KR20210145003A - Apparatus for restoring short scanning brain tomographic image using deep-learning - Google Patents

Abstract

The present invention relates to an apparatus for restoring a brain tomographic image captured in a short time using deep learning, which reduces a patient's inconvenience. More specifically, the apparatus comprises: a loss image collection unit collecting a positron emission tomography (PET) loss image including noise components acquired by performing brain PET for a set first time; an original image collection unit collecting an original PET image acquired by performing the brain PET corresponding to the PET loss image for a second time longer than the first time; a generation unit generating a plurality of similar PET images by removing the noise components of the PET loss image; a discrimination unit comparing the plurality of similar PET images generated by the generation unit with the original PET image to determine whether the similar PET image belongs to an actual image classification corresponding to the original PET image or a false image classification not corresponding to the original PET image; and a restored image output unit generating and outputting a restored image on the basis of the similar PET image determined to belong to the actual image classification in the discrimination unit.

Description

Brain tomography image restoration device taken in a short time using deep learning

The present invention relates to an apparatus for restoring a brain tomography image taken for a short time using deep learning, and more particularly, a PET image taken at a standard time from a PET image taken for a short time and a PET image taken for a short time based on deep learning. We want to remove the noise component that occurs in the image.

Amyloid positron emission tomography (PET) is a nuclear medicine imaging test that shows amyloid deposits in the brain, and is currently used to diagnose Alzheimer's disease, which is known to be caused by amyloid accumulation.

Although there are some differences in the acquisition protocol according to commercial radiopharmaceuticals for amyloid PET, in most cases for F-18 fluorbetaben (FBB), imaging should be taken for 20 minutes.

If the head moves excessively when acquiring a scan for a long time, severe motion artifacts occur in the PET image, which deteriorates diagnostic performance. Some elderly patients with dementia often require a rescan (or additional radiation exposure) due to severe head movements.

Therefore, as the use of PET in dementia patients increases, the demand for shortening the scan time is increasing.

However, PET images obtained at short scan times suffer from low signal-to-noise ratios and reduced diagnostic reliability. Recently, deep learning techniques for image restoration have been widely applied to medical images such as computed tomography (CT), magnetic resonance imaging (MRI), and PET, and some of them used deep learning techniques to restore low-dose PET images. It showed the potential to reduce noise artifacts. There are only a few studies that reduce the acquisition time of brain PET to reduce noise components and improve image quality.

Additional MR information obtained from PET/MR scanners was used to reconstruct brain PET images, but PET/MR scanners are very expensive and have not yet been widely installed. Since PET/CT scanners are used in most hospitals, there is a need for a recovery technique that uses only PET without MRI information.

US 10-2034248 B1 US 10-1975186 B1

An object of the present invention is to solve the above problem, and to provide a brain tomography image restoration device taken for a short time using deep learning for restoring a short time PET image using deep learning for a PET image taken for a short time. aim to

Objects of the present invention are not limited to the objects mentioned above, and other objects not mentioned will be clearly understood from the description below.

The apparatus for restoring a brain tomography image taken for a short time using deep learning according to an aspect of the present invention for achieving the above object is a PET loss including a noise component obtained by tomography the brain for a first preset time. A loss image collection unit for collecting images, an original image collection unit for collecting an original PET image obtained by tomography of the brain for a second time period longer than the first time, and collecting a tomographic PET image corresponding to the PET loss image; A generator unit that generates a plurality of similar PET images by removing the noise component of the lost image, and a plurality of similar PET images generated by the generator unit are compared with the original PET image so that the similar PET image corresponds to the original PET image. A discriminator unit that determines whether it belongs to the classification or a false image classification that does not correspond to the original PET image, and the discriminator unit generates and outputs a restored image based on the similar PET image determined to belong to the real image classification and a restored image output unit.

According to the present invention, there is an effect that a PET image taken for a short time can be restored into a PET image corresponding to a PET image taken at a standard time using deep learning.

According to the present invention, there is an effect that a reconstructed image obtained by reconstructing a PET image taken for a short time can be used for actual clinical reading purposes.

According to the present invention, in the encoder and decoder of a PET image taken for a short time, by performing decomposition and recomposition, high-frequency components and low-frequency components are extracted and the characteristics of the image are analyzed, thereby using pooling or unpooling. It has the effect of analyzing and extracting image features in more detail.

According to the present invention, when decomposing a PET image, a high-pass filter is used to generate different image feature maps according to the diagonal, vertical, and horizontal directions of the image, thereby providing more detailed image features than using pooling. can be extracted.

According to the present invention, it is possible to obtain a PET image from which a noise component is removed from a PET image taken for a short time without the need for a separate image noise correction device.

According to the present invention, since the brain is PET taken for a shorter time than the standard time, there is an effect that can reduce the discomfort of the patient.

The effect according to the present invention is not limited by the contents exemplified above, and more various effects are included in the present specification.

1 is a block diagram for explaining an apparatus for restoring a brain tomography image taken for a short time using deep learning according to an embodiment of the present invention.
2 is a view for explaining a PET loss image of the brain tomography image restoration apparatus taken for a short time using deep learning according to an embodiment of the present invention.
3 is a view for explaining the original PET image of the brain tomography image restoration apparatus taken for a short time using deep learning according to an embodiment of the present invention.
4 is a view for explaining a generator unit of a brain tomography image restoration apparatus taken for a short time using deep learning according to an embodiment of the present invention.
5 is a block diagram illustrating a generator unit of an apparatus for restoring a brain tomography image taken for a short time using deep learning according to an embodiment of the present invention.
6 is a view for explaining the discriminator unit of the brain tomography image restoration apparatus taken for a short time using deep learning according to an embodiment of the present invention.
7 is a block diagram for explaining a discriminator unit of a brain tomography image restoration apparatus taken for a short time using deep learning according to an embodiment of the present invention.
8 is a view for explaining a restored image of the brain tomography image restoration apparatus taken for a short time using deep learning according to an embodiment of the present invention.
9 is a PET image generated using the conventional U-net method.
10 is a graph showing changes in average SUVR according to changes in VOI of a restored image of a brain tomography image restoration apparatus taken for a short time using deep learning and a PET original image according to an embodiment of the present invention.
11 is a graph illustrating a Pearson correlation between a restored image of a brain tomography image restoration apparatus taken for a short time using deep learning and an original PET image according to an embodiment of the present invention.
12 is a graph illustrating a Bland-Altman analysis of a restored image of a brain tomography image restoration apparatus taken for a short time using deep learning and an original PET image according to an embodiment of the present invention.

Advantages and features of the present invention, and a method of achieving them, will become apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but will be implemented in various different forms, and only these embodiments allow the disclosure of the present invention to be complete, and common knowledge in the technical field to which the present invention belongs It is provided to fully inform the possessor of the scope of the invention, and the present invention is only defined by the description of the claims. On the other hand, the terms used in the present specification are for describing the embodiments, and are not intended to limit the present invention. In this specification, the singular form also includes the plural form unless otherwise specified in the phrase. Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

In addition, in the embodiments of the present invention, functionally identical components will be described with the same reference numerals, and repeated descriptions will be omitted.

Referring to FIG. 1 , the apparatus 1 for restoring a brain tomography image taken for a short time using deep learning according to an embodiment of the present invention includes a loss image collecting unit 10 , an original image collecting unit 20 , and a cropping unit. 30 , the generator unit 40 , the pixel-unit loss calculation unit 50 , the discriminator unit 60 , the error value calculation unit 70 , the false image set generation unit 80 , the restored image output unit 90 . ) is included.

