Rapid magnetic resonance multi-sequence joint imaging method, system, equipment and medium

Abstract

The invention discloses a fast magnetic resonance multi-sequence joint imaging method, system and device and a medium, and relates to the technical field of deep learning, and the method comprises the following steps: collecting a plurality of under-sampling MR images of different sequences, the plurality of under-sampling MR images sharing the same common feature, and each under-sampling MR image enjoying the corresponding specific feature; modeling a mapping relation between the plurality of undersampled MR images and hidden variables based on a hidden variable model, and constructing an optimization target of a reconstruction task in fast magnetic resonance multi-sequence joint imaging based on the mapping relation; and solving the optimization target through the deep expansion network. According to the method, features of a plurality of undersampled MR images are subdivided into common features shared among modals and specific features unique to the modals, and the sharing of the MR images and the specific features are finely decoupled, so that the problem that correlation modeling among the modals is relatively weak is solved, and introduction of redundant information or loss of sequence specific detail information can be effectively avoided.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method, system, device and medium for rapid magnetic resonance multi-sequence joint imaging. Background Art

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

[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) first performing undersampling (only acquiring part of the data points in the k-space, such as 10% or 15%, etc.), and then using prior knowledge of the image to reconstruct MRI.

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

[0005] However, the existing multi-sequence MRI joint reconstruction method still has some defects, mainly including: the complex anatomical deformation between multi-sequence images, such as motion artifacts or non-rigid deformation, causes difficulty in cross-modality feature alignment; the modeling of inter-modality correlation is weak, the decoupling of shared and specific features is not fine enough, which may introduce redundant information or lose sequence-specific detail information. SUMMARY

[0006] Based on the defects of the existing technology, the present application provides a fast magnetic resonance multi-sequence joint imaging method, system, device and medium, which solves the problems of cross-modality feature alignment difficulty and weak inter-modality correlation modeling in the existing multi-modality MRI joint reconstruction method, and the decoupling of shared and specific features is not fine enough, which may introduce redundant information or lose sequence-specific detail information.

[0007] The present application adopts the following technical solutions: In a first aspect, the present application provides a fast magnetic resonance multi-sequence joint imaging method, comprising the following steps: Collecting multiple under-sampling MR images of different sequences, wherein the multiple under-sampling MR images share the same common features, and each under-sampling MR image has its own corresponding specific features; Taking the common features and specific features as hidden variables, modeling the mapping relationship between the multiple under-sampling MR images and the hidden variables based on the hidden variable model, and constructing an optimization objective of the reconstruction task in fast magnetic resonance imaging based on the mapping relationship; Solving the optimization objective, in the solving process, transposed convolution is performed on the multiple under-sampling MR images to obtain multiple initialized specific features and common features in a high sampling domain; the multiple initialized specific features and common features in the high sampling domain are updated to obtain multiple specific features and common features in a 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; convolution dictionary operation is performed on the multiple specific features and common features in the high sampling domain, and combined with the sampled information in the under-sampling MR image, multiple current reconstruction MR images are obtained; Repeating the updating of the multiple specific features and common features of the current reconstruction MR image in the high sampling domain and in the low sampling domain and the high sampling domain to obtain multiple optimal reconstruction MR images.

[0008] Preferably, the optimization objective is specifically as follows: ; In the formula, K is the number of sequences, is the first iundersampled MR images under the sequence, a convolution dictionary for specific features under a low sampling domain, a convolution dictionary for specific features under a low sampling domain, i specific features under the sequence, a convolution dictionary for common features under a low sampling domain, C common features, a convolution dictionary for specific features under a low sampling domain, i reconstructed MR images under the sequence, a convolution dictionary for specific features under a high sampling domain, a convolution dictionary for common features under a high sampling domain, and a regularization term for the sequence and i a regularization term for the sequence and C corresponding regularization term coefficients, and corresponding regularization term coefficients, and C corresponding regularization term coefficients, a square of F-norm, a convolution operation, a Fourier transform, an inverse Fourier transform, a weight of a data consistency term, undersampling, a Hadamard product.

