MRI (Magnetic Resonance Imaging) reconstruction method and system based on double-domain deep expansion network model and medium

Through the dual-domain deep expansion network model, the regularization term gradients of the K-space domain and image domain are processed by using Mamba1 and Mamba2 networks respectively, solving the problem of low MRI reconstruction quality in the prior art and achieving high-quality images Reconstruction and aliasing artifacts are reduced.

CN118941665APending Publication Date: 2024-11-12GUANGDONG UNIV OF TECH
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410990443.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The existing MRI reconstruction method cannot effectively utilize the context information of the K-space domain and image domain, resulting in low image reconstruction quality, especially under sampling conditions below the Nyquist rate, which is prone to aliasing artifacts and sparseness transformation domains. Difficult to choose.

Method used

Using a dual-domain deep expansion network model, the regularization term gradients of the K-space domain and image domain are processed by Mamba1 and Mamba2 network models, a dual-domain deep expansion network model is constructed, and iterative optimization is used to achieve high quality MRI reconstruction.

Benefits of technology

Through the dual-domain deep expansion network model, the K-space domain and image domain information are fully utilized, the quality of MRI reconstruction is improved, the aliasing artifact is reduced, the generalization ability of the sparse transformation domain is enhanced, and the regularization parameter adjustment is simplified.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941665A_ABST
    Figure CN118941665A_ABST
Patent Text Reader

Abstract

The invention discloses an MRI (Magnetic Resonance Imaging) reconstruction method and system based on a double-domain deep expansion network model and a medium, and the method comprises the steps: obtaining simulation undersampling data which comprises K space data; a double-domain deep expansion network model is constructed based on a preset model framework, and the model framework comprises Mamba1 and Mamba2; constraining the double-domain depth expansion model based on a preset compressed sensing algorithm, wherein the constraining comprises constructing a K space and constraining MRI by an image data regularization item; and inputting the simulated under-sampling data into the double-domain depth expansion model to obtain reconstructed K space data, and obtaining a reconstructed image by adopting inverse Fourier transform. According to the method, the gradient of the K space domain regularization item and the gradient of the image domain regularization item are approximately replaced by two mutually independent network branches and are processed step by step, and the information of the K space domain and the image domain is fully utilized, so that the high-quality reconstruction effect is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of data processing and image processing, and more specifically, to an MRI reconstruction method, system and medium based on a dual-domain deep expansion network model. Background Art

[0002] Magnetic resonance imaging (MRI) is a widely used medical imaging technique for diagnosis and research. However, MRI requires a long scan time, which may cause discomfort to patients and increase the possibility of motion-related artifacts in the reconstructed image. One way to reduce the scan time of MRI is to acquire undersampled K-space data instead of acquiring full K-space data. However, sampling below the Nyquist rate in the K-space domain can lead to aliasing artifacts in the spatial domain, resulting in reduced image quality. In order to reduce the acquisition time of K-space data and improve the quality of images, researchers have developed many image reconstruction methods. Traditional methods are generally based on compressed sensing (CS) technology, which achieves accelerated imaging by acquiring a small amount of data and reconstructing using signal sparsity. However, CS methods still face some challenges and shortcomings in practical applications, especially in the selection and application of sparse transform domains; for example, different tissues and lesions may not be sparse in a specific transform domain; the generalization ability of fixed transform domains is limited; the adjustment of regularization parameters is complex, etc. In this context, deep learning technology has gradually become a research hotspot in the field of MRI image reconstruction.

[0003] Existing MRI reconstruction methods cannot explore contextual information in the K-space domain, making it difficult to use low-frequency information to guide the reconstruction of high-frequency information, or to use high-frequency information to guide the reconstruction of low-frequency information. At the same time, MRI reconstruction methods cannot explore image contextual information, making it difficult to use information from different parts of the image for image reconstruction, resulting in poor results. Effective technical solutions are urgently needed to address the above problems. Summary of the invention

[0004] The purpose of the present invention is to provide an MRI reconstruction method, system and medium based on a dual-domain deep expansion network model, which fully utilizes the information of both the K-space domain and the image domain by approximately replacing the gradient of the K-space domain regularization term and the image domain regularization gradient with two independent network branches and processing them step by step, thereby achieving high-quality reconstruction effects.

[0005] The first aspect of the present invention provides an MRI reconstruction method based on a dual-domain deep unfolding network model, comprising the following steps:

[0006] Acquire simulated under-sampling data, wherein the simulated under-sampling data includes K-space data;

[0007] A dual-domain deep unfolding network model is constructed based on a preset model framework, wherein the model framework includes Mamba1 and Mamba2;

[0008] Constraining the dual-domain depth expansion model based on a preset compressed sensing algorithm, including constructing K-space and image data regularization terms to constrain the MRI;

[0009] The simulated under-sampling data is input into the dual-domain depth expansion model to obtain reconstructed K-space data, and an inverse Fourier transform is used to obtain a reconstructed image.

[0010] In this solution, the obtaining of simulated under-sampling data specifically includes:

[0011] Acquire full-sampled K-space data;

[0012] Randomly generate a mask matrix with the same size as the K-space data, wherein the value of the mask matrix is ​​0 or 1;

[0013] The K-space data is undersampled by using a randomly generated mask to obtain the simulated undersampled data.

