An artificial intelligence multi-contrast magnetic resonance rapid imaging method

Abstract

An artificial intelligence multi-contrast magnetic resonance rapid imaging method relates to a multi-contrast magnetic resonance rapid imaging method. An artificial intelligence multi-contrast magnetic resonance rapid imaging method with fast reconstruction speed, good reconstruction quality and good interpretability is provided. Multi-channel magnetic resonance images of different contrasts are obtained, and a training set that meets the network training requirements is obtained by synthesizing multiple channels into a single channel; a complex neural network based on joint sparse constraints is built according to the traditional optimization iterative calculation process; a network loss function is established; network parameters are trained; and under-sampled multi-contrast magnetic resonance images are reconstructed using the trained network model. At the same time, the interpretability of traditional algorithms and the strong learning ability of deep networks are taken into account, and complex convolutional networks are applied to directly process complex data, and a joint sparse constraint module is designed to better utilize the structural information between different contrasts. It has the characteristics of fast reconstruction speed, good reconstruction quality and good interpretability.

Description

Technical Field

[0001] The present invention relates to a multi-contrast magnetic resonance rapid imaging method, and in particular to an artificial intelligence multi-contrast magnetic resonance rapid imaging method based on a complex network with joint sparse constraints, guided by a traditional algorithm. Background Art

[0002] Magnetic Resonance Imaging (MRI) is an important clinical auxiliary diagnostic tool. MRI often acquires multiple contrast MRI images in clinical diagnosis because different contrasts can provide richer structural information. However, the increase in the number of acquired images also directly increases the scanning time of the instrument. To this end, researchers use compressed sensing theory to undersample magnetic resonance signals to reduce sampling time, and then use reconstruction algorithms to obtain high-quality images that meet clinical diagnostic requirements, thereby achieving high-quality and fast magnetic resonance imaging. Before deep learning, researchers tried various constraints to exploit the structural similarity between different contrasts, including joint sparse features (Zongying Lai, Xinlin Zhang, Di Guo, Xiaofeng Du, Yonggui Yang, Gang Guo, Zhong Chen, Xiaobo Qu. Joint sparse reconstruction of multi-contrast MRI images with graph based redundant wavelet transform. BMC Medical Imaging, 18(1): 1-6, 2018) and joint image gradient information (Berkin Bilgic, Vivek K Goyal, Elfar Adalsteinsson, Multi-contrastreconstruction with Bayesian compressed sensing. Magnetic Resonance in Medicine, 66(6): 1601-1615, 2011). However, these traditional iterative reconstruction algorithms still take a long time to run and require finding effective prior information in advance and manually adjusting hyperparameters.

[0003] In recent years, deep learning has developed rapidly in the field of biomedical magnetic resonance (Xiaobo Qu, Yihui Huang, Hengfa Lu, Tianyu Qiu, Di Guo, Dr. Tatiana Agback, Vladislav Orekhov, Zhong Chen. Accelerated nuclear magnetic resonance spectroscopy with deep learning. Angewandte Chemie International Edition, 132(26): 10383-10386, 2020; Dichen Chen, Zi Wang, Di Guo, Vladislav Orekhov, Xiaobo Qu, Review and prospect: Deep learning in nuclear magnetic resonance spectroscopy. Chemistry - A European Journal, 26: 10391-10401, 2020; Zi Wang, Di Guo, Zhangren Tu, Yihui Huang, Yirong Zhou, Jian Wang, Liubin Feng, Donghai Lin, Yongfu You, Tatiana Agback, Vladislav Orekhov, Xiaobo Qu, XCloud - MoDern: An artificial intelligence cloud for accelerated NMR spectroscopy. arXiv:2012.14830, 2020.; Yihui Huang, Jinkui Zhao, Zi Wang, Di Guo, Xiaobo Qu, Exponential signal reconstruction with deep hankel matrix factorization. arXiv:2007.06246, 2020; Tieyuan Lu, Xinlin Zhang, Yihui Huang, Di Guo, Feng Huang, Qin Xu, Yuhan Hu, Lin Ou - Yang, Jianzhong Lin, Zhiping Yang, Xiaobo Qu, pFISTA - SENSE - ResNet for parallel MRI reconstruction.Journal of Magnetic Resonance, 318: 106790, 2020). In the field of multi-contrast MRI image reconstruction based on deep learning, Sun et al. (Liyan Sun, Zhiwen Fan, Xueyang Fu, Yue Huang, Xinghao Ding, John Paisley, A deep information sharing network for multi-contrast compressed sensing MRI reconstruction. IEEE Transactions on Image Processing, 28 (12), 6141-6153, 2019) proposed a method to use a dense residual network to achieve an end-to-end mapping method from multi-contrast undersampled images to fully sampled images. This method has lower image error and faster reconstruction speed than traditional algorithms. However, the method proposed by Sun et al. is for real-number images and is difficult to directly apply to the scene where actual MRI is complex data. At the same time, it also fails to fully utilize the structural similarity between different contrasts. . Summary of the invention

