Anatomy-aware low-field multi-contrast fast MRI joint reconstruction method

Abstract

The application discloses an anatomic structure perception low-field multi-contrast fast MRI joint reconstruction method, and is specifically implemented according to the following steps: step 1, constructing an anatomic structure perception low-field multi-contrast magnetic resonance image joint reconstruction model; step 2, constructing an anatomic structure perception low-field multi-contrast magnetic resonance image joint reconstruction network according to the reconstruction model constructed in step 1; step 3, training the anatomic structure perception low-field multi-contrast magnetic resonance image joint reconstruction network; and step 4, applying the trained anatomic structure perception low-field multi-contrast magnetic resonance image joint reconstruction network to magnetic resonance imaging. The method solves the problem of insufficient imaging optimization of a lesion part in the existing reconstruction method.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical nuclear magnetic resonance imaging, and particularly relates to a low-field multi-contrast fast MRI joint reconstruction method with anatomical structure perception. BACKGROUND

[0002] Low-field magnetic resonance imaging (Low-field MRI) is a technology with lower cost and lighter equipment, which has important application in resource-limited medical scenarios. Compared with high-field MRI, low-field MRI has the advantages of less impact on patients and more convenient equipment, but also has the problems of 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 joint imaging in clinical diagnosis, which is particularly prominent.

[0003] Multi-contrast magnetic resonance imaging (MC-MRI) can provide T1-weighted (T1w), T2-weighted (T2w) and fluid-attenuated inversion recovery (FLAIR) images of different contrasts under the same anatomy, which can provide complementary information and help improve diagnostic accuracy. 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 reduce scan time by undersampling K-space data and reconstruct complete images through algorithms, which have been widely used in high-field MRI, but due to the low signal-to-noise ratio of low-field MRI, these methods cannot be directly applied to low-field environment. In recent years, deep learning methods have provided a new way for magnetic resonance image reconstruction. Data-driven networks such as U-Net and ResNet can effectively improve image reconstruction quality by learning image features through a large amount of training data. However, these methods lack close coupling with the actual physical model, especially in low-field MRI, the complexity of deep network structure and inference time become factors to be considered. To address these challenges, model-driven deep learning methods have 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, etc. These methods 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. However, in clinical applications, a set of multi-contrast images are usually needed to provide more comprehensive diagnostic information. Based on this, in recent years, some researchers have proposed multi-contrast joint reconstruction methods. These methods use the correlation between cross-contrast images to improve reconstruction performance. For example, DenseNet and MM-GAN use high-quality guide images to enhance the reconstruction effect of target contrast images, and Hi-Net integrates multi-contrast image information through a deep learning network to obtain high-quality MR images. Although the above magnetic resonance imaging techniques have made significant progress in improving the overall image quality, existing reconstruction algorithms mainly focus on improving the overall image quality, but the imaging optimization of the lesion site is insufficient, which limits its application effect in clinical diagnosis and treatment. SUMMARY

[0004] The purpose of the present application is to provide an anatomically-aware low-field multi-contrast fast MRI joint reconstruction method to solve the problem of insufficient imaging optimization of the lesion site in existing reconstruction methods.

[0005] The technical scheme adopted by the present application is that the anatomically-aware low-field multi-contrast fast MRI joint reconstruction method is implemented according to the following steps:

[0006] Step 1: Construct a joint reconstruction model of low-field multi-contrast magnetic resonance images with anatomical structure perception;

[0007] Step 2: Construct an anatomically perceptive low-field multi-contrast magnetic resonance image joint reconstruction network based on the reconstruction model built in Step 1.

[0008] Step 3: Train the joint reconstruction network of low-field multi-contrast magnetic resonance images for anatomical structure perception;

[0009] Step 4: Apply the trained anatomical structure-aware low-field multi-contrast magnetic resonance image joint reconstruction network to perform magnetic resonance imaging.