[0014] In this solution, the dual-domain deep expansion network model is constructed based on a preset model framework, specifically including:

[0015] Based on the Mamba1 and Mamba2 network models, we obtain the shallow feature extraction module, deep feature extraction module and reconstruction module;

[0016] The data input into the model first flows to the shallow feature extraction module to obtain a shallow feature vector, wherein the feature extraction module is mainly composed of a convolution layer and a normalization layer connected in series;

[0017] Flattening the shallow feature vectors through a preset scanning method to obtain a target long sequence, wherein the first scanning method is used to process image domain data, and the second scanning method is used to process K-space domain data;

[0018] Inputting the target long sequence into a deep feature extraction module to obtain a deep feature vector;

[0019] The deep feature vector is input into a reconstruction module to obtain prior information of the input data, wherein the reconstruction module is composed of a convolutional layer.

[0020] In this solution, the dual-domain depth expansion model is constrained based on a preset compressed sensing algorithm, specifically including:

[0021] Obtain the compressed sensing model, analyze the compressed sensing model, and construct the K-space and image data regularization terms to constrain MRI reconstruction. The compressed sensing model is defined as:

[0022]

[0023] Wherein, A represents the forward operator, i.e., A=MF, F represents Fourier transform, M represents the downsampling mask; x represents the reconstructed MR image; y represents the undersampled K-space data; k represents the reconstructed K-space data; is a quadratic term used to measure the error between x after transformation A and the target y; R x (x) represents a regularization term used to constrain image data; R k (k) represents a regularization term used to constrain the K-space data, λ x and λ k Respectively represent R x (x) and R k The weight coefficient of (k) is a learnable parameter.

[0024] In this solution, the simulated under-sampling data is input into the dual-domain depth expansion model to obtain reconstructed K-space data, and an inverse Fourier transform is used to obtain a reconstructed image, which specifically includes:

[0025] The gradient of the K-space data regularization term is replaced by the Mamba1 network model to build the K-space domain branch of the dual-domain deep expansion network model;

[0026] The Mamba2 network model is used to replace the gradient of the image data regularization term to build the image domain branch of the dual-domain deep expansion network model;

[0027] A dual-domain deep unfolding network model is constructed based on the K-space domain branch and the image domain branch;

[0028] The reconstructed K-space data is obtained by iterating a preset number of times, and the reconstructed MR image is obtained by inverse Fourier transform.

[0029] In this solution, the method further includes calculating the loss between the reconstructed K-space data and the real K-space data and the reconstructed image and the real image and updating the network model parameters through back propagation, wherein the loss function of the dual-domain deep unfolding network model is constructed, specifically including:

[0030] The K-space loss function is defined as:

[0031]

[0032] The pixel loss function is defined as follows:

[0033]

[0034] The total loss function can be expressed as

[0035] Loss total =αLoss freq +βLoss pixel ;

[0036] Among them, θ is the parameter of the model, S represents the size of K space, that is, the number of pixels, and y i represents the i-th data value in the full-sampled K-space, represents the i-th data value of the reconstructed K-space data, F -1 represents the inverse Fourier transform, ∈ is a constant, α and β are Loss freq and Loss pixel The weight coefficient of .

[0037] The second aspect of the present invention further provides an MRI reconstruction system based on a dual-domain deep expansion network model, comprising a memory and a processor, wherein the memory comprises an MRI reconstruction method program based on a dual-domain deep expansion network model, and when the MRI reconstruction method program based on a dual-domain deep expansion network model is executed by the processor, the following steps are implemented:

[0038] Acquire simulated under-sampling data, wherein the simulated under-sampling data includes K-space data;

[0039] A dual-domain deep unfolding network model is constructed based on a preset model framework, wherein the model framework includes Mamba1 and Mamba2;

[0040] Constraining the dual-domain depth expansion model based on a preset compressed sensing algorithm, including constructing K-space and image data regularization terms to constrain the MRI;

[0041] The simulated under-sampling data is input into the dual-domain depth expansion model to obtain reconstructed K-space data, and an inverse Fourier transform is used to obtain a reconstructed image.

[0042] In this solution, the obtaining of simulated under-sampling data specifically includes:

[0043] Acquire full-sampled K-space data;

[0044] Randomly generate a mask matrix with the same size as the K-space data, wherein the value of the mask matrix is ​​0 or 1;

[0045] The K-space data is undersampled by using a randomly generated mask to obtain the simulated undersampled data.

[0046] In this solution, the dual-domain deep expansion network model is constructed based on a preset model framework, specifically including:

[0047] Based on the Mamba1 and Mamba2 network models, we obtain the shallow feature extraction module, deep feature extraction module and reconstruction module;

[0048] The data input into the model first flows to the shallow feature extraction module to obtain a shallow feature vector, wherein the feature extraction module is mainly composed of a convolution layer and a normalization layer connected in series;

[0049] Flattening the shallow feature vectors through a preset scanning method to obtain a target long sequence, wherein the first scanning method is used to process image domain data, and the second scanning method is used to process K-space domain data;

[0050] Inputting the target long sequence into a deep feature extraction module to obtain a deep feature vector;

[0051] The deep feature vector is input into a reconstruction module to obtain prior information of the input data, wherein the reconstruction module is composed of a convolutional layer.

