Lung gas MRI artificial intelligence reconstruction method combined with sampling matrix optimization

Abstract

The invention provides a lung gas MRI artificial intelligence reconstruction method combined with sampling matrix optimization. A Cartesian sampling matrix generator, a k-space reconstruction network and an image domain reconstruction network are constructed. The Cartesian sampling matrix generator optimizes a probability sampling matrix by constructing a differentiable binarization model, and can adaptively generate an optimal under-sampling mode according to the characteristics of a data set; the MRI reconstruction network adopts a double-domain learning mode of k space and an image domain, and abundant information contained in MRI data in the two domains is fully utilized, so that a high-quality reconstructed image is obtained; besides, the loss of the tissues in the lung and the loss of the background area are calculated respectively, and a higher weight is distributed for the loss of the tissues in the lung, so that the network pays more attention to the reconstruction quality of the areas in the lung; according to the method, the under-sampling mode and the MRI reconstruction network are optimized at the same time in an end-to-end training form, and the reconstruction quality of the lung hyperpolarized gas MRI is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of imaging technology, and in particular relates to a lung gas MRI artificial intelligence reconstruction method based on joint sampling matrix optimization. Background Art

[0002] Conventional magnetic resonance imaging (MRI) is based on hydrogen protons ( 1 H) imaging can provide information about the contour structure of the lungs. However, the lungs are hollow structures with extremely low hydrogen proton density in their internal tissues. Therefore, the lung tissues are the "blind area" of conventional MRI. Hyperpolarized gas MRI improves 129 The polarization degree of noble gases such as Xe enables visual detection of lung structure and function.

[0003] The scanning process of MRI usually takes a long time, especially the longitudinal magnetization vector of hyperpolarized gas is irreversible, the signal will decay exponentially with the scanning time, and the long scanning time will cause a significant decrease in signal intensity, thereby reducing the signal-to-noise ratio of the image. In addition, most hyperpolarized gas MRI methods require signal acquisition to be completed in one breath-holding. Long-term breath-holding will cause discomfort to the subjects, especially those with impaired respiratory function. Undersampling k-space data can effectively reduce the scanning time of MRI, but the undersampling strategy will cause serious aliasing artifacts in the image. It is necessary to apply compressed sensing, deep learning and other methods to remove aliasing artifacts and reconstruct high-quality MRI images. In the field of undersampling MRI reconstruction, the design of the sampling matrix and the optimization of the image reconstruction algorithm are two core issues. However, the existing deep learning-based undersampling lung hyperpolarized gas MRI reconstruction methods usually use a preset fixed undersampling matrix. Since the performance of the model depends largely on the selection of the undersampling matrix, the pre-set undersampling matrix is ​​not the optimal undersampling mode, which may affect the quality of the reconstructed image. Therefore, the development of joint optimization technology of sampling mode and reconstruction algorithm is expected to significantly improve the quality of the reconstructed image. Summary of the invention

[0004] The purpose of the present invention is to address the above-mentioned problems existing in the prior art and to propose a lung gas MRI artificial intelligence reconstruction method with joint sampling matrix optimization.

[0005] The above-mentioned object of the present invention is achieved by the following technical solutions:

[0006] The artificial intelligence reconstruction method of lung gas MRI based on joint sampling matrix optimization includes the following steps:

[0007] Step 1, constructing a training set, the training set includes multiple sample groups, each sample group includes 2D fully sampled lung hyperpolarized gas MRI k-space data, 2D fully sampled lung hyperpolarized gas MRI images and corresponding lung contour masks;

[0008] Step 2, constructing a Cartesian sampling matrix generator and an MRI reconstruction network, wherein the MRI reconstruction network includes a k-space reconstruction network and an image domain reconstruction network in sequence;

[0009] Step 3: Set the total loss function;

[0010] Step 4: According to the total loss function set in step 3, the Cartesian sampling matrix generator and the MRI reconstruction network are iteratively trained end-to-end using the training set generated in step 1 to obtain the optimal two-dimensional Cartesian undersampling matrix, and the parameters of the k-space reconstruction network and the image domain reconstruction network in the MRI reconstruction network are saved;

[0011] Step 5: Using the optimal two-dimensional Cartesian undersampling matrix, obtain the 2D under-sampled lung hyperpolarized gas MRI k-space data to be processed, and input the obtained 2D under-sampled lung hyperpolarized gas MRI k-space data to be processed into the MRI reconstruction network trained in step 4 to obtain the corresponding final reconstructed 2D lung hyperpolarized gas MRI image.

