Artificial intelligence reconstruction method for lung multi-core MRI based on image generation

Abstract

The present invention provides an artificial intelligence reconstruction method for multi-core MRI of the lungs based on image generation, including a cross-modality generation network, a k-space reconstruction network and an image domain reconstruction network. The cross-modality generation network uses a fully sampled reference modality image to obtain a synthesized fully sampled target modality image, and obtains synthesized fully sampled target modality k-space data through Fourier transform. The k-reconstruction network predicts the complete target modality k-space data based on the undersampled target modality k-space data and the synthesized fully sampled target modality k-space data, and then uses the image domain reconstruction network to obtain the final target modality reconstructed image. The present invention effectively introduces reference modality data as an additional information source through a cross-modality generation network, utilizes the information complementarity between multi-core MRI data, enhances the detail information of the reconstructed image, and significantly improves the reconstruction quality of the target modality.

Description

Technical Field

[0001] The present invention belongs to the field of imaging technology, and in particular relates to a lung multi-core MRI artificial intelligence reconstruction method based on image generation. Background Art

[0002] Magnetic resonance imaging (MRI) is a non-ionizing radiation and non-invasive imaging technique with good tissue contrast, so it is widely used [Mehmet Steen Moeller, Sebastian et al.Magn.Reson.Med.,2019,439:453.]. Multinuclear MRI refers to the use of different nuclei (such as 1 H, 129 Xe, etc.) to obtain more comprehensive information. 1 H MRI can provide structural information of the lungs, hyperpolarization 129 Xe MRI can provide information on lung function, such as air-air or air-blood exchange function. 1 H MRI and 129 Xe MRI can obtain quantitative parameters of lung ventilation function, which is of great significance for lung detection.

[0003] The scanning process of MRI is usually time-consuming. This problem is more significant in multi-nuclear magnetic resonance imaging because it requires the acquisition of MRI data based on different nuclei, which increases the time required for imaging. In addition, most hyperpolarized gas MRI methods require signal acquisition to be completed in one breath-hold. Long-term breath-holding can cause discomfort to the subject, especially for those with impaired respiratory function. Therefore, accelerating multi-nuclear MRI is of great significance. At present, deep learning methods have shown great potential in the field of undersampling MRI reconstruction. Deep learning-based reconstruction methods are mainly divided into two categories: data-driven end-to-end deep learning reconstruction methods and physical / model-driven unfolding iterative deep learning methods [Shanshan Wang, Taohui Xiao, Qiegen Liu, et al., Biomed. Signal Proces., 2021, 68: 102579.], among which the data-driven end-to-end deep learning reconstruction method can learn the mapping relationship between undersampled MRI data and fully sampled MRI data from a large amount of training data [Taejoon Eo, Yohan Jun, Taeseong Kim, et al., Magn. Reson. Med., 2018, 80: 2188-2201.]; the physical / model-driven unfolding iterative deep learning method takes into account the physical knowledge of MRI and the prior knowledge of images, combines the traditional MRI reconstruction model (such as compressed sensing, etc.) with the deep learning method, and uses the neural network to learn the optimal parameters of the reconstruction model [Bo Zhou, Jo Schlemper, Neel Dey, et al., Med. Image Anal., 2022, 81: 102538.]. However, existing deep learning-based lung MRI reconstruction methods are usually designed for single-core MRI data and do not utilize the correlation between multi-core MRI data [Caohui Duan, He Deng, Sa Xiao, et al., Eur. Radiol., 2022, 32: 702-713.]. 1 H MRI and 129 Xe MRI reflects similar lung anatomy and provides complementary information. 1 H MRI features for multi-core MRI reconstruction, compared with existing lung single-core MRI (such as 129 Xe MRI) reconstruction method can obtain higher quality reconstructed images. Summary of the invention

[0004] The purpose of the present invention is to effectively utilize the complementary information between multi-core MRI data to further improve the reconstruction quality, and to propose an artificial intelligence reconstruction method for lung multi-core MRI based on image generation.

