Multi-anatomy multi-contrast magnetic resonance imaging method and system

Abstract

The application provides a multi-anatomy multi-contrast magnetic resonance imaging method and system, comprising: constructing an image reconstruction model of joint sampling trajectory optimization and embedded decoupling representation. Defining a loss function based on decoupling representation learning as an optimization objective. Constructing a multi-contrast multi-anatomy magnetic resonance dataset, jointly training the sampling trajectory optimization and the image reconstruction model. Using the trained model, the sampling trajectory generated by the sampling trajectory generation network is used for data sampling, and the undersampled data is reconstructed to obtain a reconstructed image without artifacts. With the aid of decoupling representation learning and anatomy-specific sampling trajectory modeling, the application can accelerate the acquisition and reconstruction in the magnetic resonance data scene containing multiple anatomical structures and contrasts, and realize the joint optimization of the magnetic resonance sampling and reconstruction process.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular 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 (MRI) can produce images with high tissue contrast, which can be used to reveal tissue anatomy or calculate quantitative tissue parameters, providing a basis for diagnosis and treatment of related diseases. However, due to its slow acquisition speed, MRI is often time-consuming. Therefore, clinical practice often requires accelerated undersampling in k-space to shorten the sampling time. Directly performing an inverse Fourier transform on undersampled, accelerated data will produce images with undersampling artifacts. Therefore, many image reconstruction methods have been proposed for undersampled data. Traditional parallel imaging and compressed sensing methods perform well on low-magnification, undersampled data, but perform poorly at higher sampling magnifications and low signal-to-noise ratios. Recently, many deep learning-based image reconstruction methods have been proposed. By training on large amounts of MRI data, they can produce artifact-free reconstructed images from highly accelerated data. Existing deep learning methods typically assume that the training dataset is independent and identically distributed, i.e., that the training data are independent samples drawn from the same distribution. However, this assumption does not hold true for MRI data in many application scenarios. For example, the contrast of MRI images is affected by imaging sequence parameters (such as repetition time (TR), echo time (TE), and inversion time (TI)) and the manufacturer's specific MRI equipment. The anatomical structure of the image is affected by the imaged organ and anatomical orientation. Furthermore, the pattern of undersampling artifacts is influenced by multiple factors, including the acquisition signal-to-noise ratio and the k-space sampling trajectory. Many existing deep learning image reconstruction methods cannot guarantee robust results when training data exhibits domain shift, high diversity, and heterogeneity. They may experience degradation in reconstruction performance on certain samples and deviations in reconstruction performance on specific domains. Therefore, in order to advance deep learning MRI methods towards clinical application, it is necessary to propose a universal reconstruction model that performs well on diverse MRI data.

[0003] To implement this general model, it is necessary to specifically address the impact of image contrast and anatomical structure on data distribution. Within the field of deep learning, disentangled representation learning is a strategy that can be used to reduce inter-domain differences in data and achieve domain adaptation. This strategy uses deep neural networks to decouple data into a style encoding that reflects domain-specific information and a content encoding that is domain-invariant. This allows the task to be accomplished using the domain-invariant content encoding. Disentangled representation learning has been widely applied to tasks such as image classification, segmentation, and style transfer, but it remains less widely used in image reconstruction. Compared to these tasks, magnetic resonance image reconstruction presents its own unique characteristics and complexities. Unlike segmentation and classification, which only require the domain-invariant semantic information contained in the content encoding, magnetic resonance image reconstruction requires both precise content encoding for artifact removal and image anatomical structure restoration, and the integration of style encoding to restore image contrast information. Existing medical image segmentation methods based on disentangled representation learning often neglect the extraction of image texture information from the content encoding, making direct transfer to image reconstruction difficult. On the other hand, to ensure the robustness and reliability of the MRI reconstruction process, the reconstructed image must maintain fidelity with respect to the k-space sampling data. Therefore, the reconstruction method must incorporate modeling of the physical imaging process. However, existing decoupled representation learning methods treat image reconstruction as a purely data-driven image restoration task, failing to incorporate this information and failing to meet practical requirements. Therefore, it is necessary to propose a reconstruction method that can both address cross-domain image reconstruction and incorporate physical imaging models.