[0009] Preferably, the optimization objective is solved by a deep unfolding network, the deep unfolding network comprising an initialization module and a plurality of unfolding modules, each unfolding module comprising a first updating module, a second updating module and a data consistency module; wherein the initialization module transposes and convolves the plurality of undersampled MR images to obtain a plurality of initialization specific features and common features under a high sampling domain; the first updating module updates the plurality of initialization specific features and common features under the high sampling domain to obtain a plurality of specific features and common features under a low sampling domain; the second updating module updates the plurality of specific features and common features under the low sampling domain to obtain a plurality of specific features and common features under the high sampling domain; the data consistency module performs a convolution dictionary operation on the plurality of specific features and common features under the high sampling domain, and combines the sampled information in the undersampled images to obtain a plurality of current reconstructed MR images; the plurality of unfolding modules repeatedly update the plurality of specific features and common features under the low sampling domain and the high sampling domain for the plurality of current reconstructed MR images under the high sampling domain to obtain a plurality of optimal reconstructed MR images.

[0010] Preferably, the first updating module updates the plurality of initialization specific features and common features under the high sampling domain, and the specific updating process is as follows: ; ; Where, for t Phase I i The low-sampling domain-specific features of the modalities, is the proximal network operator, For update The iteration step size, For update The gradient descent iterative expansion network, for t Stage low-sampling domain common features, is the proximal operator, For update The iteration step size, For update The gradient descent iteratively unfolds the network.

[0011] Preferably, the second updating module updates the current specific features and common features in the low sampling domain, as follows: ; ; Where, for t Phase I i Highly sampled domain-specific features of the modalities, For update The gradient descent iterative expansion network, for t The common features of the high-sampling domain in the stage, For update The gradient descent iteratively unfolds the network.

[0012] Preferably, the data consistency module performs a convolution dictionary operation on multiple specific features and common features in the high sampling domain, as shown below: ; Where, for t Phase I i The reconstructed MR images of the modalities, is a matrix of all 1s, To correspond The convolutional layer, To correspond The convolutional layer.

[0013] Preferably, the multiple under-sampled MR images include T1-weighted under-sampled images, T2-weighted under-sampled images, and fluid-attenuated inversion recovery sequence under-sampled images.

[0014] In a second aspect, the present application provides a fast magnetic resonance multi-sequence joint imaging system, comprising: a collection module configured to collect a plurality of under-sampled MR images of different sequences, wherein the plurality of under-sampled MR images share a same common feature, and each under-sampled MR image has a corresponding specific feature; a mapping module configured to model a mapping relationship between the plurality of under-sampled MR images and hidden variables based on a hidden variable model, wherein the common feature and the specific feature are used as the hidden variables, and to construct an optimization objective of a reconstruction task in the fast magnetic resonance multi-sequence joint imaging based on the mapping relationship; a solving module configured to solve the optimization objective, wherein in the solving process, transpose convolution is performed on the plurality of under-sampled MR images to obtain a plurality of initialized specific features and common features in a high sampling domain; the plurality of initialized specific features and common features in the high sampling domain are updated to obtain a plurality of specific features and common features in a low sampling domain; the plurality of specific features and common features in the low sampling domain are updated to obtain a plurality of specific features and common features in the high sampling domain; convolution dictionary operation is performed on the plurality of specific features and common features in the high sampling domain, and combined with the sampled information in the under-sampled MR images to obtain a plurality of current reconstructed MR images; and the plurality of specific features and common features of the current reconstructed MR images in the high sampling domain are repeatedly updated in the low sampling domain and the high sampling domain to obtain a plurality of optimal reconstructed MR images; an iteration module configured to repeatedly update the plurality of specific features and common features of the current reconstructed MR images in the high sampling domain in the low sampling domain and the high sampling domain to obtain the plurality of optimal reconstructed MR images.

[0015] In a third aspect, the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the fast magnetic resonance multi-sequence joint imaging method described above when executing the program.

[0016] In a fourth aspect, the present application provides a computer readable storage medium, wherein the storage medium stores a computer program, and the computer program is executable on a processor to implement the fast magnetic resonance multi-sequence joint imaging method described above.

