Magnetic resonance image generation method, device, computer equipment and storage medium

Abstract

The present invention discloses a magnetic resonance image generation method, device, computer equipment and storage medium. The magnetic resonance image generation method comprises: collecting K-space data obtained by performing magnetic resonance scanning on a target to be measured; processing based on the K-space data and a reconstruction framework for image reconstruction and parameter fitting to generate an image, wherein the reconstruction framework is a sparse low-rank model embedded with a physical model and a trained neural network; and obtaining a T1 weighted image and a T1 quantized image finally output by the reconstruction framework. By adopting the present invention, accurate T1 weighted images and T1 quantized images can be obtained with high image quality.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and in particular to a method, device, computer equipment and storage medium for generating a magnetic resonance image. Background Art

[0002] As a non-invasive, non-ionizing imaging technique, magnetic resonance imaging (MRI) is widely used in physics, biology, and medicine. Although it can provide high-quality images with various contrast mechanisms, MRI is often limited to qualitative rather than quantitative imaging, and the image contrast of MRI may vary depending on the scanning equipment, imaging protocol, radiofrequency coil, and experimental environment, which hinders quantitative comparisons between different platforms. In recent years, T1 quantitative imaging has gained increasing attention and has been widely used in applications such as evaluating Parkinson's disease in patients with iron absorption, monitoring brain white matter (WM) and gray matter (GM) abnormalities, and identifying high-risk carotid atherosclerotic plaque components. However, traditional T1 mapping sequences require multiple T1-weighted images for parameter fitting, which results in a long scanning time, and long acquisition time may cause motion artifacts, image misalignment, and patient discomfort.

[0003] With the development of technology, a fast three-dimensional T1 quantitative magnetic resonance sequence technology GOAL-SNAP has been proposed. This technology uses inversion recovery pulse (IR) and three-dimensional golden angle radial sampling trajectory, so that it can obtain three-dimensional, self-registered isotropic T1-weighted images in a short time. When reconstructing the image, the signal sharing method is used to merge the adjacent radial K-space data collected in order to reduce the noise of the reconstructed image and improve the contrast. However, the contrast of the image is determined by the average signal of the shared radial K-space data, which produces contrast mixing and affects the subsequent T1 quantitative fitting; moreover, the traditional T1 quantitative fitting method is performed point by point, without considering the spatial correlation between points. The T1 quantitative image obtained by fitting has large noise and the fitting process is time-consuming. Summary of the invention

[0004] The technical problem to be solved by the present invention is that the T1 weighted images and T1 quantized images generated by the traditional technology have low accuracy and poor quality.

[0005] In order to solve the above technical problems, the present invention provides a method, an apparatus, a computer device and a storage medium for generating a magnetic resonance image.

[0006] A method for generating a magnetic resonance image, comprising:

[0007] Acquire K-space data obtained by performing magnetic resonance scanning on the target to be measured;

[0008] Processing based on the K-space data and a reconstruction framework for image reconstruction and parameter fitting to generate an image, wherein the reconstruction framework is a sparse low-rank model embedded with a physical model and a trained neural network;

[0009] The T1-weighted image and T1-quantized image finally output by the reconstruction framework are obtained.

[0010] In one embodiment, before the processing based on the K-space data and the reconstruction framework for image reconstruction and parameter fitting to generate an image, the method further includes:

[0011] The preset physical model and the trained neural network are embedded into the sparse low-rank model to obtain a reconstruction framework.

[0012] In one embodiment, before embedding the preset physical model and the trained neural network into the sparse low-rank model, the method further includes:

[0013] Acquire simulated K-space data from a digital brain phantom;

[0014] Build the initial neural network;

[0015] The initial neural network is trained to map from the T1 weighted image to the T1 quantized image according to the simulated K-space data to obtain the trained neural network.

[0016] In one embodiment, the step of acquiring simulated K-space data from a digital brain phantom includes:

[0017] The T1 quantification values ​​corresponding to different components of the brain digital phantom are collected using a three-dimensional golden radial angle acquisition method;

[0018] Processing the T1 quantization value using a signal formula of a GOAL-SNAP sequence to obtain processed data;

[0019] The processed data is subjected to Fourier transform operation and Gaussian distribution noise is added to generate simulated K-space data.

[0020] In one embodiment, the training of mapping the initial neural network from the T1 weighted image to the T1 quantized image according to the simulated K-space data to obtain the trained neural network includes:

[0021] Performing an inverse irregular Fourier transform on the simulated K-space data to obtain a first type T1-weighted image;

[0022] Inputting the simulated K-space data into a sparse low-rank reconstruction algorithm, taking the reconstructed image of the previous preset iteration to obtain a second type of T1-weighted image, and taking the image after the iteration of the sparse low-rank reconstruction algorithm to obtain a third type of T1-weighted image;

[0023] The simulated K-space data is processed using a signal formula of a GOAL-SNAP sequence, and the obtained image is used as a standard T1 quantization image;

[0024] The first type of T1 weighted image, the second type of T1 weighted image and the third type of T1 weighted image are used as inputs of the initial neural network, the standard T1 quantized image is used as output of the initial neural network, and the initial neural network is trained to obtain the trained neural network.