[0004] The influence of anatomical structure on the distribution of MRI data is also indirectly reflected in k-space data. Even under the same undersampling mode, anatomical structures from different imaging orientations and organs will produce different distributions of sampled data and undersampling artifacts. A series of works on MRI acquisition optimization aim to find the optimal k-space sampling trajectory for the entire dataset through deep learning. However, they cannot adapt to the data distribution differences caused by different anatomical structures. Inspired by this existing work, we hypothesize that each anatomical structure type has its own optimal sampling mode. Based on this, we need to propose a sampling trajectory optimization scheme for specific anatomical structures. Summary of the Invention

[0005] In view of the defects in the prior art, the present invention aims to provide a multi-anatomical structure multi-contrast magnetic resonance imaging method and system.

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

[0007] Step S1: Construct an image reconstruction model that jointly optimizes sampling trajectory and embeds decoupled representation;

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

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

[0010] Step S4: using the trained model, applying the sampling trajectory generated by the sampling trajectory generation network to perform data sampling, and performing image reconstruction on the undersampled data to obtain a reconstructed image with artifacts removed.

[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] The step S1 comprises the following steps:

[0013] Step S1.1: The sampling trajectory generation network outputs an optimized sampling trajectory based on 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 multiple coil channels and outputs the coil sensitivity map;

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

[0016] Step S1.4: The style encoding network inputs the amplitude and contrast category of the current reconstructed image 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: Decode the network input style code and content code, and output the amplitude of the reconstructed image;

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

[0020] Preferably, the loss function based on decoupled representation learning in step S2 includes style loss, content loss, auxiliary classification loss and reconstruction loss, and 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 Used to promote the correlation between style encoding and contrast categories, reconstruction loss Lrecon Used 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 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 metric.

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

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

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

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

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

[0033] Step S4.4: 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.

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

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

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

[0037] Module M3: Constructing a multi-contrast and multi-anatomical structure magnetic resonance dataset and jointly training the sampling trajectory optimization and image reconstruction model;

[0038] Module M4: Using the trained model, applying the sampling trajectory generated by the sampling trajectory generation network to perform data sampling, and reconstructing the undersampled data to obtain a reconstructed image with artifacts removed.

[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 modules:

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

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

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

[0044] Module M1.4: The style encoding network inputs the amplitude and contrast category of the current reconstructed image 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: Decodes the network input style code and content code, and outputs the amplitude of the reconstructed image;

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

[0048] Preferably, the loss function based on decoupled representation learning in the module M2 includes style loss, content loss, auxiliary classification loss and reconstruction loss, and the style loss Lstyle 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 Used to promote the correlation between style encoding and contrast categories, reconstruction loss L recon Used to promote the similarity between the reconstructed image and the full-acquisition image:

[0049]

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

[0051]

[0052] 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 metric.

[0053] Preferably, the module M3 for constructing a multi-contrast and multi-anatomical structure magnetic resonance dataset includes the following modules:

[0054] Module M3.1: Collect MRI data containing a variety of 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 training set, validation set and 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, the anatomical structure category labels of the imaging task are input into the trained sampling trajectory generation network, and the optimized sampling trajectory is output;

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

[0061] 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.

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

[0063] 1. This paper integrates decoupled representation learning into an iterative unfolding network model. By decoupling images into style and content encodings, it fully utilizes the cross-domain invariance of multi-contrast MRI data. Compared with conventional reconstruction methods, it explicitly introduces a mechanism to deal with the problem of large domain differences in data.

