A dynamic magnetic resonance intelligent imaging method

Abstract

A dynamic magnetic resonance imaging (MRI) method is disclosed, involving dynamic MRI image reconstruction. The method acquires multi-coil 3D MRI Fourier space-time data. Through inverse Fourier transform, Fourier space-time undersampling, data separation and extraction, and coil sensitivity encoding / decoding operations, a training set is established, including fully sampled merged coil 2D image space-time data, undersampled zero-padded multi-coil 2D Fourier space-time data, and corresponding 2D Fourier space-time undersampling templates. A deep learning network model based on a temporal low-rank and spatial sparse model is designed, along with the network's feedback function and loss function. The optimal parameters of the deep learning network based on the temporal low-rank and spatial sparse model are solved. The undersampled multi-coil 3D MRI Fourier space-time data to be reconstructed is input into the trained network to achieve dynamic MRI image reconstruction. The method features fast reconstruction speed, high reconstruction quality, low memory consumption, and a small amount of training data required.

Description

Technical Field

[0001] The present invention relates to a method for reconstructing dynamic magnetic resonance images, and in particular to a method for dynamic magnetic resonance intelligent imaging based on temporal low-rank and spatial sparse model expansion. Background Art

[0002] Magnetic resonance imaging plays an indispensable role in modern medical diagnosis. Dynamic MRI is a non-invasive and non-radioactive imaging technology that simultaneously reveals image anatomical structures and dynamic physiological functions with high quality. However, its long data acquisition time limits the realization of high spatiotemporal resolution acquisition, and is prone to cause patient discomfort and the introduction of motion-induced artifacts. Parallel imaging and sparse sampling shorten the scanning time by undersampling the data, achieving accelerated acquisition of dynamic imaging, but will introduce stronger image artifacts. Therefore, after accelerated acquisition, the reconstruction of dynamic images with strong artifacts is an important issue in dynamic MRI.

[0003] In the past two decades, many dynamic MRI image reconstruction methods based on prior information models have been proposed, among which sparse and low-rank priors are more commonly used, mainly including: 1) Iterative reconstruction methods by constraining the sparse characteristics of the overall image in the transform domain (L. Feng et al., "Highly accelerated real-time cardiac cine MRIusing kt SPARSE-SENSE," Magnetic Resonance in Medicine, vol. 70, pp. 64-74, 2013.; Y. Hu et al., "Spatiotemporal flexible sparse reconstruction for rapid dynamic contrast-enhanced MRI," IEEE Transactions on Biomedical Engineering, vol. 69, pp. 229-243, 2022.).2) Low-rank regularization methods using spatiotemporal partial separability or structured low-rank priors (Z. Liang, "Spatiotemporal imaging with partially separable functions", in 4th IEEE International Symposium on Biomedical Imaging (ISBI), 2007, pp. 988-991.; KH Jin, D. Lee, and JCYe, "A general framework for compressed sensing and parallel MRI using annihilating filter based low-rank Hankel matrix," IEEE Transactions on Computational Imaging, vol. 2, pp. 480-495, 2016.; X. Zhang et al. "Accelerated MRI reconstruction with separable and enhanced low-rank Hankel regularization," IEEE Transactions on Medical Imaging, vol. 41, pp. 2486-2498, 2022.; X. Qu et al. "Accelerated NMR spectroscopy with low-rank reconstruction," Angewandte Chemie International Edition, vol.54, pp.852-854, 2015.).3) In view of the good reconstruction performance and complementarity of the above methods, a method combining sparse and low-rank regularization has also emerged to achieve better performance (B. Zhao, J. P. Haldar, A. G. Christodoulou, Z. Liang, "Image reconstruction from highly undersampled (k, t)-space data with joint partial separability and sparsity constraints", IEEE Transactions on Medical Imaging, vol. 31, pp. 1809-1820, 2012.; R. Otazo, E. Candès, D. K. Sodickson, "Low-rank plus sparse matrix decomposition for accelerated dynamic MRI with separation of background and dynamic components," Magnetic Resonance in Medicine, vol. 73, pp. 1125-1136, 2015.). However, the inherent iterative process of all the above methods leads to a long reconstruction time, and the selection of algorithm parameters is also very dependent on experience.

