Magnetic resonance imaging device and image processing method

Abstract

The present invention provides a magnetic resonance imaging apparatus and an image processing method that can obtain a high Gibbs ringing correction effect for various input data. A ringing correction unit is prepared, which performs Gibbs ringing correction on an image reconstructed using a reconstruction matrix size that forms a predetermined matrix ratio relative to the matrix size of measurement data. In image reconstruction, first, the image is reconstructed using a reconstruction matrix size that forms the predetermined matrix ratio between the matrix size of measurement data and the reconstruction matrix size, thereby obtaining an intermediate reconstructed image. After the intermediate reconstructed image is subjected to ringing correction by the ringing correction unit, it is converted into k-space data and reconstructed using the final image reconstruction matrix size.

Description

Technical Field

[0001] The present invention relates to a magnetic resonance imaging device, and in particular to a ringing correction technology for images acquired by the magnetic resonance imaging device. Background Art

[0002] The magnetic resonance imaging (MRI) device uses a gradient magnetic field to give phase encoding to the nuclear magnetic resonance signals (echo signals) generated by the subject, and when sampling the nuclear magnetic resonance signals, a gradient magnetic field in a direction orthogonal to the phase encoding direction is used to collect echo signals of different frequencies within the sampling time. The data containing multiple echo signals with different phase encodings becomes measurement data of a matrix size determined by the number of phase encodings and the number of samples, i.e., k-space data. The MRI device reconstructs the image of the subject by performing Fourier transform and other processing on the k-space data.

[0003] The matrix size during measurement (measurement matrix size) varies depending on imaging conditions such as device limitations or imaging time. On the other hand, the matrix size of the reconstructed image (reconstruction matrix size) is sometimes the same as the measurement matrix size, but in order to obtain an image of desired resolution, image reconstruction is usually performed with a matrix size different from the measurement matrix size. Thus, in MRI, the measurement matrix size and reconstruction matrix size are also various.

[0004] Since the measurement matrix size is discrete and finite in the frequency space, if the measurement matrix size is different from the reconstruction matrix size and the ratio thereof is changed from 1:1, periodic or stripe-shaped artifacts (hereinafter referred to as ringing) called Gibbs ringing or truncation artifacts occur more strongly in the reconstructed image. This ringing can be suppressed by, for example, applying a k-space filter that smoothes the high-frequency region to the measurement data, but there is a disadvantage that the image clarity is reduced and the image is blurred due to the application of the k-space filter.

[0005] In contrast, a technology for correcting ringing without reducing definition has also been proposed. For example, in Non-Patent Document 1, a deep learning model such as CNN that has been learned by using images reconstructed by combining various measurement matrix sizes and reconstruction matrix sizes has been proposed. In addition, the present applicant has also proposed a technology in which a ringing artifact removal process is performed before noise processing for an image reconstructed by expanding the measurement matrix by zero padding (Patent Document 1).

[0006] Non-patent literature 1: Qianqian Zhang et al. Magn Reson Med. 2019; 82: 2133-2145

[0007] Patent Document 1: Japanese Patent Application Publication No. 2021-171490

[0008] In the technology described in Patent Document 1, the image is subjected to noise processing in a sparse space, and the measured data is zero-filled to exceed the size of the reconstruction matrix, thereby expanding. However, since ringing does not become sparse in a sparse space, the effect as a ringing correction technology is low. In addition, in the technology disclosed in Non-Patent Document 1, the expected ringing correction effect cannot be obtained based on the data input to the learned CNN. In this technology, in order to cover various combinations of matrix sizes, a large number of images are required as CNN learning data, which is unrealistic. Summary of the invention

[0009] The object of the present invention is to provide a technology that can obtain a high Gibbs ringing correction effect for various input data. In particular, the object of the present invention is to provide a technology that can learn CNN with a relatively small amount of learning data and obtain a high Gibbs ringing correction effect in a ringing correction technology using CNN.

