An unsupervised learning method for non-Cartesian magnetic resonance intelligent fast imaging

Abstract

The present invention discloses an unsupervised learning non-Cartesian magnetic resonance intelligent rapid imaging method, which relates to the fields of medical image processing and magnetic resonance imaging technology. The method comprises the following steps: first, acquiring non-Cartesian undersampled k-space data from a magnetic resonance scanner, dividing the undersampled k-space data into two complementary subsets 1 and 2 according to the order of acquisition, thereby forming a training set; then, designing a deep learning network for magnetic resonance intelligent rapid reconstruction based on unsupervised separating subspace learning for non-Cartesian sampling, the network's inference function, and a loss function; then, using the training set obtained in step S1, determining the optimal parameters of the deep learning network based on unsupervised separating subspace learning; and finally, inputting the non-Cartesian undersampled k-space data to be reconstructed into the trained deep learning network to reconstruct the magnetic resonance image. The present invention has the characteristics of not requiring full sampled data, fast reconstruction speed, and high reconstruction quality.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing and magnetic resonance imaging, and in particular to an unsupervised learning non-Cartesian magnetic resonance intelligent rapid imaging method. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and may not constitute prior art.

[0003] Magnetic resonance imaging (MRI) has the characteristics of being non-invasive, high soft tissue contrast, multiple imaging parameters and contrast, and rich image information. However, MRI faces the bottleneck of slow scanning speed and long imaging time. MRI generally uses two imaging methods: Cartesian and non-Cartesian. The latter is insensitive to motion, has a high acceleration factor, and is widely used in the diagnosis of important parts such as the human brain and heart. Traditional non-Cartesian MRI image reconstruction is mostly based on optimization algorithms, but the iterative process of multiple non-uniform Fourier transforms leads to long reconstruction times. While non-Cartesian MRI image reconstruction based on artificial intelligence can accelerate the reconstruction process, most of it relies on fully sampled k-space data as labels for network training. Therefore, how to avoid the use of fully sampled data while speeding up the reconstruction process is an important issue in fast non-Cartesian MRI.

[0004] In the past, many non-Cartesian MRI reconstruction methods using optimization algorithms have been proposed, such as: a GRAPPA algorithm for arbitrary 2D / 32D non-Cartesian sampling trajectories with rapid calibration based on cyclic ACS boundary conditions (Luo T, Noll DC, Fessler JA, Nielsen JF, A GRAPPA algorithm for arbitrary 2D / 32D non-Cartesian sampling trajectories with rapid calibration, Magnetic Resonance in Medicine, 82(3):1101-1112, 2019.), iterative golden-angle radial sparse parallel MRI reconstruction combining compressed sensing, parallel imaging and golden-angle radial sampling (Feng L, Grimm R, Block KT, Chandarana H, Kim S, Xu J, Axel L, Sodickson DK, Otazo R, Golden-angle radial sparse parallel MRI: combinationof compressed sensing, parallel imaging, and golden-angle radial sampling forfast and flexible dynamic volumetric MRI, Magnetic Resonance in Medicine, 72(3):707-717, 2014.) Other work has extended sparse prior-based reconstruction methods to non-Cartesian MRI reconstruction (Qu B, Zhang Z, Chen Y, et al. "A convergence analysis for projected fast iterative soft-thresholding algorithm under radial sampling MRI," Journal of Magnetic Resonance, 107425, 2023). These methods improve reconstruction performance through k-space learning or sparse reconstruction. However, these methods suffer from the problem of long iterative reconstruction time.