[0010] The invention is further characterized in that,

[0011] The specific process of step 1 is as follows:

[0012] Given K contrast-sampling k-space data The goal is to extract data from k-space. Reconstruct the corresponding high-quality multi-contrast image. ,in This indicates the number of k-space data samples. Indicates the number of pixels in the sampled image;

[0013] The joint reconstruction model of low-field multi-contrast magnetic resonance images with anatomical structure perception is as follows:

[0014] (1)

[0015] In equation (1), Represents sparse transformation; This represents a segmentation network used for precise localization of the target region of interest. This represents a denoising network; and This represents the weighting parameter.

[0016] The specific process of step 2 is as follows:

[0017] Step 2.1: Solve the anatomically perceptive low-field multi-contrast magnetic resonance image joint reconstruction model constructed in Step 1 using the alternating minimization algorithm, and obtain:

[0018] (9)

[0019] Step 2.2, the derivation and expansion of formula (9) is carried out into a cascaded multi-stage deep network as a joint reconstruction network for low-field multi-contrast magnetic resonance images for anatomical structure perception. Rebuild the network Including denoising networks Anatomical perception data consistency module and group sparse module GS.

[0020] In step 2.2, the denoising network is composed of seven convolution modules connected in sequence and a common convolution with a convolution kernel of 1x1, and the output of the first convolution module and the output of the sixth convolution module fuse features through Concatenation operation, the output of the second convolution module and the output of the fifth convolution module fuse features through Concatenation operation, and the output of the third convolution module and the output of the fourth convolution module fuse features through Concatenation operation; the channel numbers 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 each 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 common convolution with a convolution kernel size of , a BN layer and a Relu activation function, and the step size is 1;

[0021] The denoising network is represented as: (10).

[0022] In step 2.2, the anatomical perception data consistency module comprises a data consistency module DC and an anatomical perception module TM;

[0023] The data consistency module DC is represented as:

[0024] (11)

[0025] The anatomical perception module TM is represented as:

[0026] (12)

[0027] In the formula, is a pre-trained segmentation network;

[0028] The segmentation network is composed of eight convolution modules connected in sequence, a common convolution with a convolution kernel of 1x1 and a sigmoid function, and the output of the first convolution module and the output of the sixth convolution module fuse features through Concatenation operation, the output of the second convolution module and the output of the fifth convolution module fuse features through Concatenation operation, the output of the third convolution module and the output of the fourth convolution module fuse features through Concatenation operation, and the output of the denoising network is subjected to sparse transformation The obtained sparse features are fused with the output of the seventh convolutional module through a concatenation operation. The number of channels in the eight convolutional modules are 128, 256, 512, 1024, 512, 256, 128, and 64, respectively. Each of the eight convolutional modules consists of two common layers (CoRe), and each common layer has a kernel size of [missing information]. It consists of a regular convolution with a stride of 1, a BN layer, and a ReLU activation function;

[0029] Then the anatomical perception data consistency module The final expression is:

[0030] (13).

[0031] In step 2.2, the group sparse module GS is represented as:

[0032] (14)

[0033] The group of sparse modules GS first undergoes a learnable nonlinear transformation. Obtain high-level feature representation Then, it passes through a learnable threshold parameter. soft threshold operator That is, in formula (9) To capture group sparsity between different contrasts in the high-level representation space, and then perform a left inverse operation. MR images can then be reconstructed from a sparse representation space. and The architecture is completely symmetrical and must meet the following requirements. ,in Represents the identity operator.

[0034] The specific process of step 3 is as follows:

[0035] Step 3.1, Construct the training dataset

[0036] The training dataset was constructed based on the low-field MRI dataset M4Raw. Each group of data in the training dataset contains undersampled k-space data with different contrast levels. and the corresponding full-sample magnetic resonance images ,in The number of contrast ratios is represented by the undersampling of fully sampled k-space data with different contrast ratios in the low-field MRI dataset M4Raw during data construction. This yields the corresponding k-space undersampled data. As a reference reconstructed image As a network reconstruction Input;

