阅读说明：本技术 磁共振成像方法和装置 (Magnetic resonance imaging method and apparatus ) 是由 黄峰 孙爱琦 鞠光亮 于 2019-02-27 设计创作，主要内容包括：本申请提供了一种磁共振成像方法和装置。该方法和装置将用于描述预设K空间区域的原始采集K空间数据与共轭转置K空间数据之间的映射关系的卷积神经网络应用到未采集数据的K空间区域,如此,可以重建出该未采集数据的K空间区域的缺失K空间数据,最后,利用原始采集K空间数据和重建出的未采集数据的K空间区域的缺失K空间数据进行图像重建,得到磁共振图像。因为卷积神经网络通过一组非线性函数可以较为准确地描述原始采集K空间数据与其对应的共轭转置K空间数据的映射关系,该方法和装置重建出的图像较为准确。而且,该方法和装置能够较好地保留图像相位,并能大幅减少图像中尤其是相位变化剧烈区域处的伪影或变形。(The application provides a magnetic resonance imaging method and apparatus. The method and the device apply a convolution neural network for describing the mapping relation between the originally acquired K space data of a preset K space region and the conjugate transposed K space data to the K space region of the non-acquired data, so that missing K space data of the K space region of the non-acquired data can be reconstructed, and finally, image reconstruction is performed by using the originally acquired K space data and the reconstructed missing K space data of the K space region of the non-acquired data to obtain a magnetic resonance image. Because the convolutional neural network can more accurately describe the mapping relation between the originally acquired K space data and the corresponding conjugate transpose K space data through a group of nonlinear functions, the image reconstructed by the method and the device is more accurate. Moreover, the method and the device can better reserve the image phase and greatly reduce the artifacts or deformation in the image, especially in the area with severe phase change.)

1. A magnetic resonance imaging method, comprising:

acquiring K space data by adopting a partial Fourier acquisition mode to obtain the originally acquired K space data in an originally acquired K space region; the original acquisition K space region comprises a first K space region and a second K space region;

performing conjugate transposition on the originally acquired K space data to obtain conjugate transposition K space data in a conjugate transposition K space area; the conjugate transpose K space region includes the second K space region and a third K space region; the original acquisition K space region and the third K space region form a complete K space;

reconstructing missing K space data in a third K space region according to the conjugate transpose K space data in the third K space region and the convolutional neural network;

carrying out image reconstruction according to the originally acquired K space data and the reconstructed missing K space data in the third K space region to obtain a magnetic resonance image;

the convolutional neural network is obtained by learning a mapping relation between originally acquired K space data and conjugate transposed K space data in a preset K space region by using originally acquired K space data in the preset K space region as output data and using the conjugate transposed K space data of the originally acquired K space data in the preset K space region as input data.

2. The method according to claim 1, wherein after the conjugate transposing the original acquired K-space data to obtain the conjugate transposed K-space data and before the reconstructing missing K-space data in a third K-space region, further comprising:

training a convolutional neural network according to the originally acquired K space data and the conjugate transpose K space data;

the training of the convolutional neural network according to the originally acquired K space data and the conjugate transpose K space data specifically includes:

extracting the original collected K space data and the conjugate transposed K space data in a second K space region from the original collected K space data and the conjugate transposed K space data respectively;

and learning the mapping relation between the original collected K space data and the conjugate transposed K space data in the second K space region by taking the original collected K space data in the second K space region as output data of a pre-constructed convolutional neural network and taking the conjugate transposed K space data in the second K space region as input data of the convolutional neural network, thereby obtaining the convolutional neural network.

3. The method of claim 2, wherein prior to extracting the original acquired K-space data and the conjugate transposed K-space data in the second K-space region from the original acquired K-space data and the conjugate transposed K-space data, respectively, further comprising:

carrying out high-pass filtering processing on the originally acquired K space data and the conjugate transpose K space data;

the extracting of the original acquired K-space data and the conjugate transposed K-space data in the second K-space region from the original acquired K-space data and the conjugate transposed K-space data, respectively, specifically includes:

and extracting the original acquisition K space data and the conjugate transpose K space data in the second K space region from the original acquisition K space data and the conjugate transpose K space data which are subjected to high-pass filtering processing respectively.

4. The method of claim 2, wherein the convolutional neural network comprises a plurality of convolutional layers in series, or wherein the convolutional neural network is a single convolutional layer.

5. The method according to any one of claims 1 to 4, wherein the reconstructing an image from the originally acquired K-space data and the reconstructed missing K-space data in the third K-space region to obtain a magnetic resonance image specifically comprises:

synthesizing the originally acquired K space data and the reconstructed missing K space data in the third K space region into complete K space data;

and performing inverse Fourier transform on the complete K space data to obtain a magnetic resonance image.

6. The method according to any one of claims 1-4, wherein the second K-space region is a K-space central region, and the K-space data of the K-space central region is low-frequency K-space data.

7. A magnetic resonance imaging apparatus, characterized by comprising:

the acquisition unit is used for acquiring the K space data by adopting a partial Fourier acquisition mode to obtain the originally acquired K space data in the originally acquired K space region; the original acquisition K space region comprises a first K space region and a second K space region;

the conjugate transpose unit is used for performing conjugate transpose on the original acquired K space data to obtain conjugate transpose K space data in a conjugate transpose K space region; the conjugate transpose K space region includes the second K space region and a third K space region; the original acquisition K space region and the third K space region form a complete K space;

the data reconstruction unit is used for reconstructing missing K space data in the third K space region according to the conjugate transpose K space data in the third K space region and the convolutional neural network;

the image reconstruction unit is used for reconstructing an image according to the originally acquired K space data and the reconstructed missing K space data in the third K space region to obtain a magnetic resonance image;

the convolutional neural network is obtained by learning a mapping relation between originally acquired K space data and conjugate transposed K space data in a preset K space region by using originally acquired K space data in the preset K space region as output data and using the conjugate transposed K space data of the originally acquired K space data in the preset K space region as input data.