[0010] In order to solve the above-mentioned problems, the MRI apparatus of the present invention is prepared for a ringing correction unit that maximizes the ringing correction effect of an image reconstructed using a reconstruction matrix size that is a predetermined matrix ratio relative to the matrix size of the measurement data, and performs two stages of reconstruction in image reconstruction. In the first stage of reconstruction (first reconstruction), a reconstruction is performed so that the matrix size of the measurement data and the reconstruction matrix size become a matrix ratio that achieves the maximum effect by the ringing correction unit, to obtain an intermediate reconstructed image. After ringing correction is performed on the intermediate reconstructed image, a second stage of image reconstruction (second reconstruction) is performed as the final image reconstruction. In the second reconstruction, the reconstructed image after the ringing correction is converted into measurement data, and the measurement data is reconstructed in a manner that becomes the desired reconstruction matrix size, that is, the reconstruction matrix size set by the user.

[0011] Specifically, the MRI apparatus of the present invention comprises: an imaging unit that collects measurement data including magnetic resonance signals; an image generating unit that reconstructs the measurement data with a desired reconstruction matrix size; and a ringing correction unit that performs ringing correction on the reconstructed image.

[0012] The image generating unit comprises: a matrix size setting unit for setting a measurement matrix size of measurement data, an intermediate reconstruction matrix size having a predetermined matrix ratio relative to the measurement matrix size, and a desired reconstruction matrix size; a first reconstruction unit for reconstructing the measurement data with the intermediate reconstruction matrix size and generating an intermediate reconstructed image; and a second reconstruction unit for reconstructing k-space data of a corrected intermediate reconstructed image in which ringing correction has been performed on the intermediate reconstructed image by a ringing correction unit with the desired reconstruction matrix size.

[0013] In a preferred embodiment of the present invention, the ringing correction unit includes a CNN that has been learned using multiple image groups including a first image in which ringing does not occur and a second image in which ringing occurs, wherein the second image is an image reconstructed by resizing the first image in k-space according to the prescribed matrix ratio.

[0014] Furthermore, the image processing method of the present invention is an image processing method for reconstructing a ringing-corrected image using measurement data including magnetic resonance signals collected by a magnetic resonance imaging apparatus, and includes the following steps.

[0015] The first reconstruction step reconstructs the measurement data at a reconstruction matrix size having a predetermined matrix ratio relative to the measurement matrix size to generate an intermediate reconstructed image; the correction step performs ringing correction on the intermediate reconstructed image using a CNN that outputs an image in which ringing is corrected for an input image, the CNN learning a first image without ringing and a second image with ringing as a group; and the second reconstruction step converts the intermediate reconstructed image after ringing correction into k-space data and reconstructs it at a desired reconstruction matrix size.

[0016] The second image used in CNN learning is an image obtained by resizing the first image in k-space and reconstructing it with the reconstruction matrix size of the first image, and the reconstruction matrix size and the matrix size after resizing satisfy a predetermined matrix ratio.

[0017] Effects of the Invention

[0018] According to the present invention, after reconstruction is performed by fixing the ratio of the measurement matrix size to the reconstruction matrix size (matrix ratio), ringing correction processing is performed with the matrix ratio, and then reconstruction is performed with the final reconstruction matrix size, thereby obtaining an image with the desired matrix size (resolution) with reduced Gibbs ringing.

[0019] Furthermore, for example, when ringing correction is performed using CNN, by using only a learning image group with a fixed matrix ratio as learning data, a learned CNN capable of performing ringing correction can be efficiently constructed without significantly increasing the amount of learning data.

[0020] In the ringing correction technology of the present invention, the only condition required when constructing the ringing correction part is a fixed matrix ratio. Therefore, it is independent of the imaging part or contrast and can be applied to all spatial resolutions and matrix sizes used in the inspection. It can also be applied to either 2D imaging or 3D imaging and can be applied regardless of the sampling mode, thereby obtaining a reliable Gibbs ringing correction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1It is a diagram showing the overall structure of an MRI apparatus to which the present invention is applied.

[0022] Figure 2 This is a functional block diagram of the image processing unit.

[0023] Figure 3 It is a diagram schematically showing the steps of image processing.

[0024] Figure 4 is a diagram showing the relationship between the measurement matrix size and the reconstruction matrix size.

[0025] Figure 5 is a diagram illustrating the occurrence of Gibbs ringing.

[0026] Figure 6 This is a diagram showing an example of the structure of the CNN used in Implementation Example 1.

[0027] Figure 7 This is a diagram illustrating a group of images used for CNN learning.

[0028] Figure 8 This is a diagram showing the flow of image processing in the first embodiment.

[0029] Fig. 9 This is a diagram for explaining the processing of the first embodiment.

