CN-122023593-A - Multi-contrast magnetic resonance image reconstruction method based on frequency domain error priori guidance

CN122023593ACN 122023593 ACN122023593 ACN 122023593ACN-122023593-A

Abstract

A multi-contrast magnetic resonance image reconstruction method based on frequency domain error priori guidance utilizes self-adaptive sampling guided by frequency domain error priori, models feature decomposition strategy, space alignment and k-space data in the design of a depth expansion network and filters irrelevant background or noise in a reference image at the same time of giving definite physical constraint to the network, and the data acquisition efficiency and reconstruction quality under the obvious sub-sampling condition and the signal-to-noise ratio and edge definition of a target image.

Inventors

LI JUNCHENG
FANG XINMING
WANG JUN
SHI JUN

Assignees

上海大学

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (9)

1. A multi-contrast magnetic resonance image reconstruction method based on frequency domain error priori guidance is characterized in that the frequency domain error priori of an image is extracted from a training set through a conditional diffusion model in an off-line stage and is used for carrying out pretraining on a reconstructed network obtained by construction, and the training set after discrete mask processing carries out fine tuning on the pre-trained reconstructed network; The reconstruction network comprises an initialization module, four iteration sub-modules and a reconstruction layer, wherein each iteration sub-module performs multi-parameter alternate iteration update processing based on the expansion of an inertial block UHA algorithm according to image domain, k space and reference image feature variable information output by the initialization module or a previous iteration sub-module, respectively updates image domain and k space features by utilizing data consistency constraint and a near-end network, dynamically corrects related and unrelated components of a reference image by utilizing a space alignment network, and obtains a depth feature iteration result subjected to multistage optimization.
2. The method for reconstructing a multi-contrast magnetic resonance image based on frequency domain error prior guidance according to claim 1, wherein the initialization module performs channel expansion mapping and reference feature decomposition processing based on spatial transformation according to the input sub-sampling target contrast image and full-sampling reference image information to obtain a variable to be reconstructed including an initial image domain feature, an initial k-space feature and a reference image feature variable divided into a related component and an unrelated component.
3. The multi-contrast magnetic resonance image reconstruction method based on frequency domain error priori guidance according to claim 1, wherein the reconstruction layer performs inverse fourier transform of frequency domain features and weighted fusion of space-frequency domain features and channel dimension reduction processing according to final-stage image domain features and k-space feature information output by the iteration sub-module, so as to obtain a final high-fidelity multi-contrast magnetic resonance reconstruction image result.
4. The method for reconstructing a multi-contrast magnetic resonance image based on frequency domain error a priori guidance of claim 1 or 2, wherein the initialization module comprises three independent feature extraction networks each including a convolution residual layer, a spatial alignment network, and an output unit, wherein the first feature extraction network is configured to extract a target contrast image based on sub-sampling After 10 layers of convolution residual processing, obtaining the initial image domain characteristics The second characteristic extraction network obtains initial reference characteristics through 10 layers of convolution residual processing according to the full-sampling target contrast image Y The spatial alignment network is based on the initial image domain features With initial reference features Outputting the aligned reference features A third feature extraction network extracts the reference features according to the aligned reference features After 10 layers of convolution residual processing, an initial reference related component is output The output unit characterizes the initial image domain Obtaining initial k-space characteristics through fast Fourier transformation Alignment of the aligned reference features Subtracting the initial reference correlation component Obtaining initial reference correlation components 。
5. The method of claim 1, wherein the iterative submodules include four parallel parameter update networks, i.e., an image domain update network X-Net, a K-space update network K-Net, a reference component correlation update network S-Net, and a reference component independent update network D-Net, wherein the image domain update network X-Net is based on the previous image domain characteristics Higher order k-space features Reference correlation component of the previous stage Raw sub-sampled k-space data And a sub-sampling mask M, which respectively obtains data consistency characteristic items through a parallel data consistency branch, a k-space constraint branch and a cross-modal characteristic projection branch The k-space constraint gradient term, the image domain refinement gradient term and the third are added and then input into a first near-end network x to obtain the updated result image domain characteristics of the current iteration stage ; The data consistency branch is based on the image domain characteristics of the previous stage Raw sub-sampled k-space data And subsampling mask M, through data consistency layer pair Performing fast Fourier transform to obtain its frequency domain predicted value Then, the frequency domain predicted value And (3) with Multiplication , Is of the size of Matrix with values of all 1 is followed by original sub-sampled k-space data Obtaining data consistency characteristic items through inverse Fourier transform ; K-space update network K-Net based on previous level K-space features Inertial k-space characteristics Image domain features of the previous level Obtaining a frequency domain consistent gradient item and an intermediate k-space variable through a parallel frequency domain consistent gradient branch and a k-space inertia momentum aggregation branch respectively, adding the intermediate frequency domain consistent gradient item and the k-space variable into two first 32 dimensions which are real part features and then 32 dimensions which are imaginary part features in the channel dimension, inputting the real part features into a real part near-end network second near-end network k_real, inputting the imaginary part features into an imaginary part third near-end network k_imag, re-splicing the denoised real part features and the imaginary part features output by the two near-end networks in the channel dimension, and finally obtaining an updated result k-space feature of the current iteration stage ; Reference component correlation update network S-Net correlates reference correlation components according to the previous stage Inertial related reference component The upper level reference independent component Image domain features of the previous level Initial reference feature The updated