[0004] The purpose of the present invention is to provide an artificial intelligence multi-contrast magnetic resonance rapid imaging method with fast reconstruction speed, good reconstruction quality and good interpretability.

[0005] The present invention comprises the following steps:

[0006] 1) Obtain multi-channel magnetic resonance images with different contrasts, and obtain a training set that meets the network training requirements by synthesizing multiple channels into a single channel;

[0007] 2) Building a complex neural network based on joint sparse constraints according to the traditional optimization iterative calculation process;

[0008] 3) Establish network loss function;

[0009] 4) Training network parameters;

[0010] 5) Use the trained network model to reconstruct undersampled multi-contrast magnetic resonance images.

[0011] In step 1), the specific method of obtaining multi-channel magnetic resonance images with different contrasts and synthesizing the multi-channels into a single channel to obtain a training set that meets the network training requirements is as follows:

[0012] The complete multi-channel magnetic resonance data obtained from the magnetic resonance imager can be expressed as X after Fourier transformation. c =[x c,1 ,x c,2,...,x c,J ],in, N, J, and c represent the number of image pixels, the number of channels, and the cth contrast image, respectively. represents the complex domain; since single-channel data is processed, multiple channels need to be merged into a single channel, and the ESPIRiT (Martin Uecker, Peng Lai, Mark J. Murphy, Patrick Virtue, Michael Elad, John M. Pauly, Shreyas S. Vasanawala, Michael Lustig, ESPIRiT—an eigenvalue approach to autocalibrating parallel MRI: Where SENSE meets GRAPPA. Magnetic Resonance in Medicine, 71(3):990-1001, 2014) method is used to complete the merging of single channels. The synthesis method is as follows:

[0013]

[0014] ES(·) indicates that multiple channels are synthesized into a single channel using the ESPIRiT method. represents the composite image of the c-th contrast image, As labels during network training; use undersampling matrices and Fourier transform right Perform the operation to obtain the undersampled image μ indicates that the data is undersampled, M indicates the number of sampling points, and the specific operation is:

[0015]

[0016] in,(·) H 、(·) T Represents complex conjugate transpose and transpose respectively; through the above operations, the training set for network model training is obtained L represents the data in the training set (X u ,X combined ) logarithm,

[0017] In step 2), the complex neural network based on joint sparse constraints takes the iteration block as the core, and forms the whole network by superimposing several iteration blocks; each iteration block includes three modules: a data verification module, a sparse learning module and a joint sparse constraint module, wherein the joint sparse constraint module is the only position for exchanging information between images with different contrasts in the whole iteration block, and is also the most critical module; the whole network can be represented by a mapping function, i.e., f(X u ; Θ), Θ represents the set of network training parameters, and the specific content of each module is as follows:

[0018] 2.1 The data verification module is used to ensure the consistency between the reconstructed image and the measured data: The input of the sth iteration block can be expressed as s represents the index of the iteration block, Representing the input of the c-th contrast MRI image at the s-th iteration block, the data verification block performs the following operations:

[0019]

[0020] Each contrast image is individually verified, γ s represents the step size of the sth iteration block and satisfies γ s ∈Θ, U and F represent the undersampling matrix and Fourier transform respectively, represents the undersampled image of the cth contrast, It represents the output of the c-th contrast image after the s-th data verification module. When s=1,

[0021] 2.2 The sparse learning module is used to learn a sparse transformation, the purpose of which is to make the input image as sparse as possible: it includes forward sparse operation and reverse sparse operation, respectively. and Indicates and satisfies The output of the data verification block is used as the input of the forward sparse operation. The sparse operation is replaced by three layers of complex convolutional layers, each layer contains 24 feature maps, the convolution kernel size is 3×3, and the activation function uses the complex linear rectifier function (Complex Rectified Linear Unit, CReLU) (Elizabeth K. Cole, Joseph Y. Cheng, John M. Pauly, Shreyas S. Vasanawala, Analysis of deep complex-valued convolutional neural networks for MRI reconstruction. arXiv: 2004.01738, 2020). The forward sparse operation is as follows:

[0022]

[0023] is the output of the c-th contrast MRI image at the s-th forward sparse operation; as with the data verification module, each contrast image is operated separately; it should be noted that in the same iteration block, different contrast images share network weights when performing forward sparse operations, that is, only one set of The reverse sparse operation is similar to the forward operation. Its purpose is to simulate the inverse operation of the sparse conversion in the traditional algorithm. The number of network layers, convolution kernels, and activation functions are the same as those of the forward sparse operation. The difference is that the input of the reverse operation is the output of the joint sparse module. The operation is as follows:

[0024]

[0025] in, represents the output of the c-th contrast MRI after the s-th reverse sparse operation, is the output of the c-th contrast MRI image in the s-th joint sparse constraint module, which will be introduced in the next step 2.3; as with the forward sparse operation, the network weights are also shared in the reverse sparse operation;

[0026] 2.3 Joint sparse constraint module is used to constrain the common sparse characteristics of images with different contrasts. It is also the only module in the entire iterative block that exchanges information with different contrasts: This module contains two operations, namely the group operation G and the group soft threshold operation where λ represents the penalty parameter and satisfies λ∈Θ, ||·|| 2 Represents the two norm of the vector; the process of group operation G is as follows: For the output of the forward sparse operation After the forward sparse operation, the size is n×h×w×24 (number of training samples each time×image height×image width×number of output feature maps), that is, Then, the corresponding feature maps obtained after the forward sparse operation of the magnetic resonance images with different contrasts are combined and used Represents the characteristics of the 24 groups formed, where The vector form of the kth feature map representing the cth contrast image; finally, the pixels corresponding to each group of feature maps are combined into a vector for group threshold operation. Therefore, the joint sparse constraint module in the sth iteration block can be expressed as:

[0027]

[0028] In addition, in order to accelerate network training and improve network performance, residual structures are used within and between iteration blocks. Within the iteration block, the residual structure is defined as:

[0029]

[0030] in, represents the output of the c-th contrast magnetic resonance image in the s-th iteration block;

[0031] The residual structure between iterative blocks is reflected in the fact that the input of the s+1th iterative block is the weighted sum of the outputs of the previous two iterative blocks, and the mathematical formula is expressed as:

[0032]

[0033] Among them, μ s Used to balance the output of the first two iterative blocks and satisfy μ s ∈Θ. The parameters {γ,λ,μ} will be initialized accordingly.

[0034] In step 3), considering that the output of each iterative block affects the final reconstruction result, the loss function of the network is as follows:

[0035]

[0036] Where C and S represent the number of contrast images and the number of network iteration blocks, respectively. |·| represents the modulus of each complex number in the vector. represents the square of the vector's two-norm, Represents the output of the c-th contrast image in the s-th iteration block.

[0037] In step 4), the Adam algorithm in deep learning (Diederik P. Kingma, Jimmy Ba, Adam: A method for stochastic optimization. arXiv: 1412.6980, 2014.) is used to implement the back propagation of the network to train and update the network parameters. The trained parameter set is used express.

[0038] In step 5), the network model has been trained through step 4), that is, the network mapping function f(·) has been established; the collected multi-contrast undersampled image is input into the network, and the network finally outputs the reconstruction result of the corresponding contrast magnetic resonance image.

[0039]

[0040] in represents the multi-contrast magnetic resonance data to be reconstructed, Represents the reconstruction result corresponding to the c-th contrast magnetic resonance image.

[0041] The present invention proposes a method for building a multi-contrast magnetic resonance reconstruction network based on a traditional optimization algorithm, and designs a neural network using a joint sparse reconstruction iterative algorithm of multi-contrast images to reduce reconstruction errors.