[0005] To reduce reconstruction time, a recent effective approach is to introduce artificial intelligence, such as using a deep unrolled neural network to simulate the iterative sparse image reconstruction process of the projected fast soft thresholding algorithm (pFISTA) (Qu B, Zhang J, Kang T, Lin J, Lin M, She H, Wu Q, Wang M, Zheng G, Radial magnetic resonance image reconstruction with a deep unrolled projected fast iterative soft-thresholding network, Computers in biology and medicine, 168, 2024.). However, the reconstruction speed is affected not only by the acceleration factor, but also by the sensitivity of the channel and the density-compensated unrolled neural network NC-PDNet (Non-Cartesian Primitive Dual Network) (Ramzi Z, Chaithya G, Starck JL, Ciuciu P, NC-PDNet: A density-compensated unrolled network for 2D and 3D non-Cartesian MRI reconstruction, IEEE Transactions on Medical Imaging, 2016). 41(7):1625-1638,2022.), neural networks are used to reconstruct MR images directly from k-space data acquired by Cartesian and non-Cartesian trajectories and multi-channel RF coils, namely end-to-end recurrent neural networks (Oh C, Chung JY, Han Y, An end-to-end recurrent neural network for radial MR image reconstruction, Sensors, 22(19):7277, 2022.). However, these deep learning methods all rely on full-sampled data as network training labels for network learning. In some special imaging scenarios (such as the abdomen and heart), it is difficult to obtain a large amount of full-sampled data, resulting in reduced network training performance or even inability to train. Therefore, for unsupervised non-Cartesian MRI reconstruction, a design is designed that can reduce reconstruction time, reduce scanning time, and improve image quality without requiring full-sampled data as network training labels.

[0006] In summary, there is a lack of non-Cartesian MRI reconstruction methods based on unsupervised learning, and there is an urgent need to develop new non-Cartesian MRI reconstruction methods that do not require full sampling data as training labels to achieve fast and high-quality non-Cartesian magnetic resonance intelligent imaging reconstruction. Summary of the Invention

[0007] The purpose of the present invention is to provide a new method for high-quality and high-efficiency non-Cartesian magnetic resonance intelligent rapid imaging with unsupervised learning.

[0008] The technical solutions of the present invention are as follows:

[0009] An unsupervised learning non-Cartesian magnetic resonance intelligent fast imaging method, comprising:

[0010] Step S1: Generating training data; acquiring non-Cartesian undersampled k-space data from a magnetic resonance scanner, dividing the undersampled k-space data into two complementary subsets 1 and 2 according to the order of acquisition, and combining the undersampled k-space data, the undersampled k-space subset 1, and the subset 2 to form a training set;

[0011] Step S2: Network design; design a deep learning network for intelligent rapid reconstruction of magnetic resonance imaging based on unsupervised separation subspace learning for non-Cartesian sampling, as well as the network's reasoning function and loss function;

[0012] Step S3: training the network; using the training set obtained in step S1, solving the optimal parameters of the deep learning network based on unsupervised separating subspace learning;

[0013] Step S4: reconstructing the image; inputting the non-Cartesian undersampled k-space data to be reconstructed into the trained deep learning network to reconstruct the magnetic resonance image.

[0014] Furthermore, the step S1 includes:

[0015] Step S11: Acquire non-Cartesian sampled multi-channel magnetic resonance data from a magnetic resonance scanner;

[0016] Step S12: Generate golden angle sampling trajectory using sampling trajectory function and non-Cartesian undersampled k-space data ;

[0017] Step S13: Non-Cartesian undersampled k-space data , sampling trajectory according to the golden angle The generated time sequence is divided into complementary subsets and subcollections ;

[0018] Step S14: Undersampled k-space data , undersampled k-space subset , subcollection Together they form the training set.

[0019] Furthermore, the deep learning network for intelligent rapid reconstruction of magnetic resonance images for non-Cartesian sampling based on unsupervised separation subspace learning is composed of a preprocessing module Q and a network iteration module as the core, and is composed of N cascaded network iteration modules;

[0020] Among them, the preprocessing module Q includes two parts: multi-channel sensitivity map estimation and denoising module; the network iteration module includes a depth reconstruction module and a data consistency verification module.