parameter values, airspace gradient updating items and characteristic domain-based refined updating items are obtained through a parallel inertia momentum aggregation branch, a space alignment and decomposition gradient branch and a cross-modal characteristic consistency projection branch respectively, and then the intermediate reference component related variable, the airspace gradient updating items and the characteristic domain-based refined updating items are added and then input into a fourth near-end network s to obtain the updated result reference related component in the current iteration stage ; Reference component independent update network D-Net based on the previous level reference independent component Inertial reference independent component The upper level refers to the relevant component Image domain features of the previous level Initial reference feature After the updating parameter value and the gradient item of the updating D component are obtained through the parallel independent component inertia momentum aggregation branch and the spatial alignment and residual error decomposition gradient branch respectively, the intermediate reference component independent variable and the gradient item of the updating D component are added and then input into a fifth near-end network D to obtain the updating result reference related component of the current iteration stage 。
6. The method for reconstructing a multi-contrast magnetic resonance image based on frequency domain error a priori guidance of claim 1 or 3, wherein the reconstruction layer comprises a frequency domain-space domain conversion stitching unit and a weighted average fusion unit, wherein the frequency domain-space domain conversion stitching unit outputs k-space characteristics according to the last iteration Image domain features output after inverse fast fourier transform and last iteration The mixed characteristic tensor with the channel number of 64 is obtained by splicing the channels, and the final predicted image is obtained by performing two-dimensional convolution operation on the mixed characteristic tensor by a weighted average unit by using a 3X 3 convolution kernel, the step length of 1 and filling 1, and setting the output channel of 1 。
7. The multi-contrast magnetic resonance image reconstruction method based on frequency domain error prior guidance according to any one of claims 1-6, comprising in particular: step 1, acquiring a frequency domain error prior through back propagation of a conditional diffusion model, which specifically comprises the following steps: 1.1, a forward propagation process, namely gradually adding noise to a target contrast image in a training set until the noise is circularly updated to obtain pure Gaussian noise; 1.2, a back propagation process, namely gradually removing noise through a conditional diffusion model until a denoised target image is generated, wherein the back propagation process comprises the following steps of: Wherein: for the noisy image at time step t, Is an image at the time of t-1, For the reference contrast image, The parameters are as follows for the conditional denoising network , Sparse is scheduled for the noise of step t, In order to accumulate the coefficients of the coefficients, As a result of the standard gaussian noise, As the variance weight of the inverse process, ; 1.3 Calculating a frequency domain error prior: Wherein: Is a denoised target image, As a real image of the object, In the form of a fourier transform, Is the frequency domain error priori; Step 2, the sub-sampling mode and the conditional diffusion model are jointly optimized by utilizing the frequency domain error priori obtained in the step 1, and the method specifically comprises the following steps: 2.1, adjusting frequency domain error prior through a sampling modulation matrix and constructing a continuous sampling mask, specifically: For a Continuous sampling mask where c represents Continuous, And Respectively the slope parameters are And Is used to activate the function of the sigmoid, In order to operate the thinning-out process, For a preset target sparsity level, For normalization operation, the frequency domain error obtained in step1 is a priori Normalized to be the initial sampling probability distribution, and the matrix U obeys the interval Is uniformly distributed on the surface of the base plate, For sampling the modulation matrix, the interval is adopted Initializing the uniform distribution of the (c); 2.2, obtaining a converged fixed sampling modulation matrix through co-optimization of the continuous sampling mask and the reconstruction network, wherein the fixed sampling modulation matrix specifically comprises the following components: To train the total number of samples Is composed of parameters Defined reconstruction network To assist in the input reference modality image, For an optimal sampling modulation matrix, In order to reconstruct the optimal network parameters of the network, For the ith target contrast real image, Subsampling the image for the ith target contrast The ith continuous sampling mask; 2.3 generating continuous sampling mask distribution based on the fixed sampling modulation matrix obtained in the step 2.2 and the frequency domain error priori obtained in the step 1.3, and further converting the continuous sampling mask distribution into a binary discrete sampling mask, wherein the method specifically comprises the following steps: wherein the mask distribution is sampled continuously , , For binary search threshold operation, the average sampling density of the mask strictly accords with a preset sparsity index , To train the formula for the total number of samples, The ith continuous sampling mask; 2.4 fine tuning the reconstructed network after the common optimization in the step 2.2, specifically: Wherein: in order to reconstruct the network parameters after fine tuning, To take the following measures For the reconstruction of the network of parameters, For the i-th subsampled target contrast real image, For the i-th reference contrast image, For a discrete sub-sampling mask, d represents the discrete, The i-th target contrast real image.
8. The method for reconstructing a multi-contrast magnetic resonance image based on frequency domain error a priori guidance according to claim 7, wherein the reconstruction network is constructed by mapping inertial blocks into a neural network architecture for reconstruction based on a depth expansion concept through an inertial block main maximization algorithm.
9. The multi-contrast magnetic resonance image reconstruction method based on frequency domain error prior guidance according to claim 7, wherein the total objective function of the reconstruction network is constructed according to a compressed sensing MRI theory, and specifically comprises: Wherein: for reconstructing an image, K is K-space data of the image to be reconstructed, S is a feature of the reference contrast image that is spatially aligned to be identical to the target contrast image, D is a feature of the reference contrast image that is spatially aligned to be different from the target contrast image, In order to sub-sample the mask, In the form of a fourier transform, For the actually sampled k-space data, For spatially aligned reference contrast images, a and B are some feature transformations, Is the parameter of the ultrasonic wave to be used as the ultrasonic wave, Is an implicit regularization term.