[0037] Step 3.2. Training the anatomical structure-aware low-field multi-contrast magnetic resonance image joint reconstruction network with the training data set constructed in step 3.1 ;

[0038] The loss function in the training process is:

[0039] (15)

[0040] In the formula, is to ensure that the final output result of the reconstruction network at the T stage is consistent with the corresponding fully sampled reference image ; represents the number of magnetic resonance image contrasts; is to ensure that ; is to ensure that the reconstructed image is consistent with the fully sampled image in the anatomical structure part; represents a general dice loss, which aims to promote accurate segmentation of a specific target of interest (TOI) in the reconstruction network by the segmentation network ; represents the segmentation network outputs a segmentation probability map at the T stage; are three tuning parameters used to balance the contributions of different loss terms.

[0041] The beneficial effects of the present application are: the anatomical structure-aware low-field multi-contrast fast MRI joint reconstruction method of the present application can efficiently reconstruct high-quality multi-contrast images from undersampled k-space multi-contrast data; and can simultaneously optimize the sampling mode and adaptively reconstruct the required high-quality multi-contrast images according to clinical requirements and downstream task requirements, which can provide high-quality medical influence data for doctors to perform clinical diagnosis later, thereby improving the accuracy and efficiency of diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a flowchart of the anatomical structure-aware low-field multi-contrast fast MRI joint reconstruction method of the present application;

[0043] Figure 2 is a framework diagram of the anatomical structure-aware low-field multi-contrast magnetic resonance image joint reconstruction network constructed by the present application;

[0044] Figure 3 is a T1 contrast magnetic resonance image reconstruction example in an embodiment of the present application;

[0045] Figure 4 is a T2 contrast magnetic resonance image reconstruction example in an embodiment of the present application;

[0046] Figure 5 This is an example of FLAIR contrast magnetic resonance image reconstruction in an embodiment of the present invention. Detailed Implementation

[0047] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0048] Example 1

[0049] This invention relates to a method for rapid low-field, multi-contrast MRI reconstruction based on anatomical structure perception, such as... Figure 1 As shown, please follow these steps:

[0050] Step 1: Construct a joint reconstruction model of low-field multi-contrast magnetic resonance images with anatomical structure perception;

[0051] The specific process is as follows:

[0052] Given K contrast-sampling k-space data The goal is to extract data from k-space. Reconstruct the corresponding high-quality multi-contrast image. ,in This indicates the number of k-space data samples. Indicates the number of pixels in the sampled image;

[0053] The joint reconstruction model of low-field multi-contrast magnetic resonance images with anatomical structure perception is as follows:

[0054] (1)

[0055] In equation (1), Represents sparse transformation; This represents a segmentation network used for precise location of the TOI (Target Area of ​​Interest). This indicates a denoising network that further refines and denoises the TOI region to improve the imaging quality of the TOI in the reconstructed image. and This represents the weighting parameter, used to balance the contributions of the regularization term and the data consistency term to the model;

[0056] Formula (1) consists of three parts:

[0057] (1) For data assurance items: This represents the fidelity or consistency between the reconstructed multi-contrast image and the actual undersampled K-space measurement data, ensuring that the reconstruction result is highly consistent with the original k-space data, thereby guaranteeing the accuracy of image reconstruction under undersampled data conditions;

[0058] (2) For group sparse regularization terms: via The norm group sparsity constraint is used for multi-contrast images, and the spatial similarity between contrast images is used for information fusion to extract shared features; here the sparse transformation is learned through a convolution network , to improve the expression ability of sparse representation; this method extracts shared features of multi-contrast images in a high-level semantic space, further enhancing the relevance between each contrast;

[0059] (3) Anatomy-aware regularization term: By combining the learnable anatomical prior information, the imaging effect of the target region of interest (TOI) is strengthened; Specifically, a segmentation network is introduced to accurately locate the TOI, providing key specific anatomical location information for the denoising process of the reconstructed image; and 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.

[0060] In order to better capture the complementary information between different contrast images under the guidance of specific clinical needs, the present application combines the reconstruction process with a learnable K-space sampling pattern to improve imaging efficiency and quality.