8. The apparatus of claim 7, further comprising:

a training unit, configured to train a convolutional neural network according to the originally acquired K-space data and the conjugate transposed K-space data after the conjugate transpose unit performs conjugate transpose on the originally acquired K-space data and before the data reconstruction unit reconstructs missing K-space data in a third K-space region;

the training unit specifically comprises:

an extraction subunit, configured to extract, from the originally-acquired K-space data and the conjugate-transposed K-space data, originally-acquired K-space data and conjugate-transposed K-space data in a second K-space region, respectively;

and the learning subunit is configured to use the originally acquired K space data in the second K space region as output data of a pre-constructed convolutional neural network, use the conjugate transpose K space data in the second K space region as input data of the convolutional neural network, and learn a mapping relationship between the originally acquired K space data and the conjugate transpose K space data in the second K space region, so as to obtain the convolutional neural network.

9. The apparatus of claim 7, wherein the training unit further comprises:

a processing subunit, configured to perform high-pass filtering processing on the originally-acquired K-space data and the conjugate-transposed K-space data before the extracting subunit extracts the originally-acquired K-space data and the conjugate-transposed K-space data in a second K-space region from the originally-acquired K-space data and the conjugate-transposed K-space data, respectively;

the learning subunit is specifically: and extracting the original acquisition K space data and the conjugate transpose K space data in the second K space region from the original acquisition K space data and the conjugate transpose K space data which are subjected to high-pass filtering processing respectively.

10. The apparatus of claim 7, wherein the convolutional neural network comprises a plurality of convolutional layers in series, or wherein the convolutional neural network is a single convolutional layer.

Technical Field

The present application relates to the field of medical imaging technologies, and in particular, to a magnetic resonance imaging method and apparatus.

Background

Magnetic Resonance Imaging (MRI) has high soft tissue contrast and spatial resolution, can simultaneously obtain morphological information and functional information of an examined part, and can flexibly select Imaging parameters and an Imaging layer according to needs, thereby becoming an important means for medical image examination at present.

However, due to the constraints of nyquist sampling theorem and main magnetic field strength, the imaging speed of magnetic resonance is very slow, which greatly limits the application of magnetic resonance in clinic.

In order to shorten the time of magnetic resonance imaging, the magnetic resonance scanning speed needs to be increased. Several methods exist to increase the speed of magnetic resonance scanning. Partial Fourier acquisition is one method to increase the speed of a magnetic resonance scan. By partial fourier acquisition, it is meant that only slightly more than half of the K-space region of K-space data can be acquired, e.g., the originally acquired K-space is 50% to 60% (excluding 50%) of the full K-space. As shown in fig. 1, the acquired original acquisition K-space includes a half of K-space S near one side of K-space₁And a low-frequency part S in the other half K space_2L。

However, in the current method for reconstructing an image based on K-space data acquired by partial fourier, phase information in the final reconstructed image is lost, or the final reconstructed image has a certain degree of artifacts or deformation, or the reconstructed image has poor accuracy.

Disclosure of Invention

In view of the above, the present application provides a magnetic resonance imaging method and apparatus, which can effectively improve the phase of a reconstructed image, greatly reduce artifacts or distortions in a final reconstructed image, and improve the accuracy of the image.

In order to solve the technical problem, the following technical scheme is adopted in the application:

a magnetic resonance imaging method, comprising:

acquiring K space data by adopting a partial Fourier acquisition mode to obtain the originally acquired K space data in an originally acquired K space region; the original acquisition K space region comprises a first K space region and a second K space region;

performing conjugate transposition on the originally acquired K space data to obtain conjugate transposition K space data in a conjugate transposition K space area; the conjugate transpose K space region includes the second K space region and a third K space region; the original acquisition K space region and the third K space region form a complete K space;

reconstructing missing K space data in a third K space region according to the conjugate transpose K space data in the third K space region and the convolutional neural network;

carrying out image reconstruction according to the originally acquired K space data and the reconstructed missing K space data in the third K space region to obtain a magnetic resonance image;

the convolutional neural network is obtained by learning a mapping relation between originally acquired K space data and conjugate transposed K space data in a preset K space region by using originally acquired K space data in the preset K space region as output data and using the conjugate transposed K space data of the originally acquired K space data in the preset K space region as input data.

Optionally, after the conjugate transposing is performed on the originally acquired K space data to obtain conjugate transposed K space data and before the reconstructing of missing K space data in a third K space region, the method further includes:

training a convolutional neural network according to the originally acquired K space data and the conjugate transpose K space data;

the training of the convolutional neural network according to the originally acquired K space data and the conjugate transpose K space data specifically includes:

extracting the original collected K space data and the conjugate transposed K space data in a second K space region from the original collected K space data and the conjugate transposed K space data respectively;

and learning the mapping relation between the original collected K space data and the conjugate transposed K space data in the second K space region by taking the original collected K space data in the second K space region as output data of a pre-constructed convolutional neural network and taking the conjugate transposed K space data in the second K space region as input data of the convolutional neural network, thereby obtaining the convolutional neural network.

Optionally, before extracting the original acquired K-space data and the conjugate transposed K-space data in the second K-space region from the original acquired K-space data and the conjugate transposed K-space data, respectively, the method further includes:

carrying out high-pass filtering processing on the originally acquired K space data and the conjugate transpose K space data;

the extracting of the original acquired K-space data and the conjugate transposed K-space data in the second K-space region from the original acquired K-space data and the conjugate transposed K-space data, respectively, specifically includes:

and extracting the original acquisition K space data and the conjugate transpose K space data in the second K space region from the original acquisition K space data and the conjugate transpose K space data which are subjected to high-pass filtering processing respectively.