[0017] Compared with the prior art, the above at least one technical solution adopted by the present application can achieve the following beneficial effects: The application firstly models the mapping relationship between multiple undersampled MR images and common features and specific features based on a latent variable model, subdivides the features of the multiple undersampled MR images into common features shared among modalities and specific features unique to modalities, finely decouples the shared and specific features of the MR images, solves the problem of weak modeling of inter-modality correlation, and can effectively avoid introducing redundant information or losing sequence-specific detailed information. On this basis, an optimization objective of a reconstruction task in fast magnetic resonance imaging is constructed based on the mapping relationship. The optimization objective is solved through a deep unfolding network, and the common features and the specific features are alternately updated in the low sampling domain and the high sampling domain in the solving process, which alleviates the problem of difficult alignment of cross-modality features, and finally multiple reconstructed magnetic resonance images are obtained. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0019] Figure 1 The DU-MC-CPAE unfolding network framework of the present application; Figure 2 The comparative diagram of the reconstruction results of the traditional algorithm (joint TV regular) and the DU-MC-CPAE and the fully sampled image of the brain data of the present application embodiment under 2D 10x sampling; Wherein, Figure 2 (a) of the (a): the reconstruction result diagram of the joint TV regular method, Figure 2 (b) of the (b): the DU-MC-CPAE reconstruction result diagram, Figure 2 (c) of the (c): the fully sampled image; Figure 3 The feature visualization effect diagram of the DU-MC-CPAE in the reconstruction process of the brain data of the present application embodiment under 2D 10x sampling; Wherein, Figure 3 (a) of the (a): the T1 weighted image effect diagram, Figure 3 (b) of the (b): the T2 weighted image effect diagram, Figure 3 (c) of the (c): the FLAIR image effect diagram; Figure 4 The comparative diagram of the MC-CDic reconstruction result, the DU-MC-CPAE reconstruction result and the fully sampled image of the Siemens brain data of the present application embodiment under 1D 8x sampling; Wherein, Figure 4Fig. 2 (a): MC-CDic reconstruction results. Figure 4 Fig. 2 (b): DU-MC-CPAE reconstruction results. Figure 4 Fig. 2 (c): Full-sampling images. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0021] Related background: In single sequence MRI reconstruction networks, U-Net (Convolutional Networks for Biomedical Image Segmentation) is a popular deep learning architecture that uses a fully convolutional network with an encoder-decoder structure and is widely used for segmentation and reconstruction tasks in medical images. The MUSC (Learning Multiscale Convolutional Dictionaries for Image Reconstruction) method is a new method based on the convolutional sparse coding (CSC, Convolutional Sparse Coding) framework to address the limitations of single-scale image reconstruction models. In DCT-Net (Domain-Calibrated Translation for Portrait Stylization), a bimodal parallel reconstruction framework is designed to improve image quality by combining spatial and spectral information. In addition, probabilistic modeling is also an effective method for single sequence MRI reconstruction. Edupuganti et al. introduced a new framework that combines a variational autoencoder to integrate uncertainty estimation into the reconstruction process.

[0022] Multi-sequence MRI reconstruction takes full advantage of the information of different sequences, and plays an important role in medical diagnosis. For example, in brain imaging, T1-weighted imaging is good at visualizing anatomical structures and can clearly distinguish tissues such as gray matter, white matter and cerebrospinal fluid; T2-weighted imaging highlights water content, making it an ideal choice 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, and the core of multi-sequence MRI reconstruction is to utilize the correlation and complementarity between them to achieve better reconstruction results.

[0023] In multi-sequence MRI reference reconstruction, the input includes MR images of reference sequences and MR images of target sequences, the former is usually fully sampled images, while the latter is undersampled images, the framework hopes to supplement and reconstruct the target sequence with the detailed information in the reference sequence. For example, CDL MRI (Coupled Dictionary Learning for Multi-Contrast MRI Reconstruction) proposes a coupled dictionary representation: the reference image and the target image both contain a shared common component and a specific component. MC-CDic (Deep Unfolding Convolutional Dictionary Model for Multi-Contrast MRI Super-Resolution and Reconstruction) learns from the coupled dictionary representation and proposes a deep unfolding network based on convolutional dictionary learning. In MC-VarNet (Decomposition-Based Variational Network for Multi-Contrast MRI Super-Resolution and Reconstruction), a variational model is proposed 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 new joint registration reconstruction model, which uses a cross-modality spatial alignment term to compensate for the deviation between sequences. RNECD (Null Space Matters: Range-Null Decomposition for Consistent Multi-Contrast MRI Reconstruction) uses range-null decomposition to decompose the features into two components, which improves the ISTA unfolding framework and proposes a range-null empowered unfolding network (RUN, Range-Null Empowered Unfolding Network). In addition, the Transformer framework also plays an important role in modeling the interaction between different sequences.