[0064] 2. Compared with the existing method combining decoupled characterization and image reconstruction, the present invention combines the imaging physical model, satisfies the fidelity constraints of k-space sampling 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 for different anatomical structures respectively. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

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

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

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

[0071] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0072] Example 1:

[0073] Reference Figure 1 and Figure 2 According to the present invention, a multi-anatomical structure multi-contrast magnetic resonance imaging method is provided, the method comprising the following steps:

[0074] Step S1: Construct an image reconstruction model that combines sampling trajectory optimization and embedding 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;

[0075] Step S1.1: The sampling trajectory generation network outputs an optimized sampling trajectory based on 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 multiple coil channels and outputs the coil sensitivity map;

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

[0078] Step S1.4: The style encoding network inputs the amplitude and contrast category of the current reconstructed image 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: Decode the network input style code and content code, and output the amplitude of the reconstructed image;

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

[0082] Step S2: Define the loss function based on decoupled representation learning as the optimization target; the loss function based on decoupled representation learning 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 Used to promote the correlation between style encoding and contrast categories, reconstruction loss L recon Used to promote the similarity between the reconstructed image and the full-acquisition image:

[0083]

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

[0085]

[0086] 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 metric.

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

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

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

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

[0091] Step S4: using the trained model, applying the sampling trajectory generated by the sampling trajectory generation network to perform data sampling, and performing image reconstruction on the undersampled data to obtain a reconstructed image with artifacts removed.

[0092] The process of 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, 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;

[0095] Step S4.3: Apply sampling trajectory to obtain k-space data;

[0096] Step S4.4: 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.

[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, the system comprising the following modules:

[0100] Module M1: Construct an image reconstruction model that combines sampling trajectory optimization and embedding 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 based on the anatomical structure category label corresponding to the input magnetic resonance imaging sample;

[0102] Module M1.2: Coil sensitivity map estimation network inputs undersampled k-space data of multiple coil channels and outputs coil sensitivity maps;

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

[0104] Module M1.4: The style encoding network inputs the amplitude and contrast category of the current reconstructed image 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: Decodes the network input style code and content code, and outputs the amplitude of the reconstructed image;

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

[0108] Module M2: Define the loss function based on decoupled representation learning as the optimization target; the loss function based on decoupled representation learning 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 contentUsed to measure the similarity between the reconstructed image and the real image in terms of content, auxiliary classification loss L aux Used to promote the correlation between style encoding and contrast categories, reconstruction loss L recon Used to promote the similarity between the reconstructed image and the full-acquisition image:

[0109]

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

[0111]

[0112] 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 metric.

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

[0114] Module M3.1: Collect MRI data containing a variety of 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 training set, validation set and test set.

[0117] Module M4: Using the trained model, applying the sampling trajectory generated by the sampling trajectory generation network to perform data sampling, and reconstructing the undersampled data to obtain a reconstructed image with artifacts removed.

[0118] The process of 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, the anatomical structure category labels of the imaging task are input into the trained sampling trajectory generation network, and the optimized sampling trajectory is output;

[0121] Module M4.3: Apply sampling trajectory to obtain k-space data;

[0122] 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.

[0123] Example 3:

[0124] An embodiment of the present invention applied to cardiac magnetic resonance quantitative imaging will be described in detail below with reference to the accompanying drawings. Figure 1 is an implementation flow chart of the present invention, Figure 2 Schematic diagram of the algorithm of the present invention. Specifically, this embodiment includes the following steps:

[0125] Step S1: Joint sampling trajectory optimization and embedding of an image reconstruction model with decoupled representation. The model includes the following modules:

[0126] 1) Sampling Trajectory Generation Network: The input is the anatomical structure category label corresponding to the MRI sample, and the output of the neural network is the generated sampling trajectory. The structure category label is first one-hot encoded to obtain a vector reflecting the anatomical structure category. This vector is then input into the neural network. The network architecture is a multilayer perceptron with three hidden layers, 64, 128, and 256 hidden neurons, respectively. The activation functions for the first two layers are ReLU, and the activation function for the last layer is softmax. The output is a vector of length 256 (the number of phase-encoding lines during full acquisition). The central 16 elements of the vector (corresponding to the central 16 full-acquisition phase-encoding lines used to calculate coil sensitivity) are set to 1. This vector is then input into a sigmoid function to obtain a sampling trajectory probability vector normalized to the range of 0-1. The number of sampling lines, k, is preset by dividing the total number of phase-encoding lines by the acceleration factor (in this example, 4x and 8x acceleration are used to construct two models, respectively). The k phase-encoding lines corresponding to the k elements with the highest probability are selected as the sampling trajectory.