[0021] Furthermore, multi-channel sensitivity map estimation includes:

[0022] Undersampled k-space data According to the Nyquist sampling theorem, the acceleration factor is set, and the range of the intercept center is calculated by the formula. The center part is intercepted and retained to obtain the k-space data. ;

[0023] The k-space data of the intercept center After density compensation, the multi-channel low-frequency sensitivity map is obtained using non-uniform Fourier transform. ;

[0024] Then multi-channel low frequency sensitivity map Take the square root after the square sum to get the real image of the composite coil ;

[0025] Then the multi-channel low frequency sensitivity map Each pixel in each channel is divided by the real image The corresponding pixel points in the initialization coil sensitivity map are obtained.

[0026] Furthermore, the denoising module includes:

[0027] For undersampled k-space subsets Perform non-uniform inverse Fourier transform to generate initialized multi-channel images ;

[0028] The initialized multi-channel map The denoising process is performed on the input image and the denoised image is compared with the multi-channel image. Add to get the fused multi-channel image ;

[0029] By combining multi-channel images Multi-channel low frequency sensitivity diagram , generate undersampled k-space subsets The corresponding initialization synthetic image for subsequent reconstruction .

[0030] Furthermore, the depth reconstruction module includes forward sparse coding submodule, self-learning soft threshold submodule and inverse sparse decoding Submodules;

[0031] Forward sparse coding The submodule consists of K filters, each of which consists of a convolutional layer, a normalization function, a nonlinear activation function, and a convolutional layer cascade; forward sparse coding Submodule to initialize the synthetic image As input, sparsely coded features As output;

[0032] The self-learning soft threshold submodule converts the sparse coding features of the nth iteration block into After performing point-by-point soft thresholding on each channel, a new sparse coding feature is output. ;

[0033] Inverse sparse decoding The submodule consists of K filters, each of which consists of a convolutional layer, a normalization function, a nonlinear activation function, and a convolutional layer cascade; inverse sparse coding Submodules are sparsely coded features As input, update the map to the synthetic image as output.

[0034] Furthermore, the data consistency check module uses the composite image output by the depth reconstruction module As input, perform n iterative verifications to obtain the composite image after the nth verification .

[0035] Furthermore, the deep learning network based on unsupervised separation subspace learning is expressed as:

[0036]

[0037] in: represents the set of learnable parameters of the entire reconstruction network, Indicates the input of undersampled k-space data from the network The network finally reconstructs the synthetic image Nonlinear mapping.

[0038] Furthermore, we design the network's inference function and loss function, including:

[0039] The final synthesized image is obtained through the constraint network and undersampled k-space subsets Dual-domain loss functions Loss1 and Loss2 in image domain and k-space domain are used to achieve unsupervised learning;

[0040] The loss function Loss1 is defined as:

[0041]

[0042] The loss function Loss2 is defined as:

[0043]

[0044] in, T represents the total number of training samples, To subsample the undersampled k-space subset Using Transformation Operators The generated undersampled synthetic channel image, represents the undersampled k-space data Input network to reconstruct synthetic image The mapping process, To reconstruct the synthetic image Through the transformation operator Mapped to the undersampled k-space subset k-space data of the same trajectory;

[0045] The total loss function of the dual domain based on unsupervised learning is expressed as:

[0046]

[0047] in, and It is an adjustable coefficient used to balance the attention level of the image domain and the k-space domain.

[0048] Furthermore, the optimal parameters of the deep learning network are optimized using the Adam optimizer, and the network is trained using the training set generated in step S1 by minimizing the dual-domain total loss function Get the optimal target parameter set .

