Method and apparatus for reconstructing magnetic resonance image, and device and medium

Abstract

A method and apparatus for reconstructing a magnetic resonance image, and a device and a medium. The method comprises: acquiring magnetic resonance undersampled k-space data of a target region of interest, and inputting the magnetic resonance undersampled k-space data into a pre-trained self-supervised magnetic resonance image reconstruction model, so as to obtain the mapping between the magnetic resonance undersampled k-space data and a target magnetic resonance image and the distribution of network parameters corresponding to the self-supervised magnetic resonance image reconstruction model (S110); inputting the mapping and the distribution of the network parameters into a score-matching network constructed on the basis of a distribution feature of the network parameters corresponding to the self-supervised magnetic resonance image reconstruction model, so as to obtain the score of a target data distribution feature corresponding to the target magnetic resonance image (S120); and on the basis of the score of the target data distribution feature, reconstructing the magnetic resonance undersampled k-space data by means of preset Langevin Markov chain Monte Carlo sampling, so as to obtain a target magnetic resonance reconstructed image (S130).

Description

Magnetic resonance image reconstruction method, device, equipment and medium

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on December 27, 2023, with application number 202311821575.5, the entire contents of which are incorporated by reference into this application. Technical Field

[0002] The present application relates to the field of medical image processing technology, for example, to a magnetic resonance image reconstruction method, apparatus, device and medium. Background Art

[0003] Three-dimensional cardiovascular magnetic resonance imaging (3D-CMR) can provide more comprehensive anatomical information about the entire heart, thereby obtaining richer tissue parameter information. Deep learning models are often used in the MRI image reconstruction process for efficient image reconstruction.

[0004] However, deep learning-based image reconstruction models rely on supervised training using large amounts of fully sampled data. However, acquiring large amounts of fully sampled scan data requires long imaging scan times, which can easily lead to respiratory or heart rate instabilities, resulting in image quality degradation. Consequently, high-quality fully sampled data cannot be used as a criterion for supervised learning during the training of deep learning image reconstruction models, impacting model performance.

[0005] Summary of the Invention

[0006] The present application provides a magnetic resonance image reconstruction method, apparatus, device and medium, which can directly reconstruct high-quality magnetic resonance images from high-magnification undersampled k-space data, realizing image reconstruction based on magnetic resonance undersampled data in scenarios without full-sampled labeled data.

[0007] The present invention provides a method for reconstructing a magnetic resonance image, the method comprising:

[0008] Acquiring undersampled magnetic resonance k-space data of a target region of interest, and inputting the undersampled magnetic resonance k-space data into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain a mapping between the undersampled magnetic resonance k-space data and the target magnetic resonance image and a distribution of network parameters corresponding to the self-supervised magnetic resonance image reconstruction model;

[0009] Inputting the mapping and the distribution of the network parameters into a score matching network constructed based on the network parameter distribution characteristics of the self-supervised magnetic resonance image reconstruction model to obtain a score of the target data distribution characteristics corresponding to the target magnetic resonance image;

[0010] According to the score of the target data distribution feature, the magnetic resonance undersampled k-space data is reconstructed by a preset Langevin Markov chain Monte Carlo sampling method to obtain a target magnetic resonance reconstructed image.

[0011] The present application also provides a magnetic resonance image reconstruction device, which includes:

[0012] a first image reconstruction module configured to acquire magnetic resonance undersampled k-space data of a target region of interest, and input the magnetic resonance undersampled k-space data into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain a mapping between the magnetic resonance undersampled k-space data and a target magnetic resonance image and a distribution of network parameters corresponding to the self-supervised magnetic resonance image reconstruction model;

[0013] an image distribution feature determination module configured to input the mapping and the distribution of the network parameters into a score matching network constructed based on the network parameter distribution features of the self-supervised magnetic resonance image reconstruction model to obtain a score of the target data distribution feature corresponding to the target magnetic resonance image;

[0014] The second image reconstruction module is configured to reconstruct the magnetic resonance undersampled k-space data according to the score of the target data distribution feature through a preset Langevin Markov chain Monte Carlo sampling method to obtain a target magnetic resonance reconstructed image.

[0015] The present application also provides a computer device, comprising:

[0016] one or more processors;

[0017] a memory configured to store one or more programs;

[0018] When the one or more programs are executed by the one or more processors, the one or more processors implement the magnetic resonance image reconstruction method provided in any embodiment of the present application.

[0019] An embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the magnetic resonance image reconstruction method provided in any embodiment of the present application is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] FIG1 is a flow chart of a magnetic resonance image reconstruction method provided by an embodiment of the present application;

[0021] FIG2 is a schematic diagram of the structure of a Bayesian convolutional neural network provided in an embodiment of the present application;

[0022] FIG3 is a flow chart of a magnetic resonance image reconstruction method provided in an embodiment of the present application;

[0023] FIG4 is a reference formula for approximating the data distribution during the Bayesian convolutional neural network training process provided by an embodiment of the present application;

[0024] FIG5 is a reference formula for training a score matching network according to an embodiment of the present application;

[0025] FIG6 is a reference formula for the data diffusion process of the fractional matching network provided in an embodiment of the present application;

[0026] FIG7 is a description of a process for sampling data diffusion results based on the Langevin Markov Chain Monte Carlo sampling method provided in an embodiment of the present application;

[0027] FIG8A is a method for estimating data distribution based on a Bayesian convolutional neural network provided in an embodiment of the present application;

[0028] FIG8B is a schematic diagram of a forward diffusion and reverse diffusion process based on score matching provided in an embodiment of the present application;

[0029] FIG9 is a flow chart of a magnetic resonance image reconstruction method provided in an embodiment of the present application;

[0030] FIG10 is a schematic diagram of an example of performing multi-parameter matching based on pixel information of a magnetic resonance image provided by an embodiment of the present application;

[0031] FIG11 is a schematic structural diagram of a magnetic resonance image reconstruction device provided in an embodiment of the present application;

[0032] FIG12 is a schematic structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0033] The present application is described below in conjunction with the accompanying drawings and embodiments. The embodiments described herein are intended only to explain the present application and are not intended to limit the present application. For ease of description, the accompanying drawings only show portions related to the present application, not all structures.