[0004] Recently, deep learning has gradually become a powerful tool for dynamic magnetic resonance imaging, which can provide higher image quality and faster reconstruction speed (J.Schlemper, J.Caballero, JVHajnal, ANPrice, D.Rueckert, "Adeep cascade of convolutional neural networks for dynamic MR image reconstruction", IEEE Transactions on Medical Imaging, vol.37, pp.491-503, 2018.; Z.Ke et al., "Learned low-rank priors in dynamic MR imaging", IEEE Transactions on Medical Imaging, vol.40, pp.3698-3710, 2021.). However, the existing methods are all trained and solved for the entire high-dimensional dynamic image reconstruction problem, which has high requirements on the number of dynamic magnetic resonance image training sets and the memory of computing devices. Separable learning has been proven to effectively reduce the requirements for training data volume and computing device memory in the problem of static magnetic resonance image reconstruction (Z. Wang et al., "One-dimensional deep low-rank and sparsenetwork for accelerated MRI", IEEE Transactions on Medical Imaging, vol. 42, pp. 79-90, 2023.; Z. Wang et al., "One for multiple: Physics-informed synthetic data boosts generalizable deep learning for fast MRI reconstruction", arXiv: 2307.13220, 2023.), and is expected to be extended to dynamic magnetic resonance imaging.

[0005] In summary, in dynamic MRI image reconstruction, the reconstruction quality and computational efficiency of existing methods still need to be improved. There is no method that simultaneously utilizes the low rank of multiple time frames of dynamic MRI data and the sparsity of image space to establish a spatiotemporal deep neural network based on separable learning to achieve fast, high-quality, and low-memory dynamic MRI intelligent imaging. Summary of the invention

[0006] The present invention aims to solve the above-mentioned technical problems in the prior art and provide a dynamic magnetic resonance intelligent imaging method with fast reconstruction speed, high reconstruction quality and low memory consumption.

[0007] The present invention comprises the following steps:

[0008] 1) Acquire multi-coil three-dimensional magnetic resonance Fourier space-time data, and establish a training set including fully sampled merged coil two-dimensional image space-time data, undersampled zero-filled multi-coil two-dimensional Fourier space-time data and corresponding two-dimensional Fourier space-time undersampled templates through inverse Fourier transform, Fourier space-time undersampling, data separation and extraction, and coil sensitivity encoding and decoding operations;

[0009] 2) Design a deep learning network model based on temporal low-rank and spatial sparse models, the network’s feedback function and loss function;

[0010] 3) Using the training set obtained in step 1), solve the optimal parameters of the deep learning network based on the temporal low-rank and spatial sparse model;

[0011] 4) Inputting the undersampled multi-coil three-dimensional magnetic resonance Fourier space-time data to be reconstructed into the trained network to achieve dynamic magnetic resonance image reconstruction.

[0012] In step 1), the specific method of acquiring multi-coil three-dimensional magnetic resonance Fourier space-time data and establishing a training set including fully sampled merged coil two-dimensional image space-time data, undersampled zero-filled multi-coil two-dimensional Fourier space-time data and corresponding two-dimensional Fourier space-time undersampled templates through inverse Fourier transform, Fourier space-time undersampling, data separation and extraction, and coil sensitivity encoding and decoding operations can be:

[0013] First, fully sampled multi-coil 3D MRI Fourier space-time data are acquired from the MRI instrument and decoded by coil sensitivity. and spatial 2D inverse Fourier transform Get fully sampled merged coil 3D image space-time data represents the complex domain, M, N, and T represent the length of the frequency encoding dimension, the length of the phase encoding dimension, and the number of time frames in the data, respectively; then, the Fourier space-time undersampling operation is performed on each coil data in the fully sampled multi-coil three-dimensional magnetic resonance Fourier space-time data. It can obtain under-sampled zero-filled multi-coil two-dimensional Fourier space-time data Defined as J represents the number of coils of data, Represents the two-dimensional Fourier transform of space Indicates coil sensitivity coding; then Each coil data is transformed along the frequency encoding dimension by a one-dimensional inverse Fourier transform It can obtain under-sampled zero-filled multi-coil three-dimensional mixed space-time data Defined as represents the imaging forward coding operator, represents the one-dimensional Fourier transform along the phase encoding dimension; Divided into M rows along the frequency encoding dimension, each row contains multi-coil 2D Fourier space-time data with under-sampled zero filling Then Divided into M rows along the frequency encoding dimension, each row contains the fully sampled combined coil 2D image space-time data Then the undersampling operation The corresponding generated undersampling template is divided into M rows along the frequency encoding dimension, each row contains the corresponding two-dimensional Fourier space-time undersampling template Finally, by Z m , U m and X m Together they form the training set.

