WO2023050249A1

WO2023050249A1 - Magnetic resonance imaging method and system based on deep learning, and terminal and storage medium

Info

Publication number: WO2023050249A1
Application number: PCT/CN2021/122025
Authority: WO
Inventors: 梁栋; 程静; 贾森; 郑海荣; 刘新
Original assignee: 深圳先进技术研究院
Priority date: 2021-09-30
Filing date: 2021-09-30
Publication date: 2023-04-06

Abstract

A magnetic resonance imaging method and system based on deep learning, and a terminal and a storage medium. The method comprises: performing under-sampling on magnetic resonance fully sampled k-space data, so as to generate under-sampled k-space data; according to the under-sampled k-space data, estimating a SPIRiT convolutional kernel that fills a k-space; inputting the under-sampled k-space data and the SPIRiT convolutional kernel into a convolutional neural network for training, so as to obtain a trained image reconstruction model based on the k-space; and performing magnetic resonance image reconstruction by means of the trained image reconstruction model based on the k-space. Image reconstruction is performed by using an image reconstruction model based on a k-space, and a priori information of an image is learned by using the convolutional neural network, without the need to estimate sensitivity information of a coil, such that the reconstruction time can be greatly reduced, and a better image reconstruction effect can be obtained.

Description

Magnetic resonance imaging method, system, terminal and storage medium based on deep learning

technical field

The present application belongs to the technical field of magnetic resonance imaging, and in particular relates to a deep learning-based magnetic resonance imaging method, system, terminal and storage medium.

Background technique

Magnetic resonance imaging uses static magnetic field and radio frequency magnetic field to image human tissue. It not only provides rich tissue contrast, but also is harmless to human body. It has become a powerful tool in clinical medicine. However, the slow imaging speed has always been a major bottleneck restricting the rapid development of magnetic resonance imaging. How to increase the scanning speed and reduce the scanning time while ensuring the imaging quality is particularly important.

In terms of fast imaging, the traditional techniques are parallel imaging and compressed sensing. Parallel imaging uses the correlation between multi-channel coils to accelerate acquisition, but limited by hardware and other conditions, the acceleration is limited, and with the increase of the acceleration, the image will appear noise amplification phenomenon. Compressed sensing uses the prior information of the sparsity of the imaged object to reduce the k-space sampling points. Due to the iterative reconstruction, the reconstruction time is very long, and it is difficult to choose sparse transformation and reconstruction parameters.

Deep learning methods use neural networks to learn the optimal parameters required for reconstruction from a large amount of training data or directly learn the mapping relationship between under-sampled data and fully-sampled images, thereby achieving better imaging than parallel imaging or compressed sensing methods Quality and higher speedup. The current deep learning reconstruction method mainly uses the image-based SENSE reconstruction model, which needs to estimate the sensitivity information of the coil in advance, and the coil sensitivity information is often inaccurately estimated in places where the phase changes sharply on the image, making it difficult to remove aliasing artifacts or Residual phase singularity; In addition, for the case of image convolutions, the conventionally used coil sensitivity distribution map cannot fully represent the coil distribution, and multiple coil sensitivity distribution maps need to be estimated, which greatly increases the reconstruction time and calculation and storage pressure.

Contents of the invention

The present application provides a deep learning-based magnetic resonance imaging method, system, terminal, and storage medium, aiming to solve one of the above-mentioned technical problems in the prior art at least to a certain extent.

In order to solve the above problems, the application provides the following technical solutions:

A method of magnetic resonance imaging based on deep learning, comprising:

Under-sampling the full-acquisition k-space data of magnetic resonance to generate under-acquisition k-space data;

Estimating the SPIRiT convolution kernel filling k space according to the under-collected k space data;

The under-collected k-space data and the SPIRiT convolution kernel input convolutional neural network are trained to obtain a trained image reconstruction model based on k-space;

The magnetic resonance image reconstruction is performed through the trained image reconstruction model based on k-space.