[0025] In one embodiment, the reconstruction framework includes a sparse low-rank optimization formula, a quantization model introduced into a trained neural network, and a weighted model introduced into a physical model, and the processing based on the K-space data and the reconstruction framework for image reconstruction and parameter fitting to generate an image includes:

[0026] Using the K-space data to perform parameter initialization to obtain initialized parameters, performing sparse low-rank iterative calculation according to the initialized parameters and the sparse low-rank optimization formula to obtain a low-rank part of the next iteration and a sparse part of the next iteration;

[0027] Perform iterative calculation of the quantized image according to the quantization model to obtain the T1 quantized image of the next iteration;

[0028] Performing weighted image iterative calculation according to the K-space data, the low-rank part of the next iteration, the sparse part of the next iteration, the T1 quantized image of the next iteration and the weighted model to obtain a T1 weighted image of the next iteration;

[0029] Repeat the sparse low-rank iterative calculation, the quantized image iterative calculation and the weighted image iterative calculation until the iteration is completed, and obtain the T1 weighted image when the iteration is completed and the T1 quantized image when the iteration is completed, and use them as the final output T1 weighted image and the final output T1 quantized image.

[0030] In one embodiment, the initialization parameters include:

[0031] k=0,S=0,L=u=E H d;

[0032] The low-rank sparse optimization formula includes:

[0033]

[0034]

[0035] The quantitative model includes:

[0036] Q k+1 ≈R k+1 u k ;

[0037] The weighted model includes:

[0038]

[0039] u k+1 = k+1 / 2 - H (Eu k+1 / 2 -d);

[0040] Wherein, k represents the number of iterations, L is the initial low-rank part, S is the initial sparse part, d is the K-space data, u is the initial T1-weighted image, and E is a data conversion operator, including a multi-channel coil data fusion operation and an irregular Fourier transform operation;

[0041] L k+1 is the low-rank part of the k+1th iteration, S k+1 is the sparse part of the k+1th iteration; L k is the low-rank part of the kth iteration, S k is the sparse part of the kth iteration, Indicates the use of parameter λ1 for singular value threshold shrinkage, Indicates soft threshold shrinkage using parameter λ2, T indicates sparse transformation, t indicates iterative step size, u k represents the T1-weighted image of the kth iteration;

[0042] Q k+1 represents the T1 quantized image of the k+1th iteration, R k+1 is the trained neural network of the k+1th iteration, I is the identity matrix, μ is the predetermined parameter, and G is the physical model.

[0043] A magnetic resonance image generating device, comprising:

[0044] A data acquisition module, used for acquiring K-space data obtained by performing magnetic resonance scanning on the target to be measured;

[0045] A data processing module, configured to process the K-space data based on the K-space data and a reconstruction framework for image reconstruction and parameter fitting to generate an image, wherein the reconstruction framework is a low-rank sparse model embedded with a physical model and a trained neural network;

[0046] The image acquisition module is used to acquire the T1 weighted image and T1 quantized image finally output by the reconstruction framework.

[0047] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0048] Acquire K-space data obtained by performing magnetic resonance scanning on the target to be measured;

[0049] Processing based on the K-space data and a reconstruction framework for image reconstruction and parameter fitting to generate an image, wherein the reconstruction framework is a sparse low-rank model embedded with a physical model and a trained neural network;

[0050] The T1-weighted image and T1-quantized image finally output by the reconstruction framework are obtained.

[0051] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0052] Acquire K-space data obtained by performing magnetic resonance scanning on the target to be measured;

[0053] Processing based on the K-space data and a reconstruction framework for image reconstruction and parameter fitting to generate an image, wherein the reconstruction framework is a sparse low-rank model embedded with a physical model and a trained neural network;

[0054] The T1-weighted image and T1-quantized image finally output by the reconstruction framework are obtained.

[0055] Compared with the prior art, one or more embodiments of the above scheme may have the following advantages or beneficial effects:

[0056] The sparse low-rank model embedded with the physical model and the trained neural network is used as the reconstruction framework, and the K-space data is processed by using the reconstruction framework to obtain T1 weighted images and T1 quantized images; compared with the traditional image reconstruction method, the introduction and combination of the prior information of the trained neural network, the sparse low-rank model and the physical model can not only generate accurate T1 quantized images, but also use the sparse low-rank to constrain the spatiotemporal correlation of the T1 weighted image, and reconstruct high-quality T1 weighted images with high image quality. In addition, the introduction of the physical model can connect the learning priors of the sparse low-rank and trained neural network, so that the advantages of both can be obtained at the same time, achieving better performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The scope of the present disclosure may be better understood by reading the following detailed description of exemplary embodiments in conjunction with the accompanying drawings. The drawings included are:

[0058] Figure 1 is a schematic flow chart of a method for generating a magnetic resonance image in one embodiment;

[0059] Figure 2 A schematic diagram of a U network model in one embodiment;

[0060] Figure 3 A schematic diagram of a modular process for generating a magnetic resonance image in one embodiment;

[0061] Figure 4 FIG. 1 is a schematic diagram of the structural framework of a magnetic resonance image generating device in one embodiment. DETAILED DESCRIPTION

[0062] In order to make the objectives, technical solutions and advantages of the present invention clearer, the implementation method of the present invention will be described in detail below in conjunction with the accompanying drawings and embodiments, so that the implementation process of how the present invention applies technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0063] In the prior art, the fast three-dimensional T1 quantitative magnetic resonance sequence technology GOAL-SNAP uses a signal sharing method to merge adjacent radial K-space data when reconstructing images, in order to reduce the noise of the reconstructed image and improve the contrast. However, the contrast of the image is determined by the average signal of the shared radial K-space data, which produces contrast mixing and affects the subsequent T1 quantitative fitting; moreover, the traditional T1 quantitative fitting method is performed point by point, without considering the spatial correlation between points, and the fitted T1 quantitative image has large noise, and the fitting process takes a long time.

