KR102472464B1 - Image Processing Method and Image Processing Device using the same - Google Patents

Abstract

The present invention relates to an image processing method, and more specifically, to an image processing system based on generative adversarial networks (GAN). an image conversion step (S10) of removing noise from the first CT image, converting it into a virtual CT image, and outputting the image; an image discrimination step (S20) of determining authenticity as the virtual CT image or the 2nd CT image captured by the CT device is input to the discriminator (D); A learning step (S30) of adjusting parameters of the generator (G) or the discriminator (D) according to the authenticity output from the discriminator (D) is disclosed.

Description

Image processing method and image processing device using the same {Image Processing Method and Image Processing Device using the same}

The present invention relates to an image processing method for removing noise from a CT image and an image processing apparatus using the same.

In the medical field, computed tomography (CT, hereinafter CT) transmits X-rays through the human body and obtains images in tomograms on a cross-section. Unlike, there are no restrictions due to bones, etc., and it is in the spotlight because it can quickly photograph a wide area.

On the other hand, radiation is inevitably generated during CT imaging, and since radiation has negative effects on the human body, such as modifying DNA or causing cancer, a technique for minimizing radiation dose to patients during CT imaging is required.

Ideally, the radiation dose should minimize the amount exposed to the patient, but this may cause a CT image quality deterioration due to an increase in image noise, which may cause difficulties in medical judgment by doctors.

Therefore, in general, hospitals perform low-dose (low-radiation) CT scans that allow a certain degree of image noise.

However, in order to eliminate noise in low-dose CT, various deep learning algorithms have been proposed over the past few years, and paired datasets obtained by imaging under two conditions using a human body model (Phantom) There was a method using a method, and a method of obtaining a deep learning algorithm after obtaining paired data by generating a low-dose CT image through numerical simulation.

However, these methods are different from the actual clinical environment, so there are limitations in applying them directly to clinical practice.

An object of the present invention is to disclose an image correction deep learning algorithm using mismatched images in order to solve the above problems.

The present invention was created to achieve the object of the present invention as described above,

As an image processing system based on generative adversarial networks (GAN), as the first CT image captured by the CT device is input to the generator (G), noise is removed from the first CT image to obtain a virtual CT image. An image conversion step (S10) of converting and outputting; an image discrimination step (S20) of determining authenticity as the virtual CT image or the 2nd CT image captured by the CT device is input to the discriminator (D); Disclosed is an image processing method comprising a learning step (S30) of adjusting parameters of the generator (G) or the discriminator (D) according to the authenticity output from the discriminator (D).

The image processing method of the present invention may be characterized in that the 1st CT image and the 2nd CT image do not match.

The tube current TC2 in the CT scan of the 2nd CT image may be greater than the tube current TC1 in the CT scan of the 1st CT image.

The tube voltage (TV2) in the CT scan of the 2nd CT image may be the same as the tube voltage (TV1) in the CT scan of the 1st CT image.

The constructor (G),

은 상기 제1CT영상이고, 상기

은 상기 생성자(G)가 상기 제1CT영상으로부터 변환한 상기 가상의 CT영상이며, 상기 λ는 조정되는 파라미터이고, 상기

는 상기 가상의 CT영상에 대한 상기 판별자(D)의 판별결과이며, 상기

는 상기 제1CT영상에 대한 확률분포를 나타내며, E는 주어진 확률분포에 대한 기댓값을 나타낼 수 있다.has a network of, wherein the

is the first CT image, and

is the virtual CT image converted from the 1st CT image by the generator (G), λ is a parameter to be adjusted,

Is the discrimination result of the discriminator (D) for the virtual CT image,

represents a probability distribution for the first CT image, and E may represent an expected value for a given probability distribution.

는, 상기

의 tissue 값이 0이 되도록

로 리스케일링되되, 여기서, 상기

는 리스케일링 된 상기 제1CT영상이고, 상기

는 평균 HU(Hounsfield Unit)값이며, 상기 c는 CT의 최댓값의 역수일 수 있다.remind

, the above

so that the tissue value of is 0

Rescaled to , where, the above

is the rescaled 1 CT image,

Is an average Hounsfield Unit (HU) value, and c may be the reciprocal of the maximum value of CT.

으로부터 가상의 CT영상인

를 변환하되, 상기

는,

의 함수를 가지고, 여기서, 상기

는 함수의 합성을 나타내며, 상기

는 k번째 레이어를 나타내고, 상기 k는 조정되는 파라미터이며, k>0일 수 있다.The generator (G) is the rescaled 1CT image

virtual CT image from

Convert the above

Is,

With a function of where, the

Represents the composition of the function, wherein

denotes a k-th layer, k is a parameter to be adjusted, and may be k>0.