Referring to FIG. 2 , the loss image collecting unit 10 collects PET loss images including noise components obtained by performing F-18 florbetaben positron emission tomography of the brain for a first preset time.

Here, the PET loss image may be a tomography image of the brain for 1 to 3 minutes, and the loss image collecting unit 10 may collect PET loss images of 200 to 300 patients with a size of 400 X 400 pixels. .

On the other hand, the loss image collecting unit 10 includes a noise component obtained by performing F-18 Florbetaben Positron Emission Tomography (PET) of the brain to which the amyloid agent has been administered by the Amyloid PET Radiotracer. It may be to collect PET loss images.

On the other hand, in the PET loss image, the amyloid-deposited portion may have a color close to white.

Here, the color close to white may mean the color of pixel portions displayed in gray or white in contrast to the black portion in the PET loss image.

In the PET loss image, a border may be formed with a portion corresponding to the brain region as a boundary, and pixels of a cell or tissue portion belonging to the brain region may have a color close to white depending on the degree of amyloid deposition.

Referring to FIG. 3 , the original image collection unit 20 is obtained by performing F-18 fluorbetaben positron emission tomography of the brain for a second time period longer than the first time period, and the PET tomography corresponding to the PET loss image. Collect the original image.

The PET original image may be an image obtained by F-18 florbetaben positron emission tomography of the brain for 10 to 60 minutes, and the original image collecting unit 150 is a 400 X 400 pixel size of 200 to 300 patients. It may be collecting PET original images.

The cropping unit 30 may detect an edge in the PET loss image, extract a region inside the edge as a brain region, and remove the background region.

Since the brain region has a large difference in brightness value compared to the background region, the brain region and the background region can be easily distinguished according to edge detection.

The cropping unit 30 generates a closed curve by connecting pixels corresponding to edge components extracted as edge detection is performed on the PET loss image, and extracts the inside of the closed curve, which has a smaller radius of curvature, into the brain region. It may be to remove the background area.

The cropping unit 30 extracts the edge region by comparing the pixel values between adjacent pixels as the part corresponding to the changing boundary by comparing the pixel values between adjacent pixels since the edge occurs in a part where the brightness changes in the image. It may be to distinguish between a brain region and a background region.

The cropping unit 30 may convert the PET loss image from a size of 400 X 400 pixels to a size of 224 X 224 pixels.

The generator unit 40 removes a noise component from the lost PET image to generate a plurality of similar PET images.

4 and 5 , the generator unit 40 includes an encoder unit 100 , a decoder unit 200 , a linearization unit 300 , and an output unit 400 .

The encoder unit 100 filters the PET loss image according to the convolution operation, extracts the characteristics of the PET loss image, and uses a decomposition function that reduces the pixel size of the PET loss image relatively to the pixel size of the lost PET image. A feature map of the PET loss image with a small pixel size is generated.

The encoder unit 100 includes a first encoder unit 110 , a second encoder unit 130 , and a third encoder unit 150 .

The first encoder unit 110 receives the PET loss image, filters the PET loss image by convolution operation, extracts the characteristics of the PET loss image, and uses a wavelet-decomposition function to pixels of the lost PET image. A first feature map having a pixel size that is relatively smaller than the size is generated.

The first encoder unit 110 includes a first convolutional layer unit 111 , a first batch normalization unit 113 , a first activation unit 115 , and a first decomposition unit 117 .

The first convolutional layer unit 111 receives the PET loss image and extracts features of the PET loss image using the first convolutional layer having a preset filter kernel size.

The first convolutional layer unit 111 may extract the features of the PET loss image by performing a convolution operation while moving the convolution filter for extracting different features from the PET loss image by the stride according to the user input. .

The pixel size of the first convolutional layer may be 3×3, and the stride may be 1 pixel.

The first batch normalization unit 113 batch normalizes the distribution of pixel values belonging to the PET loss image so that the average of the pixel values belonging to the PET loss image is 0 and the variance is 1.

Batch Normalization is a method for reducing the gradient vanishing or gradient exploding that may occur in the process of learning a generative adversarial network due to a change in the distribution of pixel values belonging to an image. It means normalizing the mean to 0 and the variance to 1.

The first activator 115 generates an activated PET loss image by using an activation (Leaky ReLu) function having a preset gradient (0.2) in the batch normalized PET loss image.

Activation function is a function that converts an input signal coming into an individual neuron into an output signal in a neuron network. The ReLu function is one of the activation functions in deep-learning. It may be a function that converts only negative numbers (-) to 0 among the values to be used.

The first activator 115 generates an activated PET loss image by converting a negative number to 1/10 among values output by individual neurons of the neuron network using a Leaky ReLu function having a slope of 0.2. can

The PET loss image may have a pixel size of 224 x 224 and the number of filters changed to 64 by the first convolutional layer unit 111, the first batch normalization unit 113, and the first activation unit 115, The first activator 215 may output an activated PET loss image having a pixel size of 224 x 224 and the number of filters of 64.

The first decomposition unit 117 generates a first feature map having a pixel size relatively smaller than a pixel size of the lost PET image by using a decomposition function on the activated PET loss image.

The first decomposition unit 117 includes a first high frequency extraction unit 117a and a first low frequency extraction unit 117b.

The first high frequency extraction unit 117a decomposes the activated PET loss image by calculating one pixel value according to a difference value between at least two pixel values among the pixel values of the activated PET loss image.

The first high-frequency extraction unit 117a uses a high-pass filter among pixel values of the activated PET loss image to obtain a first high-frequency one-dimensional feature map having a difference value between at least two pixel values in a diagonal direction, in a vertical direction. may be decomposed into a second high-frequency one-dimensional feature map based on a difference value between at least two pixel values and a third high-frequency one-dimensional feature map based on a difference value between at least two pixel values along a horizontal direction.

The first low frequency extraction unit 117b calculates one pixel value based on an average between at least two pixel values among the pixel values of the activated PET loss image.

The first low-frequency extractor 117b may calculate the first low-frequency one-dimensional feature map by calculating an average value between at least two pixel values among the pixel values of the first feature map by using the low-pass filter.

The first decomposition unit 117 includes a first high-frequency one-dimensional feature map to a third high-frequency one-dimensional feature map, and a first low-frequency one-dimensional feature map based on the first high frequency extraction unit 117a and the first low frequency extraction unit 117b. By outputting the feature map, the first high-frequency one-dimensional feature map to the third high-frequency one-dimensional feature map are passed to the output of the third activation unit 255 of the third decoder unit 250 (skip), and the pixel of the PET loss image It may be to generate a first feature map that is a first low-frequency one-dimensional feature map having a pixel size that is relatively smaller than the size.

When the input image passes the low-pass filter, a low-frequency image with the left-right width or the top and bottom width reduced by 1/2 is generated, and when the input image passes the high-pass filter, a high-frequency image with the horizontal width or the top and bottom width reduced by 1/2 is generated. it could be

In other words, the first decomposition unit 117 includes a first high-frequency one-dimensional feature map through a high-pass filter in a diagonal direction of the first high-frequency extraction unit 117a, and a second high-frequency 1 through a high-pass filter in a vertical direction. It is decomposed into four feature maps consisting of a dimensional feature map, a third high-frequency one-dimensional feature map through a high-pass filter along the horizontal direction, and a first low-frequency one-dimensional feature map through a low-pass filter of the first low-frequency extraction unit 117b. may be doing

In addition, the first decomposition unit 117 outputs first to third high frequency one-dimensional feature maps having a pixel size of 112 x 112 and a number of filters of 128 to output the third decoder unit 250 . It may be skipping to the output of the third activation unit 255 of , and generating a first feature map that is a first low-frequency one-dimensional feature map having a pixel size of 112 x 112 and the number of filters of 128 may be doing

The second encoder unit 130 filters the first feature map according to the convolution operation to extract the features of the first feature map, and uses a decomposition function to obtain a pixel size that is relatively smaller than the pixel size of the first feature map. The branch generates a second feature map.