[0034] Figure 1 is a flowchart of a magnetic resonance image reconstruction method provided by an embodiment of the present application. This embodiment is applicable to magnetic resonance image reconstruction scenarios, such as image reconstruction based on undersampled magnetic resonance data in scenarios without fully acquired labeled data. The method can be performed by a magnetic resonance image reconstruction device, which can be implemented using software and / or hardware and integrated into a computer device with application development capabilities.

[0035] As shown in FIG1 , the magnetic resonance image reconstruction method of this embodiment includes the following steps.

[0036] S110. Acquire magnetic resonance undersampled k-space data of a target region of interest, and input the magnetic resonance undersampled k-space data into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain a mapping between the magnetic resonance undersampled k-space data and the target magnetic resonance image and a distribution of network parameters corresponding to the self-supervised magnetic resonance image reconstruction model.

[0037] The target region of interest may be a target tissue region whose structure is to be acquired, and the undersampled MRI k-space data is data acquired by scanning the target region of interest using an MRI device. The undersampling factor of the undersampled MRI k-space data relative to the fully sampled k-space data may be set based on an image reconstruction algorithm or image reconstruction quality requirements.

[0038] In actual MRI, obtaining high-quality reconstructed images requires full sampling during the scanning of the region of interest. However, this requires relatively long scan times, which is not ideal for some regions of interest. For example, when performing an MRI scan of the heart to analyze myocardial parameters, prolonged scan times can easily cause respiratory or heart rate instabilities, leading to a decrease in image quality. Therefore, under-sampled k-space data is typically obtained during MRI scans.

[0039] The self-supervised MRI reconstruction model is a deep learning model developed by training a Bayesian convolutional neural network model based on pre-set undersampled scan sample data and reconstructing MRI images from samples corresponding to the pre-set undersampled scan sample data. In other words, the self-supervised MRI reconstruction model learns the mapping relationship between undersampled k-space data and MRI images, enabling image reconstruction based on undersampled k-space data.

[0040] Therefore, the magnetic resonance undersampled k-space data of the target region of interest can be input into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain the mapping between the undersampled k-space data and the target magnetic resonance image and the distribution of the network parameters corresponding to the self-supervised magnetic resonance image reconstruction model.

[0041] The self-supervised magnetic resonance image reconstruction model can be the Bayesian convolutional neural network structure shown in Figure 2. y and k zrepresents the phase encoding direction and the slice direction in the three-dimensional k-space data, respectively. Inverse Fast Fourier Transform (IFFT) (·) and Fast Fourier Transform (FFT) (·) represent the inverse Fourier transform and the Fourier transform operator, respectively. The projection operator Represents the projection on Y′: Y is the undersampled k-space data, I is the all-1 matrix, Conv represents the convolution operation, ReLu is the activation function, BConv is the Bayesian convolution, N c *N mc *2 represents the number of network channels, α and γ are preset constants, M' is the undersampling operator, and Y' is the sub-undersampling data.

[0042] S120. Input the mapping and the distribution of the network parameters into a score matching network constructed based on the network parameter distribution characteristics of the self-supervised magnetic resonance image reconstruction model to obtain a score of the target data distribution characteristics corresponding to the target magnetic resonance image.

[0043] In this embodiment, the score matching network is constructed based on the mapping between the undersampled k-space data and the magnetic resonance image corresponding to the self-supervised magnetic resonance image reconstruction model and the distribution of the network parameters of the self-supervised magnetic resonance image reconstruction model. Theoretically, based on Bayes' theorem, the probability distribution of the target data can be obtained. However, this process involves an integration operation, which makes it difficult to calculate the probability distribution of the target data. Therefore, the present application uses a score matching method to approximate the probability distribution of the target data. It can be as follows: the "score" in the score matching is converted into the partial derivative of the logarithm of the distribution characteristics of the reconstructed image data under the conditions of the network parameter distribution characteristics corresponding to the self-supervised magnetic resonance image reconstruction model. Furthermore, the target data distribution characteristics corresponding to the target magnetic resonance reconstructed image can be obtained. The target data distribution characteristics are the image data distribution probability density corresponding to the final target magnetic resonance image.

[0044] S130 , reconstructing the magnetic resonance undersampled k-space data according to the score of the target data distribution feature by using a preset Langevin Markov chain Monte Carlo sampling method to obtain a target magnetic resonance reconstructed image.

[0045] Based on the score of the target data distribution feature corresponding to the target magnetic resonance image, the initial magnetic resonance reconstructed image can be diffused and then sampled and reconstructed. The diffusion process is a data noise-adding process. The object of diffusion is the initial magnetic resonance reconstructed image, which is a magnetic resonance image obtained by processing the undersampled k-space data through inverse Fourier transform. The image reconstruction process can be a process of sampling and reconstructing the noisy magnetic resonance image using the Langevin Markov chain Monte Carlo (MCMC) sampling method, thereby obtaining the target magnetic resonance reconstructed image. The target magnetic resonance reconstructed image can be the same as the target magnetic resonance image mentioned above, and both can refer to X in the formula.

