Multi-contrast magnetic resonance image reconstruction method based on frequency domain error prior guidance

Abstract

The invention discloses a multi-contrast magnetic resonance image reconstruction method based on frequency domain error priori guidance, and the method comprises the steps: carrying out the modeling of a characteristic decomposition strategy, space alignment and k-space data in the design of a deep expansion network through frequency domain error priori guidance, and giving a definite physical constraint to the network; and irrelevant background or noise in the reference image, data acquisition efficiency and reconstruction quality under a significant sub-sampling condition, and a signal-to-noise ratio and marginal definition of the target image are filtered out.

Description

Technical Field

[0001] This invention relates to a technique in the field of image processing, specifically a multi-contrast magnetic resonance image reconstruction method based on frequency domain error prior guidance. Background Technology

[0002] Magnetic resonance imaging (MRI), with its radiation-free nature and superior soft tissue contrast, plays a crucial role in the diagnosis of brain tumors and neurological diseases. Clinically, multiple contrast images (such as T1- and T2-weighted images) are typically acquired to provide complementary pathological information. Due to physical limitations, MRI scans are time-consuming. Utilizing fully sampled reference modal data to assist subsampled target modal data for rapid reconstruction has become an important means to shorten scan time and improve imaging efficiency. However, the fixed subsampling strategy of existing multi-contrast MRI reconstruction techniques limits sampling efficiency, and spatial misalignment of multiple contrast images makes traditional fusion methods prone to artifacts. Furthermore, existing depth-unfolded network features are still insufficient in terms of feature utilization and physical constraints, limiting further improvements in reconstruction accuracy. Summary of the Invention

[0003] To address the aforementioned shortcomings of existing technologies, this invention proposes a multi-contrast magnetic resonance image reconstruction method based on frequency domain error prior guidance. This method combines frequency domain error prior guidance with a deep unfolding network optimization framework for multi-contrast MRI reconstruction. By utilizing the frequency domain error prior generated by the conditional diffusion model to jointly optimize the subsampling method and reconstruction network parameters, high-quality image reconstruction of subsampling magnetic resonance images can be achieved.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a multi-contrast magnetic resonance image reconstruction method based on frequency domain error prior. In the offline stage, the frequency domain error prior of the image is extracted from the training set through a conditional diffusion model and used to pre-train the constructed reconstruction network. The training set after discrete masking is then used to fine-tune the pre-trained reconstruction network. In the online stage, the fine-tuned reconstruction network generates the reconstructed image based on the subsampled image to be processed and the reference contrast image.

[0006] The reconstruction network comprises an initialization module, four iterative sub-modules, and a reconstruction layer. The initialization module performs channel expansion mapping and spatial transformation-based reference feature decomposition based on the input subsampled target contrast image and fully sampled reference image information to obtain reconstructed variables containing initial image domain features and initial k-space features, as well as reference image feature variables divided into relevant and irrelevant components. Each iterative sub-module performs multi-parameter alternating iterative update processing based on the image domain, k-space, and reference image feature variable information output by the initialization module or the previous iterative sub-module. It updates the image domain and k-space features using data consistency constraints and a near-end network, and dynamically corrects the relevant and irrelevant components of the reference image using a spatial alignment network, resulting in a multi-level optimized depth feature iteration result. The reconstruction layer performs inverse Fourier transform of frequency domain features and weighted fusion of spatial and frequency domain features, along with channel dimensionality reduction processing, based on the final-level image domain and k-space feature information output by the iterative sub-modules, to obtain the final high-fidelity multi-contrast magnetic resonance reconstruction image result.

[0007] Technical effect

[0008] This invention achieves superior sampling efficiency by jointly optimizing the frequency domain error prior, the subsampling method, and the reconstruction network. This optimization is guided by the frequency domain, aligning the subsampling with the reconstruction network to enhance its performance. On three public datasets—IXI, Brats2018, and FastMRI—this invention demonstrates excellent reconstruction metrics compared to other state-of-the-art methods. Visually, it enables high-quality MR image reconstruction. Attached Figure Description

[0009] Figure 1 This is a flowchart of the present invention;

[0010] Figure 2 This is a schematic diagram of a conditional diffusion model;

[0011] Figure 3 A schematic diagram of the overall network structure is shown below;

[0012] Figure 4 A schematic diagram of the initialization module, spatial alignment network, and reconstruction layer structure in the reconstruction network;

[0013] Figure 5 This refers to the overall structure of the iterative submodule;

[0014] Figure 6 This is a structural diagram of the near-end network, data consistency layer, and feature transformation block in the iterative submodule;