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

[0006] The lung multi-core MRI artificial intelligence reconstruction method based on image generation includes the following steps:

[0007] Step 1, obtaining a training set and a test set, wherein the training set and the test set each include multiple sample groups, each sample group includes a 2D fully sampled reference modality image, a 2D fully sampled target modality k-space data, a 2D under-sampled target modality k-space data corresponding to the 2D fully sampled target modality k-space data, and a 2D fully sampled target modality image;

[0008] Step 2: Build a cross-modal generation network;

[0009] Step 3: Input the 2D fully sampled reference modality image in the training set generated in step 1 into the cross-modality generation network constructed in step 2 to obtain the corresponding synthesized fully sampled target modality image;

[0010] Step 4: construct a k-space reconstruction network;

[0011] Step 5: Perform a two-dimensional Fourier transform on the synthesized fully sampled target modality image obtained in step 3 to obtain synthesized fully sampled target modality k-space data, and then perform a channel splicing operation on the synthesized fully sampled target modality k-space data and the 2D under-sampled target modality k-space data in the training set generated in step 1. The obtained joint k-space data is input into the k-space reconstruction network constructed in step 4 to obtain the corresponding reconstructed target modality k-space data;

[0012] Step 6: Construct an image domain reconstruction network;

[0013] Step 7, performing a two-dimensional inverse Fourier transform on the reconstructed target modality k-space data obtained in step 5 to obtain reconstructed target modality image domain data, and then performing a channel stitching operation on the reconstructed target modality image domain data and the synthetic full-sampled target modality image obtained in step 3, and the obtained joint image data is input into the image domain reconstruction network constructed in step 6 to obtain the corresponding final reconstructed target modality image;

[0014] Step 8: Set the total loss function;

[0015] Step 9: According to the total loss function set in step 8, the cross-modality generation network, the k-space reconstruction network and the image domain reconstruction network are trained, and the parameters of all networks are saved;

[0016] Step 10: Input the 2D fully sampled reference modality image and the 2D under-sampled target modality k-space data in the test set generated in step 1 into the network trained in step 9 to obtain the final reconstructed target modality image for the test.

[0017] The sample set in step 1 above is obtained by the following steps:

[0018] Step 1.1, scanning multiple volunteers to obtain 3D full-sampling reference modality k-space data and 3D full-sampling target modality k-space data;

[0019] Step 1.2, extracting 3D fully sampled reference modality k-space data layer by layer to obtain multiple 2D fully sampled reference modality k-space data, extracting 3D fully sampled target modality k-space data layer by layer to obtain multiple 2D fully sampled target modality k-space data, and undersampling the 2D fully sampled target modality k-space data according to the under-sampling sampling matrix to obtain corresponding 2D under-sampling target modality k-space data;

[0020] Step 1.3, perform a two-dimensional inverse Fourier transform on the 2D fully sampled reference modality k-space data to obtain a 2D fully sampled reference modality image; perform a two-dimensional inverse Fourier transform on the 2D fully sampled target modality k-space data to obtain a 2D fully sampled target modality image.

[0021] As described in step 2 above, the cross-modal generation network includes an encoder and a decoder; the encoder includes multiple downsampling layers, and the decoder includes multiple upsampling layers; each downsampling layer includes multiple real-valued convolution layers and a maximum pooling layer, wherein each downsampling layer has an activation layer after the real-valued convolution layer; each upsampling layer includes multiple real-valued convolution layers and an upsampling real-valued convolution layer, wherein each upsampling layer has an activation layer after the real-valued convolution layer; the upsampling feature maps and the downsampling feature maps of the corresponding levels are channel-joined through a jump connection.

[0022] As described above, the k-space reconstruction network in step 4 includes multiple cascaded residual modules; each residual module contains multiple complex-valued convolutional layers, and each complex-valued convolutional layer except the last one is followed by a corresponding activation layer; the final output of each residual module is the pixel-by-pixel sum of the output of the last complex-valued convolutional layer in the current residual module and the input of the first complex-valued convolutional layer in the current residual module.