The technical solution adopted in the embodiment of the present application also includes: the undersampling of the magnetic resonance full-acquisition k-space data also includes:

Obtain the MRI full-acquisition k-space data from the MRI scanner;

Multiplying the magnetic resonance full-acquisition k-space data by a mask matrix to generate multi-channel under-acquisition k-space data.

The technical solution adopted in the embodiment of the present application further includes: the SPIRiT convolution kernel estimated to fill the k-space according to the undersampled k-space data includes;

Carry out SPIRiT operation to described under-acquisition K space data, estimate the SPIRiT convolution core of filling k space;

The SPIRiT operation is to reconstruct unsampled data points with the linear weighting factor calculated by the calibration data in the under-acquisition K-space data, the linear weighting factor is the SPIRiT convolution kernel, and the calibration data is the under-acquisition K-space The low-frequency full mining part of the data center;

The calculation formula of the SPIRiT convolution kernel is:

Among them, m represents the mth coil, x is the k-space data point, r represents the position, R _r is the operation of selecting all k-space points in a certain neighborhood around the position r, and g _jm represents the points obtained from all the points in the neighborhood around the position r weight vector,

is the conjugate transpose of g _jm .

The technical solution adopted in the embodiment of the present application also includes: the image reconstruction model based on k-space is:

where x is the k-space to be reconstructed, D is the undersampling mode, y is the undersampling k-space data, G is the SPIRiT operation, λ ₁ and λ ₂ are penalty parameters, and R(x) is a penalty function related to prior information .

The technical solution adopted in the embodiment of the present application also includes: the described under-acquisition k-space data and the SPIRiT convolution kernel are input into the convolutional neural network for training, and the trained image reconstruction model based on k-space includes:

Using an optimization algorithm to iteratively solve the image reconstruction model based on k-space, the solution process is:

Where n represents the number of iterations, F represents the Fourier transform; prox _{R, τ} represents the approximation operator, which is defined as:

The technical solution adopted in the embodiment of the present application also includes: said inputting the under-collected k-space data and the SPIRiT convolution kernel into the convolutional neural network for training, and obtaining the trained image reconstruction model based on k-space also includes:

The approximation operator prox _{R, τ} related to the penalty function R(x) is fitted by a convolutional neural network, and the reconstruction process is expressed as:

where Λ is a convolutional neural network.

The technical solution adopted in the embodiment of the present application also includes: the loss function of the convolutional neural network is:

Among them, m _rec is the image output by the convolutional neural network, and m _ref is the gold standard image generated by the full sampling k-space data.

Another technical solution adopted in the embodiment of the present application is: a magnetic resonance imaging system based on deep learning, including:

Undersampling module: used for undersampling the full-acquisition k-space data of magnetic resonance to generate under-acquisition k-space data;

SPIRiT processing module: for estimating the SPIRiT convolution kernel filling k space according to the under-acquisition K space data;

Model training module: for the input convolutional neural network of described under-acquisition k-space data and SPIRiT convolution kernel to train, obtain the image reconstruction model based on k-space trained well;

Image reconstruction module: for performing magnetic resonance image reconstruction through the trained k-space-based image reconstruction model.

Another technical solution adopted by the embodiment of the present application is: a terminal, the terminal includes a processor and a memory coupled to the processor, wherein,

The memory stores program instructions for implementing the deep learning-based magnetic resonance imaging method;

The processor is configured to execute the program instructions stored in the memory to control deep learning-based magnetic resonance imaging.

Another technical solution adopted in the embodiment of the present application is: a storage medium storing program instructions executable by a processor, and the program instructions are used to execute the deep learning-based magnetic resonance imaging method.

Compared with the prior art, the beneficial effect of the embodiment of the present application lies in that the deep learning-based magnetic resonance imaging method, system, terminal and storage medium of the embodiment of the present application estimate the convolution of filling the undersampled k-space from the calibration data Kernel, image reconstruction based on k-space image reconstruction model, no need to estimate the coil sensitivity information, reduce the memory space occupied, so that the reconstruction model has better robustness; and use the convolutional neural network to learn the prior information of the image and Auxiliary imaging to remove image artifacts caused by undersampling can greatly reduce reconstruction time and achieve better image reconstruction results.