[0015] Figure 7 This is a schematic diagram illustrating the effect of an example. Detailed Implementation

[0016] This embodiment addresses a set of multi-contrast MRI scan data containing a fully sampled reference modal image Y (such as a T1 or PD-weighted image) and a subsampled target modal image. Using the frequency domain error prior-guided multi-contrast MRI image reconstruction method of this invention, the target modality is first synthesized directly from the reference modality using a conditional diffusion model. Then, FFT is used to transform the synthesized image and the target image into the frequency domain space, calculating their difference—the frequency domain error prior—to characterize the region with the most significant difference between the reference and target modalities. After obtaining the frequency domain error prior, it is integrated into the multi-contrast MRI reconstruction network as prior knowledge. This promotes the joint optimization of the subsampled method and reconstruction network parameters under the guidance of the frequency domain error prior, thereby restoring the subsampled target modal image to a clear, artifact-free fully sampled target image.

[0017] like Figure 1 As shown, this embodiment specifically includes:

[0018] Step 1, as follows Figure 2 As shown, the frequency domain error prior is obtained through backpropagation using a conditional diffusion model, specifically including:

[0019] 1.1 Forward propagation process: Noise is gradually added to the target contrast images in the training set until pure Gaussian noise is obtained through cyclic updates;

[0020] The training set is preferably constructed by randomly allocating reference contrast cases and target contrast cases in a ratio of 7:1:2 to form the training set, validation set, and test set. Taking the IXI dataset as an example, a total of 570 cases are selected and constructed into the training set, validation set, and test set in a ratio of 399:57:114. After allocation, the 3D cases are sliced, and the center 20 images are taken from each slice. Taking the IXI dataset as an example, the final image ratio of the training set, validation set, and test set is 7980:1140:2280.

[0021] The aforementioned cyclical update refers to: , The image at time t, The image at time t-1, The noise scheduling parameter is a constant between 0 and 1, which varies with t. , It is Gaussian noise randomly sampled from a standard normal distribution.

[0022] The aforementioned cyclic update, after several noise additions, satisfies the average noise scheduling parameters. Approaching 0 and The distribution of is statistically equivalent to the standard normal distribution. Specifically: .

[0023] In this embodiment, the number of steps is set to T=1000.

[0024] 1.2 Backpropagation process: Noise is gradually removed using a conditional diffusion model until a denoised target image is generated, specifically as follows: ,in: For the noisy image at time step t, The image at time t-1, For reference contrast image, For conditional denoising networks, the parameters are: , For the noise scheduling at step t to be sparse, This is the cumulative coefficient. Standard Gaussian noise, The variance weights for the reverse process. .

[0025] The conditional diffusion model described herein is implemented using, but is not limited to, the techniques described in "Advances in Neural Information Processing Systems".

[0026] 1.3 Calculate the prior frequency domain error: ,in: For the target image to be denoised, For real images, For Fourier transform, This is a priori information for frequency domain error.

[0027] Step 2: Utilize the frequency domain error prior obtained in Step 1 to jointly optimize the subsampling mode and the conditional diffusion model, specifically including:

[0028] 2.1 Adjusting the frequency domain error prior and constructing a continuous sampling mask by using the sampling modulation matrix, specifically: For continuous sampling masks [where c stands for Continuous], and The slope parameters are respectively and The sigmoid activation function, For sparsification operations, The target sparsity level is set. For the normalization operation, the frequency domain error prior obtained in step 1 After normalization, it serves as the initial sampling probability distribution, and the matrix U then follows an interval... Uniform distribution on For the sampling modulation matrix, the interval is... Initialize with a uniform distribution.

[0029] 2.2 A convergent fixed-sampling modulation matrix is ​​obtained through joint optimization of the continuous sampling mask and the reconstruction network, specifically as follows: The total number of training samples For parameters Defined Reconstruction Network Reference modal image as auxiliary input, The optimal sampling modulation matrix, To reconstruct the optimal network parameters, For the i-th target, the true contrast image, For the i-th target contrast subsampled image It is the i-th consecutive sampling mask.

[0030] In this embodiment, the joint optimization training lasts for 40 epochs.

[0031] 2.3 Based on the fixed sampling modulation matrix obtained in step 2.2 and the frequency domain error prior obtained in step 1.3, after generating the continuous sampling mask distribution, it is further transformed into a binary discrete sampling mask, specifically as follows: Among them: continuous sampling mask distribution , , For the binary search threshold operation, the average sampling density of the mask strictly conforms to the preset sparsity index. , The formula for the total number of training samples is: It is the i-th consecutive sampling mask.