[0030] Fig.10 This is a diagram for explaining a part of the processing of Implementation Example 1.

[0031] Fig.11 This is a diagram for explaining another part of the processing in Implementation Example 1.

[0032] Fig.12 This is a diagram showing a configuration example of an image processing unit according to the second embodiment.

[0033] Explanation of symbols

[0034] 1-MRI device, 10-camera unit, 20-computer, 23-image processing unit, 231-image generating unit, 232-matrix size setting unit, 233-ringing correction unit, 235-CNN, 237-matrix size changing unit, 239-filter unit. DETAILED DESCRIPTION

[0035] Hereinafter, embodiments of an MRI apparatus and an image processing method implemented by the apparatus according to the present invention will be described with reference to the accompanying drawings.

[0036] First, the present invention will be described in detail with respect to the structure of the MRI apparatus 1 to which it is applied. The present invention can be applied to most types of MRI apparatuses currently in use, such as Figure 1As shown, it is roughly divided into an imaging unit 10 for collecting nuclear magnetic resonance signals, a computer 20 for controlling the imaging unit or performing various calculations, and various devices attached to the imaging unit 10 and the computer 20, such as an input device, a display device, an external storage device, etc.

[0037] The imaging unit 10 has the same structure as a general MRI apparatus, and therefore detailed description thereof is omitted, but includes a static magnetic field magnet 11 for generating a static magnetic field space in which the subject 5 is placed, a gradient magnetic field coil 12 for applying a gradient magnetic field to the static magnetic field, an RF transmission coil 13, and an RF reception coil 14, and includes a gradient magnetic field power supply 15 and a transmitter 16 for driving these coils, a receiver 17 connected to the high-frequency reception coil, and a bed device 19 for transporting the subject 5 to the static magnetic field space. In addition, the imaging unit includes a sequencer 18, and the sequencer 18 controls the signal sampling operation by the gradient magnetic field coil 12, the RF transmission coil 13, and the RF reception coil 14 according to a pulse sequence.

[0038] The computer 20 includes: an imaging control unit 21 that controls the operation of the imaging unit 10 through the sequencer 18; an image processing unit 23 that performs image reconstruction and other processing on the measurement data (k-space data) collected by the imaging unit 10; and a display control unit 25 that controls the display of the reconstructed image or GUI, etc.

[0039] The computer 20 can be composed of a general-purpose computer having a CPU or a GPU and a memory, and the functions of the above-mentioned computer are pre-programmed and realized by uploading the program to the CPU, etc. In addition, some functions of the computer are sometimes realized by hardware such as ASIC or programmable IC, and in the present invention, these hardware are referred to as computers.

[0040] The MRI apparatus 1 of this embodiment is characterized in that the computer 20 has a ringing correction unit that performs ringing correction on an image reconstructed with a specific matrix ratio. This ringing correction unit is not limited, but is typically a specific CNN that has been learned for Gibbs ringing correction, and the computer 20 is in a form that can use this ringing correction unit (CNN), for example, a specific CNN is built in or can be connected to a specific CNN. In addition, as a ringing correction unit instead of a CNN, a linear prediction process that predicts a zero-padding area of ​​k-space based on the frequency characteristics of the ringing for extrapolation can be used.

[0041] The following describes an implementation of the image processing unit 23. As an example, the structure of the image processing unit 23 in which the CNN 235 is assembled in the computer 20 is shown in FIG. Figure 2As shown in the figure, the image processing unit 23 includes: an image generating unit 231 that performs Fourier transform and other operations on the measurement data to generate an image of a desired reconstruction matrix size; and a ringing correction unit 233 that corrects ringing occurring in the image.

[0042] The ringing correction unit 233 includes a CNN 235 that has been learned using learning data of a specified matrix ratio, and corrects the Gibbs ringing of the image by applying the CNN. The image generation unit 231 includes: a first reconstruction unit 231-1 that converts the measurement data into image data using the same matrix ratio as the learning data of the CNN and generates an intermediate reconstructed image; and a second reconstruction unit 231-2 that converts the intermediate reconstructed image into measurement space data (i.e., k-space data) and reconstructs it using the desired matrix size. In addition, in order to facilitate the separate description of the functions of the image generation unit 231, the image generation unit 231 is divided into two reconstruction units 231-1 and 231-2, but this indicates that there are two functions, and as hardware, it also includes a case where one processor executes the functions of the two reconstruction units 231-1 and 231-2.

