Multi-anatomical-structure multi-contrast magnetic resonance imaging method and system

Abstract

The invention provides a multi-anatomical-structure multi-contrast magnetic resonance imaging method and system. The method comprises the following steps: constructing an image reconstruction model combining sampling trajectory optimization and embedded decoupling representation; and defining a loss function based on decoupling representation learning as an optimization target. And constructing a multi-contrast multi-anatomical structure magnetic resonance data set, and jointly training a sampling trajectory optimization and image reconstruction model. And carrying out data sampling by utilizing the trained model and applying a sampling track generated by the sampling track generation network, and carrying out image reconstruction on undersampled data to obtain a reconstructed image without artifacts. By means of decoupling representation learning and sampling trajectory modeling of anatomical structure specificity, accelerated acquisition and reconstruction can be performed in a magnetic resonance data scene containing various anatomical structures and contrast ratios, and joint optimization of magnetic resonance sampling and reconstruction processes is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and more particularly, to a multi-anatomical structure multi-contrast magnetic resonance imaging method and system, and more particularly to a multi-anatomical structure multi-contrast magnetic resonance imaging method based on decoupled representation and sampling trajectory learning. Background Art

[0002] Magnetic resonance imaging can obtain images with good tissue contrast, and then be used to reflect tissue anatomical structures or calculate quantitative parameters of tissues, providing a basis for the diagnosis and treatment of related diseases. However, limited by the slow acquisition speed, the magnetic resonance imaging process is usually very time-consuming. Therefore, in clinical practice, it is often necessary to perform accelerated undersampling in the k-space to shorten the sampling time. Direct inverse Fourier transform of the undersampled and accelerated data will result in images with undersampling aliasing artifacts. Therefore, many image reconstruction methods for undersampled data have been proposed in the current field. Traditional parallel imaging and compressed sensing methods perform well on the acquired data with low-fold undersampling, but perform poorly in the case of high sampling multiples and low signal-to-noise ratios. Many deep learning-based image reconstruction methods proposed in recent years can obtain artifact-free reconstructed images from highly accelerated data by training on a large amount of magnetic resonance data. Existing deep learning methods usually assume that the training data set is independent and identically distributed, that is, it is assumed that the training data are independent samples from the same distribution. However, magnetic resonance imaging data in many application scenarios do not satisfy this assumption. For example, the contrast of magnetic resonance images is affected by imaging sequence parameters (such as repetition time TR, echo time TE, inversion time TI) and factors of magnetic resonance imaging equipment of a specific manufacturer, and the anatomical structures of the images are affected by the imaged organs and anatomical orientations. In addition, the final undersampling artifact pattern is affected by multiple factors such as the acquisition signal-to-noise ratio and the k-space sampling trajectory. In the case of domain shift, high diversity and heterogeneity in the training data, many existing deep learning image reconstruction methods cannot guarantee robust results, and there may be a decline in reconstruction performance for some samples and a deviation in reconstruction performance for a specific domain. Therefore, in order to promote deep learning magnetic resonance imaging methods to clinical applications, it is necessary to propose a general reconstruction model that can perform well on diverse magnetic resonance data.

[0003] To implement this general model, it is necessary to specifically address the impacts of image contrast and anatomical structures on data distribution. In the field of deep learning, decoupled representation learning is a strategy that can be used to reduce the differences between data domains and thus achieve domain adaptation. This strategy decouples data into style encodings that reflect domain-specific information and content encodings that are domain-invariant through a deep neural network, so that tasks can be completed with the help of domain-invariant content encodings. Decoupled representation learning has been widely applied to tasks such as image classification, segmentation, and style transfer, yet it is still less used for image reconstruction. Compared with the above tasks, magnetic resonance image reconstruction has its own particularities and complexities. On the one hand, different from segmentation, classification, etc. which only require the domain-invariant semantic information contained in the content encoding, magnetic resonance image reconstruction requires both precise content encoding to remove artifacts and restore the image anatomical structure, and style encoding integration to restore the image contrast information. However, existing medical image segmentation methods based on decoupled representation learning usually ignore extracting image texture information in the content encoding and are difficult to be directly transferred to image reconstruction. On the other hand, for the robustness and credibility of the magnetic resonance image reconstruction process, its reconstructed images need to satisfy the fidelity with respect to k-space sampling data. Therefore, the reconstruction method needs to incorporate the modeling of the imaging physical process. However, existing decoupled representation learning methods only regard image reconstruction as a pure data-driven image restoration task and fail to combine this part of information, making it difficult to meet practical requirements. Therefore, it is necessary to propose a reconstruction method that can not only solve cross-domain image reconstruction but also combine the physical imaging model.

[0004] The impact of anatomical structures on the magnetic resonance data distribution is also indirectly reflected in the k-space data. Even under the same undersampling pattern, the sampling data and the distribution of undersampling artifacts generated by anatomical structures from different imaging orientations and organs are different. A series of works on magnetic resonance acquisition optimization hope to find the optimal k-space sampling trajectory for the entire dataset through deep learning, yet they cannot adapt to the data distribution differences brought by different anatomical structures. Inspired by these existing works, we assume that there is an optimal sampling pattern for each type of anatomical structure. Based on this, it is necessary to propose a sampling trajectory optimization scheme for specific anatomical structures. Summary of the Invention

[0005] Aiming at the defects in the prior art, the purpose of the present invention is to provide a multi-anatomical structure and multi-contrast magnetic resonance imaging method and system.

[0006] According to a multi-anatomical structure and multi-contrast magnetic resonance imaging method provided by the present invention, the method includes the following steps:

[0007] Step S1: Construct an image reconstruction model that combines sampling trajectory optimization and embedded decoupled representation;

[0008] Step S2: Define the loss function based on decoupled representation learning as the optimization objective;

[0009] Step S3: Construct a multi-contrast multi-anatomical structure magnetic resonance dataset and jointly train the sampling trajectory optimization and image reconstruction model;

[0010] Step S4: Use the trained model to perform data sampling with the sampling trajectory generated by the sampling trajectory generation network, and perform image reconstruction on the undersampled data to obtain an artifact-free reconstructed image.

[0011] Preferably, the image reconstruction model in step S1 includes a sampling trajectory generation network, a coil sensitivity map estimation network, a data fidelity module, a style encoding network, a content encoding network, a decoding network, and a phase artifact removal network;

[0012] Step S1 includes the following steps:

[0013] Step S1.1: The sampling trajectory generation network outputs an optimized sampling trajectory according to the anatomical structure category label corresponding to the input magnetic resonance imaging sample;

[0014] Step S1.2: The coil sensitivity map estimation network inputs the undersampled k-space data of the multi-coil channels and outputs the coil sensitivity map;

[0015] Step S1.3: The data fidelity module inputs the current reconstructed image, the sampling trajectory, the true sampling k-space, and the coil sensitivity map;

[0016] Step S1.4: The style encoding network inputs the amplitude of the current reconstructed image and the corresponding contrast category, and outputs the encoded style representation;

[0017] Step S1.5: The content encoding network inputs the amplitude of the current reconstructed image and outputs the encoded content representation;

[0018] Step S1.6: The decoding network inputs the style encoding and the content encoding and outputs the amplitude of the reconstructed image;

[0019] Step S1.7: The phase artifact removal network inputs the phase of the current reconstructed image and outputs the phase image after artifact removal.