[0042] The present invention first obtains single-channel undersampled and fully sampled multi-contrast magnetic resonance images as a training set for the network, builds a complex neural network based on joint sparse constraints according to the traditional optimization iterative calculation process, then uses the processed training set to train the network parameters, and finally uses the trained network model to reconstruct the undersampled multi-contrast magnetic resonance images. Compared with the existing methods, the present invention takes into account both the interpretability of traditional algorithms and the strong learning ability of deep networks, and simultaneously applies complex convolutional networks to directly process complex data, and designs a joint sparse constraint module to better utilize the structural information between different contrasts, and has the characteristics of fast reconstruction speed, good reconstruction quality and good interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is an undersampling template that samples 15% of the data in the embodiment. Figure 1 In the figure, (a) is the undersampling template of T1 contrast MRI image, and (b) is the undersampling template of T2 contrast MRI image.

[0044] Figure 2 The overall structure diagram of the multi-contrast magnetic resonance image reconstruction network model based on joint sparsity constraints and the expanded diagram of the sth iterative block.

[0045] Figure 3 is the network reconstruction result and the corresponding error graph. Figure 3In the figure, (a) and (f) are the fully sampled T1 and T2 contrast MRI images, respectively; (b) and (g) are the T1 reconstructed MRI images and T2 reconstructed MRI images based on dense connection and residual learning network proposed by Sun et al.; (d) and (i) are the corresponding error maps (Liyan Sun, Zhiwen Fan, Xueyang Fu, Yue Huang, Xinghao Ding, John Paisley, A deep information sharing network for multi-contrast compressed sensing MRI reconstruction. IEEE Transactions on Image Processing, 28(12): 6141-6153, 2019) It should be noted that since this method was proposed to process real data, its real convolutional network is changed to a complex convolutional network, and the reconstruction results show the modified results; (c) and (h) are the T1 reconstructed MRI images and T2 reconstructed MRI images of the present invention; (e) is the corresponding error map. DETAILED DESCRIPTION

[0046] The following embodiments will further illustrate the present invention in conjunction with the accompanying drawings.

[0047] The embodiment of the present invention includes the following steps:

[0048] Step 1: Obtain multi-channel magnetic resonance images with different contrasts, and synthesize the multi-channels into a single channel to obtain a training set that meets the network training requirements and a test set for testing.

[0049] The present invention uses a 3 Tesla magnetic resonance imaging device to image the brains of 5 volunteers, of which the data of 4 volunteers are used for training and the data of 1 volunteer is used for testing. The imaging parameters are (T1: repetition time = 6.9 ms, sequence echo = 2.5 ms, field of view = 256 mm 2 , thickness = 1mm; T2: repetition time = 2500ms, sequence echo = 74.8, field of view = 256mm 2 , thickness = 1mm), T1 / T2 represents two contrast images, and the data size collected for each volunteer is 256×256×186×12 (height×length×width×number of channels). The cross-sectional images and the middle 61 layers of 4 volunteers are selected as the training set. For this purpose, the 12-channel data of the contrast magnetic resonance image c can be expressed as X after Fourier transformation. c =[x c,1 ,x c,2 ,...,x c,12 ],in Represents the complex domain. c = {T1, T2}, because it is processing single channel data, it is necessary to merge multiple channels into a single channel. The present invention adopts the ESPIRiT (Martin Uecker, Peng Lai, Mark J. Murphy, Patrick Virtue, Michael Elad, John M. Pauly, Shreyas S. Vasanawala, Michael Lustig, ESPIRiT—an eigenvalue approach to autocalibrating parallelMRI: Where SENSE meets GRAPPA. Magnetic Resonance in Medicine, 71 (3): 990-1001, 2014.) method to complete the merging of single channels, and the synthesis method is as follows:

[0050]

[0051] ES(·) indicates that multiple channels are synthesized into a single channel using the ESPIRiT method. represents the composite image of the cth contrast image, In the network training process, it is used as a label. Perform the operation to obtain the undersampled image μ indicates that the data is undersampled. The embodiment uses a 15% undersampled template (e.g. Figure 1 The specific operations are as follows:

[0052]

[0053] in,(·) H 、(·) T Represents complex conjugate transpose and transpose respectively. Through the above operations, the training set for network model training is obtained 244 represents the training set data (X u ,X combined ) logarithm, In addition, a 180° rotation operation is performed on the training set to expand the data set. At this time, the number of pairs of training set data becomes 488.