[0049] Compared with the existing technology, the beneficial effects of the present invention are:

[0050] This paper provides a non-Cartesian MRI intelligent rapid imaging method based on unsupervised learning. This breakthrough, achieved in an unsupervised environment (lacking learning labels, i.e., fully sampled data), utilizes a dual-domain loss function constraint model, a MRI coil sensitivity estimation module, a data consistency check module, and a depth reconstruction module. This method iteratively expands traditional optimization methods into a neural network. This method reduces reliance on training samples and avoids the use of fully sampled data, resulting in fast reconstruction speed and high reconstruction quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is an overall reconstruction network structure of non-Cartesian magnetic resonance intelligent fast imaging based on unsupervised learning;

[0052] Figure 2 Schematic diagram of the Golden angle sampling trajectory (comprising 40 spokes) used in the embodiment;

[0053] Figure 3 This is a schematic diagram of the core parts of the reconstructed network model, the preprocessing module Q and the network iteration module;

[0054] Figure 4 : The undersampled label image of the human brain and the reconstructed image under 10 times acceleration; among them, (a) is the fully sampled label image, (b) is the undersampled image, (c) is the network reconstructed image of the present invention, (d) is the error map of the fully sampled label image itself, (e) is the error map corresponding to the fully sampled image and the undersampled image, and (f) is the error map corresponding to the reconstructed image of the present invention and the fully sampled image. DETAILED DESCRIPTION

[0055] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus comprising the element.

[0056] The features and performance of the present invention are further described in detail below with reference to the embodiments.

[0057] Example 1

[0058] The embodiment of the present invention uses multi-coil human brain data to construct a training set. In the absence of full-sampled k-space data, the optimal image and data are obtained through multiple network learning training and constraints, thereby obtaining a reconstructed magnetic resonance image.

[0059] The embodiment of the present invention includes the following steps:

[0060] Step S1: training data generation; non-Cartesian undersampled k-space data is acquired from a magnetic resonance scanner, and the undersampled k-space data is divided into two complementary subsets 1 and subset 2 according to the acquisition order. The undersampled k-space data, undersampled k-space subset 1, and subset 2 together constitute a training set.

[0061] In this embodiment, specifically, step S1 includes:

[0062] Step S11: Acquire non-Cartesian sampled multi-channel magnetic resonance data from a magnetic resonance scanner;

[0063] Step S12: Generate golden angle sampling trajectory using sampling trajectory function and non-Cartesian undersampled k-space data ;in, Indicates the channels of non-Cartesian radial line sampling k-space data, represents the complex domain, L, P, and C represent the number of L radial lines in the k-space data, the number of P sampling points per radial line, and the number of (coil) channels of the k-space data, respectively;

[0064] Step S13: Non-Cartesian undersampled k-space data , sampling trajectory according to the golden angle The generated time sequence is divided into complementary subsets and subcollections ;right 、 have:

[0065]

[0066] in, is the undersampled k-space data, For subset 1, For subset 2, for radial lines ,have ;

[0067] Step S14: Undersampled k-space data , undersampled k-space subset , subcollection Together they form the training set;

[0068] It should be noted that this embodiment uses a 3T magnetic resonance imaging machine to perform magnetic resonance imaging on the brains of 60 volunteers. The magnetic resonance imaging sequence parameters used in this embodiment are: echo time TE = 27ms, repetition time TR = 2750ms, field of view 320×320mm, and number of coils 4. Non-Cartesian sampling multi-coil magnetic resonance imaging data of the 60 volunteers' brains are collected, and the golden angle sampling trajectory is generated using the sampling trajectory function. and non-Cartesian (radial) undersampled k-space data .

[0069] Step S2: Network design; design a deep learning network for intelligent rapid reconstruction of magnetic resonance imaging based on unsupervised separation subspace learning for non-Cartesian sampling, the network's reasoning function, and the loss function.

[0070] In this embodiment, specifically, the deep learning network for intelligent rapid reconstruction of magnetic resonance for non-Cartesian sampling based on unsupervised separation subspace learning is composed of a preprocessing module Q and a network iteration module as the core, and is composed of N cascaded network iteration modules;

[0071] Among them, see Figure 1 and Figure 3 The preprocessing module Q includes two parts: multi-channel sensitivity map estimation and denoising module; the network iteration module includes a depth reconstruction module ( ) and data consistency verification module (Data consistency, DC).