[0023] As described above, the image domain reconstruction network in step 6 includes multiple cascaded dense residual modules; each dense residual module contains multiple complex-valued convolutional layers, and each complex-valued convolutional layer except the last one is followed by a corresponding activation layer, and each complex-valued convolutional layer and the corresponding activation layer serve as a first-level complex-valued convolution processing layer. In each dense residual module, the input of the first-level complex-valued convolution processing layer is the input of the dense residual module, and the input of each level of complex-valued convolution processing layer except the first-level complex-valued convolution processing layer is the result of channel concatenation of the input and output of the previous level of complex-valued convolution processing layer, and the output of the last complex-valued convolution layer is added pixel by pixel to the input of the first-level complex-valued convolution processing layer as the output of the current dense residual module.

[0024] As described above, the total loss function in step 8 is the sum of the first loss function, the second loss function and the third loss function, wherein the first loss function is the error between the fully sampled target modality image synthesized by the cross-modality generation network in step 3 and the 2D fully sampled target modality image in step 1, the second loss function is the error between the reconstructed target modality k-space data obtained by the k-space reconstruction network in step 5 and the 2D fully sampled target modality k-space data in step 1, and the third loss function is the error between the final reconstructed target modality image obtained in step 7 and the 2D fully sampled target modality image in step 1.

[0025] As mentioned above, the reference mode is 1 H MRI, target modality is 129 Xe MRI.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] The present invention effectively utilizes the lungs 129 Xe MRI vs. 1 Based on the complementary information between H MRI, a cross-modal generation network is designed to generate the full-sampled reference modality ( 1 Generate the target modality (H MRI) data 129 Xe MRI) data, although due to lung 1 HMRI contains limited information about ventilation function, which results in 129 Xe MRI data is quite different from real data, but it can still effectively generate trachea, lung edges, etc. 1 The area present in H MRI is for subsequent 129The undersampling reconstruction of Xe MRI provides useful prior knowledge such as the trachea and lung edges, which helps to enhance the detailed features of the reconstructed image, thereby improving the reconstruction accuracy. In the reconstruction network, a dual-domain (k-space and image domain) learning method is adopted, and the advantages of residual connection and dense connection are utilized to strengthen feature reuse, thereby improving the generalization ability of the network. Compared with the existing lung single-core MRI reconstruction method, this invention introduces a reference modality ( 1 H MRI) data, as an additional source of information, can significantly improve the target modality ( 129 The reconstruction quality of XeMRI. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flowchart of the artificial intelligence reconstruction method of lung multi-core MRI based on image generation.

[0029] Figure 2 It is a schematic diagram of the under-sampling sampling matrix.

[0030] Figure 3 It is a schematic diagram of the structure of the downsampling layer and the upsampling layer in the cross-modal generation network, where (a) is a schematic diagram of the structure of the downsampling layer in the cross-modal generation network; (b) is a schematic diagram of the structure of the upsampling layer in the cross-modal generation network.

[0031] Figure 4 It is a structural diagram of the residual module, where Represents a pixel-by-pixel addition operation.

[0032] Figure 5 It is a schematic diagram of the structure of the dense residual module, where Indicates the channel splicing operation, Represents a pixel-by-pixel addition operation.

[0033] Figure 6 is the test set reconstruction 129 Comparison of Xe MRI image results, where (a) is 2D full sampling 129 Xe MRI image, (b) is zero-filled 129 Xe MRI image, (c) is the final reconstruction 129 Xe MRI image.

[0034] Figure 7 It is an artificial intelligence reconstruction device for lung multi-core MRI based on image generation.

[0035] Among them, 101-acquisition module, 102-first generation module, 103-input module, 104-second generation module, 105-first change module, 106-third generation module, 107-second change module, 108-loss function construction module, 109-training module, 110-output module. DETAILED DESCRIPTION

[0036] 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.

[0037] Example 1

[0038] like Figure 1 As shown, the lung multi-core MRI artificial intelligence reconstruction method based on image generation selects multiple MRI modalities for joint reconstruction. Each MRI modality can reflect similar lung anatomical structures and the information provided by each MRI modality can complement each other. In this embodiment, the modality selected is 1 H MRI and 129 Xe MRI, which 1 H MRI is the reference modality, 129 Xe MRI is the modality of interest.