[0061] Step 2, construct an anatomy-aware low-field multi-contrast magnetic resonance image joint reconstruction network according to the reconstruction model constructed in step 1;

[0062] The specific process is as follows:

[0063] Step 2.1, solve the anatomy-aware low-field multi-contrast magnetic resonance image joint reconstruction model constructed in step 1 by an alternating minimization algorithm;

[0064] The specific process is as follows:

[0065] Step 2.1.1, for the anatomy-aware regularization term, use the Taylor expansion formula to approximate the denoising network , expressed as:

[0066] (2)

[0067] In the formula, denotes the Jacobian matrix with respect to , is the change amount in each iteration process, that is, ;

[0068] Let be , then the anatomy-aware regularization term is rewritten as follows:

[0069] (3)

[0070] Further approximation is:

[0071]

[0072] (4)

[0073] In the formula, denotes the segmentation network The Jacobian matrix of ;

[0074] When , the anatomy-aware regularization term can be approximated as:

[0075] (5)

[0076] Step 2.1.2, alternating optimization and , the expression is:

[0077] (6)

[0078] And the second sub-problem in formula (6) is solved by using the proximal gradient descent method;

[0079] The specific process is:

[0080] Step 2.1.2.1, get the intermediate updated reconstruction result by gradient descent method , denotes the intermediate reconstruction result of the th contrast, and for each contrast image, the update formula is as follows:

[0081] (7)

[0082] In the formula, represents the learnable step size, is the undersampled k-space data of the th contrast, denotes the learnable forward operator of the th contrast,

[0083]

[0084] In the formula, denotes the coil sensitivity map corresponding to the th receiving coil; denotes the one-dimensional discrete Fourier transform in the transverse direction; denotes the one-dimensional discrete Fourier transform in the vertical direction;

[0085]

[0086]

[0087] in, Represents a set of learnable position parameters. ,in , ; This represents the number of K-space samples taken in the phase encoding direction in an undersampled scenario; This represents the number of K-spaces sampled in the phase-encoding direction in a full-sample scenario; the speedup factor is defined as... ;

[0088] This invention only considers one-dimensional imaging scenarios, i.e., fixed. ,make During the reconstruction process, only the position parameters of the phase encoding direction need to be learned. ;

[0089] Step 2.1.2.2: After gradient update, the soft thresholding algorithm is used to estimate the reconstruction result of the current stage. The updated formula is:

[0090] (8)

[0091] In the formula, express The first of the matrix List, This represents a learnable sparse transformation. express The left inverse transform;

[0092] Step 2.1.3 summarizes the entire iterative optimization algorithm as follows:

[0093] (9)

[0094] Step 2.2, the derivation and expansion of formula (9) is carried out into a cascaded multi-stage deep network as a joint reconstruction network for low-field multi-contrast magnetic resonance images for anatomical structure perception. ,like Figure 2 As shown, reconstruct the network Including denoising networks Anatomical perception data consistency module GS (Sparse Modules);

[0095] (1) Denoising Network is a structure based on convolutional neural network, which is composed of seven convolutional modules connected in turn and a common convolution with a convolution kernel of 1x1, and the output of the first convolutional module and the output of the sixth convolutional module fuse features through Concatenation operation, the output of the second convolutional module and the output of the fifth convolutional module fuse features through Concatenation operation, and the output of the third convolutional module and the output of the fourth convolutional module fuse features through Concatenation operation; the channel numbers of the seven convolutional modules are 32, 64, 128, 256, 128, 64 and 32 respectively, the first convolutional module, the second convolutional module, the third convolutional module, the fifth convolutional module, the sixth convolutional module and the seventh convolutional module are all composed of two common layers CoRe, and the fourth convolutional module is composed of one common layer CoRe; each common layer is composed of a common convolution with a convolution kernel size of , a step of 1, a BN layer and a Relu activation function; the denoising network can effectively extract features and retain spatial information of the image, ensuring that important anatomical details are not lost during the denoising process;

