Low-field multi-contrast rapid MRI (Magnetic Resonance Imaging) joint reconstruction method for sensing

Abstract

The invention discloses a low-field multi-contrast rapid MRI (Magnetic Resonance Imaging) joint reconstruction method for sensing an anatomical structure. The method is specifically implemented according to the following steps: step 1, constructing a low-field multi-contrast magnetic resonance image joint reconstruction model for sensing the anatomical structure; 2, constructing a low-field multi-contrast magnetic resonance image joint reconstruction network sensed by the anatomical structure according to the reconstruction model constructed in the step 1; step 3, training a low-field multi-contrast magnetic resonance image joint reconstruction network perceived by the anatomical structure; and 4, performing magnetic resonance imaging by applying the trained anatomical structure perceived low-field multi-contrast magnetic resonance image joint reconstruction network. The method provided by the invention solves the problem of insufficient imaging optimization of the focus part in the existing reconstruction method.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical magnetic resonance imaging, and in particular relates to a low-field multi-contrast rapid MRI joint reconstruction method for anatomical structure perception. Background Art As a technology with lower cost and lighter equipment, low-field magnetic resonance imaging (Low-field MRI) has important applications in medical scenarios with limited resources. Compared with high-field MRI, low-field MRI has the advantages of less impact on patients and more convenient equipment, but it also has problems with low signal-to-noise ratio (SNR) and poor image resolution. This makes low-field MRI lag behind high-field MRI in image quality, especially in the demand for multi-contrast combined imaging in clinical diagnosis.

[0002] Multi-contrast magnetic resonance imaging (MC-MRI) can provide images with different contrasts such as T1-weighted (T1w), T2-weighted (T2w) and fluid-attenuated inversion recovery (FLAIR) under the same anatomical structure. These images can provide complementary information and help improve the accuracy of diagnosis. However, in low-field MRI, due to the limitations of signal-to-noise ratio and resolution, the effect of multi-contrast imaging is not ideal, which further limits its clinical application. Existing methods such as compressed sensing (CS) and parallel imaging have been widely used in high-field MRI by undersampling K-space data to reduce scanning time and reconstructing complete images through algorithms. However, due to the low signal-to-noise ratio of low-field MRI, these methods cannot be directly applied to low-field environments. In recent years, deep learning methods have provided a new approach for magnetic resonance image reconstruction. Data-driven networks such as U-Net and ResNet can effectively improve the quality of image reconstruction by learning image features through a large amount of training data. However, these methods lack close integration with actual physical models, especially in low-field MRI, where the complexity of deep network structures and inference time become factors that need to be considered. To address these challenges, model-driven deep learning methods have gradually emerged, combining classical reconstruction algorithms with deep learning to achieve high-quality image reconstruction through iterative optimization. Typical models include ADMM-Net, VN-Net, and ISTA-Net, which have made significant progress in high-field MRI imaging. They provide better interpretability and performance by combining physical models with data-driven learning mechanisms, but they are mainly used in single-contrast imaging methods. In clinical applications, a set of multi-contrast images is usually required to provide more comprehensive diagnostic information. Based on this, some researchers have proposed multi-contrast joint reconstruction methods in recent years. These methods use the correlation between cross-contrast images to improve reconstruction performance. For example, DenseNet and MM-GAN enhance the reconstruction effect of target contrast images by using high-quality guide images, and Hi-Net integrates the information of multi-contrast images through deep learning networks to obtain high-quality MR images. Although the above-mentioned magnetic resonance imaging technologies have made significant progress in improving global image quality, the existing reconstruction algorithms are mainly focused on improving the overall image quality, but the imaging optimization of the lesion site is insufficient, which limits their application in clinical diagnosis and treatment. Summary of the invention

[0003] The purpose of the present invention is to provide an anatomical structure-aware low-field multi-contrast rapid MRI joint reconstruction method to solve the problem of insufficient imaging optimization of lesion sites in existing reconstruction methods.

