A rapid magnetic resonance multi-sequence combined imaging method, system, device and medium

Abstract

The application discloses a kind of fast magnetic resonance multi-sequence joint imaging method, system, equipment and medium, it is related to deep learning technical field, including the following steps: acquisition different sequence multiple undersampling MR images, wherein multiple undersampling MR images share the same public feature, each undersampling MR image enjoys respective corresponding specific feature;The mapping relationship between multiple undersampling MR images and latent variable is modeled based on latent variable model, and the optimization objective of reconstruction task in fast magnetic resonance multi-sequence joint imaging is constructed based on mapping relationship;Optimization objective is solved by depth development network.The features of multiple undersampling MR images are subdivided into intermodal shared public features and modality unique specific features, the fine decoupling of the shared and specific features of MR image, solve the problem that intermodal correlation modeling is weak, and redundant information can be effectively avoided or sequence-specific detail information is lost.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a fast magnetic resonance multi-sequence joint imaging method, system, device and medium. Background Technology

[0002] Magnetic Resonance Imaging (MRI) is a widely used medical imaging technique for diagnosis. Unlike X-ray or CT (Computed Tomography) scans, MRI does not involve ionizing radiation. As a non-invasive technique, MRI is particularly useful for examining the brain, spinal cord, muscles, ligaments, and internal organs, and can accurately identify abnormalities such as tumors, inflammation, and degenerative diseases. However, due to limitations in the physical imaging principles and the Nyquist sampling theorem, MRI faces two main bottlenecks: slow acquisition speed and large data acquisition volume.

[0003] To overcome these bottlenecks, researchers have proposed a variety of fast magnetic resonance imaging methods to improve the speed and efficiency of MRI. Their work mainly focuses on three aspects: (1) designing faster acquisition sequences, such as fast spin echo sequences and gradient echo sequences; (2) using multiple coils for parallel imaging; and (3) performing undersampling (acquiring only a portion of the data points in the k-space, such as 10% or 15%) and then using prior knowledge of the image for MRI reconstruction.

[0004] In recent years, with the rapid development of deep learning technology, deep learning-based MRI reconstruction methods have been widely applied in fast magnetic resonance imaging. Currently, commonly used data-driven MRI reconstruction network structures can be broadly classified into three categories: single-sequence MRI reconstruction, multi-sequence MRI reference reconstruction, and multi-sequence MRI joint reconstruction. Among these, single-sequence MRI reconstruction methods are relatively simple and effective, improving the clarity and detail of undersampled images. However, these methods rely solely on information from a single sequence image, which may limit their ability to recover fine details, especially in cases of undersampling or poor image quality. Unlike single-sequence MRI reconstruction methods, multi-sequence reconstruction methods leverage the correlation and similarity between different MRI sequences, achieving higher performance. In multi-sequence MRI reference reconstruction, the input includes MR images of a reference sequence and MR images of a target sequence. The former is typically a fully sampled reconstructed image, while the latter is an undersampled reconstructed image. This framework aims to supplement and reconstruct the target sequence using detailed information from the reference sequence. For multi-sequence MRI joint reconstruction, the input consists of undersampled MR images from different sequences. Unlike reference reconstruction, joint reconstruction aims to utilize shared information between different sequences to improve the quality and efficiency of MRI reconstruction.

[0005] However, existing multi-sequence MRI joint reconstruction methods still have some shortcomings, mainly including: difficulty in aligning cross-modal features due to complex anatomical deformations between multiple sequence images, such as motion artifacts or non-rigid deformations; weak modeling of correlations between modalities; and insufficient decoupling of shared and specific features, which may introduce redundant information or lose sequence-specific details. Summary of the Invention

[0006] Based on the shortcomings of the existing technology, the present invention provides a fast magnetic resonance multi-sequence joint imaging method, system, device and medium, which solves the problems of cross-modal feature alignment difficulties, weak modeling of correlation between modalities, insufficient decoupling of shared and specific features, and possible introduction of redundant information or loss of sequence-specific details in multimodal MRI joint reconstruction methods.

[0007] The present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a fast magnetic resonance multi-sequence combined imaging method, comprising the following steps:

[0009] Multiple undersampled MR images of different sequences were acquired. The multiple undersampled MR images shared the same common features, while each undersampled MR image had its own corresponding specific features.

[0010] Common features and specific features are used as latent variables. The mapping relationship between multiple undersampled MR images and latent variables is modeled based on the latent variable model. Based on the mapping relationship, the optimization objective of the reconstruction task in fast magnetic resonance imaging is constructed.