Optionally, the convolutional neural network comprises a plurality of convolutional layers connected in series, or the convolutional neural network is a single convolutional layer.

Optionally, the reconstructing an image according to the originally acquired K-space data and the reconstructed missing K-space data in the third K-space region to obtain a magnetic resonance image specifically includes:

synthesizing the originally acquired K space data and the reconstructed missing K space data in the third K space region into complete K space data;

and performing inverse Fourier transform on the complete K space data to obtain a magnetic resonance image.

Optionally, the second K space region is a K space central region, and K space data of the K space central region is low-frequency K space data.

A magnetic resonance imaging apparatus comprising:

the acquisition unit is used for acquiring the K space data by adopting a partial Fourier acquisition mode to obtain the originally acquired K space data in the originally acquired K space region; the original acquisition K space region comprises a first K space region and a second K space region;

the conjugate transpose unit is used for performing conjugate transpose on the original acquired K space data to obtain conjugate transpose K space data in a conjugate transpose K space region; the conjugate transpose K space region includes the second K space region and a third K space region; the original acquisition K space region and the third K space region form a complete K space;

the data reconstruction unit is used for reconstructing missing K space data in the third K space region according to the conjugate transpose K space data in the third K space region and the convolutional neural network;

the image reconstruction unit is used for reconstructing an image according to the originally acquired K space data and the reconstructed missing K space data in the third K space region to obtain a magnetic resonance image;

the convolutional neural network is obtained by learning a mapping relation between originally acquired K space data and conjugate transposed K space data in a preset K space region by using originally acquired K space data in the preset K space region as output data and using the conjugate transposed K space data of the originally acquired K space data in the preset K space region as input data.

Optionally, the apparatus further comprises:

a training unit, configured to train a convolutional neural network according to the originally acquired K-space data and the conjugate transposed K-space data after the conjugate transpose unit performs conjugate transpose on the originally acquired K-space data and before the data reconstruction unit reconstructs missing K-space data in a third K-space region;

the training unit specifically comprises:

an extraction subunit, configured to extract, from the originally-acquired K-space data and the conjugate-transposed K-space data, originally-acquired K-space data and conjugate-transposed K-space data in a second K-space region, respectively;

and the learning subunit is configured to use the originally acquired K space data in the second K space region as output data of a pre-constructed convolutional neural network, use the conjugate transpose K space data in the second K space region as input data of the convolutional neural network, and learn a mapping relationship between the originally acquired K space data and the conjugate transpose K space data in the second K space region, so as to obtain the convolutional neural network.

Optionally, the training unit further comprises:

a processing subunit, configured to perform high-pass filtering processing on the originally-acquired K-space data and the conjugate-transposed K-space data before the extracting subunit extracts the originally-acquired K-space data and the conjugate-transposed K-space data in a second K-space region from the originally-acquired K-space data and the conjugate-transposed K-space data, respectively;

the learning subunit is specifically: and extracting the original acquisition K space data and the conjugate transpose K space data in the second K space region from the original acquisition K space data and the conjugate transpose K space data which are subjected to high-pass filtering processing respectively.

Optionally, the convolutional neural network comprises a plurality of convolutional layers connected in series, or the convolutional neural network is a single convolutional layer.

Compared with the prior art, the method has the following beneficial effects:

based on the above technical solutions, it can be known that the magnetic resonance imaging method provided by the present application applies a convolutional neural network for describing a mapping relationship between originally acquired K-space data of a preset K-space region and conjugate transposed K-space data to a K-space region where data is not acquired based on a principle that mapping relationships between originally acquired K-space data of different K-space regions and the conjugate transposed K-space data corresponding thereto are the same, so that missing K-space data of the K-space region where data is not acquired can be reconstructed, and finally, image reconstruction is performed using the originally acquired K-space data and the missing K-space data of the reconstructed K-space region where data is not acquired, so as to obtain a magnetic resonance image.

Because the convolutional neural network can more accurately describe the mapping relationship between the originally acquired K space data and the corresponding conjugate transpose K space data through a group of nonlinear functions, the missing K space data of the K space region of the uncollected data reconstructed based on the convolutional neural network is closer to the actually acquired K space data, and therefore, the image reconstructed based on the originally acquired K space data and the reconstructed K space data is more accurate. Moreover, because the image domain phase information contained in the mapping relation between the originally acquired K-space data and the corresponding conjugate transposed K-space data is acquired based on the implicit estimation method of the K-space domain, the magnetic resonance imaging method can better reserve the image phase and can greatly reduce the artifacts or deformation in the image, particularly in the region with severe phase change.

Drawings

In order that the detailed description of the present application may be clearly understood, a brief description of the drawings that will be used when describing the detailed description of the present application will be provided. It is to be understood that these drawings are merely illustrative of some of the embodiments of the application.

FIG. 1 is a schematic diagram of a K-space region acquired in a partial Fourier acquisition mode;

FIG. 2 is a schematic flowchart of a training method of a convolutional neural network according to an embodiment of the present disclosure;

fig. 3 is a schematic distribution diagram of a K-space region provided in an embodiment of the present application;

fig. 4 is a schematic flow chart of a magnetic resonance imaging method according to an embodiment of the present application;

FIG. 5 is a schematic diagram of another K-space region distribution provided in the embodiments of the present application;

FIG. 6 is a flow chart of another magnetic resonance imaging method provided by an embodiment of the present application;

fig. 7 is a schematic flowchart of a specific implementation manner of step S603 provided in an embodiment of the present application;

fig. 8 is a schematic structural diagram of a control device for executing a magnetic resonance imaging method according to an embodiment of the present application;

fig. 9 is a schematic structural diagram of a magnetic resonance imaging apparatus according to an embodiment of the present application;

fig. 10 is a schematic structural diagram of a training apparatus for a convolutional neural network according to an embodiment of the present disclosure.