[0004] The technical solution adopted by the present invention is an anatomical structure-aware low-field multi-contrast rapid MRI joint reconstruction method, which is specifically implemented according to the following steps: Step 1, constructing an anatomical structure-aware low-field multi-contrast MRI joint reconstruction model; Step 2, constructing an anatomical structure-aware low-field multi-contrast magnetic resonance image joint reconstruction network based on the reconstruction model constructed in step 1; Step 3, training an anatomically aware low-field multi-contrast MRI joint reconstruction network; Step 4: Use the trained anatomically-aware low-field multi-contrast MRI joint reconstruction network to perform MRI. The present invention is also characterized in that: The specific process of step 1 is: Given K contrast undersampled k-space data , the goal is to obtain k-space data The corresponding high-quality multi-contrast image is reconstructed ,in represents the number of sampled k-space data, Indicates the number of pixels of the sampled image; The anatomical structure-aware low-field multi-contrast MRI joint reconstruction model is: (1) In formula (1), represents a sparse transform; represents the segmentation network, which is used to accurately locate the target area of ​​interest; represents the denoising network; and Represents the weight parameter.

[0005] The specific process of step 2 is: Step 2.1, solve the anatomical structure-aware low-field multi-contrast MRI joint reconstruction model constructed in step 1 by an alternating minimization algorithm, and obtain: (9) Step 2.2, Formula (9) is derived and expanded into a cascaded multi-stage deep network as an anatomical structure-aware low-field multi-contrast MRI joint reconstruction network , rebuild the network Including denoising network , anatomical perception data consistency module and the group sparsity module GS.

[0006] In step 2.2, the denoising network It consists of seven convolution modules connected in sequence and a common convolution with a convolution kernel of 1×1. The output of the first convolution module and the output of the sixth convolution module are fused features through the concatenation operation, the output of the second convolution module and the output of the fifth convolution module are fused features through the concatenation operation, and the output of the third convolution module and the output of the fourth convolution module are fused features through the concatenation operation; the number of channels of the seven convolution modules are 32, 64, 128, 256, 128, 64, and 32 respectively. The first convolution module, the second convolution module, the third convolution module, the fifth convolution module, the sixth convolution module, and the seventh convolution module are all composed of two common layers CoRe, the fourth convolution module is composed of one common layer CoRe, and each common layer is composed of a convolution kernel size of , composed of ordinary convolution with a step size of 1, BN layer and Relu activation function; The denoising network is expressed as: (10).

[0007] In step 2.2, the anatomical perception data consistency module It includes a data consistency module DC and an anatomical perception module TM; Among them, the data consistency module DC is expressed as: (11) The anatomical perception module TM is expressed as: (12) In the formula, It is a pre-trained segmentation network; Segmentation Network It consists of eight convolution modules connected in sequence, a common convolution with a convolution kernel of 1×1, and a sigmoid function. The output of the first convolution module and the output of the sixth convolution module are fused through the concatenation operation, the output of the second convolution module and the output of the fifth convolution module are fused through the concatenation operation, and the output of the third convolution module and the output of the fourth convolution module are fused through the concatenation operation. The denoising network The output of the sparse transformation The obtained sparse features are fused with the output of the seventh convolution module through the Concatenation operation; the number of channels of the eight convolution modules are 128, 256, 512, 1024, 512, 256, 128, and 64 respectively. The eight convolution modules are composed of two common layers CoRe, and each common layer consists of a convolution kernel size of , composed of ordinary convolution with a step size of 1, BN layer and Relu activation function; Dissection perception data consistency module The final expression of is: (13).

[0008] In step 2.2, the group sparse module GS is expressed as: (14) The group sparse module GS first undergoes a learnable nonlinear transformation Obtaining high-level feature representation , followed by a learnable threshold parameter The soft threshold operator , that is, , to capture the group sparsity between different contrasts in the high-level representation space, and then after the left inverse operation The MR image can be reconstructed from the sparse representation space. and The architecture is completely symmetrical and must satisfy ,in Represents the identity operator.