[0064] In the past two years, deep learning methods have been widely used in magnetic resonance image reconstruction and physiological quantitative parameter fitting. A method for T2 parameter fitting using deep learning prior, low-rank and sparse models has been proposed. Although this method can generate accurate T2 quantification results, T2 fitting needs to be performed pixel by pixel, and the calculation time is long. An end-to-end convolutional neural network model has been proposed to efficiently reconstruct T2 parameter images using image block-based training and inference; however, this method does not apply feedback mechanisms such as data consistency, which may affect the generalization of the model. In addition, image block-based fitting cannot utilize global spatial correlation, which may affect the accuracy of the fitting results. Some people use end-to-end convolutional neural networks for T2 fitting. This method uses cyclic loss to constrain T2 parameter fitting and data consistency in K space to generate T2 quantitative images from undersampled T2 weighted images. However, this method can only reconstruct T2 quantitative images and is not optimized for undersampled T2 weighted images. Therefore, there is an urgent need for an accurate magnetic resonance image generation method that can accurately reconstruct weighted images and quantitative parameter images through image reconstruction and parameter fitting.

[0065] The present application provides a method for obtaining a T1-weighted image and a T1-quantized image with high accuracy. Figure 1 As shown, in one embodiment, the method for generating a magnetic resonance image comprises the following steps:

[0066] S110: Acquire K-space data obtained by performing magnetic resonance scanning on the target to be measured.

[0067] The target to be measured refers to the target that needs to be subjected to magnetic resonance imaging, such as the head of a patient. K-space data is data obtained by performing magnetic resonance scanning on the target to be measured. Specifically, the K-space data can be obtained by directly performing magnetic resonance scanning on the target to be measured, or the K-space data can be collected from a scanning device, which is used to perform magnetic resonance scanning on the target to be measured.

[0068] S130: Processing based on the K-space data and a reconstruction framework for image reconstruction and parameter fitting to generate an image.

[0069] Among them, the reconstruction framework is a sparse low-rank model embedded with a physical model and a trained neural network. Specifically, the physical model is a computational model that can convert a T1 quantized image into a T1 weighted image, and the trained neural network is a deep learning neural network that can map a T1 weighted image to a T1 quantized image; the physical model and the trained neural network can be regarded as mutually opposite operators. According to the K-space data, the reconstruction framework can perform image reconstruction and parameter fitting, and output T1 quantized images and T1 parameter images through the physical model, the trained neural network and the sparse low-rank model.

[0070] S150: Obtain the T1 weighted image and T1 quantized image finally output by the reconstruction framework.

[0071] Specifically, the T1 weighted image finally output by the reconstruction framework is used as the weighted image of the target to be measured, and the T1 quantized image finally output by the reconstruction framework is used as the quantized parameter image of the target to be measured.

[0072] The above-mentioned magnetic resonance image generation method uses a sparse low-rank model embedded with a physical model and a trained neural network as a reconstruction framework, and obtains a T1 weighted image and a T1 quantized image by processing the K-space data using the reconstruction framework; compared with the traditional image reconstruction method, the prior information of the trained neural network, the sparse low-rank model and the physical model are introduced and combined, which can not only generate an accurate T1 quantized image, but also use the sparse low-rank to constrain the spatiotemporal correlation of the T1 weighted image, and reconstruct a high-quality T1 weighted image with high image quality. In addition, the introduction of the physical model can connect the learning priors of the sparse low-rank and trained neural network, so that the advantages of both can be obtained at the same time, achieving better performance.

[0073] In one of the embodiments, before step S130, a framework construction step is also included: a preset physical model and a trained neural network are embedded in a sparse low-rank model to obtain a reconstructed framework.

[0074] For example, a physical model and a trained neural network may be added to a sparse low-rank model to obtain a new model as a reconstruction framework. Specifically, the framework construction step may be performed before step S110, and it is understood that the framework construction step may also be set to be performed after step S110 according to requirements.

[0075] In one embodiment, the framework building step further includes steps (a1) to (a3).

[0076] Step (a1): Acquire simulated K-space data from the brain digital phantom.

[0077] Among them, the digital brain phantom refers to a simulated digital brain model, for example, it can be a three-dimensional image of a digital brain.

[0078] Step (a2): Construct the initial neural network.

[0079] The initial neural network refers to a neural network that has not been trained or learned.

[0080] Step (a3): Train the initial neural network to map the T1 weighted image to the T1 quantized image based on the simulated K-space data to obtain a trained neural network.