Detailed Description

At present, methods for reconstructing images of K-space data acquired by partial fourier transform mainly fall into two categories:

the first type: firstly, phase information of an image reconstructed from acquired K space data is modulated, and then image reconstruction is performed by using Hermitian symmetry of Hermitian (Hermitian, Hermitian and Hermitian) of K space data corresponding to a real-value image.

The methods mainly comprise Homodyne (Homodyne) algorithm, phase correction and conjugate synthesis method and the like, and the method has the defect that the phase information of the finally reconstructed image is usually lost.

The second type: firstly, image phase information is estimated, and then the estimated phase information is used for recovering the non-collected K space data.

The method mainly comprises a Cuppen's algorithm, an iterative algorithm (POCS) based on projection to convex set theory and the like.

Compared with the first method, the method can effectively improve the modulus and the phase of the reconstructed image, but the algorithms all adopt the calculation of image phase information in an image domain, the methods generally adopt the mode of directly taking data of a central low-frequency part of K space to perform Fourier transform to the image domain, and then phase estimation is performed on the basis of the low-frequency image, and the mode can cause the obtained low-frequency image to have obvious Gibbs (Gibbs') artifacts due to the influence of selecting the data of the central area of the K space and distribute in the whole image, and is especially obvious in a place with severe phase change. Therefore, in the methods, for a local image with severe local phase change in the image, artifacts and deformation are generated to a certain extent due to a boundary effect between a low-frequency K-space region and a high-frequency K-space region, and particularly, when there is a high-frequency phase change, the artifacts and deformation are severe.

Therefore, how to eliminate artifacts or distortion in the region where the image phase changes are severe is an important direction to improve the partial fourier reconstruction algorithm.

In order to solve the above problems, the present application has made intensive studies on a magnetic resonance imaging method, and found that:

the artifact or distortion in the region where the image phase changes drastically in the second method is caused by the calculation of the image phase information in the image domain, which belongs to the display phase estimation method. Therefore, in order to eliminate artifacts or distortion in a region where the image phase changes drastically, it is necessary to avoid calculation of image phase information using the image domain.

In addition, studies have also found that phase information of magnetic resonance images can be implicitly estimated in the K-space domain, in addition to explicit phase estimation in the image domain. And by adopting the implicit estimation of the K space domain, the operation of directly taking the low-frequency region of the center of the K space to perform Fourier transform is avoided, so that the artifacts or deformation in the region with severe image phase change can be greatly reduced.

In addition, the K space region acquired by the partial fourier acquisition mode generally exceeds 50% of the complete K space, and therefore, an overlapping region exists between the original K space region acquired by the partial fourier acquisition mode and the conjugate transpose K space region corresponding to the original K space region. After the original K space data are subjected to conjugate transposition, the overlapped area is filled with the original collected K space data and also has the conjugate transposition K space data. Thus, a K-space implicit estimate of the phase can be obtained by solving a mapping function between the original acquired K-space data and the conjugate transposed K-space data for the overlapping region.

In addition, the mapping relationship between the original collected K-space data of different K-space regions and the corresponding conjugate transposed K-space data is the same. Therefore, the mapping relationship between the original acquired K-space data and the conjugate transposed K-space data in the overlapping region can be applied to the uncollected K-space region, so that the missing data of the uncollected K-space region is reconstructed according to the conjugate transposed K-space in the uncollected K-space region and the obtained mapping relationship.

Based on this, the invention concept of the magnetic resonance imaging method for the partial fourier acquisition mode provided by the present application is as follows: and performing conjugate transposition on the original acquired K space data in the acquired original acquired K space region to obtain the conjugate transposed K space data in the conjugate transposed K space region. Because the original acquisition K space area and the conjugate transposition K space area have an overlapping area. Therefore, first, the mapping relationship between the original acquired K-space data and the conjugate transposed K-space data in the overlap region is obtained, and then the mapping relationship between the original acquired K-space data and the conjugate transposed K-space data in the overlap region is applied to the other K-space regions. In this way, the missing K-space data of the K-space region where data is not acquired can be reconstructed from the mapping relationship and the conjugate transpose of the K-space region where data is not acquired. Therefore, complete K space data can be obtained according to the originally acquired K space data and the missing K space data of the K space region of the non-acquired data reconstructed by the mapping relation, and finally, image reconstruction is carried out according to the complete K space data, so that the magnetic resonance image is finally obtained.

In addition, in the embodiment of the present application, in order to obtain an accurate mapping relationship between the original acquired K-space data and the conjugate transposed K-space data, a nonlinear convolutional neural network is used in the present application to train and learn the mapping relationship between the original acquired K-space data and the conjugate transposed K-space data in the overlapping region. Moreover, the trained convolutional neural network can describe not only the corresponding image domain phase information relationship between the original acquisition K space data and the conjugate transpose K space data in the overlapping region, but also the image domain modulus relationship between the original acquisition K space data and the conjugate transpose K space data. Therefore, the method and the device can accurately acquire the corresponding image domain phase relation and module value relation between the originally acquired K space data and the conjugate transpose K space data corresponding to the originally acquired K space data, so that the reconstructed image quality and the image accuracy are higher.

In addition, the mapping relation between the originally acquired K-space data and the corresponding conjugate transpose K-space data is obtained based on an implicit estimation method of a K-space domain, so that the magnetic resonance imaging method can better reserve the image phase and can greatly reduce the artifacts or deformation in the image, particularly in the region with severe phase change.

The following detailed description of specific embodiments of the present application is provided in conjunction with the accompanying drawings.