[0072] In this embodiment, specifically, multi-channel sensitivity map estimation includes:

[0073] Undersampled k-space data According to the Nyquist sampling theorem, the acceleration factor (undersampling step) is set, and the range of the intercept center is calculated by the formula, and the central part is intercepted and retained to obtain the k-space data. It should be noted that the intercept center radius is R, and the process is expressed as:

[0074]

[0075] in, represents the radius of the sampling trajectory, represents the acceleration factor, which is expressed as ,in represents the number of fully sampled radial k-space spokes, represents the number of spokes in the undersampled radial k-space;

[0076] The k-space data of the intercept center After density compensation, the multi-channel low-frequency sensitivity map is obtained using non-uniform Fourier transform. ;

[0077] Then multi-channel low frequency sensitivity map Take the square root after the square sum to get the real image of the composite coil ;

[0078] Then the multi-channel low frequency sensitivity map Each pixel in each channel is divided by the real image The corresponding pixel points in the initialization coil sensitivity map are obtained; the process is expressed as:

[0079]

[0080]

[0081]

[0082] Among them, represents the k-space center data intercepted with the interception radius R, Indicates density compensation, represents the inverse non-uniform Fourier transform, It represents the operation of taking the square root of the sum of the squares of the multi-channel sensitivity images to obtain the composite image. is the estimated sensitivity map.

[0083] In this embodiment, the denoising module specifically includes:

[0084] First, the undersampled k-space subset Perform non-uniform inverse Fourier transform to generate initialized multi-channel images ; The process is expressed as:

[0085]

[0086] Next, initialize the multi-channel map The denoising process is performed on the input image and the denoised image is compared with the multi-channel image. Add to get the fused multi-channel image ; The process is expressed as:

[0087]

[0088] in, are the network learning parameters, Represents the nonlinear mapping of the denoising module. The network consists of P convolutional layers, each of which is connected to a normalization function (Batch Normalization, BN) and a nonlinear mapping function (Rectified Linear Unit, ReLU). The size of the convolution kernel is ;

[0089] By combining multi-channel images Multi-channel low frequency sensitivity diagram , generate undersampled k-space subsets The corresponding initialization synthetic image for subsequent reconstruction ; The process is defined as:

[0090]

[0091] in, is the complex conjugate transpose of S;

[0092] The specific calculation is ;

[0093] In summary, the preprocessing module Q can be represented by the following nonlinear mapping:

[0094]

[0095] in, is the set of learnable parameters in the preprocessing module Q.

[0096] In this embodiment, specifically, the depth reconstruction module removes image artifacts through a convolutional neural network to improve the image reconstruction quality; the depth reconstruction module includes forward sparse coding submodule, self-learning soft threshold submodule and inverse sparse decoding Submodule.

[0097] Forward sparse coding The submodule consists of K filters, each of which consists of a convolutional layer, a normalization function (Batch Normalization, BN), a nonlinear activation function (Rectified Linear Unit, ReLU) and a convolutional layer cascade. The convolution kernel size of the convolutional layer is ;

[0098] Forward sparse coding Submodule to initialize the synthetic image As input, sparsely coded features As output; the process can be described as:

[0099]

[0100] in, Parameters learned by the submodule network for the nth iteration block.

[0101] The self-learning soft threshold submodule converts the sparse coding features of the nth iteration block into After performing point-by-point soft thresholding on each channel, a new sparse coding feature is output. ; can be expressed as:

[0102]

[0103] in, represents the soft threshold operation algorithm, is the network learnable threshold, It is the sparse coding feature after the soft threshold operation.