[0012] The Cartesian sampling matrix generator generates a two-dimensional Cartesian undersampling matrix M as described above;

[0013] The two-dimensional Cartesian undersampling matrix M undersamples the 2D fully sampled lung hyperpolarized gas MRI k-space data in the training set obtained in step 1 to obtain 2D undersampled lung hyperpolarized gas MRI k-space data;

[0014] The 2D under-sampled lung hyperpolarized gas MRI k-space data is input into the k-space reconstruction network to obtain the reconstructed 2D lung hyperpolarized gas MRI k-space data;

[0015] The reconstructed 2D lung hyperpolarized gas MRI k-space data is subjected to a two-dimensional inverse Fourier transform to obtain a preliminary reconstructed image, which is then input into an image domain reconstruction network to obtain the corresponding final reconstructed 2D lung hyperpolarized gas MRI image.

[0016] As mentioned above, step 1 specifically includes the following steps:

[0017] Step 1.1 Scan multiple subjects and obtain 3D fully sampled lung hyperpolarized gas MRI spatial data and 3D fully sampled lung data for each subject. 1 H MRI k-space data;

[0018] Step 1.2, extracting 3D fully sampled lung hyperpolarized gas MRI k-space data layer by layer, each 3D fully sampled lung hyperpolarized gas MRI k-space data corresponding to a plurality of 2D fully sampled lung hyperpolarized gas MRI k-space data; extracting 3D fully sampled lung layer by layer 1 H MRI k-space data, each 3D fully sampled lung 1 H MRI k-space data corresponding to multiple 2D fully sampled lungs 1 H MRI k-space data;

[0019] Step 1.3, perform a two-dimensional Fourier inverse transform on the 2D full-sampling lung hyperpolarized gas MRI k-space data to obtain a 2D full-sampling lung hyperpolarized gas MRI image; 1 H MRI k-space data is transformed by two-dimensional Fourier inverse transform to obtain 2D fully sampled lung 1 H MRI images;

[0020] Step 1.4: Segmentation of 2D full-sample lungs 1 H MRI image to obtain the lung contour mask;

[0021] Step 1.5: The 2D fully sampled lung hyperpolarized gas MRI k-space data, the 2D fully sampled lung hyperpolarized gas MRI image and the corresponding lung contour mask of a subject are taken as a sample group. The training set includes multiple sample groups.

[0022] As described above in step 2, the operation steps of the Cartesian sampling matrix generator to generate the two-dimensional Cartesian undersampling matrix M are as follows:

[0023] Step 2.1.1, set a probability sampling matrix P, the size of the probability sampling matrix P is set to 1×N x , where N x is the total number of columns of the 2D fully sampled lung hyperpolarized gas MRI k-space data obtained in step 1; P(i) represents the probability value of the i-th column of k-space data being sampled, and the initial value of the probability sampling matrix P is a 1×N matrix. x The vector O is used to parameterize, P = σ t (O), σ t represents the element-wise sigmoid function with slope t;

[0024] Step 2.1.2: Normalize the probability sampling matrix P:

[0025]

[0026] Among them, P nor is the normalized probability sampling matrix, α is the set sampling rate, p ais the mean of all elements in the probability sampling matrix;

[0027] Step 2.1.3, element-by-element comparison of the normalized probability sampling matrix P nor With the size of the threshold variable U, and using the sigmoid function to obtain a continuous, approximately binary intermediate matrix M1:

[0028] M1=σ s (P nor -U)

[0029] Among them, σ s represents an element-by-element sigmoid function with a slope of s; the size of the threshold variable U is the same as that of the probability sampling matrix P, and it obeys a uniform random distribution in the interval [0,1]; the size of the intermediate matrix M1 is the same as that of the probability sampling matrix P;

[0030] Step 2.1.4: Copy the intermediate matrix M1 to N y times, forming a size N y ×N x The two-dimensional Cartesian undersampling matrix M, N y is the total number of rows of 2D fully sampled lung hyperpolarized gas MRI k-space data obtained in step 1.

[0031] As described above, the k-space reconstruction network includes multiple cascaded coding enhancement modules; each coding enhancement module includes multiple coding enhancement complex-valued convolution processing layers in turn, and the last coding enhancement complex-valued convolution processing layer includes a coding enhancement complex-valued convolution layer; except for the last coding enhancement complex-valued convolution processing layer, each of the remaining coding enhancement complex-valued convolution processing layers includes a coding enhancement complex-valued convolution layer and an activation layer in turn; the main component unit of the coding enhancement complex-valued convolution layer is a coding enhancement complex-valued convolution kernel, and the size of the coding enhancement complex-valued convolution kernel is 1×N x .

[0032] As described above, the image domain reconstruction network includes a plurality of cascaded dense residual modules; each dense residual module includes a plurality of complex-valued convolution processing layers in sequence, and the last complex-valued convolution processing layer includes a complex-valued convolution layer; except for the last complex-valued convolution processing layer, each of the remaining complex-valued convolution processing layers includes a complex-valued convolution layer and an activation layer in sequence;

[0033] In each dense residual module: the input of the complex-valued convolution layer in the first complex-valued convolution processing layer is the input of the current dense residual module; except for the first complex-valued convolution processing layer, the input of the complex-valued convolution layer in each complex-valued convolution processing layer is the result of channel-wise concatenation of the output features of the activation layers in all previous complex-valued convolution processing layers; the output of the dense residual module is the sum of the output of the complex-valued convolution layer in the last complex-valued convolution processing layer of the current dense residual module and the input of the complex-valued convolution layer in the first complex-valued convolution processing layer.