[0009] The specific process of step 3 is: Step 3.1: Build a training dataset A training dataset is constructed based on the low-field MRI dataset M4Raw. Each set of data in the training dataset contains undersampled k-space data with different contrasts. And the corresponding fully sampled magnetic resonance image ,in Represents the number of contrasts. During the data construction process, the full-sampled k-space data of different contrasts in the low-field MRI dataset M4Raw are undersampled to obtain the corresponding k-space undersampled data. As a reference reconstructed image, As a reconstruction network Input; Step 3.2: Use the training data set constructed in step 3.1 to train the anatomical structure-aware low-field multi-contrast MRI joint reconstruction network. ; The loss function during training is: (15) In the formula, This is to ensure that the reconstruction network is finally output in the T stage. And the corresponding fully sampled reference image Stay consistent. Indicates the number of contrasts in magnetic resonance images; To ensure ; This is to ensure the consistency of the reconstructed image and the fully sampled image in terms of anatomical structure; Represents the general dice loss, which aims to segment the network Facilitates accurate segmentation of specific objects of interest (TOIs) in the reconstructed network, Represents the segmentation network The segmentation probability map output at the Tth stage; are three tuning parameters used to balance the contribution of different loss terms.

[0010] The beneficial effects of the present invention are as follows: the anatomical structure-aware low-field multi-contrast fast MRI joint reconstruction method of the present invention can efficiently reconstruct high-quality multi-contrast images from under-sampled k-space multi-contrast data; and can simultaneously optimize the sampling mode and can adaptively reconstruct the required high-quality multi-contrast images for specific areas according to clinical needs and downstream task requirements, which can provide high-quality medical impact data for doctors to perform clinical diagnosis in the later stage, thereby improving diagnostic accuracy and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 A flowchart of the anatomical structure-aware low-field multi-contrast fast MRI joint reconstruction method of the present invention; Figure 2 A framework diagram of the anatomical structure-aware low-field multi-contrast magnetic resonance image joint reconstruction network constructed for the present invention; Figure 3 This is a diagram showing an example of T1 contrast magnetic resonance image reconstruction in an embodiment of the present invention; Figure 4 This is a diagram showing an example of T2 contrast magnetic resonance image reconstruction in an embodiment of the present invention; Figure 5 This is an example diagram of FLAIR contrast magnetic resonance image reconstruction in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0013] Example 1 The present invention provides an anatomical structure-aware low-field multi-contrast rapid MRI joint reconstruction method, such as Figure 1 As shown, the specific implementation steps are as follows: Step 1, constructing an anatomical structure-aware low-field multi-contrast MRI joint reconstruction model; The specific process is: Given K contrast undersampled k-space data , the goal is to obtain k-space data The corresponding high-quality multi-contrast image is reconstructed ,in represents the number of sampled k-space data, Indicates the number of pixels of the sampled image; The anatomical structure-aware low-field multi-contrast MRI joint reconstruction model is: (1) In formula (1), represents a sparse transform; represents the segmentation network, which is used to accurately locate the TOI (target region of interest); represents a denoising network, which further refines and denoises the TOI area to improve the imaging quality of the reconstructed image TOI; and Represents the weight parameter, which is used to balance the contribution of the regularization term and the data consistency term to the model; Formula (1) consists of three parts: (1) Data assurance item: It indicates the fidelity or consistency between the reconstructed multi-contrast image and the actual under-sampled K-space measurement data, ensuring that the reconstruction result is highly consistent with the original K-space data, thereby ensuring the accuracy of image reconstruction under the condition of under-sampled data; (2) is the group sparse regularization term: The norm performs group sparse constraints on multi-contrast images, uses the spatial similarity between contrast images to fuse information, and extracts shared features; here, the convolutional network is used to learn the sparse transformation , to improve the expressive power of sparse representation; this method extracts the shared features of multi-contrast images in a high-level semantic space, further enhancing the correlation between each contrast; (3) Anatomy-aware regularization: This is done by combining learnable anatomical prior information to enhance the imaging of the target region of interest (TOI); specifically, a segmentation network is introduced. , which is used to accurately locate TOI and provide key specific anatomical structure location information for the denoising process of reconstructed images; then through the denoising network The TOI region is further refined and denoised to improve the imaging quality of the TOI region of the reconstructed image.

