Hyperpolarized gas lung MRI reconstruction method based on self-supervised learning

Abstract

The invention discloses a hyperpolarized gas lung MRI reconstruction method based on self-supervised learning. The method comprises the following steps: firstly, obtaining 2D undersampled hyperpolarized gas MRI k space data as a training set, constructing a combined sampling self-supervision reconstruction network, and then training the combined sampling self-supervision reconstruction network through a set loss function; the self-supervised training process is optimized through a sampling sub-network in the combined sampling self-supervised reconstruction network, and the image reconstructed by noise and artifacts is reduced by using weighted loss, so that the reconstruction quality is further improved. According to the method, only the undersampled magnetic resonance image data is taken as the training set, so that the consumption of human resources during full-sampled magnetic resonance image collection is reduced, and the method can be applied to an actual scene in which high-quality full-sampled magnetic resonance images are difficult to obtain. After the training of the combined sampling self-supervised reconstruction network is completed, additional manual optimization is not needed, and the reconstruction effect is excellent.

Description

Technical Field

[0001] The present invention is based on the field of artificial intelligence technology, and specifically relates to a hyperpolarized gas lung MRI reconstruction method based on self-supervised learning. Background Art

[0002] Hyperpolarized gas Magnetic Resonance Imaging (HPMRI) increases the polarization of inert gas and uses the hyperpolarized gas as a contrast agent, making it possible to directly image the lung cavity structure. However, the longitudinal magnetization vector of the hyperpolarized gas cannot be restored during the sampling process, so long-term imaging is easily affected by noise and artifacts, resulting in a decrease in the image signal-to-noise ratio. In addition, most hyperpolarized gas MRI methods need to be completed within one breath-hold, and a longer breath-hold time will place a great burden on the subject. Therefore, the development of fast imaging technology is of great significance to hyperpolarized gas MRI.

[0003] Compressed sensing (CS) reconstruction technology uses the sparsity of signals to collect unrelated signals at a sampling rate lower than the Nyquist rate, and then accurately reconstructs the original signal through a nonlinear reconstruction algorithm. Ajraoui et al. first applied compressed sensing CS to hyperpolarization. 3 He lung MRI. However, the CS method has some disadvantages, such as the difficulty in adjusting parameters, the inadequacy of commonly used sparse transforms in describing complex biological tissue structures, and the long calculation time of nonlinear reconstruction algorithms.

[0004] Deep learning reconstruction methods show great potential in accelerating hyperpolarized gas MRI. Duan et al. used deep learning methods for the first time to reconstruct hyperpolarized gas MRI. They used a cascaded U-Net model combined with the lung contour prior information of proton images to quickly and accurately reconstruct undersampled lungs. 129 Xe images. They also used a cascaded dense residual network to accelerate multi-b-value hyperpolarized gas diffusion MRI. However, these methods are all supervised training and require high-quality fully sampled data. However, due to the influence of noise and artifacts, high-quality fully sampled data is difficult to obtain, which limits the further application of deep learning methods in hyperpolarized gas MRI reconstruction.

[0005] Researchers have proposed some reconstruction methods to get rid of the dependence on full-sampled data: Gan et al. proposed a reconstruction network based on the Noise2noise principle. This method collects two undersampled images of the same target and learns from each other. However, collecting two images of the same target is not easy to achieve in practical applications. Yaman et al. proposed a physically guided self-supervised training network. This method is based on the Noiser2noise principle and divides the undersampled k-space data into two subsets, one for training and one for calculating the loss. The effectiveness of the method was proved by experiments on public data sets. However, none of these methods can be directly applied to hyperpolarized gas reconstruction. They do not consider the impact of noise and artifacts on training during reconstruction, and use a fixed k-space division mode in the self-supervised training process without optimization for different data. The final reconstruction result may not be optimal. In addition, commonly used undersampling modes, such as variable density sampling, will densely sample the center of k-space and sparsely sample the edge of k-space, resulting in the network over-expressing the central part of k-space during the self-supervision process, which makes the center of k-space more sensitive to noise and artifacts.

[0006] In summary, the prior art has the following problems:

[0007] 1. Most current self-supervised reconstruction techniques use a fixed k-space undersampling mode without considering the impact of the undersampling mode on reconstruction;

[0008] 2. The current deep learning reconstruction technology of hyperpolarized gas MRI does not consider the impact of noise and artifacts on training;

[0009] 3. Currently, deep learning reconstruction techniques for hyperpolarized gas MRI all use supervised training, but high-quality hyperpolarized gas data are difficult to obtain. Summary of the invention

[0010] The present invention aims at the above-mentioned problems existing in the prior art and provides a hyperpolarized gas lung MRI reconstruction method based on self-supervised learning.