[0081] The initial neural network is trained by using simulated K-space data so that the initial neural network learns the mapping from T1 weighted images to T1 quantized images, thereby obtaining a trained neural network that can map T1 weighted images to T1 quantized images.

[0082] In one embodiment, step (a1) includes steps (a11) to (a13).

[0083] Step (a11): using a three-dimensional golden radial angle acquisition method to acquire T1 quantification values ​​corresponding to different components of the brain digital phantom.

[0084] The different components of the brain digital phantom include gray matter and white matter. Specifically, the gray matter and white matter of the brain digital phantom are assigned corresponding T1 quantization values, and the data are collected in conjunction with the three-dimensional golden radial angle acquisition method.

[0085] Step (a12): The T1 quantization value is processed using the signal formula of the GOAL-SNAP sequence to obtain processed data.

[0086] Step (a13): Perform Fourier transform operation on the processed data and add Gaussian distribution noise to generate simulated K-space data.

[0087] The operation of adding Gaussian distribution noise is a data processing operation, for example, it can be to add noise value to the data obtained by Fourier transform. By cooperating with the three-dimensional golden radial angle acquisition method, using the signal formula of the GOAL-SNAP sequence and Fourier transform, Gaussian distribution noise is added to generate simulated K-space data with high accuracy.

[0088] In one embodiment, the initial neural network is a U network, specifically a U-Net convolutional neural network. The U network is used as the basic network, and cross-layer connections are added inside each encoding module, decoding module, and bridge module. The convolution kernel size of the encoding module is 3*3, the pooling operation is maximum pooling, and the activation function is ReLU; the convolution kernel size of the decoding module is 3*3, the upsampling operation is transposed convolution, the convolution kernel size is 3*3, and the activation function is ReLU; the convolution kernel size of the bridge module is 3*3, and the activation function is ReLU. The encoding module and the decoding module have 4 layers respectively, and the corresponding encoding module and decoding module use splicing operations to merge feature maps. The number of feature maps output at each level is 32, 64, 128, and 256.

[0089] For example, Figure 2 The U network model with cross-layer connection is shown. 201 is the feature map concatenation operation between the corresponding encoding module and the decoding module, 202 is a four-layer encoding layer, 203 is a four-layer decoding layer, and 204 is a bridge layer. The encoding layer is composed of a 3*3 convolution layer, a batch normalization layer, and a ReLU activation function as a basic unit, which is repeated twice, and then a cross-layer connection derived from the input is added, that is, 205, and finally connected to the maximum pooling layer; the decoding layer is initially connected to the upsampling operation, which is a transposed convolution layer, and then a 3*3 convolution layer, a batch normalization layer, and a ReLU activation function are composed of a basic unit, which is repeated twice, and then a cross-layer connection derived from the input is added; the bridge layer is similar to the encoding layer, only the maximum pooling layer is reduced. The number of output feature maps corresponding to each layer of the encoding layer and the decoding layer is 32, 64, 128, and 256.

[0090] Specifically, step (a3) ​​can be to obtain a T1 weighted image and a T1 quantized image respectively according to the simulated K-space data, and use the T1 weighted image corresponding to the simulated K-space data as the input of the initial neural network and the T1 quantized image corresponding to the simulated K-space data as the output of the initial neural network to train the initial neural network.

[0091] Further, in one embodiment, step (a3) ​​may include steps (a31) to (a34).

[0092] Step (a31): performing an inverse irregular Fourier transform on the simulated K-space data to obtain a first type T1-weighted image.

[0093] The simulated K-space data is subjected to inverse irregular Fourier transform to obtain a T1-weighted image with artifacts and noise as the first type of T1-weighted image.

[0094] Step (a32): Input the simulated K-space data into the sparse low-rank reconstruction algorithm, take the reconstructed image of the previous preset iteration to obtain the second type T1 weighted image, and take the image after the sparse low-rank reconstruction algorithm is iterated to obtain the third type T1 weighted image.

[0095] Specifically, the preset number of times can be 6. For example, the simulated K-space data is input into the sparse low-rank reconstruction algorithm, and the reconstructed images of the first 6 iterations are taken, and these reconstructed T1-weighted images contain artifacts and noise as the second type of T1-weighted images. Similarly, the high-quality images after the sparse low-rank reconstruction algorithm is iterated are taken as the third type of T1-weighted images.

[0096] Step (a33): The simulated K-space data is processed using the signal formula of the GOAL-SNAP sequence, and the obtained image is used as a standard T1 quantization image.

[0097] The simulated K-space data were processed according to the signal formula of the GOAL-SNAP sequence to obtain an artifact-free and noise-free image as a standard T1 quantization image.

[0098] Step (a34): using the first type T1 weighted image, the second type T1 weighted image and the third type T1 weighted image as inputs of the initial neural network, using the standard T1 quantized image as output of the initial neural network, and training the initial neural network to obtain a trained neural network.

[0099] Specifically, the back propagation algorithm can be used to update the weights of the neural network to complete the training of the neural network. The loss function of the neural network is selected as a linear weighted combination of the 1-norm and 2-norm of the difference between the labeled data (standard T1 quantized image) and the network output, with weights of 1 and 0.1 respectively.