[0014] In order to better capture the complementary information between images with different contrasts under the guidance of specific clinical needs, the present invention combines the reconstruction process with a learnable K-space sampling mode to improve imaging efficiency and quality.

[0015] Step 2, constructing an anatomical structure-aware low-field multi-contrast magnetic resonance image joint reconstruction network based on the reconstruction model constructed in step 1; The specific process is: Step 2.1, solving the anatomical structure-aware low-field multi-contrast MRI joint reconstruction model constructed in step 1 by an alternating minimization algorithm; The specific process is: Step 2.1.1: For the anatomical structure-aware regularization term, use the Taylor expansion formula to denoise the network. Approximately, it can be expressed as: (2) In the formula, Indicates about The Jacobian matrix of is the change in each iteration, that is ; Will Recorded as , then the anatomical structure-aware regularization term is rewritten as follows: (3) Further approximation:

[0016] (4) In the formula, Represents the segmentation network about The Jacobian matrix of ; when When , the anatomical structure-aware regularization term can be approximated as: (5) Step 2.1.2, Alternating Optimization and , the expression is: (6) And for the second sub-problem in formula (6) The proximal gradient descent method is used for solution; The specific process is: Step 2.1.2.1, obtain the intermediate updated reconstruction result by gradient descent method , Indicates The intermediate reconstruction result of each contrast is updated as follows for each contrast image: (7) In the formula, represents the learnable step size, For the The undersampled k-space data of the contrast, Indicates A contrast-based learnable forward operator,

[0017] In the formula, Indicates Coil sensitivity diagram corresponding to each receiving coil; represents the one-dimensional discrete Fourier transform of the transverse direction; Represents the one-dimensional discrete Fourier transform in the vertical direction;

[0018]

[0019] in, represents a set of learnable position parameters, ,in , ; Indicates the number of K-space samples in the phase encoding direction in the under-sampling scenario; represents the number of K-space samples in the phase encoding direction in the full sampling scene; the acceleration factor is defined as ; The present invention only considers one-dimensional imaging scenes, i.e. fixed ,make , during the reconstruction process, only the position parameters of the phase encoding direction need to be learned ; Step 2.1.2.2, after gradient update, use the soft threshold algorithm to estimate the reconstruction result of the current stage , the update formula is: (8) In the formula, express The matrix List, represents a learnable sparse transformation, express The left inverse transform of Step 2.1.3 summarizes the entire iterative optimization algorithm as follows: (9) Step 2.2, Formula (9) is derived and expanded into a cascaded multi-stage deep network as an anatomical structure-aware low-field multi-contrast MRI joint reconstruction network ,like Figure 2 As shown, the reconstructed network Including denoising network , anatomical perception data consistency module and group sparse module GS; (1) Denoising network It is a structure based on convolutional neural network, which consists of seven convolution modules connected in sequence and a common convolution with a convolution kernel of 1×1. The output of the first convolution module and the output of the sixth convolution module are fused features through the concatenation operation, the output of the second convolution module and the output of the fifth convolution module are fused features through the concatenation operation, and the output of the third convolution module and the output of the fourth convolution module are fused features through the concatenation operation; the number of channels of the seven convolution modules are 32, 64, 128, 256, 128, 64, and 32 respectively. The first convolution module, the second convolution module, the third convolution module, the fifth convolution module, the sixth convolution module, and the seventh convolution module are all composed of two common layers CoRe, and the fourth convolution module is composed of one common layer CoRe. Each common layer is composed of a convolution kernel size of , a normal convolution with a step size of 1, a BN layer and a Relu activation function; this denoising network can effectively extract features and retain the spatial information of the image, ensuring that important anatomical details are not lost during the denoising process; Denoising Network Used to reconstruct the results of the previous stage Remove noise and artifacts; According to formula (1) and formula (5), the denoising network is expressed as: (10) (2) Anatomical perception data consistency module It includes a data consistency module DC and an anatomical perception module TM; The function of the data consistency module DC is to use the observed k-space data to force the reconstruction result to be consistent with it, ensuring that the reconstructed image does not deviate from the actual observed k-space data. The expression is: (11) The role of the anatomical perception module TM is to quantify the reconstruction results of the previous stage Compared with the current denoising result The difference between , especially in the anatomical region of interest, is expressed as: (12) In the formula, is a pre-trained segmentation network. During the reconstruction network training process, some parameters are frozen to generate a probability map of the target of interest (TOI) to help locate the anatomical structure of interest. By correcting the difference in the TOI, it is ensured that the reconstructed image can capture the anatomical structure information and improve the anatomical accuracy of the reconstruction result.