[0043] The image processing unit 23 further includes: a matrix size setting unit 232 for setting the matrix size required for the processing of the first reconstruction unit 231-1, i.e., the intermediate reconstruction matrix size, and the matrix size required for the processing of the second reconstruction unit 231-2, i.e., the final reconstruction matrix size; and a matrix size changing unit 237 for performing zero padding (hereinafter referred to as zero padding or ZF) or shearing on the measurement data to change the matrix size of the measurement data. In addition, in a preferred embodiment, a filter unit 239 is provided, which applies a k-space filter to the k-space data before the final reconstruction.

[0044] The outline of the processing of the image processing unit 23 having the above configuration is shown in FIG. Figure 3 . In addition, in the following description or referenced drawings, "SM" is used to represent the measurement matrix (size), "RM" is used to represent the reconstruction matrix (size), "I" is used to represent the image, and after "SM", "RM", and "I", lowercase letters such as "a, b, c, d, gt, in" are used to represent symbols used to distinguish each matrix and image.

[0045] If the image processing unit 23 receives the measurement data (k-space data) from the camera unit 10, the matrix size setting unit 232 first sets the three matrix sizes (S1) required for the processing of the image generating unit 231. One of the three matrix sizes is the measurement matrix size SMa of the received k-space data, which can be the matrix size of the measurement data itself, or the matrix size of the k-space data after preprocessing and cutting the measurement data. Another is the second matrix size whose ratio to the first matrix size is the same as the matrix ratio of the data set used in the learning of CNN235, and is the intermediate reconstruction matrix size RMb. The third is the reconstruction matrix size RMd desired by the user, and is a size set by the user via an input device, etc., or a size set by default according to the inspection process and can be changed by the user. Specific examples of these settings will be described in detail in the following embodiments.

[0046] Next, the first reconstruction unit 231-1 reconstructs the measurement data of the first matrix size with the intermediate reconstruction matrix size RMb (S2). The intermediate reconstructed image Ib of the intermediate reconstruction matrix size can be obtained by reconstruction. Next, CNN235 is applied to the intermediate reconstructed image Ib to correct the ringing occurring in the intermediate reconstructed image Ib (S3). The CNN235 uses an image data set having the same ratio of the matrix size of the measurement data to the reconstruction matrix size and the ratio of SMa to RMb for learning to output an image reconstructed with the matrix ratio and corrected for ringing as a correct image. Therefore, by applying CNN235, an intermediate reconstructed image Ic with corrected ringing can be obtained.

[0047] The intermediate reconstructed image Ic is converted into k-space data, and changed to the final reconstruction matrix size RMd in k-space (S4). Finally, the k-space data is converted into real space data to obtain the final reconstructed image (S5). The ringing of the intermediate reconstructed image is corrected, and the high-frequency components that cause the ringing are removed from the converted k-space data. The change in the matrix size of the k-space data varies depending on whether the intermediate reconstruction matrix size is larger or smaller than the final reconstruction matrix size. In the case of being smaller than the final reconstruction matrix size, for example, zero padding is performed. In the case of being larger than the final reconstruction matrix size, the size is changed by k-space shearing, and an appropriate k-space filter is applied according to the matrix size. In this way, the ringing that occurs during the second reconstruction is corrected.

[0048] The final reconstructed image is produced using k-space data from which the components causing the ringing are removed by ringing correction, and thus the ringing is significantly suppressed. Figure 4 and Figure 5 , the occurrence of ringing is explained.

[0049] Figure 4The left side of is the case where the measurement matrix size is the same as the reconstruction matrix size (both 128×128 in this example), but Figure 4 As shown on the right side of , when the 128×128 measurement data is reconstructed after zero padding to 256×256, the matrix size of the reconstructed image becomes 256×256, and the apparent resolution is improved. Figure 5 As shown in FIG. 1 , the point spread function (PSF) obtained by Fourier transforming the shearing waveform of the k-space at the time of zero padding becomes a waveform represented by a sinc function with respect to the rectangular wave (the shearing waveform of the k-space data).