[0034] As mentioned in step 3, the total loss function L total The settings are as follows:

[0035] L total =L k +L im

[0036] Among them, L k is the k-space loss function, L im is the image domain loss function;

[0037] L k =L′(k rec ,k full )

[0038] L′ is the error function of the k-space loss function, k rec represents the reconstructed 2D lung hyperpolarized gas MRI k-space data, k full represents 2D fully sampled lung hyperpolarized gas MRI k-space data;

[0039] L im =λL lun +(1-λ)L bac

[0040] Among them, L lun Indicates the loss in the lungs, L bac represents background loss, λ is a weighting parameter, λ>0.5;

[0041] L lun =L″(mask×I full ,mask×I rec )

[0042] L bac =L″((1-mask)×I full ,(1-mask)×I rec )

[0043] I rec Represents the final reconstructed 2D lung hyperpolarized gas MRI image, I fullrepresents a 2D fully sampled lung hyperpolarized gas MRI image, L″ represents the error function of the image domain loss function, and mask represents the lung contour mask.

[0044] As mentioned above, the error function L″ of the image domain loss function is the sum of the mean absolute value error loss and the mean square error loss.

[0045] As mentioned above in step 4, except for the first round of iterative training, each of the remaining rounds of iterative training will take the first row of the two-dimensional Cartesian under-sampling matrix M obtained after the previous round of iterative training as the probability sampling matrix P of this round of iterative training and execute steps 2.1.2-step 2.1.4 to complete the construction of the Cartesian sampling matrix generator in this round of iterative training.

[0046] A computer device comprises a memory and a processor, wherein a computer program is stored in the memory, and steps 2-5 of the reconstruction method described above are implemented when the processor executes the computer program.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] Existing lung hyperpolarized gas MRI reconstruction methods all use a preset fixed undersampling matrix. The present invention takes into account the influence of the sampling mode on the reconstruction quality, optimizes the sampling mode and the reconstruction algorithm, and can adaptively generate the optimal two-dimensional Cartesian undersampling matrix according to the characteristics of the data set, which helps to achieve higher quality lung hyperpolarized gas MRI reconstruction.

[0049] The present invention improves the existing random sampling matrix generation method and limits the sampling mode to Cartesian sampling, which is more practical.

[0050] In addition, the present invention calculates the losses of the lung tissue and the background area respectively, and assigns a higher weight to the lung tissue in a weighted manner, so that the network pays more attention to the reconstruction accuracy of key areas in the lung. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flowchart of the artificial intelligence reconstruction method of lung gas MRI with joint sampling matrix optimization.

[0052] Figure 2 It is a structural diagram of the coding enhancement module in the k-space reconstruction network.

[0053] Figure 3 It is a schematic diagram of the structure of the dense residual module in the image domain reconstruction network, where Indicates the channel splicing operation, Represents an element-by-element addition operation.

[0054] Figure 4: This is a comparison chart of the hyperpolarized gas MRI reconstruction results in the test set of Example 1, where (a) is the optimal two-dimensional Cartesian undersampling matrix generated by the Cartesian sampling matrix generator, (b) is a 2D fully sampled lung hyperpolarized gas MRI image, (c) is a zero-filled lung hyperpolarized gas MRI image, and (d) is the final reconstructed 2D lung hyperpolarized gas MRI image. PSNR is the Peak Signal-to-Noise Ratio, and SSIM is the Structural Similarity. DETAILED DESCRIPTION

[0055] In order to make the objectives, technical solutions and advantages of the present invention clearer, thereby facilitating those skilled in the art to understand and implement the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. The implementation examples described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0056] Example 1

[0057] like Figure 1 As shown, the lung gas MRI artificial intelligence reconstruction method based on joint sampling matrix optimization specifically includes the following steps:

[0058] Step 1: construct a training set and a test set. The training set and the test set both include multiple sample groups. Each sample group includes 2D fully sampled lung hyperpolarized gas MRI k-space data, 2D fully sampled lung hyperpolarized gas MRI images, and corresponding lung contour masks. In this embodiment, the hyperpolarized gas is selected from 129 Xe, the specific method is as follows:

[0059] Step 1.1, scan multiple subjects, and obtain 3D full-sampled lung hyperpolarized gas MRI k-space data and 3D full-sampled lung data for each subject. 1 H MRI k-space data.