[0011] The optimization objective is solved by transposing and convolving multiple undersampled MR images to obtain multiple initial specific features and common features in the high-sampling domain. These initial specific features and common features in the high-sampling domain are then updated to obtain multiple specific features and common features in the low-sampling domain. These low-sampling features and common features are then updated to obtain multiple specific features and common features in the high-sampling domain. Finally, convolutional dictionary operations are performed on these multiple specific features and common features in the high-sampling domain, and combined with the sampled information in the undersampled MR images to obtain multiple currently reconstructed MR images.

[0012] The multiple specific and common features of the current reconstructed MR image in the high-sampling domain are repeatedly updated in both the low-sampling and high-sampling domains to obtain multiple optimal reconstructed MR images.

[0013] Preferably, the optimization objective is as follows:

[0014]

[0015] ;

[0016] In the formula, K The number of sequences. For the first i Undersampled MR images from a sequence A convolution dictionary for specific features in the low-sampling domain. For the first i Specific features of a sequence A convolutional dictionary of common features in the low-sampling domain. C As a public feature, For the first i Reconstructed MR images from a sequence, A convolutional dictionary for specific features in the high-sampling domain. A convolutional dictionary of common features in the high-sampling domain. and For the first i a sequence and C Implicit regularization terms, and They are respectively and C The corresponding implicit regularization coefficients, The square of the F-norm, For convolution operations, For Fourier transform, This is the inverse Fourier transform. The weights of the data consistency items, This is due to undersampling. For Hadama accumulation.

[0017] Preferably, the optimization objective is solved using a deep unfolding network, which includes an initialization module and multiple unfolding modules. Each unfolding module includes a first update module, a second update module, and a data consistency module. Specifically, the initialization module performs transposed convolution on multiple undersampled MR images to obtain multiple initialized specific features and common features in the high-sampling domain. The first update module updates the multiple initialized specific features and common features in the high-sampling domain to obtain multiple specific features and common features in the low-sampling domain. The second update module updates the multiple specific features and common features in the low-sampling domain to obtain multiple specific features and common features in the high-sampling domain. The data consistency module performs convolution dictionary operations on the multiple specific features and common features in the high-sampling domain and combines them with the sampled information in the undersampled images to obtain multiple currently reconstructed MR images. The multiple unfolding modules repeatedly update the multiple specific features and common features of the currently reconstructed MR images in the high-sampling domain under both low-sampling and high-sampling domains to obtain multiple optimal reconstructed MR images.

[0018] Preferably, the first update module updates multiple initial specific features and common features under the high-sampling domain, as shown below:

[0019] ;

[0020] ;

[0021] In the formula, for t Phase 1 i Low-sampling-domain specific features of each modality, For near-end network operators, For updates Iteration step size, For updates Gradient descent iterative unfolding of the network, for t Common features of the low-sampling domain in the phase, For proximal operators, For updates Iteration step size, For updates The gradient descent iterative unfolding network.

[0022] Preferably, the second update module updates the current specific features and common features in the low-sampling domain, as shown below:

[0023] ;

[0024] ;

[0025] In the formula, for t Phase 1 i High-sampling-domain specific features of each modality For updates Gradient descent iterative network expansion, for t Common features of the high-sampling domain in the phase, For updates The gradient descent iterative unfolding network.

[0026] Preferably, the data consistency module performs convolutional dictionary operations on multiple specific features and common features in the high-sampling domain, as shown below:

[0027] ;

[0028] In the formula, for t Phase 1 i Reconstructed MR images of each modality, It is a matrix of all ones. For the corresponding convolutional layers, For the corresponding The convolutional layer.

[0029] Preferably, the multiple undersampled MR images include T1-weighted undersampled images, T2-weighted undersampled images, and liquid attenuation inversion recovery sequence undersampled images.

[0030] Secondly, the present invention provides a fast magnetic resonance multi-sequence combined imaging system, comprising:

[0031] The acquisition module is used to acquire multiple undersampled MR images from different sequences. These multiple undersampled MR images share the same common features, while each undersampled MR image has its own corresponding specific features.

[0032] The mapping module is used to treat common features and specific features as latent variables, model the mapping relationship between multiple undersampled MR images and latent variables based on the latent variable model, and construct the optimization objective for the reconstruction task in fast magnetic resonance multi-sequence joint imaging based on the mapping relationship.