[0127] 2) Coil Sensitivity Map Estimation Module: This module is used to generate coil sensitivity maps from k-space data obtained from multi-coil parallel imaging. Its input is the multi-coil parallel imaging k-space, and its output is the coil sensitivity map. Specifically, this module first extracts the 16 fully acquired phase-encoded lines in the center of the multi-coil k-space using a mask, transforms it to the image domain using an inverse Fourier transform, and inputs it into a U-Net network. Each block in the encoder portion of the U-Net network sequentially performs two convolutions, instance normalization, and a LeakyReLU activation function, followed by one average pooling. The first block has 8 convolutional channels, and the number of channels is doubled after each subsequent pooling downsampling, for a total of four downsampling steps. The decoder portion sequentially performs deconvolution with a kernel size of 2 and a stride of 2, followed by LeakyReLU activation. The output is obtained after four deconvolution upsampling steps and skip connections.

[0128] 3) Data Fidelity Module: This module takes as input the current reconstructed image, the sampling trajectory, the true sampled k-space, and the coil sensitivity map. It converts the current image to k-space by element-by-element multiplication of the coil sensitivity map. After data fidelity processing, it is converted back to the image domain via channel merging and outputs an image constrained by data fidelity. The data fidelity processing first subtracts the current reconstructed k-space data from the true undersampled k-space data. The subtraction is then masked by the sampling trajectory and multiplied by the iteration step size (set as a learnable variable) for that iteration. The result is then added to the current reconstructed k-space before processing to produce the k-space data.

[0129] 4) Style Encoding Network: This network takes as input the amplitude and contrast category of the reconstructed image and outputs the encoded style representation. The network architecture is a convolutional neural network based on conditional convolution. The conditional convolution accepts contrast category labels, making the parameters of the final convolution kernel dependent on contrast. This convolutional neural network uniformly samples the input image to a fixed size and passes it through five conditional convolutional layers with output channels of 16, 32, 64, 128, and 128, followed by the LeakyReLU activation function. The resulting features are then passed through an MLP with 32 and 16 output neurons, respectively, to produce a style code of length 16.

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

[0131] 6) Decoder Network: This network inputs the style and content codes obtained from the encoding described above and outputs a reconstructed image. One branch of the network uses a U-Net variant architecture, feeding the style code into the U-Net encoder to generate intermediate features. The other branch feeds the content code into a 4-layer multilayer perceptron with 128 hidden neurons per layer to generate mapped features. These features are then repeated in both the width and height dimensions and concatenated with the intermediate features of the same size in the channel dimension. The concatenated features are then fed into the U-Net decoder to generate the output image. The encoder-decoder architecture is identical to the coil sensitivity estimation network.

[0132] 7) Phase Artifact Removal Network: This network takes as input 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 decoupling representation image reconstruction model includes the following steps:

[0134] First, the anatomical structure information category label of the image to be imaged (multiple categories are preset, such as coronal, sagittal, axial, and cardiac short axis) is input into the sampling trajectory generation network to generate a sampling trajectory, which is used to undersample the fully acquired 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, and the image amplitude and data contrast category are input into the style encoding network. The image amplitude is also input into the content encoding network. The style encoding network and the content encoding network output style and content codes respectively; (3) the obtained style code and content code 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-artifacted, and the next iteration is repeated.