[0032] 2.4 Fine-tune the reconstructed network optimized in step 2.2, specifically as follows: ,in: To fine-tune and reconstruct the network parameters, For Reconstructing the network for parameters, For the i-th subsampled target contrast real image, For the i-th reference contrast image, For discrete subsampling masks, d represents discrete. Let i be the true contrast image of the i-th target.

[0033] The overall objective function of the reconstruction network is constructed based on compressed sensing MRI theory, and is as follows: ,in: To reconstruct the image, K represents the k-space data of the image to be reconstructed, S represents the features of the reference contrast image that are identical to the target contrast image after spatial alignment, and D represents the features of the reference contrast image that are different from the target contrast image after spatial alignment. For subsampling mask, For Fourier transform, This refers to the actual sampled k-space data. Here is a spatially aligned reference contrast image, where A and B represent some kind of feature transformation. For hyperparameters, This is an implicit regularization term.

[0034] The reconstruction network is constructed by mapping inertial blocks to a neural network architecture for reconstruction using the inertial block principal maximization minimization algorithm based on the idea of ​​deep expansion.

[0035] like Figure 3 As shown, the reconstruction network includes: an initialization module, four sequentially cascaded iterative sub-modules, and a reconstruction layer. The initialization module is connected to the first iterative sub-module and maps the input single-channel or multi-channel image to a high-dimensional feature space. Adjacent iterative sub-modules are connected by passing state variables, with the output of the previous iterative sub-module serving as the input of the next iterative sub-module. The reconstruction layer is connected to the last iterative sub-module and is used to map the high-dimensional features back to the image space.

[0036] The mapping to the high-dimensional feature space refers to: the initialization module based on the sampled target contrast image The initial image domain features are compared with the contrast image Y of the fully sampled target. Initial k-space features Initial reference correlation components Initial reference correlation components .

[0037] The state variables include: image domain features X, k-space features K, reference correlated components S, reference uncorrelated components D, and the momentum state at the previous time step;

[0038] The iterative submodules share the weight parameters of the proximal network and the feature transformation network during the iteration process; the first iterative submodule is based on the input of the initialization module. , , , ,make Complete the input for the first iterative submodule.

[0039] like Figure 4As shown, the initialization module specifically includes: three independent feature extraction networks, each containing convolutional residual layers, a spatial alignment network, and an output unit, wherein: the first feature extraction network is based on the subsampled target contrast image. After 10 layers of convolutional residual processing, the initial image domain features are obtained. The second feature extraction network obtains initial reference features based on the fully sampled target contrast image Y, after 10 layers of convolutional residual processing. Spatial alignment networks are based on initial image domain features. Compared with the initial reference features Output aligned reference features The third feature extraction network extracts features based on the aligned reference features. After 10 layers of convolutional residual processing, the initial reference correlation components are output. The output unit will generate the initial image domain features. The initial k-space features are obtained through Fast Fourier Transform. Aligned reference features Subtract the initial reference correlation component Obtain the initial reference correlation components .

[0040] The feature extraction network described herein employs a deep residual network structure consisting of 10 cascaded convolutional residual layers without batch normalization. Each convolutional residual layer includes: a first convolutional layer, a nonlinear activation layer, a second convolutional layer, and a skip connection path. Specifically: the first convolutional layer performs convolution operations using a 3×3 convolutional kernel based on the input features, with the number of output channels matching the number of input channels (32 dimensions), a stride of 1, and padding of 1; the nonlinear activation layer uses the ReLU function to perform a nonlinear transformation on the output of the first convolutional layer; the second convolutional layer also performs convolution operations using a 3×3 convolutional kernel with a stride of 1 and padding of 1 based on the activated features; the skip connection path directly passes the input features of the residual block to the end, adding them element-wise with the output of the second convolutional layer to obtain the output features of the current residual block; by stacking 10 identical convolutional residual layers, deep semantic features of the image are extracted while maintaining the stability of the feature distribution.

[0041] The spatial alignment network employs an architecture combining offset estimation and deformable convolution to perform pixel-level elastic calibration of the reference image at the feature level, eliminating spatial misalignment between modalities. This spatial alignment network includes an offset estimation subnetwork and a deformable resampling unit, wherein: the offset estimation subnetwork obtains an offset field based on the target image features and reference image features in the current iteration; the deformable resampling unit obtains aligned reference features based on the original reference image features and the offset field.

