KR20190038333A - Apparatus and Method for Reconstructing Magnetic Resonance Image use Learning, Under-sampling Apparatus and Method about it, and Recording Medium thereof - Google Patents

Abstract

Disclosed are an under-sampling apparatus for restoring a magnetic resonance image using machine learning and a method thereof, an apparatus for restoring a magnetic resonance image using machine learning and a method thereof, and a recording medium therefor. The under-sampling apparatus includes: an area setting unit setting an area corresponding to the center of a first k-space image as a first area and setting the remaining area as a second area; and an under-sampling unit performing full-sampling for the first area and performing under-sampling for the second area to perform under-sampling for the first k-space image, wherein the under-sampling performed on the second area includes selecting lines at predetermined intervals and then performing full-sampling only for the selected lines. The under-sampling apparatus can obtain a high-quality restored magnetic resonance image while reducing a photographing time.

Description

TECHNICAL FIELD [0001] The present invention relates to an apparatus and method for recovering a magnetic resonance image using a magnetic resonance imaging (MRI) it, and Recording Medium thereof}

The present invention relates to an under-sampling apparatus and method for restoring a magnetic resonance image, an apparatus and method for restoring a magnetic resonance image, and a recording medium therefor, and more particularly, to an under-sampling apparatus and method for restoring a magnetic resonance image using learning And a recording medium therefor. 2. Description of the Related Art

Magnetic resonance imaging (MRI) is a representative medical image that can obtain non-invasive and high-quality images of the human body. Generally, as the image quality of the mechanically acquired image increases, the data to be acquired increases, which means that the imaging time is longer in the MRI image.

Because the imaging time of magnetic resonance imaging is very long, imaging a magnetic resonance imaging for a long time for high quality imaging is a very painful process for the patient. In addition, it takes a considerable time to acquire data, thereby giving a considerable economic burden to a patient using a magnetic resonance imaging apparatus.

For this reason, a method of acquiring an under-sampled magnetic resonance image to acquire an image similar to a high-quality image obtained by reconstructing the under-sampled image to reduce the imaging time of the magnetic resonance imaging is used. To obtain a magnetic resonance image, a k-space image is first acquired through photography, and the acquired k-space image is acquired to obtain a magnetic resonance image. In the k-space image of the undersampled magnetic resonance image, no specific line (or region) is encoded and no data exists.

In order to reconstruct the under-sampled magnetic resonance image, there has been a compression sensing method. This method restores the image using random sampling data and total variation minimization-based least squares method. It has a fundamental disadvantage of erasing the fine spatial changes of images when reconstructing images, and it is very difficult to inject into clinics due to the nature of the medical imaging field, which is important to detect small variants (eg, early cancers) .

Recently, a method of performing image reconstruction using a deep learning network such as CNN (Convolutional Neural Network) has been proposed. As a representative prior art, "A Deep Cascade of Convolutional Neural Networks for MR Image Reconstruction" proposed by Jp Schlemper et al. .

However, the conventional reconstruction methods using the deep learning network are a method of performing image conversion by connecting CNNs only in series, and this method also has a problem in that artifacts generated when reconstructing a magnetic resonance image can not be efficiently removed.

In order to solve the problems of the prior art as described above, it is an object of the present invention to provide an under-sampling apparatus and method for restoring a magnetic resonance image using a machine learning method capable of obtaining a high-quality reconstructed magnetic resonance image, A magnetic resonance image reconstruction apparatus and method using the same, and a recording medium therefor.

According to an aspect of the present invention, there is provided an apparatus for reconstructing a magnetic resonance image using learning, the method comprising: performing full sampling on a first region corresponding to a center of a first k-space image; An undersampling unit for undersampling the first k-space image by performing an undersampling for a second region excluding the first k-space image, and a magnetic resonance image based on the first k-space image through a previously learned neural network Space image based on the first k-space image in an area corresponding to the first spatial image in a second k-space image based on the output magnetic resonance image, and corrects the second k- And a magnetic resonance image reconstruction unit for generating a spatial image and acquiring a magnetic resonance image based on the third k-space image.

Wherein the MRI image reconstruction unit includes a first image transform unit for transforming the first k-space image to obtain a first MRI image, a second magnetic resonance imaging unit for receiving the first MRI image, A second image reconstruction unit for acquiring an image of the first k-space image, a second image transformation unit for transforming the second MRI image to acquire the second k-space image, A second image reconstruction unit that generates the third k-space image by correcting a portion overlapping the first k-space image with the first k-space image, and a second image reconstruction unit that transforms the third k- And a third image converting unit for acquiring the image.

The first image transform unit and the third image transform unit may use Fourier inverse transform, and the second image transform unit may use Fourier transform.

The first image reconstruction unit may perform a full sampling of a center region of a k-space image corresponding to a magnetic resonance image and convert a reference magnetic resonance image converted from a k-space image obtained by undersampling the remaining region of the k- And the magnetic resonance image converted from the reference-pool sampled k-space image may be learned in advance as a label.

The first area may be set to an area corresponding to 3% to 8% of the center of the first k-space image.

The undersampling performed in the second area may be undersampling for performing encoding by selecting one line for at least four lines.

The apparatus for reconstructing a magnetic resonance image using the learning may further include an area setting unit setting an area corresponding to the center of the first k-space image as the first area and setting the remaining area as the second area have.

According to another aspect of the present invention, there is provided an apparatus for reconstructing a magnetic resonance image using learning, comprising: a decomposition unit that generates nk1 low-resolution input images using k1 first filters; And a reconstruction unit that generates a high-resolution output image using k2 second-pair filters, wherein the high-resolution input image is a k-space image Space image by performing under sampling on a second region excluding the first region in the k-space image, and the k2-space image generated by k2 The second pair of filters generates a reference high resolution label image for learning of the neural network as nk2 low resolution label images It may be a filter for ssang k2 of the second filter.