[0052] In this solution, the dual-domain depth expansion model is constrained based on a preset compressed sensing algorithm, specifically including:

[0053] Obtain the compressed sensing model, analyze the compressed sensing model, and construct the K-space and image data regularization terms to constrain MRI reconstruction. The compressed sensing model is defined as:

[0054]

[0055] Wherein, A represents the forward operator, i.e., A=MF, F represents Fourier transform, M represents the downsampling mask; x represents the reconstructed MR image; y represents the undersampled K-space data; k represents the reconstructed K-space data; is a quadratic term used to measure the error between x after transformation A and the target y; R x (x) represents a regularization term used to constrain image data; R k (k) represents a regularization term used to constrain the K-space data, λ x and λ k Respectively represent R x (x) and R k The weight coefficient of (k) is a learnable parameter.

[0056] In this solution, the simulated under-sampling data is input into the dual-domain depth expansion model to obtain reconstructed K-space data, and an inverse Fourier transform is used to obtain a reconstructed image, which specifically includes:

[0057] The gradient of the K-space data regularization term is replaced by the Mamba1 network model to build the K-space domain branch of the dual-domain deep expansion network model;

[0058] The Mamba2 network model is used to replace the gradient of the image data regularization term to build the image domain branch of the dual-domain deep expansion network model;

[0059] A dual-domain deep unfolding network model is constructed based on the K-space domain branch and the image domain branch;

[0060] The reconstructed K-space data is obtained by iterating a preset number of times, and the reconstructed MR image is obtained by inverse Fourier transform.

[0061] In this solution, the method further includes calculating the loss between the reconstructed K-space data and the real K-space data and the reconstructed image and the real image and updating the network model parameters through back propagation, wherein the loss function of the dual-domain deep unfolding network model is constructed, specifically including:

[0062] The K-space loss function is defined as:

[0063]

[0064] The pixel loss function is defined as follows:

[0065]

[0066] The total loss function can be expressed as

[0067] Loss total =αLoss freq +βLoss pixel ;

[0068] Among them, θ is the parameter of the model, S represents the size of K space, that is, the number of pixels, and y i represents the i-th data value in the full-sampled K-space, represents the i-th data value of the reconstructed K-space data, F -1 represents the inverse Fourier transform, ∈ is a constant, α and β are Loss freq and Loss pixel The weight coefficient of .

[0069] The third aspect of the present invention provides a computer-readable storage medium, which includes a machine-readable MRI reconstruction method program based on a dual-domain deep expansion network model. When the MRI reconstruction method program based on the dual-domain deep expansion network model is executed by a processor, the steps of an MRI reconstruction method based on a dual-domain deep expansion network model as described in any one of the above items are implemented.

[0070] The present invention discloses an MRI reconstruction method, system and medium based on a dual-domain deep expansion network model. The gradient of the K-space domain regularization term and the image domain regularization gradient are approximately replaced by two independent network branches and processed step by step, thereby fully utilizing the information of both the K-space domain and the image domain, thereby achieving high-quality reconstruction effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A flowchart of an MRI reconstruction method based on a dual-domain deep unfolding network model of the present invention is shown;

[0072] Figure 2 A schematic diagram of a Mamba network model based on a dual-domain deep expansion network model of the present invention is shown;

[0073] Figure 3 A schematic diagram of a dual-domain deep expansion network model based on the dual-domain deep expansion network model of the present invention is shown;

[0074] Figure 4 A block diagram of an MRI reconstruction system based on a dual-domain deep expansion network model of the present invention is shown. DETAILED DESCRIPTION

[0075] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0076] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0077] Figure 1 A flowchart of an MRI reconstruction method based on a dual-domain deep expansion network model of the present application is shown.

[0078] like Figure 1 As shown, the present application discloses an MRI reconstruction method based on a dual-domain deep unfolding network model, comprising the following steps:

[0079] S102, acquiring simulated under-sampling data, where the simulated under-sampling data includes K-space data;

[0080] S104, constructing a dual-domain deep unfolding network model based on a preset model framework, wherein the model framework includes Mamba1 and Mamba2;

[0081] S106, constraining the dual-domain depth expansion model based on a preset compressed sensing algorithm, including constructing K-space and image data regularization terms to constrain MRI;

[0082] S108, inputting the simulated under-sampling data into the dual-domain depth expansion model to obtain reconstructed K-space data, and using inverse Fourier transform to obtain a reconstructed image.

[0083] It should be noted that, in this embodiment, the MRI reconstruction model function of the dual-domain deep expansion network model is defined as:

[0084] x i =recon(k i );

[0085] Among them, k i ∈R N is the under-sampled K-space data, which is the input of the model; i ∈R N is the reconstructed image; recon(·) is the dual-domain deep expansion network model that maps the undersampled K-space data to the reconstructed image.

[0086] Specifically, firstly, simulated under-sampling data is acquired, wherein the simulated under-sampling data includes K-space data, and then a dual-domain deep expansion network model is constructed based on a preset model framework, wherein the model framework includes Mamba1 and Mamba2, and the dual-domain deep expansion model is constrained based on a preset compressed sensing algorithm, including constructing K-space and image data regularization terms to constrain MRI, and finally, the simulated under-sampling data is input into the dual-domain deep expansion model to obtain reconstructed K-space data, and an inverse Fourier transform is used to obtain a reconstructed image.

[0087] According to an embodiment of the present invention, obtaining simulated under-sampling data specifically includes:

[0088] Acquire full-sampled K-space data;

[0089] Randomly generate a mask matrix with the same size as the K-space data, wherein the value of the mask matrix is ​​0 or 1;

[0090] The K-space data is undersampled by using a randomly generated mask to obtain the simulated undersampled data.