The second encoder unit 130 includes a second convolutional layer unit 131 , a second batch normalization unit 133 , a second activation unit 135 , and a second decomposition unit 137 .

The second convolutional layer unit 131 receives the first feature map and extracts features of the first feature map by using a second convolutional layer having a preset filter kernel size (3 X 3).

The pixel size of the second convolutional layer may be 3×3, and the stride may be 1 pixel.

The second arrangement normalization unit 133 arrangement normalizes the distribution of pixel values belonging to the first feature map such that the average of the pixel values belonging to the first feature map is 0 and the variance is 1.

The second activator 135 generates the activated first feature map using an activation (Leaky ReLu) function having a preset gradient (0.2) on the batch normalized first feature map.

The second decomposition unit 137 generates a second feature map having a pixel size that is relatively smaller than a pixel size of the first feature map by using a decomposition function on the activated first feature map.

The second decomposition unit 137 includes a second high frequency extraction unit 137a and a second low frequency extraction unit 137b.

The second high frequency extractor 137a decomposes the second feature map by calculating one pixel value based on a difference between at least two pixel values among the pixel values of the activated first feature map.

The second high frequency extractor 137a uses a high pass filter among pixel values of the activated first feature map to obtain a first high frequency 2D feature map based on a difference value between at least two pixel values in a diagonal direction, in a vertical direction. may be decomposed into a second high-frequency two-dimensional feature map based on a difference value between at least two pixel values and a third high-frequency two-dimensional feature map based on a difference value between at least two pixel values along a horizontal direction.

The second low frequency extractor 137b calculates one pixel value based on an average between at least two pixel values among the pixel values of the activated first feature map.

The second low-frequency extractor 137b may output the first low-frequency two-dimensional feature map by calculating an average value between at least two pixel values among the pixel values of the activated first feature map using a low-pass filter. .

The second decomposition unit 137 includes a first high-frequency two-dimensional feature map, a third high-frequency two-dimensional feature map, and a first low-frequency two-dimensional feature map based on the second high frequency extraction unit 137a and the second low frequency extraction unit 137b. By outputting the feature map, the first high-frequency two-dimensional feature map to the third high-frequency two-dimensional feature map are passed to the output of the second activation unit 235 of the second decoder unit 230 (skip), and The second feature map, which is the first low-frequency two-dimensional feature map, may be generated with a pixel size that is relatively smaller than the pixel size.

In addition, the second decomposition unit 137 outputs the first high frequency 2D feature map to the third high frequency 2D feature map having a pixel size of 56 x 56 and the number of filters of 256 to output the second decoder unit 230 . It may be skipping to the output of the second activation unit 235 of , and generating a second feature map that is a first low-frequency two-dimensional feature map having a pixel size of 56 x 56 and the number of filters of 256 may be doing

The third encoder unit 150 filters the second feature map according to the convolution operation to extract the features of the second feature map, and uses a decomposition function to obtain a pixel size that is relatively smaller than the pixel size of the second feature map. The branch generates a feature map of the PET loss image.

The third encoder unit 150 includes a third convolutional layer unit 151 , a third batch normalization unit 153 , a third activation unit 155 , and a third decomposition unit 157 .

The third convolutional layer unit 151 receives the second feature map and extracts the features of the second feature map using a third convolutional layer having a preset filter kernel size (3 X 3).

The pixel size of the third convolutional layer may be 3×3, and the stride may be 1 pixel.

The third arrangement normalization unit 153 arrangement normalizes the distribution of pixel values belonging to the second feature map such that the average of the pixel values belonging to the second feature map is 0 and the variance is 1.

The third activator 155 generates the activated second feature map using an activation (Leaky ReLu) function having a preset gradient (0.2) on the batch normalized second feature map.

The third decomposition unit 157 generates a feature map of a lost PET image having a pixel size relatively smaller than a pixel size of the second feature map by using a decomposition function on the activated second feature map.

The third decomposition unit 157 includes a third high frequency extraction unit 157a and a third low frequency extraction unit 157b.

The third high frequency extraction unit 157a decomposes the activated second feature map by calculating one pixel value based on a difference between at least two pixel values among the pixel values of the activated second feature map.

The third high frequency extraction unit 157a converts the pixel values of the activated second feature map to the first high frequency 3D feature map based on the difference value between at least two pixel values along the diagonal direction using a high pass filter, in the vertical direction. Accordingly, it may be decomposed into a second high frequency 3D feature map based on a difference value between at least two pixel values and a third high frequency 3D feature map based on a difference value between at least two pixel values along a horizontal direction.

The third low-frequency extractor 157b calculates one pixel value based on an average between at least two pixel values among the pixel values of the activated second feature map.

The third low-frequency extractor 157b may calculate the first low-frequency 3D feature map by calculating an average value between at least two pixel values among the pixel values of the activated second feature map using a low-pass filter. .

The third decomposition unit 157 includes a first high frequency 3D feature map to a third high frequency 3D feature map and a first low frequency 3D feature map based on the third high frequency extraction unit 157a and the third low frequency extraction unit 157b. By outputting the feature map, the first high-frequency three-dimensional feature map to the third high-frequency three-dimensional feature map are passed to the output of the first activation unit 215 of the first decoder unit 210 (skip), and rather than the second feature map It may be to generate a feature map of a PET loss image that is a first low-frequency three-dimensional feature map having a relatively small pixel size.

The third decomposition unit 157 outputs the first high frequency 3D feature map to the third high frequency 3D feature map having a pixel size of 28 x 28 and the number of filters of 256, and outputs the first to third high frequency 3D feature maps of the first decoder unit 210 . 1 It may be skipping to the output of the activation unit 215 and making a splicing connection, and generating a feature map of a PET loss image that is a first low-frequency three-dimensional feature map having a pixel size of 28 x 28 and the number of filters of 256 may be doing

The decoder unit 200 receives the feature map of the lost PET image, filters the feature map of the lost PET image by convolution operation, extracts the feature of the feature map of the lost PET image, and the pixel size of the feature map of the lost PET image. A similar PET image with the same pixel size as that of the PET loss image is generated using a recomposition function that enlarges the PET image.

The decoder unit 200 includes a first decoder unit 210 , a second decoder unit 230 , and a third decoder unit 250 .

The first decoder unit 210 filters the feature map of the lost PET image according to the convolution operation to extract the features of the feature map of the lost PET image, and uses a recomposition function to be larger than the pixel size of the feature map of the lost PET image. A first similar image having a relatively large pixel size is generated.

The first decoder unit 210 includes a first convolutional layer unit 211 , a first placement normalization unit 213 , a first activation unit 215 , and a first recomposition unit 217 .

The first convolutional layer unit 211 receives the feature map of the lost PET image and extracts features of the feature map of the lost PET image using the first convolutional layer having a preset filter kernel size (3 X 3). .

The pixel size of the first convolutional layer may be 3 x 3, and the stride may have 1 pixel.