[0020] Preferably, the loss function based on decoupled representation learning in step S2 includes a style loss, a content loss, an auxiliary classification loss, and a reconstruction loss. The style loss L style is used to measure the similarity in style between different reconstructed images of the same contrast type. The content loss L content is used to measure the similarity in content between the reconstructed image and the true image. The auxiliary classification loss L aux is used to promote the correlation between the style encoding and the contrast category. The reconstruction loss Lrecon To promote the similarity between the reconstructed image and the full acquisition image:

[0021]

[0022] L aux (s N , i) = CE(MLP(s N ), i)

[0023]

[0024] where s is the style encoding, the superscripts p and q represent the indices of different samples, cos represents the cosine similarity, c is the content encoding, l1 represents the L1 norm, the subscript N represents the Nth iteration, i represents the contrast category of the image, CE is the cross-entropy loss, and MLP is a classifier based on a multi-layer perceptron. is the reconstructed image, x GT is the real image, and error is an arbitrary difference metric.

[0025] Preferably, the step of constructing the multi-contrast multi-anatomical structure magnetic resonance dataset in step S3 includes the following steps:

[0026] Step S3.1: Collect magnetic resonance imaging data containing multiple anatomical structures and contrasts;

[0027] Step S3.2: Preprocess the collected data, including normalization and denoising;

[0028] Step S3.3: Divide the preprocessed data into a training set, a validation set, and a test set.

[0029] Preferably, the process of data sampling and image reconstruction in step S4 includes the following steps:

[0030] Step S4.1: Save the trained deep learning model;

[0031] Step S4.2: During magnetic resonance imaging, input the anatomical structure category label of the imaging task into the trained sampling trajectory generation network to output an optimized sampling trajectory;

[0032] Step S4.3: Apply the sampling trajectory to sample k-space data;

[0033] Step S4.4: Input the sampled data into a reconstruction model composed of a trained coil sensitivity map estimation network, a data fidelity module, a style encoding network, a content encoding network, a decoding network, and a phase artifact removal network to obtain an artifact-removed reconstructed image.

[0034] The present invention also provides a multi - anatomical - structure and multi - contrast magnetic resonance imaging system, which includes the following modules:

[0035] Module M1: Construct an image reconstruction model that combines sampling trajectory optimization and embedded decoupled representation;

[0036] Module M2: Define a loss function based on decoupled representation learning as the optimization objective;

[0037] Module M3: Construct a multi - contrast and multi - anatomical - structure magnetic resonance data set, and jointly train the sampling trajectory optimization and image reconstruction model;

[0038] Module M4: Use the trained model, apply the sampling trajectory generated by the sampling trajectory generation network for data sampling, and perform image reconstruction on the undersampled data to obtain an artifact - removed reconstructed image.

[0039] Preferably, the image reconstruction model in the module M1 includes a sampling trajectory generation network, a coil sensitivity map estimation network, a data fidelity module, a style encoding network, a content encoding network, a decoding network, and a phase artifact - removal network;

[0040] The module M1 includes the following sub - modules:

[0041] Module M1.1: The sampling trajectory generation network outputs an optimized sampling trajectory according to the anatomical structure category label corresponding to the input magnetic resonance imaging sample;

[0042] Module M1.2: The coil sensitivity map estimation network inputs the undersampled k - space data of the multi - coil channels and outputs the coil sensitivity map;

[0043] Module M1.3: The data fidelity module inputs the current reconstructed image, the sampling trajectory, the true sampling k - space, and the coil sensitivity map;

[0044] Module M1.4: The style encoding network inputs the amplitude of the current reconstructed image and the corresponding contrast category, and outputs the encoded style representation;

[0045] Module M1.5: The content encoding network inputs the amplitude of the current reconstructed image and outputs the encoded content representation;

[0046] Module M1.6: The decoding network inputs the style encoding and the content encoding and outputs the amplitude of the reconstructed image;

[0047] Module M1.7: The phase artifact - removal network inputs the phase of the current reconstructed image and outputs the phase image after artifact removal.

[0048] Preferably, the loss function based on decoupled representation learning in the module M2 includes a style loss, a content loss, an auxiliary classification loss, and a reconstruction loss. The style loss Lstyle Used to measure the similarity in style between reconstructed images with different reconstruction images of the same type of contrast, content loss L content Used to measure the similarity in content between the reconstructed image and the real image, auxiliary classification loss L aux Used to promote the correlation between the style encoding and the contrast category, reconstruction loss L recon Used to promote the similarity between the reconstructed image and the fully sampled image:

[0049]

[0050] L aux (s N ,i) = CE(MLP(s N ),i)

[0051]

[0052] where s is the style encoding, the superscripts p and q represent the indices of different samples, cos represents the cosine similarity, c is the content encoding, l1 represents the L1 norm, the subscript N represents the Nth iteration, i represents the contrast category of the image, CE is the cross-entropy loss, and MLP is a classifier based on a multi-layer perceptron, is the reconstructed image, x GT is the real image, and error is an arbitrary difference metric.

[0053] Preferably, the construction of the multi-contrast multi-anatomical structure magnetic resonance dataset in the module M3 includes the following modules:

[0054] Module M3.1: Collect magnetic resonance imaging data containing multiple anatomical structures and contrasts;

[0055] Module M3.2: Preprocess the collected data, including normalization and denoising;

[0056] Module M3.3: Divide the preprocessed data into a training set, a validation set, and a test set.

[0057] Preferably, the process of data sampling and image reconstruction in the module M4 includes the following modules:

[0058] Module M4.1: Save the trained deep learning model;

[0059] Module M4.2: During magnetic resonance imaging, input the anatomical structure category label of the imaging task into the trained sampling trajectory generation network to output an optimized sampling trajectory;

[0060] Module M4.3: Apply the sampling trajectory to sample k-space data;

[0061] Module M4.4: Input the sampled data into a reconstruction model composed of a trained coil sensitivity map estimation network, a data fidelity module, a style encoding network, a content encoding network, a decoding network, and a phase artifact removal network to obtain an artifact-removed reconstructed image.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] 1. The present invention integrates decoupled representation learning in an iterative unfolding network model. By decoupling an image into style and content encodings, it fully utilizes the cross-domain invariant information of multi-contrast magnetic resonance data, and explicitly introduces a mechanism for dealing with the problem of large domain differences in data compared with conventional reconstruction methods.

[0064] 2. Compared with existing methods that combine decoupled representation and image reconstruction, the present invention combines an imaging physical model, satisfies the fidelity constraint of k-space sampled data, and has better reliability.