[0046] The technical solution of this embodiment obtains magnetic resonance undersampled k-space data of a target region of interest and inputs the magnetic resonance undersampled k-space data into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain a mapping between the magnetic resonance undersampled k-space data and the target magnetic resonance image and a distribution of network parameters corresponding to the self-supervised magnetic resonance image reconstruction model; the mapping and the distribution of network parameters are input into a score matching network constructed based on the network parameter distribution characteristics of the self-supervised magnetic resonance image reconstruction model to obtain a score of the target data distribution characteristics corresponding to the target magnetic resonance image; based on the score of the target data distribution characteristics, the magnetic resonance undersampled k-space data is reconstructed using a preset Langevin Markov chain Monte Carlo sampling method to obtain a target magnetic resonance reconstructed image. The technical solution of the embodiment of the present application solves the problem of low image reconstruction quality based on magnetic resonance undersampled data, and can directly reconstruct high-quality magnetic resonance images from high-magnification undersampled k-space data, realizing image reconstruction based on magnetic resonance undersampled data in a scenario without full-sampled labeled data.

[0047] Figure 3 is a flowchart of a magnetic resonance image reconstruction method provided in an embodiment of the present application. This embodiment shares the same concept as the magnetic resonance image reconstruction method described in the previous embodiment and describes the process of training a self-supervised magnetic resonance image reconstruction model. This method can be performed by a magnetic resonance image reconstruction device, which can be implemented using software and / or hardware and integrated into a computer device with application development capabilities.

[0048] As shown in FIG3 , the magnetic resonance image reconstruction method of this embodiment includes the following steps.

[0049] S210 : Undersample the preset undersampled scan sample data using a preset undersampled operator to obtain sub-undersampled sample data.

[0050] The preset undersampled scan sample data is the scan data obtained during an actual MRI scan. If obtaining full-sampled data is inconvenient, the preset undersampled scan sample data can be used as a relative reference to the full-sampled data. Accordingly, the reconstructed image corresponding to the preset undersampled scan sample data can be used as the learning target for the training model.

[0051] Assuming that Y and X are the undersampled k-space data and the magnetic resonance image corresponding to the undersampled k-space data, the reconstruction model of X reconstructed from the undersampled data can be expressed as: Where λ is a preset value, A is the encoding operator, which can be expressed as A=MFS, M is the data undersampling operator, F is the Fourier operator, S is the multi-channel coil sensitivity matrix, and R(X) is the regularization operator. When the trained self-supervised MRI image reconstruction model is a Bayesian Convolutional Neural Network (BCNN), R(X) can be considered as the distribution prior p(X) of the image data.

[0052] In this step, the undersampling operator M' is used to undersample the undersampled k-space data Y to obtain sub-undersampled data Y', which can be expressed as Y'=M'Y.

[0053] M is the data undersampling operator, which is the undersampling operator used in conventional MRI scanning. M' is the undersampling operator used in this embodiment to perform undersampling based on M. The undersampling multiples corresponding to these two undersampling operators can be set according to the actual MRI scanning imaging requirements.

[0054] S220. Using the sub-undersampled sample data as model training input samples, and using the sample-reconstructed magnetic resonance image corresponding to the preset undersampled scan sample data as a sample label, a Bayesian convolutional neural network model is trained to obtain a Bayesian convolutional neural network model for establishing a mapping relationship between the magnetic resonance undersampled k-space data and the magnetic resonance image corresponding to the magnetic resonance undersampled k-space data.

[0055] In this step, the Bayesian convolutional neural network estimates the mapping from Y to X, and then approximately estimates p(X). The BCNN trains the network model using the undersampled data Y′ as training data to obtain the mapping between Y′ and the filled image X′ corresponding to Y, where X′ can be obtained by inverse Fourier transform of Y. It can be expressed as: X′=f θ (Y′)+n1. Among them, f θ The mapping with the parameter θ learned by BCNN, f θ Follows the distribution p(θ). f θUsed to convert undersampled k-space data Y into image X corresponding to Y: X = f θ (Y)+n2.

[0056] Where n1 and n2 represent Gaussian noise with scales γ1 and γ2, respectively, and assume that where n t and θ s Express expectations, is the variance of the characterization scale.

[0057] In this embodiment, the Kullback-Leibler divergence analysis algorithm (KL divergence) is used to perform a constraint analysis on the model learning loss between the output results of the Bayesian convolutional neural network model and the sample labels corresponding to the output results. When the corresponding divergence value in the Kullback-Leibler divergence analysis algorithm meets the preset constraints, the model training process is terminated, and the trained Bayesian convolutional neural network model and the data distribution characteristics corresponding to the trained Bayesian convolutional neural network model are obtained.

[0058] The preset constraint condition can be the formula shown in FIG4. When the formula in FIG4 reaches the minimum value, an approximate estimate q(θ) can be obtained. In the formula, N represents the number of training sample data Y; const represents a constant. q(θ) is an approximate estimate of p(θ), which is the distribution of network parameters. In the formula, μ θ represents expectation, σ θ represents the variance, σ S represents the variance of noise of different scales, and s is the sth random distribution.

[0059] S230 , constructing a score matching network based on the network parameter distribution characteristics of the self-supervised magnetic resonance image reconstruction model.