[0020] The segmentation network It is a simple U-Net network, which consists of eight convolution modules connected in sequence, a common convolution with a convolution kernel of 1×1 and a sigmoid function. The output of the first convolution module and the output of the sixth convolution module are fused through the concatenation operation, the output of the second convolution module and the output of the fifth convolution module are fused through the concatenation operation, and the output of the third convolution module and the output of the fourth convolution module are fused through the concatenation operation. Denoising network The output of the sparse transformation The obtained sparse features are fused with the output of the seventh convolution module through the Concatenation operation; the number of channels of the eight convolution modules are 128, 256, 512, 1024, 512, 256, 128, and 64 respectively. The eight convolution modules are composed of two common layers CoRe, and each common layer consists of a convolution kernel size of , composed of ordinary convolution with a step size of 1, BN layer and Relu activation function; Dissection perception data consistency module The final expression of is: (13) Formula (13) combines data consistency and anatomical perception to ensure that the reconstructed image can accurately capture the anatomically interesting region and improve the reconstruction quality of the region while maintaining the consistency of the observed data. The purpose of module design is to combine the denoising results obtained in the reconstruction process, the reconstruction results of the previous stage, and the undersampled k-space observation data to ensure the comprehensive optimization of the reconstructed image in terms of data consistency and anatomical perception; (3) Reconstruction result of the group sparse module GS in the middle Apply based on The group sparsity of the norm is used to further enhance the MRI image reconstruction effect, especially the anatomical structure. Then the expression of the group sparse module GS is: (14) The group sparse module GS first undergoes a learnable nonlinear transformation Obtaining high-level feature representation , followed by a learnable threshold parameter The soft threshold operator , that is, , to capture the group sparsity between different contrasts in the high-level representation space, and then after the left inverse operation The MR image can be reconstructed from the sparse representation space. and The architecture is completely symmetrical and must satisfy ,in represents the identity operator; The present invention forces sparse transformation With segmentation network A weight sharing strategy is implemented between the top-level codes, which effectively establishes group sparsity constraints in a meaningful high-level semantic space, promotes the fusion of complementary information between various contrasts, and thus improves the imaging quality of specific anatomical structures. Step 3, training an anatomically aware low-field multi-contrast MRI joint reconstruction network; The specific process is: Step 3.1: Build a training dataset A training dataset is constructed based on the low-field MRI dataset M4Raw. Each set of data in the training dataset contains undersampled k-space data with different contrasts. And the corresponding fully sampled magnetic resonance image ,in Represents the number of contrasts. During the data construction process, the full-sampled k-space data of different contrasts in the low-field MRI dataset M4Raw are undersampled to obtain the corresponding k-space undersampled data. As a reference reconstructed image, As a reconstruction network Input; The low-field MRI dataset M4Raw includes fully sampled k-space data of three contrasts: T1, T2, and FLAIR, and the corresponding fully sampled images; Step 3.2: Use the training data set constructed in step 3.1 to train the anatomical structure-aware low-field multi-contrast MRI joint reconstruction network. ; Rebuilding the network The parameters to be learned can be summarized as ; The loss function during training is: (15) In the formula, This is to ensure that the reconstruction network is finally output in the T stage. And the corresponding fully sampled reference image Stay consistent. Indicates the number of contrasts in magnetic resonance images; To ensure ; This is to ensure the consistency of the reconstructed image and the fully sampled image in terms of anatomical structure; Represents the general dice loss, which aims to segment the network Facilitates accurate segmentation of specific objects of interest (TOIs) in the reconstructed network, Represents the segmentation network The segmentation probability map output at the Tth stage; are three tuning parameters used to balance the contribution of different loss terms; The training of the reconstruction network is carried out on a computing device based on the pytorch framework, equipped with the Ubuntu 20.04 operating system and the RTX 3090TiGPU. The reconstruction network is set to a deep network containing 6 cascaded modules, i.e., T=6. The parameters of each module are initialized by the Xavier random initialization strategy to ensure the reasonable distribution of the initial parameters, so as to improve the stability and convergence speed of the network training. In order to further optimize the training effect, the Adam optimizer is used for parameter training in the present invention, which has good convergence. During the training process, the iteration cycle is set to 100 times, 4 small batches are used to process data in each iteration, and the learning rate is set to 1×10 -4 , to achieve progressive optimization of the network; In terms of model input, the present invention converts the initial input image into The image is set to zero padding to ensure that the initial input does not carry any prior information interference. For the three weighting parameters α, β and γ involved in formula (11), the present invention sets them to 0.01, 0.1 and 1 respectively through a large number of experiments and tuning processes. These parameter values ​​play a key role in weighing different loss terms and network training objectives, and can effectively improve the accuracy and quality of the reconstructed image, while enhancing the stability and robustness of the model.