[0065] 3. The present invention models the sampling trajectory as a sampling trajectory specific to different anatomical structures, and can be optimized separately for different anatomical structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] By reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings, other features, objects, and advantages of the present invention will become more apparent:

[0067] Figure 1 is a flowchart of an embodiment of the method for joint sampling trajectory optimization and image reconstruction of magnetic resonance imaging based on decoupled representation learning of the present invention;

[0068] Figure 2 is a schematic diagram of the algorithm principle of the method for joint sampling trajectory optimization and image reconstruction of magnetic resonance imaging based on decoupled representation learning of the present invention;

[0069] Figure 3 is a schematic diagram of the results of the method for joint sampling trajectory optimization and image reconstruction of cardiac magnetic resonance imaging based on decoupled representation learning of an embodiment of the present invention;

[0070] Figure 4 is a schematic diagram of the results of the method for joint sampling trajectory optimization and image reconstruction of multi-contrast magnetic resonance imaging of the brain based on decoupled representation learning of an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0071] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any way. It should be noted that those of ordinary skill in the art can make several changes and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0072] Example 1:

[0073] Refer to Figure 1 and Figure 2 , a multi - anatomical - structure and multi - contrast magnetic resonance imaging method provided by the present invention, the method comprises the following steps:

[0074] Step S1: Construct an image reconstruction model that combines sampling trajectory optimization and decoupled representation embedding; the image reconstruction model includes a sampling trajectory generation network, a coil sensitivity map estimation network, a data fidelity module, a style encoding network, a content encoding network, a decoding network, and a phase artifact removal network;

[0075] Step S1.1: The sampling trajectory generation network outputs an optimized sampling trajectory according to the anatomical structure category label corresponding to the input magnetic resonance imaging sample;

[0076] Step S1.2: The coil sensitivity map estimation network inputs the undersampled k - space data of the multi - coil channels and outputs the coil sensitivity map;

[0077] Step S1.3: The data fidelity module inputs the current reconstructed image, the sampling trajectory, the true sampling k - space, and the coil sensitivity map;

[0078] Step S1.4: The style encoding network inputs the amplitude of the current reconstructed image and the corresponding contrast category, and outputs the encoded style representation;

[0079] Step S1.5: The content encoding network inputs the amplitude of the current reconstructed image and outputs the encoded content representation;

[0080] Step S1.6: The decoding network inputs the style encoding and the content encoding and outputs the amplitude of the reconstructed image;

[0081] Step S1.7: The phase artifact removal network inputs the phase of the current reconstructed image and outputs the phase image after artifact removal.

[0082] Step S2: Define a loss function based on decoupled representation learning as the optimization objective; the loss function based on decoupled representation learning includes a style loss, a content loss, an auxiliary classification loss, and a reconstruction loss. The style loss L style is used to measure the similarity in style between different reconstructed images of the same contrast type. The content loss L content is used to measure the similarity in content between the reconstructed image and the true image. The auxiliary classification loss L aux is used to promote the correlation between the style encoding and the contrast category. The reconstruction loss L recon is used to promote the similarity between the reconstructed image and the fully sampled image:

[0083]

[0084] L aux (s N , i) = CE(MLP(s N ), i)

[0085]

[0086] where s is the style encoding, the superscripts p and q represent the indices of different samples, cos represents cosine similarity, c is the content encoding, l1 represents the L1 norm, the subscript N represents the Nth iteration, i represents the contrast class of the image, CE is the cross-entropy loss, and MLP is a classifier based on a multi-layer perceptron. is the reconstructed image, x GT is the ground truth image, and error is an arbitrary difference metric.

[0087] Step S3: Construct a multi-contrast multi-anatomical structure magnetic resonance dataset and jointly train the sampling trajectory optimization and image reconstruction models; constructing a multi-contrast multi-anatomical structure magnetic resonance dataset includes the following steps:

[0088] Step S3.1: Collect magnetic resonance imaging data containing multiple anatomical structures and contrasts;

[0089] Step S3.2: Preprocess the collected data, including normalization and denoising;

[0090] Step S3.3: Divide the preprocessed data into a training set, a validation set, and a test set.

[0091] Step S4: Use the trained model to perform data sampling using the sampling trajectory generated by the sampling trajectory generation network and perform image reconstruction on the undersampled data to obtain an artifact-free reconstructed image.

[0092] The process of performing data sampling and image reconstruction includes the following steps:

[0093] Step S4.1: Save the trained deep learning model;

[0094] Step S4.2: During magnetic resonance imaging, input the anatomical structure class label of the imaging task into the trained sampling trajectory generation network to output an optimized sampling trajectory;

[0095] Step S4.3: Use the sampling trajectory to sample k-space data;

[0096] Step S4.4: Input the sampled data into the reconstruction model composed of a trained coil sensitivity map estimation network, a data fidelity module, a style encoding network, a content encoding network, a decoding network, and a phase artifact removal network to obtain an artifact-free reconstructed image.

[0097] The present invention also provides a multi-anatomical-structure multi-contrast magnetic resonance imaging system, which can be implemented by executing the process steps of the multi-anatomical-structure multi-contrast magnetic resonance imaging method. That is, those skilled in the art can understand the multi-anatomical-structure multi-contrast magnetic resonance imaging method as a preferred embodiment of the multi-anatomical-structure multi-contrast magnetic resonance imaging system.

[0098] Example 2:

[0099] The present invention also provides a multi-anatomical-structure multi-contrast magnetic resonance imaging system, which includes the following modules:

[0100] Module M1: Construct an image reconstruction model that combines sampling trajectory optimization and embedded decoupled representation; the image reconstruction model includes a sampling trajectory generation network, a coil sensitivity map estimation network, a data fidelity module, a style encoding network, a content encoding network, a decoding network, and a phase artifact removal network;

[0101] Module M1.1: The sampling trajectory generation network outputs an optimized sampling trajectory according to the anatomical structure category label corresponding to the input magnetic resonance imaging sample.

[0102] Module M1.2: The coil sensitivity map estimation network inputs the undersampled k-space data of the multi-coil channels and outputs the coil sensitivity map.

[0103] Module M1.3: The data fidelity module inputs the current reconstructed image, the sampling trajectory, the true sampling k-space, and the coil sensitivity map.

[0104] Module M1.4: The style encoding network inputs the amplitude of the current reconstructed image and the corresponding contrast category, and outputs the encoded style representation.

[0105] Module M1.5: The content encoding network inputs the amplitude of the current reconstructed image and outputs the encoded content representation.

[0106] Module M1.6: The decoding network inputs the style encoding and the content encoding and outputs the amplitude of the reconstructed image.

[0107] Module M1.7: The phase artifact removal network inputs the phase of the current reconstructed image and outputs the phase image after artifact removal.