[0014] In step 2), the deep learning network model takes the iteration block as the core, and stacks several iteration blocks as the overall network structure. Each iteration block contains three sub-blocks; the network structure of a single iteration block is as follows:

[0015] a) Sub-block R is used to update the network intermediate quantity b obtained based on the temporal low-rank prior solution m,n ; It consists of L containing the residual structure R The convolutional neural network consists of a one-dimensional convolutional neural network with a convolution kernel size of I. R ; The first layer input is in represents the final output of the previous iteration block, ε n Indicates that the extraction is along The nth vector of the first dimension can also be obtained by the inverse operation Σ represents sum operation; 1st to Lth R The layers are connected by a linear rectified unit (ReLU), and the input of each layer is the output of the previous network layer; finally, the final calculation result b of the sub-block R is obtained m,n ; The complete sub-block R is represented by the following nonlinear mapping function:

[0016]

[0017] Among them, Θ Rrepresents the internal parameters of the sub-block R in the iteration block, f R (·) represents the nonlinear mapping trained by sub-block R, Represents L in sub-block R R Layer one-dimensional convolutional neural network;

[0018] b) Sub-block G is used to update the network intermediate quantity d obtained based on the spatial sparse prior solution m,t ; It consists of two L G The convolutional neural network consists of a one-dimensional convolutional neural network with a convolution kernel size of I. G ; The first layer input of the first one-dimensional convolutional neural network is Indicates that the extraction is along The tth vector of the second dimension can also be recovered by the inverse operation Σ represents sum operation; 1st to Lth G The layers are connected by a linear rectified unit (ReLU), and the input of each layer is the output of the previous network layer; the output of the last layer of the first one-dimensional convolutional neural network is soft-thresholded and then input into the first layer of the second one-dimensional convolutional neural network; the 1st to Lth layers of the second one-dimensional convolutional neural network are G The layers are connected by a linear rectified unit (ReLU), and the input of each layer is the output of the previous network layer; finally, the final calculation result d of the sub-block G is obtained. m,t ; The complete sub-block G is represented by the following nonlinear mapping function:

[0019]

[0020] Among them, Θ G represents the internal parameters of the sub-block G in the iteration block, f G (·) represents the nonlinear mapping trained by sub-block G, and Represent the first and second L in sub-block G respectively G The soft threshold operation soft is defined as soft(x;θ)=max{|x|-θ}·x / |x|, where θ represents the trainable threshold.

[0021] c) Sub-block D is used to update the output of the current iterative block obtained based on the data verification operation. Data verification requires that each output must maintain consistency with the sampled data to ensure that the output result is aligned with the collected data; the complete sub-block D is represented by the following function:

[0022]

[0023] Among them, Θ D represents the internal parameters of the sub-block D in the iteration block, f D (·) represents the linear mapping trained by sub-block D; its specific operation can be expressed as:

[0024]

[0025] in, p represents the data index, Ω represents the position of the sampled data point, and λ 1 ,λ 2 Both represent regularization parameters;

[0026] In summary, by cascading the above three sub-blocks R, G and D, the entire single iteration block can be represented by the following nonlinear mapping function group:

[0027]

[0028] in, Represents the set of internal parameters of all sub-blocks in the iteration block; F 1 (·) indicates the concatenation of sub-blocks; F 2 (·) indicates the parallel connection of sub-blocks; Indicates the trained arrive The nonlinear mapping is the nonlinear mapping of each sub-block f D (Z m ,b m,n ,d m,t |Θ D )

[0029] The designed deep learning network model based on temporal low-rank and spatial sparse model can be expressed as a whole:

[0030]

[0031] Among them, Θ represents the set of internal parameters of the overall network; F overall (Z m |Θ) represents the trained multi-coil two-dimensional Fourier space-time data Z from under-sampling and zero-filling m To the final output value of the network Non-linear mapping of all iterative blocks cascaded;

[0032] The feedback function of the network is an important process for the network to solve the target value. In the process of network model construction, the output value of the network is Combined coil 2D image space-time data with full sampling X mCompare and feed back the gradient to update the parameters of the iterative module, so that the network output value is closer to the fully sampled artifact-free dynamic magnetic resonance image;

[0033] The loss function can be defined as:

[0034]

[0035] Among them, Θ represents the set of internal parameters of the overall network, ||·|| 2 represents the bi-norm term, k represents the kth iteration block, k = 1, 2, ..., K, K represents the total number of iteration blocks, C represents the total number of training data, and Σ represents the summation operation.