It should be noted that the convolutional neural network used in the embodiment of the present application may be obtained in an offline training manner, or may be obtained in an online training manner. The off-line training mode may use multiple sets of training data, which may be K-space data for acquiring multiple scan objects. The number of iterations of the off-line training can be increased, so that the convolutional neural network obtained by training has better fitting performance and generalization.

The online training mode can be obtained by taking the model parameters of the convolutional neural network which is trained offline as initialization parameters, setting a smaller learning rate and performing iterative optimization for a smaller number of times. As an example, the learning rate during offline training may be set to 0.001, and the learning rate during online adjustment may be reduced to 0.0001.

First, a specific implementation of training the convolutional neural network by an off-line training method is described below.

Referring to fig. 2, a training method of a convolutional neural network provided in an embodiment of the present application includes the following steps:

s201: and constructing a convolutional neural network.

By way of example, this step may be to serially connect convolutional layers of a plurality of small convolutional kernels to construct a convolutional neural network. For example, the convolutional neural network may be constructed by concatenating convolutional layers of 3 convolutional kernels by 3 x 3.

As another example, the convolution layer with one large convolution kernel may be connected in series to construct a convolution neural network, for example, the convolution neural network may be constructed with 1 convolution kernel with 7 × 7 convolution layers.

S202: a training set of the convolutional neural network is obtained.

The step may specifically be: acquiring original K space data in a partial Fourier acquisition mode, performing conjugate transpose on the original K space data to obtain conjugate transposed K space data, and taking the original K space data and the conjugate transposed K space data in a preset K space area as training data of a convolutional neural network.

The preset K-space region may be an overlapping region of the original K-space region and the conjugate transpose K-space region.

As an example, as shown in fig. 3, the raw K-space data acquired by the partial fourier acquisition method is K-space data in the K-space region a. And performing conjugate transpose on the K space data in the K space area A to obtain conjugate transpose K space data in a conjugate transpose K space area B, wherein an overlapping area exists between the K space area A and the conjugate transpose K space area B, namely a K space area a. The original collected K space data in the K space area a and the corresponding conjugate transpose K space data form a group of training data.

In addition, since the data in the K space central region has a higher signal-to-noise ratio, in order to improve the accuracy of the convolutional neural network obtained by training, the preset K space region may be the K space central region. And the K space data of the central area of the K space is low-frequency K space data.

S203: and learning the mapping relation between the original collected K space data and the conjugate transposed K space data by taking the original collected K space data in the training set as the output data of the convolutional neural network and taking the conjugate transposed K space data in the training set as the input data of the convolutional neural network, thereby obtaining the convolutional neural network.

In this step, the originally acquired K space data in the preset K space region in the plurality of sets of training data may be used as the output data of the convolutional neural network, the conjugate transposed K space data corresponding to the originally acquired K space data may be used as the input data of the convolutional neural network, and the mapping relationship between the originally acquired K space data in the preset K space region in the plurality of sets of training data and the conjugate transposed K space data corresponding thereto may be learned through multiple iterations.

The mapping relationship may be a nonlinear mapping relationship, which may describe not only a corresponding image domain modulus relationship between the originally acquired K-space data and its corresponding conjugate transpose K-space data, but also an image domain phase relationship between the two.

It should be noted that the mapping relationship between the original acquired K-space data and the conjugate transposed K-space data can be represented as a convolution kernel of a convolutional neural network.

Therefore, after the mapping relation between the originally acquired K space data and the conjugate transpose K space data is obtained, the convolution kernel of the convolutional neural network is obtained, and therefore the model parameters of the convolutional neural network can be determined according to the mapping relation, and the trained convolutional neural network is obtained.

The above is a specific implementation manner of the training method of the convolutional neural network provided in the embodiment of the present application. In the above training method of the convolutional neural network, training is performed based on K-space data of a K-space domain. Specifically, the original collected K space data in a preset K space region is used as output data, and the conjugate transpose K space data of the original collected K space data in the preset K space region is used as input data for training to obtain the K space data. The trained model parameters or convolution kernels of the convolutional neural network can describe not only the corresponding image domain phase relationship between the originally acquired K space data and the conjugate transpose K space data corresponding to the originally acquired K space data, but also the image domain modulus relationship between the originally acquired K space data and the conjugate transpose K space data.

Therefore, the convolutional neural network can more accurately describe the mapping relation between the originally acquired K space data and the corresponding conjugate transpose K space data through a group of nonlinear functions, and further the K space data of the non-acquired K space area reconstructed based on the convolutional neural network is closer to the actual K space data. Therefore, an image reconstructed based on the originally acquired K-space data and the reconstructed K-space data is accurate, and because the image phase information contained in the mapping relation between the originally acquired K-space data and the corresponding conjugate transpose K-space data is acquired based on the implicit estimation method of the K-space domain, the magnetic resonance imaging method based on the convolutional neural network can better reserve the image phase and can greatly reduce the artifacts or deformation in the image, particularly in the region with severe phase change.

Based on the convolutional neural network obtained by training in the embodiment, the application also provides a magnetic resonance imaging method. The imaging method is used for reconstructing images of K space data obtained by adopting a partial Fourier acquisition mode.

Referring to fig. 4, an embodiment of the present application provides a magnetic resonance imaging method including the following steps:

s401: and acquiring the K space data by adopting a partial Fourier acquisition mode to obtain the originally acquired K space data in the originally acquired K space region.

As shown in FIG. 5, K-space data is acquired by partial Fourier acquisition to obtain an originally acquired K-space region S₀Raw K space data K in_P. Original acquisition K space region S₀A first K space region S including a peripheral region near K space₁And a second K space region S near the central region of K space₂。

Wherein, is located in the first K space region S₁The K space data may be high frequency data, located in a second K space region S₂The K-space data within may be low frequency data.