[0060] Step 1.2: Extract 3D fully sampled lung hyperpolarized gas MRI k-space data layer by layer, and obtain multiple 2D fully sampled lung hyperpolarized gas MRI k-space data corresponding to each 3D fully sampled lung hyperpolarized gas MRI k-space data. 1 H MRI k-space data, composed of each 3D fully sampled lung 1 H MRI k-space data corresponding to multiple 2D fully sampled lungs 1 H MRI k-space data.

[0061] Step 1.3: Perform a two-dimensional inverse Fourier transform on the 2D fully sampled lung hyperpolarized gas MRI k-space data to obtain a 2D fully sampled lung hyperpolarized gas MRI image. 1 H MRI k-space data is transformed by two-dimensional Fourier inverse transform to obtain 2D fully sampled lung 1 H MRI images.

[0062] Step 1.4: Segmentation of 2D full-sample lungs 1 The lung contour mask is obtained from the H MRI image.

[0063] Step 1.5: Take a subject's 2D fully sampled lung hyperpolarized gas MRI k-space data, 2D fully sampled lung hyperpolarized gas MRI image, and the corresponding lung contour mask 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.

[0064] In this embodiment, 3D bSSFP sequence was used to scan 85 subjects to obtain 3D fully sampled lung hyperpolarized gas MRI k-space data, and 3D FLASH sequence was used to scan the same 85 subjects to obtain 3D fully sampled lung 1 H MRI k-space data. 3D fully sampled lung hyperpolarized gas MRI k-space data and 3D fully sampled lung 1 The matrix size of H MRI k-space data is 96 × 96 × 24. 3D full-sampled lung hyperpolarized gas MRI k-space data is extracted layer by layer to obtain 2D full-sampled lung hyperpolarized gas MRI k-space data. 1 H MRI k-space data to obtain 2D fully sampled lung 1 HMRI k-space data. Perform a two-dimensional inverse Fourier transform on the 2D fully sampled lung hyperpolarized gas MRI k-space data to obtain a 2D fully sampled lung hyperpolarized gas MRI image. 1 The H MRI k-space data is transformed by two-dimensional Fourier inverse transformation to obtain a 2D fully sampled lung 1 H MRI image. 2D full-sampled lung segmentation using seeded region growing combined with manual correction 1 H MRI images, and obtain the lung contour mask. A sample group of 68 subjects was randomly selected as the training set, and a sample group of 17 subjects was randomly selected as the test set.

[0065] Step 2, constructing a Cartesian sampling matrix generator and an MRI reconstruction network, wherein the MRI reconstruction network includes a k-space reconstruction network and an image domain reconstruction network in sequence;

[0066] Among them, the Cartesian sampling matrix generator generates a two-dimensional Cartesian undersampling matrix M,

[0067] The two-dimensional Cartesian undersampling matrix M undersamples the 2D fully sampled lung hyperpolarized gas MRI k-space data in the training set obtained in step 1 to obtain 2D undersampled lung hyperpolarized gas MRI k-space data;

[0068] The 2D under-sampled lung hyperpolarized gas MRI k-space data is input into the k-space reconstruction network to obtain the reconstructed 2D lung hyperpolarized gas MRI k-space data;

[0069] The reconstructed 2D lung hyperpolarized gas MRI k-space data is subjected to a two-dimensional inverse Fourier transform to obtain a preliminary reconstructed image, which is then input into an image domain reconstruction network to obtain the corresponding final reconstructed 2D lung hyperpolarized gas MRI image.

[0070] The construction methods of the Cartesian sampling matrix generator, the k-space reconstruction network, and the image domain reconstruction network are as follows:

[0071] Step 2.1 constructs a Cartesian sampling matrix generator to generate a two-dimensional Cartesian undersampling matrix M. The specific steps are as follows:

[0072] Step 2.1.1. Set a probability sampling matrix P. In order to limit the sampling mode to Cartesian sampling, the size of the probability sampling matrix P is set to 1×N x , where N x is the total number of columns of 2D fully sampled lung hyperpolarized gas MRI k-space data obtained in step 1, that is, the phase encoding number. P(i) represents the probability value of the i-th column of k-space data being sampled.

[0073] The initial value of the probability sampling matrix P is a 1×N x The vector O is parameterized, that is, P = σ t (O), σ t represents an element-by-element sigmoid function with a slope of t. Therefore, each element in the probability sampling matrix P is in the interval (0,1), and the probability sampling matrix P is adaptively adjusted with each round of end-to-end iterative training.

[0074] Step 2.1.2: To ensure that the probability sampling matrix P satisfies the given sampling rate constraint (the given sampling rate is recorded as the set sampling rate α), the probability sampling matrix P is normalized:

[0075]

[0076] Among them, P nor is the normalized probability sampling matrix, α is the set sampling rate, p ais the mean of all elements in the probability sampling matrix. The sampling rate α is set to be fixed, and the probability sampling matrix P is adaptively adjusted with each round of end-to-end iterative training. Through the normalization operation, it is ensured that no matter how the probability sampling matrix P is adjusted, the normalized probability sampling matrix P nor The mean of all elements in is equal to the set sampling rate α.