[0091] It should be noted that, in this embodiment, the frequency domain data in the K space is randomly downsampled at a sampling rate of 10%, and the missing data is filled with zeros to obtain the frequency domain data in the simulated undersampled K space, thereby obtaining the simulated undersampled data.

[0092] According to an embodiment of the present invention, the dual-domain deep expansion network model is constructed based on a preset model framework, specifically including:

[0093] Based on the Mamba1 and Mamba2 network models, we obtain the shallow feature extraction module, deep feature extraction module and reconstruction module;

[0094] The data input into the model first flows to the shallow feature extraction module to obtain a shallow feature vector, wherein the feature extraction module is mainly composed of a convolution layer and a normalization layer connected in series;

[0095] Flattening the shallow feature vectors through a preset scanning method to obtain a target long sequence, wherein the first scanning method is used to process image domain data, and the second scanning method is used to process K-space domain data;

[0096] Inputting the target long sequence into a deep feature extraction module to obtain a deep feature vector;

[0097] The deep feature vector is input into a reconstruction module to obtain prior information of the input data, wherein the reconstruction module is composed of a convolutional layer.

[0098] It should be noted that, in this embodiment, specifically, Figure 2 As shown in the figure, the Mamba1 network model and the Mamba2 network model of the MRI reconstruction method based on the dual-domain deep expansion network model, the Mamba1 network is K-space frequency domain data, in order to ensure the symmetry of the Fourier domain data, only the upper half is selected when scanning its features, the scanning method is characterized by obtaining a sequence from low frequency to high frequency and a sequence data block from high frequency to low frequency, wherein the Mamba1 and Mamba2 network models are mainly composed of three modules: shallow feature extraction module, deep feature extraction module and reconstruction module; the input data first flows to the shallow feature extraction module to obtain the shallow The feature vector, the feature extraction module is mainly composed of a convolution layer (Conv) and a normalization layer (Norm-Layer) connected in series; the shallow feature vector is flattened into a small-sized long sequence through a special scanning method. There are mainly two scanning methods. The first scanning method is used to process image domain data. This scanning method is from left to right and from top to bottom, which is a Z-shaped scan; the second scanning method is used to process K-space domain data. In order to ensure the symmetry of Fourier domain data, this scanning method only scans half of the symmetrical data, according to the scanning strategy from low frequency to high frequency and from high frequency to low frequency.

[0099] Furthermore, the shallow feature vector is flattened by one of the two scanning methods mentioned above as input, and then input into the deep feature extraction module to obtain the deep feature vector. The deep feature extraction module is mainly connected by the depth separable convolution (DWConv) -> SiLU activation function -> two-dimensional state space model (2D-SSM) -> normalization layer (Norm-Layer);

[0100] The deep feature vector is input into the reconstruction module to obtain the prior information of the input data. The reconstruction module consists of a convolutional layer.

[0101] According to an embodiment of the present invention, constraining the dual-domain depth expansion model based on a preset compressed sensing algorithm specifically includes:

[0102] Obtain the compressed sensing model, analyze the compressed sensing model, and construct the K-space and image data regularization terms to constrain MRI reconstruction. The compressed sensing model is defined as:

[0103]

[0104] Wherein, A represents the forward operator, i.e., A=MF, F represents Fourier transform, M represents the downsampling mask; x represents the reconstructed MR image; y represents the undersampled K-space data; k represents the reconstructed K-space data; is a quadratic term used to measure the error between x after transformation A and the target y; R x (x) represents a regularization term used to constrain image data; R k (k) represents a regularization term used to constrain the K-space data, λ x and λ k Respectively represent R x (x) and R k The weight coefficient of (k) is a learnable parameter.

[0105] It should be noted that, in this embodiment, R k (k) and R x (x) The regularization term is used to introduce certain constraints, such as sparsity and image prior knowledge. Specifically, a compressed sensing model is obtained, the compressed sensing model is analyzed, and K-space and image data regularization terms are constructed to constrain the MRI reconstruction optimization problem. The compressed sensing model is defined as: λ x R x (x)+λ k R k (k), where A represents the forward operator, i.e., A=MF, F represents Fourier transform, M represents the downsampling mask; x represents the reconstructed MR image; y represents the undersampled K-space data; k represents the reconstructed K-space data; is a quadratic term used to measure the error between x after transformation A and the target y; R x (x) represents a regularization term used to constrain image data; R k (k) is a regularization term used to constrain the K-space data; λ x and λ k Respectively represent R x (x) and R k The weight coefficient of (k) is a learnable parameter.

[0106] According to an embodiment of the present invention, the step of inputting the simulated under-sampling data into the dual-domain depth expansion model to obtain reconstructed K-space data, and using inverse Fourier transform to obtain a reconstructed image specifically includes:

[0107] The gradient of the K-space data regularization term is replaced by the Mamba1 network model to build the K-space domain branch of the dual-domain deep expansion network model;

[0108] The Mamba2 network model is used to replace the gradient of the image data regularization term to build the image domain branch of the dual-domain deep expansion network model;

[0109] A dual-domain deep unfolding network model is constructed based on the K-space domain branch and the image domain branch;

[0110] The reconstructed K-space data is obtained by iterating a preset number of times, and the reconstructed MR image is obtained by inverse Fourier transform.