[0042] The offset estimation subnet adopts an encoder-decoder structure. Its input layer uses a combined feature tensor obtained by concatenating the target image features and reference image features in the current iteration stage along the channel dimension. The number of channels in this combined feature tensor is twice the number of channels in the input features. The encoder consists of three sets of downsampling convolutional modules. Each set contains a convolutional layer with a kernel size of 3×3 and padding of 1, and a LeakyReLU activation function. By setting a stride of 2 for convolution, the feature map resolution is gradually reduced to extract multi-scale spatial location features. The decoder consists of three sets of upsampling convolutional modules. Each set contains a bilinear interpolation upsampling layer and a convolutional layer. The output features of each layer of the decoder are concatenated with the features of the corresponding layer of the encoder to fuse deep semantic information and shallow spatial information. The output layer outputs an offset field through the last convolutional layer. The number of channels is twice the product of the number of deformable convolutional groups and the kernel size, corresponding to the offset vectors in the x and y directions.

[0043] The deformable resampling unit employs a split deformable convolution. Based on the original reference image features and the estimated offset field, it dynamically interpolates and samples the reference features according to the position indicated by the offset field, and outputs spatially aligned reference features.

[0044] In the offset estimation subnet and deformable resampling unit, except for the 1×1 convolution, all convolutional layers use 3×3 convolution kernels, with padding of 1, a basic feature channel count of 32, and a deformable convolution group count of 4.

[0045] like Figure 5 As shown, the iterative submodule includes four parallel parameter update networks: the image domain update network X-Net, the k-space update network K-Net, the reference component-dependent update network S-Net, and the reference component-independent update network D-Net.

[0046] The image domain update network X-Net updates the image domain based on the features of the previous level image domain. Features of the next higher k-space The previous level references the relevant components. Original subsampled k-space data The subsampling mask M is then processed through a parallel data consistency branch, a k-space constraint branch, and a cross-modal feature projection branch to obtain data consistency feature terms. The k-space constraint gradient term, the image domain refinement gradient term, and the sum of these three are input into the first proximal network x to obtain the update result of the current iteration stage: image domain features. ;

[0047] The data consistency branch is based on the features of the previous level image domain. Original subsampled k-space data and subsampling mask M, through such Figure 6 The data consistency layer shown The frequency domain prediction value is obtained by performing a fast Fourier transform. Then, the frequency domain prediction value and Multiplication , For size The matrix with all values ​​of 1 is then compared with the original subsampled k-space data. The data consistency feature term is obtained after inverse Fourier transform. .

[0048] The k-space constraint branch utilizes the inverse transformation of k-space variables to constrain image domain variables, maintaining consistency between the two domains. This is based on the features of the previous-level image domain. Features of the next higher k-space First of all Perform inverse Fourier transform and then with The difference is multiplied by the weighting factor. We obtain the k-space constrained gradient term, where: For trade-off parameters.

[0049] The aforementioned cross-modal feature projection branch utilizes the correlation component S of the reference image to guide the restoration of texture details in the target image. Based on the features of the previous image domain... The previous level references the relevant components. The processing procedure is to The input is fed into the first feature transformation block, which maps it to a high-dimensional latent feature space; The input is fed into the second feature transformation block, which also maps it to a high-dimensional latent feature space. After calculating the difference between the two in the latent feature space, the difference is input into the third feature transformation block. Then, the difference in the image domain is multiplied by a weighting coefficient. The image domain refinement gradient term is obtained, where: For trade-off parameters.

[0050] The k-space update network K-Net updates based on the features of the previous k-space. Inertial k-space characteristics and features of the next higher level image domain The frequency-domain consistent gradient term and intermediate k-space variables are obtained by passing through parallel frequency-domain consistent gradient branches and k-space inertial momentum aggregation branches respectively. The intermediate frequency-domain consistent gradient term and k-space variables are added together and divided into two along the channel dimension: the first 32 dimensions are real features and the last 32 dimensions are imaginary features. The real features are input into the second proximal network k_real, and the imaginary features are input into the third proximal network k_imag. The denoised real features and imaginary features output from the two proximal networks are then concatenated along the channel dimension to obtain the updated k-space features for the current iteration. .

[0051] The frequency-domain consistent gradient branch calculates the frequency domain deviation between the current k-space features and the current image domain features, using this deviation as the update direction for gradient descent. Based on the previous k-space features... Compared with the features of the previous level image domain First of all Perform a Fast Fourier Transform to map it to the frequency domain, denoted as Multiply by the step size parameter The frequency domain consistent gradient term is obtained, where: This is a learnable step size parameter.

[0052] The k-space inertial momentum convergence branch is based on the characteristics of the previous k-space. Inertial k-space characteristics , Multiply by a coefficient , Multiply by a coefficient Then, summing these two terms yields the intermediate k-space variables, where: Let K be the inertial weight parameter of the k-space feature K. This is a learnable step size parameter.