[0060] Among them, the score in the score matching network can be understood as the data distribution probability feature corresponding to the reconstructed image, which can be expressed as That is, it is expressed by the partial derivative of the logarithm of the data distribution prior p(X). The score matching network can use S φ (X), where φ is the network parameter.

[0061] According to Bayesian theory, it can be expressed as This equation can be obtained by training the formula shown in Figure 5. For f θ (Y) Apply different scales Gaussian noise data.

[0062] Exemplarily, according to the output result of the Bayesian convolutional neural network model with the sub-under-sampled sample data as input data and the distribution characteristics of the network parameters, noise of different scales can be gradually added to the initial magnetic resonance reconstruction image of the sub-under-sampled sample data to obtain a set of noisy magnetic resonance images; then, the fractional matching network is trained based on the above set of noisy magnetic resonance images and a preset fractional matching algorithm to obtain a target fractional matching network; wherein, the initial magnetic resonance reconstruction image is an image obtained by performing an inverse Fourier transform based on the sub-under-sampled sample data.

[0063] After the score matching network is determined, the Langevin Markov chain Monte Carlo (MCMC) sampling method can be used to perform sampling reconstruction using the formula in FIG6 and the process shown in FIG7 to obtain the final reconstructed image. i Indicates the step size, z i It obeys the standard normal distribution and can be regarded as a random variable that obeys the standard normal distribution. t represents a random variable, X t represents the image, and ∈ represents the noise.

[0064] In an optional implementation, a score matching network may be constructed using U-net.

[0065] In one example, the training process of the BCNN network and the training process of the score matching network can refer to the contents shown in Figures 8A and 8B. As shown in Figures 8A and 8B, in Figure 8A, the relationship between the undersampled data Y' and the image corresponding to the corresponding data Y is established through BCNN, and the approximate estimate of the network parameter distribution q(θ|μ θ ,σ θ ). Wherein, Y′ is the data obtained by undersampling from Y. In FIG8B , Forward represents the diffusion process of the forward image data denoising, and Reverse represents the process of image reconstruction using the Langevin Markov chain Monte Carlo (MCMC) sampling method, thereby obtaining the target magnetic resonance reconstruction image. In the forward denoising process, and Y t The data distribution characteristics are shown in the figure. Where T represents the number of iterations and L represents the noise intensity. is the input undersampled data, and is the magnetic resonance image determined by inverse Fourier transform.

[0066] S240. Obtain scan data to be reconstructed, and input the scan data to be reconstructed into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain a mapping between the scan data to be reconstructed and a target magnetic resonance image and a distribution of network parameters corresponding to the self-supervised magnetic resonance image reconstruction model.

[0067] After completing the model training of the target function through the above steps, it can be applied to the analysis process of undersampled k-space data.

[0068] The scan data to be reconstructed may be data acquired by performing magnetic resonance scanning on any part of interest, and is usually undersampled k-space data.

[0069] S250. Input the mapping and the distribution of the network parameters into a score matching network constructed based on the network parameter distribution characteristics of the self-supervised magnetic resonance image reconstruction model to obtain a score of the target data distribution characteristics corresponding to the target magnetic resonance image.

[0070] S260 , reconstructing the magnetic resonance undersampled k-space data by a preset Langevin Markov chain Monte Carlo sampling method according to the score of the target data distribution feature to obtain a target magnetic resonance reconstructed image.

[0071] In an optional embodiment, steps S250 and S260 may be implemented by a model that combines a score matching function with an image reconstruction function. Specifically, the mapping and the distribution of the network parameters are input into the model to directly obtain an optimized target magnetic resonance reconstructed image with higher image quality.

[0072] The technical solution of this embodiment is to undersample preset undersampled scan sample data using a preset undersampling operator to obtain a number of sub-undersampled samples; use the sub-undersampled sample data as model training input samples, and use the sample-reconstructed magnetic resonance images corresponding to the preset undersampled scan sample data as sample labels to train a Bayesian convolutional neural network to obtain a Bayesian convolutional neural network model for establishing a mapping relationship between magnetic resonance undersampled k-space data and the magnetic resonance image corresponding to the magnetic resonance undersampled k-space data; construct a fractional matching network based on the network parameter distribution characteristics corresponding to the self-supervised magnetic resonance image reconstruction model; obtain scan data to be reconstructed, and input the scan data to be reconstructed into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain a mapping between the scan data to be reconstructed and a target magnetic resonance image and a distribution of network parameters corresponding to the self-supervised magnetic resonance image reconstruction model; The mapping and the distribution of the network parameters are input into a score matching network constructed based on the network parameter distribution characteristics of the self-supervised magnetic resonance image reconstruction model to obtain a score of the target data distribution characteristics corresponding to the target magnetic resonance image; based on the score of the target data distribution characteristics, the undersampled k-space data is reconstructed using a preset Langevin Markov chain Monte Carlo sampling method to obtain a target magnetic resonance reconstructed image. The technical solution of the embodiment of the present application solves the problem of low image reconstruction quality based on magnetic resonance undersampled data. It can directly reconstruct high-quality magnetic resonance images from high-magnification undersampled k-space data, realizing image reconstruction based on magnetic resonance undersampled data in scenarios without fully sampled labeled data.

[0073] Figure 9 is a flowchart of a magnetic resonance image reconstruction method provided in an embodiment of the present application. This embodiment shares the same concept as the magnetic resonance image reconstruction method described above, illustrating the process of analyzing tissue parameters based on magnetic resonance reconstructed images. This method can be performed by a magnetic resonance image reconstruction device, which can be implemented using software and / or hardware and integrated into a computer device with application development capabilities.