[0104] Inverse sparse decoding The submodule consists of K filters, each of which consists of a convolutional layer, a normalization function (Batch Normalization, BN), a nonlinear activation function (Rectified Linear Unit, ReLU) and a convolutional layer cascade. The convolution kernel size of the convolutional layer is ; Inverse sparse coding Submodules are sparsely coded features As input, update the map to the synthetic image As output; the process can be described as:

[0105]

[0106] in, The parameters learned by the submodule network for the nth iteration block, It is the synthetic image output by the n-th iteration sparse reconstruction module.

[0107] The network iteration blocks through , and The three cascade structures realize the iteration of the network. The nonlinear mapping of the reconstruction process of the nth iterative block can be expressed as:

[0108]

[0109] in is the set of parameters learned by the deep reconstruction module network in the nth iteration block.

[0110] In this embodiment, specifically, the data consistency check module is used to enhance the consistency of the reconstructed image data and the measured k-space data, improve the quality of the reconstructed image, accelerate the convergence, and use the synthetic image output by the depth reconstruction module to obtain the same image as the original image. As input, the following form is used for verification:

[0111]

[0112] Where n represents the nth iteration, is the inverse NUFFT operator, represents the synthetic image before the nth verification, represents the composite image after the nth verification;

[0113] Therefore, a single iteration block can be represented as the following mapping form:

[0114]

[0115] in, represents the set of network training parameters for the kth iteration, Represents network training synthetic images To update the composite image Nonlinear mapping.

[0116] Therefore, the deep learning network based on unsupervised separating subspace learning is expressed as:

[0117]

[0118] in, Represents the set of learnable parameters of the entire reconstruction network, N represents the Nth iteration, Indicates undersampled data input from the network The network finally reconstructs the synthetic image Nonlinear mapping.

[0119] In this embodiment, the reasoning function and loss function of the design network are specifically designed, including:

[0120] The final synthesized image is obtained through the constraint network and undersampled k-space subsets The dual-domain loss functions Loss1 and Loss2 in the image domain and k-space domain realize unsupervised learning (no need for fully sampled k-space data or fully sampled images as labels for network training).

[0121] Loss function constraint Loss1 implementation steps:

[0122] Using a subset Using Transformation Operators Generate undersampled synthetic channel images , the formula is expressed as:

[0123]

[0124] in, represents the data of the undersampled k-space subset 2, Indicates density compensation, is the inverse non-uniform Fourier transform, Sensitivity diagram The complex conjugate transpose of ;

[0125] The final composite image reconstructed by the network With subcollections Generated undersampled synthetic channel map Loss function constraints are performed in the image domain; Loss1 is defined as:

[0126]

[0127] in, represents the two-norm term, n represents the nth iteration block, N represents the total number of iteration blocks, t represents the tth sample, T represents the total number of training samples, Represents a sum operation.

[0128] Loss function constraint Loss2 implementation steps:

[0129] The network learns to reconstruct the graph Through the transformation operator k-space data mapped to the same trajectory as subset 2 , the process can be described as:

[0130]

[0131] in, is the inverse non-uniform Fourier transform, represents the sensitivity graph;

[0132] Will and The constraint Loss2 of the loss function is defined as:

[0133]

[0134] in, represents the undersampled k-space, represents the k-space subset 2, represents the undersampled k-space data Input network to reconstructed graph The mapping process, n represents the nth iteration block, N represents the total number of iteration blocks, t represents the tth sample, T represents the total number of training samples, Represents a sum operation.

[0135] In summary, the total loss function of dual-domain constraints based on unsupervised learning can be expressed as:

[0136]

[0137] in, and It is an adjustable coefficient used to balance the attention level of the image domain and the k-space domain.

[0138] The final reconstructed image of the undersampled k-space subset 1 passing through the network Loss function constraints are performed on the undersampled k-space subset 2 in the image domain and the k-space domain to realize a non-Cartesian intelligent fast MRI reconstruction method that does not require full sampling data as training labels.