[0036] In step 3), the optimal parameters of the deep learning network based on the temporal low-rank and spatial sparse model can be solved by using the Adam optimizer with good performance in deep learning, and the network training is performed using the training set generated in step 1), by minimizing the loss function in step 2). To obtain the optimal target parameter set

[0037] In step 4), the under-sampled multi-coil three-dimensional magnetic resonance Fourier space-time data to be reconstructed is input into the trained network to realize dynamic magnetic resonance image reconstruction; the under-sampled multi-coil three-dimensional magnetic resonance Fourier space-time data to be reconstructed is input into the trained network to realize dynamic magnetic resonance image reconstruction; First, do a one-dimensional inverse Fourier transform along the frequency encoding dimension get All rows of form a batch as the input of the trained network, and the network is forward propagated according to the trained relatively optimal parameters, the batch is reconstructed in parallel, and they are stitched together in sequence to obtain the final reconstructed magnetic resonance image. in, The network reconstruction process for each row of can be expressed as:

[0038]

[0039] The present invention proposes a dynamic magnetic resonance intelligent imaging method. The method first collects magnetic resonance Fourier space-time data, then performs Fourier inverse transform, Fourier space-time undersampling, data separation and extraction, and coil sensitivity encoding and decoding operations according to signal characteristics, and constructs a training set composed of fully sampled merged coil two-dimensional image space-time data, undersampled zero-filled multi-coil two-dimensional Fourier space-time data and corresponding two-dimensional Fourier space-time undersampling templates, the former is the label of the author network, and the latter two are used as inputs; then, a deep learning neural network structure is designed according to the process of iterative solution of temporal low-rank and spatial sparse models, and the reconstruction model composed of the optimal parameters of the network is solved using the above training set; finally, the undersampled magnetic resonance Fourier space-time data to be reconstructed is input into the trained network to realize dynamic magnetic resonance image reconstruction.

[0040] In terms of training data, the present invention uses a high-dimensional data separation strategy based on the sampling characteristics of dynamic magnetic resonance images to achieve significant expansion of the data set and alleviate the bottleneck problem of insufficient training data. In terms of network design, the present invention uses traditional optimization methods as a guide to design a spatiotemporal deep neural network by simultaneously constraining temporal low rank and spatial sparsity, which has the characteristics of fast reconstruction speed, high reconstruction quality and low memory consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a Fourier space-time under-sampling template with a sampling rate of 17% adopted in the embodiment.

[0042] Figure 2 It is a deep learning network structure based on temporal low-rank and spatial sparse models. (a) and (b) are schematic diagrams of the training and reconstruction phases, the overall network structure, and the kth iteration block structure; (c) is a schematic diagram of the network structure of each sub-block.

[0043] Figure 3 It is a heart full sampling label image and a reconstructed image at a sampling rate of 17%. Among them, (a) is a heart full sampling label image under multiple time frames, and (b) is a heart reconstructed image under multiple time frames of the present invention. DETAILED DESCRIPTION

[0044] The following embodiments will further illustrate the present invention in conjunction with the accompanying drawings. In the embodiments of the present invention, a training set is constructed using multi-coil cardiac movie dynamic data, and the optimal network parameters are obtained through several iterative trainings. Finally, the undersampled multi-channel cardiac movie dynamic data to be reconstructed is input into the trained deep learning network model to obtain a reconstructed dynamic magnetic resonance image.

[0045] The embodiment of the present invention includes the following steps:

[0046] Step 1: Obtain multi-coil 3D MRI Fourier space-time data, and establish a training set including fully sampled merged coil 2D image space-time data, undersampled zero-filled multi-coil 2D Fourier space-time data, and corresponding 2D Fourier space-time undersampled templates through inverse Fourier transform, Fourier space-time undersampling, data separation and extraction, and coil sensitivity encoding and decoding operations.