Wherein the learning unit includes a neural network that is learned so that the n1 reference low resolution input images generated by repeating the k1 filters n times in the reference high resolution input image are nk2 reference low resolution label images obtained through in-depth learning, nk2 reference low resolution label images can be generated by repeating the k2 filters n times for a high resolution label image.

The learning unit may be different from the neural network corresponding to the first low-resolution input image among the k1 low-resolution input images and the neural network corresponding to the second low-resolution input image among the k1 low-resolution input images.

According to another aspect of the present invention, there is provided a method for reconstructing a magnetic resonance image using learning, the method comprising: performing full sampling on a first region corresponding to a center of the first k-space image; Sampling the first k-space image by performing undersampling on a second region excluding the first region, outputting a magnetic resonance image based on the first k-space image through a pre-learned neural network Space image based on the first k-space image in an area corresponding to the first spatial image in a second k-space image based on the output magnetic resonance image, and corrects the second k- Generating a spatial image and acquiring a magnetic resonance image based on the third k-space image.

The step of outputting a magnetic resonance image based on the first k-space image through a previously learned neural network may include the steps of acquiring a first magnetic resonance image by converting the first k-space image, And acquiring the second magnetic resonance image through the learned neural network.

The obtaining of the first magnetic resonance image may include a step of Fourier inversely transforming the first k-space image to obtain the first magnetic resonance image.

Wherein the generating the third k-space image comprises: transforming the output MRI image to obtain the second k-space image, and generating the first k-space image in the second k- And generating the third k-space image by correcting a portion overlapping the first k-space image with the first k-space image.

The step of acquiring the second k-space image may include a step of acquiring the second k-space image by Fourier transforming the output MRI image.

The pre-learned neural network may include a means for performing a full sampling of a center region of a k-space image corresponding to a magnetic resonance image and inputting a reference magnetic resonance image converted from a k-space image obtained by undersampling the remaining region of the k- Value, and the magnetic resonance image converted from the reference pool sampled k-space image may be learned in advance by using the label.

The method may further include setting a region corresponding to the center of the first k-space image to the first region and setting the remaining region to the second region .

According to another aspect of the present invention, there is provided a method for reconstructing a magnetic resonance image using learning, the method comprising: performing full sampling on a first region corresponding to a center of a k-space image; Converting the k-space image subjected to undersampling to generate a high-resolution input image, decomposing the high-resolution input image into nk1 low-resolution input images using k1 first filters, generating a nk2 low-resolution output image by applying a corresponding neural network according to nk1 low-resolution input images, and generating a high-resolution output image using the nk2 low-resolution output images using nk2 second pair filters, Wherein the k2 second pair filters are configured to classify the reference high resolution label image for learning of the neural network into nk2 low resolution labels, Generating a filter may be ssang for k2 of the second filter.

The neural network learns nk1 reference low resolution input images generated by repeating the k1 filters n times in the reference high resolution input image to be nk2 reference low resolution label images obtained through in-depth learning, and nk2 reference low resolution The label image may be generated by repeating the k2 filters n times for a high resolution label image.

Wherein generating the nk2 low resolution output images comprises applying a first neural network corresponding to a first low resolution input image of the nk1 low resolution input images and generating a second low resolution input image corresponding to the second low resolution input image among the nk1 low resolution input images And applying a second neural network.

The recording medium according to another aspect of the present invention may be recorded with a program for performing a magnetic resonance image reconstruction method using the above-described learning.

The present invention is advantageous in that a high quality reconstructed magnetic resonance image can be obtained while reducing a photographing time.

1 is a diagram illustrating a magnetic resonance image transformed according to the acquired k-space image.
FIG. 2 is a view for explaining a process of undersampling and MRI reconstruction according to a preferred embodiment of the present invention. Referring to FIG.
3 is a structural diagram of an under-sampling apparatus for reconstructing a magnetic resonance image using learning according to an embodiment of the present invention.
FIG. 4 is a diagram illustrating a comparison between a full sampled k-space image and a k-space image subjected to undersampling and a first k-space image.
5 is a diagram illustrating a magnetic resonance image obtained by converting a full sampled k-space image, a k-space image subjected to undersampling, and a first k-space image, respectively.
6 is a structural diagram of a magnetic resonance image reconstruction apparatus using learning according to a preferred embodiment of the present invention.
7 is a detailed structural diagram of a magnetic resonance image reconstruction unit of the present invention.
FIG. 8 is a view for explaining a procedure for acquiring a reconstructed magnetic resonance image of the MRI reconstruction unit of the present invention.
FIG. 9 is a diagram for explaining a combined product neural network algorithm. FIG.
10 is a diagram for explaining a convolution method of a composite-object-based neural network.
11 is a view for explaining a downsampling method of a composite-articulated network.
FIG. 12 is a flowchart illustrating an undersampling method for restoring a magnetic resonance image using learning according to a preferred embodiment of the present invention, according to time.
FIG. 13 is a flowchart illustrating a method for restoring a magnetic resonance image using learning according to a preferred embodiment of the present invention, according to the flow of time.
FIG. 14 shows an example of comparison of magnetic resonance images generated through embodiments of the present invention.
15 is a diagram illustrating a depth learning apparatus according to another embodiment of the present invention.
16 is a view illustrating a magnetic resonance image reconstruction apparatus according to another embodiment of the present invention.
17 is a block diagram of an apparatus for reconstructing a magnetic resonance image according to another embodiment of the present invention.

While the invention is susceptible to various modifications and alternative forms, specific embodiments thereof are shown by way of example in the drawings and will herein be described in detail. It should be understood, however, that the invention is not intended to be limited to the particular embodiments, but includes all modifications, equivalents, and alternatives falling within the spirit and scope of the invention. Like reference numerals are used for like elements in describing each drawing.