[0137] After 12 iterations, the image obtained from the last iteration is converted to the k-space domain using the coil sensitivity map, then converted to the image domain using an inverse Fourier transform. A square sum operation is then performed on the channel dimension to create a reconstructed magnitude image. The style code output by the style encoding network in the last iteration is then input to the auxiliary classifier, which then outputs the classification result for the corresponding contrast category. The style code and content code generated at each iteration, as well as the final reconstructed image and classification result, are used to calculate the training loss function 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 reconstructed image output by the model and the full gold standard image x GT The difference metric between them is used as the reconstruction loss. The difference metric uses 10 times the absolute value error of each pixel and 1 times the structural similarity error, that is, 1 minus the structural similarity metric (SSIM):

[0140]

[0141] (2) Style loss (Lstyle): During training, the similarity between the style codes s obtained in each iteration is calculated between two samples p and q 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, the value of 1 minus the cosine value between the two vectors is minimized to make them in the same direction. The results of each iteration are summed to obtain the final style loss:

[0142]

[0143] (3) Auxiliary classification loss (Laux): In order to further promote the style encoder to extract contrast-related losses, this loss is designed as an auxiliary classification task to calculate the accuracy of contrast classification with the help of style encoding. During the training process, a classification module based on a multi-layer perceptron (MLP) is introduced and trained synchronously with the network. Its input is the style encoding obtained in the last iteration. It contains 6 hidden layers (number of neurons: 256, 128, 64, 32, 16), and the activation function is ReLU (first 5 layers) and sigmoid (last layer). It outputs the contrast category i to which the predicted image belongs. The multi-class cross entropy (CE) loss between the predicted category and the true contrast category is calculated, which is the auxiliary classification loss:

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

[0145] (4) Content loss (Lcont): During training, the similarity between the content encodings c obtained at each iteration is calculated for two samples p and q in a batch. If the two images are from different contrast images scanned from the same layer, the similarity between their content encodings is maximized; otherwise, if the contents of the two images are not paired, the content loss value is 0. Similarity is measured by the absolute value error l1:

[0146]

[0147] The above four losses are weighted and summed by 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 MRI dataset, it is necessary to preprocess the multi-contrast multi-anatomical structure cardiac MRI k-space data, including short-axis T1mapping, T2mapping, sagittal / transverse aortic imaging, short-axis / two-chamber heart / three-chamber heart / four-chamber heart / aortic outflow tract cardiac cine imaging, including: (1) normalizing 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) using a channel compression algorithm to compress data samples with too many coil channels to 10 virtual coil channels; (3) assigning different contrast categories to the data according to different sampling parameters. The contrast of aortic imaging and cine imaging data is relatively consistent, and they are each classified as a separate category. The T1mapping data acquired by the MOLLI sequence has different inversion times (TI), so the sampling data is divided into three contrast groups 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 three contrast groups according to this parameter. In this way, data containing 8 different contrast categories are constructed; (4) according to the anatomical structure category of each data, it is marked as sagittal, transverse, short axis, two-chamber heart, three-chamber heart, four-chamber heart, and aortic outflow tract, a total of 8 different anatomical structure categories.

[0149] Step S3.2: Jointly train the model using preprocessed fully sampled k-space data: Input the sampling trajectory into the sampling trajectory generation network using anatomical structure category labels to obtain a sampling trajectory. Apply the sampling trajectory to obtain undersampled k-space, which is then input into the reconstruction model. Based on the outputs of the relevant modules, a loss function is calculated as defined in step S2. The learnable parameters of the joint sampling trajectory optimization and decoupled representation image reconstruction model are updated by minimizing the loss function until the reconstruction loss on the validation set converges, resulting in a trained deep learning model. In this embodiment, the Adam optimizer is used with an initial learning rate of 0.0002, which is decayed by 0.5 times every 10 training rounds, for a total of 60 rounds.

[0150] Step S4: The data sampling and image reconstruction process includes: saving the trained deep learning model. For short-axis imaging, for example, the anatomical structure category labels corresponding to the short-axis position are input into the trained sampling trajectory generation network, which outputs an optimized sampling trajectory. The sampling trajectory is then used to sample k-space data, which is then input 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 produce an artifact-free reconstructed image. For T2 mapping, the final T2 quantitative map can be generated by fitting the reconstructed image based on the different T2 preparation times during acquisition.