[0033] The solution module is used to solve the optimization objective. During the solution process, multiple undersampled MR images are transposed and convolved to obtain multiple initial specific features and common features in the high-sampling domain; these initial specific features and common features in the high-sampling domain are updated to obtain multiple specific features and common features in the low-sampling domain; these low-sampling features and common features are updated to obtain multiple specific features and common features in the high-sampling domain; convolution dictionary operations are performed on the multiple specific features and common features in the high-sampling domain, and combined with the sampled information in the undersampled MR images to obtain multiple currently reconstructed MR images; and the multiple specific features and common features of the currently reconstructed MR images in the high-sampling domain are repeatedly updated in both the low-sampling and high-sampling domains to obtain multiple optimal reconstructed MR images.

[0034] The iterative module is used to repeatedly update multiple specific and common features of the current reconstructed MR image in the high-sampling domain under both the low-sampling and high-sampling domains, to obtain multiple optimal reconstructed MR images.

[0035] Thirdly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described fast magnetic resonance multi-sequence combined imaging method.

[0036] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described fast magnetic resonance multi-sequence combined imaging method.

[0037] Compared with the prior art, the above-mentioned at least one technical solution adopted by the present invention can achieve the following beneficial effects:

[0038] This invention first models the mapping relationship between multiple undersampled MR images and common and specific features based on a latent variable model. It subdivides the features of multiple undersampled MR images into common features shared across modalities and specific features unique to each modality. This fine decoupling of shared and specific features in MR images solves the problem of weak inter-modal correlation modeling and effectively avoids introducing redundant information or losing sequence-specific details. Based on this, an optimization objective for the reconstruction task in fast magnetic resonance imaging is constructed based on the mapping relationship. The optimization objective is solved using a deep unfolded network. During the solution process, common and specific features are alternately updated in the undersampled and high-sampled domains, alleviating the difficulty of cross-modal feature alignment, and ultimately obtaining multiple reconstructed MR images. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a diagram of the DU-MC-CPAE unfolded network framework of the present invention;

[0041] Figure 2 This is a comparison chart of the reconstruction results of the traditional algorithm (joint TV regularization), the reconstruction results of DU-MC-CPAE, and the fully sampled image under 2D 10x sampling of United Imaging brain data in an embodiment of the present invention.

[0042] in, Figure 2 (a): Schematic diagram of the reconstruction results using the joint TV regularization method. Figure 2 (b): Schematic diagram of DU-MC-CPAE reconstruction results. Figure 2 (c): Fully sampled image;

[0043] Figure 3 This is a visualization of the features of the DU-MC-CPAE during the reconstruction process of the brain image data in 2D 10x sampling, according to an embodiment of the present invention.

[0044] in, Figure 3 (a): The effect of the T1-weighted image. Figure 3 (b): The effect of the T2-weighted image. Figure 3 (c): A rendering of a FLAIR image;

[0045] Figure 4 This is a comparison of the MC-CDic reconstruction results, DU-MC-CPAE reconstruction results, and fully sampled images of Siemens brain data under 1D 8x sampling, as an embodiment of the present invention.

[0046] in, Figure 4 (a): Schematic diagram of MC-CDic reconstruction results. Figure 4 (b): Schematic diagram of DU-MC-CPAE reconstruction results. Figure 4 (c): Fully sampled image. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Background: In single-sequence MRI reconstruction networks, U-Net (Convolutional Networks for Biomedical Image Segmentation) is a popular deep learning architecture that utilizes a fully convolutional network with an encoder-decoder structure, widely used for medical image segmentation and reconstruction tasks. MUSC (Learning Multiscale Convolutional Dictionaries for Image Reconstruction) is a novel approach built on the Convolutional Sparse Coding (CSC) framework to address the limitations of single-scale image reconstruction models. DCT-Net (Domain-Calibrated Translation for Portrait Stylization) employs a bimodal parallel reconstruction framework, combining spatial and spectral information to improve image quality. Furthermore, probabilistic modeling is also an effective method for processing single-sequence MRI reconstruction. Edupuganti et al. introduced a novel framework that incorporates variational autoencoders to integrate uncertainty estimation into the reconstruction process.

[0049] Multi-sequence MRI reconstruction fully utilizes information from different sequences, playing a crucial role in medical diagnosis. Taking the brain as an example, T1-weighted imaging excels at visualizing anatomical structures, clearly distinguishing tissues such as gray matter, white matter, and cerebrospinal fluid; T2-weighted imaging highlights water content, making it ideal for detecting edema, inflammation, and certain pathologies; T2-FLAIR (T2-weighted Fluid Attenuated Inversion Recovery) imaging enhances the visibility of lesions by suppressing signals from cerebrospinal fluid, improving the detection of abnormalities such as multiple sclerosis plaques, tumors, and subarachnoid hemorrhage. Each sequence emphasizes different aspects of tissue characteristics; the core of multi-sequence MRI reconstruction lies in utilizing their correlation and complementarity to achieve better reconstruction results.