[0108] Module M2: Define a loss function based on decoupled representation learning as the optimization objective; the loss function based on decoupled representation learning includes a style loss, a content loss, an auxiliary classification loss, and a reconstruction loss. The style loss L style is used to measure the similarity in style between different reconstructed images of the same contrast category. The content loss L contentUsed to measure the similarity in content between the reconstructed image and the ground truth image, assisting the classification loss L aux Used to promote the correlation between the style encoding and the contrast category, the reconstruction loss L recon Used to promote the similarity between the reconstructed image and the fully sampled image:

[0109]

[0110] L aux (s N , i) = CE(MLP(s N ), i)

[0111]

[0112] where s is the style encoding, the superscripts p, q represent the indices of different samples, cos represents the cosine similarity, c is the content encoding, l1 represents the L1 norm, the subscript N represents the Nth iteration, i represents the contrast category of the image, CE is the cross-entropy loss, and MLP is a classifier based on a multi-layer perceptron is the reconstructed image, x GT is the ground truth image, and error is an arbitrary difference metric

[0113] Module M3: Construct a multi-contrast multi-anatomical structure magnetic resonance dataset and jointly train the sampling trajectory optimization and image reconstruction models; constructing a multi-contrast multi-anatomical structure magnetic resonance dataset includes the following modules:

[0114] Module M3.1: Collect magnetic resonance imaging data containing multiple anatomical structures and contrasts;

[0115] Module M3.2: Preprocess the collected data, including normalization and denoising;

[0116] Module M3.3: Divide the preprocessed data into a training set, a validation set, and a test set

[0117] Module M4: Use the trained model to perform data sampling using the sampling trajectory generated by the sampling trajectory generation network and perform image reconstruction on the undersampled data to obtain an artifact-free reconstructed image

[0118] The process of performing data sampling and image reconstruction includes the following modules:

[0119] Module M4.1: Save the trained deep learning model;

[0120] Module M4.2: During magnetic resonance imaging, input the anatomical structure category label of the imaging task into the trained sampling trajectory generation network to output an optimized sampling trajectory;

[0121] Module M4.3: Sampling k-space data using the application sampling trajectory.

[0122] Module M4.4: Input the sampled data into a reconstruction model composed of a trained coil sensitivity map estimation network, a data fidelity module, a style encoding network, a content encoding network, a decoding network, and a phase artifact removal network to obtain an artifact-removed reconstructed image.

[0123] Example 3:

[0124] The following will introduce in detail an embodiment of the present invention applied to cardiac magnetic resonance quantitative imaging in conjunction with the accompanying drawings. Figure 1 It is a flowchart of the embodiment of the present invention. Figure 2 It is a schematic diagram of the algorithm of the present invention. Specifically, this embodiment includes the following steps:

[0125] Step S1: A joint sampling trajectory optimization and embedded decoupled representation image reconstruction model, which model includes the following modules:

[0126] 1) Sampling trajectory generation network: The input is the anatomical structure category label corresponding to the magnetic resonance imaging sample, and the output is the generated sampling trajectory through a neural network. First, perform one-hot encoding on the structure category label to obtain a vector reflecting the anatomical structure category, and then input it into the neural network. The network architecture is a multi-layer perceptron with 3 hidden layers, the number of hidden neurons being 64, 128, and 256 respectively, the activation functions of the first 2 layers being ReLU, and the activation function of the last layer being softmax. Output a vector of length 256 (the number of phase encoding lines during full sampling). Set 1 for 16 elements in the central region of the vector (corresponding to the 16 central full-sampling phase encoding lines for calculating the coil sensitivity), and then input it into the sigmoid function to obtain a sampling trajectory probability vector normalized to the range of 0-1. Preset the number of sampling lines k according to the total number of phase encoding lines divided by the acceleration factor (in this example, 4-fold and 8-fold accelerations are used respectively to construct 2 models), and take the k phase encoding lines corresponding to the k elements with the largest probability as the sampling trajectory.

[0127] 2) Coil Sensitivity Map Estimation Module: This module is used to generate a coil sensitivity map from the k-space data obtained by multi-coil parallel imaging. Its input is the multi-coil parallel imaging k-space, and the output is the coil sensitivity map. Specifically, this module first extracts the region of 16 fully sampled phase-encoding lines at the center of the multi-coil k-space through a mask, applies the inverse Fourier transform to transform it into the image domain, and inputs it into a U-Net network. In each block of the encoder part of this U-Net network, 2 convolutional-instance normalization-LeakyReLU activation functions are sequentially executed, followed by an average pooling. The number of channels in the convolutional layer of the first block is 8, and the number of channels doubles after each pooling downsampling. A total of 4 downsamplings are performed. The decoder part sequentially executes transposed convolution with a kernel size of 2 and a stride of 2 and LeakyReLU activation. After 4 transposed convolution upsamplings and skip connections, the output is obtained.

[0128] 3) Data Fidelity Module: This module inputs the current reconstructed image, sampling trajectory, true sampled k-space, and coil sensitivity map. It converts it to the k-space by multiplying the current image and the coil sensitivity map element by element, performs data fidelity processing, and then converts it to the image domain through channel merging, and outputs the image after data fidelity constraint. Among them, for data fidelity processing, first subtract the current reconstructed k-space data from the true undersampled k-space data, then apply the mask of the sampling trajectory to the result of the subtraction, multiply it by the iteration step size of this iteration (set as a learnable variable), and add the result to the current reconstructed k-space before processing to obtain the k-space data.

[0129] 4) Style Encoding Network: This network inputs the amplitude and contrast category of the current reconstructed image and outputs the encoded style representation. The network architecture is a convolutional neural network based on conditional convolution. Conditional convolution accepts the contrast category label, making the parameters of the finally applied convolutional kernel depend on the contrast. This convolutional neural network uniformly samples the input image to a fixed size. After sequentially passing through 5 conditional convolutional layers with 16, 32, 64, 128, and 128 output channels and LeakyReLU activation functions, the obtained features pass through an MLP with 32 and 16 output neurons respectively to obtain a style encoding of length 16.

[0130] 5) Content Encoding Network: This network inputs the amplitude of the current reconstructed image and outputs the encoded content representation. Its network architecture is a U-Net network, which has the same structure as the coil sensitivity estimation network, and the number of output channels in the last layer is changed to 4, so as to obtain a feature map with the same width and height as the image and 4 channels as the content encoding.

[0131] 6) Decoding network: This network inputs the style and content codes obtained by the above encoding and outputs the reconstructed image. One branch of the architecture adopts a U-Net variant, inputs the style code into the encoder of the U-Net network to obtain the intermediate features, and the other branch inputs the content code into a 4-layer multilayer perceptron with 128 hidden neurons in each layer to obtain the mapped features, which are repeated in the width and height dimensions and then spliced ​​with the intermediate features of the same size in the channel dimension. The spliced ​​features are then input into the U-Net decoder to obtain the output image. The encoder-decoder architecture is the same as the coil sensitivity estimation network.

[0132] 7) Phase artifact removal network: This network inputs the phase of the current reconstructed image and outputs the phase after artifact removal. Its network architecture is the same as the coil sensitivity estimation network.