[0054] The same method can be used to process the 61-layer data from another volunteer to obtain the network test set

[0055] Step 2: Build a complex neural network based on joint sparse constraints according to the traditional optimization iterative calculation process

[0056] The complex network model based on joint sparse constraint takes the iteration block as the core, and the whole network is constructed by superimposing several iteration blocks. Each iteration block contains three modules: data verification module, sparse learning module and joint sparse constraint module, among which the joint sparse constraint module is the only location for exchanging information between images with different contrast in the whole iteration block, and it is also the most critical module. The whole network can be represented by a mapping function, that is, f(X u ; Θ), Θ represents the set of network training parameters. Figure 2 Given the expansion diagram of the sth iteration block, the following is based on Figure 2 Explain each module in the sth iteration block:

[0057] a) The data verification module is used to ensure the consistency between the reconstructed image and the measured data: the input of the sth iteration block can be expressed as s represents the index of the iteration block, Represents the input of the T1 contrast magnetic resonance image at the sth iteration block. The data verification block performs the following operations:

[0058]

[0059] Each contrast image is individually verified, γ s represents the step size of the sth iteration block and satisfies γ s ∈Θ, U and F represent the undersampling matrix and Fourier transform respectively, represents the undersampled image of the cth contrast, It represents the output of the c-th contrast image after the s-th data verification module. When s=1,

[0060] b) The sparse learning module is used to learn a sparse transformation: it includes a forward sparse operation and a reverse sparse operation, respectively. and Indicates and satisfies The output of the data verification block is used as the input of the forward sparse operation. The sparse operation is replaced by three layers of complex convolutional layers, each layer contains 24 feature maps, the convolution kernel size is 3×3, and the activation function uses the complex linear rectifier function (Complex Rectified Linear Unit, CReLU) (Elizabeth K. Cole, Joseph Y. Cheng, John M. Pauly, Shreyas S. Vasanawala, Analysis of deep complex-valued convolutional neural networks for MRI reconstruction.arXiv:2004.01738,2020.). The forward sparse operation is as follows:

[0061]

[0062] is the output of the c-th contrast MRI image in the s-th forward sparse operation. As with the data verification module, each contrast image is operated separately. It should be noted that in the same iteration block, different contrast images share network weights when performing forward sparse operations, that is, only one set of The reverse sparse operation is similar to the forward operation, and its purpose is to simulate the inverse operation of the sparse conversion in the traditional algorithm. The number of network layers, convolution kernels, and activation functions are the same as the forward sparse operation. The difference is that the input of the reverse operation is the output of the joint sparse constraint module. The operation is as follows:

[0063]

[0064] in, represents the output of the c-th contrast MRI after the s-th reverse sparse operation, is the output of the c-th contrast MRI in the s-th joint sparse constraint module. The joint sparse constraint module will be introduced in the next step c). As with the forward sparse operation, the network weights are also shared in the reverse sparse operation.

[0065] c) The joint sparse constraint module is used to constrain the common sparse characteristics of images with different contrasts. It is also the only module in the entire iterative block that exchanges information with different contrasts: This module contains two operations, namely the group operation G and the group soft threshold operation where λ represents the penalty parameter and satisfies λ∈Θ, ||·|| 2 Represents the two-norm of the vector. The process of group operation G is as follows: For the output of the forward sparse operation Its size is n×h×w×24 (number of training samples each time×image height×image width×number of output feature maps), that is Then, the corresponding feature maps obtained after the forward sparse operation of the magnetic resonance images with different contrasts are combined and used Indicates the characteristics of the 24 groups formed, The vector form of the kth feature map representing the T1 contrast image. Finally, the pixel points corresponding to each group of feature maps are subjected to group threshold operation. Therefore, the joint sparse constraint module in the sth iteration block can be expressed as:

[0066]

[0067] In addition, in order to accelerate network training and improve network performance, residual structures are used in the iteration block kernel and between iteration blocks. In the iteration block, the residual structure is defined as:

[0068]

[0069] in, Represents the output of the c-th contrast magnetic resonance image in the s-th iteration block.

[0070] The residual structure between iterative blocks is reflected in the fact that the input of the s+1th iterative block is the weighted sum of the outputs of the previous two iterative blocks, and the mathematical formula is expressed as:

[0071]

[0072] Among them, μ s Used to balance the output of the first two iterative blocks and satisfy μ s ∈Θ. The parameters {γ,λ,μ} are initialized accordingly. The parameters {γ,λ,μ} are initialized to {1,0.001,0.5}.