[0139] Step S3: training the network; using the training set obtained in step S1, solving the optimal parameters of the deep learning network based on unsupervised separating subspace learning;

[0140] In this embodiment, the optimal parameters of the deep learning network are determined by using the Adam optimizer with good performance. The network is trained using the training set generated in step S1, and the dual-domain total loss function is minimized. Get the optimal target parameter set .

[0141] Step S4: reconstructing the image; inputting the non-Cartesian undersampled k-space data to be reconstructed into the trained deep learning network to reconstruct the magnetic resonance image; the network reconstruction process can be expressed as:

[0142]

[0143] In the embodiment, the input of the network is a radial sampling trajectory with an acceleration factor of 10 (the sampling trajectory diagram is shown in FIG. Figure 2 The undersampled multi-coil human brain data (shown in Figure 2) has a data dimension of 640×50×4. The fully sampled label image, undersampled image, and reconstructed output image of the network are shown in Figure 2. Figure 4 (a), Figure 4 (b) and Figure 4 (c); The error diagram of the undersampled image and the fully sampled image is Figure 4 (e) The error diagram between the reconstructed image of the present invention and the fully sampled image is Figure 4 (f).

[0144] Compared to existing technologies, this method, based on an unsupervised learning design model, uses the unsupervised division of undersampled non-Cartesian MRI k-space data into two complementary subsets based on the acquisition order of the sampling trajectory, without full sampling data. The k-space data in one subspace (subset) is used as network input, while the data in the other subspace serves as network training labels. The network learning process is constrained by the equality of the two subspace images, and the neural network learning is iteratively developed using traditional optimization methods. This method features fast reconstruction speed, high reconstruction quality, and no need for full sampling data.

[0145] The above-described embodiments merely represent specific implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of protection of the present application. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the technical concept of the present application, and all such variations and improvements fall within the scope of protection of the present application.

[0146] This background section is provided to generally present the context of the invention, and the work of the presently named inventors, the work to the extent described in this background section, and aspects of the description in this section that did not constitute prior art at the time of filing are neither explicitly nor implicitly admitted to be prior art to the present invention.

Claims

1. A non-Cartesian magnetic resonance intelligent rapid imaging method based on unsupervised learning, characterized in that: include: Step S1: training data generation; Acquire non-Cartesian undersampled k-space data from a magnetic resonance scanner, divide the undersampled k-space data into two complementary subsets 1 and 2 according to the order of acquisition, and form a training set with the undersampled k-space data, the undersampled k-space subset 1, and the subset 2; Step S2: Network design; design a deep learning network for intelligent rapid reconstruction of magnetic resonance imaging based on unsupervised separation subspace learning for non-Cartesian sampling, as well as the network's reasoning function and loss function; Step S3: training the network; using the training set obtained in step S1, solving the optimal parameters of the deep learning network based on unsupervised separating subspace learning; Step S4: reconstructing the image; inputting the non-Cartesian undersampled k-space data to be reconstructed into the trained deep learning network to reconstruct the magnetic resonance image; The deep learning network based on unsupervised separation subspace learning is expressed as: in: represents the set of learnable parameters of the entire reconstruction network, Indicates the input of undersampled k-space data from the network The network finally reconstructs the synthetic image Nonlinear mapping of ; Design the network's inference capabilities and loss functions, including: The final synthesized image is obtained through the constraint network and undersampled k-space subsets Dual-domain loss functions Loss1 and Loss2 in image domain and k-space domain are used to achieve unsupervised learning; The loss function Loss1 is defined as: The loss function Loss2 is defined as: in, T represents the total number of training samples, To subsample the undersampled k-space subset Using Transformation Operators The generated undersampled synthetic channel image, represents the undersampled k-space data Input network to reconstruct synthetic image The mapping process, To reconstruct the synthetic image Through the transformation operator Mapped to the undersampled k-space subset k-space data of the same trajectory; Indicates the total number of iteration blocks; The total loss function of the dual domain based on unsupervised learning is expressed as: in, and It is an adjustable coefficient used to balance the attention level of the image domain and the k-space domain.