[0111] It should be noted that, in this embodiment, specifically, Figure 3 As shown, it is a dual-domain deep expansion network model of the MRI reconstruction method based on the dual-domain deep expansion network model, in which a K-space domain branch module is constructed and the Mamba1 network model is used to approximate and replace Data consistency module is -α t M(k (t) -y); According to the K-space domain branch module, data consistency module and under-input module, the K-space domain branch of the dual-domain deep expansion network model is constructed; the image domain branch module is constructed, and the Mamba2 network model is used to approximately replace α t λ The data consistency module is A * (Ax (t)-y); according to the image domain branch module, the data consistency module and the under-input module, build the image domain branch of the dual-domain deep expansion network model; according to the K-space domain branch and the image domain branch, build the dual-domain deep expansion network model; iterate t times to obtain the reconstructed K-space data, and obtain the reconstructed MR image through inverse Fourier transform; calculate the loss between the reconstructed K-space data and the real K-space data and the reconstructed image and the real image, and update the network model parameters through back propagation.

[0112] Furthermore, using the gradient descent algorithm to solve, the optimized iterative algorithm can be obtained as follows:

[0113]

[0114] k (t) represents the K-space data restored by the t-th iteration, α t represents the iteration step, M represents the random mask, k (t) -y represents the measurement error of the current K-space data; represents the gradient of the regularization term of the K-space data; x (t) represents the MR image restored by the tth iteration, A * represents the conjugate transpose of operator A, that is, A * =F -1 M, Ax (t) -y represents the measurement error of the current reconstructed image data, represents the gradient of the regularization term for the image data.

[0115] According to an embodiment of the present invention, the method further includes calculating the loss between the reconstructed K-space data and the real K-space data and the reconstructed image and the real image and updating the network model parameters through back propagation, wherein the loss function of the dual-domain deep unfolding network model is constructed, specifically including:

[0116] The K-space loss function is defined as:

[0117]

[0118] The pixel loss function is defined as follows:

[0119]

[0120] The total loss function can be expressed as

[0121] Loss total =αLoss freq +βLoss pixel ;

[0122] Among them, θ is the parameter of the model, S represents the size of K space, that is, the number of pixels, and y irepresents the i-th data value in the full-sampled K-space, represents the i-th data value of the reconstructed K-space data, F -1 represents the inverse Fourier transform, ∈ is a constant, α and β are Loss freq and Loss pixel The weight coefficient of .

[0123] It should be noted that in this embodiment, the loss function Loss is designed for the MRI reconstruction model network. freq +Loss pixel The undersampled K-space data and the undersampled images obtained by inverse Fourier transform of the K-space data are input into the dual-domain deep expansion network model. The Mamba1 network is used to learn the sparsity of the frequency domain data in the K-space, and the Mamba2 network is used to learn the prior knowledge of the image. After multiple rounds of iterations, the reconstructed K-space data is obtained. The reconstructed image obtained by inverse Fourier transform of the reconstructed K-space data is used as a label, and the fully sampled K-space data and the fully sampled MR image are combined with Loss. freq +Loss pixel , the network model parameters are updated through back propagation to obtain the trained dual-domain deep expansion network model.

[0124] Figure 4 A block diagram of an MRI reconstruction system based on a dual-domain deep expansion network model of the present invention is shown.

[0125] like Figure 4 As shown, the present invention discloses an MRI reconstruction system based on a dual-domain deep expansion network model, including a memory and a processor, wherein the memory includes an MRI reconstruction method program based on a dual-domain deep expansion network model, and when the MRI reconstruction method program based on a dual-domain deep expansion network model is executed by the processor, the following steps are implemented:

[0126] Acquire simulated under-sampling data, wherein the simulated under-sampling data includes K-space data;

[0127] A dual-domain deep unfolding network model is constructed based on a preset model framework, wherein the model framework includes Mamba1 and Mamba2;

[0128] Constraining the dual-domain depth expansion model based on a preset compressed sensing algorithm, including constructing K-space and image data regularization terms to constrain MRI;

[0129] The simulated under-sampling data is input into the dual-domain depth expansion model to obtain reconstructed K-space data, and an inverse Fourier transform is used to obtain a reconstructed image.

[0130] It should be noted that, in this embodiment, the MRI reconstruction model function of the dual-domain deep expansion network model is defined as:

[0131] x i =recon(k i );

[0132] Among them, k i ∈R N is the under-sampled K-space data, which is the input of the model; i ∈R N is the reconstructed image; recon(·) is the dual-domain deep expansion network model that maps the undersampled K-space data to the reconstructed image.

[0133] Specifically, firstly, simulated under-sampling data is acquired, wherein the simulated under-sampling data includes K-space data, and then a dual-domain deep expansion network model is constructed based on a preset model framework, wherein the model framework includes Mamba1 and Mamba2, and the dual-domain deep expansion model is constrained based on a preset compressed sensing algorithm, including constructing K-space and image data regularization terms to constrain MRI, and finally, the simulated under-sampling data is input into the dual-domain deep expansion model to obtain reconstructed K-space data, and an inverse Fourier transform is used to obtain a reconstructed image.

[0134] According to an embodiment of the present invention, obtaining simulated under-sampling data specifically includes:

[0135] Acquire full-sampled K-space data;

[0136] Randomly generate a mask matrix with the same size as the K-space data, wherein the value of the mask matrix is ​​0 or 1;

[0137] The K-space data is undersampled by using a randomly generated mask to obtain the simulated undersampled data.