[0021] Step 4, using the trained anatomically-aware low-field multi-contrast MRI joint reconstruction network to perform MRI; Through the above training process, the present invention can obtain the optimal parameters of the anatomical structure-aware multi-contrast magnetic resonance reconstruction network. Based on the trained network, multiple contrast low-field k-space data are input to generate corresponding high-quality multi-contrast magnetic resonance images. Since the parameters of the network have been fully optimized and adjusted, even at a lower sampling rate, the network can still generate reconstruction results that are almost consistent with standard full-sampling nuclear magnetic resonance images. This network can effectively compensate for the information loss caused by undersampling, thereby still providing high-quality image reconstruction effects under fast imaging. This method effectively improves the image quality under low sampling conditions, thereby accelerating the process of magnetic resonance imaging and improving its practicality in clinical applications.

[0022] Example 2 In order to verify the effectiveness of the present invention, the numerical experiments used 158 human brain magnetic resonance images with three contrasts, T1, T2 and FLAIR, to train and test the network. For each individual, the size of each contrast image is . The present invention undersamples the 2D slice images of T1, T2 and FLAIR contrast at different sampling rates according to the 1D Cartesian sampling mode, thereby obtaining 158 sets of brain magnetic resonance imaging data required for the experiment. Among them, 128 sets of data are selected for training, and 30 sets of data are used to evaluate the performance of the trained network. This embodiment samples the k-space data at a sampling rate of 1 / 8 to conduct experiments and evaluations on the reconstruction network.

[0023] Example 3 The difference from Example 2 is that the k-space data is sampled at a sampling rate of 1 / 10 to perform experiments and evaluations on the reconstruction network.