[0050] It can be seen from the waveform that when the measurement matrix size is set to SM and the reconstruction matrix size is set to RM, the point spread function PSF vibrates with a period of (RM / SM)*2. The frequency component of this vibration becomes a striped artifact (ringing) and appears in the reconstructed image. That is, the period of the striped artifact (ringing) is determined by the ratio of the measurement matrix size to the reconstruction matrix size, that is, the matrix ratio. In this embodiment, the characteristic that the period of the ringing depends on the matrix ratio is utilized, and a group of images with a fixed matrix ratio is used to learn the CNN. After the ringing correction is performed using the CNN, the conversion to the final reconstruction matrix is ​​performed. As a result, the CNN can be simply constructed using a relatively small learning data set, and the ringing correction effect can be reliably performed regardless of the combination of the measurement matrix and the reconstruction matrix.

[0051] According to this embodiment, by adopting a structure in which CNN learning and application is performed with only a fixed matrix ratio, it can be applied to all spatial resolutions and matrix sizes used in inspections, and a ringing correction effect can be uniformly obtained without depending on the imaging site or contrast.

[0052] Hereinafter, a specific CNN structure and an embodiment of processing by an image processing unit using the structure will be described. In the following embodiments, common drawings among the drawings referred to in the description of the structure of the above-mentioned device and computer are directly cited.

[0053] <Implementation Method 1>

[0054] In the present embodiment, a case will be described in which a final image having a desired matrix size larger than the matrix size is reconstructed using measurement data captured by an arbitrary measurement matrix.

[0055] First, the specific structure and learning steps of CNN235 for ringing correction are explained.

[0056] Generally, CNN has a structure in which multiple convolution layers and pooling layers are stacked between the input layer and the output layer, and the coefficients or weights of the layers are changed by repeated learning, thereby obtaining the desired output. In this embodiment, this general CNN can also be used, but since complex images can be obtained in MRI, a complex corresponding CNN is used. By setting the CNN corresponding to the complex image, processing can be performed without being affected by the intensity or phase of the signal.

[0057] An example of a complex corresponding CNN used in this embodiment is shown below. Figure 6 .

[0058] The CNN 235 is a CNN with a nine-layer structure, each layer of which is composed of blocks including a convolution layer, an activation function (Re LU) and a pooling layer, and has a structure in which the real part and the imaginary part of the input image are combined in the first layer, and the real part and the imaginary part are separated in the ninth layer. In addition, the kernel size or the number of channels of each layer is not particularly limited, but for example, the kernel size can be set to "3×3" to "9×9", etc., and the number of channels can be set to "16ch" to "128ch", etc.

[0059] In addition, in this embodiment, although the example of a simple CNN is shown, CNN is not limited to this. For example, a CNN with a changed resolution such as U-net may also be used.

[0060] In the learning step of CNN, a plurality of correct images (plural) and input images (plural) are prepared. Figure 7 As shown on the left end of , the correct image (first image) is an image 701 without ringing. The input image (second image) 705 is an image generated by performing Fourier transform (FT) on the correct image as k-space data (first measurement data) 702, cutting out its low-frequency region with a predetermined size, zero-padding (ZF) the high-frequency region of the cut data 703 to restore it to the original size, and then performing inverse Fourier transform (IFT) on the data (second measurement data) 704 after the matrix size is changed ( Figure 7 That is, the input image is an image reconstructed with a reconstruction matrix size different from the measurement matrix size of the first measurement data. In this way, a combination of a plurality of correct images 701 and an input image 702 generated therefrom is used as a learning image group.

[0061] Here, the matrix size of the correct image is not limited, and images of various matrix sizes can be used, but in any group, the ratio of the matrix size of the original measurement data (first measurement data) 702 to the matrix size of the clipped data (before size change) 703 is set to a fixed value. This ratio (fixed value) corresponds to the following ratio, that is, when the clipped data is regarded as the measurement data (first measurement data) of the second image (input image), the matrix ratio (reconstruction matrix size / measurement matrix size) when reconstructed with a reconstruction matrix size that becomes a predetermined matrix ratio with respect to its measurement matrix size.

[0062] The CNN of this embodiment is produced using learning data with a constant matrix ratio. For example, when the matrix ratio = 2, there are the following combinations.

[0063] Measurement matrix size: 256, reconstruction matrix size: 512

[0064] Measurement matrix size: 384, reconstruction matrix size: 768

[0065] Measurement matrix size: 300, reconstruction matrix size: 600

[0066] In addition, only the dimensions in a single axis direction are shown in this example, but the matrix is ​​not only a square, but sometimes a rectangle, or a circle or ellipse inscribed therein, or even 3D, and is not limited thereto. There can be multiple matrix ratios depending on the shape or dimension of the matrix, but they are all set to fixed values.