[0024] For multi-contrast MRI joint reconstruction, the input is under-sampled MR images of different sequences. For example, jVN (Joint Multi-Contrast Variational Network Reconstruction with Application to Rapid 2D and 3D Imaging) integrates multi-contrast data into a variational network framework, allowing the exchange of shared anatomical features between different sequences. Of course, the deep unfolding framework is also widely used for joint reconstruction. MD-GraphFormer (A Model-Driven Graph Transformer for Fast Multi-Contrast MR Imaging) combines the physical constraints of MRI and uses a graph structure and attention mechanism to learn the complementary information between different sequences.

[0025] Based on the task target of multi-contrast MRI joint reconstruction, the present application first constructs a multi-contrast constrained probabilistic auto-encoder (MC-CPAE, Muti-Contrast Constrained Probabilistic Auto-Encoder) framework from the perspective of probabilistic modeling. In the encoder stage, the point estimate of the latent variable is obtained by maximum a posteriori (MAP, Maximum A Posteriori) estimation. In the decoder stage, the learnable parameters are updated by the maximum log-likelihood method. The optimization objective of the encoder can be combined with the reconstruction objective as the total optimization objective. The present application uses semi-quadratic splitting and proximal gradient descent method to expand it into a multi-layer deep unfolding network (DU-MC-CPAE, Deep Unfolding Muti-Contrast Constrained Probabilistic Auto-Encoder), and learns the network parameters in an end-to-end manner. Finally, on the under-sampled brain and knee data, the DU-MC-CPAE model achieves good reconstruction performance and shows a certain degree of interpretability in feature visualization. The present application provides a fast multi-contrast MRI joint imaging method, comprising the following steps:

[0026] S1: Collect multiple under-sampled MR images of different sequences.

[0027] During data acquisition, the present invention collects MR data from different sequences. For the brain, for example, three different MR sequences are used: T1-weighted, T2-weighted, and fluid-attenuated inversion recovery (FLAIR) . For each sequence, the present invention first estimates the coil sensitivity map using the ESPIRiT (An Eigenvalue Approach to Autocalibrating Parallel MRI) algorithm. Then, after k-space sampling mask processing, the corresponding T1-weighted undersampled image, T2-weighted undersampled image, and fluid-attenuated inversion recovery sequence undersampled image are generated. Multiple undersampled MR images share the same common features, while each undersampled MR image possesses its own specific features.

[0028] 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, and the optimization objective of the reconstruction task in fast magnetic resonance imaging is constructed based on the mapping relationship.

[0029] Constructing a probabilistic latent variable model: referencing coupled dictionary learning to distinguish latent variables into common features and unique characteristics Different MR image sequences share the same common features, which can be used to model the correlation between sequences. However, their specific features are separated from each other, which can be used to model the differences between sequences. Assuming that the conditional distribution of the image with respect to the latent variable obeys a Gaussian distribution, the maximum a posteriori probability (MAP) is used to approximate the latent variable of the E-step of the Expectation-Maximization (EM) algorithm as a point estimate, and a multiple sequence constrained probabilistic autoencoder (MC-CPAE) is obtained. The optimization objective of the encoder is:

[0030] ; Where, K is the number of sequences, For the i A series of undersampled MR images, is the convolution dictionary of the specific features in the low sampling domain, For the i The specific characteristics of the sequence, is the convolution dictionary of common features in the low-sampling domain, C For public features, For the i A series of reconstructed MR images, is the convolution dictionary of the specific features in the high sampling domain, is the convolution dictionary of common features in the high-sampling domain, and are the implicit regularization terms of the i th sequence and C th sequence respectively, and are the corresponding implicit regularization term coefficients, and C are the corresponding implicit regularization term coefficients, is the square of F-norm, is the convolution operation.