[0133] Based on the above modules, the computational process of the joint sampling trajectory optimization and embedded decoupled representation image reconstruction model includes the following steps:

[0134] First, the category label of the anatomical structure information of the image to be imaged (multiple categories are preset, such as coronal, sagittal, transverse, cardiac short axis, etc.) is input into the sampling trajectory generation network to generate a sampling trajectory, which is used to undersample the fully sampled k-space data to obtain undersampled k-space data.

[0135] Next, the undersampled k-space data is input into the coil sensitivity map estimation module to obtain the coil sensitivity map.

[0136] Then, the following iterative process is repeated for the image: (1) the image and coil sensitivity map are input into the data fidelity module, and the image after data fidelity constraint is output; (2) the obtained image is decomposed into phase and amplitude, the image phase is input into the image phase de-artifacting network, the image amplitude and data contrast category are input into the style encoding network, and the image amplitude is also input into the content encoding network. The style encoding network and the content encoding network output style and content coding respectively; (3) the obtained style coding and content coding are input into the decoding network to obtain the output reconstructed amplitude image. The reconstructed amplitude image is recombined with the reconstructed phase image output by the phase de-artifacting network to obtain the current image with de-artifacting, and the next iteration is repeated.

[0137] After repeating 12 iterations, the image obtained from the last iteration is converted to the k-space domain using the coil sensitivity map, and then converted to the image domain by inverse Fourier transform. The reconstructed amplitude image is obtained by merging in the channel dimension through square sum operation. The style code output by the style encoding network in the last iteration is then input into the auxiliary classifier, and the classification result of the corresponding contrast category is output. The style code, content code, and final reconstructed image and classification result generated in each iteration during the calculation process will be used to calculate the training loss function value in step S2.

[0138] Step S2: Define the loss function based on decoupled representation learning as the optimization objective, which includes the following parts:

[0139] (1) Reconstruction loss (Lrecon): Calculate the difference metric between the reconstructed image output by the model and the full-gold-standard image x GT as the reconstruction loss. The difference metric uses 10 times the per-pixel absolute error and 1 time the structural similarity error, that is, 1 minus the structural similarity metric (SSIM):

[0140]

[0141] (2) Style loss (Lstyle): During training, between two samples p and q in a batch, calculate the similarity between the style encodings s obtained in each iteration: If the two samples have the same contrast, maximize the similarity between their style encodings; if the two samples have different contrasts, minimize the similarity between their style encodings. Here, the style encoding similarity can be calculated by cosine similarity. To minimize the similarity, minimize the absolute value of the cosine value between the two vectors to make them orthogonal; to maximize the similarity, minimize 1 minus the cosine value between the two vectors to make them in the same direction. Sum the results calculated for each iteration to obtain the final style loss:

[0142]

[0143] (3) Auxiliary classification loss (Laux): To further promote the style encoder to extract contrast-related losses, this loss designs an auxiliary classification task to calculate the accuracy of contrast classification using the style encoding. During training, introduce a classification module based on a multi-layer perceptron (MLP) to train synchronously with the network. Its input is the style encoding obtained in the last iteration, which contains 6 hidden layers (number of neurons: 256, 128, 64, 32, 16), and the activation functions are ReLU (the first 5 layers) and sigmoid (the last layer). The output predicts the contrast class i to which the image belongs, and calculate the multi-class cross-entropy (CE) loss between the predicted class and the true contrast class, which is the auxiliary classification loss:

[0144] L aux (s N ,i) = CE(MLP(s N ),i)

[0145] (4) Content loss (Lcont): During training, between two samples p and q in a batch, calculate the similarity between the corresponding content encodings c obtained in each iteration: If the two images are different-contrast images from the same layer scan, maximize the similarity between their content encodings; conversely, if the content of the two images is not paired, the value of the content loss is 0. The similarity is measured using the absolute error l1:

[0146]

[0147] Weighted sum the above four losses through weight coefficients to form a complete loss function. Among them, the weight coefficients are hyperparameters.

[0148] Step S3.1: To construct a multi-contrast multi-anatomical structure magnetic resonance dataset, it is necessary to preprocess the multi-contrast multi-anatomical structure cardiac magnetic resonance k-space data containing short-axis T1 mapping, T2 mapping, sagittal / transverse aortic imaging, and cardiac cine imaging of short-axis / two-chamber / three-chamber / four-chamber / aortic outflow tract, including: (1) Normalize it according to its maximum value range in the image domain so that the amplitude distribution of the entire data in the image domain is within the range of [0,1]; (2) For data samples with too many coil channels, use a channel compression algorithm to compress them to 10 virtual coil channels; (3) According to different sampling parameters, assign different contrast categories to the data. The contrast of aortic imaging and cine imaging data is relatively consistent, and they are used as separate categories. Since the T1 mapping data acquired by the MOLLI sequence has different inversion times (TI), its sampled data is divided into 3 groups of contrasts according to the inversion time; the T2 mapping data acquired by the FLASH sequence has different T2 preparation times (0 / 35 / 55 milliseconds), so it is divided into 3 groups of contrasts according to this parameter. In this way, a data containing 8 different contrast categories is constructed; (4) According to the anatomical structure category of each data, label it as 8 different anatomical structure categories including sagittal, transverse, short-axis, two-chamber, three-chamber, four-chamber, and aortic outflow tract.

[0149] Step S3.2: Using the preprocessed fully sampled k-space data to jointly train the model: Input the anatomical structure category label into the sampling trajectory generation network to obtain the sampling trajectory. Apply the sampling trajectory to obtain the undersampled k-space. Input the undersampled k-space into the reconstruction model. Calculate the loss function according to the definition method in Step S2 based on the output of relevant modules. Update the learnable parameters of the joint sampling trajectory optimization and decoupled representation image reconstruction model by minimizing the loss function until the reconstruction loss on the validation set converges, and obtain the trained deep learning model. In this embodiment, the Adam optimizer is used, with an initial learning rate of 0.0002, which decays to 0.5 times the original every 10 training rounds, and a total of 60 training rounds are performed.

[0150] Step S4: The process of data sampling and image reconstruction includes: Saving the trained deep learning model. Taking short-axis imaging as an example, input the anatomical structure category label corresponding to the short axis into the trained sampling trajectory generation network to output the optimized sampling trajectory; Apply the sampling trajectory to sample the k-space data, and input the sampled data into the reconstruction model composed of the trained coil sensitivity map estimation network, data fidelity module, style encoding network, content encoding network, decoding network, and phase artifact removal network to obtain the artifact-removed reconstructed image. For T2 mapping, according to different T2 preparation times during acquisition, the final T2 quantitative map can be fitted from the reconstructed image.