The first batch normalization unit 213 batch-normalizes the distribution of pixel values belonging to the feature map of the lost PET image such that the average of the pixel values belonging to the feature map of the lost PET image is 0 and the variance is 1.

The first activator 215 generates a feature map of the activated PET loss image by using an activation (Leaky ReLu) function having a preset gradient (0.2) on the feature map of the batch normalized PET loss image.

The first recomposition unit 217 generates a first similar image having a pixel size relatively larger than a pixel size of a feature map of the lost PET image by using a recomposition function on the feature map of the activated PET loss image.

The first recomposition unit 217 receives the feature map of the PET loss image and the first high frequency 3D feature map to the third high frequency 3D feature map from the third decomposition unit 157, and receives a feature map of the PET loss image. , the first high-frequency 3D feature map to the third high-frequency 3D feature map may be used to generate a first similar image using a recomposition function.

The first recomposition unit 217 may generate a first similar image having a pixel size of 56 x 56 and the number of filters of 512.

The second decoder unit 230 filters the first similar image according to the convolution operation to extract features of the first similar image, and uses a recomposition function to obtain a pixel size that is relatively larger than the pixel size of the first similar image. The branch creates a second similar image.

The second decoder unit 230 includes a second convolutional layer unit 231 , a second placement normalization unit 233 , a second activation unit 235 , and a second recomposition unit 237 .

The second convolutional layer unit 231 receives the first likeness image and extracts features of the first likeness image by using a second convolutional layer having a preset filter kernel size (3 X 3).

The pixel size of the second convolutional layer may be 3 x 3, and the stride may have 1 pixel.

The second arrangement normalization unit 233 arrangement normalizes the distribution of pixel values belonging to the first similar image by setting the average of pixel values belonging to the first similar image to be 0 and the variance to be 1.

The second activator 235 generates an activated first likeness image using an activation (Leaky ReLu) function having a preset gradient (0.2) on the batch normalized first likeness image.

The second recomposition unit 237 generates a second likeness image having a pixel size that is relatively larger than a pixel size of the first likeness image by using a recomposition function on the activated first likeness image.

The second recomposition unit 237 may generate a second similar image having a pixel size of 112 x 112 and a pixel size of 256 .

The third decoder unit 250 filters the second similar image according to the convolution operation, extracts features of the second similar image, and uses a recomposition function to determine the pixel size of the lost PET image received from the encoder unit 100 . A similar PET image with the same pixel size as

The third decoder unit 250 includes a third convolutional layer unit 251 , a third batch normalization unit 253 , a third activation unit 255 , and a third recomposition unit 257 .

The third convolutional layer unit 251 receives the second likeness image and extracts features of the second likeness image by using a third convolutional layer having a preset filter kernel size (3 X 3).

The pixel size of the third convolutional layer may be 3 x 3, and the stride may have 1 pixel.

The third arrangement normalization unit 253 arrangement normalizes the distribution of pixel values belonging to the second similar image by setting the average of pixel values belonging to the second similar image to be 0 and the variance to be 1.

The third activator 255 generates an activated second likeness image using an activation (Leaky ReLu) function having a preset gradient (0.2) on the batch normalized second likeness image.

The third recomposition unit 257 generates a similar PET image having the same pixel size as that of the lost PET image by using a recomposition function on the activated second similar image.

The third recomposition unit 257 may generate a similar PET image having a pixel size of 224 x 224 and the number of filters of 128.

The linearization unit 300 generates a linearized similar PET image by adjusting the size of the received PET loss image using a convolutional layer to the similar PET image.

The pixel size of the convolutional layer may be 1 x 1, and the stride may be 1 pixel.

The similar PET image linearized by the linearization unit 300 may be changed to a single-channel image having a pixel size of 224 x 224 and a filter number of 1.

The output unit 400 generates and outputs a similar PET image having 400 x 400 pixels by adding (zero-padding) rows and columns of 0 to the top, bottom, left, and right sides of the linearized similar PET image of a single channel. .

The generator unit 40 generates a plurality of similar PET images by adding or subtracting an error value equal to or less than a pixel unit loss value to at least one pixel value among pixel values of the PET loss image using the lost PET image. it could be

The pixel unit loss calculator 50 calculates a pixel unit loss value based on the mean square error between the PET loss image and the similar PET image.

The pixel unit loss calculator 50 may calculate a pixel unit loss value between the similar PET image and the PET loss image based on a mean squared error according to Equation 1 below.

는 G(z)에 대한 확률 분포를 의미하며, E는 주어진 확률 분포에 대한 기댓값을 의미하고,

는 G(z)와 z의 minimum difference 값을 의미하는 것일 수 있다.Here, z means the normalized value for the PET loss image, G(z) means the normalized value for the similar PET image,

is the probability distribution for G(z), E is the expected value for a given probability distribution,

may mean a value of the minimum difference between G(z) and z.

The discriminator unit 60 compares the plurality of similar PET images generated by the generator unit 40 with the original PET image to determine whether the similar PET image belongs to an actual image classification corresponding to the original PET image or to the original PET image. It is determined whether it belongs to the false image classification that does not correspond.

6 and 7 , the discriminator unit 60 includes a first characteristic map generation unit 610 , a second characteristic map generation unit 630 , a third characteristic map generation unit 650 , and a linearization unit ( 670).

The first characteristic map generating unit 610 receives the PET original image, extracts the characteristics of the PET original image, and generates a first characteristic map.

The first characteristic map generating unit 610 includes a first convolutional layer unit 611 , a first arrangement normalization unit 613 , and a first activation unit 615 .

The first convolutional layer unit 611 receives the original PET image and extracts features of the original PET image by using the first convolutional layer having a preset pixel size.

The pixel size of the first convolutional layer may be 4 x 4, and the stride may be 2 pixels.

The first batch normalization unit 613 batch normalizes the distribution of pixel values belonging to the original PET image so that the average of pixel values belonging to the original PET image is 0 and the variance is 1.

The first activator 615 activates the batch normalized original PET image by using an activation (Leaky ReLu) function having a preset gradient (0.2) in the batch normalized original PET image, and generates a first characteristic map.

The first feature map generator 610 sets the pixel size of 224 x 224 and the number of filters 64 according to the first convolutional layer unit 611 , the first arrangement normalizer 613 , and the first activator 615 . The branch may be to generate the first characteristic map.

The second characteristic map generator 630 receives the first characteristic map, extracts the characteristic of the first characteristic map, and generates a second characteristic map.

The second characteristic map generating unit 630 includes a second convolutional layer unit 631 , a second arrangement normalization unit 633 , and a second activation unit 635 .

The second convolutional layer unit 631 receives the first feature map and extracts features of the first feature map by using the second convolutional layer having a preset pixel size.

The pixel size of the second convolutional layer may be 4×4, and the stride may be 2 pixels.

The second arrangement normalization unit 633 arrangement normalizes the distribution of pixel values pertaining to the first characteristic map such that the average of the pixel values belonging to the first characteristic map is 0 and the variance is 1.

The second activator 635 activates the batch normalized first feature map and generates a second feature map by using an activation (Leaky ReLu) function having a preset gradient (0.2) in the batch normalized first feature map.

The second characteristic map generator 630 sets the pixel size to 112 x 112 and the number of filters to 128 according to the second convolutional layer unit 631 , the second arrangement normalization unit 633 , and the second activator 635 . The branch may be to generate the second characteristic map.