The terms first, second, etc. may be used to describe various components, but the components should not be limited by the terms. The terms are used only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be referred to as a second component, and similarly, the second component may also be referred to as a first component. Hereinafter, embodiments according to the present invention will be described in detail with reference to the accompanying drawings.

A magnetic resonance imaging (MRI) device obtains a k-space image through photography and converts it to obtain a magnetic resonance image.

1 is a diagram illustrating a magnetic resonance image transformed according to the acquired k-space image.

Referring to FIG. 1 (a), if a full-sampled k-space image is transformed, an accurate magnetic resonance image can be obtained, but it takes a long time to acquire an image, which is inefficient.

Referring to FIG. 1 (b), if the under-sampled k-space image is used to reduce the image acquisition time, the converted MRI image becomes inaccurate.

Therefore, when an undersampled k-space image is used to reduce the shooting time, a reconstruction process is required to obtain a high-quality magnetic resonance image.

In the embodiment according to the present invention, undersampling is performed at regular intervals, and the reason thereof is shown in Fig. For example, when one line is selected for every four lines as shown in FIG. 1B, the information of the four folded image can be obtained. The embodiment learns and restores the folded image information and the unfolded low resolution blurred image information obtained in the center 3% - 8% region.

The present invention discloses an under-sampling and a magnetic resonance image reconstruction process for obtaining an accurate magnetic resonance image using an undersampled k-space image.

For example, the embodiment uses a deep learning method of a convolutional neural network to recover a desired image using an under-sampled k-space image. The modified image obtained from the undersampled data is input to the composite neural network, and the composite neural network is learned to output the original image obtained from the full sampled data. The embodiment additionally measures not only uniform sampling but also data of an additional low frequency region in order to efficiently acquire an original image, adds the data to the data, and learns the composite neural network using the modified image derived from the data. FIG. 2 is a view for explaining a process of undersampling and MRI reconstruction according to a preferred embodiment of the present invention. Referring to FIG.

Referring to FIG. 2, the present invention can obtain a reconstructed magnetic resonance image through (A) undersampling, (B) deep-running, and (C) correction.

(A) In the undersampling process, full sampling may be performed on the first region and undersampling may be performed on the second region to acquire the first K-space image S.

Full sampling implies 100% encoding of all parts of the k-space image, and undersampling implies encoding only part of the k-space image. For example, undosampling can be performed by selecting and encoding only one line per four lines in a k-space image. The selected lines may be selected at regular intervals. That is, if one line is encoded, undosampling can be performed by skipping the next three lines and selecting and encoding the fourth line.

The K-space image can be converted into a magnetic resonance image using the Fourier inverse transform, and the MRI image can be transformed into the K-space image using the Fourier transform.

(B) During the deep learning process, a magnetic resonance image obtained by inverse Fourier transforming the first k-space image S is output to the first magnetic resonance image through the learned neural network. The learned neural network may be a composite neural network, and the composite neural network may use a magnetic resonance image converted from a reference first k-space image as an input value, and a magnetic resonance image converted from a reference full sampling sampled k-space image as a label Can be learned in advance. The learned neural network of the present invention is not limited to the composite neural network but other neural networks that can provide the same functions and actions can be applied.

(C) In the correction process, (B) the first magnetic resonance image obtained in the deep learning process is converted back to the second K-space image, and then the first K- Acquires a third k-space image by correcting the second K-space image based on the first k-space image (S). Finally, the third k-space image is transformed using the inverse Fourier transform to obtain the reconstructed second MRI image.

Now, the structure of an undersampling apparatus for restoring a magnetic resonance image using learning according to a preferred embodiment of the present invention will be described in detail.

3 is a structural diagram of an under-sampling apparatus for reconstructing a magnetic resonance image using learning according to an embodiment of the present invention.

Referring to FIG. 3, an under-sampling apparatus 400 for reconstructing a magnetic resonance image using learning according to an embodiment of the present invention may include an area setting unit 420 and an undersampling unit 430.

The undersampling device 400 acquires a K-space image through magnetic resonance imaging. In order to acquire K-space images by full sampling, it takes a long time to take a picture, resulting in an increase in shooting cost and side effects such as claustrophobia. Therefore, the undersampling apparatus 400 of the present invention can be operated so that an accurate magnetic resonance image can be reconstructed with only undersampled K-space images.

The region setting unit 420 may set a region corresponding to the center of the first k-space image to be acquired as a first region and a remaining region as a second region.

For example, the first region may be set to a region corresponding to 3% to 8% of the center of the k-space image, and the remaining region may be set to the second region.

It is not easy to perform the learning to accurately correct the converted MRI image because the separation condition is not satisfied if a magnetic resonance image converted into only the k-space image uniformly undersampled over the entire region is used in the learning process. Therefore, according to the present invention, undersampling is performed on the second area, but full sampling is performed in the first area corresponding to the center part, so that a more accurate magnetic resonance image can be restored.

The undersampling unit 430 performs full sampling for the first area and undersampling for the second area. Here, full sampling implies 100% encoding of all parts of the k-space image, and undersampling implies encoding only a part of the k-space image. For example, undosampling can be performed by selecting and encoding only one line per four lines in a k-space image. The selected lines may be selected at regular intervals. That is, if one line is encoded, undosampling can be performed by skipping the next three lines and selecting and encoding the fourth line.

For example, the undersampling unit 430 performs under-sampling on the second area as shown in FIG. 1 (b), and performs full sampling on the first area corresponding to the center part. As a result, the undersampling unit 430 may perform undersampling including a first region and a second region in a total encoding of 27% to 33%.

FIG. 4 illustrates a comparison between a full sampled k-space image and a k-space image and a first k-space image subjected to undersampling, FIG. 5 illustrates a case where a full sampled k-space image and undersampling are performed A k-space image, and a first k-space image, respectively.