[0024] In order to quantitatively evaluate the reconstruction performance, the present invention uses two commonly used image quality evaluation indicators, namely, peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). In addition to evaluating the overall PSNR and SSIM metrics of the entire image, specific TOI regions are also specifically evaluated. To this end, the present invention also measures the PSNR and SSIM values ​​of specific TOI regions, expressed as TOI-PSNR and TOI-SSIM, respectively, in order to more accurately evaluate the performance of the reconstruction network in the lesion area or key anatomical parts. Specifically, for the M4Raw dataset, TOI evaluation was performed in the subcortical region of the brain.

[0025] Through the evaluation of these TOI-specific regions, the superiority of the reconstruction network in anatomical structure perception and local detail restoration can be verified in more detail, especially in key lesion sites or anatomical structures.

[0026] Tables 1 and 2 show the performance comparison of the anatomical structure-aware low-field multi-contrast MRI joint reconstruction method of the present invention with a variety of current mainstream single-contrast and multi-contrast reconstruction methods: (1) The single-contrast method specifically includes the traditional Grappa as a baseline, the pure data-driven network U-Net, the model-unfolded deep network ISTA-Net, and the multi-task network IDMHV that combines segmentation tasks to improve image reconstruction performance. (2) The multi-contrast method includes the pure data-driven MTrans, which introduces an improved transformer to capture the complementary information between different contrasts, and the unfolded model MC-J-MoDL, which simultaneously optimizes the sampling mode and the reconstruction network to improve performance. Compared with other methods, the reconstruction network A²MRI proposed in the present invention shows the best reconstruction accuracy under different sampling modes and sampling rates, especially in complex structures and lesion areas. Even at low sampling rates, A²MRI still maintains high PSNR and SSIM, significantly outperforming other methods. Figure 3 , Figure 4 and Figure 5 The visualization results show that A²MRI can reconstruct clear details without artifacts in the lesion area, which is superior to other methods in detail preservation and artifact suppression, especially in the anatomical structure area of ​​interest.

[0027] Table 1: Comparison of different methods on the brain data test set at 1 / 10 sampling rate of 1D Cartesian (TOI: subcortical area)

[0028] Table 2: Comparison of different methods on the brain data test set at 1D Cartesian 1 / 8 sampling rate (TOI: subcortical area) .

Claims

1. An anatomically structure-aware low-field multi-contrast fast MRI joint reconstruction method, characterized in that: Follow the steps below to implement it: Step 1, constructing an anatomical structure-aware low-field multi-contrast MRI joint reconstruction model; Step 2, constructing an anatomical structure-aware low-field multi-contrast magnetic resonance image joint reconstruction network based on the reconstruction model constructed in step 1; Step 3, training an anatomically aware low-field multi-contrast MRI joint reconstruction network; Step 4: Use the trained anatomically-aware low-field multi-contrast MRI joint reconstruction network to perform MRI.

2. The anatomical structure-aware low-field multi-contrast fast MRI joint reconstruction method according to claim 1, characterized in that: The specific process of step 1 is: Given K contrast undersampled k-space data , the goal is to obtain k-space data The corresponding high-quality multi-contrast image is reconstructed ,in represents the number of sampled k-space data, Indicates the number of pixels of the sampled image; The anatomical structure-aware low-field multi-contrast MRI joint reconstruction model is: （1） In formula (1), represents a sparse transform; represents the segmentation network, which is used to accurately locate the target area of ​​interest; represents the denoising network; and Represents the weight parameter.

3. The anatomical structure-aware low-field multi-contrast fast MRI joint reconstruction method according to claim 2, characterized in that: The specific process of step 2 is: Step 2.1, solve the anatomical structure-aware low-field multi-contrast MRI joint reconstruction model constructed in step 1 by an alternating minimization algorithm, and obtain: （9） Step 2.2, Formula (9) is derived and expanded into a cascaded multi-stage deep network as an anatomical structure-aware low-field multi-contrast MRI joint reconstruction network , rebuild the network Including denoising network , anatomical perception data consistency module and the group sparsity module GS.