Description of drawings

Fig. 1 is the flowchart of the magnetic resonance imaging method based on deep learning of the embodiment of the present application;

Fig. 2 is the schematic diagram of the SPIRiT operation of the embodiment of the present application;

Fig. 3 is a schematic diagram of the convolutional neural network iterative training process of the embodiment of the present application;

4 is a schematic structural diagram of a deep learning-based magnetic resonance imaging system according to an embodiment of the present application;

FIG. 5 is a schematic structural diagram of a terminal according to an embodiment of the present application;

FIG. 6 is a schematic structural diagram of a storage medium according to an embodiment of the present application.

Detailed ways

In order to make the purpose, technical solution and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present application, not to limit the present application.

Please refer to FIG. 1 , which is a flowchart of a deep learning-based magnetic resonance imaging method according to an embodiment of the present application. The magnetic resonance imaging method based on deep learning of the embodiment of the present application comprises the following steps:

S1: Obtain full MRI k-space data from the MRI scanner;

S2: Under-sampling the full-sampled k-space data of magnetic resonance to generate multi-channel under-sampled k-space data;

In this step, the under-sampling specifically includes multiplying the full-sampled k-space data of the magnetic resonance by a mask matrix to generate multi-channel under-sampled k-space data.

S3: Carry out SPIRiT operation on the undersampled k-space data, and estimate the SPIRiT convolution kernel that fills the k-space;

In this step, as shown in FIG. 2 , it is a schematic diagram of the SPIRiT operation of the embodiment of the present application. The SPIRiT operation is to reconstruct unsampled data points in the undersampled K-space data with the linear weighting factor calculated from the self-calibration data. The linear weighting factor is the SPIRiT convolution kernel, where the calibration data is the center of the undersampled K-space data Low frequency full sampling part. Specifically, this application obtains the weight coefficient by solving the following formula:

Where m represents the mth coil, x is the k-space data point, r represents the position, R _r is an operation to select all k-space points in a neighborhood around the position r, and g _jm is obtained from all points in the neighborhood around the position r The weight vector of

is the conjugate transpose of g _jm .

S4: Input the undersampled k-space data and the estimated SPIRiT convolution kernel into the built convolutional neural network for iterative training, and obtain the trained image reconstruction model based on k-space;

In this step, the convolutional neural network is a three-layer convolutional neural network, the convolution kernel size is 3, and the feature maps are 32, 32, 2, of which 2 channels are used to represent the real part and the imaginary part of the image.

This application adopts an image reconstruction model based on k-space, and the model is expressed as follows:

Among them, x is the k-space to be reconstructed, D is the undersampling mode, y is the undersampling k-space data, G is the SPIRiT operation, which is to estimate the relationship between the undersampling k-space data and the surrounding channel data from the correction data (represented by kernel ), and then apply the kernel to the entire k-space, λ ₁ and λ ₂ are penalty parameters, R(x) is a penalty function related to prior information, and the penalty function includes but is not limited to Tikhonov regularization and image sparse regularization.

Since the image reconstruction model (2) based on k-space needs to be solved iteratively, it usually requires more iterations to converge, resulting in a longer time for image reconstruction. In addition, the selection of reconstruction parameters such as penalty function and penalty parameters is difficult, and improper selection will lead to failure of image reconstruction. Therefore, this application expands the solution process of the k-space-based image reconstruction model to the trained convolutional neural network, and finally obtains a k-space-based image reconstruction model through network training.

First, an appropriate optimization algorithm is used to solve the k-space-based image reconstruction model (2). The embodiment of the present application takes the convex set projection algorithm (projections onto convex sets, referred to as POCS) as an example to minimize the problem (2) It can be solved by the following iterative process:

Among them, n represents the number of iterations, F represents the Fourier transform; prox _{R, τ} represents the approximation operator, which is defined as follows:

The third step in the solution process (3) is to transform to the image domain because the prior information based on the image is easier to express. Although prior information based on the image domain is easier to express, it is still difficult to select appropriate prior information. Therefore, the embodiment of the present application uses a convolutional neural network to learn prior information from a large amount of training data. Combined with the solution process (3), the convolutional neural network is used to fit the approximation operator prox _{R, τ} related to the penalty function R(x), that is, the reconstruction process can be expressed as the following iterative process:

where Λ is a convolutional neural network.