The third characteristic map generation unit 650 receives the second characteristic map, extracts the characteristics of the second characteristic map, and generates a third characteristic map.

The third feature map generating unit 650 includes a third convolutional layer unit 651 , a third arrangement normalization unit 653 , and a third activation unit 655 .

The third convolutional layer unit 651 receives the second feature map and extracts features of the first feature map by using the third convolutional layer having a preset pixel size.

The pixel size of the third convolutional layer may be 4×4, and the stride may be 2 pixels.

The third arrangement normalization unit 653 arrangement normalizes the distribution of pixel values belonging to the second characteristic map such that the average of the pixel values belonging to the second characteristic map is 0 and the variance is 1.

The third activator 655 activates the batch normalized second feature map and generates a third feature map by using an activation (Leaky ReLu) function having a preset gradient (0.2) in the batch normalized second feature map.

The third feature map generator 650 sets the pixel size of 56 x 56 and the number of filters 256 according to the third convolutional layer unit 651 , the third arrangement normalization unit 653 , and the third activation unit 655 . The branch may be to generate the third characteristic map.

The linearization unit 670 uses a convolutional layer having a filter kernel size (1 X 1) preset in the third characteristic map and adjusts it to match the size of the received original PET image to generate a linearized original PET image.

The pixel size of the convolutional layer may be 1 x 1, and the stride may be 1 pixel.

The original PET image linearized by the linearization unit 670 may be changed to a single-channel image having a pixel size of 224 x 224 and a filter number of 1.

The error value calculation unit 70 calculates an error value between the similar PET image determined to belong to the false image classification by the delimiter unit 60 and the original PET image.

)과, 디스크리미네이터부(60)에 대한 오차값(

)을 산출하는 것일 수 있다.The error value calculating unit 70 calculates an error value between the similar PET image and the original PET image based on Equations 2 and 3 below, but the error value (

) and the error value for the discriminator unit 60 (

) can be calculated.

는 제네레이터(40)부에 대한 오차값의 최솟값을 의미하고, z는 PET 손실 이미지에 대한 정규화된 값을 의미하며, G(z)는 유사 PET 이미지에 대한 정규화된 값을 의미하고,

는 G(z)에 대한 확률 분포를 의미하며, E는 주어진 확률 분포에 대한 기댓값을 의미하고, D(G(z))는 디스크리미네이터부(60)에서 유사 PET 이미지에 대해 판별한 결과값을 의미하며,

는 G(z)와 z의 minimum difference 값을 의미하는 것일 수 있다.here,

denotes the minimum value of the error value for the generator 40, z denotes a normalized value for the PET loss image, and G(z) denotes a normalized value for the similar PET image,

denotes a probability distribution with respect to G(z), E denotes an expected value for a given probability distribution, and D(G(z)) denotes a result value determined by the discriminator unit 60 for a similar PET image. means,

may mean a value of the minimum difference between G(z) and z.

는 디스크리미네이터부(60)에 대한 오차값의 최솟값을 의미하고, z는 PET 손실 이미지에 대한 정규화된 값을 의미하며, x는 PET 원본 이미지에 대한 정규화된 값을 의미하고, G(z)는 유사 PET 이미지에 대한 정규화된 값을 의미하며,

는 원본 PET 이미지에 대한 확률 분포를 의미하며, E는 주어진 확률 분포에 대한 기댓값을 의미하고, D(x)는 디스크리미네이터부(60)에서 원본 PET 이미지에 대해 판별한 결과값을 의미하며, D(G(z))는 디스크리미네이터부(60)에서 유사 PET 이미지에 대해 판별한 결과값을 의미하는 것일 수 있다.here,

denotes the minimum value of the error value for the discriminator unit 60, z denotes a normalized value for the PET loss image, x denotes a normalized value for the PET original image, and G(z) is the normalized value for similar PET images,

denotes a probability distribution for the original PET image, E denotes an expected value for a given probability distribution, and D(x) denotes a result value determined by the discriminator unit 60 on the original PET image, D(G(z)) may mean a result value determined by the discriminator 60 for a similar PET image.

)과, 디스크리미네이터부(60)에 대한 오차값(

)의 합이 극소화되도록 PET 손실 이미지를 제네레이터부(40)와 디스크리미네이터부(60)에 기초하여 반복 학습시키는 것일 수 있다.The apparatus for restoring a brain tomography image taken for a short time using deep learning according to an embodiment of the present invention has an error value (

) and the error value for the discriminator unit 60 (

) may be repeated learning of the PET loss image based on the generator unit 40 and the discriminator unit 60 so that the sum of ) is minimized.

The false image set generating unit 80 compares the error values of each of the plurality of similar PET images determined to belong to the false image classification by the discriminator 60, and the similar PET with an error value equal to or less than a preset threshold value. A plurality of false PET image sets are generated based on a preset generative adversarial neural network model by removing images and storing similar PET images exceeding a threshold.

The discriminator unit 60 compares a plurality of similar PET images generated by the generator unit 40 with the original PET image, as a result of the error value calculated by the error value calculation unit 70 , and determines that the similar PET image is a PET image. It may be determined by calculating a first binary data value if the original image belongs to an actual image classification corresponding to the original image, and a second binary data value if the similar PET image belongs to a false image classification.

The generator unit 40 uses the false PET image set generated by the false image set generation unit 80 to correct by adding or subtracting an error value less than or equal to at least one pixel unit loss value among pixel values of the PET loss image. It may be to generate a plurality of similar PET images according to the

The discriminator unit 60 compares the original PET image with a plurality of similar PET images generated by using the false PET image set in the generator unit 40 so that the similar PET image is classified as an actual image corresponding to the original PET image. It may be to determine whether it belongs to or belongs to a false image classification that does not correspond to the original PET image.

The restored image output unit 90 generates and outputs a restored image based on the similar PET image determined by the discriminator 60 to belong to the actual image classification.

8 shows a restored image of a brain tomography image restoration apparatus taken for a short time using deep learning according to an embodiment of the present invention.

The restored image output unit 90 adds 0 rows and columns to the top, bottom, left, and right sides of the single-channel image having a size of 224 x 224 pixels determined to belong to the actual image classification in the discriminator unit 60 . Thus, a restored image having a size of 400 x 400 pixels may be generated and output.

Hereinafter, in order to evaluate the performance of a brain tomography image restoration apparatus taken for a short time using deep learning according to an embodiment of the present invention, the validity of a deep learning-based restoration approach through a short time PET scan was reviewed.

9 is a PET image generated using the conventional U-net method.

The lost PET image according to FIG. 2 has a severe noise component and poor image quality, and the reconstructed image generated according to the present invention according to FIG. 8 has more anatomical details than the PET image generated from the U-net according to FIG. It can be seen that it reflects better.

Although both the present invention and the U-net method significantly reduce the noise component, the U-net produces a more blurry PET image than the present method.

For quantitative comparison between the method of the present invention and the U-net method, average PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural SIMilarity), NRMSE (Normalized Root Mean Square Error) for a test data set (58 patients) was calculated.

Parameters PET loss image restored image U-net method
created
PET image PSNR 33.035 35.826 34.600 SSIM 0.884 0.882 0.869 NRMSE 15.421 11.286 12.921

Table 1 shows a PET loss image, a restored image generated by a brain tomography image restoration apparatus taken for a short time using deep learning according to an embodiment of the present invention, PSNR, SSIM of a PET image generated by the U-net method, It represents NRMSE.