Wherein, the signal-to-noise ratio of the image reconstructed by the high-frequency data is lower, and the signal-to-noise ratio of the image reconstructed by the low-frequency data is higher.

It should be noted that S401 collects the original collected K space region S₀Each row of data within.

S402: and performing conjugate transposition on the original acquired K space data in the original acquired K space region to obtain the conjugate transposition K space data.

As also shown in FIG. 5, K-space is divided into a first K-space region S₁A second K space region S₂And a third K space region S₃. Wherein the first K space region S₁And a second K space region S₂Forming an original acquisition K space region S₀Thus, the first K space region S₁And a second K space region S₂Already populated with the raw acquisition K-space data. And a third K space region S₃Not filled with K-space data.

In addition, the original acquisition K space region S₀Is transformed into a K space region S_Conjugation. The conjugate transpose K space region S_ConjugationA third K space region S including a peripheral region near K space₃And a second K space region S near the central region of K space₂。

Thus, the second K space region S₂For the original acquisition of K space region S₀And conjugate transpose K space S_ConjugationThe overlapping area of (a). And a second K space region S₂For the original acquisition of K space region S₀Adjacent to the third K space region S₃A partial region of (a); original acquisition K space region S₀And a third K space region S₃Constituting the complete K space.

As an example, the step may specifically be: for original acquisition K space region S₀Raw K space data K in_PPerforming conjugate transposition to obtain conjugate transposition K space data

The conjugate transpose of K space data

Including a second K space region S₂Inner conjugate transpose of K space data

And a third K space region S₃Inner conjugate transpose of K space data

Since the data in the central region of K-space is low frequency data and the signal-to-noise ratio is high, the second spatial region S is taken as an example₂May be the K-space central region.

S403: and reconstructing missing K space data in the third K space region according to the conjugate transpose K space data in the third K space region and the convolutional neural network.

The step may specifically be: and inputting the conjugate transpose K space data KPH,3 in the third K space region into the convolutional neural network, and obtaining an output result which is the reconstructed K space data in the third K space region.

It should be noted that, in the embodiment of the present application, the convolutional neural network is a convolutional neural network obtained by training in the above embodiment. Therefore, the convolutional neural network can describe the mapping relation between the original K space acquisition data in the preset K space region and the corresponding conjugate transpose K space data.

In addition, in the K space, the mapping relationship between the original K space acquisition data of different K space regions and the conjugate transpose K space acquisition data is the same, so that the obtained convolutional neural network for describing the mapping relationship between the original K space acquisition data in the preset K space region and the conjugate transpose K space acquisition data can also describe the mapping relationship between the original K space acquisition data in other K space regions and the conjugate transpose K space acquisition data.

Therefore, the conjugate transposed K-space data in the third K-space region S3 is input to the convolutional neural network, which operates to output the missing K-space data K in the reconstructed third K-space region_P,3。

S404: and reconstructing an image according to the originally acquired K space data and the reconstructed missing K space data in the third K space region to obtain a magnetic resonance image.

The step may specifically be: firstly, integrating originally acquired K space data with missing K space data in a reconstructed third K space region to obtain complete K space data; and then carrying out inverse Fourier transform on the complete K space data, and reconstructing to obtain a magnetic resonance image.

The foregoing is a specific implementation manner of the magnetic resonance imaging method provided in the embodiments of the present application. In the specific implementation manner, based on the principle that the mapping relationship between the originally acquired K-space data of different K-space regions and the corresponding conjugate transposed K-space data is the same, a convolutional neural network for describing the mapping relationship between the originally acquired K-space data of a preset K-space region and the conjugate transposed K-space data is applied to the K-space region where data is not acquired, so that missing K-space data of the K-space region where data is not acquired can be reconstructed, and finally, image reconstruction is performed by using the originally acquired K-space data and the missing K-space data of the reconstructed K-space region where data is not acquired, so as to obtain a magnetic resonance image.

Because the convolutional neural network can more accurately describe the mapping relationship between the originally acquired K space data and the corresponding conjugate transpose K space data through a group of nonlinear functions, the missing K space data of the K space region of the uncollected data reconstructed based on the convolutional neural network is closer to the actually acquired K space data, and therefore, the image reconstructed based on the originally acquired K space data and the reconstructed K space data is more accurate. Moreover, because the image phase information contained in the mapping relation between the originally acquired K-space data and the corresponding conjugate transposed K-space data is acquired based on the implicit estimation method of the K-space domain, the magnetic resonance imaging method can better reserve the image phase and can greatly reduce the artifacts or deformation in the image, particularly in the region with severe phase change.

In the magnetic resonance imaging method provided by the above embodiment, the convolutional neural network is a convolutional neural network trained offline. As another implementation mode of the application, the convolutional neural network can also be obtained by an online training mode, and the convolutional neural network obtained by the online training mode can be applied immediately. Accordingly, the present application also provides another implementation of a magnetic resonance imaging method.

Referring to fig. 6, another implementation manner of the magnetic resonance imaging method provided by the embodiment of the present application includes the following steps:

s601 to S602 are the same as S401 to S402, and for the sake of brevity, will not be described in detail here. For specific information, please refer to the related descriptions of S401 to S402.

S603: and training the convolutional neural network according to the originally acquired K space data and the conjugate transpose K space data.

In the embodiment of the application, the trained convolutional neural network is used for describing the mapping relationship between the original acquired K-space data and the conjugate transposed K-space data. Therefore, when the convolutional neural network is trained, the adopted training data are the original collected K space data and the conjugate transposed K space data in the same K space region. In order to satisfy the condition of training data, the original acquisition K-space data and the conjugate transpose K-space data in the overlapping region of the original acquisition K-space region and the conjugate transpose K-space region may be used for training.