4 (a) is a full sampled k-space image, FIG. 4 (b) is a k-space image subjected to undersampling and FIG. 4 (c) is a first k-space image. 5A, 5B, and 5C are magnetic resonance images obtained by converting each of FIGS. 4A, 4B, and 4C. FIG.

Referring to FIGS. 4 and 5, FIG. 5B includes the characteristics of the image, but it is difficult to accurately determine the location of important information such as cancer cell information indicated by a red arrow in the drawing. On the other hand, FIG. 5 (c) can accurately identify the location of important information such as cancer cell information indicated by a red arrow on the drawing while well-containing features of the image.

Now, a structure of a magnetic resonance image reconstruction apparatus using learning according to a preferred embodiment of the present invention will be described in detail.

6 is a structural diagram of a magnetic resonance image reconstruction apparatus using learning according to a preferred embodiment of the present invention.

6, an apparatus for reconstructing a magnetic resonance image using learning according to an exemplary embodiment of the present invention may include an area setting unit 105, an undersampling unit 110, and a magnetic resonance image reconstructing unit 120 have.

The area setting unit 105 and the undersampling unit 110 may have the same configuration as the area setting unit 420 and the undersampling unit 430 of the undersampling apparatus 400 described above. The undersampling unit 110 can undersample the first k-space image.

The magnetic resonance image reconstruction unit 120 acquires a reconstructed magnetic resonance image using the first k-space image.

FIG. 7 is a detailed structural diagram of a magnetic resonance image reconstruction unit of the present invention, and FIG. 8 is a view for explaining a reconstruction magnetic resonance image acquisition procedure of the magnetic resonance image reconstruction unit of the present invention.

7, the magnetic resonance image reconstruction unit 120 includes a first image transformation unit 121, a first image reconstruction unit 122, a second image transformation unit 123, a second image reconstruction unit 124, And a third image conversion unit 125. [

7 and 8, the first image transform unit 121 may obtain the first MRI image 20 by inverse Fourier transforming the first k-space image 10, The second magnetic resonance imaging apparatus 120 may acquire the second MRI image 30 by receiving the first MRI image 20. The first image reconstruction unit 122 uses the composite neural network, and the composite neural network transforms the magnetic resonance image converted from the reference first k-space image to the input value, and converts the reference full sampled k- It can be learned in advance by using the resonance image as a label.

The second image transforming unit 123 may Fourier transform the second MRI image 30 to obtain the second k-space image 40. The second image reconstructing unit 124 may obtain the second K- -Space image 40 based on the first K-space image 10 by correcting the second k-space image 40 in the region corresponding to the first k-space image 10 in the spatial image 40, The image 50 can be acquired.

Finally, the third image transform unit 125 transforms the third k-space image 50 to acquire the reconstructed magnetic resonance image 60.

Now, each part of the magnetic resonance image reconstruction unit 120 will be described in more detail.

The first image transform unit 121 transforms the first k-space image 10 to acquire the first MRI image 20. The first k-space image 10 is transformed into a first magnetic resonance image 20 using Fourier inverse transform.

The first image reconstruction unit 122 receives the first MRI image 20 and acquires the second MRI image 30. The first image reconstruction unit 122 may acquire the second MRI image 30 using the learned neural network.

For learning, the present invention may utilize a deep learning algorithm and, as an example, use a Convolutional Neural Network (CNN) algorithm. The composite neural network is a model that simulates human brain function based on the assumption that when a person recognizes an object, it extracts the basic features of the object and then recognizes the object based on the result after complicated calculation in the brain. Recently, it is widely used in image recognition and speech recognition. In the composite neural network, various filters for extracting features of images through convolution (Conv) operation are basically used together with pooling or non-linear activation functions for adding nonlinear characteristics. do.

In the following, a brief description of the composite neural network algorithm will now be given, with reference to the accompanying drawings, of embodiments according to the present invention.

FIG. 9 is a diagram for explaining a combined product neural network algorithm. FIG. FIG. 10 is a view for explaining a convolution method of a convolutional neural network, and FIG. 11 is a diagram for explaining a downsampling method of a convolutional neural network.

Referring to FIG. 9, the composite neural network algorithm extracts a feature map for an input image through convolution and downsampling for an input image, and identifies or classifies an input image through a feature map (classification). The feature map includes feature information on the input image. For extracting the feature map, convolutions (C1, C2, C3) and downsampling (MP1, MP2) are repeated, and the number of repetitions can be variously determined according to the embodiment.

9 and 10, when the size of the filter (or the kernel 210) used for the convolution is determined, a weighted sum of products between weight values assigned to the respective pixels of the filter and pixel values of the input image 200, The convolution is performed. That is, the pixel value 230 of the convolution layer can be determined by multiplying the pixel value of the pixel corresponding to the specific region of the input image where the filter overlaps with the weight of the filter, and then adding all the multiplied values of the region where the filter overlaps .

As shown in FIG. 10, in a specific region of the input image 200 overlapping with the weights (4, 0, 0, 0, 0, 0, 0, 0, A sum is performed on the multiplication result between the pixel values (0, 0, 0, 0, 1, 1, 0, 1, 2) to determine the pixel value 230 of the final -8. The size of the input image 200 is 7 × 7, the size of the filter 210 is 3 × 3 (the size of the input image 200 is 3 × 3) A convolution layer of size 5X5 can be generated.

Since the pixel value according to the convolution becomes the pixel value 230 of the center pixel of the overlapped area, the size of the convolutional layer, i.e., the convoluted image, relative to the input image decreases. However, when padding an outer area of an input image with a specific pixel value, a convolution layer having a size of 7 × 7 equal to the size of the input image can be generated. The number of convolution layers is determined by the number of filters used.

Referring to FIGS. 9 and 11, downsampling is performed to reduce the size of the convolution layer, that is, to reduce the resolution. A commonly used method for downsampling is max-pooling (MP). A maximum pulling layer smaller than the size of the convolution layer can be generated by taking the maximum value among the pixel values of the convolution layer included in the kernel used for downsampling.