[0151] Figure 3 It is a schematic diagram of the results of this embodiment. Among them, the first three rows respectively correspond to the images with T2 preparation times of 0 / 35 / 55 milliseconds for T2 mapping, and the fourth row corresponds to the T2 quantitative map fitted from the multi-contrast images; The first column is the fully sampled image reconstruction and quantitative results, and the second to seventh columns are the images reconstructed using different control methods under 4-fold acceleration and the quantitative maps fitted; The eighth column is the image reconstructed using the method of the present invention under 4-fold acceleration and the quantitative map, the ninth to fourteenth columns are the images reconstructed by the control method under 8-fold acceleration and the quantitative maps, and the fifteenth column is the image reconstructed by the method of the present invention under 8-fold acceleration and the quantitative map. After reconstruction by the control method, most of the images do not have serious undersampling artifacts. However, residual artifacts and blurring appear in the reconstructed images of individual contrasts, resulting in artifacts and blurring in the T2 quantitative map. In contrast, after reconstruction using the method of the present invention, the image quality index of the reconstructed image with respect to the fully sampled gold standard has been improved. On this basis, the image reconstructed by the method of the present invention is used to fit the T2 quantitative map, and the myocardial region in the figure is complete, the boundary is clear, there are no obvious undersampling artifacts, and the overall quality is improved.

[0152] Example 4:

[0153] As Figure 1, the present invention discloses a method for jointly optimizing sampling trajectories and image reconstruction based on decoupled representation learning, and the method comprises the following steps:

[0154] Step S1: Construct an image reconstruction model for jointly optimizing sampling trajectories and embedding decoupled representations, and the model comprises the following modules:

[0155] 1) Sampling trajectory network: The input of this network is the anatomical structure category label corresponding to the magnetic resonance imaging sample, and the output is the generated sampling trajectory. First, one-hot encode the structure category label to obtain a vector reflecting the anatomical structure category, and then input it into the network. The network architecture is a multi-layer perceptron with 3 layers, the number of hidden neurons being 64, 128, and 256 respectively, and the activation function being ReLU except for the last layer. Output a vector of length 256 (the number of phase-encoding lines during full sampling), add a relatively large value to 16 elements in the central region of the vector (corresponding to the 16 central full-sampling phase-encoding lines for calculating coil sensitivity), and then input it into the sigmoid function to obtain a sampling trajectory probability distribution vector normalized to the range of 0-1. According to the preset number of sampling lines k, take the k phase-encoding lines corresponding to the k elements with the largest probability as the sampling trajectory.

[0156] 2) Coil sensitivity map estimation module: This module is used to generate a coil sensitivity map from the k-space data obtained by multi-coil parallel imaging. Its input is the multi-coil parallel imaging k-space, and the output is the coil sensitivity map. Specifically, this module first extracts the region of the 16 central full-sampling phase-encoding lines of the multi-coil k-space through a mask, applies inverse Fourier transform to transform it into the image domain, and inputs it into a U-Net network. In each block of the encoder part of this U-Net network, 2 consecutive convolutions - instance normalization - LeakyReLU activation functions and one average pooling are performed in sequence. The number of channels of the convolutional layer in the first block is 8, and the number of channels doubles after each pooling downsampling. A total of 4 times of downsampling are performed. The decoder part performs transposed convolution with a kernel size of 2 and a stride of 2 and LeakyReLU activation in sequence, and obtains the output through 4 times of transposed convolution upsampling and skip connections.

[0157] 3) Data fidelity module: The input of this module is the current reconstructed image, sampling trajectory, true sampled k-space, and coil sensitivity map. It multiplies the current image element-wise with the coil sensitivity map to transform it into the k-space, performs data fidelity processing, and then transforms it back to the image domain through channel merging, and outputs the image after data fidelity constraint. Among them, for data fidelity processing, first subtract the current reconstructed k-space data from the true undersampled k-space data, then apply the mask of the sampling trajectory to the result of the subtraction, multiply it by the iteration step size of this iteration (set as a learnable variable), and add the result to the current reconstructed k-space before processing to obtain the k-space data.

[0158] 4) Style Encoding Network: This network takes the amplitude and contrast category of the current reconstructed image as input and outputs the encoded style representation. Its network architecture is a convolutional neural network based on conditional convolution. Conditional convolution accepts the contrast category label, and the final applied convolutional kernel parameters are different for different contrasts. This convolutional neural network uniformly samples the input image to a fixed size. After passing through 5 conditional convolutional layers with 16, 32, 64, 128, and 128 output channels and the LeakyReLU activation function in sequence, the obtained features pass through an MLP with 32 and 16 output neurons respectively to obtain a style encoding of length 16.

[0159] 5) Content Encoding Network: This network takes the amplitude of the current reconstructed image as input and outputs the encoded content representation. Its network architecture is a U-Net network, which has the same structure as the coil sensitivity estimation network, and the number of output channels of the last layer is changed to 4, so as to obtain a feature map with the same width and height as the image and 4 channels as the content encoding.

[0160] 6) Decoding Network: This network takes the style and content encodings obtained above as input and outputs the reconstructed image. The architecture of one branch adopts a U-Net variant. The style encoding is input into the encoder of the U-Net network to obtain intermediate features. The other branch inputs the content encoding into a multi-layer perceptron with 4 layers and 128 hidden neurons in each layer to obtain the mapped features. After repeating them in the width and height dimensions, they are concatenated with the intermediate features of the same size in the channel dimension. The concatenated features are then input into the decoder of the U-Net to obtain the output image. The architecture of the encoder-decoder is the same as that of the coil sensitivity estimation network.

[0161] 7) Phase Artifact Removal Network: This network takes the phase of the current reconstructed image as input and outputs the phase after artifact removal. Its network architecture is the same as that of the coil sensitivity estimation network.

[0162] Based on the above modules, the calculation process of the joint sampling trajectory optimization and embedded decoupled representation image reconstruction model includes the following steps:

[0163] First, input the anatomical structure category to which the imaging data belongs into the sampling trajectory generation network, output the sampling trajectory, and apply this trajectory to sample the fully sampled k-space training data to obtain the undersampled k-space data.

[0164] Next, input the undersampled k-space data into the coil sensitivity map estimation module to obtain the coil sensitivity map, and use the coil sensitivity map to merge the multi-channel k-space data to obtain an image.

[0165] Then, the following iterative process is repeated for the image: (1) the image and coil sensitivity map are input into the data fidelity module, and the image after data fidelity constraint is output; (2) the obtained image is decomposed into phase and amplitude, the image phase is input into the image phase de-artifacting network, the image amplitude and data contrast category are input into the style encoding network, and the image amplitude is also input into the content encoding network. The style encoding network and the content encoding network output style and content coding respectively; (3) the obtained style coding and content coding are input into the decoding network to obtain the output reconstructed amplitude image. The reconstructed amplitude image is recombined with the reconstructed phase image output by the phase de-artifacting network to obtain the current image with de-artifacting, and the next iteration is repeated.