[0077] Step 2.1.3, element-by-element comparison of the normalized probability sampling matrix P nor With the size of the threshold variable U, and using the sigmoid function to obtain a continuous, approximately binary intermediate matrix M1:

[0078] M1=σ s (P nor -U)

[0079] Among them, σ s represents the element-by-element sigmoid function with slope s. The size of the threshold variable U is the same as that of the probability sampling matrix P and follows a uniform random distribution in the interval [0,1]. The size of the intermediate matrix M1 is the same as that of the probability sampling matrix P, which is 1×N x .

[0080] Step 2.1.4: Copy the intermediate matrix M1 to N y times, forming a size N y ×N x The two-dimensional Cartesian undersampling matrix M, N y is the total number of rows of the 2D fully sampled lung hyperpolarized gas MRI k-space data obtained in step 1, that is, the number of frequency codes.

[0081] In this embodiment, the phase encoding number N of the hyperpolarized gas MRI is x is 96, the frequency code number N y is 96, the slope t and slope s are 5 and 200 respectively, the sampling rate α is set to 0.25, and the corresponding acceleration factor is 4.

[0082] Step 2.2: Construct the k-space reconstruction network.

[0083] The k-space reconstruction network includes multiple cascaded coding enhancement modules; each coding enhancement module includes multiple coding enhancement complex-valued convolution processing layers in turn; the last coding enhancement complex-valued convolution processing layer includes a coding enhancement complex-valued convolution layer; except for the last coding enhancement complex-valued convolution processing layer, each of the remaining coding enhancement complex-valued convolution processing layers includes a coding enhancement complex-valued convolution layer and an activation layer in turn. The main component of the coding enhancement complex-valued convolution layer is the coding enhancement complex-valued convolution kernel, and the size of the coding enhancement complex-valued convolution kernel is 1×N x .

[0084] In this embodiment, the k-space reconstruction network is composed of two cascaded coding enhancement modules. The structural diagram of the coding enhancement module is as follows: Figure 2 As shown in the figure, each coding enhancement module includes five coding enhancement complex-valued convolutional layers in sequence. The main component of the coding enhancement complex-valued convolutional layer is the coding enhancement complex-valued convolutional kernel, and the size of the coding enhancement complex-valued convolutional kernel is 1×96. Except for the last coding enhancement complex-valued convolutional layer, each of the remaining coding enhancement complex-valued convolutional layers is followed by a Tanh activation layer.

[0085] Step 2.3: Construct an image domain reconstruction network.

[0086] The image domain reconstruction network includes multiple cascaded dense residual modules; each dense residual module includes multiple complex-valued convolution processing layers in turn; the last complex-valued convolution processing layer includes a complex-valued convolution layer; except for the last complex-valued convolution processing layer, each of the remaining complex-valued convolution processing layers includes a complex-valued convolution layer and an activation layer in turn. In each dense residual module: the input of the complex-valued convolution layer in the first complex-valued convolution processing layer is the input of the current dense residual module; except for the first complex-valued convolution processing layer, the input of the complex-valued convolution layer in each complex-valued convolution processing layer is the result of channel-wise concatenation of the output features of the activation layers in all previous complex-valued convolution processing layers, thereby achieving full flow and utilization of features within the dense residual module; the output of the dense residual module is the pixel-by-pixel sum of the output of the complex-valued convolution layer in the last complex-valued convolution processing layer of the current dense residual module and the input of the complex-valued convolution layer in the first complex-valued convolution processing layer of the current dense residual module.

[0087] In this embodiment, the image domain reconstruction network is composed of 15 cascaded dense residual modules. The structural diagram of the dense residual module is as follows: Figure 3 As shown in Figure 2, each dense residual module consists of 5 complex-valued convolutional layers, and each complex-valued convolutional layer except the last one is followed by a ReLu activation layer.

[0088] Step 3: Set the total loss function L total The total loss function is the k-space loss function L k And the image domain loss function L im sum:

[0089] L total =L k +L im

[0090] The k-space loss function L k The error between the reconstructed 2D lung hyperpolarized gas MRI k-space data and the 2D fully sampled lung hyperpolarized gas MRI k-space data in step 1 is calculated:

[0091] Lk =L′(k rec ,k full )

[0092] L′ is the error function of the k-space loss function, k rec represents the reconstructed 2D lung hyperpolarized gas MRI k-space data, k full Represents 2D fully sampled lung hyperpolarized gas MRI k-space data.