[0039] The specific steps include:

[0040] Step 1: Get the training set and the test set, where both the training set and the test set include multiple sample groups, each sample group includes 2D full sampling 1 H MRI image, 2D full sampling 129 Xe MRI k-space data and 2D full sampling 129 2D undersampling of Xe MRI k-space data 129 Xe MRI k-space data and 2D full sampling 129 Xe MRI images, the specific method is as follows:

[0041] Step 1.1: Scan multiple volunteers to obtain 3D full sampling 1 H MRI k-space data, and 3D full sampling 129 Xe MRIk spatial data.

[0042] Step 1.2: Extract 3D full sampling layer by layer 1 H MRI k-space data, obtain multiple 2D full sampling 1 H MRI k-space data; extract 3D full sampling layer by layer 129Xe MRI k-space data, obtaining multiple 2D full samples 129 Xe MRI k-space data, 2D full sampling according to the under-sampling sampling matrix 129 Xe MRI k-space data is undersampled to obtain the corresponding 2D undersampling 129 Xe MRI k-space data.

[0043] Step 1.3: Full 2D sampling 1 H MRI k-space data is transformed by two-dimensional Fourier inverse transform to obtain 2D full sampling 1 HMRI images; 2D full sampling 129 Xe MRI k-space data is transformed into 2D full sampling by 2D inverse Fourier transform 129 Xe MRI image.

[0044] Step 1.4: Full 2D sampling of a volunteer 1 H MRI image, 2D full sampling 129 Xe MRI k-space data, 2D undersampling 129 Xe MRI k-space data and 2D full sampling 129 Xe MRI images are taken as a sample group, and all sample groups are divided into training sets and test sets. The training set and the test set both include multiple sample groups.

[0045] In this embodiment, 3D bSSFP sequence was used to scan 94 volunteers to obtain 3D full sampling 1 H MRI k-space data and 3D full sampling 129 Xe MRI k-space data. 3D full sampling 1 H MRI k-space data and 3D full sampling 129 The matrix size of Xe MRI k-space data is 96×96, and the number of layers is 24. Target modality 129 The Xe MRI sampling acceleration factor is 4 times, and the undersampling method is Cartesian sampling. The schematic diagram of the undersampling sampling matrix is ​​as follows Figure 2 As shown. Extract 3D full sampling layer by layer 1 H MRI k-space data, and 3D full sampling 129 Xe MRI k-space data, 2D full sampling 1 H MRI k-space data and 2D full sampling 129 Xe MRI k-space data. 2D full sampling based on undersampling sampling matrix 129 Xe MRI k-space data is undersampled to obtain the corresponding 2D undersampling 129 Xe MRI k-space data. 2D full sampling 1H MRI k-space data and 2D full sampling 129 Xe MRI k-space data is transformed into 2D full sampling by 2D inverse Fourier transform 1 H MRI images and 2D full sampling 129 Xe MRI images. A sample group of 84 volunteers was randomly selected as the training set, and a sample group of 10 volunteers was randomly selected as the test set.

[0046] Step 2: Construct a cross-modal generative network. The cross-modal generative network is based on the U-Net framework and includes an encoder and a decoder. The encoder includes multiple downsampling layers, and the decoder includes multiple upsampling layers. Each downsampling layer includes multiple real-valued convolution layers and a maximum pooling layer, wherein each downsampling layer has an activation layer after the real-valued convolution layer, and the downsampling layer uses continuous real-valued convolution operations and a maximum pooling operation to continuously reduce the spatial dimension of the image while improving its feature depth. Each upsampling layer includes multiple real-valued convolution layers and an upsampling real-valued convolution layer, wherein each upsampling layer has an activation layer after the real-valued convolution layer, and the upsampling layer uses continuous real-valued convolution operations and an upsampling real-valued convolution operation to gradually restore the spatial dimension of the image while gradually reducing the feature depth. The feature map of the upsampling layer at the corresponding level is channel-joined with the feature map of the downsampling layer through a jump connection.