[0050] In multi-sequence MRI reference reconstruction, the input includes MR images of a reference sequence and a target sequence. The former is typically a fully sampled image, while the latter is an undersampled image. This framework aims to supplement and reconstruct the target sequence using detailed information from the reference sequence. For example, CDLMRI (Coupled Dictionary Learning for Multi-Contrast MRI Reconstruction) proposes a coupled dictionary representation: both the reference and target images contain a shared common component and a specific component. MC-CDic (Deep Unfolding Convolutional Dictionary Model for Multi-Contrast MRI Super-Resolution and Reconstruction) borrows from the coupled dictionary representation and proposes a deep unfolding network based on convolutional dictionary learning. Meanwhile, MC-VarNet (Decomposition-Based Variational Network for Multi-Contrast MRI Super-Resolution and Reconstruction) proposes a variational model to observe consistent and inconsistent information in different sequence images. DUN-SA (Deep Unfolding Network with Spatial Alignment for Multi-Modal MRI Reconstruction) proposes a novel joint registration and reconstruction model that uses a cross-modal spatial alignment term to compensate for sequence biases. RNECD (Null Space Matters: Range-Null Decomposition for Consistent Multi-Contrast MRI Reconstruction) uses range-null decomposition to decompose features into two components; it improves the ISTA unfolding framework by proposing the Range-Null Empowered Unfolding Network (RUN). Furthermore, the Transformer framework has played a crucial role in modeling interactions between different sequences.

[0051] For joint reconstruction of multi-sequence MRI, the input consists of undersampled MR images from different sequences. For example, jVN (Joint Multi-Contrast Variational Network Reconstruction with Application to Rapid2D and 3D Imaging) integrates multi-sequence data into a variational network framework, allowing the exchange of shared anatomical features between different sequences. Of course, depth unfolding frameworks are also widely used for joint reconstruction; MD-GraphFormer (A Model-Driven Graph Transformer for Fast Multi-Contrast MR Imaging) incorporates the physical constraints of MRI, utilizing graph structure and attention mechanisms to learn complementary information between different sequences.

[0052] Based on the objective of multi-sequence MRI joint reconstruction, this invention first constructs a multi-sequence constrained probabilistic auto-encoder (MC-CPAE) framework from a probabilistic modeling perspective. In the encoder stage, point estimates of latent variables are obtained through maximum a posteriori (MAP) estimation. In the decoder stage, learnable parameters are updated using maximum log-likelihood. The encoder's optimization objective can be combined with the reconstruction objective to form a total optimization objective. This invention employs a semi-quadratic split and proximal gradient descent method to unfold it into a multi-layer deep unfolding network (DU-MC-CPAE), and learns the network parameters in an end-to-end manner. Finally, the DU-MC-CPAE model achieves good reconstruction performance on undersampled brain and knee joint data, and demonstrates a certain degree of interpretability in feature visualization. This invention provides a fast multi-sequence MRI joint imaging method, comprising the following steps:

[0053] S1: Acquire multiple undersampled MR images of different sequences.

[0054] During data acquisition, this invention collects MR data of different sequences. Taking the brain as an example, this involves three different MR sequences: T1-weighted, T2-weighted, and Fluid Attenuation Inversion Recovery (FLAIR). For each sequence image, this invention first uses the ESPIRiT (An Eigenvalue Approach to Autocalibrating Parallel MRI) algorithm to estimate its coil sensitivity map. Then, after processing with a sampling mask in k-space, corresponding T1-weighted undersampled images, T2-weighted undersampled images, and FLAIR undersampled images are generated. Multiple undersampled MR images share the same common features, while each undersampled MR image possesses its own specific features.

[0055] S2: Common features and specific features are used as latent variables. The mapping relationship between multiple undersampled MR images and latent variables is modeled based on the latent variable model. Based on the mapping relationship, the optimization objective of the reconstruction task in fast magnetic resonance imaging is constructed.

[0056] Constructing a probabilistic latent variable model: Referring to coupled dictionary learning, latent variables are distinguished into common features. and special features Different sequence MR images share the same common features, thus modeling the correlation between sequences. However, their specific features are also distinct, thus modeling the differences between sequences. Assuming that the conditional distributions of the latent variables in the images all follow a Gaussian distribution, using the maximum a posteriori probability (MAP) to approximate the latent variables at step E of the expectation-maximization (EM) algorithm as point estimates, a multi-sequence restricted probabilistic autoencoder (MC-CPAE) can be obtained. The optimization objective of the encoder is:

[0057]

[0058] ;

[0059] In the formula, K The number of sequences. For the first i Undersampled MR images of a sequence, A convolution dictionary for specific features in the low-sampling domain. For the first i Specific features of a sequence, A convolutional dictionary of common features in the low-sampling domain. C As a public feature, For the first i Reconstructed MR images of a sequence, A convolutional dictionary for specific features in the high-sampling domain. A convolutional dictionary of common features in the high-sampling domain. and For the first i a sequence and C Implicit regularization terms, and They are respectively and C The corresponding implicit regularization coefficients, The square of the F-norm, This is a convolution operation.