[0096] Denoising network is used for removing noise and artifacts from the reconstruction result of the previous stage;

[0097] According to formula (1) and formula (5), the denoising network is expressed as:

[0098] (10)

[0099] (2) Anatomical perception data consistency module includes a data consistency module DC and an anatomical perception module TM;

[0100] Wherein, the role of the data consistency module DC is to force the reconstruction result to be consistent with the observed k-space data by using the observed k-space data, to ensure that the reconstructed image does not deviate from the actual observed k-space data, and the expression is:

[0101] (11)

[0102] The role of the anatomical perception module TM is to quantify the difference between the reconstruction result of the previous stage and the current denoising result , especially in the region of interest of the anatomical structure, and the expression is:

[0103] (12)

[0104] In the formula, is a pre-trained segmentation network, and the expression is: ​Some parameters are frozen during the training of the reconstruction network to generate a probability map of a target of interest (TOI), which helps to locate the anatomical structure of interest, and through the difference correction of the TOI part, it is ensured that the reconstructed image can capture the information of the anatomical structure part, and the anatomical accuracy of the reconstruction result is improved.

[0105] The segmentation network is a simple U-Net network composed of eight convolution modules connected in turn, a normal convolution with a convolution kernel of 1x1, and a sigmoid function, and the output of the first convolution module and the output of the sixth convolution module are fused by Concatenation operation, the output of the second convolution module and the output of the fifth convolution module are fused by Concatenation operation, the output of the third convolution module and the output of the fourth convolution module are fused by Concatenation operation, and the output of the denoising network is obtained by sparse transformation The sparse features obtained are fused with the output of the seventh convolution module by Concatenation operation; the channel numbers of the eight convolution modules are 128, 256, 512, 1024, 512, 256, 128, and 64, respectively, and each of the eight convolution modules is composed of two common layers CoRe, each of which is composed of a normal convolution with a convolution kernel size of 3x3 and a step of 1, a BN layer and a Relu activation function.

[0106] The final expression of the anatomical perception data consistency module is:

[0107] (13)

[0108] Formula (13) integrates data consistency and anatomical perception, ensuring that the reconstructed image can accurately capture the region of interest in anatomy and improve the reconstruction quality of the part while maintaining the consistency of the observation data;

[0109] The purpose of the module design is to combine the denoising results obtained during 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;

[0110] (3) The group sparsity module GS is applied to the middle reconstruction result based on the group sparsity of the norm to further enhance the MRI image reconstruction effect, especially the anatomical structure part;

[0111] The expression for the sparse module GS is:

[0112] (14)

[0113] The group of sparse modules GS first undergoes a learnable nonlinear transformation. Obtain high-level feature representation Then, it passes through a learnable threshold parameter. soft threshold operator That is, in formula (9) To capture group sparsity between different contrasts in the high-level representation space, and then perform a left inverse operation. MR images can then be reconstructed from a sparse representation space. and The architecture is completely symmetrical and must meet the following requirements. ,in Represents the identity operator;

[0114] This invention forces sparse transformation With segmentation network A weight-sharing strategy was implemented between the top-level encoders. This strategy effectively established group sparsity constraints in a meaningful high-level semantic space, promoted the fusion of complementary information between various contrasts, and thus improved the imaging quality of specific anatomical structures.

[0115] Step 3: Train the joint reconstruction network of low-field multi-contrast magnetic resonance images for anatomical structure perception;

[0116] The specific process is as follows:

[0117] Step 3.1, Construct the training dataset

[0118] The training dataset was constructed based on the low-field MRI dataset M4Raw. Each group of data in the training dataset contains undersampled k-space data with different contrast levels. and the corresponding full-sample magnetic resonance images ,in The number of contrast ratios is represented by the undersampling of fully sampled k-space data with different contrast ratios in the low-field MRI dataset M4Raw during data construction. This yields the corresponding k-space undersampled data. As a reference reconstructed image As a network reconstruction Input;

[0119] The low-field MRI dataset M4Raw includes fully sampled k-space data and corresponding fully sampled images for three contrast levels: T1, T2, and FLAIR.