4. The anatomical structure-aware low-field multi-contrast fast MRI joint reconstruction method according to claim 2, characterized in that: In step 2.2, the denoising network It consists of seven convolution modules connected in sequence and a common convolution with a convolution kernel of 1×1. The output of the first convolution module and the output of the sixth convolution module are fused features through the concatenation operation, the output of the second convolution module and the output of the fifth convolution module are fused features through the concatenation operation, and the output of the third convolution module and the output of the fourth convolution module are fused features through the concatenation operation; the number of channels of the seven convolution modules are 32, 64, 128, 256, 128, 64, and 32 respectively. The first convolution module, the second convolution module, the third convolution module, the fifth convolution module, the sixth convolution module, and the seventh convolution module are all composed of two common layers CoRe, the fourth convolution module is composed of one common layer CoRe, and each common layer is composed of a convolution kernel size of , composed of ordinary convolution with a step size of 1, BN layer and Relu activation function; The denoising network is expressed as: (10).

5. The anatomical structure-aware low-field multi-contrast fast MRI joint reconstruction method according to claim 2, characterized in that: In step 2.2, the anatomical perception data consistency module It includes a data consistency module DC and an anatomical perception module TM; Among them, the data consistency module DC is expressed as: （11） The anatomical perception module TM is expressed as: （12） In the formula, It is a pre-trained segmentation network; Segmentation Network It consists of eight convolution modules connected in sequence, a common convolution with a convolution kernel of 1×1, and a sigmoid function. The output of the first convolution module and the output of the sixth convolution module are fused through the concatenation operation, the output of the second convolution module and the output of the fifth convolution module are fused through the concatenation operation, and the output of the third convolution module and the output of the fourth convolution module are fused through the concatenation operation. The denoising network The output of the sparse transformation The obtained sparse features are fused with the output of the seventh convolution module through the Concatenation operation; the number of channels of the eight convolution modules are 128, 256, 512, 1024, 512, 256, 128, and 64 respectively. The eight convolution modules are composed of two common layers CoRe, and each common layer consists of a convolution kernel size of , composed of ordinary convolution with a step size of 1, BN layer and Relu activation function; Dissection perception data consistency module The final expression of is: （13）。 6. The anatomical structure-aware low-field multi-contrast fast MRI joint reconstruction method according to claim 2, characterized in that: In step 2.2, the group sparse module GS is expressed as: （14） The group sparse module GS first undergoes a learnable nonlinear transformation Obtaining high-level feature representation , followed by a learnable threshold parameter The soft threshold operator , that is, , to capture the group sparsity between different contrasts in the high-level representation space, and then after the left inverse operation The MR image can be reconstructed from the sparse representation space. and The architecture is completely symmetrical and must satisfy ,in Represents the identity operator.

7. The anatomical structure-aware low-field multi-contrast fast MRI joint reconstruction method according to claim 2, characterized in that: The specific process of step 3 is: Step 3.1: Build a training dataset A training dataset is constructed based on the low-field MRI dataset M4Raw. Each set of data in the training dataset contains undersampled k-space data with different contrasts. And the corresponding fully sampled magnetic resonance image ,in Represents the number of contrasts. During the data construction process, the full-sampled k-space data of different contrasts in the low-field MRI dataset M4Raw are undersampled to obtain the corresponding k-space undersampled data. As a reference reconstructed image, As a reconstruction network Input; Step 3.2: Use the training data set constructed in step 3.1 to train the anatomical structure-aware low-field multi-contrast MRI joint reconstruction network. ; The loss function during training is: （15） In the formula, This is to ensure that the reconstruction network is finally output in the T stage. And the corresponding fully sampled reference image Stay consistent. Indicates the number of contrasts in magnetic resonance images; To ensure ; This is to ensure the consistency of the reconstructed image and the fully sampled image in terms of anatomical structure; Represents the general dice loss, which aims to segment the network Facilitates accurate segmentation of specific objects of interest (TOIs) in the reconstructed network, Represents the segmentation network The segmentation probability map output at the Tth stage; are three tuning parameters used to balance the contribution of different loss terms.

Images