[0060] Construct the overall optimization objective.

[0061] By combining the optimization objective of the encoder in MC-CPAE with the constraints (data consistency term), the overall optimization objective of the reconstruction task is obtained:

[0062]

[0063] ;

[0064] In the formula, For Fourier transform, This is the inverse Fourier transform. The weights of the data consistency items, This is due to undersampling. This is the Hadamard product. The first two terms are dictionary representation losses, which can be viewed as an explicit regularization term based on coupled dictionary learning; the last two terms are for common features. and special features Implicit regularization terms (coefficients represented by a dictionary). This modeling approach combines explicit and implicit regularization, offering both interpretability and flexibility.

[0065] S3: Construct an alternating iterative optimization algorithm.

[0066] Referring to the Half-Quadratic Splitting (HQS) method, auxiliary variables are introduced into the overall optimization objective. and We get the following formula:

[0067]

[0068]

[0069] ;

[0070] In the formula, For measuring auxiliary variables and and C andU Weights for similarity For the first i Specific features of the high-sampling domain in each modality These are common features of the high-sampling domain.

[0071] By decoupling through this HQS approach, common features are made possible. and special features The update can be performed separately in the low-quality domain (low-sampling domain) and the high-quality domain (high-sampling domain), thus splitting the above equation into three sub-problems. For each sub-problem, this invention uses the proximal gradient algorithm to update the variables. The entire optimization process is an alternating iterative format, including three steps:

[0072] (1) Update .

[0073] ;

[0074] ;

[0075] in, for t Phase 1 i Low-sampling-domain specific features of each modality, For proximal operators, For updates Iteration step size, for t -1 Phase i Low-sampling-domain specific features of each modality, For gradient operators, for t -1 Phase i High-sampling-domain specific features of each modality for t Common features of the low-sampling domain in the phase, For updates Iteration step size, for t -1 stage low-sampling domain common features, for t -1 stage high sampling domain common features.

[0076] Dictionary representation of loss for each mode in the low-quality domain:

[0077] ,

[0078] ;

[0079] In the formula, This is a transpose convolution operation.

[0080] The dictionary representation loss represents the loss that is coupled together by subtracting the low-quality domain images of each modality from the specific components:

[0081] ,

[0082] ;

[0083] In the formula, and Features are concatenated after subtracting low-quality domain images and specific components for each modality. ， .

[0084] (2) Update .

[0085] The update process in this step is similar to (1):

[0086] ;

[0087] ;

[0088] In the formula, for t Phase 1 i High-sampling-domain specific features of each modality for t Common characteristics of the high-sampling domain in the phase.

[0089] (3) Data consistency module.

[0090] The subproblem containing the data consistency term has a closed-form solution:

[0091] ;

[0092] In the formula, for t Phase 1 i Reconstructed (high-sampling domain) images of each modality. It is a matrix of all 1s.

[0093] S4: Constructing a multi-sequence MRI joint reconstruction network: Naturally, the convolution dictionary and proximal operators in the iterative process can be approximated by neural networks. Based on this, this invention designs an unfolded network DU-MC-CPAE for the above optimization process. The overall unfolded framework is as follows: Figure 1 As shown.

[0094] Figure 1 middle, x 1. x 2 and x3 shows undersampled MR images under different sequences. , and To and x 1. x 2 and x 3. Update the corresponding transposed convolution dictionary. Gradient descent iterative unfolding network, UGD_ For updates Gradient descent iterative unfolding network, UGD_C is the update Gradient descent iterative unfolding network, UGD_ For updates Gradient descent iterative network expansion, , and To and x 1. x 2 and x The optimal reconstructed image corresponding to 3.

[0095] The DU-MC-CPAE network includes feature initialization and T repeated unrolling stages. In feature initialization, this invention obtains the initialized specific features through a transposed convolutional dictionary (convolutional layer). and public characteristics In the repeated T unfolding phases, each phase contains 3 modules: update ,renew And the data consistency module. They correspond to the three steps of the alternating optimization process in S3, which are expansions of the iterative process, starting with the first... t Taking each stage as an example:

[0096] (1) Update .