[0067] The CNN is trained using the training image group with the matrix ratio fixed in this way so that the output image when the CNN is applied to the input image is close to the correct image.

[0068] By using the CNN learned in this way, for measurement data of various sizes, the ringing effect correction of the image reconstructed with the reconstruction matrix size having a fixed matrix ratio with the measurement matrix size is realized. That is, after the measurement data of various different matrix sizes are resized with the same matrix ratio as the learning image group, they are applied as the input image of the learned CNN235, thereby outputting a ringing-corrected image with a different resolution from the original image.

[0069] Next, refer to Figure 8 and Figures 9 to 11 , the processing of the image processing unit 23 of this embodiment including the ringing correction using the CNN 235 is described. Figure 8 In, with Figure 3 The same symbols are used to represent the same processing contents, and repeated descriptions are omitted.

[0070] <s1>

[0071] If the image processing unit 23 receives the measurement data (measured k-space data) from the camera unit 10, the matrix size setting unit 232 first sets the three matrix sizes required for the processing of the image generating unit 231, namely, the matrix size of the measurement data (measurement matrix size) SMa, the intermediate reconstruction matrix size RMBb, and the final reconstruction matrix size RMd (S1). The ratio (matrix ratio) RMb / SMa (=N) between the measurement matrix size SMa and the intermediate reconstruction matrix size RMBb is the same as the matrix ratio of the data (image group) used by CNN235 in learning. In addition, in the present embodiment, the final reconstruction matrix size RMd is larger than the measurement matrix size SMa (RMd>SMa), and the intermediate reconstruction matrix size RMBb is larger than the measurement matrix size SMa.

[0072] <s2>

[0073] The matrix size changer 237 changes the measurement data to an intermediate reconstruction matrix size RMb. Here, RMb>RMa, such as Fig. 9 As shown, the size change is extended by zero padding (ZF) the high frequency region for the measurement data 901 of an arbitrary measurement matrix. In addition, when the image data 900 is already obtained, it can be Fourier transformed (FT) to obtain the measurement data 901 .

[0074] <s3>

[0075] The image generation unit 231 (first reconstruction unit 231-1) performs inverse Fourier transform (IFT) on the measurement data 902 after the matrix size change as real space data 903, and performs ringing correction based on CNN235. That is, the real space data 903 is used as the input image of CNN235, and CNN235 is applied. Through this processing, an image after ringing correction can be obtained as an output image 904 of CNN235.

[0076] <S40~S44>

[0077] The peripheral area in the k-space of the output image 904 of the CNN 235 is extrapolated to become an image with a higher resolution than the original image data 900, but it is not an image with the resolution desired by the user. Therefore, the matrix size changer 237 reconstructs the image after ringing correction after matching the final reconstruction matrix size RMd on the k-space data. That is, the output image 904 is Fourier transformed (FT) to obtain the k-space data 905, and the matrix size is changed to perform the final reconstruction (second reconstruction).

[0078] Here, a process for preventing new ringing from occurring is performed by changing the matrix size. This process differs depending on whether the intermediate reconstruction matrix size RMb is larger or smaller than the final reconstruction matrix size RMd.

[0079] <s42>

[0080] In the case of RMb>RMd, if Fig.10 As shown, after the k-space data 906 of the final reconstruction matrix size is clipped from the k-space data 905 of the intermediate reconstruction matrix size, a k-space filter (frequency filter) 907 of the same size as the final reconstruction matrix size RMd is set and applied.

[0081] As an example, if it is as follows,

[0082] - Measurement matrix size SMa: 192

[0083] Intermediate reconstruction matrix size RMB: 384

[0084] Final reconstruction matrix size RMd: 256,

[0085] By cutting a 256-bit matrix from a 384-bit matrix, even if the data (signal) at the edge of the matrix is ​​extrapolated by applying the CNN 235, the signal is disconnected at the end of the reconstructed matrix to form a step difference, thereby causing ringing. This step difference can be smoothed by using a k-space filter of the same size as the cut k-space data 906, that is, the same size as the final reconstructed matrix size RMd, thereby preventing new ringing due to the change in matrix size.