[0073] Step 3: Establish network loss function

[0074] Considering that the output of each iterative block affects the final reconstruction result, the loss function of the network is as follows:

[0075]

[0076] Where C and S represent the number of contrast images and the number of network iteration blocks, respectively. |·| represents the modulus of each complex number in the vector. represents the square of the vector's two-norm, It represents the output of the c-th contrast image in the s-th iteration block, taking into account both image reconstruction quality and reconstruction time, and setting S=8.

[0077] Step 4: Train network parameters

[0078] The Adam algorithm in deep learning (Diederik P. Kingma, Jimmy Ba, Adam: A method for stochastic optimization. arXiv: 1412.6980, 2014) is used to implement the back propagation of the network to achieve the purpose of training and updating network parameters. The trained parameter set is used express.

[0079] Step 5: Reconstruct undersampled multi-contrast MRI images using the trained network model

[0080] After the fourth step, the network model has been trained, that is, the network mapping function f(·) has been established. The collected multi-contrast undersampled magnetic resonance images are input into the network, and the final output of the network represents the reconstruction result of the corresponding contrast magnetic resonance image:

[0081]

[0082] in, They represent the reconstruction results corresponding to the T1 / T2 contrast MRI undersampled images.

[0083] In the embodiment, the T1 / T2 contrast full-sampling magnetic resonance image and the reconstructed magnetic resonance image at a sampling rate of 15% and their corresponding error graphs are shown in FIG. Figure 3 shown.

[0084] It can be seen that an artificial intelligence multi-contrast magnetic resonance imaging method can quickly reconstruct high-quality magnetic resonance images, and is superior to the reconstruction results of the dense residual convolutional neural network mentioned in the introduction (Liyan Sun, Zhiwen Fan, Xueyang Fu, Yue Huang, Xinghao Ding, John Paisley, A deep information sharing network for multi-contrast compressed sensing MRI reconstruction. IEEE Transactions on Image Processing, 28(12): 6141-6153, 2019).

[0085] The present invention proposes an artificial intelligence multi-contrast magnetic resonance fast imaging method. This method builds a neural network based on traditional iterative reconstruction, constructs a joint sparse constraint module to utilize structural information of different contrasts, and uses a complex convolutional network to directly process complex data. The invented method has the characteristics of fast reconstruction speed, good reconstructed image quality, and good network interpretability.

Claims

1. An artificial intelligence multi-contrast magnetic resonance rapid imaging method, characterized in that The following steps are involved: 1) Obtain multi-channel magnetic resonance images with different contrasts, and obtain a training set that meets the network training requirements by synthesizing multiple channels into a single channel; 2) Building a complex neural network based on joint sparse constraints according to the traditional optimization iterative calculation process; The complex neural network based on joint sparse constraint takes the iteration block as the core, and forms the whole network by superimposing several iteration blocks; each iteration block contains three modules: data verification module, sparse learning module and joint sparse constraint module, wherein the joint sparse constraint module is the only position for exchanging information between images with different contrast in the whole iteration block, and is also the most critical module; the whole network is represented by a mapping function, i.e., f(X u ; Θ), Θ represents the set of network training parameters, and the specific content of each module is as follows: 2.1 The data verification module is used to ensure the consistency between the reconstructed image and the measured data: The input of the sth iteration block is expressed as s represents the index of the iteration block, Representing the input of the c-th contrast MRI image at the s-th iteration block, the data verification block performs the following operations: Each contrast image is individually verified, γ s represents the step size of the sth iteration block and satisfies γ s ∈Θ, U and F represent the undersampling matrix and Fourier transform respectively, represents the undersampled image of the cth contrast, represents the output of the c-th contrast image after the s-th data verification module; when s=1, 2.2 The sparse learning module is used to learn a sparse transformation, the purpose of which is to make the input image as sparse as possible: it includes forward sparse operation and reverse sparse operation, respectively. and Indicates and satisfies The output of the data verification block is used as the input of the forward sparse operation. The sparse operation is replaced by three layers of complex convolution layers. Each layer contains 24 feature maps. The convolution kernel size is 3×3. The activation function uses a complex linear rectification function. The forward sparse operation is as follows: is the output of the c-th contrast MRI image at the s-th forward sparse operation; as with the data verification module, each contrast image is operated separately; it should be noted that in the same iteration block, different contrast images share network weights when performing forward sparse operations, that is, only one set of The reverse sparse operation is similar to the forward operation. Its purpose is to simulate the inverse operation of the sparse conversion in the traditional algorithm. The number of network layers, convolution kernels, and activation functions are consistent with the forward sparse operation. The difference is that the input of the reverse operation is the output of the joint sparse module. The operation is as follows: in, represents the output of the c-th contrast MRI after the s-th reverse sparse operation, is the output of the c-th contrast MRI image in the s-th joint sparse constraint module; as with the forward sparse operation, the network weights are also shared in the reverse sparse operation; 2.3 Joint sparse constraint module is used to constrain the common sparse characteristics of images with different contrasts. It is also the only module in the entire iterative block that exchanges information with different contrasts: This module contains two operations, namely the group operation G and the group soft threshold operation Where λ represents the penalty parameter and satisfies λ∈Θ, ||·||2 represents the bi-norm of the vector; the process of the group operation G is as follows: For the output of the forward sparse operation After the forward sparse operation, the size is n×h×w×24, where n is the number of training samples each time, h is the image height, w is the image width, and 24 is the number of output feature maps, that is, Then, the corresponding feature maps obtained after the forward sparse operation of the magnetic resonance images with different contrasts are combined and used Represents the characteristics of the 24 groups formed, where The vector form of the kth feature map representing the cth contrast image; finally, the pixels corresponding to each group of feature maps are combined into a vector for group threshold operation. Therefore, the joint sparse constraint module in the sth iteration block is expressed as: In addition, in order to accelerate network training and improve network performance, residual structures are used within and between iteration blocks. Within the iteration block, the residual structure is defined as: in, represents the output of the c-th contrast magnetic resonance image in the s-th iteration block; The residual structure between iterative blocks is reflected in the fact that the input of the s+1th iterative block is the weighted sum of the outputs of the previous two iterative blocks, and the mathematical formula is expressed as: Among them, μ s Used to balance the output of the first two iterative blocks and satisfy μ s ∈Θ, the parameters {γ,λ,μ} will be initialized accordingly; 3) Establish network loss function; 4) Training network parameters; 5) Use the trained network model to reconstruct undersampled multi-contrast magnetic resonance images.