[0151] Figure 3 The results of this example are shown in Figure 1. The first three rows correspond to T2 mapping images with T2 preparation times of 0, 35, and 55 milliseconds, respectively. The fourth row corresponds to T2 quantitative maps obtained by fitting multi-contrast images. The first column shows the reconstruction and quantitative results of fully sampled images. Columns 2-7 show images reconstructed using different control methods at 4x acceleration and the resulting quantitative maps. Column 8 shows images and quantitative maps reconstructed using the method of the present invention at 4x acceleration. Columns 9-14 show images and quantitative maps reconstructed using the control method at 8x acceleration. Column 15 shows images and quantitative maps reconstructed using the method of the present invention at 8x acceleration. After reconstruction using the control method, most images showed no significant undersampling artifacts. However, residual artifacts and blurring appeared in the reconstructed images of certain contrasts, leading to artifacts and blurring in the T2 quantitative maps. In contrast, after reconstruction using the method of the present invention, the image quality indicators of the reconstructed images relative to the full sampling gold standard were significantly improved. Furthermore, images reconstructed using the method of the present invention were used to fit the T2 quantitative maps. The images showed intact myocardial regions with clear boundaries and no obvious undersampling artifacts, indicating improved overall quality.

[0152] Example 4:

[0153] like Figure 1The present invention discloses a method for joint sampling trajectory optimization and image reconstruction of magnetic resonance imaging based on decoupled representation learning, which comprises the following steps:

[0154] Step S1: Construct an image reconstruction model that combines sampling trajectory optimization and embedding decoupled representation. The model contains the following modules:

[0155] 1) Sampling Trajectory Network: This network inputs the anatomical structure category label corresponding to the MRI sample, and outputs the generated sampling trajectory. The structure category label is first one-hot encoded to obtain a vector reflecting the anatomical structure category, which is then input into the network. The network architecture consists of a three-layer multilayer perceptron with 64, 128, and 256 hidden neurons, respectively, and a Reluctant Unit (ReLU) activation function, except for the last layer. The output is a vector of length 256 (the number of phase-encoding lines during full acquisition). A large value is added to the 16 elements in the central region of the vector (corresponding to the central 16 phase-encoding lines used to calculate coil sensitivity). This vector is then fed into a sigmoid function to obtain a sampling trajectory probability distribution vector normalized to the range 0-1. Based on the preset number of sampling lines, k, the k phase-encoding lines corresponding to the k elements with the highest probability are selected as the sampling trajectory.

[0156] 2) Coil Sensitivity Map Estimation Module: This module is used to generate coil sensitivity maps from k-space data obtained from multi-coil parallel imaging. Its input is the multi-coil parallel imaging k-space, and its output is the coil sensitivity map. Specifically, this module first extracts the 16 fully acquired phase-encoded lines in the center of the multi-coil k-space using a mask, transforms it to the image domain using an inverse Fourier transform, and inputs it into a U-Net network. Each block in the encoder portion of the U-Net network sequentially performs two convolutions, instance normalization, and a LeakyReLU activation function, followed by one average pooling. The first block has 8 convolutional channels, and the number of channels is doubled after each subsequent pooling downsampling, for a total of four downsampling steps. The decoder portion sequentially performs deconvolution with a kernel size of 2 and a stride of 2, followed by LeakyReLU activation. The output is obtained after four deconvolution upsampling steps and skip connections.

[0157] 3) Data Fidelity Module: This module takes as input the current reconstructed image, the sampling trajectory, the true sampled k-space, and the coil sensitivity map. It converts the current image to k-space by element-by-element multiplication of the coil sensitivity map. After data fidelity processing, it is converted back to the image domain via channel merging and outputs an image constrained by data fidelity. The data fidelity processing first subtracts the current reconstructed k-space data from the true undersampled k-space data. The subtraction is then masked by the sampling trajectory and multiplied by the iteration step size (set as a learnable variable) for that iteration. The result is then added to the current reconstructed k-space before processing to produce the k-space data.