[0074] As shown in FIG9 , the magnetic resonance image reconstruction method of this embodiment includes the following steps.

[0075] S310. Obtain magnetic resonance undersampled k-space data of a target region of interest, and input the magnetic resonance undersampled k-space data into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain a mapping between the magnetic resonance undersampled k-space data and the target magnetic resonance image and a distribution of network parameters corresponding to the self-supervised magnetic resonance image reconstruction model.

[0076] S320: Input the mapping and the distribution of the network parameters into a score matching network constructed based on the network parameter distribution characteristics of the self-supervised magnetic resonance image reconstruction model to obtain a score of the target data distribution characteristics corresponding to the target magnetic resonance image.

[0077] S330 , reconstructing the undersampled k-space data according to the score of the target data distribution feature by using a preset Langevin Markov Chain Monte Carlo sampling method to obtain a target magnetic resonance reconstructed image.

[0078] S340 , performing signal evolution trajectory matching on the signal evolution trajectory formed by the target magnetic resonance reconstructed image and a preset magnetic resonance signal evolution trajectory set to obtain a target magnetic resonance signal evolution trajectory.

[0079] In clinical practice, magnetic resonance images with different contrast ratios can reflect different tissue information. For example, T1-weighted (T1 represents longitudinal relaxation time) can reflect differences in longitudinal tissue relaxation, enabling better visualization of anatomical structures, while T2-weighted (T2 represents transverse relaxation time) can reflect differences in transverse tissue relaxation, enabling better visualization of tissue pathology. When the training sample images for the self-supervised magnetic resonance image reconstruction model are multi-contrast reconstructed images, the target magnetic resonance image ultimately output by the self-supervised magnetic resonance image reconstruction model is also a multi-contrast image. Each pixel in the target magnetic resonance image can contain multiple tissue parameters.

[0080] In this embodiment, data matching can be performed based on the information variation characteristics of each pixel in the target MRI reconstructed image to obtain the target parameters of the tissue corresponding to each pixel, that is, different contrast parameters. The data matching object can be a pre-constructed collection of different contrast signal data.

[0081] Each pixel in the target MRI reconstructed image can be combined with the pixel information of the pixel at the same position in the temporally associated MRI reconstructed image to form a pixel signal evolution trajectory. Furthermore, the signal evolution trajectory formed by the target MRI reconstructed image can be matched with a preset set of MRI signal evolution trajectories to obtain the target MRI signal evolution trajectory.

[0082] S350 : Determine quantization results of the preset parameters of the target magnetic resonance reconstructed image according to the quantization mapping relationship between the target magnetic resonance signal evolution trajectory and the preset parameters.

[0083] The target magnetic resonance signal evolution trajectory is a signal evolution trajectory from a preset set of magnetic resonance signal evolution trajectories, and the contrast parameter corresponding to the target magnetic resonance signal evolution trajectory is known. The quantization result of the preset parameter (any contrast parameter) corresponding to the signal evolution trajectory constituting the target magnetic resonance reconstructed image can be determined by using the quantization mapping relationship between the target magnetic resonance signal evolution trajectory and the preset parameter.

[0084] In one embodiment, as shown in FIG10 , taking T1 and T1ρ (a tissue relaxation time similar to T1, representing T1 in the rotation frame, referred to as T1ρ) as an example, after motion correction is performed on the reconstructed multi-contrast image, a dictionary matching method is used to determine the T1 and T1ρ data corresponding to the pixel signal evolution trajectory of the target magnetic resonance reconstructed image.

[0085] For example, the evolution trajectory of the signal under different inversion times (TI) and spin-lock times (TSL) under a balance steady-state free precession (bSSFP) sequence can be simulated in advance, a data dictionary within a certain range of T1, T2 and T1ρ values ​​can be constructed, and the reconstructed image can be parameter inverted pixel by pixel using a pattern matching method to obtain the T1 and T1ρ values ​​corresponding to the tissue at each pixel.

[0086] The technical solution of this embodiment is to obtain magnetic resonance undersampled k-space data of a target region of interest, and input the undersampled k-space data into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain a mapping between the undersampled k-space data and the target magnetic resonance image and the distribution of corresponding network parameters; input the mapping and the distribution of the network parameters into a score matching network constructed based on the network parameter distribution characteristics of the self-supervised magnetic resonance image reconstruction model to obtain the score of the target data distribution characteristics corresponding to the target magnetic resonance image; based on the score of the target data distribution characteristics, the undersampled k-space data is reconstructed using a preset Langevin Markov chain Monte Carlo sampling method to obtain a target magnetic resonance reconstructed image; signal evolution trajectory matching is performed on the signal evolution trajectory constituted by the target magnetic resonance reconstructed image with a preset magnetic resonance signal evolution trajectory set to obtain a target magnetic resonance signal evolution trajectory; and based on the quantitative mapping relationship between the target magnetic resonance signal evolution trajectory and the preset parameters, the quantization results of the preset parameters of the target magnetic resonance reconstructed image are determined. The technical solution of the embodiment of the present application solves the problem that the multi-parameter quantitative fitting model cannot accurately characterize the signal evolution trajectory. It can directly reconstruct high-quality magnetic resonance images from high-magnification under-sampled k-space data and obtain accurate quantitative values ​​of tissue parameters.