[0011] The above-mentioned object of the present invention is achieved by the following technical solutions:

[0012] The hyperpolarized gas lung MRI reconstruction method based on self-supervised learning includes the following steps:

[0013] Step 1, obtaining a training set, the training set comprising a plurality of sample groups, each sample group comprising 2D undersampled hyperpolarized gas MRI k-space data;

[0014] Step 2: construct a joint sampling self-supervised reconstruction network, which includes a sampling subnetwork and a reconstruction subnetwork;

[0015] The preset undersampling matrix is ​​divided into a training subset sampling matrix Θ and a loss subset sampling matrix Λ through a sampling subnetwork; the 2D undersampling hyperpolarized gas MRI k-space data are sampled respectively according to the training subset sampling matrix Θ and the loss subset sampling matrix Λ, and a training subset and a loss subset of the 2D undersampling hyperpolarized gas MRI k-space data are correspondingly obtained, and the training subset of the 2D undersampling hyperpolarized gas MRI k-space data is input into a reconstruction subnetwork to generate a reconstructed image;

[0016] Step 3, setting a loss function Loss, where the loss function Loss is based on a loss subset of the reconstructed image and the 2D undersampled hyperpolarized gas MRI k-space data;

[0017] Step 4: According to the loss function Loss set in step 3, the joint sampling self-supervised reconstruction network constructed in step 2 is trained end-to-end;

[0018] Step 5: Input the 2D undersampled hyperpolarized gas MRI k-space data to be processed into the reconstruction subnetwork of the joint sampling self-supervised reconstruction network trained in step 4 to obtain the final reconstructed hyperpolarized gas MRI image.

[0019] As described above, the sampling subnetwork includes, in order, a sampling weight matrix layer, an activation layer, a normalization layer, a random matrix transformation layer, a binarization layer, and an undersampling matrix partitioning module;

[0020] The sampling weight matrix layer outputs the sampling weight matrix to the activation layer of the sampling sub-network. The output P of the activation layer of the sampling sub-network is then input to the normalization layer. The output of the normalization layer is input to the random matrix transformation layer. The random matrix transformation layer subtracts the output of the normalization layer from a random matrix with a uniform distribution of element values ​​in the range of [0,1]. The output of the random matrix transformation layer is then input to the binarization layer, which outputs the sampling mask.

[0021] The sampling mask, the preset under-sampling matrix and the 2D under-sampling hyperpolarized gas MRI k-space data are all input into the under-sampling matrix partitioning module. The under-sampling matrix partitioning module divides the preset under-sampling matrix into a training subset sampling matrix Θ and a loss subset sampling matrix Λ. The 2D under-sampling hyperpolarized gas MRI k-space data are sampled according to the training subset sampling matrix Θ and the loss subset sampling matrix Λ, respectively, to obtain the training subset and the loss subset of the 2D under-sampling hyperpolarized gas MRI k-space data.

[0022] As mentioned above, the output of the normalization layer is:

[0023]

[0024] Among them, N α(P) is the output of the normalization layer, P is the output of the activation layer of the sampling subnetwork, P is in matrix form; α is the set sampling rate; is the mean of all elements in the output of the activation layer of the sampling subnetwork.

[0025] As mentioned above, the binarization layer uses a sigmoid-like function B(x):

[0026]

[0027] λ is the function parameter and x is the input of the binarization layer.

[0028] As described above, the sampling weight matrix layer outputs a sampling weight matrix, and the random matrix transformation layer outputs a random matrix. Both the sampling weight matrix and the random matrix have the same size as the 2D undersampled hyperpolarized gas MRI k-space data.

[0029] As described above, the reconstruction subnetwork includes a two-dimensional inverse Fourier transform module and a plurality of cascaded residual networks, each of which sequentially cascades an input layer, a plurality of residual modules, a first residual network convolutional layer, a second residual network convolutional layer, and a data consistency layer;

[0030] The input layer includes an input layer convolution layer and an input layer activation layer. The input of the input layer is added to the output of the first residual network convolution layer through a residual connection and then input into the second residual network convolution layer.

[0031] The residual module includes a first residual module convolution layer, a residual module activation layer, a second residual module convolution layer and a constant multiplication layer. The input of the first residual module convolution layer and the output of the constant multiplication layer are residually connected and added as the output of the residual module.

[0032] As described above, the input of the input layer of the first residual network in the reconstruction subnetwork is the training subset image; the input of the input layer of the remaining residual networks in the reconstruction subnetwork except the first residual network is the output of the previous residual network; the input of the data consistency layer of each residual network in the reconstruction subnetwork also includes the training subset of 2D undersampled hyperpolarized gas MRI k-space data;

[0033] In the data consistency layer, the feature image input to the data consistency layer is first subjected to a two-dimensional Fourier transform and then subjected to a consistency process. The result of the consistency process is then subjected to a two-dimensional inverse Fourier transform and then output. The feature image output by the data consistency layer is the feature image output by the residual network. The consistency process is as follows:

[0034]

[0035] x recis the result of the uniformity processing, that is, the k-space data corresponding to the feature image output by the residual network where the data uniformity layer is located; x cnn represents the k-space data obtained by two-dimensional Fourier transform of the feature image input to the data consistency layer, x0 is the training subset of 2D undersampled hyperpolarized gas MRI k-space data, and x0 is output by the sampling subnetwork; j is the k-space sampling point, ω Θ is the set of all sampling points contained in the training subset sampling matrix, with coefficient β∈[0,+∞].