[0100] By processing the simulated K-space data in various ways, when converting the simulated K-space data to images, various types of images are obtained, with images with artifacts and noise as input and images without artifacts and noise as standard output, so that the initial neural network can not only learn the mapping from T1 weighted images to T1 quantization images, but also learn the ability to remove artifacts and noise. Through the learning and training of the neural network, the neural network prior information is introduced, providing the ability to remove artifacts and noise, greatly shortening the number of iteration cycles and accelerating the reconstruction speed.

[0101] In one embodiment, the reconstruction framework includes a sparse low-rank optimization formula, a quantization model that introduces a trained neural network, and a weighted model that introduces a physical model. The quantization model is a model for generating T1 quantized images, and the weighted model is a model for generating T1 weighted images; the quantization model introduces a trained neural network, which means that the calculation formula of the quantization model includes the trained neural network, and the physical model is introduced into the weighted model, which means that the calculation formula of the weighted model includes the physical model. Specifically, step S130 includes steps (b1) to (b4).

[0102] Step (b1): Use K-space data to perform parameter initialization to obtain initialized parameters, perform sparse low-rank iterative calculation based on the initialized parameters and the sparse low-rank optimization formula to obtain the low-rank part of the next iteration and the sparse part of the next iteration.

[0103] Taking the initial number of iterations as 0 as an example, according to the initialized parameters and the sparse low-rank optimization formula, the low-rank part of the first iteration and the sparse part of the first iteration are calculated.

[0104] Step (b2): Perform iterative calculation of the quantized image according to the quantization model to obtain the T1 quantized image of the next iteration.

[0105] Step (b3): ​​performing weighted image iterative calculation based on the K-space data, the low-rank part of the next iteration, the sparse part of the next iteration, the T1 quantized image of the next iteration and the weighted model to obtain the T1 weighted image of the next iteration.

[0106] Step (b4): Repeat the sparse low-rank iterative calculation, the quantized image iterative calculation and the weighted image iterative calculation until the iteration is completed, and obtain the T1 weighted image when the iteration is completed and the T1 quantized image when the iteration is completed, and use them as the final output T1 weighted image and the final output T1 quantized image.

[0107] That is, the T1 weighted image and the T1 quantized image finally output by the reconstruction framework are the iteratively completed T1 weighted image and T1 quantized image. Specifically, repeating the sparse low-rank iterative calculation is to substitute the low-rank part obtained in the previous time and the sparse part obtained in the previous time into the sparse low-rank optimization formula for sparse low-rank iterative calculation to obtain the low-rank part of the next iteration and the sparse part of the next iteration. By using the sparse low-rank optimization formula, quantization model and weighted model for iterative calculation to obtain the final output T1 quantized image and T1 weighted image, compared to the end-to-end deep neural network reconstruction, this embodiment uses a physical model to convert the high-quality T1 quantized map obtained by mapping into a T1 weighted image, and fuse it with the intermediate result image of the sparse low-rank iterative algorithm to further improve the reconstruction effect and stability.

[0108] According to the research of the inventors, the optimization objective function of the traditional sparse low-rank model is Formula 1:

[0109] min(‖L‖ * + ‖TS‖1),stE(L+S)=d

[0110] Among them, L is the low-rank part, S is the sparse part, T is the sparse transform, which is the Fourier transform in the time dimension, E is the data conversion operator, including multi-channel coil data fusion operation and irregular Fourier transform operation, and d is the K-space data. This application introduces deep neural network prior information and physical model constraints into the sparse low-rank model, and the new optimization objective function is Formula 2:

[0111]

[0112] Among them, G is a physical model that can convert T1 quantized images into T1 weighted images, and R is a deep neural network that can map T1 weighted images into T1 quantized images. G and R can be regarded as opposite operators. Auxiliary variables are introduced and mathematical solutions are performed.

[0113] Formula 2 introduces the auxiliary variable u, where the auxiliary variable u refers to the T1-weighted image, and formula 3 is obtained:

[0114]

[0115] stL+S=u

[0116] The unconstrained form of Equation 3 is Equation 4:

[0117]

[0118] Formula 4 can be split into three sub-problems to solve, as shown in Formula 5:

[0119]

[0120]

[0121]

[0122] Among them, the first two equations can be solved using the iterative shrinkage threshold algorithm, as shown in Formula 6:

[0123]

[0124]

[0125] in, represents the use of parameter λ1 for singular value threshold shrinkage, It represents the soft threshold shrinkage using parameter λ2, where λ1 and λ2 are parameters, which are set to 0.01 and 0.01 here.

[0126] The solution of the third equation of Formula 5 can introduce an auxiliary variable Q, which refers to the T1 quantized image.

[0127] As shown in formula 7:

[0128]

[0129] stRu=Q

[0130] Converted to unconstrained form, as shown in Formula 7:

[0131]

[0132] In order to solve R, u and Q, three sub-problems can be obtained as shown in Formula 8:

[0133]

[0134]

[0135]

[0136] The solution to the first sub-problem of Formula 8 can be regarded as the training process of a deep neural network, as shown in Formula 9:

[0137] R k+1 =NET

[0138] NET is the trained neural network.