[0120] Step 3.2. Training the anatomy-aware low-field multi-contrast magnetic resonance image joint reconstruction network with the training data set constructed in step 3.1 ;

[0121] reconstruction network The parameters to be learned in the reconstruction network can be summarized as ;

[0122] The loss function in the training process is:

[0123] (15)

[0124] wherein, is to ensure that the final output result of the reconstruction network at the T stage is consistent with the corresponding fully sampled reference image , represents the number of magnetic resonance image contrasts; is to ensure that ; is to ensure that the reconstructed image and the fully sampled image are consistent in the anatomical structure part; represents a general dice loss, which aims to promote accurate segmentation of a specific target of interest (TOI) in the reconstruction network by a segmentation network , represents the segmentation probability map output by the segmentation network at the T stage; are three tuning parameters used to balance the contributions of different loss terms;

[0125] The training of the reconstruction network is carried out on a computing device based on a pytorch framework, equipped with an Ubuntu 20.04 operating system and an RTX 3090Ti GPU, 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 a Xavier random initialization strategy to ensure a reasonable distribution of initial parameters, so as to improve the stability and convergence speed of network training; in order to further optimize the training effect, the present application adopts an Adam optimizer for parameter training, which has good convergence; in the training process, the iteration period is set to 100 times, 4 small batches of data are used in each iteration, and the learning rate is set to 1×10⁻ 4 , so as to realize progressive optimization of the network;

[0126] In terms of input of the model, the present application uses the initial input image Set to zero padding image, ensure the initial input without any prior information interference; for the three weighted parameters alpha, beta and gamma involved in formula (11), the application sets them to 0.01, 0.1 and 1 respectively through a large number of experiments and tuning process; these parameter values play a key role in balancing different loss terms and the training target of the network, which can effectively improve the accuracy and quality of the reconstructed image, and enhance the stability and robustness of the model.

[0127] Step 4, applying the trained anatomical structure perception low-field multi-contrast magnetic resonance image joint reconstruction network to perform magnetic resonance imaging;

[0128] Through the above training process, the optimal parameters of the anatomical structure perception multi-contrast magnetic resonance reconstruction network can be obtained, and based on the trained network, a plurality of contrast low-field k-space data can be 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 identical to standard full-sampling nuclear magnetic resonance images. This network can effectively compensate for the information loss caused by undersampling, thereby providing high-quality image reconstruction results even in the case of fast imaging. This method effectively improves the image quality under low sampling conditions, thereby speeding up the magnetic resonance imaging process and improving its clinical application practicability.

[0129] Embodiment 2

[0130] In order to verify the effectiveness of the application, numerical experiments use 158 registered T1, T2 and FLAIR contrast brain magnetic resonance images of human body to train and test the network. For each individual, the size of each contrast image is The application undersamples the 2D slice images of T1, T2 and FLAIR contrast according to the 1D Cartesian sampling mode at different sampling rates, thereby obtaining 158 sets of brain magnetic resonance imaging data required for experiments. Among them, 128 sets of data are used for training, and 30 sets of data are used for evaluating the performance of the trained network. In this embodiment, the k-space data is sampled at a rate of 1 / 8 to perform network reconstruction experiments and evaluation.

[0131] Embodiment 3

[0132] The difference from embodiment 2 is that the k-space data is sampled at a rate of 1 / 10 to perform network reconstruction experiments and evaluation.

[0133] To quantitatively evaluate the reconstruction performance, the present application adopts two commonly used image quality evaluation indicators, namely the peak signal-to-noise ratio (PSNR) and the 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. For this purpose, the present application also measures the PSNR and SSIM values of the specific TOI region, denoted as TOI-PSNR and TOI-SSIM, respectively, in order to more accurately evaluate the performance of the reconstruction network in the lesion area or the key anatomical site. Specifically, for the M4Raw data set, TOI evaluation is performed in the subcortical region of the brain.