[0093] In order to make the network pay more attention to the reconstruction accuracy of the lung tissue, the present invention uses the image domain loss function L im L lun With background loss L bac (i.e., loss outside the lungs) are divided into two parts, and the specific calculation formula is as follows:

[0094] L im =λL lun +(1-λ)L bac

[0095] L lun =L″(mask×I full ,mask×I rec )

[0096] L bac =L″((1-mask)×I full ,(1-mask)×I rec )

[0097] λ is the weighting parameter (λ>0.5), I rec Represents the final reconstructed 2D lung hyperpolarized gas MRI image, I full represents the 2D fully sampled lung hyperpolarized gas MRI image in step 1, L″ represents the error function of the image domain loss function, and mask represents the lung contour mask obtained in step 1.

[0098] In this embodiment, the error function L′ of the k-space loss function uses the mean square error (L2) loss. The error function L″ of the image domain loss function uses the sum of the mean absolute value error (L1) loss and the mean square error (L2) loss. λ is 0.8.

[0099] Step 4: According to the total loss function set in step 3, the Cartesian sampling matrix generator and the MRI reconstruction network are trained end-to-end iteratively using the training set generated in step 1. The learning rate of the MRI reconstruction network is initialized to 0.0002, the batch size is set to 10, and the Cartesian sampling matrix generator and the MRI reconstruction network are trained on the Tensorflow2.0 platform using the Adam optimizer. The training is stopped after the total number of iterations reaches 200, and the parameters of the k-space reconstruction network and the image domain reconstruction network as well as the optimal two-dimensional Cartesian undersampling matrix are saved.

[0100] Among them, for the Cartesian sampling matrix generator: except for the first round of iterative training, each of the remaining rounds of iterative training will take the first row of the two-dimensional Cartesian under-sampling matrix M obtained after the previous round of iterative training as the probability sampling matrix P of this round of iterative training and execute subsequent steps 2.1.2-step 2.1.4 to complete the construction of the Cartesian sampling matrix generator in this round of iterative training.

[0101] Step 5: Using the optimal two-dimensional Cartesian undersampling matrix obtained in step 4, obtain the 2D under-sampled lung hyperpolarized gas MRI k-space data to be processed, and input the obtained 2D under-sampled lung hyperpolarized gas MRI k-space data to be processed into the MRI reconstruction network trained in step 4 to obtain the final reconstructed 2D lung hyperpolarized gas MRI image.

[0102] According to the optimal two-dimensional Cartesian under-sampling matrix obtained in step 4, 3D under-sampled lung hyperpolarized gas MRI k-space data are obtained by scanning with an MRI scanner, and the 3D under-sampled lung hyperpolarized gas MRI k-space data are extracted layer by layer, and a plurality of 2D under-sampled lung hyperpolarized gas MRI k-space data to be processed are correspondingly obtained, and the obtained 2D under-sampled lung hyperpolarized gas MRI k-space data to be processed are input into the MRI reconstruction network trained in step 4, and a final reconstructed 2D lung hyperpolarized gas MRI image is correspondingly obtained.

[0103] In this embodiment, in order to verify the reconstruction effect, an optimal two-dimensional Cartesian under-sampling matrix is ​​obtained according to the Cartesian sampling matrix generator trained in step 4; the 2D fully sampled lung hyperpolarized gas MRI k-space data in the test set generated in step 1 is under-sampled using the optimal two-dimensional Cartesian under-sampling matrix to obtain the unprocessed 2D under-sampled lung hyperpolarized gas MRI k-space data for testing, and the unprocessed 2D under-sampled lung hyperpolarized gas MRI k-space data for testing is input into the MRI reconstruction network trained in step 4 to obtain the final reconstructed 2D lung hyperpolarized gas MRI image of the test set.

[0104] In this embodiment, the final reconstructed 2D lung hyperpolarized gas MRI image of the test set is compared with the 2D fully sampled lung 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.

[0105] Figure 4 An example of reconstruction result in the test set is shown, and the PSNR / SSSIM value is marked at the bottom of the image. Among them, (a) is the optimal two-dimensional Cartesian undersampling matrix generated by the network. The optimal two-dimensional Cartesian undersampling matrix obtained in this embodiment is a Cartesian undersampling matrix with an acceleration factor of 4 times, (b) is a 2D fully sampled lung hyperpolarized gas MRI image, (c) is a zero-filled lung hyperpolarized gas MRI image, and (d) is the final reconstructed lung hyperpolarized gas MRI image. It can be seen from the result graph that the joint sampling matrix optimized lung gas MRI artificial intelligence reconstruction method provided by the present invention, wherein the two-dimensional Cartesian undersampling matrix is ​​capable of adaptively generating an undersampling matrix and obtaining high-quality hyperpolarized gas MRI reconstructed images.

[0106] Example 2

[0107] Utilizing the joint sampling matrix optimized lung gas MRI artificial intelligence reconstruction method described in Example 1, this embodiment provides a joint sampling matrix optimized lung gas MRI artificial intelligence reconstruction device, including a network construction module, a loss function construction module, a training module and an application module.