Description

Multi-contrast magnetic resonance image reconstruction method based on frequency domain error priori guidance Technical Field The invention relates to a technology in the field of image processing, in particular to a multi-contrast magnetic resonance image reconstruction method based on frequency domain error priori guidance. Background Magnetic Resonance Imaging (MRI) plays a key role in the diagnosis of brain tumors and neurological disorders by virtue of its radiopacity and excellent soft tissue contrast. A variety of contrast images (e.g., T1, T2 weighted images) are typically acquired clinically to provide complementary pathology information. Because of the long time consumption of MRI scanning caused by physical limitation, the sub-sampled target mode data is assisted by the fully sampled reference mode data to reconstruct rapidly, and the method becomes an important means for shortening the scanning time and improving the imaging efficiency. However, the fixed sub-sampling strategy of the existing multi-contrast MRI reconstruction technique results in limited sampling efficiency, and the misalignment of multiple contrast image spaces makes the conventional fusion approach prone to artifact. In addition, the existing deep expansion network has the defects of characteristic utilization rate and physical constraint, and further improvement of reconstruction accuracy is limited. Disclosure of Invention Aiming at the defects in the prior art, the invention provides a multi-contrast magnetic resonance image reconstruction method based on frequency domain error priori guidance, a sub-sampling technology based on frequency domain error priori guidance and a depth expansion network are used for multi-contrast MRI reconstruction, and the sub-sampling method and reconstruction network parameters are jointly optimized by using the frequency domain error priori generated by a conditional diffusion model, so that high-quality image reconstruction of a sub-sampled magnetic resonance image is realized. The invention is realized by the following technical scheme: The invention relates to a multi-contrast magnetic resonance image reconstruction method based on frequency domain error priori guidance, which is used for extracting the frequency domain error priori of an image from a training set through a conditional diffusion model in an off-line stage, pre-training a constructed reconstruction network, fine-tuning the pre-trained reconstruction network through the training set after discrete mask processing, and generating a reconstruction image according to a sub-sampling image to be processed and a reference contrast image through the fine-tuned reconstruction network in the on-line stage. The reconstruction network comprises an initialization module, four iteration sub-modules and a reconstruction layer, wherein the initialization module performs channel expansion mapping and reference feature decomposition processing based on space transformation according to an input sub-sampling target contrast image and full-sampling reference image information to obtain a to-be-reconstructed variable containing initial image domain features, initial k-space features and reference image feature variables divided into related components and irrelevant components, each iteration sub-module performs multi-parameter alternating iteration update processing based on the expansion of an inertia block optimal-Ha minimization algorithm according to image domain, k-space and reference image feature variable information output by the initialization module or a previous iteration sub-module, updates image domain and k-space features respectively by utilizing data consistency constraint and a near-end network, and dynamically corrects related and irrelevant components of a reference image by utilizing a space alignment network to obtain a depth feature iteration result subjected to multi-level optimization, and the reconstruction layer performs inverse Fourier transformation of frequency domain features and weighted fusion and channel dimension reduction processing of the frequency domain features according to final-level image domain features and k-space feature information output by the iteration sub-module to obtain a final high-fidelity magnetic resonance image reconstruction result. Technical effects The invention optimizes the frequency domain received by the sub-sampling method through the frequency domain error prior, the sub-sampling method and the reconstruction network together, so that the sampling efficiency is improved, and the sub-sampling and the reconstruction network are aligned, so that the reconstruction network can exert better performance. At IXI, brats, 2018 and FastMRI, good reconstruction indexes are obtained for other most advanced methods, and from the visual effect, the invention can realize high-quality MR image reconstruction. Drawings FIG. 1 is a flow chart of the present invention; FIG. 2 is a schematic diagra