[0053] The reference component correlation update network S-Net updates the reference component based on the previous level correlation reference component. Inertial related reference components The higher-level reference is an irrelevant component. Upper-level image domain features and initial reference features After passing through the parallel correlated component inertial momentum aggregation branch, spatial alignment and decomposition gradient branch, and cross-modal feature consistency projection branch to obtain the updated parameter values, spatial gradient update term, and feature domain-based refined update term, respectively, the intermediate reference component correlated variables, spatial gradient update term, and feature domain-based refined update term are summed and then input into the fourth proximal network s to obtain the update result of the current iteration stage, the reference correlated component. .

[0054] The related component inertial momentum convergence branch is based on the related component of the previous level related reference. Inertial related reference components ,make Multiply by a coefficient , Multiply by a coefficient The intermediate reference component is obtained by summing the results, where: The inertial weight parameter is used to reference the relevant component S. For learnable step size parameters, For trade-off parameters.

[0055] The spatial alignment and decomposition gradient branches are based on the features of the previous image domain. Initial reference features Components independent of the previous level reference First, and Input to a spatial alignment network to obtain spatially aligned reference features Then, calculate and The difference is then multiplied by a coefficient. The spatial gradient update term is obtained.

[0056] The cross-modal feature consistency projection branch is based on the features of the previous level image domain. Related components of the previous level ,Will Input to feature transformation block A, The inputs are fed into feature transformation block B, which maps both features to the same high-dimensional latent feature space. The difference between the two features in the latent feature space is then calculated and fed back into the feature transformation block. Multiply by a coefficient Finally, a refined update term based on the feature domain is obtained.

[0057] The reference component-independent update network D-Net updates based on the reference component of the previous level. Inertial reference independent components The higher level references the relevant components. Upper-level image domain features and initial reference features After obtaining the updated parameter values ​​and the gradient terms of the updated D component through parallel unrelated component inertial momentum aggregation branches and spatial alignment and residual decomposition gradient branches, the unrelated variables of the intermediate reference components are added to the gradient terms of the updated D component and then input into the fifth proximal network d to obtain the update result of the reference related components in the current iteration stage. .

[0058] The aforementioned irrelevant component inertial momentum convergence branch is based on the previous level reference irrelevant component. Components independent of inertial reference . Multiply by a coefficient , Multiply by a coefficient The intermediate reference component, which is independent of the variables, is obtained by summing them together. The inertial weighting parameter is used to reference the irrelevant component D. This is a learnable step size parameter.

[0059] The spatial alignment and residual decomposition gradient branch are based on the features of the previous image domain. Initial reference features and related references from the previous level. ,Will and Input to a spatial alignment network to obtain spatially aligned reference features Then, calculate and The difference is then multiplied by a coefficient. The gradient term for updating the D component is obtained.

[0060] The inertia weight parameters are all initialized to 0.25 during training, i.e. , 0.25, 0.25.

[0061] The learnable step size parameter is initialized to 1 during training, i.e. , 1, 1, 1.

[0062] The aforementioned tradeoff parameter is initialized to 1 during training, i.e. , 1, 1.

[0063] During the network training phase, these parameters are not treated as fixed constants, but participate in the backpropagation process. Their values ​​are adaptively updated based on the gradient information of the loss function in order to learn the optimal parameter configuration for the current iteration phase.

[0064] like Figure 6 As shown, the first feature conversion block, the second feature conversion block, and the third feature conversion block All layers employ a five-layer chain structure consisting of three convolutional layers and two nonlinear activation layers stacked alternately. Specifically: the first convolutional layer is a feature enhancement convolutional layer with a 3×3 kernel, a stride of 1, and padding of 1, used to map the number of input feature channels to a preset feature channel dimension; the second convolutional layer is a ReLU nonlinear activation layer, used to introduce nonlinear features; the third convolutional layer is a feature mapping convolutional layer, also using a 3×3 kernel, a stride of 1, and padding of 1, used for deep feature transformation in the latent space; the first nonlinear activation layer is a ReLU nonlinear activation layer; the second nonlinear activation layer is a feature reconstruction convolutional layer, whose kernel size, stride, and padding parameters are consistent with the aforementioned convolutional layers, used to output the final transformed features. Furthermore, this layer's output is not connected to an activation function to fully preserve the negative value information of the features, thus supporting subsequent gradient residual calculations.