[0036] As mentioned above, the loss function Loss is set as:

[0037] Loss = W × L (y Λ ,E Λ (f θ (y Θ ,E Θ )))

[0038] Among them, × represents multiplication, L is the error loss calculation function, y Θ is the training subset of 2D undersampled hyperpolarized gas MRI k-space data, y Λ is the loss subset of 2D undersampled hyperpolarized gas MRI k-space data, Θ is the training subset sampling matrix, Λ is the loss subset sampling matrix, Ω=Θ∪Λ and Θ∩Λ={}, Ω is the preset undersampling matrix; the encoding operator E Θ It is obtained by two-dimensional Fourier transform and the corresponding training subset sampling matrix Θ, that is, first perform two-dimensional Fourier transform, and multiply the result of two-dimensional Fourier transform and the training subset sampling matrix Θ; the encoding operator E Λ It is obtained by two-dimensional Fourier transform and loss subset sampling matrix Λ, that is, two-dimensional Fourier transform is first performed, and the result of two-dimensional Fourier transform is multiplied by loss subset sampling matrix Λ; The parameters are The reconstructed subnetwork; W is the weight matrix of the loss function.

[0039] As described above, the loss function weight matrix W is updated in each training iteration as follows: the sampling frequency of each sampling point in the k-space is updated according to the sampling points contained in the training subset sampling matrix Θ in each training iteration, and then the inverse of the updated sampling frequency of each sampling point is used as the value of the corresponding sampling point in the loss function weight matrix W.

[0040] In step 4 as described above, the sampling weight matrix of the first round of iterative training is randomly initialized; in each round of iterative training except the first round of iterative training, the sampling weight matrix obtained after the previous round of iterative training is used as the sampling weight matrix of the current round of iterative training.

[0041] Compared with the prior art, the present invention has the following advantages:

[0042] 1. The present invention optimizes the partitioning strategy of under-sampling k-space in the self-supervised training process by building a sampling subnetwork, and adaptively obtains the best partitioning strategy according to the characteristics of the data, thereby improving the reconstruction quality.

[0043] 2. The present invention improves the robustness of the model to noise and artifacts by using weighted self-supervised training loss. In undersampling, the center of the k-space is often oversampled, making the center more sensitive to noise and artifacts. Therefore, improving the robustness of the model to noise and artifacts can further improve the reconstruction quality.

[0044] 3. The present invention builds a joint sampling self-supervised reconstruction network, and uses the set self-supervised loss function to train the joint sampling self-supervised reconstruction network in the absence of high-quality fully sampled magnetic resonance images, and can ensure reconstruction performance. Compared with the existing magnetic resonance imaging reconstruction technology based on deep learning, the present invention only requires 2D under-sampled hyperpolarized gas MRI k-space data as a training set, reducing the human resources consumed in collecting high-quality fully sampled magnetic resonance images, and the present invention can be applied to scenes where it is difficult to obtain high-quality fully sampled magnetic resonance images, and no additional optimization is required during application, which is of use value. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of the present invention;

[0046] Figure 2a The overall structure of the network is reconstructed for joint sampling and self-supervision;

[0047] Figure 2b It is a schematic diagram of the structure of the sampling sub-network;

[0048] Figure 2c Schematic diagram of the structure of the residual module in the reconstruction sub-network;

[0049] Figure 3 Comparison of the results obtained by the present invention and different self-supervised reconstruction methods. DETAILED DESCRIPTION

[0050] In order to facilitate ordinary technicians in this field to understand and implement the present invention, the present invention is further described in detail below in conjunction with the embodiments and drawings. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention, and do not limit the present invention.

[0051] The hyperpolarized gas lung MRI reconstruction method based on self-supervised learning specifically includes the following steps:

[0052] Step 1: Obtain a training set and a test set, wherein both the training set and the test set include multiple sample groups, each sample group includes 2D under-sampled hyperpolarized gas MRI k-space data. In this embodiment, in order to facilitate subsequent testing and verification of the reconstruction effect, each sample group also includes a 2D fully sampled hyperpolarized gas MRI image corresponding to the 2D under-sampled hyperpolarized gas MRI k-space data; the 2D fully sampled hyperpolarized gas MRI image uses a 2D fully sampled hyperpolarized gas MRI image. 129 Xe MRI image, 2D undersampled hyperpolarization 129 XeMRI k-space data using 2D undersampling hyperpolarization 129 Xe MRI k-space data, specifically including the following steps:

[0053] Step 1.1, scan multiple subjects to obtain corresponding 3D full-sampling hyperpolarized gas MRI k-space data.

[0054] Step 1.2, extracting 3D fully sampled hyperpolarized gas MRI k-space data layer by layer to obtain multiple 2D fully sampled hyperpolarized gas MRI k-space data, and undersampling the 2D fully sampled hyperpolarized gas MRI k-space data according to a preset undersampling matrix to obtain corresponding 2D undersampling hyperpolarized gas MRI k-space data.

[0055] Step 1.3, performing a two-dimensional inverse Fourier transform on the 2D full-sampling hyperpolarized gas MRI k-space data to obtain a 2D full-sampling hyperpolarized gas MRI image.