[0031] The total optimization objective is constructed.

[0032] The optimization objective of the encoder in MC-CPAE is combined with the constraint condition (data consistency term) to obtain the overall optimization objective of the reconstruction task: ; where, is the Fourier transform, is the inverse Fourier transform, is the weight of the data consistency term, is the under-sampling, is the Hadamard product. The first two terms are the dictionary representation loss, which can be regarded as an explicit regularization term based on coupled dictionary learning; the last two terms are implicit regularization terms for the common feature and the specific feature (dictionary representation coefficients). This modeling method combines explicit regularization and implicit regularization, which has certain interpretability and flexibility.

[0033] S3: Construct an alternating iterative optimization algorithm.

[0034] Referring to the Half-Quadratic Splitting (HQS) method, auxiliary variables and are introduced into the total optimization objective to obtain the following formula: ; where, is the weight for measuring the closeness of the auxiliary variables and to C and U , is the specific feature of the high sampling domain under the i th modality, is the common feature of the high sampling domain.

[0035] By decoupling in this HQS way, the common features and unique characteristics The update can be performed in the low-quality domain (low-sampling domain) and the high-quality domain (high-sampling domain), that is, the above equation is divided into three sub-problems. For each sub-problem, the present invention uses the proximal gradient algorithm to update the variables. The entire optimization process is an alternating iterative format, including three steps:

[0036] (1) Update .

[0037] ; ; in, for t Phase I i The low-sampling domain-specific features of the modalities, is the proximal operator, For update The iteration step size, for t -1st stage i The low-sampling domain-specific features of the modalities, is the gradient operator, for t -1st stage i Highly sampled domain-specific features of the modalities, for t Stage low-sampling domain common features, For update The iteration step size, for t -1 stage low-sampling domain common features, for t -1 stage high sampling domain common features.

[0038] The dictionary representation loss for each modality in the low-quality domain is: , ; Where, is the transposed convolution operation.

[0039] The dictionary representation loss is the difference between the low-quality domain images of each modality and the unique components coupled together: , ; Where, and The features of the cascade after subtraction of the low-quality domain images of each modality and the specific components, ， .

[0040] (2) Update .

[0041] The update process of this step is similar to (1): ; ; wherein, is the high-sampling domain specific feature of the m-th modality in the l-th stage, t is the high-sampling domain common feature in the l-th stage. i t (3) Data consistency module.

[0042] The sub-problem containing the data consistency term has a closed-form solution:

[0043] wherein, ; wherein, is the reconstructed (high-sampling domain) image of the m-th modality in the l-th stage, t is an all-one matrix. i S4: Constructing a multi-sequence MRI joint reconstruction network: Naturally, the convolution dictionary and the proximal operator in the iterative process can be approximated by a neural network, based on which, the present application designs an unfolding network DU-MC-CPAE for the above optimization process, and the overall unfolding framework is shown in .

[0044] Figure 1 wherein,

[0045] 1, Figure 1 2 and x 3 are undersampled MR images under different sequences, x , x and are the transpose convolution dictionaries corresponding to 1, 2 and x 3, the unfolding network of the gradient descent iterative formula for updating x is UGD_1, x the unfolding network of the gradient descent iterative formula for updating is UGD_C, and the unfolding network of the gradient descent iterative formula for updating is UGD_2. the unfolding network of the gradient descent iterative formula for updating is UGD_1, the unfolding network of the gradient descent iterative formula for updating is UGD_C, and the unfolding network of the gradient descent iterative formula for updating ,​​ and For x 1. x 2 and x 3 corresponds to the optimal reconstructed image.

[0046] In the DU-MC-CPAE network, there are T stages of feature initialization and repetition. In feature initialization, the present invention obtains the initialized specific features by transposing the convolution dictionary (convolution layer) and common features In the repeated T expansion phases, each phase contains 3 modules: Update ,renew and data consistency modules. They correspond to the three steps of the alternating optimization process in S3, which are the expansion of the iterative process. t Take this stage as an example:

[0047] (1) Update .

[0048] ; ; Where, for t Phase I i The low-sampling domain-specific features of the modalities, is the proximal network operator, For update The gradient descent iterative expansion network, For update The gradient descent iteratively unfolds the network.