[0065] like Figure 6 As shown, the first proximal network x, the second proximal network k_real, the third proximal network k_imag, the fourth proximal network s, and the fifth proximal network d all adopt a cascaded residual block structure without batch normalization, wherein: each residual block contains a feature extraction main path and a skip connection path. The main path sequentially contains a first convolutional layer, a ReLU activation layer, and a second convolutional layer. The first and second convolutional layers are both configured with 3×3 convolutional kernels, a stride of 1, and padding of 1 to ensure that the spatial resolution of the feature map remains unchanged during transmission; the skip connection path directly transmits the input signal of the residual block to the end without loss. The output feature of the main path and the input feature of the skip connection path are added element-wise to obtain the output of the current residual block.

[0066] like Figure 4 As shown, the reconstruction layer includes: a frequency-to-spatial conversion stitching unit and a weighted average fusion unit, wherein: the frequency-to-spatial conversion stitching unit is based on the k-space features output from the last iteration. Image domain features after performing inverse Fast Fourier Transform and the output of the last iteration The channels are concatenated to obtain a mixed feature tensor with 64 channels; the weighted averaging unit performs a 2D convolution operation on the mixed feature tensor using a 3×3 convolution kernel, a stride of 1, and padding of 1, with the output channel set to 1, to obtain the final predicted image. .

[0067] Through practical application experiments, the above-mentioned multi-contrast magnetic resonance image reconstruction based on frequency domain error prior guidance was performed on the RTX 3090 GPU platform and PyTorch framework. For the conditional diffusion model, the Adam optimizer was used for training, with a fixed learning rate. No weight decay was set. To ensure training stability, an exponential moving average (weights set to 0.999) was used. The model was trained for 300 epochs with a diffusion step size of T=1000. For the reconstruction model, the Adam optimizer was also used with a learning rate of 10^(-4) and a batch size of 2, for a total of 30 epochs.

[0068] This embodiment uses three evaluation metrics to assess the model's training results: Peak Signal-to-Noise Ratio (PSNR), Structure Consistency Index (SSIM), and Mean Squared Error (RMSE). PSNR is the most commonly used objective metric for measuring image reconstruction quality; it is based on the logarithm of the error between pixels, and a higher value indicates better reconstruction quality. SSIM comprehensively evaluates the similarity between two images from three dimensions: brightness, contrast, and structure; a value closer to 1 indicates better reconstruction performance. RMSE measures the deviation between predicted and true values; a value closer to 0 indicates more accurate model predictions.

[0069] The present invention was compared with zero-value filling (Zeo-Filled), classical MRI reconstruction methods, UNet, Multi-UNet, CUNet, MDUNet, Restomer, and the two latest multi-contrast MRI reconstruction methods, VANet and MC-DuDoN. Quantitative comparison data are shown in Table 1, and visual comparison data are shown below. Figure 6 As shown.

[0070]

[0071] Compared with existing technologies, this invention utilizes a conditional diffusion model (CDM) to generate a frequency domain error prior (FEP) for the target mode, reflecting the recovery difficulty of each frequency point in the k-space. By integrating the FEP into the joint optimization framework, the network is forced to prioritize the acquisition of frequency data with larger errors (i.e., difficult to recover), thereby significantly improving the sampling effectiveness and reconstruction details. A spatial alignment and feature decomposition module is proposed. Embedding a spatial alignment network into the deep unfolded network can adaptively correct the spatial deviation between the reference mode and the target mode. Simultaneously, through a feature decomposition strategy, reference features are separated into relevant and irrelevant features, using only relevant features to assist reconstruction, effectively avoiding the negative impact of interference information. A deep unfolded network based on multivariate iterative updates is constructed. This invention is based on the Inertial Block Superiority Minimization (TITAN) algorithm, unfolding the optimization algorithm into a deep neural network, and alternately updating the image domain, k-space, and feature domain variables in each iteration stage. This design not only gives the network a clear physical meaning but also significantly improves the fidelity of image reconstruction through multi-domain constraints.

[0072] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for reconstructing multi-contrast magnetic resonance images based on frequency domain error prior guidance, characterized in that, In the offline stage, the frequency domain error prior of the image is extracted from the training set through the conditional diffusion model. This prior is used to pre-train the constructed reconstruction network. Then, the training set after discrete masking is used to fine-tune the pre-trained reconstruction network. In the online phase, a reconstructed image is generated from the subsampled image to be processed and the reference contrast image using a finely tuned reconstruction network. The reconstruction network includes an initialization module, four iterative sub-modules, and a reconstruction layer. Each iterative sub-module performs multi-parameter alternating iterative update processing based on the inertial block UHA minimization algorithm, according to the image domain, k-space, and reference image feature variable information output by the initialization module or the previous iterative sub-module. It updates the image domain and k-space features using data consistency constraints and a near-end network, and dynamically corrects the relevant and irrelevant components of the reference image using a spatial alignment network to obtain the multi-level optimized depth feature iteration result.