[0047] In this embodiment, the encoder includes 4 downsampling layers, and the decoder includes 4 upsampling layers. The structural schematic diagrams of the upsampling layers and the downsampling layers are as follows: Figure 3 (a) and 3(b). Each downsampling layer includes three real-valued convolutional layers with a kernel size of 3×3 and a maximum pooling layer, where each real-valued convolutional layer of the downsampling layer is followed by a ReLu activation layer. Each upsampling layer includes three real-valued convolutional layers with a kernel size of 3×3 and an upsampling real-valued convolutional layer with a kernel size of 3×3, where each real-valued convolutional layer of the upsampling layer is followed by a ReLu activation layer. The feature maps of the upsampling layer at the corresponding level are channel-joined with the feature maps of the downsampling layer through a skip connection.

[0048] Step 3: Fully sample the training set in step 1 in 2D 1 The H MRI image is input into the cross-modal generation network constructed in step 2 to obtain the corresponding synthetic full sampling 129 Xe MRI images;

[0049] Step 4: construct a k-space reconstruction network, which includes multiple cascaded residual modules. Each residual module contains multiple complex-valued convolutional layers, and each complex-valued convolutional layer except the last one is followed by a corresponding activation layer. The final output of each residual module is the pixel-by-pixel sum of the output of the last complex-valued convolutional layer in the current residual module and the input of the first complex-valued convolutional layer in the current residual module.

[0050] In this embodiment, the k-space reconstruction network is composed of two cascaded residual modules. The structural diagram of the residual module is as follows: Figure 4 As shown in the figure. Each residual module contains 5 complex-valued convolutional layers. Except for the last complex-valued convolutional layer, each complex-valued convolutional layer is followed by a corresponding ReLu activation layer. The final output of each residual module is the sum of the output of the last complex-valued convolutional layer and the input of the first complex-valued convolutional layer pixel by pixel.

[0051] Step 5: The synthesized full sample obtained in step 3 129 Xe MRI images are transformed into synthetic full samples by two-dimensional Fourier transform 129 Xe MRI k-space data, and then synthesized full sampling 129 Xe MRI k-space data with 2D undersampling in the training set generated in step 1 129 The Xe MRI k-space data is spliced ​​through channels, and the obtained joint k-space data is input into the k-space reconstruction network constructed in step 4 to obtain the reconstructed 129 Xe MRI k-space data.

[0052] Step 6, construct an image domain reconstruction network, which includes multiple cascaded dense residual modules. Each dense residual module contains multiple complex-valued convolutional layers. Except for the last complex-valued convolutional layer, each complex-valued convolutional layer is followed by a corresponding activation layer. Each complex-valued convolutional layer and the corresponding activation layer serve as a first-level complex-valued convolutional processing layer. The dense residual module adopts a dense connection method. In each dense residual module, the input of the first-level complex-valued convolutional processing layer is the input of the dense residual module. Except for the first-level complex-valued convolutional processing layer, the input of each level of complex-valued convolutional processing layer is the result of channel concatenation of the input and output of the previous level of complex-valued convolutional processing layer. The output of the last complex-valued convolutional layer is added pixel by pixel to the input of the first-level complex-valued convolutional processing layer as the output of the current dense residual module.

[0053] The input of the first complex-valued convolutional layer of the first dense residual module is the reconstructed 129 Xe MRIk spatial data is reconstructed by two-dimensional Fourier inverse transform 129 Xe MRI image domain data and the synthetic full sampling obtained in step 3129 The joint data is obtained by channel stitching of Xe MRI images; the input of the first complex-valued convolutional layer of other dense residual modules except the first dense residual module is the final output of the previous dense residual module.

[0054] 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 5 As shown in Figure 2, each dense residual module contains 5 complex-valued convolutional layers, and each complex-valued convolutional layer except the last one is followed by a corresponding ReLu activation layer.