2. An artificial intelligence multi-contrast magnetic resonance rapid imaging method as claimed in claim 1, characterized in that in step 1), the specific method of obtaining multi-channel magnetic resonance images with different contrasts and synthesizing multiple channels into a single channel to obtain a training set that meets the network training requirements is as follows: The complete multi-channel magnetic resonance data obtained from the magnetic resonance imager is represented by X after Fourier transformation. c =[x c,1 ,x c,2 ,...,x c,J ],in, N, J, and c represent the number of image pixels, the number of channels, and the cth contrast image, respectively. Represents the complex domain; since single-channel data is processed, multiple channels need to be merged into a single channel. The ESPIRiT method is used to complete the merging of multiple channels. The synthesis method is as follows: ES(·) indicates that multiple channels are synthesized into a single channel using the ESPIRiT method. represents the composite image of the c-th contrast image, As labels during network training; then use the undersampling matrix and Fourier transform right Perform the operation to obtain the undersampled image μ indicates that the data is undersampled, M indicates the number of sampling points, and the specific operation is: in,(·) H 、(·) T Represents complex conjugate transpose and transpose respectively; through the above operations, the training set for network model training is obtained L represents the data in the training set (X u ,X combined ) logarithm, 3. The artificial intelligence multi-contrast magnetic resonance rapid imaging method as claimed in claim 1, characterized in that In step 3), considering that the output of each iterative block affects the final reconstruction result, the loss function of the network is as follows: Where C and S represent the number of contrast images and the number of network iteration blocks, respectively. |·| represents the modulus of each complex number in the vector. represents the square of the vector's two-norm, Represents the output of the c-th contrast image in the s-th iteration block.

4. The artificial intelligence multi-contrast magnetic resonance rapid imaging method as claimed in claim 1, characterized in that In step 4), the Adam algorithm in deep learning is used to implement the back propagation of the network to achieve the purpose of training and updating network parameters. The trained parameter set is used express.

5. The artificial intelligence multi-contrast magnetic resonance rapid imaging method as claimed in claim 1, characterized in that In step 5), the network model has been trained through step 4), that is, the network mapping function f(·) has been established; the collected multi-contrast undersampled image is input into the network, and the network finally outputs the reconstruction result of the corresponding contrast magnetic resonance image. in represents the multi-contrast magnetic resonance data to be reconstructed, Represents the reconstruction result corresponding to the c-th contrast magnetic resonance image.

Images