For example, when a 2X2 kernel is applied to a convolution layer 310 of 4X4 size, 6, 8, 3 and 4 are determined to be maximum values for 2X2 regions indicated by different colors, .

9, the feature map is input to the fully-connected neural network, and the learning of the parameters of the combined-product neural network is performed according to the difference value between the label for the given full-sampling input image and the output value of the neural network. Is performed.

As described above, the first image reconstructing unit 122 reconstructs a reference first MRI image generated by the reference first k-space image as an input value, and generates a magnetic resonance image generated by the reference full sampled k-space image It may be learned in advance by label.

The second image transforming unit 123 may transform the second MRI image 30 to obtain the second k-space image 40. The second k-space image 40 is acquired for the next correction. A Fourier transform can be used for the transformation.

The second image reconstructing unit 124 may generate the fourth k-space image 50 by correcting the second k-space image 40 using the first k-space image 10. Since the second MRI image 30 is acquired from the first MRI image 20 by learning, the second k-space image 40, which has been converted from the second MRI image 30, Some of the information of the image 10 may be deformed. Therefore, the second image reconstructing unit 124 reconstructs the third k-space image 10 in the second k-space image 40 by replacing the portion overlapping the first k-space image 10 with the first k-space image 10, k-space image 50 as shown in FIG.

Finally, the third image transforming unit 125 transforms the third k-space image 50 to acquire the reconstructed magnetic resonance image 60. Fourier inverse transform can be used for the transform.

As described above, the present invention performs full sampling in the first area and undersampling in the second area to undersample the k-space image. Therefore, the imaging time of the magnetic resonance image is short, The image can be obtained.

FIG. 12 is a flowchart illustrating an undersampling method for restoring a magnetic resonance image using learning according to a preferred embodiment of the present invention, according to time.

12, an undersampling method for restoring a magnetic resonance image using learning according to an exemplary embodiment of the present invention may include an area setting step S510 and a first k-space image obtaining step S520 .

The area setting step S510 is a step of setting the first area and the second area in the area setting part 420. [

The first k-space image acquiring step (S520) is a step of undersampling the first k-space image by the undersampling unit (430).

FIG. 13 is a flowchart illustrating a method for restoring a magnetic resonance image using learning according to a preferred embodiment of the present invention, according to the flow of time.

Referring to FIG. 13, a method for reconstructing a magnetic resonance image using learning according to an exemplary embodiment of the present invention includes a region setting step S610, a first k-space image acquiring step S620, (S630), a second MRI image acquisition step (S640), a second k-space image acquisition step (S650), a third k-space image acquisition step (S660), and a reconstructed MRI image acquisition step (S670) .

The region setting step S610 and the first k-space image acquiring step S620 may include the area setting step S510 and the first k-space image acquiring step in the undersampling method for restoring the magnetic resonance image using the above- (S520).

The first MRI image acquisition step S630 is a step of acquiring the first MRI image 20 by converting the first k-space image 10 in the first image conversion unit 121. [

The second MRI image acquisition step S640 is a step of acquiring the second MRI image 30 using the learned neural network by receiving the first MRI image 20 from the first image reconstruction unit 122 .

The second k-space image acquisition step S650 is a step of acquiring the second k-space image 40 by converting the second MRI image 30 in the second image conversion unit 123. [

The third k-space image acquisition step S660 is a step of acquiring the second k-space image 10 from the second image reconstructing unit 124 based on the first k-space image 10, - acquiring the third k-space image (50) by correcting the spatial image (40).

The reconstructed MRI image acquisition step S670 is a step of acquiring a reconstructed MRI image 60 by transforming the third k-space image 50 in the third image transformation unit 125. [

FIG. 14 shows an example of comparison of magnetic resonance images generated through embodiments of the present invention.

14, the resolution of the image is 256X256, FIG. 14A is a magnetic resonance image implemented through full sampling, FIG. 14B is a magnetic resonance image implemented through conventional undersampling, And is a magnetic resonance image implemented according to an embodiment of the present invention.

As shown in FIG. 14, the magnetic resonance image c provided by the embodiment of the present invention is an image quality much improved compared to the conventional (b), and the image quality close to (a).

를 적용하여, 차원이 축소된 256X256 픽셀의 자기공명영상

이 생성될 수 있다. 여기에서 {

}

는 웨이블렛 풀링 필터에 해당하며,

는 스트라이드 2 합성곱 연산자이며,

이다. For a high-resolution image, the dimension (number of pixels) of the input data is high, and deep learning may not be possible. An in-depth learning method using wavelet pooling can be applied to the present invention to solve the dimensionality problem of the input data. This can be a method of applying dimensionality reduction of input data using wavelet pulling and then performing deep learning on data with reduced dimension to restore the magnetic resonance image. For example, let the number of pixels of a magnetic resonance image to be reconstructed be 512 512. At this time, the magnetic resonance image modified in the undersampling data is referred to as a first magnetic resonance image (I). A first magnetic resonance image (I) of 512.times.512 pixels is applied to a wavelet-pulling function

, A 256x256 pixel magnetic resonance image with reduced dimensions

Can be generated. From here {

}

Corresponds to a wavelet pulling filter,

Is a stride 2 composite product operator,

to be.

라 하면 다음 수학식 1을 만족한다. The corresponding restoration sites

The following equation (1) is satisfied.

[Equation 1]

}는 {

}의 쌍 필터(dual filter)에 해당하며,

는 스트라이드 1 합성곱 연산자이고,

는 업 샘플링(up-sampling) 연산자이다.From here, {

} Is {

}, Which corresponds to a dual filter,

Is a stride 1 composite product operator,

Is an up-sampling operator.