[0158] 4) Style Encoding Network: This network takes as input the amplitude and contrast category of the reconstructed image and outputs the encoded style representation. Its architecture is a convolutional neural network based on conditional convolutions, which accept contrast category labels and apply different convolution kernel parameters for different contrasts. This convolutional neural network uniformly samples the input image to a fixed size and passes it through five conditional convolutional layers with output channels of 16, 32, 64, 128, and 128, followed by a LeakyReLU activation function. The resulting features are then passed through an MLP with 32 and 16 output neurons, respectively, to produce a style code of length 16.

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

[0160] 6) Decoder Network: This network inputs the style and content codes obtained from the encoding described above and outputs a reconstructed image. One branch of the network uses a U-Net variant architecture, feeding the style code into the U-Net encoder to generate intermediate features. The other branch feeds the content code into a 4-layer multilayer perceptron with 128 hidden neurons per layer to generate mapped features. These features are then repeated in both the width and height dimensions and concatenated with the intermediate features of the same size in the channel dimension. The concatenated features are then fed into the U-Net decoder to generate the output image. The encoder-decoder architecture is identical to the coil sensitivity estimation network.

[0161] 7) Phase Artifact Removal Network: This network takes as input 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.

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

[0163] First, the anatomical structure category to which the imaging data belongs is input into the sampling trajectory generation network, which outputs the sampling trajectory. The trajectory is applied to sample the fully sampled k-space training data to obtain undersampled k-space data.

[0164] Next, the undersampled k-space data is input into the coil sensitivity map estimation module to obtain the coil sensitivity map, and the coil sensitivity map is used to merge the multi-channel k-space data to obtain the 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, and the image amplitude and data contrast category are input into the style encoding network. The image amplitude is also input into the content encoding network. The style encoding network and the content encoding network output style and content codes respectively; (3) the obtained style code and content code 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-artifacted, and the next iteration is repeated.

[0166] After 12 iterations, the image obtained from the last iteration is converted to the k-space domain using the coil sensitivity map, then converted to the image domain using an inverse Fourier transform. A square sum operation is then performed on the channel dimension to create a reconstructed magnitude image. The style code output by the style encoding network in the last iteration is then input to the auxiliary classifier, which then outputs the classification result for the corresponding contrast category. The style code and content code generated at each iteration, as well as the final reconstructed image and classification result, are used to calculate the training loss function 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 sampled 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 at 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 between the two vectors is minimized to make them orthogonal; to maximize the similarity, the value of 1 minus the cosine 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 is designed as an auxiliary classification task to calculate the accuracy of contrast classification with the help of style encoding. During the training process, a classification module based on a multi-layer perceptron is introduced and trained synchronously with the network. Its input is the style encoding obtained in the last iteration. It contains 6 hidden layers (number of neurons: 256, 128, 64, 32, 16), and the activation function is ReLU (first 5 layers) and sigmoid (last layer). The output is the contrast category to which the predicted image belongs. The multi-class cross entropy loss between the predicted category and the true contrast category is calculated, which is the auxiliary classification loss.

[0171] (4) Content loss (Lcont): During training, the similarity between the content encodings of two samples in a batch is calculated at each iteration. If the two images are from the same layer but with different contrast, the similarity between their content encodings is maximized. Conversely, if the contents of the two images are not paired, the content loss is 0. Similarity is measured using the absolute error.

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

[0173] Step S3.1: In order to construct a multi-contrast multi-anatomical structure k-space full sampling dataset, in addition to incorporating the cardiac MRI dataset obtained by preprocessing according to Example 1, it is also necessary to preprocess the multi-contrast brain MRI k-space data including T1-weighted, T2-weighted, and fluid-attenuated inversion recovery (FLAIR) data, including: (1) normalizing them according to their 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) using a channel compression algorithm to compress data samples with too many coil channels to 10 virtual coil channels; (3) unifying the matrix size of all data to 320*384; (4) assigning different contrast categories to the data according to different sampling sequence parameters. T1-weighted, T2-weighted, and FLAIR are divided into three contrast groups, and these three contrast categories are added on the basis of the original contrast categories. (5) This dataset only contains one anatomical structure type, "brain cross-section", and this one anatomical structure category is added on the basis of the original anatomical structure category.