[0056] Step 1.4: Take a subject's 2D fully sampled hyperpolarized gas MRI image and the corresponding 2D under-sampled hyperpolarized gas MRI k-space data as a sample group, and divide all sample groups into training sets and test sets, and the training set and the test set both include multiple sample groups.

[0057] In this implementation case, 94 subjects were scanned using a 3D bSSFP sequence to obtain a set of 3D fully sampled hyperpolarized gas MRI k-space data. The matrix size of the 3D fully sampled hyperpolarized gas MRI k-space data is 96×84, and the number of layers is 24. The undersampling mode is variable density undersampling, and the undersampling acceleration factor is 4. The 3D fully sampled hyperpolarized gas MRI k-space data is extracted layer by layer to obtain 2D fully sampled hyperpolarized gas MRI k-space data. The 2D fully sampled hyperpolarized gas MRI k-space data is undersampled according to the undersampling matrix to obtain the corresponding 2D undersampled hyperpolarized gas MRI k-space data. A two-dimensional inverse Fourier transform is performed on the 2D fully sampled 2D hyperpolarized gas MRI k-space data to obtain a 2D fully sampled hyperpolarized gas MRI image. A sample group of 84 subjects was randomly selected as a training set, and a sample group of 10 subjects was selected as a test set.

[0058] In this example, 2D fully sampled hyperpolarized gas MRI k-space data is collected and retrospectively undersampled to obtain 2D undersampled hyperpolarized gas MRI k-space data. During the application process, prospective 2D undersampled hyperpolarized gas MRI k-space data can also be collected for reconstruction. The 2D fully sampled hyperpolarized gas MRI image is used in this implementation case to calculate the evaluation index with the reconstructed image to evaluate the effect of the present invention.

[0059] Step 2: Construct a joint sampling self-supervised reconstruction network, which includes 1 sampling sub-network and 1 reconstruction sub-network.

[0060] The 2D under-sampled hyperpolarized gas MRI k-space data in the sample group and the preset under-sampling matrix are input into a joint sampling self-supervised reconstruction network. In the sampling subnetwork of the joint sampling self-supervised reconstruction network, the preset under-sampling matrix is ​​divided into a training subset sampling matrix Θ and a loss subset sampling matrix Λ. The 2D under-sampled hyperpolarized gas MRI k-space data are sampled according to the training subset sampling matrix Θ and the loss subset sampling matrix Λ, respectively, to obtain a training subset and a loss subset of the 2D under-sampled hyperpolarized gas MRI k-space data correspondingly. The training subset of the 2D under-sampled hyperpolarized gas MRI k-space data is input into the reconstruction subnetwork to generate a reconstructed image. The reconstructed image is subjected to a two-dimensional Fourier transform to obtain the reconstructed k-space data, and the loss function is calculated based on the reconstructed k-space data and the loss subset of the 2D under-sampled hyperpolarized gas MRI k-space data.

[0061] (1) Sampling subnetwork

[0062] The sampling subnetwork includes a sampling weight matrix layer, an activation layer, a normalization layer, a random matrix transformation layer, a binarization layer, and an undersampling matrix partitioning module in order. In the sampling subnetwork: the sampling weight matrix output by the sampling weight matrix layer and the random matrix output by the random matrix transformation layer have the same size as the 2D undersampling hyperpolarized gas MRI k-space data.

[0063] In this example, the sampling subnetwork includes a sampling weight matrix layer, an activation layer, a normalization layer, a random matrix transformation layer, a binarization layer, and an undersampling matrix partitioning module that are cascaded in sequence. The sampling weight matrix output by the sampling weight matrix layer is 96×84 in size, and the activation layer of the sampling subnetwork uses a sigmoid activation function. The sampling weight matrix of the sampling weight matrix layer is randomly initialized, and then the sampling weight matrix is ​​input to the activation layer of the sampling subnetwork. The output P of the activation layer of the sampling subnetwork is then input to the normalization layer. The output of the normalization layer is input to the random matrix transformation layer. The random matrix transformation layer subtracts a random matrix with a uniform distribution of element values ​​in the range of [0,1] from the output of the normalization layer. Finally, the output of the random matrix transformation layer is input to the binarization layer. The binarization layer outputs a sampling mask, which is used to divide the preset undersampling matrix into a training subset sampling matrix and a loss subset sampling matrix.

[0064] Among them, the normalization layer is used to ensure that the sampling rate of the sampling subnetwork is consistent with the set sampling rate:

[0065]

[0066] Among them, N α (P) is the output of the normalization layer, P is the output of the activation layer of the sampling subnetwork, P is in matrix form; α is the set sampling rate; is the mean of all elements in the output of the activation layer of the sampling subnetwork. Through the normalization layer, all elements in the output P of the activation layer of the sampling subnetwork are processed and limited to [0,1], and the mean of all elements in the output of the normalization layer is limited to α.