을 생성하고, 복수의 자기공명영상

에 심층학습을 적용하여 복원된 자기공명영상들을 종합하여 고해상도 자기공명영상을 제공할 수 있다.Another embodiment of the present invention is a method of reconstructing a plurality of magnetic resonance images of a lower dimension than a dimension of an undersampled first magnetic resonance image (I)

And generates a plurality of magnetic resonance images

And a high-resolution magnetic resonance image can be provided by synthesizing the reconstructed magnetic resonance images.

와 {

}라고 하고, 출력 영상에 적용한 웨이블렛 풀링 작용소

의 복원 작용소와 쌍 필터를

, {

} 라고 하자. 그리고, 적용한 심층학습을"U-net"이라 하면, 웨이블릿 풀링과 심층 학습을 이용한 복원 과정은 수학식 2로 나타낼 수 있다.The wavelet pooling function applied to the input image and the filter

Wow {

}, And a wavelet pulling function applied to the output image

And a pair filter

, {

}. If the applied in-depth learning is referred to as " U-net ", the reconstruction process using wavelet pulling and in-depth learning can be expressed by Equation (2).

&Quot; (2) "

15 is a diagram illustrating a depth learning apparatus according to another embodiment of the present invention.

15, the depth learning apparatus 700 includes a first decomposition unit 710 for decomposing an undersampled reference high-resolution input image x, a decomposition unit 710 for decomposing a full sampled reference high-resolution label image y, 2 decomposition unit 720, and a learning unit 730. [ The high-resolution input image (x) is an image including an evenly sampled image and an image in a low-frequency domain.

}를 n번 반복하여 nk1개의 레퍼런스 저해상도 입력 영상(

)을 생성한다. The first decomposition unit 710 decomposes the reference high resolution input image xr into k1 filters {

} Is repeated n times to obtain nk1 reference low resolution input images (

).

}를 n번 반복하여 nk2개의 레퍼런스 저해상도 레이블 영상(

)를 생성한다.The second decomposition unit 720 decomposes the high resolution label image yr into k2 filters {

} Is repeated n times and nk2 reference low-resolution label images (

).

)과 nk2개의 레퍼런스 저해상도 레이블 영상(

) 간의 관계를 딥러닝을 통해 학습한다. 딥러닝 방식은 앞서 언급한 합성곱 신경망일 수 있다. The learning unit 730 includes nk1 reference low resolution input images (

) And nk2 reference low resolution label images (

) Through deep learning. The deep learning method may be the aforementioned composite neural network.

}를 이용하여 해상도 2^k-1 ×2^{k- 1} 의 k1개의 영상(711)을 생성한다. 제1 분해부(710)는 다시 k1개의 필터 {

}를 이용하여 해상도 2^k-2 ×2^{k- 2} 의 2k1개의 영상(712)을 생성한다. 제1 분해부(710)는 위와 같은 동작을 반복하여, k1 개의 필터 {

}를 n번 반복하여 nk1개의 레퍼런스 저해상도 입력 영상(

)(713)을 생성한다. Specifically, the resolution of the original image of high resolution is 2 ^k × 2 ^k , the first decomposition unit 710 includes k 1 filters {

} To generate a k1 of images 711 of the resolution of 2 × 2 ^{k1 k1} used. The first decomposing unit 710 further includes k1 filters {

} To generate a 2k1 of images 712 of the resolution of ^{2 k-2} × 2 ^k- 2 used. The first decomposing unit 710 repeats the above operation so that k1 filters {

} Is repeated n times to obtain nk1 reference low resolution input images (

) ≪ / RTI >

}를 이용하여 해상도 2^k-1 ×2^{k- 1} 의 k2개의 영상(721)을 생성한다. 제2 분해부(720)는 다시 k2개의 필터 {

}를 이용하여 해상도 2^k-2 ×2^{k- 2} 의 2k2개의 영상(722)을 생성한다. 제2 분해부(720)는 위와 같은 동작을 반복하여, k2 개의 필터 {

}를 n번 반복하여 nk2개의 레퍼런스 저해상도 레이블 영상(

)(723)을 생성한다. Similarly, if the resolution of the high-resolution label image is 2 ^k x 2 ^k, then the second decomposing unit 720 may include k 2 filters {

} K2 to generate a single image 721 of the resolution of 2 ^k-1 × 2 ^{k- 1} used. The second decomposing unit 720 further includes k2 filters {

} To generate a 2k2 of images 722 of the resolution of ^{2 k-2} × 2 ^k- 2 used. The second decomposing unit 720 repeats the above operation to obtain k2 filters {

} Is repeated n times and nk2 reference low-resolution label images (

) ≪ / RTI >

)(713)에 대하여 심층학습으로 통해 얻은 결과가 레퍼런스 저해상도 레이블 영상(

)(723)이 되도록 학습되는 신경망을 포함할 수 있다. 이 때, 학습부(730)는 nk1개의 레퍼런스 저해상도 입력 영상(

)(713)에 따라 nk2개의 레퍼런스 저해상도 레이블 영상 (

)(723) 중 대응하는 레퍼런스 저해상도 레이블 영상을 다르게 심층학습을 수행할 수 있다. 예를 들어, 학습부(730)는 nk1개의 레퍼런스 저해상도 입력 영상 (

)(713) 중 일부에 대해서 nk2개의 레퍼런스 저해상도 레이블 영상 (

)(723) 중 대응하는 일부의 레퍼런스 저해상도 레이블 영상이 되도록 학습을 수행할 수도 있다.The learning unit 730 receives the reference low resolution input image (

(713) is a reference low-resolution label image

) ≪ / RTI > (723). At this time, the learning unit 730 generates nk1 reference low-resolution input images (

) ≪ / RTI > (713)

) 723 of the corresponding reference low-resolution label image. For example, the learning unit 730 may include nk1 reference low resolution input images (

(713), nk2 reference low-resolution label images

) 723 of the reference low-resolution label image.

 The learning unit 730 learned in this manner can be used to restore a high-resolution input image to a magnetic resonance image.