[0087] Figure 11 is a structural schematic diagram of a magnetic resonance image reconstruction device provided in an embodiment of the present application. This embodiment is applicable to scenarios of magnetic resonance image reconstruction, such as obtaining a high-quality magnetic resonance reconstructed image based on undersampled data. The magnetic resonance image reconstruction device can be implemented by software and / or hardware and integrated into a computer terminal device with application development capabilities.

[0088] As shown in FIG11 , the magnetic resonance image reconstruction apparatus includes: a first image reconstruction module 410 , an image distribution feature determination module 420 and a second image reconstruction module 430 .

[0089] Among them, the first image reconstruction module 410 is configured to obtain magnetic resonance undersampled k-space data of a target region of interest, and input the magnetic resonance undersampled k-space data into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain a mapping between the magnetic resonance undersampled k-space data and the target magnetic resonance image and a distribution of network parameters corresponding to the self-supervised magnetic resonance image reconstruction model; the image distribution feature determination module 420 is configured to input the mapping and the distribution of the network parameters into a score matching network constructed based on the network parameter distribution characteristics of the self-supervised magnetic resonance image reconstruction model to obtain a score of the target data distribution feature corresponding to the target magnetic resonance image; the second image reconstruction module 430 is configured to reconstruct the magnetic resonance undersampled k-space data according to the score of the target data distribution feature through a preset Langevin Markov chain Monte Carlo sampling method to obtain a target magnetic resonance reconstructed image.

[0090] The technical solution of this embodiment obtains the scan data to be reconstructed and inputs the scan data to be reconstructed into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain an initial magnetic resonance reconstructed image; inputs the initial magnetic resonance reconstructed image into a fractional matching network constructed based on the data distribution characteristics corresponding to the self-supervised magnetic resonance image reconstruction model to obtain the target data distribution characteristics corresponding to the initial magnetic resonance reconstructed image; and based on the target data distribution characteristics, samples and reconstructs the initial magnetic resonance reconstructed image using a preset Langevin Markov chain Monte Carlo sampling method to obtain a target magnetic resonance reconstructed image. The technical solution of the embodiment of the present application solves the problem of low image reconstruction quality based on magnetic resonance undersampled data and can improve the quality of images reconstructed directly from high-magnification undersampled k-space data.

[0091] In an optional embodiment, the self-supervised magnetic resonance image reconstruction model is a deep learning model obtained by training a Bayesian convolutional neural network model based on preset under-sampling scan sample data and sample reconstructed magnetic resonance images corresponding to the preset under-sampling scan sample data.

[0092] In an optional embodiment, the magnetic resonance image reconstruction apparatus further includes a model training module configured to train a self-supervised magnetic resonance image reconstruction model based on preset undersampled scan sample data and corresponding sample reconstructed magnetic resonance images, including:

[0093] Undersampling the preset undersampled scan sample data using a preset undersampling operator to obtain sub-undersampled sample data;

[0094] The sub-undersampled sample data is used as a model training input sample, and the sample reconstructed magnetic resonance image is used as a sample label to train a Bayesian convolutional neural network model to obtain a Bayesian convolutional neural network model for establishing a mapping relationship between the magnetic resonance undersampled k-space data and the magnetic resonance image corresponding to the magnetic resonance undersampled k-space data.

[0095] In an optional embodiment, the model training module is further configured to:

[0096] Based on the Kullback-Leibler divergence analysis algorithm, a constraint analysis is performed on the model learning loss between the output results of the Bayesian convolutional neural network model and the sample labels corresponding to the output results;

[0097] When the corresponding divergence value in the Kullback-Leibler divergence analysis algorithm meets the preset constraint conditions, the model training process is terminated, and the trained Bayesian convolutional neural network model and the network parameter distribution characteristics corresponding to the trained Bayesian convolutional neural network model are obtained.

[0098] In an optional embodiment, the model training module is further configured to: construct a score matching network based on the network parameter distribution characteristics of the self-supervised magnetic resonance image reconstruction model.

[0099] A score matching network is constructed based on the network parameter distribution characteristics of the self-supervised magnetic resonance image reconstruction model, including:

[0100] gradually adding noise of different scales to the initial magnetic resonance reconstruction images of the sub-undersampled sample data according to the output result of the Bayesian convolutional neural network model using the sub-undersampled sample data as input data and the distribution characteristics of the network parameters to obtain a set of noisy magnetic resonance images;

[0101] training the score matching network according to the set of noisy magnetic resonance images and a preset score matching algorithm to obtain a target score matching network;

[0102] The initial magnetic resonance reconstructed image is an image obtained by performing inverse Fourier transform based on the sub-undersampled sample data.

[0103] In an optional embodiment, the target data distribution feature is the image data distribution probability density corresponding to the target magnetic resonance image.

[0104] In an optional embodiment, the magnetic resonance image reconstruction apparatus further includes a magnetic resonance multi-parameter determination module configured to:

[0105] performing signal evolution trajectory matching on the signal evolution trajectory formed by the target magnetic resonance reconstructed image and a preset magnetic resonance signal evolution trajectory set to obtain a target magnetic resonance signal evolution trajectory;

[0106] According to the quantitative mapping relationship between the target magnetic resonance signal evolution trajectory and the preset parameters, the quantization results of the preset parameters of the target magnetic resonance reconstructed image are determined.

[0107] The magnetic resonance image reconstruction device provided in the embodiments of the present application can execute the magnetic resonance image reconstruction method provided in any embodiment of the present application, and has the functional modules and effects corresponding to the execution method.