의 함수를 가지며, 여기서, 상기

는 상기 제2CT영상이며, 상기

는 상기 가상의 CT영상에 대한 상기 판별자(D)의 판별결과이고, 상기

는 상기 제2CT영상에 대한 상기 판별자(D)의 판별결과이며, 상기

는 상기 제2CT영상에 대한 확률분포를 나타내고, 상기

는 상기 제1CT영상에 대한 확률분포이며, E는 주어진 확률분포에 대한 기댓값을 나타낼 수 있다.The discriminator (D),

Has a function of, where the

is the second CT image, and

Is the discrimination result of the discriminator (D) for the virtual CT image,

Is the discrimination result of the discriminator (D) for the 2nd CT image,

Represents a probability distribution for the second CT image,

Is a probability distribution for the first CT image, and E may represent an expected value for a given probability distribution.

는, 상기

의 tissue의 값이 0이 되도록

로 리스케일링되되, 여기서, 상기

는 리스케일링 된 상기 제2CT영상이고, 상기

는 평균 HU(Hounsfield Unit)값이며, 상기 c는 CT의 최댓값의 역수일 수 있다.remind

, the above

so that the value of tissue in is 0

Rescaled to , where, the above

is the rescaled 2 CT image,

Is an average Hounsfield Unit (HU) value, and c may be the reciprocal of the maximum value of CT.

In addition, an image processing device including a memory and a processor for storing the program of the present invention, wherein the image processing method according to any one of claims 1 to 9 is performed when the programs are executed by the processor. Initiate.

According to the present invention, by using a patient image obtained by matching tube voltage and increasing tube current as a reference image, brightness in a low-dose image can be maintained while noise can be removed as in a high-dose image.

In addition, the present invention has an advantage in that it can be directly applied to existing CT equipment and medical treatment environments because image processing methods can be learned from actual clinical images that do not match.

In addition, the present invention has the advantage of efficiently denoising training by placing tanh(kx) at the end of the rescaling and generator (G) networks.

In addition, the present invention has an advantage in that image quality can be improved by applying it to a conventional CT device that does not have a low-dose image noise removal technology.

In addition, the present invention has an advantage of improving the correction performance by configuring the network structure of the image correction algorithm and the image provided thereto to be suitable for a low-dose CT environment.

In addition, the present invention has the possibility of being extended to image correction by various image deterioration factors (i, e., motion artifacts, metal artifacts) in addition to noise removal.

1 is a conceptual diagram showing the training flow of the image processing method of the present invention.
FIG. 2 is a conceptual diagram showing a flow of image processing performed in an actual clinical environment using an image processing method trained according to the image processing method of FIG. 1 .
3 is a table showing tube current, tube voltage, and radiation dose indicators (CTDI, DLP) of a CT image used in the image processing method of the present invention.
4 is a graph showing a generator (G) using the learning method of the image processing method of the present invention.
5A is a table showing the noise removal efficiency improvement effect by the learning method of the image processing method of the present invention.
5B is a photograph showing the effect of improving the noise removal efficiency by the learning method of the image processing method of the present invention.

Hereinafter, an image correction system according to the present invention and an image correction method using the same will be described with reference to the accompanying drawings.

A key feature of the present invention is to provide a technology for reducing noise in an image using an artificial neural network learned based on deep learning, and in particular, to learn an image processing method with generative adversarial networks (GANs). .

That is, the present invention is an image processing method using a network based on generative adversarial networks (GAN). an image conversion step (S10) of removing noise, converting the image into a virtual CT image, and outputting the image; an image discrimination step (S20) of determining authenticity as the virtual CT image or the 2nd CT image captured by the CT device is input to the discriminator (D); Disclosed is an image processing method comprising a learning step (S30) of adjusting the parameters of the generator (G) according to the authenticity output from the discriminator (D).

Here, the image can be variously interpreted as a concept including a static image (image) and a dynamic image.

The image processing method is a method of correcting image noise, and can be corrected regardless of the type of image such as X-ray, Molecular Imaging and computed tomography (MICT), and Magnetic Resonance imaging (MRI). Images are preferably used for CT images of the abdomen, chest, genitourinary system, spine, etc. taken by a CT device.

Here, CT images can be classified into low dose computed tomography (LDCT) and standard dose computed tomography (SDCT) according to the relative radiation dose to which the patient is exposed during imaging, and each has different characteristics.