[0139] The solution to the second sub-problem can be obtained by taking the derivative of the original problem and setting the derivative to 0, as shown in the following example:

[0140] Formula 10:

[0141]

[0142] Making mathematical approximations, (EG) H EG≈I,G H E H d≈Q≈R k+1 u k , I is the identity matrix, then Formula 10 can be simplified to:

[0143] Q k+1 ≈R k+1 u k

[0144] Similarly, the solution to the third sub-problem can also be obtained by differentiating the original problem and setting the derivative to 0, as shown in Formula 11:

[0145]

[0146] R k+1 H It can be approximately replaced by G, and mathematically approximated to obtain R k+1 H R k+1 ≈I. In order to increase the stability of the iteration, the constraint of data consistency is added, and Formula 11 can be simplified to:

[0147]

[0148] u k+1 =u k+1 / 2 -E H (Eu k+1 / 2 -d)

[0149] Among them, μ is a parameter, which can be set to 1 here.

[0150] After the above research, the present application proposes a reconstruction framework, including Formula 6, Formula 10 and Formula 11. Specifically, in one embodiment, the initialization parameters include:

[0151] k=0,S=0,L=u=E H d;

[0152] The low-rank sparse optimization formula includes:

[0153]

[0154]

[0155] Quantitative models include:

[0156] Q k+1 ≈R k+1 u k ;

[0157] The weighted model includes:

[0158]

[0159] u k+1 =u k+1 / 2 -E H (Eu k+1 / 2 -d);

[0160] Where k represents the number of iterations, L is the initial low-rank part, S is the initial sparse part, d is the K-space data, u is the initial T1-weighted image, and E is the data conversion operator, including multi-channel coil data fusion operation and irregular Fourier transform operation. k+1 is the low-rank part of the k+1th iteration, S k+1 is the sparse part of the k+1th iteration; L k is the low-rank part of the kth iteration, S k is the sparse part of the kth iteration, Indicates the use of parameter λ1 for singular value threshold shrinkage, Indicates that the soft threshold shrinkage is performed using parameter λ2, T indicates sparse transformation, t indicates the iteration step size, which can be set to 1, u k represents the T1-weighted image of the kth iteration; u k+1 is the T1-weighted image of the k+1th iteration, Q k+1 represents the T1 quantized image of the k+1th iteration, R k+1 is the trained neural network of the k+1th iteration, I is the identity matrix, μ is a predetermined parameter, and G is the physical model. Specifically, in this embodiment, μ can be set to 1.

[0161] Specifically, step (b1) uses the low-rank sparse optimization formula to obtain L k+1 and S k+1 , using the calculation formula of the quantitative model to get Q k+1 , use the calculation formula of the weighted model to get u k+1 , determine whether the convergence condition is reached, if not, continue iterative optimization, if yes, end the iterative process and output the T1 quantized image Q k+1 and T1-weighted image u k+1 By introducing the data consistency term: u k+1 =u k+1 / 2 -E H (Eu k+1 / 2 -d), increasing the algorithm stability and improving the generalization ability of the entire reconstruction framework.

[0162] Take an application example to illustrate. Figure 3As shown, the input of the reconstruction framework is K-space data, i.e., 101. Then, according to formula 6, the low-rank part, i.e., 102, and the sparse part, i.e., 103, are iteratively optimized. Formula 10 is used to input the trained neural network, i.e., 104, to obtain a T1 quantized image. Finally, formula 11 is used to complete an iterative process through a physical model, i.e., 105. An iterative module, i.e., 106, is repeated k times to obtain a high-quality T1 weighted image and a T1 quantized image output, i.e., 107.

[0163] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple steps or multiple stages. These steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed in turn or alternately with other steps or at least part of the steps or stages in other steps.

[0164] In one embodiment, a magnetic resonance image generating apparatus is provided, such as Figure 4 As shown, the magnetic resonance image generating device includes: a data acquisition module 410 , a data processing module 430 and an image acquisition module 450 .

[0165] The data acquisition module 410 is used to acquire K-space data obtained by performing magnetic resonance scanning on the target to be measured; the data processing module 430 is used to process the K-space data and a reconstruction framework for image reconstruction and parameter fitting to generate an image, and the reconstruction framework is a low-rank sparse model embedded with a physical model and a trained neural network; the image acquisition module 450 is used to obtain the T1 weighted image and T1 quantized image that are finally output by the reconstruction framework.

[0166] The magnetic resonance image generation device uses a sparse low-rank model embedded with a physical model and a trained neural network as a reconstruction framework, and processes the K-space data using the reconstruction framework to obtain a T1 weighted image and a T1 quantized image; compared with the traditional image reconstruction method, the prior information of the trained neural network, the sparse low-rank model and the physical model are introduced and combined, which can not only generate an accurate T1 quantized image, but also use the sparse low-rank to constrain the spatiotemporal correlation of the T1 weighted image, and reconstruct a high-quality T1 weighted image with high image quality. In addition, the introduction of the physical model can connect the learning priors of the sparse low-rank and trained neural network, so that the advantages of both can be obtained at the same time, achieving better performance.