[0067] In this embodiment, the binarization layer uses a sigmoid-like function B(x):

[0068]

[0069] λ is a function parameter, and x is the input of the binarization layer. The binarization layer will continuously optimize the learnable parameters (such as the function parameter λ in this embodiment) during the training process to fit the non-learnable transition function (such as the sigmoid-like function B(x) in this embodiment). The function parameter λ in the function used by the binarization layer is learned through the network to ensure that the transition function is fitted while having the ability to learn.

[0070] The sampling mask output by the binarization layer, the undersampling matrix preset in step 1.2, and the 2D undersampling hyperpolarized gas MRI k-space data are all input into the undersampling matrix partitioning module. The undersampling matrix partitioning module divides the undersampling matrix preset in step 1.2 into a training subset sampling matrix Θ and a loss subset sampling matrix Λ. The 2D undersampling hyperpolarized gas MRI k-space data are sampled according to the training subset sampling matrix Θ to obtain the training subset of the 2D undersampling hyperpolarized gas MRI k-space data; the loss subset sampling matrix Λ samples the 2D undersampling hyperpolarized gas MRI k-space data to obtain the loss subset of the 2D undersampling hyperpolarized gas MRI k-space data.

[0071] (2) Reconstructing the subnetwork

[0072] The reconstruction subnetwork includes a two-dimensional inverse Fourier transform module and multiple cascaded residual networks. The training subset of 2D undersampled hyperpolarized gas MRI k-space data is input into the two-dimensional inverse Fourier transform module. After two-dimensional inverse Fourier transform, the training subset image is obtained and input into the first residual network.

[0073] In each residual network, an input layer, multiple residual modules, two residual network convolutional layers (i.e., the first residual network convolutional layer and the second residual network convolutional layer) and a data consistency layer are cascaded in sequence; the input layer includes an input layer convolutional layer and an input layer activation layer; the residual module includes a first residual module convolutional layer, a residual module activation layer, a second residual module convolutional layer and a constant multiplication layer.

[0074] The input of the input layer of the first residual network in the reconstruction subnetwork is the training subset image; the input of the input layer of the remaining residual networks in the reconstruction subnetwork except the first residual network is the output of the previous residual network;

[0075] The input to the data consistency layer of each residual network in the reconstruction subnetwork also includes a training subset of 2D undersampled hyperpolarized gas MRI k-space data.

[0076] In each residual network: the input of the input layer of each residual network and the output of the first residual network convolutional layer of the residual network are added through a residual connection, and then input into the second residual network convolutional layer; the number of input channels of each residual module, the number of output channels of each residual module, the number of output channels of the first residual network convolutional layer, the number of input channels of the second residual network convolutional layer, and the number of output channels of the second residual network convolutional layer are all equal to the number of input channels of the input layer of the residual network.

[0077] In this implementation case, the reconstruction subnetwork includes 5 residual networks, each of which includes 1 input layer, 10 residual modules, a first residual network convolution layer, a second residual network convolution layer and 1 data consistency layer cascaded in sequence:

[0078] The input layer of each residual network includes an input convolution layer and an input activation layer. The convolution kernel size of the input convolution layer is 3×3, the number of input channels is 2, the number of output channels is 64, the input feature map size is 96×84, and the output feature map size is 96×84. The activation function of the input activation layer is the ReLU function.

[0079] Each residual module includes a first residual module convolution layer, a residual module activation layer, a second residual module convolution layer and a constant multiplication layer cascaded in sequence. The input of the first residual module convolution layer and the output of the constant multiplication layer are residually connected and added as the output of the residual module. In this example, in each residual module: the convolution kernel size of the first residual module convolution layer is 3×3, the number of input channels is 64, and the number of output channels is 64; the convolution kernel size of the second residual module convolution layer is 3×3, the number of input channels is 64, and the number of output channels is 2; the output of the first residual module convolution layer is input to the residual module activation layer, and the output of the second residual module convolution layer is input to the constant multiplication layer; the constant multiplication layer multiplies the input of the constant multiplication layer by a constant to adjust the size of the input of the constant multiplication layer, and the output of the constant multiplication layer is the output of the residual module; the activation function of the residual module activation layer is the ReLU function.

[0080] The data consistency layer of the residual network is used to ensure that the corresponding position of the k-space data corresponding to the feature image output by the residual network is consistent with the corresponding position of the 2D undersampled hyperpolarized gas MRI k-space data. In the data consistency layer, the feature image input by the data consistency layer is first subjected to a two-dimensional Fourier transform, and then subjected to a consistency process. The result of the consistency process is then subjected to a two-dimensional inverse Fourier transform and then output. The feature image output by the data consistency layer is the feature image output by the residual network. The consistency process is as follows:

[0081]

[0082] x rec is the result of the uniformity processing, that is, the k-space data corresponding to the feature image output by the residual network where the data uniformity layer is located; x cnn represents the k-space data obtained by two-dimensional Fourier transform of the feature image input to the data consistency layer, x0 is the training subset of 2D undersampled hyperpolarized gas MRI k-space data, and x0 is output by the sampling subnetwork; j is the k-space sampling point, ω Θ is the set of all sampling points contained in the training subset sampling matrix, ω ΘCorresponding to x0, the coefficient β∈[0,+∞] is used to control the composition ratio of the 2D undersampled MRI k-space data corresponding to the training subset image input by the reconstruction subnetwork to the k-space corresponding to the feature image output by the residual network. In this embodiment, the coefficient β takes infinity, that is, in the set of all sampling points ω contained in the training subset sampling matrix Θ Replace x with x0 at the sampling point in cnn .