[0108] Wherein, the network construction module is used to implement step 2 in embodiment 1;

[0109] The loss function construction module is used to implement step 3 in Example 1;

[0110] The training module is used to implement step 4 in Example 1.

[0111] The application module is used to implement step 5 in Example 1.

[0112] Example 3

[0113] 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.

[0114] Example 4

[0115] A computer-readable storage medium stores a computer program, which implements steps 2-5 in embodiment 1 when executed by a processor.

[0116] Example 5

[0117] 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.

[0118] The specific embodiments 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 embodiments 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 lung gas MRI artificial intelligence reconstruction method based on joint sampling matrix optimization, characterized in that: The steps include: Step 1, constructing a training set, the training set includes multiple sample groups, each sample group includes 2D fully sampled lung hyperpolarized gas MRI k-space data, 2D fully sampled lung hyperpolarized gas MRI images and corresponding lung contour masks; Step 2, constructing a Cartesian sampling matrix generator and an MRI reconstruction network, wherein the MRI reconstruction network includes a k-space reconstruction network and an image domain reconstruction network in sequence; Step 3: Set the total loss function; Step 4: According to the total loss function set in step 3, the Cartesian sampling matrix generator and the MRI reconstruction network are iteratively trained end-to-end using the training set generated in step 1 to obtain the optimal two-dimensional Cartesian undersampling matrix, and the parameters of the k-space reconstruction network and the image domain reconstruction network in the MRI reconstruction network are saved; Step 5: Using the optimal two-dimensional Cartesian undersampling matrix, obtain the 2D under-sampled lung hyperpolarized gas MRI k-space data to be processed, and input the obtained 2D under-sampled lung hyperpolarized gas MRI k-space data to be processed into the MRI reconstruction network trained in step 4 to obtain the corresponding final reconstructed 2D lung hyperpolarized gas MRI image.

2. The artificial intelligence reconstruction method for lung gas MRI based on joint sampling matrix optimization according to claim 1 is characterized in that: The Cartesian sampling matrix generator generates a two-dimensional Cartesian under-sampling matrix M; The two-dimensional Cartesian undersampling matrix M undersamples the 2D fully sampled lung hyperpolarized gas MRI k-space data in the training set obtained in step 1 to obtain 2D undersampled lung hyperpolarized gas MRI k-space data; The 2D under-sampled lung hyperpolarized gas MRI k-space data is input into the k-space reconstruction network to obtain the reconstructed 2D lung hyperpolarized gas MRI k-space data; The reconstructed 2D lung hyperpolarized gas MRI k-space data is subjected to a two-dimensional inverse Fourier transform to obtain a preliminary reconstructed image, which is then input into an image domain reconstruction network to obtain the corresponding final reconstructed 2D lung hyperpolarized gas MRI image.

3. The artificial intelligence reconstruction method for lung gas MRI with joint sampling matrix optimization according to claim 1 is characterized in that: The step 1 specifically comprises the following steps: Step 1.1 Scan multiple subjects and obtain 3D full-sampled lung hyperpolarized gas MRI k-space data and 3D full-sampled lung data for each subject. 1 H MRI k-space data; Step 1.2, extracting 3D fully sampled lung hyperpolarized gas MRI k-space data layer by layer, each 3D fully sampled lung hyperpolarized gas MRI k-space data corresponding to a plurality of 2D fully sampled lung hyperpolarized gas MRI k-space data; extracting 3D fully sampled lung layer by layer 1 H MRI k-space data, each 3D fully sampled lung 1 H MRI k-space data corresponding to multiple 2D fully sampled lungs 1 H MRI k-space data; Step 1.3, perform a two-dimensional Fourier inverse transform on the 2D full-sampling lung hyperpolarized gas MRI k-space data to obtain a 2D full-sampling lung hyperpolarized gas MRI image; 1 H MRI k-space data is transformed by two-dimensional Fourier inverse transform to obtain 2D fully sampled lung 1 H MRI images; Step 1.4: Segmentation of 2D full-sample lungs 1 H MRI image to obtain the lung contour mask; Step 1.5: The 2D fully sampled lung hyperpolarized gas MRI k-space data, the 2D fully sampled lung hyperpolarized gas MRI image and the corresponding lung contour mask of a subject are taken as a sample group. The training set includes multiple sample groups.