[0108] Figure 12 is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. Figure 12 shows a block diagram of an exemplary computer device 12 suitable for implementing the embodiments of the present application. The computer device 12 shown in Figure 12 is merely an example and should not limit the functionality and scope of use of the embodiments of the present application. The computer device 12 can be any terminal device with computing capabilities, such as an intelligent controller, server, mobile phone, or other terminal device.

[0109] 12 , computer device 12 is implemented as a general-purpose computing device. Components of computer device 12 may include: one or more processors or processing units 16 , system memory 28 , and a bus 18 that connects various system components (including system memory 28 and processing units 16 ).

[0110] Bus 18 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processor, or a local bus using any of a variety of bus architectures. Examples of these architectures include the Industry Subversive Alliance (ISA) bus, the Micro Channel Architecture (MCA) bus, an enhanced ISA bus, a Video Electronics Standards Association (VESA) local bus, and a Peripheral Component Interconnect (PCI) bus.

[0111] The computer device 12 includes a variety of computer system readable media. These media can be any available media that can be accessed by the computer device 12, including volatile and non-volatile media, removable and non-removable media.

[0112] System memory 28 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache memory 32. Computer device 12 may include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 34 may be configured to read and write to non-removable, non-volatile magnetic media (not shown in FIG. 12 , commonly referred to as a “hard drive”). Although not shown in FIG. 12 , a magnetic disk drive may be provided for reading and writing to a removable non-volatile magnetic disk (e.g., a “floppy disk”), as well as an optical disk drive for reading and writing to a removable non-volatile optical disk (e.g., a compact disc read-only memory (CD-ROM), a digital versatile disc read-only memory (DVD-ROM), or other optical media). In these cases, each drive may be connected to bus 18 via one or more data media interfaces. The system memory 28 may include at least one program product having a set (eg, at least one) of program modules configured to perform the functions of various embodiments of the present application.

[0113] A program / utility 40 having a set (at least one) of program modules 42 may be stored, for example, in system memory 28. Such program modules 42 may include an operating system, one or more application programs, other program modules, and program data, each of which, or a combination thereof, may include an implementation of a network environment. Program modules 42 generally implement the functions and / or methods of the embodiments described herein.

[0114] The computer device 12 can also communicate with one or more external devices 14 (e.g., a keyboard, pointing device, display 24, etc.), one or more devices that enable a user to interact with the computer device 12, and / or any device that enables the computer device 12 to communicate with one or more other computing devices (e.g., a network card, a modem, etc.). This communication can occur via an input / output (I / O) interface 22. Furthermore, the computer device 12 can communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network such as the Internet) via a network adapter 20. As shown, the network adapter 20 communicates with other modules of the computer device 12 via a bus 18. Although not shown in FIG. 12 , other hardware and / or software modules can be used in conjunction with the computer device 12, including microcode, device drivers, redundant processing units, external disk drive arrays, Redundant Arrays of Independent Drives (RAID) systems, tape drives, and data backup storage systems.

[0115] The processing unit 16 executes various functional applications and data processing by running programs stored in the system memory 28, such as implementing the magnetic resonance image reconstruction method provided by the embodiment of the present invention, which includes:

[0116] Acquiring undersampled magnetic resonance k-space data of a target region of interest, and inputting the undersampled magnetic resonance k-space data into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain a mapping between the undersampled magnetic resonance k-space data and the target magnetic resonance image and a distribution of network parameters corresponding to the self-supervised magnetic resonance image reconstruction model;

[0117] Inputting the mapping and the distribution of the network parameters into a score matching network constructed based on the network parameter distribution characteristics of the self-supervised magnetic resonance image reconstruction model to obtain a score of the target data distribution characteristics corresponding to the target magnetic resonance image;

[0118] According to the score of the target data distribution feature, the magnetic resonance undersampled k-space data is reconstructed by a preset Langevin Markov chain Monte Carlo sampling method to obtain a target magnetic resonance reconstructed image.

[0119] An embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the magnetic resonance image reconstruction method provided in any embodiment of the present application is implemented, the method comprising:

[0120] Acquiring undersampled magnetic resonance k-space data of a target region of interest, and inputting the undersampled magnetic resonance k-space data into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain a mapping between the undersampled magnetic resonance k-space data and the target magnetic resonance image and a distribution of network parameters corresponding to the self-supervised magnetic resonance image reconstruction model;

[0121] Inputting the mapping and the distribution of the network parameters into a score matching network constructed based on the network parameter distribution characteristics of the self-supervised magnetic resonance image reconstruction model to obtain a score of the target data distribution characteristics corresponding to the target magnetic resonance image;

[0122] According to the score of the target data distribution feature, the magnetic resonance undersampled k-space data is reconstructed by a preset Langevin Markov chain Monte Carlo sampling method to obtain a target magnetic resonance reconstructed image.

[0123] The computer storage medium of the embodiment of the present application can adopt any combination of one or more computer-readable media. Computer-readable media can be computer-readable signal media or computer-readable storage media. Computer-readable storage media can be, for example: electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or devices, or any combination of the above. Computer-readable storage media include: electrical connections with one or more wires, portable computer disks, hard disks, RAM, read-only memories (ROM), erasable programmable read-only memories (EPROM), flash memories, optical fibers, portable CD-ROMs, optical storage devices, magnetic storage devices, or any suitable combination of the above. In this document, computer-readable storage media can be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. The storage medium can be a non-transitory storage medium.

[0124] A computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries computer-readable program code. Such propagated data signals may take various forms, including electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transfer a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0125] The program code contained on the computer-readable medium can be transmitted using any appropriate medium, including wireless, wire, optical cable, radio frequency (RF), etc., or any suitable combination of the above.