[0055] Step 7: Reconstruct the 129 Xe MRI k-space data is reconstructed by two-dimensional inverse Fourier transform 129 Xe MRI image domain data, and then reconstruct 129 Xe MRI image domain data and the synthetic full sampling obtained in step 3 129 The Xe MRI images are spliced ​​together and the resulting joint image data is input into the image domain reconstruction network constructed in step 6 to obtain the final reconstructed 129 Xe MRI images;

[0056] Step 8: Set the total loss function. The total loss function is the sum of the first loss function, the second loss function, and the third loss function. The first loss function is the synthetic full sampling obtained by the cross-modal generation network in step 3. 129 Xe MRI image with 2D full sampling in step 1 129 The error between Xe MRI images. In this embodiment, the error corresponding to the first loss function uses the mean square error (L2) loss. The second loss function is the reconstruction obtained by the k-space reconstruction network in step 5. 129 Xe MRI k-space data and 2D full sampling in step 1 129 The error between the Xe MRI k-space data, in this embodiment, the error corresponding to the second loss function uses the L2 loss. The third loss function is the final reconstruction obtained in step 7 129 Xe MRI image with 2D full sampling in step 1 129 The error between Xe MRI images, in this embodiment, the error corresponding to the third loss function adopts L2 loss.

[0057] Step 9. According to the total loss function set in step 8, the cross-modal generation network constructed in step 2, the k-space reconstruction network constructed in step 4, and the image domain reconstruction network constructed in step 6 are end-to-end trained using the training set generated in step 1. The learning rate of the network is initialized to 0.0001, the batch size is set to 10, and the network is trained on the Tensorflow 2.0 platform using the ADAM optimizer. The training is stopped after the total number of iterations reaches 200, and the parameters of all networks are saved.

[0058] Step 10: Fully sample the 2D in the test set generated in step 1 1 H MRI image, 2D undersampling 129 Xe MRI k-space data is input into the trained network to obtain the final reconstruction for testing 129 Xe MRI image.

[0059] Step 11: Reconstruct the test result obtained in step 10 129 Xe MRI image and 2D full sampling of the test set obtained in step 1 129 Xe MRI images were compared and Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) indicators were calculated.

[0060] Figure 6 Shows a 2D undersampling in the test set 129 The reconstruction results of Xe MRI k-space data, with PSNR / SSSIM values ​​marked at the bottom of the image. (a) is 2D full sampling 129 Xe MRI image, (b) is zero-filled 129 Xe MRI image, (c) is the final reconstruction obtained by the method of the present invention 129 Xe MRI image. As can be seen from the result figure, the lung multi-core MRI artificial intelligence reconstruction method based on image generation provided by the present invention can be used to reconstruct the image from high-power 2D undersampling. 129 High-quality reconstruction of Xe MRI k-space data 129 Xe MRI image.

[0061] Example 2

[0062] like Figure 7As shown, an embodiment of the present invention provides a lung multi-core MRI artificial intelligence reconstruction device based on image generation, which includes: an acquisition module 101, a first generation module 102, an input module 103, a second generation module 104, a first change module 105, a third generation module 106, a second change module 107, a loss function construction module 108, a training module 109 and an output module 110.

[0063] The acquisition module 101 is used to acquire a training set and a test set, wherein the training set and the test set each include multiple sample groups, each sample group includes a 2D full sample 1 H MRI image, 2D full sampling 129 Xe MRI k-space data and 2D full sampling 129 2D undersampling of Xe MRI k-space data 129 Xe MRI k-space data and 2D full sampling 129 Xe MRI image.

[0064] The first generation module 102 is used to construct a cross-modal generation network, and the cross-modal generation network is the same as that in Example 1.

[0065] The input module 103 is used to obtain the 2D full sampling in the training set obtained by the acquisition module 101 1 The H MRI image is input into the cross-modal generation network of the first generation module 102 to obtain a synthetic full sampling 129 Xe MRI image.

[0066] The second generation module 104 is used to construct a k-space reconstruction network. The k-space reconstruction network is the same as that in the first embodiment.