[0086] As the k-space filter 907 , a commonly available frequency filter, for example, a Gaussian filter, a Fermi filter, a Hanning filter, a truncation filter such as a well-known filter having only a Hanning filter at the end, etc. can be used.

[0087] <s43>

[0088] Then, the image generator 231 (second reconstruction unit 231 - 2 ) performs inverse Fourier transform (IFT) on the k-space data 908 after the filter is applied, and generates a final reconstructed image 909 .

[0089] <s44>

[0090] In the case of RMb≤RMd, if Fig.11 As shown, firstly, the k-space data 905 of the intermediate reconstruction matrix size is zero-filled (ZF) and the size is expanded to the final reconstruction matrix size, and then the k-space filter 917 is applied to the expanded k-space data 916 .

[0091] As an example, if it is as follows,

[0092] - Measurement matrix size SMa: 192

[0093] Intermediate reconstruction matrix size RMB: 384

[0094] Final reconstruction matrix size RMd: 512

[0095] The periphery of the 384-bit matrix is ​​padded with zeros and expanded to a 512-bit matrix. This expansion causes a step difference between the end of the intermediate reconstruction matrix and the periphery padded with zeros, which can also cause ringing. In order to eliminate this step difference, smoothing is performed using a k-space filter of the same size as the matrix size of the k-space data 905 before expansion, that is, the intermediate reconstruction matrix size RMb. This prevents new ringing from occurring due to a change in the matrix size.

[0096] <s43>

[0097] Then, the image generator 231 (second reconstruction unit 231 - 2 ) performs inverse Fourier transform (IFT) on the filtered k-space data 918 to generate a final reconstructed image 919 , which is the same as the case of RMb>RMd.

[0098] As described above, according to this embodiment, first, the measurement data obtained by the camera unit with a predetermined matrix size is resized with a constant matrix ratio and then reconstructed (intermediate reconstruction), and the intermediate reconstructed image is subjected to ringing correction using a CNN that has been trained using an image data set with the constant matrix ratio. Then, resolution adjustment including resizing and filtering is performed in real space to obtain a final image with a final reconstruction matrix size.

[0099] This makes it possible to significantly suppress the occurrence of blurring compared to the case where the k-space filter is directly applied to the measurement data.

[0100] Furthermore, there is no need to use various combinations of different measurement matrix sizes and reconstruction matrix sizes as learning data as in the previous ringing correction using CNN. Therefore, CNN can be constructed simply and a high ringing suppression effect can be obtained.

[0101] <Implementation Method 2>

[0102] In the first embodiment, a case where a CNN 235 is learned using a fixed matrix ratio between a reconstruction matrix and a measurement matrix is ​​described. Fig.12 As shown, two or more CNNs 235 may be prepared for different matrix ratios, and one of them may be selected to perform ringing correction. Therefore, the image generation unit 231 of this embodiment includes a CNN selection unit 236.

[0103] As a reference for selecting a CNN by the CNN selection unit 236, for example, the CNN 235 whose matrix ratio is closest to the ratio of the final reconstruction matrix to the measurement matrix is ​​selected. In this way, the difference between the size of the intermediate reconstruction matrix and the size of the final reconstruction matrix can be minimized, such as Figure 5 As shown, the vibration frequency of the point distribution function that causes ringing can be set on the high-frequency side to a level where the ringing is not noticeable, and the occurrence of ringing during the final reconstruction can be suppressed.

[0104] In this embodiment, it is necessary to construct multiple CNNs 235, but for the learning image data set, various images can be used without restrictions on their matrix size, contrast, camera position, etc., and the same original image can be used to construct them. Therefore, compared with the previous ringing correction CNN, the learning steps can be simplified.

[0105] The above describes the embodiments of the MRI apparatus and the image processing method implemented by the MRI apparatus of the present invention. However, part of the functions of the image processing unit of the MRI apparatus, in particular, each process of the image processing method of the present invention can be installed in a computer different from the MRI apparatus. Such an image processing apparatus is also included in the present invention.