[0083] Step 3, set the loss function. The loss function is the weighted self-supervised training reconstruction loss. The training subset of the obtained 2D under-sampled hyperpolarized gas MRI k-space data is transformed into the image domain through two-dimensional inverse Fourier transform and then input into the reconstruction subnetwork to obtain the reconstructed image. The reconstructed image is transformed from the image domain to the k-space domain through two-dimensional Fourier transform, multiplied by the loss subset sampling matrix Λ, and then the mean square error is calculated with the loss subset of the 2D under-sampled hyperpolarized gas MRI k-space data. The loss function Loss is set as:

[0084]

[0085] Where × represents multiplication; L is the error loss calculation function, such as the commonly used mean square error or mean absolute error; y Θ is the training subset of 2D undersampled hyperpolarized gas MRI k-space data, y Λ is the loss subset of 2D undersampled hyperpolarized gas MRI k-space data, Θ is the training subset sampling matrix, Λ is the loss subset sampling matrix, Ω=Θ∪Λ and Θ∩Λ={}, Ω is the preset undersampling matrix; E Θ With E Λ are all encoding operators, E Θ It is obtained by two-dimensional Fourier transform and the corresponding training subset sampling matrix Θ, that is, first perform two-dimensional Fourier transform, and multiply the result of two-dimensional Fourier transform and the training subset sampling matrix Θ, which is characterized by performing two-dimensional Fourier transform on the corresponding position of the training subset sampling matrix Θ; E Λ It is obtained by two-dimensional Fourier transform and loss subset sampling matrix Λ, that is, two-dimensional Fourier transform is performed first, and the result of two-dimensional Fourier transform is multiplied by loss subset sampling matrix Λ, which is characterized by two-dimensional Fourier transform of the corresponding position of loss subset sampling matrix Λ; W is the loss function weight matrix, The parameters are In this example, L selects the mean square error, and the loss function weight matrix W is updated in each training iteration as follows: the sampling frequency of each sampling point in the k-space is updated according to the sampling points contained in the training subset sampling matrix Θ in each training iteration, and then the inverse of the updated sampling frequency of each sampling point is used as the value of the corresponding sampling point in the loss function weight matrix W.

[0086] Step 4: According to the loss function Loss set in step 3, the joint sampling self-supervised reconstruction network constructed in step 2 is trained end-to-end using the training set generated in step 1.

[0087] In each round of training iteration, the 2D under-sampled hyperpolarized gas MRI k-space data in the training set generated in step 1 and the preset under-sampling matrix are input into the joint sampling self-supervised reconstruction network constructed in step 2. The sampling subnetwork samples the preset under-sampling matrix to generate the training subset sampling matrix Θ (sampled part) and the loss subset sampling matrix Λ (unsampled part). The 2D under-sampled hyperpolarized gas MRI k-space data are sampled according to the training subset sampling matrix Θ and the loss subset sampling matrix Λ, respectively, to obtain the training subset and loss subset of the 2D under-sampled hyperpolarized gas MRI k-space data. The training subset of the 2D under-sampled hyperpolarized gas MRI k-space data is input into the reconstruction subnetwork to obtain a reconstructed image.

[0088] In the first round of iterative training, the sampling weight matrix O is randomly initialized; except for the first round of iterative training, each round of iterative training uses the sampling weight matrix obtained after the previous round of iterative training as the sampling weight matrix of the current round of iterative training. The undersampling matrix input to the sampling subnetwork is the same in each round.

[0089] In this embodiment, the loss function is calculated in each round of iteration: the reconstructed image output by the reconstruction subnetwork is subjected to a two-dimensional Fourier transform to obtain reconstructed 2D hyperpolarized gas MRI k-space data, the reconstructed 2D hyperpolarized gas MRI k-space data is multiplied by the loss subset sampling matrix Λ to obtain the loss subset corresponding to the reconstructed k-space, and the loss subset corresponding to the reconstructed k-space is multiplied by the loss subset y of the original 2D under-sampled hyperpolarized gas MRI k-space data. Λ Calculate the mean square error and multiply it by the loss function weight matrix W.

[0090] In this example, the learning rate of the network is initialized to 0.0001, the batch size is set to 8, and the network is trained on the PyTorch platform using the Adam optimizer. The training is stopped after the total number of iterations reaches 200, and the parameters of the corresponding joint sampling self-supervised reconstruction network are saved.

[0091] Step 5: Input the 2D under-sampled hyperpolarized gas MRI k-space data to be processed into the reconstruction sub-network trained in step 4 to obtain the final reconstructed hyperpolarized gas MRI image. In this example, the 2D under-sampled hyperpolarized gas MRI k-space data in the test set generated in step 1 is input into the reconstruction sub-network of the joint sampling self-supervised reconstruction network trained in step 4 to obtain the final reconstructed hyperpolarized gas MRI image for the test. In this step, the sampling sub-network will be frozen. All 2D under-sampled hyperpolarized gas MRI k-space data will be input into the reconstruction sub-network.