4. The artificial intelligence reconstruction method for lung gas MRI with joint sampling matrix optimization according to claim 2 is characterized in that: In step 2, the operation steps of the Cartesian sampling matrix generator to generate the two-dimensional Cartesian undersampling matrix M are as follows: Step 2.1.1, set a probability sampling matrix P, the size of the probability sampling matrix P is set to 1×N x , where N x is the total number of columns of the 2D fully sampled lung hyperpolarized gas MRI k-space data obtained in step 1; P(i) represents the probability value of the i-th column of k-space data being sampled, and the initial value of the probability sampling matrix P is a 1×N matrix. x The vector O is used to parameterize, P = σ t (O), σ t represents the element-wise sigmoid function with slope t; Step 2.1.2: Normalize the probability sampling matrix P: Among them, P nor is the normalized probability sampling matrix, α is the set sampling rate, p a is the mean of all elements in the probability sampling matrix; Step 2.1.3, element-by-element comparison of the normalized probability sampling matrix P nor With the size of the threshold variable U, and using the sigmoid function to obtain a continuous, approximately binary intermediate matrix M1: M1=σ s (P nor -U) Among them, σ s represents an element-by-element sigmoid function with a slope of s; the size of the threshold variable U is the same as that of the probability sampling matrix P, and it obeys a uniform random distribution in the interval [0,1]; the size of the intermediate matrix M1 is the same as that of the probability sampling matrix P; Step 2.1.4: Copy the intermediate matrix M1 to N y times, forming a size N y ×N x The two-dimensional Cartesian undersampling matrix M, N y is the total number of rows of 2D fully sampled lung hyperpolarized gas MRI k-space data obtained in step 1.

5. The artificial intelligence reconstruction method for lung gas MRI with joint sampling matrix optimization according to claim 2 is characterized in that: The k-space reconstruction network includes a plurality of cascaded coding enhancement modules; each coding enhancement module includes a plurality of coding enhancement complex-valued convolution processing layers in sequence, and the last coding enhancement complex-valued convolution processing layer includes a coding enhancement complex-valued convolution layer; except for the last coding enhancement complex-valued convolution processing layer, each of the remaining coding enhancement complex-valued convolution processing layers includes a coding enhancement complex-valued convolution layer and an activation layer in sequence; the main component unit of the coding enhancement complex-valued convolution layer is a coding enhancement complex-valued convolution kernel, and the size of the coding enhancement complex-valued convolution kernel is 1×N x .

6. The artificial intelligence reconstruction method for lung gas MRI with joint sampling matrix optimization according to claim 2 is characterized in that: The image domain reconstruction network includes a plurality of cascaded dense residual modules; each dense residual module includes a plurality of complex-valued convolution processing layers in sequence, and the last complex-valued convolution processing layer includes a complex-valued convolution layer; except for the last complex-valued convolution processing layer, each of the remaining complex-valued convolution processing layers includes a complex-valued convolution layer and an activation layer in sequence; In each dense residual module: the input of the complex-valued convolution layer in the first complex-valued convolution processing layer is the input of the current dense residual module; except for the first complex-valued convolution processing layer, the input of the complex-valued convolution layer in each complex-valued convolution processing layer is the result of channel-wise concatenation of the output features of the activation layers in all previous complex-valued convolution processing layers; the output of the dense residual module is the sum of the output of the complex-valued convolution layer in the last complex-valued convolution processing layer of the current dense residual module and the input of the complex-valued convolution layer in the first complex-valued convolution processing layer.

7. The artificial intelligence reconstruction method for lung gas MRI with joint sampling matrix optimization according to claim 2 is characterized in that: In step 3, the total loss function L total The settings are as follows: L total =L k +L im Among them, L k is the k-space loss function, L im is the image domain loss function; L k =L′(k rec ,k full ) L′ is the error function of the k-space loss function, k rec represents the reconstructed 2D lung hyperpolarized gas MRI k-space data, k full represents 2D fully sampled lung hyperpolarized gas MRI k-space data; THE im =λL lun +(1-λ)L bac Among them, L lun Indicates the loss in the lungs, L bac represents background loss, λ is a weighting parameter, λ>0.5; L lun =L″(mask×I full ,mask×I rec ) L bac =L″((1-mask)×I full ,(1-mask)×I rec ) I rec Represents the final reconstructed 2D lung hyperpolarized gas MRI image, I full represents a 2D fully sampled lung hyperpolarized gas MRI image, L″ represents the error function of the image domain loss function, and mask represents the lung contour mask.

8. The artificial intelligence reconstruction method for lung gas MRI with joint sampling matrix optimization according to claim 7 is characterized in that: The error function L″ of the image domain loss function is the sum of the mean absolute value error loss and the mean square error loss.

9. The artificial intelligence reconstruction method for lung gas MRI with joint sampling matrix optimization according to claim 4 is characterized in that: In step 4, except for the first round of iterative training, each of the remaining rounds of iterative training takes the first row of the two-dimensional Cartesian under-sampling matrix M obtained after the previous round of iterative training as the probability sampling matrix P of the current round of iterative training and executes steps 2.1.2 to 2.1.4 to complete the construction of the Cartesian sampling matrix generator in the current round of iterative training.

10. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, steps 2-5 of the reconstruction method described in any one of claims 1-9 are implemented.

Images