[0166] After repeating 12 iterations, the image obtained from the last iteration is converted to the k-space domain using the coil sensitivity map, and then converted to the image domain by inverse Fourier transform. The reconstructed amplitude image is obtained by merging in the channel dimension through square sum operation. The style code output by the style encoding network in the last iteration is then input into the auxiliary classifier, and the classification result of the corresponding contrast category is output. The style code, content code, and final reconstructed image and classification result generated in each iteration during the calculation process will be used to calculate the training loss function value in step S2.

[0167] Step S2: To jointly train the sampling trajectory and image optimization model, it is necessary to predefine the training loss function, which mainly includes the following parts:

[0168] (1) Reconstruction loss (Lrecon): The difference measure between the reconstructed image output by the calculation model and the fully mined gold standard image is used as the reconstruction loss. The difference measure uses 10 times the absolute value error of each pixel and 1 times the structural similarity error (1 minus the structural similarity measure).

[0169] (2) Style loss (Lstyle): During training, the similarity between the style codes obtained in each iteration is calculated between two samples in a batch: if the two samples have the same contrast, the similarity between their style codes is maximized; if the two samples have different contrasts, the similarity between their style codes is minimized. Here, the style code similarity can be calculated using cosine similarity. To minimize the similarity, the absolute value of the cosine value between the two vectors is minimized to make them orthogonal; to maximize the similarity, 1 minus the cosine value between the two vectors is minimized to make them in the same direction. The final style loss is obtained by summing the results of each iteration.

[0170] (3) Auxiliary classification loss (Laux): To further promote the style encoder to extract contrast-related losses, this loss designs an auxiliary classification task to calculate the accuracy of contrast classification using style encoding. During training, a classification module based on a multi-layer perceptron is introduced and trained synchronously with the network. Its input is the style encoding obtained from the last iteration, which consists of 6 hidden layers (number of neurons: 256, 128, 64, 32, 16), with ReLU (for the first 5 layers) and sigmoid (for the last layer) as activation functions. It outputs the predicted contrast category to which the image belongs, and calculates the multi-class cross-entropy loss between the predicted category and the true contrast category, which is the auxiliary classification loss.

[0171] (4) Content loss (Lcont): During training, between two samples in a batch, calculate the similarity between the content encodings obtained in each iteration: If the two images are different-contrast images from the same layer scan, maximize the similarity between their content encodings; otherwise, if the content of the two images is not paired, the value of the content loss is 0. The similarity is measured using the absolute error.

[0172] The above four losses are weighted and summed up by weight coefficients to form a complete loss function. Among them, the weight coefficients are hyperparameters.

[0173] Step S3.1: To construct a multi-contrast multi-anatomical structure k-space fully sampled dataset, in addition to including the cardiac magnetic resonance dataset preprocessed according to Example 1, it is also necessary to preprocess the multi-contrast brain magnetic resonance k-space data including T1-weighted, T2-weighted, and fluid-attenuated inversion recovery (FLAIR), including: (1) Normalize it according to its maximum value range in the image domain so that the amplitude distribution of the entire data in the image domain is within the range of [0,1]; (2) For data samples with too many coil channels, use a channel compression algorithm to compress them to 10 virtual coil channels; (3) Unify the matrix size of all data to 320*384; (4) According to different sampling sequence parameters, assign different contrast categories to the data. Divide T1-weighted, T2-weighted, and FLAIR into 3 groups of contrasts, and add these 3 contrast categories on the basis of the original contrast categories. (5) This dataset only contains 1 type of anatomical structure type, namely "brain transverse section", and add this 1 type of anatomical structure category on the basis of the original anatomical structure categories.

[0174] Step S3.2: Using the preprocessed fully sampled k-space data to jointly train the model: Generate a sampling trajectory from the anatomical structure category label by inputting it into the sampling trajectory generation network. Apply the sampling trajectory to obtain the undersampled k-space. Input the undersampled k-space into the reconstruction model. Calculate the loss function according to the definition method in Step S2 based on the outputs of relevant modules. Update the learnable parameters of the joint sampling trajectory optimization and decoupled representation image reconstruction model by minimizing the loss function until the reconstruction loss on the validation set converges, obtaining the trained deep learning model. In this embodiment, the Adam optimizer is used, with an initial learning rate of 0.0002, which decays to 0.5 times the original every 10 training rounds, and a total of 60 training rounds are performed.

[0175] Figure 4 It is a schematic diagram of the results of this embodiment. Among them, the three rows respectively correspond to FLAIR, T2-weighted, and T1-weighted images; the first column is the fully sampled image reconstruction and quantitative results, and the second to seventh columns are the images reconstructed using different control methods under 8-fold acceleration; the eighth column is the image reconstructed using the method of the present invention under 8-fold acceleration. After reconstruction by the control methods, most images do not have serious undersampling artifacts. However, residual artifacts and blurring appear in the reconstructed images of individual contrasts, especially in the T1-weighted images. In contrast, after reconstruction using the method of the present invention, the reconstructed images have no obvious artifacts, the image details are restored, and the peak signal-to-noise ratio and structural similarity metric with respect to the fully sampled gold standard are improved, laying a foundation for subsequent applications of the images to downstream post-processing and analysis tasks.

[0176] Given any number of images of contrasts and anatomical structure categories, based on the present invention, not only can a sampling trajectory optimized specifically for the anatomical structure category be obtained, but also artifact-free images can be obtained by accelerating the acquisition and reconstruction based on the subsampled trajectory. Compared with the existing conventional deep learning reconstruction models, the present invention increases the robustness to the diversity of magnetic resonance data and has better practical value for diverse data.

[0177] Those skilled in the art can understand this embodiment as a more specific illustration of Embodiment 1 and Embodiment 2.

[0178] Those skilled in the art know that, in addition to implementing the system and its various devices, modules, and units provided by the present invention in the form of pure computer-readable program code, the method steps can be logically programmed to enable the system and its various devices, modules, and units provided by the present invention to be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers, etc. to achieve the same functions. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a kind of hardware component, and the devices, modules, and units included therein for implementing various functions can also be regarded as the structures within the hardware component; the devices, modules, and units for implementing various functions can also be regarded as either software modules for implementing the method or structures within the hardware component.

[0179] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. Without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A multi-anatomical structure multi-contrast magnetic resonance imaging method, characterized in that: The method comprises the following steps: Step S1: construct an image reconstruction model that jointly optimizes sampling trajectories and embeds decoupled representations; Step S2: define a loss function based on decoupled representation learning as an optimization objective; Step S3: constructing a multi-contrast and multi-anatomical structure magnetic resonance dataset, and jointly training the sampling trajectory optimization and image reconstruction model; Step S4: using the trained model, applying the sampling trajectory generated by the sampling trajectory generation network to perform data sampling, and reconstructing the under-sampled data to obtain a reconstructed image with artifacts removed.