[0138] It should be noted that, in this embodiment, the frequency domain data in the K space is randomly downsampled at a sampling rate of 10%, and the missing data is filled with zeros to obtain the frequency domain data in the simulated undersampled K space, thereby obtaining the simulated undersampled data.

[0139] According to an embodiment of the present invention, the dual-domain deep expansion network model is constructed based on a preset model framework, specifically including:

[0140] Based on the Mamba1 and Mamba2 network models, we obtain the shallow feature extraction module, deep feature extraction module and reconstruction module;

[0141] The data input into the model first flows to the shallow feature extraction module to obtain a shallow feature vector, wherein the feature extraction module is mainly composed of a convolution layer and a normalization layer connected in series;

[0142] Flattening the shallow feature vectors through a preset scanning method to obtain a target long sequence, wherein the first scanning method is used to process image domain data, and the second scanning method is used to process K-space domain data;

[0143] Inputting the target long sequence into a deep feature extraction module to obtain a deep feature vector;

[0144] The deep feature vector is input into a reconstruction module to obtain prior information of the input data, wherein the reconstruction module is composed of a convolutional layer.

[0145] It should be noted that, in this embodiment, specifically, Figure 2 As shown in the figure, the Mamba1 network model and the Mamba2 network model of the MRI reconstruction method based on the dual-domain deep expansion network model, the Mamba1 network is K-space frequency domain data, in order to ensure the symmetry of the Fourier domain data, only the upper half is selected when scanning its features, the scanning method is characterized by obtaining a sequence from low frequency to high frequency and a sequence data block from high frequency to low frequency, wherein the Mamba1 and Mamba2 network models are mainly composed of three modules: shallow feature extraction module, deep feature extraction module and reconstruction module; the input data first flows to the shallow feature extraction module to obtain the shallow The feature vector, the feature extraction module is mainly composed of a convolution layer (Conv) and a normalization layer (Norm-Layer) connected in series; the shallow feature vector is flattened into a small-sized long sequence through a special scanning method. There are mainly two scanning methods. The first scanning method is used to process image domain data. This scanning method is from left to right and from top to bottom, which is a Z-shaped scan; the second scanning method is used to process K-space domain data. In order to ensure the symmetry of Fourier domain data, this scanning method only scans half of the symmetrical data, according to the scanning strategy from low frequency to high frequency and from high frequency to low frequency.

[0146] Furthermore, the shallow feature vector is flattened by one of the two scanning methods mentioned above as input, and then input into the deep feature extraction module to obtain the deep feature vector. The deep feature extraction module is mainly connected by the depth separable convolution (DWConv) -> SiLU activation function -> two-dimensional state space model (2D-SSM) -> normalization layer (NormLayer);

[0147] The deep feature vector is input into the reconstruction module to obtain the prior information of the input data. The reconstruction module consists of a convolutional layer.

[0148] According to an embodiment of the present invention, constraining the dual-domain depth expansion model based on a preset compressed sensing algorithm specifically includes:

[0149] Obtain the compressed sensing model, analyze the compressed sensing model, and construct K-space and image data regularization terms to constrain MRI reconstruction. The compressed sensing model is defined as:

[0150]

[0151] Wherein, A represents the forward operator, i.e., A=MF, F represents Fourier transform, M represents the downsampling mask; x represents the reconstructed MR image; y represents the undersampled K-space data; k represents the reconstructed K-space data; is a quadratic term used to measure the error between x after transformation A and the target y; R x (x) represents a regularization term used to constrain image data; R k (k) represents a regularization term used to constrain the K-space data, λ x and λ k Respectively represent R x (x) and R k The weight coefficient of (k) is a learnable parameter.

[0152] It should be noted that, in this embodiment, R k (k) and R x (x) The regularization term is used to introduce certain constraints, such as sparsity and image prior knowledge. Specifically, a compressed sensing model is obtained, the compressed sensing model is analyzed, and K-space and image data regularization terms are constructed to constrain the MRI reconstruction optimization problem. The compressed sensing model is defined as: Wherein, A represents the forward operator, i.e., A=MF, F represents Fourier transform, M represents the downsampling mask; x represents the reconstructed MR image; y represents the undersampled K-space data; k represents the reconstructed K-space data; is a quadratic term used to measure the error between x after transformation A and the target y; R x (x) represents a regularization term used to constrain image data; R k (k) is a regularization term used to constrain the K-space data; λ x and λ k Respectively represent R x (x) and R k The weight coefficient of (k) is a learnable parameter.

[0153] According to an embodiment of the present invention, the step of inputting the simulated under-sampling data into the dual-domain depth expansion model to obtain reconstructed K-space data, and using inverse Fourier transform to obtain a reconstructed image specifically includes:

[0154] The gradient of the K-space data regularization term is replaced by the Mamba1 network model to build the K-space domain branch of the dual-domain deep expansion network model;

[0155] The Mamba2 network model is used to replace the gradient of the image data regularization term to build the image domain branch of the dual-domain deep expansion network model;

[0156] A dual-domain deep unfolding network model is constructed based on the K-space domain branch and the image domain branch;

[0157] The reconstructed K-space data is obtained by iterating a preset number of times, and the reconstructed MR image is obtained by inverse Fourier transform.