[0167] In one of the embodiments, the magnetic resonance image generating apparatus further comprises a framework establishing module for embedding a preset physical model and a trained neural network into a sparse low-rank model to obtain a reconstruction framework.

[0168] In one of the embodiments, the magnetic resonance image generating device further includes a network training module for acquiring simulated K-space data from a digital brain phantom; constructing an initial neural network; and training the initial neural network to map from a T1 weighted image to a T1 quantized image based on the simulated K-space data to obtain a trained neural network.

[0169] In one embodiment, the network training module obtains simulated K-space data from the digital brain phantom, specifically including: the network training module uses a three-dimensional golden radial angle acquisition method to collect T1 quantization values ​​corresponding to different components of the digital brain phantom, uses the signal formula of the GOAL-SNAP sequence to process the T1 quantization values ​​to obtain processed data, performs Fourier transform operation on the processed data and adds Gaussian distribution noise to generate simulated K-space data.

[0170] In one of the embodiments, the network training module trains the initial neural network to map the T1 weighted image to the T1 quantized image based on the simulated K-space data, specifically including: performing an inverse irregular Fourier transform on the simulated K-space data to obtain a first type of T1 weighted image; inputting the simulated K-space data into a sparse low-rank reconstruction algorithm, taking the reconstructed image of the previous preset iteration to obtain a second type of T1 weighted image, and taking the image after the sparse low-rank reconstruction algorithm is iterated to obtain a third type of T1 weighted image; processing the simulated K-space data using the signal formula of the GOAL-SNAP sequence, and using the obtained image as a standard T1 quantized image; using the first type of T1 weighted image, the second type of T1 weighted image, and the third type of T1 weighted image as the input of the initial neural network, and using the standard T1 quantized image as the output of the initial neural network, and training the initial neural network to obtain a trained neural network.

[0171] In one embodiment, the reconstruction framework includes a sparse low-rank optimization formula, a quantization model that introduces a trained neural network, and a weighted model that introduces a physical model. The data processing module 430 is used to: use K-space data to perform parameter initialization to obtain initialized parameters, perform sparse low-rank iterative calculations based on the initialized parameters and the sparse low-rank optimization formula to obtain the low-rank part of the next iteration and the sparse part of the next iteration; perform quantized image iterative calculations based on the quantization model to obtain the T1 quantized image of the next iteration; perform weighted image iterative calculations based on K-space data, the low-rank part of the next iteration, the sparse part of the next iteration, the T1 quantized image of the next iteration, and the weighted model to obtain the T1 weighted image of the next iteration; repeat the sparse low-rank iterative calculations, the quantized image iterative calculations, and the weighted image iterative calculations until the iteration is completed, and obtain the T1 weighted image when the iteration is completed and the T1 quantized image when the iteration is completed, and use them as the final output T1 weighted image and the final output T1 quantized image.

[0172] In one embodiment, the initialization parameters include:

[0173] k=0,S=0,L=u=E H d;

[0174] The low-rank sparse optimization formula includes:

[0175]

[0176]

[0177] Quantitative models include:

[0178] Q k+1 ≈R k+1 u k ;

[0179] The weighted model includes:

[0180]

[0181] u k+1 =u k+1 / 2 -E H (Eu k+1 / 2 -d);

[0182] For the specific definition of the magnetic resonance image generating device, please refer to the definition of the magnetic resonance image generating method above, which will not be repeated here. Each module in the above-mentioned magnetic resonance image generating device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules. It should be noted that the division of modules in the embodiment of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation.

[0183] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps of the method in each of the above embodiments when executing the computer program.

[0184] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method in the above embodiments are implemented.

[0185] The above-mentioned computer device and computer-readable storage medium can implement the steps of the methods in the above-mentioned embodiments, and similarly, can generate accurate and high-quality T1-weighted images and T1 quantized images.

[0186] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0187] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0188] Although the embodiments disclosed in the present invention are as above, the contents described are only embodiments adopted for facilitating the understanding of the present invention and are not intended to limit the present invention. Any technician in the technical field to which the present invention belongs can make any modifications and changes in the form and details of the implementation without departing from the spirit and scope disclosed in the present invention, but the protection scope of the present invention shall still be subject to the scope defined in the attached claims.

Claims

1. A method for generating a magnetic resonance image, characterized in that: include: Inputting K-space data, wherein the K-space data is obtained by performing magnetic resonance scanning on a target to be measured; Processing based on the K-space data and a reconstruction framework for image reconstruction and parameter fitting to generate an image, wherein the reconstruction framework is a sparse low-rank model embedded with a physical model for converting a quantized image into a weighted image and a trained neural network; Obtaining a T1-weighted image and a T1-quantized image finally output by the reconstruction framework; The training process of the trained neural network includes: Acquiring simulated K-space data from a digital brain phantom; specifically comprising: acquiring T1 quantization values ​​corresponding to different components of the digital brain phantom using a three-dimensional golden radial angle acquisition method; processing the T1 quantization values ​​using a signal formula of a GOAL-SNAP sequence to obtain processed data; performing a Fourier transform operation on the processed data and adding Gaussian distribution noise to generate simulated K-space data; Build the initial neural network; The initial neural network is trained to map from T1 weighted image to T1 quantized image according to the simulated K-space data to obtain the trained neural network; this step specifically includes: performing inverse irregular Fourier transform on the simulated K-space data to obtain a first type of T1 weighted image; inputting the simulated K-space data into a sparse low-rank reconstruction algorithm, taking the reconstructed image of the previous preset iteration to obtain a second type of T1 weighted image, and taking the image after the iteration of the sparse low-rank reconstruction algorithm to obtain a third type of T1 weighted image; processing the simulated K-space data using the signal formula of the GOAL-SNAP sequence, and using the obtained image as a standard T1 quantized image; using the first type of T1 weighted image, the second type of T1 weighted image and the third type of T1 weighted image as inputs of the initial neural network, and using the standard T1 quantized image as output of the initial neural network, and training the initial neural network to obtain the trained neural network.