Claims

1. A magnetic resonance imaging device, characterized in that: have: An imaging unit for collecting measurement data including magnetic resonance signals; an image generating unit, reconstructing the measurement data with a desired reconstruction matrix size; and A ringing correction unit, performing ringing correction on the reconstructed image; The image generation unit comprises: a matrix size setting unit configured to set a measurement matrix size of the measurement data, an intermediate reconstruction matrix size having a predetermined matrix ratio relative to the measurement matrix size, and the desired reconstruction matrix size; a first reconstruction unit, reconstructing the measurement data with the intermediate reconstruction matrix size and generating an intermediate reconstructed image; and The second reconstruction unit reconstructs k-space data of the corrected intermediate reconstructed image obtained by performing ringing correction on the intermediate reconstructed image by the ringing correction unit, using the desired reconstruction matrix size.

2. The magnetic resonance imaging apparatus according to claim 1, wherein: The ringing correction unit includes a CNN that has been learned using a plurality of image groups including a first image in which ringing does not occur and a second image in which ringing occurs, wherein the second image is an image reconstructed by resizing the first image in k-space at the prescribed matrix ratio.

3. The magnetic resonance imaging apparatus according to claim 2, characterized in that: The second image used in the learning of the CNN is an image obtained by inverse Fourier transforming the 2k spatial data. In the 2k spatial data, the high-frequency area is zero-filled with data of a low-frequency area cut out from the 1k spatial data obtained by inverse Fourier transforming the first image, and the matrix size is set to be the same as that of the 1k spatial data.

4. The magnetic resonance imaging apparatus according to claim 2, characterized in that: The CNN is a complex corresponding CNN.

5. The magnetic resonance imaging apparatus according to claim 1, characterized in that: The desired reconstruction matrix size set in the matrix size setting unit is larger than the measurement matrix size, and the predetermined matrix ratio, ie, intermediate reconstruction matrix size / measurement matrix size, is larger than 1.

6. The magnetic resonance imaging apparatus according to claim 2, characterized in that: The image generation unit further includes a matrix size changer configured to change the size of the intermediate reconstructed image to which the CNN is applied in a measurement space.

7. The magnetic resonance imaging apparatus according to claim 6, characterized in that: The matrix size changer converts the intermediate reconstructed image to which the CNN is applied into k-space data and then performs size change by performing at least one of a cropping process and a zero padding process.

8. The magnetic resonance imaging apparatus according to claim 6, characterized in that: The image generating unit further includes a filtering unit configured to apply a frequency filter after the matrix size change unit has performed the size change.

9. The magnetic resonance imaging apparatus according to claim 8, characterized in that: The intermediate reconstruction matrix size set in the matrix size setting unit is larger than the desired reconstruction matrix size, The size of the frequency filter applied by the filter unit is the same as the size of the k-space data after the size change.

10. The magnetic resonance imaging apparatus according to claim 8, characterized in that: The intermediate reconstruction matrix size set in the matrix size setting unit is the same as or smaller than the desired reconstruction matrix size, The size of the frequency filter applied by the filter unit is the same as the size of the k-space data before the size change.

11. The magnetic resonance imaging apparatus according to claim 2, characterized in that: The ringing correction unit includes a plurality of CNNs having different predetermined matrix ratios used in learning.

12. The magnetic resonance imaging apparatus according to claim 11, characterized in that: The matrix size setting unit selects a CNN among the plurality of CNNs whose predetermined matrix ratio is closest to the ratio of the desired reconstruction matrix size to the measurement matrix size, and determines the intermediate reconstruction matrix size using the matrix ratio.

13. An image processing method, which uses measurement data including magnetic resonance signals collected by a magnetic resonance imaging device to reconstruct a ringing-corrected image, the image processing method being characterized by comprising: A first reconstruction step of reconstructing the measurement data with a reconstruction matrix size having a predetermined matrix ratio relative to a measurement matrix size of the measurement data to generate an intermediate reconstructed image; A correction step of correcting the ringing of the intermediate reconstructed image using a CNN that outputs an image in which the ringing of the input image is corrected, wherein the CNN learns a first image in which the ringing does not occur and a second image in which the ringing occurs as a group; and In the second reconstruction step, the intermediate reconstructed image after ringing correction is converted into k-space data and reconstructed with the desired reconstruction matrix size. The second image used in the CNN learning is an image obtained by resizing the first image in k-space and reconstructing it with a reconstruction matrix size of the first image, and the reconstruction matrix size and the matrix size after resizing satisfy a predetermined matrix ratio.

14. The image processing method according to claim 13, characterized in that: The second reconstruction step includes the steps of converting the intermediate reconstruction image into k-space data, performing size change in k-space, and applying a frequency filter.

Images