[0049] UGD It is a gradient descent iterative unfolding network. Its structure is based on the iterative form of S3, replacing the convolution dictionary and transposed convolution dictionary with convolution layers and transposed convolution layers: ; ; ; ; Where, For update The expanded form of the gradient term is, To correspond The transposed convolutional layer, To correspond The convolutional layer, To correspond The convolutional layer, For update The expanded form of the gradient term is, To correspond The transposed convolutional layer, To correspond The convolutional layer, For The convolutional layer, An operation that concatenates along the channel dimension.

[0050] ProxNet It is a proximal operator network, whose structure is a series connection of multiple residual channel attention modules, that is, channel attention is integrated into the residual module to enhance the interaction of information between different sequences in the network. Through this step, it is possible to update specific features in the low-sampling domain. and common features .

[0051] (2) Update .

[0052] ; ; Where, for t Phase I i Highly sampled domain-specific features of the modalities, For update The gradient descent iterative expansion network, for t The common features of the high-sampling domain in the stage, For update The gradient descent iteratively unfolds the network.

[0053] The structures of ProxNet and UGD in this step are similar to (1), except that the undersampled images are Replace the reconstructed image with the middle layer , replace the dictionary (convolution) of the low-sampling domain with the dictionary (convolution) of the high-sampling domain. This step can be used to update the specific features in the low-sampling domain. and common features .

[0054] (3) Data consistency module.

[0055] ; In the formula, in the formula, for t Phase I i The reconstructed MR images of the modalities, is a matrix of all 1s, To correspond The convolutional layer, To correspond The convolutional layer.

[0056] The data consistency module applies the closed-form solution derived in S3 based on the frequency domain consistency constraint. k Undersampled data in space and network reconstruction results Weighted, to achieve the reconstruction result of the middle layer Updates.

[0057] In the above repetition T After the expansion stage, the final output of the network is T The results of reconstruction after the data consistency module in each stage are high-quality MR image.

[0058] S5: DU-MC-CPAE network end-to-end training: This paper implements the DU-MC-CPAE code in PyTorch and uses L1 norm error as the loss function for network training. : ; in, is the output of the DU-MC-CPAE network, For the i Under the mode j Fully sampled image of the data, Represents the parameters of the reconstruction network (specifically including the parameters of the convolutional layer, transposed convolutional layer and proximal operator network), K and B The present invention uses two NVIDIA RTX4090 graphics processors to train the network, adopts the Adam optimizer with a batch size of 2, and the learning rate is from Gradually decrease to , the number of iterations is 200,000 times.

[0059] S6: Apply the trained DU-MC-CPAE network to MRI image reconstruction: the input is k The output of the spatial undersampled data is the reconstructed MR image.

[0060] 2. Evidence of the effects of the embodiments: The embodiments of the present invention have achieved some positive effects during the development or use process, and indeed have great advantages over the existing technology. The following content describes them with the data and charts of the experimental process.

[0061] In numerical experiments, the present application tests its performance indicators with contrast algorithms on UnionBrain brain dataset and Siemens brain dataset. The present application uses the average peak signal-to-noise ratio (PSNR), the structural similarity index (SSIM) and the normalized root mean square error (NRMSE) to evaluate the performance of the model, the higher the PSNR and SSIM, the lower the NRMSE, the better the result. In the contrast algorithm, the single sequence reconstruction method includes: Zero-Filling and Restormer; the multi-sequence reconstruction method includes: joint TV regularization, MTrans (Multimodal transformer for accelerated MR imaging), MC-CDic (Deep Unfolding Convolutional Dictionary Model for Multi-Contrast MRI Super-Resolution and Reconstruction) and PromptMR (Prompting for dynamic and multi-contrast mri reconstruction). In addition, in the DU-MC-CPAE, the common feature C and its corresponding common dictionary are removed, and the method is degenerated into a single sequence reconstruction algorithm, which is referred to as DU-SC-CPAE (Deep Unfolding Single Contrast Constrained Probabilistic Auto Encoder) here.