[0067] The first changing module 105 is used to convert the synthesized full sample 129 Xe MRI images are transformed into synthetic full samples by two-dimensional Fourier transform 129 Xe MRI k-space data, and then synthesized full sampling 129 Xe MRI k-space data and the corresponding 2D undersampling in the training set obtained by acquisition module 101 129 The Xe MRI k-space data is channel-joined, and the obtained joint k-space data is input into the k-space reconstruction network constructed by the second generation module 104 to obtain a reconstructed 129 Xe MRI k-space data.

[0068] The third generation module 106 is used to construct an image domain reconstruction network. The image domain reconstruction network is the same as that in the first embodiment.

[0069] The second change module 107 is used to reconstruct 129Xe MRI k-space data is reconstructed by two-dimensional inverse Fourier transform 129 Xe MRI image domain data, and then reconstruct 129 Full sampling of Xe MRI image domain data and synthesis 129 The Xe MRI images are spliced ​​through channels to obtain joint image data and the joint image data is input into the image domain reconstruction network to obtain the final reconstructed 129 Xe MRI image.

[0070] The loss function construction module 108 is used to set the total loss function. The total loss function is the sum of the first loss function, the second loss function and the third loss function. The first loss function is the synthetic full sampling output by the cross-modal generation network in the input module 103 129 Xe MRI Image and Acquisition Module 101 2D Full Sampling 129 The second loss function is the reconstructed k-space reconstruction network obtained in the first variation module 105. 129 Xe MRI k-space data and 2D full sampling in acquisition module 101 129 The third loss function is the final reconstruction obtained by the image domain reconstruction network in the second transformation module 107. 129 Xe MRI Image and Acquisition Module 101 2D Full Sampling 129 The error between Xe MRI images. The total loss function of this embodiment is set the same as that of embodiment 1.

[0071] The training module 109 is used to perform end-to-end training on the constructed cross-modal generation network, the image domain reconstruction network, and the image domain reconstruction network according to the total loss function constructed by the loss function construction module 108 and using the training set of the acquisition module 101, and output the trained network.

[0072] The output module 110 is used to convert the 2D full sampling in the test set 1 H MRI image, 2D undersampling 129 Xe MRI k-space data is input into the trained network to obtain the final reconstruction for testing 129 Xe MRI image.

[0073] 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. An artificial intelligence reconstruction method for lung multi-core MRI based on image generation, characterized in that: The steps include: Step 1, obtaining a training set and a test set, wherein the training set and the test set each include multiple sample groups, each sample group includes a 2D fully sampled reference modality image, a 2D fully sampled target modality k-space data, a 2D under-sampled target modality k-space data corresponding to the 2D fully sampled target modality k-space data, and a 2D fully sampled target modality image; Step 2: Build a cross-modal generation network; Step 3: Input the 2D fully sampled reference modality image in the training set generated in step 1 into the cross-modality generation network constructed in step 2 to obtain the corresponding synthesized fully sampled target modality image; Step 4: construct a k-space reconstruction network; Step 5: Perform a two-dimensional Fourier transform on the synthesized fully sampled target modality image obtained in step 3 to obtain synthesized fully sampled target modality k-space data, and then perform a channel splicing operation on the synthesized fully sampled target modality k-space data and the 2D under-sampled target modality k-space data in the training set generated in step 1. The obtained joint k-space data is input into the k-space reconstruction network constructed in step 4 to obtain the corresponding reconstructed target modality k-space data; Step 6: Construct an image domain reconstruction network; Step 7, performing a two-dimensional inverse Fourier transform on the reconstructed target modality k-space data obtained in step 5 to obtain reconstructed target modality image domain data, and then performing a channel stitching operation on the reconstructed target modality image domain data and the synthetic full-sampled target modality image obtained in step 3, and the obtained joint image data is input into the image domain reconstruction network constructed in step 6 to obtain the corresponding final reconstructed target modality image; Step 8: Set the total loss function; Step 9: According to the total loss function set in step 8, the cross-modality generation network, the k-space reconstruction network and the image domain reconstruction network are trained, and the parameters of all networks are saved; Step 10, input the 2D fully sampled reference modality image and the 2D under-sampled target modality k-space data in the test set generated in step 1 into the network trained in step 9 to obtain the final reconstructed target modality image for the test; The reference mode is 1 H MRI, target modality is 129 Xe MRI.