[0134] Through the evaluation of these specific TOI regions, the superiority of the reconstruction network in the anatomical structure perception ability and the restoration of local details can be more meticulously verified, especially in the key lesion site or the anatomical structure.

[0135] Tables 1 and 2 show the performance comparison of the anatomical structure perception low-field multi-contrast magnetic resonance image joint reconstruction method of the present application with a variety of current mainstream single-contrast and multi-contrast reconstruction methods: (1) Single-contrast methods specifically include the traditional Grappa as the baseline, the pure data-driven network U-Net, the model-unfolded deep network ISTA-Net, and the multi-task network IDMHV that combines the segmentation task to improve the image reconstruction performance. (2) Multi-contrast methods include the pure data-driven MTrans, which introduces an improved transformer to capture complementary information between different contrasts, and the unfolded model MC-J-MoDL, which simultaneously optimizes the sampling pattern and the reconstruction network to improve performance. Compared with other methods, the reconstruction network A²MRI proposed by the present application shows the best reconstruction accuracy under different sampling patterns and sampling rates, especially in complex structures and lesion areas. Even at a low sampling rate, A²MRI still maintains a high PSNR and SSIM, significantly better than other methods. Figure 3 、 Figure 4 and Figure 5 The visualization results of and show that A²MRI can reconstruct clear details and has no artifacts in the lesion area, which is superior to other methods in detail preservation and artifact suppression, especially in the anatomical structure region of interest.

[0136] Table 1: Comparison results of different methods on the brain data test set under 1D Cartesian 1 / 10 sampling rate (TOI: subcortical region of the brain)

[0137]

[0138] Table 2: Comparison results of different methods on the brain data test set under 1D Cartesian 1 / 8 sampling rate (TOI: subcortical region of the brain)

[0139]

Claims

1. A method for combined low-field, multi-contrast, rapid MRI reconstruction based on anatomical structure perception, characterized in that: The specific steps are as follows: Step 1: Construct a joint reconstruction model of low-field multi-contrast magnetic resonance images with anatomical structure perception; The specific process of step 1 is as follows: Given K contrast-sampling k-space data The goal is to extract data from k-space. Reconstruct the corresponding high-quality multi-contrast image. ,in This indicates the number of k-space data samples. Indicates the number of pixels in the sampled image; The joint reconstruction model of low-field multi-contrast magnetic resonance images with anatomical structure perception is as follows: （1） In equation (1), Represents sparse transformation; This represents a segmentation network used for precise localization of the target region of interest. This represents a denoising network; and Indicates the weighting parameter; Indicates the first A learnable forward operator with contrast; Represents a set of learnable position parameters; Step 2: Construct an anatomically perceptive low-field multi-contrast magnetic resonance image joint reconstruction network based on the reconstruction model built in Step 1. The specific process of step 2 is as follows: Step 2.1: Solve the anatomically perceptive low-field multi-contrast magnetic resonance image joint reconstruction model constructed in Step 1 using the alternating minimization algorithm, and obtain: （9） In equation (9), This indicates the denoising result of the denoising network; Indicates the first Intermediate reconstruction results with varying contrast; Represents the learnable step size; Indicates the consistency of anatomical perception data module The processing results; This indicates that it has a learnable threshold parameter; Step 2.2, the derivation and expansion of formula (9) is carried out into a cascaded multi-stage deep network as a joint reconstruction network for low-field multi-contrast magnetic resonance images for anatomical structure perception. Rebuild the network Including denoising networks Anatomical perception data consistency module Sparse modules (GS) and groups; Step 3: Train the joint reconstruction network of low-field multi-contrast magnetic resonance images for anatomical structure perception; Step 4: Apply the trained anatomical structure-aware low-field multi-contrast magnetic resonance image joint reconstruction network to perform magnetic resonance imaging.