[0097] ;

[0098] ;

[0099] In the formula, for t Phase 1 i Low-sampling-domain specific features of each modality, For near-end network operators,

[0100] For updates Gradient descent iterative network expansion, For updates The gradient descent iterative unfolding network.

[0101] UGDThis is a gradient descent iterative unfolding network. Its structure is based on the iterative form of S3, but the convolution dictionary and transposed convolution dictionary are replaced with convolutional layers and transposed convolutional layers:

[0102] ;

[0103] ;

[0104] ;

[0105] ;

[0106] In the formula, For updates The expansion of the gradient term, For the corresponding The transposed convolutional layer, For the corresponding convolutional layers, For the corresponding convolutional layers, For updates The expansion of the gradient term, For the corresponding The transposed convolutional layer, For the corresponding convolutional layers, To convolutional layers, This refers to operations that cascade along the channel dimension.

[0107] ProxNet This is a near-end operator network, structured as a series of residual channel attention modules. Channel attention is integrated into the residual modules to enhance the interaction of information between different sequences within the network. This step enables the updating of specific features in the low-sampling domain. and public characteristics .

[0108] (2) Update .

[0109] ;

[0110] ;

[0111] In the formula, for t Phase 1 i High-sampling-domain specific features of each modality For updates Gradient descent iterative network expansion, for t Common features of the high-sampling domain in the phase, For updates The gradient descent iterative unfolding network.

[0112] In this step, the structures of ProxNet and UGD are similar to (1), except that the undersampled image is... Replace with an image reconstructed from an intermediate layer. This involves replacing the dictionary (convolution) of the low-sampling domain with the dictionary (convolution) of the high-sampling domain. This step allows for the updating of specific features in the low-sampling domain. and public characteristics .

[0113] (3) Data consistency module.

[0114] ;

[0115] In the formula, in the formula, for t Phase 1 i Reconstructed MR images of each modality, It is a matrix of all ones. For the corresponding convolutional layers, For the corresponding The convolutional layer.

[0116] The data consistency module, based on frequency domain consistency constraints, applies the closed-form solution derived in S3. Through this step, in... k undersampled data in space and network reconstruction results Weighting was used to reconstruct the intermediate layer results. Update.

[0117] In the above repetition T After the first expansion phase, the final output of the network is the [number]th [stage]. T The reconstructed results after passing through the data consistency module in each stage provide high-quality data for each sequence. MR image.

[0118] S5: End-to-end training of the DU-MC-CPAE network: This invention implements the DU-MC-CPAE code in PyTorch, using L1 norm error as the loss function for network training. :

[0119] ;

[0120] in, For the output of the DU-MC-CPAE network, For the first i The th mode jFull sampled images of each data point. This represents the parameters of the reconstructed network (specifically including the parameters of convolutional layers, transposed convolutional layers, and proximal operator networks). K and B These represent the number of sequences and the batch size, respectively. This invention uses two NVIDIA RTX4090 graphics processors to train the network, employs the Adam optimizer with a batch size of 2, and a learning rate from... Gradually decrease to The number of iterations was 200,000.

[0121] S6: Applying the trained DU-MC-CPAE network for MRI image reconstruction: Input is... k The undersampled data in the space is output as the reconstructed MR image.

[0122] II. Evidence of the Relevant Effects of the Embodiments. The embodiments of the present invention have achieved some positive effects during research and development or use, and indeed possess significant advantages compared to existing technologies. The following description, combined with data and charts from the experimental process, illustrates these advantages.

[0123] In numerical experiments, the performance metrics of this invention compared to the contrast algorithms were tested on the United Imaging Brain Dataset and the Siemens Brain Dataset. This invention uses the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Normalized Root Mean Square Error (NRMSE) to evaluate the model's performance; higher PSNR and SSIM, and lower NRMSE, indicate better results. In the contrast algorithms, single-sequence reconstruction methods include Zero-Filling and Restormer; multi-sequence reconstruction methods include joint TV regularization, MTrans (Multimodal transformer for accelerated MR imaging), MC-CDic (DeepUnfolding Convolutional Dictionary Model for Multi-Contrast MRI Super-Resolution and Reconstruction), and PromptMR (Prompting for dynamic and multi-contrast MRI reconstruction). Furthermore, in DU-MC-CPAE, by removing the common feature C and its corresponding common dictionary, the method degenerates into a single sequence reconstruction algorithm, which is referred to here as DU-SC-CPAE (Deep Unfolding SingleContrast Constrained Probabilistic Auto Encoder).