Specifically, as shown in FIG. 3 , it is a schematic diagram of an iterative training process of a convolutional neural network according to an embodiment of the present application. Among them, SPIRiT operation, data fidelity DC and convolutional neural network correspond to the three steps in the iterative process (5), and IFT (Inverse Fourier Transform) and FT (Fourier Transform, IFT) represent the inverse Fourier transform and Fourier transform respectively , RSS means coil channel merge. The input of the convolutional neural network is the under-acquired k-space data and the SPIRiT convolution kernel. The image output by the network is compared with the gold standard image generated by the fully-acquired k-space data. The difference between the generated gold standard images calculates the loss function, namely:

J(Θ)＝min _Θ ||m _rec -m _ref || ₁ (6)

Among them, m _rec is the image output by the network, and m _ref is the gold standard image generated by all k-space data. Through the optimization algorithm, the loss function is minimized and the network parameters are continuously updated until the training converges to obtain the final image based on k-space. Rebuild the model.

S5: Perform magnetic resonance image reconstruction through the trained image reconstruction model based on k-space.

Based on the above, the deep learning-based magnetic resonance imaging method of the embodiment of the present application estimates the convolution kernel filling the undersampled k-space from the calibration data, and uses the image reconstruction model based on the k-space for image reconstruction without estimating the coil sensitivity information , to reduce the memory space occupied, so that the reconstruction model has better robustness; and use the convolutional neural network to learn the prior information of the image and assist in imaging, so as to remove the image artifacts caused by undersampling, which can greatly reduce the reconstruction time, And get better image reconstruction effect.

Please refer to FIG. 4 , which is a schematic structural diagram of a deep learning-based magnetic resonance imaging system according to an embodiment of the present application. The deep learning-based magnetic resonance imaging system 40 of the embodiment of the present application includes:

Under-sampling module 41: used for under-sampling the full-sampling k-space data of magnetic resonance, and generating under-sampling k-space data;

SPIRiT processing module 42: for estimating and filling the SPIRiT convolution kernel of k space according to underacquisition K space data;

Model training module 43: used to input the convolutional neural network with under-collected k-space data and SPIRiT convolution kernel for training, and obtain a trained image reconstruction model based on k-space;

Image reconstruction module 44: used for performing magnetic resonance image reconstruction through a trained k-space-based image reconstruction model.

Please refer to FIG. 5 , which is a schematic diagram of a terminal structure in an embodiment of the present application. The terminal 50 includes a processor 51 and a memory 52 coupled to the processor 51 .

The memory 52 stores program instructions for implementing the above-mentioned deep learning-based magnetic resonance imaging method.

The processor 51 is used to execute the program instructions stored in the memory 52 to control the magnetic resonance imaging based on deep learning.

Wherein, the processor 51 may also be referred to as a CPU (Central Processing Unit, central processing unit). The processor 51 may be an integrated circuit chip with signal processing capabilities. The processor 51 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components . A general-purpose processor may be a microprocessor, or the processor may be any conventional processor, or the like.

Please refer to FIG. 6 , which is a schematic structural diagram of a storage medium according to an embodiment of the present application. The storage medium of the embodiment of the present application stores a program file 61 capable of realizing all the above-mentioned methods, wherein the program file 61 can be stored in the above-mentioned storage medium in the form of a software product, and includes several instructions to make a computer device (which can It is a personal computer, a server, or a network device, etc.) or a processor (processor) that executes all or part of the steps of the methods in various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disc, etc., which can store program codes. , or terminal devices such as computers, servers, mobile phones, and tablets.

The above description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in this invention may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention will not be limited to these embodiments shown in the present invention, but will conform to the widest scope consistent with the principles and novel features disclosed in the present invention.