2. The multi-anatomical structure multi-contrast magnetic resonance imaging method according to claim 1, characterized in that: The image reconstruction model in step S1 includes a sampling trajectory generation network, a coil sensitivity map estimation network, a data fidelity module, a style encoding network, a content encoding network, a decoding network and a phase artifact removal network; The step S1 comprises the following steps: Step S1.1: The sampling trajectory generation network outputs an optimized sampling trajectory according to the anatomical structure category label corresponding to the input magnetic resonance imaging sample; Step S1.2: the coil sensitivity map estimation network inputs the undersampled k-space data of the multi-coil channels and outputs the coil sensitivity map; Step S1.3: the data fidelity module inputs the current reconstructed image, sampling trajectory, real sampling k-space and coil sensitivity map; Step S1.4: The style encoding network inputs the amplitude and contrast category of the current reconstructed image and outputs the encoded style representation; Step S1.5: the content encoding network inputs the amplitude of the current reconstructed image and outputs the encoded content representation; Step S1.6: Decode the network input style code and content code, and output the amplitude of the reconstructed image; Step S1.7: The phase de-artifacting network inputs the phase of the current reconstructed image and outputs the phase image after de-artifacting.

3. The multi-anatomical structure multi-contrast magnetic resonance imaging method according to claim 1, characterized in that: The loss function based on decoupled representation learning in step S2 includes style loss, content loss, auxiliary classification loss and reconstruction loss. The style loss L style Used to measure the similarity in style of reconstructed images with different contrast ratios, content loss L content Used to measure the similarity between the reconstructed image and the real image in terms of content, auxiliary classification loss L aux To promote the correlation between style encoding and contrast categories, the reconstruction loss L recon Used to promote the similarity between the reconstructed image and the full-acquisition image: L aux (s N ,i)=CE(MLP(s N ),i) Where s is the style code, superscripts p and q represent the indexes of different samples, cos represents cosine similarity, c is the content code, l1 represents the L1 norm, subscript N represents the Nth iteration, i represents the contrast category of the image, CE is the cross entropy loss, and MLP is a classifier based on a multi-layer perceptron. To reconstruct the image, x GT is the real image, and error is an arbitrary difference measure.

4. The multi-anatomical structure multi-contrast magnetic resonance imaging method according to claim 1, characterized in that: The step S3 of constructing a multi-contrast multi-anatomical structure magnetic resonance data set comprises the following steps: Step S3.1: collecting magnetic resonance imaging data containing multiple anatomical structures and contrasts; Step S3.2: preprocessing the collected data, including normalization and denoising; Step S3.3: Divide the preprocessed data into training set, validation set and test set.

5. The multi-anatomical structure multi-contrast magnetic resonance imaging method according to claim 1, characterized in that: The process of data sampling and image reconstruction in step S4 comprises the following steps: Step S4.1: Save the trained deep learning model; Step S4.2: During magnetic resonance imaging, the anatomical structure category label of the imaging task is input into the trained sampling trajectory generation network, and the optimized sampling trajectory is output; Step S4.3: Apply sampling trajectory to obtain k-space data; Step S4.4: Input the sampled data into a reconstruction model consisting of a trained coil sensitivity map estimation network, a data fidelity module, a style encoding network, a content encoding network, a decoding network and a phase artifact removal network to obtain an artifact-free reconstructed image.

6. A multi-anatomical structure multi-contrast magnetic resonance imaging system, characterized in that: The system includes the following modules: Module M1: Construct an image reconstruction model with joint sampling trajectory optimization and embedded decoupled representation; Module M2: Define the loss function based on decoupled representation learning as the optimization objective; Module M3: construct a multi-contrast and multi-anatomical structure magnetic resonance dataset, and jointly train the sampling trajectory optimization and image reconstruction model; Module M4: using the trained model, applying the sampling trajectory generated by the sampling trajectory generation network to perform data sampling, and reconstructing the under-sampled data to obtain a reconstructed image with artifacts removed.

7. The multi-anatomical structure multi-contrast magnetic resonance imaging system according to claim 6, characterized in that: The image reconstruction model in the module M1 includes a sampling trajectory generation network, a coil sensitivity map estimation network, a data fidelity module, a style encoding network, a content encoding network, a decoding network and a phase artifact removal network; The module M1 includes the following modules: Module M1.1: The sampling trajectory generation network outputs an optimized sampling trajectory according to the anatomical structure category label corresponding to the input magnetic resonance imaging sample; Module M1.2: The coil sensitivity map estimation network inputs the undersampled k-space data of multiple coil channels and outputs the coil sensitivity map; Module M1.3: The data fidelity module inputs the current reconstructed image, sampling trajectory, true sampling k-space and coil sensitivity map; Module M1.4: The style encoding network inputs the amplitude and contrast category of the current reconstructed image and outputs the encoded style representation; Module M1.5: The content encoding network inputs the amplitude of the current reconstructed image and outputs the encoded content representation; Module M1.6: Decode the network input style code and content code, and output the amplitude of the reconstructed image; Module M1.7: The phase de-artifacting network inputs the phase of the current reconstructed image and outputs the phase image after de-artifacting.

8. The multi-anatomical structure multi-contrast magnetic resonance imaging system according to claim 6, characterized in that: The loss function based on decoupled representation learning in the module M2 includes style loss, content loss, auxiliary classification loss and reconstruction loss. The style loss L style Used to measure the similarity in style of reconstructed images with different contrast ratios, content loss L content Used to measure the similarity between the reconstructed image and the real image in terms of content, auxiliary classification loss L aux To promote the correlation between style encoding and contrast categories, the reconstruction loss L recon Used to promote the similarity between the reconstructed image and the full-acquisition image: L aux (s N ,i)=CE(MLP(s N ),i) Where s is the style code, superscripts p and q represent the indexes of different samples, cos represents cosine similarity, c is the content code, l1 represents the L1 norm, subscript N represents the Nth iteration, i represents the contrast category of the image, CE is the cross entropy loss, and MLP is a classifier based on a multi-layer perceptron. To reconstruct the image, x GT is the real image, and error is an arbitrary difference measure.

9. The multi-anatomical structure multi-contrast magnetic resonance imaging system according to claim 6, characterized in that: The module M3 constructs a multi-contrast multi-anatomical structure magnetic resonance data set, including the following modules: Module M3.1: Collect MRI data containing a variety of anatomical structures and contrasts; Module M3.2: Preprocess the collected data, including normalization and denoising; Module M3.3: Divide the preprocessed data into training set, validation set and test set.

10. The multi-anatomical structure multi-contrast magnetic resonance imaging system according to claim 6, characterized in that: The process of data sampling and image reconstruction in the module M4 includes the following modules: Module M4.1: Save the trained deep learning model; Module M4.2: During magnetic resonance imaging, the anatomical structure category label of the imaging task is input into the trained sampling trajectory generation network, and the optimized sampling trajectory is output; Module M4.3: Apply sampling trajectory to obtain k-space data; Module M4.4: Input the sampled data into the reconstruction model composed of the trained coil sensitivity map estimation network, data fidelity module, style encoding network, content encoding network, decoding network and phase artifact removal network to obtain the artifact-free reconstructed image.

Images