[0124] The United Imaging brain data included three sequences: FLAIR, T1, and T2. The training dataset consisted of 7661 data points, and the test dataset consisted of 1860 data points, using a 10×2D Cartesian sampling pattern. The results are shown in Table 1. It can be seen that the DU-MC-CPAE network designed in this invention achieved optimal performance across different sequences. Meanwhile, in... Figure 2 In the visualization results, compared with the results of joint TV regularization, the DU-MC-CPAE results also have clearer structure and texture, and less artifacts and noise. Furthermore, the reconstruction process of DU-MC-CPAE has a certain degree of interpretability, allowing visualization of common and specific components between different sequences, such as... Figure 3 As shown in the figure, the left side of the equation is the reconstructed image, and the right side of the equation is the undersampled image, common features, and specific features.

[0125] This is enlightening for understanding the correlations and differences between sequences.

[0126] The Siemens brain dataset also includes three sequences: FLAIR, T1, and T2. The training dataset contains 2020 sequences, and the test dataset contains 324 sequences. It uses an 8×1D Cartesian sampling pattern with a center sampling rate of 8% and high-frequency random selection. Because the dimensions of different sequences in the original data vary, this invention aligns the dimensions by padding with zeros in the image domain and K-space. The performance indicators are shown in Table 2, and the visualization results are as follows: Figure 4 As shown, the method of the present invention achieves the best reconstruction accuracy and provides clearer detail information.

[0127] Table 1. Comparison of different methods of 2D 10x sampling in Simian Imaging brain data

[0128]

[0129] Table 2 Comparison of different 1D 8x sampling methods in Siemens brain data

[0130]

[0131] This invention uses dictionary representation to facilitate interaction of multimodal image features in the feature domain, which alleviates the difficulty of cross-modal feature alignment to some extent. This invention also solves the problem of weak intermodal correlation modeling by subdividing features under dictionary representation into common features shared between modalities and unique features unique to each modality. Furthermore, this invention alleviates the problem of a large number of model parameters to some extent by expanding the network framework and sharing dictionary parameters at different stages.

[0132] Based on the same concept, the present invention also provides a fast magnetic resonance multi-sequence joint imaging system, including an acquisition module, a mapping module and a solution module.

[0133] The acquisition module is used to acquire multiple undersampled MR images from different sequences. These multiple undersampled MR images share the same common features, while each undersampled MR image has its own corresponding specific features.

[0134] The mapping module is used to treat common and specific features as latent variables, model the mapping relationship between multiple undersampled MR images and latent variables based on the latent variable model, and construct the optimization objective for the reconstruction task in fast magnetic resonance multi-sequence joint imaging based on the mapping relationship.

[0135] The solution module is used to solve the optimization objective. During the solution process, multiple undersampled MR images are transposed and convolved to obtain multiple initial specific features and common features in the high-sampling domain; the multiple initial specific features and common features in the high-sampling domain are updated to obtain multiple specific features and common features in the low-sampling domain; the multiple specific features and common features in the low-sampling domain are updated to obtain multiple specific features and common features in the high-sampling domain; the multiple specific features and common features in the high-sampling domain are convolved with a dictionary operation and combined with the sampled information in the undersampled MR images to obtain multiple currently reconstructed MR images.

[0136] The iterative module is used to repeatedly update multiple specific and common features of the current reconstructed MR image in the high-sampling domain under both the low-sampling and high-sampling domains, to obtain multiple optimal reconstructed MR images.

[0137] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described fast magnetic resonance multi-sequence combined imaging method.

[0138] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described fast magnetic resonance multi-sequence combined imaging method.