[0092] In this embodiment, the final reconstructed hyperpolarized gas MRI image of the test obtained in step 5 is compared with the 2D fully sampled hyperpolarized gas MRI image of the test set obtained in step 1, and the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) indicators are calculated.

[0093] Figure 3 This is a comparison between the present invention and different self-supervised reconstruction methods. Zero-filling is an under-sampled zero-filled image, A2A and SSDU are two self-supervised reconstruction algorithms, Ours is the method of the present invention, and Superivsed is the supervised training of the network of the method of the present invention. The arrows mark the differences in the reconstruction results of different methods. The PSNR of under-sampled zero-filling Zero-filling is 24.40, and the SSIM is 0.772; the PSNR of the self-supervised reconstruction algorithm A2A is 27.09, and the SSIM is 0.840; the PSNR of the self-supervised reconstruction algorithm SSDU method is 30.22, and the SSIM is 0.879; the PSNR of the method of the present invention is 30.30, and the SSIM is 0.890; the PSNR of Supervised is 30.49, and the SSIM is 0.895. It can be seen that compared with other self-supervised reconstruction methods, the present invention can better restore image details and obtain results comparable to those of supervised reconstruction methods.

[0094] Example 2

[0095] Hyperpolarized gas lung MRI reconstruction device based on self-supervised learning, including:

[0096] A network construction module, used to implement step 2 described in embodiment 1;

[0097] A loss function construction module, used to implement step 3 described in Example 1;

[0098] A training module, used to implement step 4 described in Example 1;

[0099] The application module is used to implement step 5 described in Example 1.

[0100] Example 3

[0101] A computer device includes a memory and a processor, wherein a computer program is stored in the memory, and steps 2-5 in the above embodiment 1 are implemented when the processor executes the computer program.

[0102] Example 4

[0103] A computer-readable storage medium stores a computer program, which implements steps 2-5 in the above-mentioned embodiment 1 when executed by a processor.

[0104] Example 5

[0105] A computer program product includes a computer program, and when the computer program is executed by a processor, steps 2-5 in the above embodiment 1 are implemented.

[0106] The specific implementation examples described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific implementation examples described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A hyperpolarized gas lung MRI reconstruction method based on self-supervised learning, characterized in that: The steps include: Step 1, obtaining a training set, the training set comprising a plurality of sample groups, each sample group comprising 2D undersampled hyperpolarized gas MRI k-space data; Step 2: construct a joint sampling self-supervised reconstruction network, which includes a sampling subnetwork and a reconstruction subnetwork; The preset undersampling matrix is ​​divided into a training subset sampling matrix Θ and a loss subset sampling matrix Λ through a sampling subnetwork; the 2D undersampling hyperpolarized gas MRI k-space data are sampled respectively according to the training subset sampling matrix Θ and the loss subset sampling matrix Λ, and a training subset and a loss subset of the 2D undersampling hyperpolarized gas MRI k-space data are correspondingly obtained, and the training subset of the 2D undersampling hyperpolarized gas MRI k-space data is input into a reconstruction subnetwork to generate a reconstructed image; Step 3, setting a loss function Loss, where the loss function Loss is based on a loss subset of the reconstructed image and the 2D undersampled hyperpolarized gas MRI k-space data; Step 4: According to the loss function Loss set in step 3, the joint sampling self-supervised reconstruction network constructed in step 2 is trained end-to-end; Step 5: Input the 2D undersampled hyperpolarized gas MRI k-space data to be processed into the reconstruction subnetwork of the joint sampling self-supervised reconstruction network trained in step 4 to obtain the final reconstructed hyperpolarized gas MRI image.

2. The hyperpolarized gas lung MRI reconstruction method based on self-supervised learning according to claim 1, characterized in that: The sampling subnetwork includes, in order, a sampling weight matrix layer, an activation layer, a normalization layer, a random matrix transformation layer, a binarization layer, and an undersampling matrix partitioning module; The sampling weight matrix layer outputs the sampling weight matrix to the activation layer of the sampling sub-network. The output P of the activation layer of the sampling sub-network is then input to the normalization layer. The output of the normalization layer is input to the random matrix transformation layer. The random matrix transformation layer subtracts the output of the normalization layer from a random matrix with a uniform distribution of element values ​​in the range of [0,1]. The output of the random matrix transformation layer is then input to the binarization layer, which outputs the sampling mask. The sampling mask, the preset under-sampling matrix and the 2D under-sampling hyperpolarized gas MRI k-space data are all input into the under-sampling matrix partitioning module. The under-sampling matrix partitioning module divides the preset under-sampling matrix into a training subset sampling matrix Θ and a loss subset sampling matrix Λ. The 2D under-sampling hyperpolarized gas MRI k-space data are sampled according to the training subset sampling matrix Θ and the loss subset sampling matrix Λ, respectively, to obtain the training subset and the loss subset of the 2D under-sampling hyperpolarized gas MRI k-space data.