2. The method for combined low-field multi-contrast rapid MRI reconstruction based on anatomical structure perception according to claim 1, characterized in that, In step 2.2, the denoising network It consists of seven convolutional modules connected sequentially and a regular convolutional layer with a 1×1 kernel. The outputs of the first and sixth convolutional modules are fused using a concatenation operation; the outputs of the second and fifth convolutional modules are fused using a concatenation operation; and the outputs of the third and fourth convolutional modules are fused using a concatenation operation. The number of channels in the seven convolutional modules are 32, 64, 128, 256, 128, 64, and 32, respectively. The first, second, third, fifth, sixth, and seventh convolutional modules each consist of two common layers (CoRe), while the fourth convolutional module consists of one common layer (CoRe). Each common layer consists of a convolutional layer with a kernel size of [missing information]. It consists of a regular convolution with a stride of 1, a BN layer, and a ReLU activation function; The denoising network is represented as: (10).

3. The method for combined low-field multi-contrast rapid MRI reconstruction based on anatomical structure perception according to claim 2, characterized in that, In step 2.2, the anatomical perception data consistency module Includes the Data Consistency Module (DC) and the Anatomical Sensing Module (TM); The data consistency module (DC) is represented as follows: （11） The Anatomical Sensing Module TM is represented as: （12） In the formula, It is a pre-trained segmentation network; Segmentation network The denoising network consists of eight sequentially connected convolutional modules, a 1×1 convolutional kernel, and a sigmoid function. The outputs of the first and sixth convolutional modules are fused using a concatenation operation; the outputs of the second and fifth convolutional modules are fused using the same operation; and the outputs of the third and fourth convolutional modules are fused using the same operation. The output is subjected to sparse transformation The obtained sparse features are fused with the output of the seventh convolutional module through a concatenation operation. The number of channels in the eight convolutional modules are 128, 256, 512, 1024, 512, 256, 128, and 64, respectively. Each of the eight convolutional modules consists of two common layers (CoRe), and each common layer has a kernel size of [missing information]. It consists of a regular convolution with a stride of 1, a BN layer, and a ReLU activation function; Then the anatomical perception data consistency module The final expression is: （13）。 4. The method for combined low-field multi-contrast rapid MRI reconstruction based on anatomical structure perception according to claim 3, characterized in that, In step 2.2, the group sparse module GS is represented as: （14） The group of sparse modules GS first undergoes a learnable nonlinear transformation. Obtain high-level feature representation Then it passes through a learnable threshold parameter. soft threshold operator That is, in formula (9) To capture group sparsity between different contrasts in the high-level representation space, and then perform a left inverse operation. MR images can then be reconstructed from a sparse representation space. and The architecture is completely symmetrical and must meet the following requirements. ,in Represents the identity operator.

5. The method for combined low-field multi-contrast rapid MRI reconstruction based on anatomical structure perception according to claim 4, characterized in that, The specific process of step 3 is as follows: Step 3.1, Construct the training dataset The training dataset was constructed based on the low-field MRI dataset M4Raw. Each group of data in the training dataset contains undersampled k-space data with different contrast levels. and the corresponding full-sample magnetic resonance images ,in The number of contrast ratios is represented by the undersampling of fully sampled k-space data with different contrast ratios in the low-field MRI dataset M4Raw during data construction. This yields the corresponding k-space undersampled data. As a reference reconstructed image As a network reconstruction Input; Step 3.2: Train the anatomical structure-aware low-field multi-contrast magnetic resonance image joint reconstruction network using the training dataset constructed in Step 3.

1. ; The loss function during training is: （15） In the formula, This is to ensure the final output of the reconstruction network in phase T. With the corresponding fully sampled reference image Maintain consistency Indicates the number of contrast features in a magnetic resonance image; In order to ensure ; This is to ensure consistency between the reconstructed image and the fully sampled image at the anatomical structures. This represents the general Dice loss, which aims to improve network segmentation. Facilitates accurate segmentation of targets of specific interest (TOIs) in the reconstructed network. Represents a segmented network The segmentation probability map output at stage T; These are three tuning parameters used to balance the contributions of different loss terms.

Images