For the K-space shown in fig. 5, the second K-space region is an overlapping region of the original K-space region and the conjugate transposed K-space region. Thus, as an example, as shown in fig. 7, the present step may specifically include the following steps:

s6031: and carrying out high-pass filtering processing on the originally acquired K space data and the conjugate transpose K space data.

In order to enable the trained convolutional neural network to be better applied to other K space regions, in the embodiment of the present application, high-pass filtering processing may be performed on the originally acquired K space data in the originally acquired K space region and the conjugate transpose K space data in the conjugate transpose K space corresponding to the originally acquired K space data. The high-pass filtering processing can inhibit a strong signal in the K space data, so that the mapping relation between the original acquisition K space data in the overlapping area and the corresponding conjugate transpose K space data can be well matched to other K space areas.

S6032: and extracting the original acquisition K space data and the conjugate transpose K space data in the second K space region from the original acquisition K space data and the conjugate transpose K space data which are subjected to high-pass filtering processing respectively.

The step may specifically be: extracting the original K space data in the second K space region from the original collected K space data after the high-pass filtering processing, and extracting the conjugate transpose K space data in the second K space region from the conjugate transpose K space data after the high-pass filtering processing.

As the K space central region is low frequency data, the signal to noise ratio is higher, and as an example, the second K space region may be the K space central region, so that a more accurate mapping relationship is obtained through training. In this example, S6032 may specifically be: extracting original K space data on a K space central area from the original collected K space data after the high-pass filtering processing, and extracting conjugate transpose K space data on the K space central area from the conjugate transpose K space data after the high-pass filtering processing.

S6033: and learning the mapping relation between the original collected K space data and the conjugate transposed K space data in the second K space region by taking the original collected K space data in the second K space region as output data of a pre-constructed convolutional neural network and taking the conjugate transposed K space data in the second K space region as input data of the convolutional neural network, thereby obtaining the convolutional neural network.

In this application, the pre-constructed convolutional neural network may be a convolutional neural network obtained by an offline training mode. Through the online learning of S6033, the convolutional neural network trained offline in advance can be optimized online, so that the optimized convolutional neural network can better describe the mapping relationship between the current originally acquired K space data and the corresponding conjugate transpose K space data, and the reconstructed image has higher quality.

The online training or online optimization of the convolutional neural network is completed through the above steps S6031 to S6033.

S604 to S605 are the same as S403 to S404 in the above embodiment, and for the sake of brevity, will not be described in detail here.

The above is another implementation manner of the magnetic resonance imaging method provided by the embodiment of the application. In this implementation manner, in addition to the beneficial effects of the magnetic resonance imaging method described in the above embodiment, online training and immediate application of the convolutional neural network can be realized, and offline training and online optimization of the convolutional neural network can also be realized. The whole-process GPU (Graphics Processing Unit) acceleration can be realized in a mode of combining offline training and online optimization, the online reconstruction time is short, and particularly, the operation time can be obviously reduced when bidirectional partial Fourier K space data is reconstructed, so that the magnetic resonance imaging method can be applied to clinical practice. Therefore, the magnetic resonance imaging method provided by the embodiment of the application is a quick, accurate and stable magnetic resonance imaging method suitable for clinical use.

The foregoing is a specific implementation manner of the magnetic resonance imaging method provided in the embodiments of the present application.

The magnetic resonance imaging method provided by the above-mentioned embodiment can be executed by the control device shown in fig. 8. The control device shown in fig. 8 includes a processor (processor)810, a communication Interface (Communications Interface)820, a memory (memory)830, and a bus 840. Processor 810, communication interface 820, and memory 830 communicate with each other via bus 840.

The memory 830 may store logic instructions for magnetic resonance imaging, and the memory may be a non-volatile memory (non-volatile memory), for example. The processor 810 may invoke logic instructions to perform magnetic resonance imaging in the memory 830 to perform the magnetic resonance imaging methods described above. As an embodiment, the logic instruction of the magnetic resonance imaging may be a program corresponding to control software, and when the processor executes the instruction, the control device may correspondingly display a functional interface corresponding to the instruction on the display interface.

The functionality of the logic instructions of the magnetic resonance imaging, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present disclosure may be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the methods according to the embodiments of the present application. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes.

The above-mentioned logic instructions for magnetic resonance imaging may be referred to as "magnetic resonance imaging apparatus", and the apparatus may be divided into various functional modules. See in particular the examples below.

Referring to fig. 9, an embodiment of the present application provides a magnetic resonance imaging apparatus, including:

the acquisition unit 91 is configured to acquire K-space data in a partial fourier acquisition manner to obtain originally acquired K-space data in an originally acquired K-space region; the original acquisition K space region comprises a first K space region and a second K space region;

a conjugate transpose unit 92, configured to perform conjugate transpose on the originally acquired K space data to obtain conjugate transpose K space data in a conjugate transpose K space region; the conjugate transpose K space region includes the second K space region and a third K space region; the original acquisition K space region and the third K space region form a complete K space;

a data reconstruction unit 93, configured to reconstruct missing K space data in the third K space region according to the conjugate transposed K space data in the third K space region and the convolutional neural network;

an image reconstruction unit 94, configured to perform image reconstruction according to the originally acquired K space data and the reconstructed missing K space data in the third K space region, so as to obtain a magnetic resonance image;

the convolutional neural network is obtained by learning a mapping relation between originally acquired K space data and conjugate transposed K space data in a preset K space region by using originally acquired K space data in the preset K space region as output data and using the conjugate transposed K space data of the originally acquired K space data in the preset K space region as input data.