2. The method according to claim 1, characterized in that Before the processing based on the K-space data and the reconstruction framework for image reconstruction and parameter fitting to generate an image, the method further includes: The preset physical model and the trained neural network are embedded into the sparse low-rank model to obtain a reconstruction framework.

3. The method according to claim 1, characterized in that The reconstruction framework includes a sparse low-rank optimization formula, a quantization model that introduces a trained neural network, and a weighted model that introduces a physical model. The processing based on the K-space data and the reconstruction framework for image reconstruction and parameter fitting to generate an image includes: Using the K-space data to perform parameter initialization to obtain initialized parameters, performing sparse low-rank iterative calculation according to the initialized parameters and the sparse low-rank optimization formula to obtain a low-rank part of the next iteration and a sparse part of the next iteration; Perform iterative calculation of the quantized image according to the quantization model to obtain the T1 quantized image of the next iteration; Performing weighted image iterative calculation according to the K-space data, the low-rank part of the next iteration, the sparse part of the next iteration, the T1 quantized image of the next iteration and the weighted model to obtain a T1 weighted image of the next iteration; Repeat the sparse low-rank iterative calculation, the quantized image iterative calculation and the weighted image iterative calculation until the iteration is completed, and obtain the T1 weighted image when the iteration is completed and the T1 quantized image when the iteration is completed, and use them as the final output T1 weighted image and the final output T1 quantized image.

4. The method according to claim 3, characterized in that The initialization parameters include: k=0,S=0,L=u=E H d; The low-rank sparse optimization formula includes: The quantitative model includes: Q k+1 ≈R k+1 u k ; The weighted model includes: u k+1 =u k+1 / 2 -AND H (I k+1 / 2 -d); Wherein, k represents the number of iterations, L is the initial low-rank part, S is the initial sparse part, d is the K-space data, u is the initial T1-weighted image, and E is a data conversion operator, including a multi-channel coil data fusion operation and an irregular Fourier transform operation; L k+1 is the low-rank part of the k+1th iteration, S k+1 is the sparse part of the k+1th iteration; L k is the low-rank part of the kth iteration, S k is the sparse part of the kth iteration, Indicates the use of parameter λ1 for singular value threshold shrinkage, Indicates soft threshold shrinkage using parameter λ2, T indicates sparse transformation, t indicates iterative step size, u k represents the T1-weighted image of the kth iteration; Q k+1 represents the T1 quantized image of the k+1th iteration, R k+1 is the trained neural network of the k+1th iteration, I is the identity matrix, μ is the predetermined parameter, and G is the physical model.

5. A magnetic resonance image generating device, characterized in that: include: A data acquisition module, used for acquiring K-space data obtained by performing magnetic resonance scanning on the target to be measured; A data processing module, configured to process the K-space data based on the K-space data and a reconstruction framework for image reconstruction and parameter fitting to generate an image, wherein the reconstruction framework is a low-rank sparse model embedded with a physical model for converting a quantized image into a weighted image and a trained neural network; An image acquisition module, used to acquire the T1 weighted image and T1 quantized image finally output by the reconstruction framework; The training process of the trained neural network includes: Acquiring simulated K-space data from a digital brain phantom; specifically comprising: acquiring T1 quantization values ​​corresponding to different components of the digital brain phantom using a three-dimensional golden radial angle acquisition method; processing the T1 quantization values ​​using a signal formula of a GOAL-SNAP sequence to obtain processed data; performing a Fourier transform operation on the processed data and adding Gaussian distribution noise to generate simulated K-space data; Build the initial neural network; The initial neural network is trained to map from T1 weighted image to T1 quantized image according to the simulated K-space data to obtain the trained neural network; this step specifically includes: performing inverse irregular Fourier transform on the simulated K-space data to obtain a first type of T1 weighted image; inputting the simulated K-space data into a sparse low-rank reconstruction algorithm, taking the reconstructed image of the previous preset iteration to obtain a second type of T1 weighted image, and taking the image after the iteration of the sparse low-rank reconstruction algorithm to obtain a third type of T1 weighted image; processing the simulated K-space data using the signal formula of the GOAL-SNAP sequence, and using the obtained image as a standard T1 quantized image; using the first type of T1 weighted image, the second type of T1 weighted image and the third type of T1 weighted image as inputs of the initial neural network, and using the standard T1 quantized image as output of the initial neural network, and training the initial neural network to obtain the trained neural network.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 4 are implemented.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images