Referring to Table 1, it can be seen that the PSNR and SSIM of the reconstructed image are the highest, and the NRMSE is the lowest.

10 is an average standardized uptake value ratio (SVR) according to a change in VOI (Volume Of Interest) of a restored image of a brain tomography image restoration apparatus taken for a short time using deep learning and a PET original image according to an embodiment of the present invention; ) is a graph showing the change in

As shown in FIG. 10 , it can be seen that very similar values are displayed between the reconstructed image 1001 and the original PET image 1001 in most areas.

11 is a graph illustrating a Pearson correlation between a restored image of a brain tomography image restoration apparatus taken for a short time using deep learning and an original PET image according to an embodiment of the present invention.

11 , it can be seen that there is a very strong positive correlation between the SUVR of the reconstructed image and the image of the original image (r = 0.998, p < 0.001).

12 is a graph illustrating a Bland-Altman analysis of a restored image of a brain tomography image restoration apparatus taken for a short time using deep learning and an original PET image according to an embodiment of the present invention.

The Bland-Altman analysis is a graph showing the error of the values obtained by the two measurement methods or whether there is a difference between the estimated value and the actual measured value.

Referring to FIG. 12 , a very small SUVR mean difference of 0.016 (95% CI (confidence interval): 0.013, 0.019) appears between the reconstructed image and the original PET image in the Bland-Altman analysis.

To investigate the feasibility of a deep learning-based restoration method through a short-time PET scan of the brain tomography image restoration apparatus taken for a short time using deep learning according to an embodiment of the present invention, the brain was subjected to F-18 flow for 2 minutes. The PET loss image from betaben positron emission tomography and the original PET image from F-18 fluorbetaben positron emission tomography of the brain for 20 min were used as input images and comparison images, respectively.

It has been shown that an efficient reconstruction image is generated from a PET loss image by the apparatus for restoring a brain tomography image taken for a short time using deep learning according to an embodiment of the present invention.

Image quality metrics for model evaluation (PSNR, SSIM, NRMSE, etc.) between the reconstructed image and the original PET image were calculated.

The apparatus for restoring a brain tomography image taken for a short time using deep learning according to an embodiment of the present invention improves image quality by suppressing a noise component of a PET loss image taken with a short time brain.

와,

에 따라 달라지는데 여기서 L은 픽셀 값의 동적 범위이며, K는 일정한 상수이다. 복원 이미지에 대한 평균 SSIM 지수는

,

가

,

에서

,

로 증가하였을 때, 0.8818에서 0.9939로 증가하였다. 그러나, 이 경우, SSIM 지수의 변동은 매우 낮았다.The Structural SIMilarity (SSIM) index, which evaluates image quality, is a parameter in the SSIM formula

Wow,

where L is the dynamic range of pixel values, and K is a constant constant. The average SSIM index for the restored image is

,

go

,

at

,

, increased from 0.8818 to 0.9939. However, in this case, the fluctuation of the SSIM index was very low.

The present invention employs a generative adversarial network (GAN) as an additional mean square error between the reconstructed image and the PET loss image. The performance of the present invention was compared with U-net. U-net minimizes only pixel loss between the reconstructed image and the original PET image, thereby overloading the image, whereas the reconstructed image according to the embodiment of the present invention has the advantage of clearly reconstructing the detailed structure of the brain.

In the measurement of image quality such as PSNR, NRMSE and SSIM, the present invention outperformed U-net. The time taken to generate a reconstructed image from a PET loss image was within a few ms on the GPU system, which would make it very useful for clinical use for the present invention.

Since amyloid positron emission tomography (PET) images are intended to treat patients with memory impairment, deep learning-generated images should be available for clinical readout purposes.

The apparatus for restoring a brain tomography image taken for a short time using deep learning according to an embodiment of the present invention uses various methods to determine whether the restored image can be used clinically.

Tests were conducted, such as whether the doctor can distinguish the original PET image from the reconstructed image, the difference in visual interpretation results, and the difference in the statistical analysis using the Standardized Uptake Value Ratio (SUVR) in the two images.

Readers Test 1 Test 2 4yrs experienced physician 26/58 (44.8%) 28/58 (48.3%) Over 15yrs experienced physician 1 37/58 (63.8%) 35/58 (60.3%) Over 15yrs experienced physician 2 26/58 (44.8%) 32/58 (54.2%)

Table 2 shows the results of Test 1, which is a test to distinguish between a restored image and a fake or real image, and Test 2, a test to select the original PET image from the restored image and the original PET image of the same patient. will be.

Referring to Table 2, Test 1 showed that the overall accuracy was not high regardless of the length of clinical reading experience in nuclear medicine.

Also, in Test 2, the more experienced doctors had in clinical reading, the higher the frequency (48.3% to 60.3%) of selecting the original PET image. However, overall, the clinician could not distinguish the reconstructed image from the original PET image.

Preferably, when the PET original image and the reconstructed image were presented simultaneously to three nuclear medicine specialists who were clinically reading, the accuracy of the original image selection was within 40% to 60%. This suggests that the reconstructed image generated by the present invention is hardly distinguishable from the original PET image.

Next, a BAPL (Brain Amyloid Plaque Load) score test was performed to evaluate intra-observer agreement and diagnostic accuracy.

Metric Reader 1 Reader 2 Reader 3 Mean Accuracy 91.4% (81.0, 97.1) 89.7% (78.8, 96.1) 86.2% (74.6, 93.9) 89.1% Sensitivity 95.2% (83.8, 99.4) 88.1% (74.4, 96.0) 90.5% (77.4, 97.3) 91.3% Specificity 81.3% (54.4, 96.0) 93.8% (69.8, 99.8) 75.0% (47.6, 92.7) 83.3%

Table 3 shows the Accuracy, Sensitivity, and Specificity of three doctors in clinical readings using BAPL scores for reconstructed images. In addition, the data in parentheses in Table 3 are 95% confidence intervals.

Three doctors evaluated the reconstructed images according to the BAPL score system, and there were no inappropriate images that were difficult to interpret. In 5, 6, and 8 of 58 patients, each physician evaluated the BAPL score differently from the Ground-Truth score.

Overall, the mean values for Accuracy, Sensitivity, and Specificity were 89.1%, 91.3%, and 83.3%, respectively.

Original Image (Ground-Truth) Reader 1 Reader 2 Reader 3 BS1 BS2 BS3 Total BS1 BS2 BS3 Total BS1 BS2 BS3 Total restore
image sBS1 13 2 0 15 15 5 0 20 12 4 0 16 sBS2 3 16 0 19 One 13 0 14 4 14 0 18 sBS3 0 0 24 24 0 0 24 24 0 0 24 24 Total 16 18 24 58 16 18 24 58 16 18 24 58

Table 4 shows the confusion metrics for the PET image reading using the BAPL score between the original PET image and the reconstructed image.

Here, BS means the BAPL score of Ground-Truth, sBS means the BAPL score of the restored image, GT means the Ground-Truth score, BS1 is BAPL 1 (negative reaction), BS2 is BAPL 2 (weak positive), BS3 may mean BAPL 3 (strongly positive).

For BAPL 3 (strong positive), all 3 doctors showed 100% accuracy, but for BAPL 1 (negative) and BAPL 2 (weak positive), 5-8 out of 58 patients were false positives or false positives. . Interpretation of how positive amyloid uptake is in the visual reading that distinguishes BAPL 1 and BAPL 2 is subjective.

To compensate for the weakness of visual readings, SUVR is used as a quantitative indicator in routine practice to infer disease severity or prognosis.