[0047] This embodiment uses a magnetic resonance apparatus with a magnetic field strength of 3 Tesla to perform dynamic movie imaging on the hearts of 100 volunteers. The magnetic resonance imaging sequence parameters used in this embodiment are: the echo time of the sequence is TE = 1.6ms, the repetition time is TR = 3.6ms, and the spatial resolution is 2×2mm 2 , the time resolution is 50ms, the layer thickness is 8mm, and the number of coils is 10. 1000 cardiac movie multi-coil Fourier space-time data of 100 volunteers scanned by magnetic resonance imaging were used as the source of the network training set.

[0048] The above fully sampled multi-coil 3D MRI Fourier space-time data is decoded by coil sensitivity and spatial 2D inverse Fourier transform Get fully sampled merged coil 3D image space-time data The length of the frequency encoding dimension in the data is 192, the length of the phase encoding dimension is 192, and the number of time frames is 12; then the Fourier space-time undersampling operation is performed on each coil data in the fully sampled multi-coil three-dimensional magnetic resonance Fourier space-time data (Fourier space-time undersampling template such as Figure 1 As shown. Figure 1 In the undersampling template, the white ones are sampling points, indicating that the data corresponding to the position is sampled; the black ones are points that are not sampled, and the data corresponding to the position is lost). The undersampling zero-filled multi-coil two-dimensional Fourier space-time data can be obtained. Defined as Among them, 1920=192×10 means that the number of coils of the data is 10. Represents the two-dimensional Fourier transform of space Indicates coil sensitivity coding; then Each coil data is transformed along the frequency encoding dimension by a one-dimensional inverse Fourier transform It can obtain under-sampled zero-filled multi-coil three-dimensional mixed space-time data Defined as represents the imaging forward coding operator, represents the one-dimensional Fourier transform along the phase encoding dimension; Divided into 192 rows along the frequency encoding dimension, each row contains the multi-coil 2D Fourier space-time data of under-sampled zero-filled Then The frequency encoding dimension is divided into 192 rows, each row contains the fully sampled combined coil 2D image space-time data Then the undersampling operation The corresponding generated undersampling template is divided into 192 rows along the frequency encoding dimension, each row contains the corresponding two-dimensional Fourier space-time undersampling template Finally, by Z m , U m and X m Together they form the training set. The total number of training sets is 1000×192=192000.

[0049] Step 2: Design a deep learning network model based on temporal low-rank and spatial sparse models, the network’s feedback function and loss function

[0050] The deep learning network model is based on the iteration block as the core. By superimposing several iteration blocks as the overall network structure, each iteration block contains three sub-blocks. Taking the kth iteration block as an example, the network structure is explained as follows:

[0051] a) Sub-block R is used to update the network intermediate quantity obtained based on the temporal low-rank prior solution It consists of a 6-layer one-dimensional convolutional neural network with a residual structure, and the convolution kernel size is 3; the first layer input is in represents the final output of the previous iteration block (i.e., the k-1th iteration block), ε n Indicates that the extraction is along The nth vector of the first dimension can also be obtained by the inverse operation Σ represents the summation operation; the 1st to 6th layers are connected by the linear rectified linear unit (ReLU), and the input of each layer is the output of the previous network layer; finally, the final calculation result of the sub-block R is obtained The complete sub-block R is represented by the following nonlinear mapping function:

[0052]

[0053] in, represents the internal parameters of the sub-block R in the kth iteration block, f R (·) represents the nonlinear mapping trained by sub-block R, Represents the 6-layer one-dimensional convolutional neural network in sub-block R;

[0054] b) Sub-block G is used to update the network intermediate quantity obtained based on the spatial sparse prior solution It consists of two 3-layer one-dimensional convolutional neural networks, and the convolution kernel size is 3; the first layer input of the first one-dimensional convolutional neural network is Indicates that the extraction is along The tth vector of the second dimension can also be recovered by the inverse operation Σ represents the summation operation; the 1st to 3rd layers are connected by a linear rectified linear unit (ReLU), and the input of each layer is the output of the previous network layer; the output of the last layer of the first one-dimensional convolutional neural network is soft-thresholded and then input into the first layer of the second one-dimensional convolutional neural network; the 1st to 3rd layers of the second one-dimensional convolutional neural network are connected by a linear rectified linear unit (ReLU), and the input of each layer is the output of the previous network layer; finally, the final calculation result of the sub-block G is obtained. The complete sub-block G is represented by the following nonlinear mapping function:

[0055]

[0056] in, represents the internal parameters of the sub-block G in the kth iteration block, f G (·) represents the nonlinear mapping trained by sub-block G, and denote the first and second 3-layer one-dimensional convolutional neural networks in sub-block G, respectively, and define the soft threshold operation soft as soft(x;θ)=max{|x|-θ}·x / |x|, θ (k) represents the trainable threshold in the kth iteration block;

[0057] c) Sub-block D is used to update the output of the current iterative block obtained based on the data verification operation. Data verification requires that each output must maintain consistency with the sampled data to ensure that the output result is aligned with the collected data. The complete sub-block D is represented by the following function:

[0058]

[0059] in, represents the internal parameters of the sub-block D in the iteration block, f D (·) represents the linear mapping trained by sub-block D. Its specific operation can be expressed as:

[0060]

[0061] in, p represents the data index, Ω represents the position of the sampled data point, and λ1 ,λ 2 Both represent regularization parameters;

[0062] In summary, by cascading the above three sub-blocks R, G and D, the kth iteration block as a whole can be represented by the following nonlinear mapping function group:

[0063]

[0064] Among them, Θ (k) Represents the set of internal parameters of all sub-blocks in the iteration block; F 1 (·) indicates the concatenation of sub-blocks; F 2 (·) indicates the parallel connection of sub-blocks; Indicates the trained arrive The nonlinear mapping is the nonlinear mapping of each sub-block combination of;

[0065] The specific structure of the deep learning network model designed based on temporal low-rank and spatial sparse model is as follows: Figure 2 As shown. It can be expressed as follows:

[0066]

[0067] Among them, Θ represents the set of internal parameters of the overall network; F overall (Z m |Θ) represents the trained multi-coil two-dimensional Fourier space-time data Z from under-sampling and zero-filling m To the final output value of the network All iterative blocks of the nonlinear mapping are cascaded.

[0068] The feedback function of the network is an important process for the network to solve the target value. In the process of network model construction, the output value of the network is Combined coil 2D image space-time data with full sampling X m Compare and feed back the gradient to update the parameters of the iterative module, so that the network output value is closer to the fully sampled artifact-free dynamic magnetic resonance image;

[0069] The loss function of a deep learning network is defined as follows:

[0070]

[0071] Among them, Θ represents the set of internal parameters of the overall network, ||·|| 2 represents the bi-norm term, k represents the kth iteration block, k=1,2,...,K, K represents the total number of iteration blocks, in this embodiment K=10, C represents the total number of training data, in this embodiment C=192000, and Σ represents a summation operation.

[0072] Step 3: Train the optimal parameters of the deep learning network based on temporal low-rank and spatial sparse model

[0073] The Adam optimizer (D. Kingma, J. Ba, "Adam: A method for stochastic optimization," arXiv:1412.6980, 2014.), which performs well in deep learning, is used to minimize the loss function in the second step. Training can obtain the optimal target parameter set

[0074] Step 4: Reconstruct the undersampled magnetic resonance Fourier space-time data to obtain the reconstructed dynamic magnetic resonance image

[0075] The undersampled multi-coil 3D MRI Fourier space-time data to be reconstructed First, do a one-dimensional inverse Fourier transform along the frequency encoding dimension get All rows of form a batch as the input of the trained network, and the network is forward propagated according to the trained relatively optimal parameters, the batch is reconstructed in parallel, and they are stitched together in sequence to obtain the final reconstructed magnetic resonance image. in, The network reconstruction process for each row of can be expressed as:

[0076]

[0077] In this embodiment, the input of the network is 17% sampling rate (Fourier space-time undersampling template as shown in Figure 1 As shown. Figure 1 The under-sampling template shows the sampling point in white, indicating that the data corresponding to the position is sampled; the black point indicates that the data corresponding to the position is lost). The heart full sampling label image under multiple time frames and the heart reconstruction image under multiple time frames of the present invention at a sampling rate of 17% are shown as follows: Figure 3 Figure (a) and (b) in Figure 1. It can be seen that the deep learning dynamic magnetic resonance intelligent imaging method based on temporal low-rank and spatial sparse models can reconstruct high-quality dynamic magnetic resonance images.