2. The method for artificial intelligence reconstruction of lung multi-core MRI based on image generation according to claim 1, characterized in that: The sample group of step 1 is obtained by the following steps: Step 1.1, scanning multiple volunteers to obtain 3D full-sampling reference modality k-space data and 3D full-sampling target modality k-space data; Step 1.2, extracting 3D fully sampled reference modality k-space data layer by layer to obtain multiple 2D fully sampled reference modality k-space data, extracting 3D fully sampled target modality k-space data layer by layer to obtain multiple 2D fully sampled target modality k-space data, and undersampling the 2D fully sampled target modality k-space data according to the under-sampling sampling matrix to obtain corresponding 2D under-sampling target modality k-space data; Step 1.3, perform a two-dimensional inverse Fourier transform on the 2D fully sampled reference modality k-space data to obtain a 2D fully sampled reference modality image; perform a two-dimensional inverse Fourier transform on the 2D fully sampled target modality k-space data to obtain a 2D fully sampled target modality image.

3. The method for artificial intelligence reconstruction of lung multi-core MRI based on image generation according to claim 1, characterized in that: In step 2, the cross-modal generation network includes an encoder and a decoder; the encoder includes multiple downsampling layers, and the decoder includes multiple upsampling layers; each downsampling layer includes multiple real-valued convolution layers and a maximum pooling layer, wherein each downsampling layer has an activation layer after the real-valued convolution layer; each upsampling layer includes multiple real-valued convolution layers and an upsampling real-valued convolution layer, wherein each upsampling layer has an activation layer after the real-valued convolution layer; the upsampling feature maps and the downsampling feature maps of the corresponding levels are channel-joined through a jump connection.

4. The method for artificial intelligence reconstruction of lung multi-core MRI based on image generation according to claim 1, characterized in that: The k-space reconstruction network in step 4 includes multiple cascaded residual modules; each residual module includes multiple complex-valued convolutional layers, and a corresponding activation layer is added after each complex-valued convolutional layer except the last complex-valued convolutional layer; the final output of each residual module is the pixel-by-pixel sum of the output of the last complex-valued convolutional layer in the current residual module and the input of the first complex-valued convolutional layer in the current residual module.

5. The method for artificial intelligence reconstruction of lung multi-core MRI based on image generation according to claim 1, characterized in that: The image domain reconstruction network in step 6 includes multiple cascaded dense residual modules; each dense residual module includes multiple complex-valued convolutional layers, and each complex-valued convolutional layer except the last one is followed by a corresponding activation layer; each complex-valued convolutional layer and the corresponding activation layer serve as a first-level complex-valued convolution processing layer, and in each dense residual module, the input of the first-level complex-valued convolution processing layer is the input of the dense residual module, and the input of each level of complex-valued convolution processing layer except the first-level complex-valued convolution processing layer is the result of channel concatenation of the input and output of the previous level of complex-valued convolution processing layer, and the result of pixel-by-pixel addition of the output of the last complex-valued convolutional layer and the input of the first-level complex-valued convolution processing layer serves as the output of the current dense residual module.

6. The method for artificial intelligence reconstruction of lung multi-core MRI based on image generation according to claim 1, characterized in that: The total loss function in step 8 is the sum of the first loss function, the second loss function and the third loss function, wherein the first loss function is the error between the fully sampled target modality image synthesized by the cross-modality generation network in step 3 and the 2D fully sampled target modality image in step 1, the second loss function is the error between the reconstructed target modality k-space data obtained by the k-space reconstruction network in step 5 and the 2D fully sampled target modality k-space data in step 1, and the third loss function is the error between the final reconstructed target modality image obtained in step 7 and the 2D fully sampled target modality image in step 1.

Images