In the restored image of the brain tomography image restoration apparatus taken for a short time using deep learning according to an embodiment of the present invention, the SUVR of 52 brain regions is very high with the value of the original PET image (r = 0.998, p < 0.001). showed a correlation.

In the Bland-Altman analysis, the average of the difference values was 0.016 (1.6%), and the range of variation in the SUVR value was very small. That is, the present invention can generate a reconstructed image having an SUVR value comparable to the original PET image.

Taken together, these results suggest that the restored image generated by the brain tomography image restoration apparatus taken for a short time using deep learning according to an embodiment of the present invention can be used for clinical reading purposes.

As mentioned above, although the present invention has been described in detail through preferred embodiments, the present invention is not limited thereto and may be practiced in various ways within the scope of the claims. Since the foregoing has described rather broadly the features and technical strengths of the present invention in order to better understand the claims of the present invention, the concepts and specific embodiments of the present invention described above are not intended to be used for other purposes similar to the present invention. It should be recognized by those skilled in the art that it can be used immediately as a basis for designing or changing the shape.

It will be understood that the above-described embodiment is only one embodiment according to the present invention, and may be implemented in various changes and modifications within the scope of the technical spirit of the present invention by those of ordinary skill in the art. will be able Accordingly, the disclosed embodiments should be considered in an illustrative rather than a restrictive viewpoint, and such various changes and modifications are also shown in the claims of the present invention to be described later as belonging to the scope of the technical spirit of the present invention, and the scope equivalent thereto All differences therein should be construed as being included in the present invention.

1: A brain tomography image restoration device taken for a short time using deep learning,
10: loss image collection unit;
20: original image collection unit;
30: cropping unit,
40: generator unit,
50: pixel unit loss calculator;
60: disc liminator unit,
70: error value calculation unit;
80: false image set generating unit;
90: restored image output unit,
100: encoder unit;
110: a first encoder unit;
111: a first convolutional layer unit of the first encoder unit;
113: a first batch normalization unit of the first encoder unit;
115: a first activation unit of the first encoder unit;
117: a first decomposition unit;
117a: a first high-frequency extraction unit,
117b: a first low-frequency extraction unit,
130: a second encoder unit;
131: a second convolutional layer unit of the second encoder unit;
133: a second batch normalization unit of the second encoder unit;
135: a second activation unit of the second encoder unit;
137: a second decomposition unit;
137a: a second high-frequency extraction unit,
137b: a second low-frequency extraction unit,
150: a third encoder unit;
151: a third convolutional layer unit of the third encoder unit;
153: a third batch normalization unit of the third encoder unit;
155: a third activation unit of the third encoder unit;
157: a third decomposition unit;
157a: a third high-frequency extraction unit,
157b: a third low-frequency extraction unit,
200: decoder unit;
210: a first decoder unit;
211: a first convolutional layer unit of the first decoder unit;
213: a first batch normalization unit of the first decoder unit;
215: a first activation unit of the first decoder unit;
217: a first recomposition unit;
230: a second decoder unit;
231: a second convolutional layer unit of the second decoder unit;
233: a second batch normalization unit of the second decoder unit;
235: a second activation unit of the second decoder unit;
237: a second recomposition unit;
250: a third decoder unit;
251: a third convolutional layer unit of the third decoder unit;
253: a third batch normalization unit of the third decoder unit;
255: a third activation unit of the third decoder unit;
257: a third recomposition unit;
300: a linearization unit of the generator unit;
400: output unit,
610: a first characteristic map generator;
611: a first convolutional layer unit of the first characteristic map generation unit;
613: a first batch normalization unit of the first characteristic map generation unit;
615: a first activation unit of the first characteristic map generation unit;
630: a second characteristic map generator;
631: a second convolutional layer unit of the second characteristic map generating unit;
633: a second batch normalization unit of the second characteristic map generation unit;
635: a second activation unit of the second characteristic map generation unit;
650: a third characteristic map generator;
651: a third convolutional layer unit of the third feature map generating unit;
653: a third batch normalization unit of the third characteristic map generation unit;
655: a third activation unit of the third characteristic map generation unit;
670: linearization unit of the discriminator,
1001: change in average SUVR according to change in VOI of the restored image,
1003: Change in average SUVR according to change in VOI of the original PET image.

Claims (19)