[0158] It should be noted that, in this embodiment, specifically, Figure 3 As shown, it is a dual-domain deep expansion network model of the MRI reconstruction method based on the dual-domain deep expansion network model, in which a K-space domain branch module is constructed and the Mamba1 network model is used to approximate and replace Data consistency module is -α t M(k (t) -y); According to the K-space domain branch module, data consistency module and under-input module, the K-space domain branch of the dual-domain deep expansion network model is constructed; the image domain branch module is constructed, and the Mamba2 network model is used to approximately replace α t λ The data consistency module is A * (Ax (t) -y); according to the image domain branch module, the data consistency module and the under-input module, build the image domain branch of the dual-domain deep expansion network model; according to the K-space domain branch and the image domain branch, build the dual-domain deep expansion network model; iterate t times to obtain the reconstructed K-space data, and obtain the reconstructed MR image through inverse Fourier transform; calculate the loss between the reconstructed K-space data and the real K-space data and the reconstructed image and the real image, and update the network model parameters through back propagation.

[0159] Furthermore, using the gradient descent algorithm to solve, the optimized iterative algorithm can be obtained as follows:

[0160]

[0161] k (t) represents the K-space data restored by the t-th iteration, α t represents the iteration step, M represents the random mask, k (t) -y represents the measurement error of the current K-space data; represents the gradient of the regularization term of the K-space data; x (t)represents the MR image restored by the tth iteration, A * represents the conjugate transpose of operator A, that is, A * =F -1 M, Ax (t) -y represents the measurement error of the current reconstructed image data, represents the gradient of the regularization term for the image data.

[0162] According to an embodiment of the present invention, the method further includes calculating the loss between the reconstructed K-space data and the real K-space data and the reconstructed image and the real image and updating the network model parameters through back propagation, wherein the loss function of the dual-domain deep unfolding network model is constructed, specifically including:

[0163] The K-space loss function is defined as:

[0164]

[0165] The pixel loss function is defined as follows:

[0166]

[0167] The total loss function can be expressed as

[0168] Loss total =αLoss freq +βLoss pixel ;

[0169] Among them, θ is the parameter of the model, S represents the size of K space, that is, the number of pixels, and y i represents the i-th data value in the full-sampled K-space, represents the i-th data value of the reconstructed K-space data, F -1 represents the inverse Fourier transform, ∈ is a constant, α and β are Loss freq and Loss pixel The weight coefficient of .

[0170] It should be noted that in this embodiment, the loss function Loss is designed for the MRI reconstruction model network. freq +Loss pixel The undersampled K-space data and the undersampled images obtained by inverse Fourier transform of the K-space data are input into the dual-domain deep expansion network model. The Mamba1 network is used to learn the sparsity of the frequency domain data in the K-space, and the Mamba2 network is used to learn the prior knowledge of the image. After multiple rounds of iterations, the reconstructed K-space data is obtained. The reconstructed image obtained by inverse Fourier transform of the reconstructed K-space data is used as a label, and the fully sampled K-space data and the fully sampled MR image are combined with Loss. freq +Loss pixel, the network model parameters are updated through back propagation to obtain the trained dual-domain deep expansion network model.

[0171] The third aspect of the present invention provides a computer-readable storage medium, which includes an MRI reconstruction method program based on a dual-domain deep expansion network model. When the MRI reconstruction method program based on a dual-domain deep expansion network model is executed by a processor, the steps of an MRI reconstruction method based on a dual-domain deep expansion network model as described in any one of the above items are implemented.

[0172] The present invention discloses an MRI reconstruction method, system and medium based on a dual-domain deep expansion network model. The gradient of the K-space domain regularization term and the image domain regularization gradient are approximately replaced by two independent network branches and processed step by step, thereby fully utilizing the information of both the K-space domain and the image domain, thereby achieving high-quality reconstruction effects.

[0173] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0174] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0175] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0176] A person of ordinary skill in the art can understand that: all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiments; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks or optical disks, and other media that can store program codes.

[0177] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

Claims

1. An MRI reconstruction method based on a dual-domain deep unfolding network model, characterized in that: The following steps are involved: Acquire simulated under-sampling data, wherein the simulated under-sampling data includes K-space data; A dual-domain deep unfolding network model is constructed based on a preset model framework, wherein the model framework includes Mamba1 and Mamba2; Constraining the dual-domain depth expansion model based on a preset compressed sensing algorithm, including constructing K-space and image data regularization terms to constrain MRI; The simulated under-sampling data is input into the dual-domain depth expansion model to obtain reconstructed K-space data, and an inverse Fourier transform is used to obtain a reconstructed image.

2. The MRI reconstruction method based on the dual-domain deep unfolding network model according to claim 1, characterized in that: The obtaining of simulated under-sampling data specifically includes: Acquire full-sampled K-space data; Randomly generate a mask matrix with the same size as the K-space data; The K-space data is undersampled by using a randomly generated mask to obtain the simulated undersampled data.

3. The MRI reconstruction method based on the dual-domain deep unfolding network model according to claim 2, characterized in that: The dual-domain deep expansion network model is constructed based on a preset model framework, specifically including: Based on the Mamba1 and Mamba2 network models, we obtain the shallow feature extraction module, deep feature extraction module and reconstruction module; The data input into the model first flows to the shallow feature extraction module to obtain a shallow feature vector, wherein the feature extraction module is mainly composed of a convolution layer and a normalization layer connected in series; Flattening the shallow feature vectors through a preset scanning method to obtain a target long sequence, wherein the first scanning method is used to process image domain data, and the second scanning method is used to process K-space domain data; Inputting the target long sequence into a deep feature extraction module to obtain a deep feature vector; The deep feature vector is input into a reconstruction module to obtain prior information of the input data, wherein the reconstruction module is composed of a convolutional layer.