[0062] UnionBrain brain data includes FLAIR, T1 and T2 three sequences, the training data amount is 7661, the test data amount is 1860, and 10x2D Cartesian sampling mode is used. The results are shown in Table 1, it can be seen that the DU-MC-CPAE network designed by the present application reaches the optimal under different sequences. At the same time in the visualization results of Figure 2 compared with the results of joint TV regularization, the results of DU-MC-CPAE also have clearer structure and texture, and less artifacts and noise. In addition, the reconstruction process of DU-MC-CPAE has certain interpretability, and the common components and specific components between different sequences can be visualized, as shown in Figure 3 the figure, the left side of the equation is the reconstructed image, and the right side of the equation is the undersampling image, the common feature and the specific feature.

[0063] This has an enlightening effect on understanding the correlation and difference between sequences.

[0064] The Siemens brain dataset also includes three sequences of FLAIR, T1 and T2, the training data amount is 2020, the test data amount is 324, an 8x1D Cartesian sampling mode is used, the center sampling rate is 8%, and the high frequency is randomly selected. Figure 4 As shown in Table 2 and the visualization results as shown in

[0065] Table 1 Comparison results of different methods 2D 10x sampling on Union brain data Table 2 Comparison results of different methods 1D 8x sampling on Siemens brain data The present application solves the problem of weak correlation modeling between modalities by subdividing the features under the dictionary representation into public features shared between modalities and specific features unique to modalities.

[0066] Based on the same concept, the present application also provides a fast magnetic resonance multi-sequence joint imaging system, which comprises an acquisition module, a mapping module and a solving module.

[0067] The acquisition module is used for acquiring multiple undersampled MR images of different sequences, wherein the multiple undersampled MR images share the same public features, and each undersampled MR image has its own corresponding specific features.

[0068] The mapping module is used for modeling the mapping relationship between the multiple undersampled MR images and the hidden variables based on the hidden variable model, and constructing the optimization objective of the reconstruction task in the fast magnetic resonance multi-sequence joint imaging based on the mapping relationship.

[0069] The solution module is used to solve the optimization target. During the solution process, a plurality of undersampled MR images are transposed and convolved to obtain a plurality of initialized specific features and common features in the high sampling domain; the plurality of initialized specific features and common features in the high sampling domain are updated to obtain a plurality of specific features and common features in the low sampling domain; the plurality of specific features and common features in the low sampling domain are updated to obtain a plurality of specific features and common features in the high sampling domain; a convolution dictionary operation is performed on the plurality of specific features and common features in the high sampling domain, and the features are combined with the sampled information in the undersampled MR images to obtain a plurality of current reconstructed MR images.

[0070] The iterative module is used to repeatedly update multiple specific features and common features of the current reconstructed MR image in the high sampling domain in the low sampling domain and the high sampling domain to obtain multiple optimal reconstructed MR images.

[0071] The present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the above-mentioned rapid magnetic resonance multi-sequence joint imaging method is implemented.

[0072] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned rapid magnetic resonance multi-sequence joint imaging method is implemented.

[0073] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0074] Obviously, those skilled in the art may make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if such modifications and variations fall within the scope of the claims and their equivalents, the present invention is intended to include such modifications and variations.

Claims

1. A rapid magnetic resonance multi-sequence combined imaging method, characterized in that: The following steps are involved: Acquire multiple undersampled MR images of different sequences, wherein the multiple undersampled MR images share the same common features and each undersampled MR image has its own corresponding specific features; The 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. The optimization objective of the reconstruction task in fast magnetic resonance imaging is constructed based on the mapping relationship. Solving the optimization objective, during the solving process, performing transposed convolution on multiple undersampled MR images to obtain multiple initialized specific features and common features in a high-sampling domain; updating the multiple initialized specific features and common features in the high-sampling domain to obtain multiple specific features and common features in a low-sampling domain; updating the multiple specific features and common features in the low-sampling domain to obtain multiple specific features and common features in a high-sampling domain; performing a convolution dictionary operation on the multiple specific features and common features in the high-sampling domain, and combining them with sampled information in the undersampled MR images to obtain multiple current reconstructed MR images; The multiple specific features and common features of the current reconstructed MR image in the high sampling domain are repeatedly updated in the low sampling domain and the high sampling domain to obtain multiple optimal reconstructed MR images.