[0139] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0140] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A fast magnetic resonance multi-sequence joint imaging method, characterized in that, Includes the following steps: Multiple undersampled MR images of different sequences were acquired. The multiple undersampled MR images shared the same common features, while each undersampled MR image had its own corresponding specific features. Common features and specific features are used as latent variables. The mapping relationship between multiple undersampled MR images and latent variables is modeled based on the latent variable model. Based on the mapping relationship, the optimization objective of the reconstruction task in fast magnetic resonance imaging is constructed. The optimization objective is solved by transposing and convolving multiple undersampled MR images to obtain multiple initial specific features and common features in the high-sampling domain. These initial specific features and common features in the high-sampling domain are then updated to obtain multiple specific features and common features in the low-sampling domain. These low-sampling features and common features are then updated to obtain multiple specific features and common features in the high-sampling domain. Finally, convolutional dictionary operations are performed on these multiple specific features and common features in the high-sampling domain, and combined with the sampled information in the undersampled MR images to obtain multiple currently reconstructed MR images. The multiple specific and common features of the current reconstructed MR image in the high-sampling domain are repeatedly updated in both the low-sampling and high-sampling domains to obtain multiple optimal reconstructed MR images; The specific optimization objectives are as follows: ; In the formula, K The number of sequences. For the first i Undersampled MR images from a sequence A convolution dictionary for specific features in the low-sampling domain. For the first i Specific features of a sequence A convolutional dictionary of common features in the low-sampling domain. C As a public feature, For the first i Reconstructed MR images from a sequence, A convolutional dictionary for specific features in the high-sampling domain. A convolutional dictionary of common features in the high-sampling domain. and For the first i a sequence and C Implicit regularization terms, and They are respectively and C The corresponding implicit regularization coefficients, The square of the F-norm, For convolution operations, For Fourier transform, For inverse Fourier transform, The weights of the data consistency items, This is due to undersampling. For Hadamah accumulation; The optimization objective is solved using a deep unfolding network, which includes an initialization module and multiple unfolding modules. Each unfolding module includes a first update module, a second update module, and a data consistency module. Specifically, the initialization module performs transposed convolution on multiple undersampled MR images to obtain multiple initialized specific features and common features in the high-sampling domain. The first update module updates these initialized specific features and common features in the high-sampling domain to obtain multiple specific features and common features in the low-sampling domain. The second update module updates these specific features and common features in the low-sampling domain to obtain multiple specific features and common features in the high-sampling domain. The data consistency module performs convolution dictionary operations on the multiple specific features and common features in the high-sampling domain and combines them with the sampled information in the undersampled images to obtain multiple currently reconstructed MR images. The multiple unfolding modules repeatedly update the multiple specific features and common features of the currently reconstructed MR images in the high-sampling domain under both low-sampling and high-sampling domains to obtain multiple optimal reconstructed MR images. The first update module updates multiple initial specific features and common features in the high-sampling domain, as shown below: ; ; In the formula, for t Phase 1 i Low-sampling-domain specific features of each modality, For near-end network operators, For updates Iteration step size, For updates Gradient descent iterative network expansion, for t Common features of the low-sampling domain in the phase, For proximal operators, For updates Iteration step size, For updates Gradient descent iterative unfolding network; The second update module updates the current specific features and common features in the low-sampling domain, as shown below: ; ; In the formula, for t Phase 1 i High-sampling-domain specific features of each modality For updates Gradient descent iterative network expansion, for t Common features of the high-sampling domain in the phase, For updates The gradient descent iterative unfolding network.

2. The fast magnetic resonance multi-sequence joint imaging method as described in claim 1, characterized in that, The data consistency module performs convolutional dictionary operations on multiple specific and common features in the high-sampling domain, as shown below: ; In the formula, for t Phase 1 i Reconstructed MR images of each modality, It is a matrix of all ones. For the corresponding Convolutional layers, For the corresponding The convolutional layer.

3. The fast magnetic resonance multi-sequence joint imaging method as described in claim 1, characterized in that, Multiple undersampled MR images include T1-weighted undersampled images, T2-weighted undersampled images, and liquid attenuation inversion recovery sequence undersampled images.

4. A joint imaging system based on the fast magnetic resonance multi-sequence joint imaging method of claim 1, characterized in that, include: The acquisition module is used to acquire multiple undersampled MR images from different sequences. These multiple undersampled MR images share the same common features, while each undersampled MR image has its own corresponding specific features. The mapping module is used to treat common features and specific features as latent variables, model the mapping relationship between multiple undersampled MR images and latent variables based on the latent variable model, and construct the optimization objective for the reconstruction task in fast magnetic resonance multi-sequence joint imaging based on the mapping relationship. The solution module is used to solve the optimization objective. During the solution process, multiple undersampled MR images are transposed and convolved to obtain multiple initial specific features and common features in the high-sampling domain; these initial specific features and common features in the high-sampling domain are updated to obtain multiple specific features and common features in the low-sampling domain; these low-sampling features and common features are updated to obtain multiple specific features and common features in the high-sampling domain; convolution dictionary operations are performed on the multiple specific features and common features in the high-sampling domain, and combined with the sampled information in the undersampled MR images to obtain multiple currently reconstructed MR images. The iterative module is used to repeatedly update multiple specific and common features of the current reconstructed MR image in the high-sampling domain under both the low-sampling and high-sampling domains, to obtain multiple optimal reconstructed MR images.

5. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the fast magnetic resonance multi-sequence combined imaging method according to any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the fast magnetic resonance multi-sequence combined imaging method according to any one of claims 1-3.

Images