2. The unsupervised learning non-Cartesian magnetic resonance intelligent rapid imaging method according to claim 1, characterized in that: The step S1 includes: Step S11: Acquire non-Cartesian sampled multi-channel magnetic resonance data from a magnetic resonance scanner; Step S12: Generate golden angle sampling trajectory using sampling trajectory function and non-Cartesian undersampled k-space data ; Step S13: Non-Cartesian undersampled k-space data , sampling trajectory according to the golden angle The generated time sequence is divided into complementary subsets and subcollections ; Step S14: Undersampled k-space data , undersampled k-space subset , subcollection Together they form the training set.

3. The unsupervised learning non-Cartesian magnetic resonance intelligent rapid imaging method according to claim 2, characterized in that: The deep learning network for intelligent rapid reconstruction of magnetic resonance imaging for non-Cartesian sampling based on unsupervised separation subspace learning is composed of a preprocessing module Q and a network iteration module as the core, and is composed of N cascaded network iteration modules. Among them, the preprocessing module Q includes two parts: multi-channel sensitivity map estimation and denoising module; the network iteration module includes a depth reconstruction module and a data consistency verification module.

4. The unsupervised learning non-Cartesian magnetic resonance intelligent rapid imaging method according to claim 3, characterized in that: Multi-channel sensitivity map estimation, including: Undersampled k-space data According to the Nyquist sampling theorem, the acceleration factor is set, and the range of the intercept center is calculated by the formula. The center part is intercepted and retained to obtain the k-space data. ; The k-space data of the intercept center After density compensation, the multi-channel low-frequency sensitivity map is obtained using non-uniform Fourier transform. ; Then multi-channel low frequency sensitivity map Take the square root after the square sum to get the real image of the composite coil ; Then the multi-channel low frequency sensitivity map Each pixel in each channel is divided by the real image The corresponding pixel points in the initialization coil sensitivity map are obtained.

5. The unsupervised learning non-Cartesian magnetic resonance intelligent rapid imaging method according to claim 4, characterized in that: Denoising module, including: For undersampled k-space subsets Perform non-uniform inverse Fourier transform to generate initialized multi-channel images ; The initialized multi-channel map The denoising process is performed on the input image and the denoised image is compared with the multi-channel image. Add to get the fused multi-channel image ; By combining multi-channel images Multi-channel low frequency sensitivity diagram , generate undersampled k-space subsets The corresponding initialization synthetic image for subsequent reconstruction .

6. The unsupervised learning non-Cartesian magnetic resonance intelligent rapid imaging method according to claim 5, characterized in that: The depth reconstruction module includes forward sparse coding submodule, self-learning soft threshold submodule and inverse sparse decoding Submodules; Forward sparse coding The submodule consists of K filters, each of which consists of a convolutional layer, a normalization function, a nonlinear activation function, and a convolutional layer cascade; forward sparse coding Submodule to initialize the synthetic image As input, sparsely coded features As output; The self-learning soft threshold submodule converts the sparse coding features of the nth iteration block into After performing point-by-point soft thresholding on each channel, a new sparse coding feature is output. ; Inverse sparse decoding The submodule consists of K filters, each of which consists of a convolutional layer, a normalization function, a nonlinear activation function, and a convolutional layer cascade; inverse sparse coding Submodules are sparsely coded features As input, update the map to the synthetic image as output.

7. The unsupervised learning non-Cartesian magnetic resonance intelligent rapid imaging method according to claim 6, characterized in that: The data consistency check module uses the composite image output by the depth reconstruction module As input, perform n iterative verifications to obtain the composite image after the nth verification .

8. The unsupervised learning non-Cartesian magnetic resonance intelligent rapid imaging method according to claim 1, characterized in that: The optimal parameters of the deep learning network are obtained by using the Adam optimizer and the training set generated in step S1 to train the network by minimizing the dual-domain total loss function. Get the optimal target parameter set .

Images