16 is a view illustrating a magnetic resonance image reconstruction apparatus according to another embodiment of the present invention.

The magnetic resonance image reconstruction apparatus 800 includes a decomposition unit 810, a learning unit 820, and a reconstruction unit 830.

}를 n회 적용하여 저해상도 입력 영상

을 생성한다.The decomposition unit 810 decomposes the high resolution input image Xi into a plurality of filters {

} ≪ / RTI > is applied n times,

.

에 따라 적용되는 심층학습 신경망을 다르게 할 수 있다. 학습부(820)는 저해상도 입력 영상

에 대응하는 심층학습 신경망을 적용하여

를 근사하는 저해상도 출력 영상

을 생성한다. 이 때, 심층학습 신경망은 합성곱 신경망일수 있다. 이 때,

는 입력 영상 Xi에 필터{

}를 n번 적용하여 변환한 저해상도 영상이다.The learning unit 820 includes an in-depth learning neural network constructed by the learning unit 730 through learning. The learning unit 820 includes at least two deep learning neural networks,

The depth learning neural network applied can be different. The learning unit 820 includes a low-

To apply the deep learning neural network corresponding to

A low-resolution output image

. In this case, the deep learning neural network may be a composite neural network. At this time,

Lt; RTI ID = 0.0 > {

} N times.

에 제2 분해부(720)의 필터{

}의 쌍필터(dual filter) {

를 이용하여 합성 연산을 n번 반복하여 고해상도 입력 영상 Xi에 대응되는 풀샘플링에 의해 생성된 고해상도 라벨 영상 Yi를 근사하는 복원된 고해상도 출력 영상(

)을 생성한다. 고해상도 라벨 영상 Yi는 풀샘플링한 라벨 영상이다. The reconstruction unit 830 reconstructs the low-resolution output image

The filter of the second decomposing unit 720

} Dual filter of {

, The composite operation is repeated n times to obtain a reconstructed high-resolution output image (" Yi ") that approximates the high-resolution label image Yi generated by full sampling corresponding to the high-

). The high-resolution label image Yi is a full-sampled label image.

17 is a block diagram of an apparatus for reconstructing a magnetic resonance image according to another embodiment of the present invention.

가 같은 2D 하르 필터로 k1과 k2는 4이다.17, the high resolution input image and the high resolution label image have a resolution of 512 (= 2 ⁹ ) X512, n is 1,

Are the same 2D filter, k1 and k2 are 4.

)를 1번 적용하여 256(=2⁸)X256의 저해상도 입력

를 생성한다.As shown in FIG. 17, the decomposition unit 810 decomposes the high-resolution input image Xi into four filters (

) Is applied once to obtain a low-resolution input of 256 (= 2 ⁸ ) X256

.

을 심층학습 신경망(811)에 입력하여

을 근사하는 저해상도 출력 영상(

)을 생성하며, 저해상도 입력

을 심층학습 신경망(812)에 입력하여

을 근사하는 저해상도 출력 영상 (

)을 생성한다. 도 17에서 보여주는 예시는 발명을 설명하기 위한 일 예시일 뿐, 본 발명이 이에 한정되는 것은 아니다.The learning unit 820 includes two deep learning neural networks 811 and 812,

Into the deep learning neural network 811

Resolution output image (

), And a low-resolution input

Is input to the deep learning neural network 812

Resolution output image (

). The example shown in FIG. 17 is only an example for explaining the invention, but the present invention is not limited thereto.

)에 쌍필터

를 적용하여 풀샘플링에 의해 생성된 고해상도 라벨 Yi를 근사하는 512X512의 고해상도 출력 영상(

)을 생성한다.The reconstruction unit 830 reconstructs the low-

) Pair filter

Resolution high-resolution output image (512x512) that approximates the high-resolution label Yi generated by full sampling

).

The above-described technical features may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, and the like, alone or in combination. The program instructions recorded on the medium may be those specially designed and constructed for the embodiments or may be available to those skilled in the art of computer software. Examples of computer-readable media include magnetic media such as hard disks, floppy disks and magnetic tape; optical media such as CD-ROMs and DVDs; magnetic media such as floppy disks; Magneto-optical media, and hardware devices specifically configured to store and execute program instructions such as ROM, RAM, flash memory, and the like. Examples of program instructions include machine language code such as those produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

As described above, the present invention has been described with reference to particular embodiments, such as specific elements, and specific embodiments and drawings. However, it should be understood that the present invention is not limited to the above- Those skilled in the art will appreciate that various modifications and changes may be made thereto without departing from the scope of the present invention. Accordingly, the spirit of the present invention should not be construed as being limited to the embodiments described, and all of the equivalents or equivalents of the claims, as well as the following claims, belong to the scope of the present invention .

Claims (21)