For example, the lower the radiation dose, the higher the brightness level of the image, but the higher the noise, and the higher the radiation dose, the lower the brightness level but the lower the noise.

Here, the radiation dose is affected by various factors such as tube voltage, tube current, pitch, beam width control, scan range and number of times.

Meanwhile, an image processing method to be described later learns an image correction method using the 1st CT image and the 2nd CT image acquired from a patient, and the virtual CT image output by converting the 1st CT image.

Here, the first CT image is an image data set captured by a CT imaging device, and may consist of one or more images.

The first CT image is an image taken of an actual patient, and may be a low dose CT image that allows only readable image noise.

On the other hand, the virtual CT image is an image output by removing noise from the first CT image by the creator (G) and converting it into a virtual CT image, and may be composed of one or more images.

Specifically, the virtual CT image is not an image taken of an actual patient, but a virtual CT image generated by the creator (G) by removing noise from the first CT image.

At this time, various methods for removing noise from the 1st CT image may be used.

For example, spatial domain filter, nonlocal main filter, bilateral filter, NLBayes (Non-Local Bayes), BM3D (Block-matching and 3D) as filtering-based noise removal methods filtering) WNNM (Weighted Nuclear Norm Minimization), etc. As a deep learning-based denoising method, various denoising methods such as DnCNN (Denoising Convolutional Neural Network) that removes Gaussian noise can be used.

Meanwhile, the 2nd CT image is an image data set captured by a CT imaging device, and may consist of one or more images, and may be an image with reduced noise compared to the 1st CT image.

In order to reduce the noise of the 2nd CT image compared to the 1st CT image, radiation dose can be increased during imaging, and for this purpose, tube voltage (TV) and tube current (Tube Current, TC) can be adjusted.

Conventionally, for image correction, an image in which noise was reduced by increasing tube voltage during CT scan was used as a reference image (the second CT image of the present invention). The image also had a problem in that the brightness level of the image was lowered.

That is, as shown in FIG. 5A, the virtual CT image model (Group2) obtained by increasing the tube voltage of the 2nd CT image than that of the 1st CT image has a higher attenuation (shading, brightness) value than that of the low-dose image (Group1). It can be seen by confirming that the brightness level of the Group2 image is lower than that of the Group1 image in FIG. 5B.

Accordingly, in the present invention, as shown in FIG. 3, the tube voltage of the low-dose image (LDCT) corresponding to the 1st CT image and the standard-dose image (SDCT) corresponding to the 2nd CT image is the same as 80 kVp, and the low-dose image ( The tube current of LDCT) was 85 mAs, and the tube current of standard dose image (SDCT) was 160 mAs, so that the second CT image could be obtained while maintaining the brightness level while increasing the radiation dose.

In other words, the tube current (TC2) in the CT scan of the 2nd CT image is made larger than the tube current (TC1) in the CT scan of the 1st CT image, and the tube voltage (TV2) in the CT scan of the 2nd CT image is set to the first. By setting the tube voltage (TV1) in the 1CT image CT scan to be the same, it was possible to acquire the 2nd CT image with reduced noise while maintaining the brightness level.

Meanwhile, the 2nd CT image may not be matched with the 1st CT image.

That is, the 2nd CT image and the 1st CT image may be non-matched images taken from different patients, and this is because it is ethically unacceptable to take CT scans more than once on a single patient, and paired data This is because it is difficult to obtain from real patients.

Meanwhile, the gist of the present invention is to generate a virtual CT image, that is, a virtual standard-dose CT image, by removing noise from a low-dose CT image using a generative adversarial neural network-based network.

In a generative adversarial network-based network, generators (G) and discriminators (D) are alternately learned, generators (G) generate virtual data, and discriminators (D) distinguish virtual data from real data. The generator (G) generates virtual data more elaborately so that the discriminator (D) cannot discriminate, and the discriminator (D) discriminates the virtual data more precisely. As a result, the generator (G) generates data similar to reality.

That is, in the present invention, the generator (G) generates a virtual CT image from the 1st CT image, and the discriminator (D) discriminates between the virtual CT image and the 2nd CT image, and through learning, the generator (G) more elaborately A virtual CT image is generated, and the discriminator (D) discriminates the virtual CT image more precisely.

Here, the generative adversarial neural network has no constraints on image generation, so there is a concern that the morphological information of the image to be corrected may be lost. The constraints of the generative adversarial network can be configured to minimize the distance between the image through the deep learning generator (G) and the original image.