Claims

A method for magnetic resonance imaging based on deep learning, characterized in that, comprising:

Under-sampling the full-acquisition k-space data of magnetic resonance to generate under-acquisition k-space data;

Estimating the SPIRiT convolution kernel filling k space according to the under-collected k space data;

The under-collected k-space data and the SPIRiT convolution kernel input convolutional neural network are trained to obtain a trained image reconstruction model based on k-space;

The magnetic resonance image reconstruction is performed through the trained image reconstruction model based on k-space.
The magnetic resonance imaging method based on deep learning according to claim 1, wherein the undersampling of the magnetic resonance full-acquisition k-space data comprises:

Obtain the MRI full-acquisition k-space data from the MRI scanner;

Multiplying the magnetic resonance full-acquisition k-space data by a mask matrix to generate multi-channel under-acquisition k-space data.
The magnetic resonance imaging method based on deep learning according to claim 2, wherein the SPIRiT convolution kernel that fills the k-space according to the under-acquisition K-space data estimation comprises:

Carry out SPIRiT operation to described under-acquisition K space data, estimate the SPIRiT convolution core of filling k space;

The SPIRiT operation is to reconstruct unsampled data points with the linear weighting factor calculated by the calibration data in the under-acquisition K-space data, the linear weighting factor is the SPIRiT convolution kernel, and the calibration data is the under-acquisition K-space The low-frequency full mining part of the data center;

The calculation formula of the SPIRiT convolution kernel is:

Among them, m represents the mth coil, x is the k-space data point, r represents the position, R r is the operation of selecting all k-space points in a certain neighborhood around the position r, and g jm represents the points obtained from all the points in the neighborhood around the position r weight vector,
is the conjugate transpose of g jm .
The magnetic resonance imaging method based on deep learning according to any one of claims 1 to 3, wherein the image reconstruction model based on k-space is:

where x is the k-space to be reconstructed, D is the undersampling mode, y is the undersampling k-space data, G is the SPIRiT operation, λ 1 and λ 2 are penalty parameters, and R(x) is a penalty function related to prior information .
The magnetic resonance imaging method based on deep learning according to claim 4, wherein the described under-acquisition k-space data and the SPIRiT convolution kernel input convolutional neural network are trained to obtain a well-trained k-space-based The image reconstruction models include:

Using an optimization algorithm to iteratively solve the image reconstruction model based on k-space, the solution process is:

Where n represents the number of iterations, F represents the Fourier transform; prox R, τ represents the approximation operator, which is defined as:
The magnetic resonance imaging method based on deep learning according to claim 5, wherein the described under-acquisition k-space data and the SPIRiT convolution kernel input convolutional neural network are trained to obtain a well-trained k-space-based The image reconstruction model also includes:

The approximation operator prox R, τ related to the penalty function R(x) is fitted by a convolutional neural network, and the reconstruction process is expressed as:

where Λ is a convolutional neural network.
The magnetic resonance imaging method based on deep learning according to claim 6, wherein the loss function of the convolutional neural network is:

Among them, m rec is the image output by the convolutional neural network, and m ref is the gold standard image generated by the full sampling k-space data.
A magnetic resonance imaging system based on deep learning, characterized in that, comprising:

Undersampling module: used for undersampling the full-acquisition k-space data of magnetic resonance to generate under-acquisition k-space data;

SPIRiT processing module: for estimating the SPIRiT convolution kernel filling k space according to the under-acquisition K space data;

Model training module: for the input convolutional neural network of described under-acquisition k-space data and SPIRiT convolution kernel to train, obtain the image reconstruction model based on k-space trained well;

Image reconstruction module: for performing magnetic resonance image reconstruction through the trained k-space-based image reconstruction model.
A terminal, characterized in that the terminal includes a processor and a memory coupled to the processor, wherein,

The memory is stored with program instructions for realizing the magnetic resonance imaging method based on deep learning according to any one of claims 1-7;

The processor is configured to execute the program instructions stored in the memory to control deep learning-based magnetic resonance imaging.
A storage medium, characterized in that it stores program instructions executable by a processor, and the program instructions are used to execute the deep learning-based magnetic resonance imaging method according to any one of claims 1 to 7.