[0078] The present invention proposes a deep learning fast magnetic resonance intelligent imaging method that simultaneously utilizes the low rank of multiple time frames of dynamic magnetic resonance data and the sparsity of image space. This spatial-temporal deep learning dynamic magnetic resonance intelligent imaging method based on separable learning has fast reconstruction speed, high reconstruction quality, low memory consumption and small amount of required training data.

[0079] The above embodiments are only preferred embodiments of the present invention and cannot be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A dynamic magnetic resonance intelligent imaging method, Features The following steps are involved: 1) Obtain multi-coil three-dimensional magnetic resonance Fourier space-time data, and establish a training set including fully sampled merged coil two-dimensional image space-time data, undersampled zero-filled multi-coil two-dimensional Fourier space-time data and corresponding two-dimensional Fourier space-time undersampled templates through inverse Fourier transform, Fourier space-time undersampling, data separation and extraction, and coil sensitivity encoding and decoding operations. The specific method is as follows: Acquisition of fully sampled multi-coil 3D MRI Fourier space-time data from MRI instruments and decoding by coil sensitivity and spatial 2D inverse Fourier transform Get fully sampled merged coil 3D image space-time data represents the complex domain, M, N, T represent the length of the frequency encoding dimension, the length of the phase encoding dimension and the number of time frames in the data respectively; Fourier space-time undersampling operation is performed on each coil data in the fully sampled multi-coil three-dimensional magnetic resonance Fourier space-time data Get undersampled zero-filled multi-coil 3D Fourier space-time data Defined as J represents the number of coils of data, represents the spatial two-dimensional Fourier transform, S represents the coil sensitivity code; then Each coil data is transformed along the frequency encoding dimension by a one-dimensional inverse Fourier transform Get undersampled zero-filled multi-coil three-dimensional hybrid space-time data Defined as represents the imaging forward coding operator, represents the one-dimensional Fourier transform along the phase encoding dimension; Divided into M rows along the frequency encoding dimension, each row contains multi-coil 2D Fourier space-time data with under-sampled zero filling Then Divided into M rows along the frequency encoding dimension, each row contains the fully sampled combined coil 2D image space-time data Then the undersampling operation The corresponding generated undersampling template is divided into M rows along the frequency encoding dimension, each row contains the corresponding two-dimensional Fourier space-time undersampling template Finally, by Z m , U m and X m Together they form a training set; 2) Design a deep learning network model based on temporal low-rank and spatial sparse models, the network’s feedback function and loss function; 3) Using the training set obtained in step 1), solve the optimal parameters of the deep learning network based on the temporal low-rank and spatial sparse model; 4) Inputting the undersampled multi-coil three-dimensional magnetic resonance Fourier space-time data to be reconstructed into the trained network to achieve dynamic magnetic resonance image reconstruction.