3. The hyperpolarized gas lung MRI reconstruction method based on self-supervised learning according to claim 2, characterized in that: The output of the normalization layer is: Among them, N α (P) is the output of the normalization layer, P is the output of the activation layer of the sampling subnetwork, P is in matrix form; α is the set sampling rate; is the mean of all elements in the output of the activation layer of the sampling subnetwork.

4. The hyperpolarized gas lung MRI reconstruction method based on self-supervised learning according to claim 2, characterized in that: The binarization layer uses a sigmoid-like function B(x): λ is the function parameter and x is the input of the binarization layer.

5. The hyperpolarized gas lung MRI reconstruction method based on self-supervised learning according to claim 2, characterized in that: The sampling weight matrix layer outputs a sampling weight matrix, and the random matrix transformation layer outputs a random matrix. The sampling weight matrix and the random matrix are both the same size as the 2D under-sampled hyperpolarized gas MRI k-space data.

6. The hyperpolarized gas lung MRI reconstruction method based on self-supervised learning according to claim 1, characterized in that: The reconstruction subnetwork includes a two-dimensional inverse Fourier transform module and a plurality of cascaded residual networks, each of which sequentially cascades an input layer, a plurality of residual modules, a first residual network convolutional layer, a second residual network convolutional layer, and a data consistency layer; The input layer includes an input layer convolution layer and an input layer activation layer. The input of the input layer is added to the output of the first residual network convolution layer through a residual connection and then input into the second residual network convolution layer. The residual module includes a first residual module convolution layer, a residual module activation layer, a second residual module convolution layer and a constant multiplication layer. The input of the first residual module convolution layer and the output of the constant multiplication layer are residually connected and added as the output of the residual module.

7. The hyperpolarized gas lung MRI reconstruction method based on self-supervised learning according to claim 6, characterized in that: The input of the input layer of the first residual network in the reconstruction subnetwork is the training subset image; the input of the input layer of the remaining residual networks in the reconstruction subnetwork except the first residual network is the output of the previous residual network; the input of the data consistency layer of each residual network in the reconstruction subnetwork also includes the training subset of 2D undersampled hyperpolarized gas MRI k-space data; In the data consistency layer, the feature image input to the data consistency layer is first subjected to a two-dimensional Fourier transform and then subjected to a consistency process. The result of the consistency process is then subjected to a two-dimensional inverse Fourier transform and then output. The feature image output by the data consistency layer is the feature image output by the residual network. The consistency process is as follows: x rec is the result of the uniformity processing, that is, the k-space data corresponding to the feature image output by the residual network where the data uniformity layer is located; x cnn represents the k-space data obtained by two-dimensional Fourier transform of the feature image input to the data consistency layer, x0 is the training subset of 2D undersampled hyperpolarized gas MRI k-space data, and x0 is output by the sampling subnetwork; j is the k-space sampling point, ω Θ is the set of all sampling points contained in the training subset sampling matrix, with coefficient β∈[0,+∞].

8. The hyperpolarized gas lung MRI reconstruction method based on self-supervised learning according to claim 1, characterized in that: The loss function Loss is set as: Loss=W×L(y Λ ,E Λ (f θ (y Θ ,E Θ ))) Among them, × represents multiplication, L is the error loss calculation function, y Θ is the training subset of 2D undersampled hyperpolarized gas MRI k-space data, y Λ is the loss subset of 2D undersampled hyperpolarized gas MRI k-space data, Θ is the training subset sampling matrix, Λ is the loss subset sampling matrix, Ω=Θ∪Λ and Θ∩Λ={}, Ω is the preset undersampling matrix; the encoding operator E Θ It is obtained by two-dimensional Fourier transform and the corresponding training subset sampling matrix Θ, that is, first perform two-dimensional Fourier transform, and multiply the result of two-dimensional Fourier transform and the training subset sampling matrix Θ; the encoding operator E Λ It is obtained by two-dimensional Fourier transform and loss subset sampling matrix Λ, that is, two-dimensional Fourier transform is first performed, and the result of two-dimensional Fourier transform is multiplied by loss subset sampling matrix Λ; The parameters are The reconstructed subnetwork; W is the weight matrix of the loss function.

9. The hyperpolarized gas lung MRI reconstruction method based on self-supervised learning according to claim 8, characterized in that: The loss function weight matrix W is updated in each training iteration as follows: the sampling frequency of each sampling point in the k-space is updated according to the sampling points contained in the training subset sampling matrix Θ in each training iteration, and then the inverse of the updated sampling frequency of each sampling point is used as the value of the corresponding sampling point in the loss function weight matrix W.

10. The hyperpolarized gas lung MRI reconstruction method based on self-supervised learning according to claim 5, characterized in that: In step 4, the sampling weight matrix of the first round of iterative training is randomly initialized; in each round of iterative training except the first round of iterative training, the sampling weight matrix obtained after the previous round of iterative training is used as the sampling weight matrix of the current round of iterative training.

Images