[0126] The computer program code for performing the operation of the present application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and also conventional procedural programming languages ​​such as "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a LAN or WAN, or can be connected to an external computer (e.g., using an Internet service provider to connect via the Internet).

[0127] The multiple modules or steps of the present application described above can be implemented using a general-purpose computing device. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Alternatively, they can be implemented using program code executable by a computer device, so that they can be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into multiple integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. Thus, the present application is not limited to any specific combination of hardware and software.

Claims

1. A magnetic resonance image reconstruction method, comprising: Obtaining magnetic resonance undersampled k-space data of a target region of interest, and inputting the magnetic resonance undersampled k-space data into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain a mapping between the magnetic resonance undersampled k-space data and a target magnetic resonance image and a distribution of network parameters corresponding to the self-supervised magnetic resonance image reconstruction model; Inputting the mapping and the distribution of network parameters into a score matching network constructed based on the distribution characteristics of the network parameters of the self-supervised magnetic resonance image reconstruction model to obtain a score of the target data distribution characteristics corresponding to the target magnetic resonance image; Reconstructing the magnetic resonance undersampled k-space data by a preset Langevin Markov chain Monte Carlo sampling method according to the score of the target data distribution characteristics to obtain a target magnetic resonance reconstructed image.

2. The method according to claim 1, wherein, The self-supervised magnetic resonance image reconstruction model is a deep learning model obtained by training a Bayesian convolutional neural network model based on preset undersampled scan sample data and a sample reconstructed magnetic resonance image corresponding to the preset undersampled scan sample data.

3. The method according to claim 2, wherein, The training process of the self-supervised magnetic resonance image reconstruction model includes: Undersampling the preset undersampled scan sample data by a preset undersampling operator to obtain sub-undersampled sample data; Using the sub-undersampled sample data as a model training input sample and the sample reconstructed magnetic resonance image as a sample label to train the Bayesian convolutional neural network model to obtain a Bayesian convolutional neural network model for establishing a mapping relationship between magnetic resonance undersampled k-space data and a magnetic resonance image corresponding to the magnetic resonance undersampled k-space data.

4. The method according to claim 3, wherein, The using the sub-undersampled sample data as a model training input sample and the sample reconstructed magnetic resonance image as a sample label to train the Bayesian convolutional neural network model to obtain a Bayesian convolutional neural network model for establishing a mapping relationship between magnetic resonance undersampled k-space data and a magnetic resonance image corresponding to the magnetic resonance undersampled k-space data includes: Based on the Kullback-Leibler divergence analysis algorithm, performing constraint analysis on the model learning loss between the output result of the Bayesian convolutional neural network model and the sample label corresponding to the output result; When the divergence value in the Kullback-Leibler divergence analysis algorithm satisfies a preset constraint condition, ending the model training process to obtain a trained Bayesian convolutional neural network model and a distribution characteristic of network parameters corresponding to the trained Bayesian convolutional neural network model.

5. The method according to claim 4, wherein The process of constructing the score matching network based on the distribution characteristics of the network parameters of the self-supervised magnetic resonance image reconstruction model includes: According to the output result of the Bayesian convolutional neural network model with the sub-undersampled sample data as input data and the distribution characteristics of the network parameters, gradually adding noises with different scales to the initial magnetic resonance reconstruction image of the sub-undersampled sample data to obtain a group of noise-added magnetic resonance images; Training the score matching network according to the set of noisy magnetic resonance images and a preset score matching algorithm to obtain a target score matching network; Wherein, the initial magnetic resonance reconstruction image is an image obtained by performing inverse Fourier transform on the sub-undersampled sample data.

6. The method according to claim 1, wherein The target data distribution feature is the probability density of the image data distribution corresponding to the target magnetic resonance image.

7. The method according to any one of claims 1 to 6, further comprising: Performing signal evolution trajectory matching on the signal evolution trajectory formed by the target magnetic resonance reconstruction image and a preset set of magnetic resonance signal evolution trajectories to obtain a target magnetic resonance signal evolution trajectory; Determining a quantization result of a preset parameter of the target magnetic resonance reconstruction image according to a quantization mapping relationship between the target magnetic resonance signal evolution trajectory and the preset parameter.

8. A magnetic resonance image reconstruction apparatus, comprising: A first image reconstruction module configured to acquire magnetic resonance undersampled k-space data of a target region of interest and input the magnetic resonance undersampled k-space data into a pre-trained self-supervised magnetic resonance image reconstruction model to obtain a mapping between the magnetic resonance undersampled k-space data and a target magnetic resonance image and a distribution of network parameters corresponding to the self-supervised magnetic resonance image reconstruction model; An image distribution feature determination module configured to input the mapping and the distribution of network parameters into a score matching network constructed based on a network parameter distribution feature of the self-supervised magnetic resonance image reconstruction model to obtain a score of a target data distribution feature corresponding to the target magnetic resonance image; A second image reconstruction module configured to reconstruct the magnetic resonance undersampled k-space data by a preset Langevin Markov chain Monte Carlo sampling method according to the score of the target data distribution feature to obtain a target magnetic resonance reconstruction image.

9. A computer device, comprising: One or more processors; A memory configured to store one or more programs; When the one or more programs are executed by the one or more processors, enabling the one or more processors to implement the magnetic resonance image reconstruction method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, wherein, The computer program, when executed by a processor, implements the magnetic resonance image reconstruction method according to any one of claims 1 to 7.

Images