Spatial image, performing a full sampling on a first region corresponding to a center of the first k-space image, performing under sampling on a second region excluding the first region in the first k-space image, An undersampling unit for undersampling the spatial image; And
And outputting a magnetic resonance image based on the first k-space image through a previously learned neural network, and generating an area corresponding to the first spatial image in a second k-space image based on the output magnetic resonance image, Space image, and generating a third k-space image by correcting the second k-space image based on the first k-space image, and acquiring a magnetic resonance image based on the third k- ,
Magnetic Resonance Imaging System Using Learning. The method according to claim 1,
Wherein the MRI reconstruction unit comprises:
A first image transform unit for transforming the first k-space image to obtain a first MRI image;
A first image reconstruction unit that receives the first MRI image and acquires a second MRI image through the learned neural network;
A second image conversion unit for converting the second MRI image to acquire the second k-space image;
A second image reconstruction unit for generating the third k-space image by correcting a portion overlapping the first k-space image with the first k-space image in the second k-space image; And
And a third image transformer for transforming the third k-space image to obtain a reconstructed magnetic resonance image,
Magnetic Resonance Imaging System Using Learning. 3. The method of claim 2,
Wherein the first image transform unit and the third image transform unit use a Fourier transform and the second image transform unit uses a Fourier transform.
Magnetic Resonance Imaging System Using Learning. The method according to claim 1,
Wherein the first image reconstructing unit comprises:
A center area of a k-space image corresponding to a magnetic resonance image is sampled and a reference magnetic resonance image transformed from a k-space image obtained by undersampling the remaining area of the k-space image is used as an input value, - a magnetic resonance image transformed from a spatial image is learned in advance with a label,
Magnetic Resonance Imaging System Using Learning. The method according to claim 1,
Wherein the first region is set to an area corresponding to 3% to 8% of the center of the first k-space image.
Magnetic Resonance Imaging System Using Learning. The method according to claim 1,
Wherein the undersampling performed in the second area is undersampling for performing encoding by selecting one line for at least four lines.
Magnetic Resonance Imaging System Using Learning. The method according to claim 1,
Further comprising an area setting unit setting an area corresponding to the center of the first k-space image as the first area and setting the remaining area as the second area.
Magnetic Resonance Imaging System Using Learning. A decomposition unit for generating nk1 low-resolution input images using k1 first filters;
A learning unit for generating nk2 low-resolution output images by applying corresponding neural networks according to the nk1 low-resolution input images; And
And a reconstruction unit for generating k2 low-resolution output images by using k2 second-pair filters,
Wherein the high-resolution input image performs full sampling on a first region corresponding to a center of the k-space image, and performs k-space on the k-space image by performing undersampling on a second region excluding the first region, Image,
Wherein the k2 second pair filters are pair filters for k2 second filters that generate reference high resolution label images for learning of the neural network into nk2 low resolution label images.
Magnetic Resonance Imaging System Using Learning. 9. The method of claim 8,
Wherein,
And a neural network in which n1 reference low resolution input images generated by repeating the k1 filters n times in a reference high resolution input image are nk2 reference low resolution label images obtained through depth learning,
Wherein the nk2 reference low resolution label images are generated by repeating the k2 filters n times for a high resolution label image.
Magnetic Resonance Imaging System Using Learning. 9. The method of claim 8,
Wherein,
Wherein the neural network corresponding to the first low-resolution input image among the k1 low-resolution input images and the neural network corresponding to the second low-resolution input image among the k1 low-resolution input images are different from each other.
Magnetic Resonance Imaging System Using Learning. Performing a full sampling on a first region corresponding to a center of the first k-space image, performing under sampling on a second region excluding the first region in the first k-space image, sampling the k-space image;
Outputting a magnetic resonance image based on the first k-space image through a pre-learned neural network;
The second k-space image is corrected based on the first k-space image in an area corresponding to the first spatial image in a second k-space image based on the output magnetic resonance image, Generating an image; And
And acquiring a magnetic resonance image based on the third k-space image. 12. The method of claim 11,
The step of outputting a magnetic resonance image based on the first k-space image through a pre-learned neural network includes:
Transforming the first k-space image to obtain a first MRI image; And
And acquiring the second magnetic resonance image through the learned neural network by receiving the first magnetic resonance image,
Magnetic resonance image reconstruction method using learning. 13. The method of claim 12,
Wherein acquiring the first magnetic resonance image comprises:
Space image; and inversely transforming the first k-space image to obtain the first magnetic resonance image.
Magnetic resonance image reconstruction method using learning. 12. The method of claim 11,
Wherein the generating the third k-space image comprises:
Transforming the output MRI image to obtain the second k-space image; And
Generating a third k-space image by correcting a portion of the second k-space image that overlaps the first k-space image with the first k-space image.
Magnetic resonance image reconstruction method using learning.
15. The method of claim 14,
Wherein acquiring the second k-space image comprises:
And acquiring the second k-space image by Fourier transforming the output MRI image.
Magnetic resonance image reconstruction method using learning. 12. The method of claim 11,
Wherein the pre-learned neural network comprises:
Sampling a center region of a k-space image corresponding to a magnetic resonance image and converting a reference magnetic resonance image converted from a k-space image obtained by undersampling the remaining region of the k-space image into an input value, and a magnetic resonance image transformed from the k-space image is learned in advance with the label.
Magnetic resonance image reconstruction method using learning. 12. The method of claim 11,
Further comprising setting an area corresponding to the center of the first k-space image as the first area and setting the remaining area as the second area.
Magnetic resonance image reconstruction method using learning. space image is transformed into a k-space image by performing full sampling on a first region corresponding to the center of the k-space image, and a k-space image obtained by performing undersampling on the second region excluding the first region in the k- Generating an input image;
Decomposing the high-resolution input image into nk1 low-resolution input images using k1 first filters;
Generating nk2 low-resolution output images by applying a corresponding neural network according to the nk1 low-resolution input images; And
And generating a high-resolution output image using the nk2 second-pair filters of the nk2 low-resolution output images,
Wherein the k2 second pair filters are pair filters for k2 second filters that generate reference high resolution label images for learning of the neural network into nk2 low resolution label images.
Magnetic resonance image reconstruction method using learning. 19. The method of claim 18,
The neural network,
The nk1 reference low resolution input images generated by repeating the k1 filters n times in the reference high resolution input image are learned so as to become nk2 reference low resolution label images obtained through the deep learning,
Wherein the nk2 reference low resolution label images are generated by repeating the k2 filters n times for a high resolution label image.
Magnetic resonance image reconstruction method using learning. 19. The method of claim 18,
Wherein the step of generating the nk2 low-
Applying a first neural network corresponding to a first low-resolution input image among the nk1 low-resolution input images; And
And applying a second neural network corresponding to a second low resolution input image of the nk1 low resolution input images.
Magnetic resonance image reconstruction method using learning. A computer-readable recording medium on which a program for performing the magnetic resonance image reconstruction method according to any one of claims 11 to 20 is recorded.

Images