4. The MRI reconstruction method based on the dual-domain deep unfolding network model according to claim 3, characterized in that: The constraining the dual-domain depth expansion model based on a preset compressed sensing algorithm specifically includes: Obtain the compressed sensing model, analyze the compressed sensing model, and construct K-space and image data regularization terms to constrain MRI reconstruction. The compressed sensing model is defined as: Wherein, A represents the forward operator, i.e., A=MF, F represents Fourier transform, M represents the downsampling mask; x represents the reconstructed MR image; y represents the undersampled K-space data; k represents the reconstructed K-space data; is a quadratic term used to measure the error between x after transformation A and the target y; R x (x) represents a regularization term used to constrain image data; R k (k) represents a regularization term used to constrain the K-space data, λ x and λ k Respectively represent R x (x) and R k The weight coefficient of (k) is a learnable parameter.

5. The MRI reconstruction method based on the dual-domain deep unfolding network model according to claim 4, characterized in that: The step of inputting the simulated under-sampling data into the dual-domain depth expansion model to obtain reconstructed K-space data, and using inverse Fourier transform to obtain a reconstructed image specifically includes: The gradient of the K-space data regularization term is replaced by the Mamba1 network model to build the K-space domain branch of the dual-domain deep expansion network model; The Mamba2 network model is used to replace the gradient of the image data regularization term to build the image domain branch of the dual-domain deep expansion network model; A dual-domain deep unfolding network model is constructed based on the K-space domain branch and the image domain branch; The reconstructed K-space data is obtained by iterating a preset number of times, and the reconstructed MR image is obtained by inverse Fourier transform.

6. The MRI reconstruction method based on the dual-domain deep unfolding network model according to claim 5, characterized in that: The method further includes calculating the loss between the reconstructed K-space data and the real K-space data and between the reconstructed image and the real image and updating the network model parameters through back propagation, wherein the loss function of the dual-domain deep unfolding network model is constructed, specifically including: The K-space loss function is defined as: The pixel loss function is defined as follows: The total loss function can be expressed as Loss total =αLoss freq +βLoss pixel ; Among them, θ is the parameter of the model, S represents the size of K space, that is, the number of pixels, and y i represents the ith data value in the full-sampled K-space, represents the i-th data value of the reconstructed K-space data, F -1 represents the inverse Fourier transform, ∈ is a constant, α and β are Loss freq and Loss pixel The weight coefficient of .

7. An MRI reconstruction system based on a dual-domain deep unfolding network model, characterized in that: The invention comprises a memory and a processor, wherein the memory comprises an MRI reconstruction method program based on a dual-domain deep expansion network model, and the MRI reconstruction method program based on a dual-domain deep expansion network model is executed by the processor to implement the following steps: Acquire simulated under-sampling data, wherein the simulated under-sampling data includes K-space data; A dual-domain deep unfolding network model is constructed based on a preset model framework, wherein the model framework includes Mamba1 and Mamba2; Constraining the dual-domain depth expansion model based on a preset compressed sensing algorithm, including constructing K-space and image data regularization terms to constrain MRI; The simulated under-sampling data is input into the dual-domain depth expansion model to obtain reconstructed K-space data, and an inverse Fourier transform is used to obtain a reconstructed image.

8. The MRI reconstruction system based on the dual-domain deep unfolding network model according to claim 7, characterized in that: The obtaining of simulated under-sampling data specifically includes: Acquire full-sampled K-space data; Randomly generate a mask matrix with the same size as the K-space data; The K-space data is undersampled by using a randomly generated mask to obtain the simulated undersampled data.

9. The MRI reconstruction system based on the dual-domain deep unfolding network model according to claim 8, characterized in that: The dual-domain deep expansion network model is constructed based on a preset model framework, specifically including: Based on the Mamba1 and Mamba2 network models, we obtain the shallow feature extraction module, deep feature extraction module and reconstruction module; The data input into the model first flows to the shallow feature extraction module to obtain a shallow feature vector, wherein the feature extraction module is mainly composed of a convolution layer and a normalization layer connected in series; Flattening the shallow feature vectors through a preset scanning method to obtain a target long sequence, wherein the first scanning method is used to process image domain data, and the second scanning method is used to process K-space domain data; Inputting the target long sequence into a deep feature extraction module to obtain a deep feature vector; The deep feature vector is input into a reconstruction module to obtain prior information of the input data, wherein the reconstruction module is composed of a convolutional layer.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes an MRI reconstruction method program based on a dual-domain deep expansion network model. When the MRI reconstruction method program based on a dual-domain deep expansion network model is executed by a processor, the steps of an MRI reconstruction method based on a dual-domain deep expansion network model as described in any one of claims 1 to 6 are implemented.

Citation Information

Cited By

  • MRI (Magnetic Resonance Imaging) reconstruction method based on iterative shrinkage threshold algorithm and state space module

    CN119338943A

  • Double-domain Swin Mama-based generative adversarial network MRI reconstruction method

    CN120635248A