and continuously dynamically changing interventional features are decomposed into a low-rank matrix L and a sparse matrix S, i.e. x = L + S. By utilizing the spatial sparsity properties of L and S, a model based on low-rank and sparse decomposition is used for magnetic resonance interventional image reconstruction. Step S1.2: Unfold the model based on low-rank and sparse decomposition into a deep neural network; the deep unfolded neural network consists of multiple iterative blocks, each iterative block corresponding to one iteration in the iterative algorithm based on low-rank and sparse decomposition. The sparse transformation ψ in the iterative algorithm based on low-rank and sparse decomposition is replaced by a structure composed of multiple convolutional layers and activation functions. The weight parameters of the convolutional kernels in the convolutional layers are learned during the training of the deep unfolded neural network, and the regularization parameters in the iterative algorithm based on low-rank and sparse decomposition are learned during the training of the deep unfolded neural network. Step S1.3: The first input of the first iteration block is the parameters after network initialization. The first input of the second and subsequent iteration blocks is the output of the previous iteration block. The second input of each iteration block includes the k-space data obtained by radial sampling and the coil sensitivity map. Finally, after passing through all iteration blocks, the reconstructed image is output.

2. The real-time magnetic resonance image reconstruction method based on a deep unfolded neural network according to claim 1, characterized in that, In step S1.1, the model based on low-rank and sparse decomposition for magnetic resonance imaging reconstruction is as follows: Where E = ΩFC is the k-space encoding operator, C is the coil sensitivity, F is the Fourier transform, Ω is the sampling method, and d is the acquired k-space data; It is the TV transform along the time direction S; ψ is the sparse transform, λ L , λ s , as well as It is the regularization parameter; k-space represents the Fourier space.

3. The real-time magnetic resonance image reconstruction method based on a deep unfolded neural network according to claim 1, characterized in that, Step S2 includes the following steps: Step S2.1: Use the interventional image sequence obtained through simulation as the training set, and perform a non-uniform Fourier transform on the interventional image sequence to obtain the simulated radial sampling k-space data; Step S2.2: Use the simulated radial sampling k-space data and coil sensitivity map as inputs to the deep unfolded neural network to be trained, and output the reconstructed image from the deep unfolded neural network; Step S2.3: Calculate the loss between the reconstructed image output by the deep unfolded neural network and the intervention image sequence. Update the parameters in the deep unfolded neural network through backpropagation based on the loss. The parameters in the deep unfolded neural network include the weights of the convolutional kernels in the convolutional layers and the regularization parameters. After multiple rounds of training and updating, the trained network parameters are obtained, which is the trained deep unfolded neural network model.

4. The real-time magnetic resonance image reconstruction method based on a deep unfolded neural network according to claim 1, characterized in that, Step S3 includes the following steps: Step S3.1: Before performing real-time imaging, acquire a fully sampled reference image to estimate the coil sensitivity map; Step S3.2: During real-time interventional imaging, radial k-space data is acquired in real time. The acquired k-space data and the estimated coil sensitivity map are used as inputs to a pre-trained deep unfolded neural network, thereby reconstructing the magnetic resonance interventional image in real time.

5. A real-time magnetic resonance image reconstruction system based on a deep unfolded neural network, characterized in that, Includes the following modules: Module M1: Construct a deep unfolded neural network based on low-rank and sparse decomposition; Module M2: A deep unfolded neural network trained using simulated data; Module M3: Real-time image reconstruction using a trained deep unfolded neural network; Module M1 includes the following modules: Module M1.1: In magnetic resonance imaging, the changing background and continuously dynamically changing interventional features in the interventional image sequence x are decomposed into a low-rank matrix L and a sparse matrix S, i.e., x = L + S. By utilizing the spatial sparsity properties of L and S, a model based on low-rank and sparse decomposition is used for magnetic resonance interventional image reconstruction. Module M1.2: Unfolds the low-rank and sparse decomposition-based model into a deep neural network; the deep unfolded neural network consists of multiple iterative blocks, each corresponding to one iteration in the iterative algorithm based on low-rank and sparse decomposition. The sparse transformation ψ in the iterative algorithm based on low-rank and sparse decomposition is replaced by a structure composed of multiple convolutional layers and activation functions. The weight parameters of the convolutional kernels in the convolutional layers are learned during the training of the deep unfolded neural network, and the regularization parameters in the iterative algorithm based on low-rank and sparse decomposition are learned during the training of the deep unfolded neural network. Module M1.3: The first input of the first iteration block is the parameters after network initialization. The first input of the second and subsequent iteration blocks is the output of the previous iteration block. The second input of each iteration block includes the k-space data obtained by radial sampling and the coil sensitivity map. Finally, after passing through all iteration blocks, the reconstructed image is output.

6. The real-time magnetic resonance image reconstruction system based on a deep unfolded neural network according to claim 5, characterized in that, In module M1.1, the model based on low-rank and sparse decomposition for magnetic resonance imaging reconstruction is as follows: Where E = ΩFC is the k-space encoding operator, C is the coil sensitivity, F is the Fourier transform, Ω is the sampling method, and d is the acquired k-space data; It is the TV transform along the time direction S; ψ is the sparse transform, λ L , λ s , as well as It is the regularization parameter; k-space represents the Fourier space.

7. The real-time magnetic resonance image reconstruction system based on a deep unfolded neural network according to claim 5, characterized in that, Module M2 includes the following modules: Module M2.1: The intervention image sequence obtained through simulation is used as the training set. Non-uniform Fourier transform is performed on the intervention image sequence to obtain the simulated radial sampling k-space data. Module M2.2: The simulated radial sampling k-space data and coil sensitivity map are used as inputs to the deep unfolded neural network to be trained, and the deep unfolded neural network outputs the reconstructed image; Module M2.3: The loss is calculated between the reconstructed image and the intervention image sequence output by the deep unfolded neural network. The parameters in the deep unfolded neural network are updated through backpropagation based on the loss. The parameters in the deep unfolded neural network include the weights of the convolutional kernels in the convolutional layers and the regularization parameters. After multiple rounds of training and updating, the trained network parameters are obtained, which is the trained deep unfolded neural network model.

8. The real-time magnetic resonance image reconstruction system based on a deep unfolded neural network according to claim 5, characterized in that, Module M3 includes the following modules: Module M3.1: Acquires a fully sampled reference image before real-time imaging to estimate the coil sensitivity map; Module M3.2: During real-time interventional imaging, radial k-space data is acquired in real time. The acquired k-space data and the estimated coil sensitivity map are used as inputs to a pre-trained deep unfolded neural network, thereby reconstructing the magnetic resonance interventional image in real time.

Images