The foregoing is a specific implementation manner of a magnetic resonance imaging apparatus provided in the embodiments of the present application. In this specific implementation manner, based on the principle that the mapping relationships between the originally acquired K-space data of different K-space regions and the conjugate transposed K-space data corresponding thereto are the same, a convolutional neural network for describing the mapping relationship between the originally acquired K-space data of a preset K-space region and the conjugate transposed K-space data is applied to the K-space region where no data is acquired, so that the data reconstruction unit 93 can reconstruct missing K-space data of the K-space region where no data is acquired, and finally, the image reconstruction unit 94 performs image reconstruction using the originally acquired K-space data and the reconstructed missing K-space data of the K-space region where no data is acquired, to obtain a magnetic resonance image.

Because the convolutional neural network can more accurately describe the mapping relationship between the originally acquired K-space data and the conjugate transpose K-space data corresponding to the originally acquired K-space data through a set of nonlinear functions, the missing K-space data of the K-space region of the uncollected data reconstructed by the data reconstruction unit 93 based on the convolutional neural network is closer to the actually acquired K-space data, and therefore, the image reconstructed by the data reconstruction unit 93 based on the originally acquired K-space data and the reconstructed K-space data is more accurate. Moreover, because the image phase information contained in the mapping relation between the originally acquired K-space data and the corresponding conjugate transposed K-space data is acquired based on the implicit estimation method of the K-space domain, the magnetic resonance imaging device can better reserve the image phase and can greatly reduce the artifacts or deformation in the image, particularly in the region with severe phase change.

As a possible implementation manner of the present application, the magnetic resonance imaging apparatus described above may further include:

a training unit 95, configured to train a convolutional neural network according to the originally acquired K space data and the conjugate transposed K space data after the conjugate transpose unit 92 performs conjugate transpose on the originally acquired K space data and before the data reconstruction unit 93 reconstructs missing K space data in a third K space region;

wherein, the training unit 95 may specifically include:

an extracting sub-unit 951, configured to extract the originally acquired K-space data and the conjugate transposed K-space data in the second K-space region from the originally acquired K-space data and the conjugate transposed K-space data, respectively;

a learning subunit 952, configured to use the originally acquired K space data in the second K space region as output data of a pre-constructed convolutional neural network, use the conjugate transpose K space data in the second K space region as input data of the convolutional neural network, and learn a mapping relationship between the originally acquired K space data and the conjugate transpose K space data in the second K space region, so as to obtain the convolutional neural network.

In addition, as another embodiment of the present application, in order to further improve the accuracy of the trained convolutional neural network, the training unit 95 may further include:

a processing subunit 953, configured to perform high-pass filtering on the originally acquired K-space data and the conjugate transposed K-space data before the extracting subunit extracts the originally acquired K-space data and the conjugate transposed K-space data in a second K-space region from the originally acquired K-space data and the conjugate transposed K-space data, respectively;

the learning subunit 952 is specifically: and extracting the original acquisition K space data and the conjugate transpose K space data in the second K space region from the original acquisition K space data and the conjugate transpose K space data which are subjected to high-pass filtering processing respectively.

The training unit 95 can be used for realizing the online training and the instant application of the convolutional neural network, and can also be used for realizing the offline training and the online optimization of the convolutional neural network. The whole-process GPU (Graphics Processing Unit) acceleration can be realized in a mode of combining offline training and online optimization, the online reconstruction time is short, and particularly, the operation time can be obviously reduced when bidirectional partial Fourier K space data is reconstructed, so that the magnetic resonance imaging method can be applied to clinical practice. Therefore, the magnetic resonance imaging device provided by the embodiment of the application is a fast, accurate and stable magnetic resonance imaging device suitable for clinical use.

In addition, based on the training method of the convolutional neural network provided by the above embodiment (shown in fig. 2), the present application also provides a specific implementation manner of the training apparatus of the convolutional neural network.

Referring to fig. 10, the training apparatus of the convolutional neural network includes:

a construction unit 101, configured to construct a convolutional neural network;

an obtaining unit 102, configured to obtain a training set of the convolutional neural network; the training set comprises original collected K space data in a preset K space region and conjugate transpose K space data obtained by conjugate transpose of the original collected K space data;

the learning unit 103 is configured to learn a mapping relationship between the originally acquired K space data and the conjugate transposed K space data by using the originally acquired K space data as output data of the convolutional neural network and using the conjugate transposed K space data as input data of the convolutional neural network, so as to obtain the convolutional neural network.

The above is a specific implementation manner of the training apparatus of the convolutional neural network provided in the embodiment of the present application. In the above convolutional neural network training device, training is performed based on K-space data in a K-space domain. Specifically, the original collected K space data in a preset K space region is used as output data, and the conjugate transpose K space data of the original collected K space data in the preset K space region is used as input data for training to obtain the K space data. The trained model parameters or convolution kernels of the convolutional neural network can describe not only the corresponding image domain phase relationship between the originally acquired K space data and the conjugate transpose K space data corresponding to the originally acquired K space data, but also the image domain modulus relationship between the originally acquired K space data and the conjugate transpose K space data.

Therefore, the convolutional neural network can more accurately describe the mapping relation between the originally acquired K space data and the corresponding conjugate transpose K space data through a group of nonlinear functions, and further the K space data of the non-acquired K space area reconstructed based on the convolutional neural network is closer to the actual K space data. Therefore, an image reconstructed based on the originally acquired K-space data and the reconstructed K-space data is accurate, and because the image phase information contained in the mapping relation between the originally acquired K-space data and the corresponding conjugate transpose K-space data is acquired based on the implicit estimation method of the K-space domain, the magnetic resonance imaging method based on the convolutional neural network can better reserve the image phase and can greatly reduce the artifacts or deformation in the image, particularly in the region with severe phase change.

The above provides a specific implementation manner for the embodiment of the present application.