2. A dynamic magnetic resonance intelligent imaging method as claimed in claim 1, Features In step 2), the deep learning network model takes the iteration block as the core, and stacks several iteration blocks as the overall network structure. Each iteration block contains three sub-blocks; the network structure of a single iteration block is as follows: a) Sub-block R is used to update the network intermediate quantity b obtained based on the temporal low-rank prior solution m,n ; It consists of L containing the residual structure R The convolutional neural network consists of a one-dimensional convolutional neural network with a convolution kernel size of I. R ; The first layer input is in Represents the final output of the previous iteration block, E n Indicates that the extraction is along The nth vector of the first dimension of , or retrieved by the inverse operation Σ represents sum operation; 1st to Lth R The layers are connected by a linear rectified unit (ReLU), and the input of each layer is the output of the previous network layer; finally, the final calculation result b of the sub-block R is obtained m,n ; The complete sub-block R is represented by the following nonlinear mapping function: Among them, Θ R represents the internal parameters of the sub-block R in the iteration block, f R (·) represents the nonlinear mapping trained by sub-block R, Represents L in sub-block R R Layer one-dimensional convolutional neural network; b) Sub-block G is used to update the network intermediate quantity d obtained based on the spatial sparse prior solution m,t ; It consists of two L G The convolutional neural network consists of a one-dimensional convolutional neural network with a convolution kernel size of I. G ; The first layer input of the first one-dimensional convolutional neural network is Indicates that the extraction is along The tth vector of the second dimension of , or the inverse operation to get Σ represents sum operation; 1st to Lth G The layers are connected by a linear rectified unit (ReLU), and the input of each layer is the output of the previous network layer; the output of the last layer of the first one-dimensional convolutional neural network is soft-thresholded and then input into the first layer of the second one-dimensional convolutional neural network; the 1st to Lth layers of the second one-dimensional convolutional neural network are G The layers are connected by a linear rectified unit (ReLU), and the input of each layer is the output of the previous network layer; finally, the final calculation result d of the sub-block G is obtained. m,t ; The complete sub-block G is represented by the following nonlinear mapping function: Among them, Θ G represents the internal parameters of the sub-block G in the iteration block, f G (·) represents the nonlinear mapping trained by sub-block G, and Represent the first and second L in sub-block G respectively G The soft threshold operation soft is defined as soft(x;θ)=max{|x|-θ}·x / |x|, where θ represents the trainable threshold. c) Sub-block D is used to update the output of the current iterative block obtained based on the data verification operation. Data verification requires that each output must maintain consistency with the sampled data to ensure that the output result is aligned with the collected data; the complete sub-block D is represented by the following function: Among them, Θ D represents the internal parameters of the sub-block D in the iteration block, f D (·) represents the linear mapping trained by sub-block D; its specific operation is expressed as: in, p represents the data index, Ω represents the position of the sampled data point, and λ 1 ,λ 2 Both represent regularization parameters; In summary, the above three sub-blocks R, G and D are cascaded, and the single iterative block as a whole is represented by the following nonlinear mapping function group: in, Represents the set of internal parameters of all sub-blocks in the iteration block; F 1 (·) indicates the concatenation of sub-blocks; F 2 (·) indicates the parallel connection of sub-blocks; Indicates the trained arrive The nonlinear mapping is the nonlinear mapping of each sub-block f D (Z m ,b m,n ,d m,t |Θ D ) The overall representation of the designed deep learning network model based on temporal low-rank and spatial sparse model expansion is: Among them, Θ represents the set of internal parameters of the overall network; F overall (Z m |Θ) represents the trained multi-coil two-dimensional Fourier space-time data Z from under-sampling and zero-filling m To the final output value of the network All iterative blocks of the nonlinear mapping are cascaded.

3. A dynamic magnetic resonance intelligent imaging method as claimed in claim 1, Features In step 2), the feedback function of the network is an important process for the network to solve the target value. In the process of building the network model, the output value of the network is Combined coil 2D image space-time data with full sampling X m The comparison is made and the gradient is fed back to update the parameters of the iterative module, so that the network output value is closer to the fully sampled artifact-free dynamic magnetic resonance image.

4. A dynamic magnetic resonance intelligent imaging method as claimed in claim 1, Features In step 2), the loss function is defined as: Among them, Θ represents the set of internal parameters of the overall network, ||·|| 2 represents the bi-norm term, k represents the kth iteration block, k = 1, 2, ..., K, K represents the total number of iteration blocks, C represents the total number of training data, and Σ represents the summation operation.

5. A dynamic magnetic resonance intelligent imaging method as claimed in claim 1, Features In step 3), the optimal parameters of the deep learning network based on the temporal low-rank and spatial sparse model are solved by using the Adam optimizer with good performance in deep learning, and the network is trained using the training set generated in step 1), by minimizing the loss function in step 2). To obtain the optimal target parameter set 6. A dynamic magnetic resonance intelligent imaging method as claimed in claim 1, Features In step 4), the under-sampled multi-coil three-dimensional magnetic resonance Fourier space-time data to be reconstructed is input into the trained network to realize dynamic magnetic resonance image reconstruction; the under-sampled multi-coil three-dimensional magnetic resonance Fourier space-time data to be reconstructed is input into the trained network to realize dynamic magnetic resonance image reconstruction; First, do a one-dimensional inverse Fourier transform along the frequency encoding dimension get All rows of form a batch as the input of the trained network, and the network is forward propagated according to the trained relatively optimal parameters, and the batch is reconstructed in parallel, and they are spliced ​​in sequence to obtain the final reconstructed magnetic resonance image in, The network reconstruction process for each row of is expressed as:

Images