2. The multi-contrast magnetic resonance image reconstruction method based on frequency domain error prior guidance according to claim 1, characterized in that, The initialization module performs channel expansion mapping and reference feature decomposition based on spatial transformation according to the input subsampled target contrast image and fully sampled reference image information, to obtain the variables to be reconstructed containing initial image domain features and initial k-space features, as well as reference image feature variables divided into relevant and irrelevant components.

3. The multi-contrast magnetic resonance image reconstruction method based on frequency domain error prior guidance according to claim 1, characterized in that, The reconstruction layer performs inverse Fourier transform of frequency domain features and weighted fusion of spatial and frequency domain features and channel dimensionality reduction based on the final-level image domain features and k-space feature information output by the iterative submodule, to obtain the final high-fidelity multi-contrast magnetic resonance reconstruction image result.

4. The multi-contrast magnetic resonance image reconstruction method based on frequency domain error prior guidance according to claim 1 or 2, characterized in that, The initialization module includes: three independent feature extraction networks, each containing convolutional residual layers, a spatial alignment network, and an output unit, wherein: the first feature extraction network is based on the subsampled target contrast image. After 10 layers of convolutional residual processing, the initial image domain features are obtained. The second feature extraction network obtains initial reference features based on the fully sampled target contrast image Y, after 10 layers of convolutional residual processing. Spatial alignment networks are based on initial image domain features. Compared with the initial reference features Output aligned reference features The third feature extraction network extracts features based on the aligned reference features. After 10 layers of convolutional residual processing, the initial reference correlation components are output. The output unit will generate the initial image domain features. The initial k-space features are obtained through Fast Fourier Transform. Aligned reference features Subtract the initial reference correlation component Obtain the initial reference correlation components .

5. The multi-contrast magnetic resonance image reconstruction method based on frequency domain error prior guidance according to claim 1, characterized in that, The iterative submodule specifically comprises four parallel parameter update networks: an image domain update network X-Net, a k-space update network K-Net, a reference component-correlated update network S-Net, and a reference component-independent update network D-Net. The image domain update network X-Net updates parameters based on the features of the previous image domain. Features of the next higher k-space The previous level references the relevant components. Original subsampled k-space data The subsampling mask M is then processed through a parallel data consistency branch, a k-space constraint branch, and a cross-modal feature projection branch to obtain data consistency feature terms. The k-space constraint gradient term, the image domain refinement gradient term, and the sum of these three are input into the first proximal network x to obtain the update result of the current iteration stage: image domain features. ; Data consistency branches are based on the features of the previous image domain. Original subsampled k-space data and the subsampling mask M, through the data consistency layer The frequency domain prediction value is obtained by performing a fast Fourier transform. Then, the frequency domain prediction value and Multiplication , For size The matrix with all values ​​of 1 is then compared with the original subsampled k-space data. The data consistency feature term is obtained after inverse Fourier transform. ; K-Net, a k-space update network, updates its network based on the characteristics of the previous k-space. Inertial k-space characteristics and features of the next higher level image domain The frequency-domain consistent gradient term and intermediate k-space variables are obtained by passing through parallel frequency-domain consistent gradient branches and k-space inertial momentum aggregation branches respectively. The intermediate frequency-domain consistent gradient term and k-space variables are added together and divided into two along the channel dimension: the first 32 dimensions are real features and the last 32 dimensions are imaginary features. The real features are input into the second proximal network k_real, and the imaginary features are input into the third proximal network k_imag. The denoised real features and imaginary features output from the two proximal networks are then concatenated along the channel dimension to obtain the updated k-space features for the current iteration. ; The S-Net, a network that updates reference components based on the previous level's related reference components, updates the network based on the previous level's related reference components. Inertial related reference components The higher-level reference is an irrelevant component. Upper-level image domain features and initial reference features After passing through parallel inertial momentum aggregation branch, spatial alignment and decomposition gradient branch, and cross-modal feature consistency projection branch to obtain updated parameter values, spatial gradient update terms, and feature domain-based refined update terms, the intermediate reference component related variables, spatial gradient update terms, and feature domain-based refined update terms are added together and then input into the fourth proximal network s to obtain the update result of the current iteration stage, the reference related components. ; Reference component-independent update network D-Net updates based on the reference-independent components of the previous level. Inertial reference independent components The higher level references the relevant components. Upper-level image domain features and initial reference features After obtaining the updated parameter values ​​and the gradient terms of the updated D component through parallel unrelated component inertial momentum aggregation branches and spatial alignment and residual decomposition gradient branches, the unrelated variables of the intermediate reference components are added to the gradient terms of the updated D component and then input into the fifth proximal network d to obtain the update result of the reference related components in the current iteration stage. .