[0174] Step S3.2: Jointly train the model using preprocessed fully sampled k-space data: Input the sampling trajectory into the sampling trajectory generation network using anatomical structure category labels to obtain a sampling trajectory. Apply the sampling trajectory to obtain undersampled k-space, which is then input into the reconstruction model. Based on the outputs of the relevant modules, a loss function is calculated as defined in step S2. The learnable parameters of the joint sampling trajectory optimization and decoupled representation image reconstruction model are updated by minimizing the loss function until the reconstruction loss on the validation set converges, resulting in a trained deep learning model. In this embodiment, the Adam optimizer is used with an initial learning rate of 0.0002, which is decayed by 0.5 times every 10 training rounds, for a total of 60 rounds.

[0175] Figure 4 The figure is a schematic diagram of the results of this embodiment. Among them, the three rows correspond to FLAIR, T2-weighted, and T1-weighted images respectively; the first column is the full sampling image reconstruction and quantitative results, the second to seventh columns are images reconstructed using different control methods at 8 times acceleration; the eighth column is the image reconstructed using the method of the present invention at 8 times acceleration. After reconstruction using the control method, most images have no serious undersampling artifacts, however, residual artifacts and blurring appear in the reconstructed images of individual contrasts, especially in T1-weighted images. In contrast, after reconstruction using the method of the present invention, the reconstructed image has no obvious artifacts, the image details are restored, and the peak signal-to-noise ratio and structural similarity metric of the full sampling gold standard are improved, laying the foundation for the subsequent application of images to downstream post-processing and analysis tasks.

[0176] Given any given set of contrast and anatomical structure categories, the present invention not only generates a sampling trajectory optimized for specific anatomical structure categories, but also accelerates acquisition and reconstruction of artifact-free images based on subsampling trajectories. Compared to existing conventional deep learning reconstruction models, the present invention increases robustness to the diversity of magnetic resonance data, offering greater practical value for diverse data.

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

[0178] Those skilled in the art will appreciate that, in addition to implementing the system and its various devices, modules, and units provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same functions of the system and its various devices, modules, and units provided by the present invention in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; the devices, modules, and units for implementing various functions can also be considered as both software modules implementing the method and structures within the hardware component.

[0179] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

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 trajectory and embeds decoupled representation; Step S2: define a loss function based on decoupled representation learning as the 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 undersampled data to obtain a reconstructed image with artifacts removed; 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 based on 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 multiple coil channels and outputs the coil sensitivity map; Step S1.3: The data fidelity module inputs the current reconstructed image, sampling trajectory, true 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. The process of data sampling and image reconstruction in step S4 includes 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 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.

2. 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 Used to promote the correlation between style encoding and contrast categories, 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 metric.

3. The multi-anatomical structure multi-contrast magnetic resonance imaging method according to claim 1, characterized in that: Constructing the multi-contrast and multi-anatomical structure magnetic resonance dataset in step S3 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.

4. 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 that combines sampling trajectory optimization and embedding decoupled representation; Module M2: Define the loss function based on decoupled representation learning as the optimization objective; Module M3: Constructing a multi-contrast and multi-anatomical structure magnetic resonance dataset and jointly training 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 undersampled data to obtain a reconstructed image with artifacts removed; 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 based on the anatomical structure category label corresponding to the input magnetic resonance imaging sample; Module M1.2: Coil sensitivity map estimation network inputs undersampled k-space data of multiple coil channels and outputs coil sensitivity maps; 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: Decodes the network input style code and content code, and outputs the amplitude of the reconstructed image; Module M1.7: Phase artifact removal network inputs the phase of the current reconstructed image and outputs the phase image after artifact removal; 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 labels of the imaging task are 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.

5. The multi-anatomical structure multi-contrast magnetic resonance imaging system according to claim 4, 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 Used to promote the correlation between style encoding and contrast categories, 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 metric.

6. The multi-anatomical structure multi-contrast magnetic resonance imaging system according to claim 4, characterized in that: The module M3 constructs a multi-contrast multi-anatomical structure magnetic resonance dataset 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.

Images