a loss image collecting unit for collecting a PET loss image including a noise component obtained by tomography the brain for a first preset time;
an original image collecting unit which is acquired by tomography the brain for a second time period longer than the first time period, and collects an original tomographic PET image corresponding to the PET loss image;
a generator unit for generating a plurality of similar PET images by removing the noise component of the lost PET image;
By comparing the plurality of similar PET images generated by the generator with the original PET image, whether the similar PET image belongs to an actual image classification corresponding to the original PET image or a false image that does not correspond to the original PET image a discriminator unit that determines whether it belongs to a classification; and
a restored image output unit for generating and outputting a restored image based on the similar PET image determined by the discriminator to belong to the actual image classification;
A brain tomography image restoration device taken for a short time using deep learning, including a. According to claim 1,
The loss image collection unit,
collecting images of the PET loss by F-18 fluorbetaben positron emission tomography of the brain during the first hour;
The original image collection unit,
collecting the PET raw images by subjecting the brain to the F-18 florbetaben positron emission tomography during the second time period;
A brain tomography image restoration device taken for a short time using deep learning. 3. The method of claim 2,
a pixel unit loss calculator for calculating a pixel unit loss value based on a mean square error between the PET loss image and the original PET image;
The generator unit,
Generating a plurality of similar PET images by correcting at least one pixel value among pixel values of the PET loss image by adding or subtracting an error value equal to or less than the loss value per pixel by using the lost PET image
A brain tomography image restoration device taken for a short time using deep learning. 3. The method of claim 2,
an error value calculation unit for calculating an error value between the similar PET image determined to belong to the false image classification by the discriminator unit and the original PET image; and
The discriminator compares the error value of each of the plurality of similar PET images determined to belong to the false image classification, removes the similar PET image with the error value equal to or less than a preset threshold value, and sets the threshold value. Further comprising; a false image set generator that generates a plurality of false PET image sets based on a preset generative adversarial neural network model according to the excessively similar PET images
A brain tomography image restoration device taken for a short time using deep learning. 5. The method of claim 4,
The generator unit,
Using the false PET image set generated by the false image set generator, adding or subtracting an error value less than or equal to the pixel loss value based on the mean square error to at least one pixel value among the pixel values of the PET loss image. Generating a plurality of similar PET images by modification
A brain tomography image restoration device taken for a short time using deep learning. 3. The method of claim 2,
The discriminator unit,
a convolutional layer unit for receiving the original PET image and extracting a feature map of the original PET image using a first convolutional layer having a preset filter kernel size;
a batch normalization unit for batch normalizing the distribution of pixel values belonging to the feature map such that the pixel values belonging to the feature map have an average of 0 and a variance of 1;
an activator configured to generate an activated feature map using an activation function having a preset gradient for the batch normalized feature map; and
A second convolutional layer having a filter kernel size that is relatively smaller than that of the first convolutional layer is used in the activated feature map to match the size of the received original PET image to generate a linearized PET image. Linearization unit; including
A brain tomography image restoration device taken for a short time using deep learning. 7. The method of claim 6,
The discriminator unit,
By comparing the plurality of similar PET images generated by the generator with the original linearized PET image, it is determined whether an error value between the similar PET image and the linearized original PET image is less than or equal to a preset threshold, Determining whether a similar PET image corresponds to the linearized original PET image
A brain tomography image restoration device taken for a short time using deep learning. 3. The method of claim 2,
The generator unit,
The PET loss image is filtered according to a convolution operation to extract features of the lost PET image, and a pixel size relatively smaller than the pixel size of the lost PET image using a decomposition function that reduces the pixel size of the lost PET image. an encoder unit for generating a feature map of the PET loss image having a size; and
Receives the feature map of the lost PET image, filters the feature map of the lost PET image by convolutional operation, extracts features of the feature map of the lost PET image, and enlarges the pixel size of the feature map of the lost PET image A decoder unit that generates a similar PET image having the same pixel size as the pixel size of the lost PET image by using a recomposition function
A brain tomography image restoration device taken for a short time using deep learning. 9. The method of claim 8,
The encoder unit,
The PET loss image is received and the PET loss image is filtered according to a convolution operation to extract the characteristics of the PET loss image, and a pixel size relatively smaller than the pixel size of the PET loss image is obtained by using the decomposition function. a first encoder unit for generating a first feature map with branches;
A second feature that filters the first feature map according to a convolution operation to extract a feature of the first feature map, and uses the decomposition function to have a pixel size that is relatively smaller than the pixel size of the first feature map a second encoder unit for generating a map; and
The PET loss having a smaller pixel size than the pixel size of the second feature map is extracted by filtering the second feature map according to the convolution operation, and extracting the feature of the second feature map using the decomposition function. A third encoder unit that generates a feature map of the image;
A brain tomography image restoration device taken for a short time using deep learning. 9. The method of claim 8,
The decoder unit,
The feature map of the lost PET image is filtered according to the convolution operation to extract the feature of the feature map of the lost PET image, and the pixel size of the feature map of the lost PET image is relatively larger than the pixel size of the feature map of the lost image by using the recomposition function. a first decoder unit for generating a first likeness image having a size;
Filtering the first similar image according to a convolution operation to extract features of the first similar image, and using the recomposition function, a second similarity having a relatively larger pixel size than the pixel size of the first similar image a second decoder unit for generating an image; and
Filtering the second similar image according to a convolution operation to extract features of the second similar image, and using the recomposition function to have the same pixel size as the pixel size of the lost PET image received from the encoder unit What includes; a third decoder unit that generates a similar PET image
A brain tomography image restoration device taken for a short time using deep learning. 10. The method of claim 9,
The first encoder unit,
a first convolutional layer unit receiving the PET loss image and extracting features of the PET loss image using a first convolutional layer having a preset filter kernel size;
a first batch normalization unit for batch normalizing the distribution of pixel values belonging to the PET loss image such that the average of the pixel values belonging to the PET loss image is 0 and the variance is 1;
a first activator configured to generate an activated PET loss image by using an activation function having a preset gradient on the batch normalized PET loss image; and
A first decomposition unit for generating a first feature map having a pixel size relatively smaller than a pixel size of the lost PET image by using the decomposition function in the activated PET loss image;
A brain tomography image restoration device taken for a short time using deep learning. 12. The method of claim 11,
The second encoder unit,
a second convolutional layer unit receiving the first feature map and extracting features of the first feature map using a second convolutional layer having a preset filter kernel size;
a second arrangement normalizer for arrangement normalizing the distribution of pixel values pertaining to the first feature map such that the pixel values belonging to the first feature map have an average of 0 and a variance of 1;
a second activator configured to generate an activated first feature map using an activation function having a preset gradient on the batch normalized first feature map; and
Including;
A brain tomography image restoration device taken for a short time using deep learning. 13. The method of claim 12,
The third encoder unit,
a third convolutional layer unit receiving the second feature map and extracting features of the second feature map using a third convolutional layer having a preset filter kernel size;
a third arrangement normalization unit for arrangement normalizing the distribution of pixel values pertaining to the second feature map such that the pixel values belonging to the second feature map have an average of 0 and a variance of 1;
a third activator configured to generate an activated second feature map using an activation function having a preset gradient for the batch normalized second feature map; and
a third decomposition unit generating a feature map of the PET loss image having a pixel size relatively smaller than a pixel size of the second feature map by using the decomposition function in the activated second feature map; thing
A brain tomography image restoration device taken for a short time using deep learning. 11. The method of claim 10,
The first decoder unit,
a first convolutional layer unit receiving the feature map of the lost PET image and extracting features of the feature map of the lost PET image using a first convolutional layer having a preset filter kernel size;
a first batch normalization unit for batch normalizing the distribution of pixel values belonging to the feature map of the lost PET image by making the average of the pixel values belonging to the feature map of the lost PET image to be a value of 0 and a value having a variance of 1;
a first activator configured to generate a feature map of an activated PET loss image by using an activation function having a preset gradient for the batch normalized feature map of the PET loss image; and
a first recomposition unit for generating a first similar image having a pixel size relatively larger than a pixel size of a feature map of the lost PET image by using the recomposition function in the feature map of the activated PET loss image; to do
A brain tomography image restoration device taken for a short time using deep learning. 15. The method of claim 14,
The second decoder unit,
a second convolutional layer unit receiving the first likeness image and extracting features of the first likeness image by using a second convolutional layer having a preset filter kernel size;
a second batch normalization unit for batch normalizing the distribution of pixel values belonging to the first similar image by setting the pixel values belonging to the first similar image to have an average of 0 and a variance of 1;
a second activator configured to generate an activated first likeness image using an activation function having a preset gradient on the batch normalized first likeness image; and
a second recomposition unit configured to generate a second likeness image having a pixel size that is relatively larger than a pixel size of the first likeness image by using the recomposition function on the activated first likeness image;
A brain tomography image restoration device taken for a short time using deep learning. 16. The method of claim 15,
The third decoder unit,
a third convolutional layer unit receiving the second likeness image and extracting features of the second likeness image by using a third convolutional layer having a preset filter kernel size;
a third batch normalization unit for batch normalizing the distribution of pixel values belonging to the second similar image by setting an average of 0 and a variance of 1 of the pixel values belonging to the second similar image;
a third activator configured to generate a second likeness image activated by using an activation function having a preset gradient on the batch normalized second likeness image; and
a third recomposition unit for generating the pseudo PET image having the same pixel size as the pixel size of the lost PET image by using the recomposition function on the activated second similar image;
A brain tomography image restoration device taken for a short time using deep learning. 12. The method of claim 11,
The first decomposition unit,
a first high frequency extraction unit for calculating one pixel value based on a difference between at least two pixel values among the pixel values of the first feature map; and
A first low-frequency extraction unit for calculating one pixel value based on an average between at least two pixel values among the pixel values of the first feature map;
A brain tomography image restoration device taken for a short time using deep learning. 10. The method of claim 9,
The second decomposition unit,
a second high frequency extraction unit for calculating one pixel value based on a difference between at least two pixel values among the pixel values of the second feature map; and
A second low-frequency extraction unit for calculating one pixel value based on an average between at least two pixel values among the pixel values of the second feature map;
A brain tomography image restoration device taken for a short time using deep learning. 10. The method of claim 9,
The third decomposition unit,
a third high frequency extraction unit for calculating one pixel value based on a difference between at least two pixel values among pixel values of the feature map of the PET loss image; and
A third low-frequency extractor for calculating one pixel value based on an average between at least two pixel values among the pixel values of the feature map of the PET loss image;
A brain tomography image restoration device taken for a short time using deep learning.

Images