6. The multi-contrast magnetic resonance image reconstruction method based on frequency domain error prior guidance according to claim 1 or 3, characterized in that, The reconstruction layer includes: a frequency-to-spatial conversion stitching unit and a weighted average fusion unit, wherein: the frequency-to-spatial conversion stitching unit is based on the k-space features output from the last iteration. Image domain features after performing inverse Fast Fourier Transform and the output of the last iteration The channels are concatenated to obtain a mixed feature tensor with 64 channels; the weighted averaging unit performs a 2D convolution operation on the mixed feature tensor using a 3×3 convolution kernel, a stride of 1, and padding of 1, with the output channel set to 1, to obtain the final predicted image. .

7. The multi-contrast magnetic resonance image reconstruction method based on frequency domain error prior guidance according to any one of claims 1-6, characterized in that, specifically include: Step 1: Obtain the frequency domain error prior through backpropagation using the conditional diffusion model, specifically including: 1.1 Forward propagation process: Noise is gradually added to the target contrast images in the training set until pure Gaussian noise is obtained through cyclic updates; 1.2 Backpropagation process: Noise is gradually removed using a conditional diffusion model until a denoised target image is generated, specifically as follows: ,in: For the noisy image at time step t, The image at time t-1, For reference contrast image, For conditional denoising networks, the parameters are: , For the noise scheduling at step t to be sparse, This is the cumulative coefficient. Standard Gaussian noise, The variance weights for the reverse process. ; 1.3 Calculate the prior frequency domain error: ,in: For the target image to be denoised, For real images, For Fourier transform, Frequency domain error prior; Step 2: Utilize the frequency domain error prior obtained in Step 1 to jointly optimize the subsampling mode and the conditional diffusion model, specifically including: 2.1 Adjusting the frequency domain error prior and constructing a continuous sampling mask by using the sampling modulation matrix, specifically: For continuous sampling masks [where c stands for Continuous], and The slope parameters are respectively and The sigmoid activation function, For sparsification operations, The target sparsity level is set. For the normalization operation, the frequency domain error prior obtained in step 1 After normalization, it serves as the initial sampling probability distribution, and the matrix U then follows an interval... Uniform distribution on For the sampling modulation matrix, the interval is... Initialize with a uniform distribution; 2.2 A convergent fixed-sampling modulation matrix is ​​obtained through joint optimization of the continuous sampling mask and the reconstruction network, specifically as follows: The total number of training samples For parameters Defined Reconstruction Network Reference modal image as auxiliary input, The optimal sampling modulation matrix, To reconstruct the optimal network parameters, For the i-th target, the true contrast image, For the i-th target contrast subsampled image This is the i-th consecutive sampling mask; 2.3 Based on the fixed sampling modulation matrix obtained in step 2.2 and the frequency domain error prior obtained in step 1.3, after generating the continuous sampling mask distribution, it is further transformed into a binary discrete sampling mask, specifically as follows: Among them: continuous sampling mask distribution , , For the binary search threshold operation, the average sampling density of the mask strictly conforms to the preset sparsity index. , The formula for the total number of training samples is: This is the i-th consecutive sampling mask; 2.4 Fine-tune the reconstructed network optimized in step 2.2, specifically as follows: ,in: To fine-tune and reconstruct the network parameters, For Reconstructing the network for parameters, For the i-th subsampled target contrast real image, For the i-th reference contrast image, For discrete subsampling masks, d represents discrete. Let i be the true contrast image of the i-th target.

8. The multi-contrast magnetic resonance image reconstruction method based on frequency domain error prior guidance according to claim 7, characterized in that, The reconstruction network is constructed by mapping inertial blocks to a neural network architecture for reconstruction using the inertial block principal maximization minimization algorithm based on the idea of ​​deep expansion.

9. The multi-contrast magnetic resonance image reconstruction method based on frequency domain error prior guidance according to claim 7, characterized in that, The overall objective function of the reconstruction network is constructed based on compressed sensing MRI theory, and is as follows: ,in: To reconstruct the image, K represents the k-space data of the image to be reconstructed, S represents the features of the reference contrast image that are identical to the target contrast image after spatial alignment, and D represents the features of the reference contrast image that are different from the target contrast image after spatial alignment. For subsampling mask, For Fourier transform, This refers to the actual sampled k-space data. Here is a spatially aligned reference contrast image, where A and B represent some kind of feature transformation. For hyperparameters, This is an implicit regularization term.

Images