2. The rapid magnetic resonance multi-sequence combined imaging method according to claim 1, characterized in that: The optimization objectives are specifically as follows: ; Where, K is the number of sequences, For the i Undersampled MR images in the sequence, is the convolution dictionary of the specific features in the low sampling domain, For the i The unique features of the sequence, is the convolution dictionary of common features in the low-sampling domain, C For public features, For the i The reconstructed MR images under the sequence, is the convolution dictionary of the specific features in the high sampling domain, is the convolution dictionary of common features in the high-sampling domain, and For the i Sequences and C The implicit regularization term of and They are and C The corresponding implicit regularization coefficient is, is the square of the F-norm, is the convolution operation, is the Fourier transform, is the inverse Fourier transform, is the weight of the data consistency item, is under-sampling, It is Hadamard.

3. The rapid magnetic resonance multi-sequence combined imaging method according to claim 2, characterized in that: The optimization objective is solved by a deep unfolding network, which includes an initialization module and multiple unfolding modules, each of which includes a first update module, a second update module and a data consistency module; wherein 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 image to obtain multiple current reconstructed MR images; the multiple unfolding modules repeatedly update the multiple specific features and common features of the current reconstructed MR image in the high sampling domain in the low sampling domain and the high sampling domain to obtain multiple optimal reconstructed MR images.

4. The rapid magnetic resonance multi-sequence combined imaging method according to claim 3, wherein: The first updating module updates multiple initialized specific features and common features in the high sampling domain, as shown below: ; ; Where, for t Phase i The low-sampling domain-specific features of the modalities, is the proximal network operator, For update The iteration step size, For update The gradient descent iterative expansion network, for t Stage low-sampling domain common features, is the proximal operator, For update The iteration step size, For update The gradient descent iteratively unfolds the network.

5. The rapid magnetic resonance multi-sequence combined imaging method according to claim 4, characterized in that: The second updating module updates the current specific features and common features in the low sampling domain, as shown below: ; ; Where, for t Phase i Highly sampled domain-specific features of the modalities, For update The gradient descent iterative expansion network, for t The common features of the high-sampling domain in the stage, For update The gradient descent iteratively unfolds the network.

6. The rapid magnetic resonance multi-sequence combined imaging method according to claim 5, characterized in that: The data consistency module performs convolution dictionary operations on multiple specific features and common features in the high sampling domain, as shown below: ; Where, for t Phase i The reconstructed MR images of the modalities, is a matrix of all 1s, To correspond The convolutional layer, To correspond The convolutional layer.

7. The rapid magnetic resonance multi-sequence combined imaging method according to claim 1, wherein: The multiple undersampled MR images include T1-weighted undersampled images, T2-weighted undersampled images, and fluid-attenuated inversion recovery sequence undersampled images.

8. A rapid magnetic resonance multi-sequence combined imaging system, characterized in that: include: An acquisition module, configured to acquire a plurality of undersampled MR images of different sequences, wherein the plurality of undersampled MR images share the same common features and each undersampled MR image has its own corresponding specific features; A mapping module is used to use 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 target of the reconstruction task in fast MRI multi-sequence joint imaging based on the mapping relationship; A solution module is used to solve the optimization target. During the solution process, a plurality of undersampled MR images are transposed and convolved to obtain a plurality of initialized specific features and common features in a high-sampling domain; the plurality of initialized specific features and common features in the high-sampling domain are updated to obtain a plurality of specific features and common features in a low-sampling domain; the plurality of specific features and common features in the low-sampling domain are updated to obtain a plurality of specific features and common features in a high-sampling domain; a convolution dictionary operation is performed on the plurality of specific features and common features in the high-sampling domain, and the features are combined with the sampled information in the undersampled MR images to obtain a plurality of current reconstructed MR images; The iterative module is used to repeatedly update multiple specific features and common features of the current reconstructed MR image in the high sampling domain in the low sampling domain and the high sampling domain to obtain multiple optimal reconstructed MR images.

9. A computer device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for rapid magnetic resonance multi-sequence joint imaging according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by a processor, the method for rapid magnetic resonance multi-sequence joint imaging according to any one of claims 1 to 7 is implemented.

Images