As shown in FIG. 1, the image processing method of the present invention using the generative adversarial neural network largely includes an image conversion step (S10), an image discrimination step (S20), and a learning step (S30), each step Of course, the order of each operation can be changed within the range where the execution order can be logically changed as needed, and some steps can be added or deleted.

First, before each step is performed, it is necessary to perform data scaling so that characteristics such as an image noise value can be calculated by a formula, and to make the value ranges of each characteristic similar.

으로 스케일링하였다.Data scaling has various methods such as normalization and standardization. Conventionally, data normalization method is used to

scaled by

는 스케일링된 CT영상을 의미한다.Here, x is a CT image,

denotes a scaled CT image.

However, the present invention proposes a rescaling method suitable for tissue noise removal in the 1st CT image or low-dose CT image so that noise removal can be performed more efficiently.

로 데이터를 스케일링하는 방식을 제안한다. That is, in the present invention, when scaling data

We propose a method for scaling the data with

가 될 수 있도록 하며, 이는 종래 데이터 정규화 방식과 달리 tissue를 0 근방에 위치되게 할 수 있는 이점이 있다. The present invention through such rescaling

, and this has the advantage of being able to position the tissue near 0, unlike the conventional data normalization method.

는 스케일링된 CT영상을 의미하며,

는 평균 HU(Hounsfield Unit)값이며, c는 CT의 최대값의 역수이다.Here, x is a CT image,

denotes a scaled CT image,

is the average Hounsfield Unit (HU) value, and c is the reciprocal of the maximum value of CT.

은, 상기

의 tissue 값이 0이 되도록

로 리스케일링되며, 제2CT영상인 상기

는, 상기

의 tissue의 값이 0이 되도록

로 리스케일링 될 수 있는 것이다.For example, the first CT image

silver, above

so that the tissue value of is 0

It is rescaled to , and the second CT image

, the above

so that the value of tissue in is 0

It can be rescaled with

5A to 5B are data showing performance comparison results of each image processing method, and Group 1 to Group 4 compare performance of images processed under various conditions such as low-dose image, standard-dose image, data normalization learning method, and rescaling learning method, respectively. Show results.

Here, Group 4 is a model using the image processing method and rescaling learning method of the present invention, and Group 3 is a model of a virtual CT image obtained by learning with a conventionally used data normalization method.

That is, in FIG. 5a, the SNR (Signal-to-Noise Ratio) value was 14.8 ± 3.1 in liver, portal, and aorta, respectively, in the case of Group 4 (model using image processing method and rescaling learning method) of the present invention. , 25.7±6.5, 29.4±7.8, and Group3 (model learned using data normalization) was 11.6±2.3, 20.9±4.7, 24.5±6.5 in liver, portal, and aorta, respectively, and CNR (contrast-to-noise ratio, Contrast-to-Noise Ratio) values are 7.8±2.2, 18.8±5.8, 22.4±7.0 for Group 4, and 5.9±1.7, 15.2±4.2, 18.8±5.9 for Group 3, which means that the SNR and CNR values of Group 4 are It can be seen that it is significantly improved compared to the SNR and CNR values of Group3.

From this, it can be seen that the method using the rescaling of the present invention reduces noise more efficiently than the conventional scaling method using data normalization.

Meanwhile, an image processing method as shown in FIG. 1 may be performed using the data rescaled by the above method.

The image processing method includes an image conversion step (S10) using a generative adversarial neural network, an image discrimination step (S20), and a learning step (S30). Of course, it may be changed or some steps may be deleted.

In addition, between each step, a storage step including a storage unit capable of storing the 1st CT image, 2nd CT image, virtual CT image data, and the result values of the generator (G) network and discriminator (D) network can be configured. is of course

The image conversion step (S10) is a step of converting and outputting a virtual CT image by removing noise from the first CT image as the first CT image captured by the CT device is input to the generator (G) network.

At this time, in the generator (G) network, various deep learning algorithms may be used, for example, convolution, Rectified Linear Unit (ReLU), pooling, unplooling, and the like.

In the image discrimination step (S20), the virtual CT image or the 2nd CT image captured by the CT device is input to the network of the discriminator (D) using the virtual CT image generated by the creator (G). As a step of determining whether the virtual CT image and the second CT image are genuine or not and outputting a resultant value, various configurations are possible.

Finally, in the learning step (S30), the generator (G) and the discriminator ( As a step of adjusting and updating the parameters of D), various configurations are possible.

Through each of the above steps, the generator (G) and the discriminator (D) are learned to have optimal performance, and the trained image processing includes the sufficiently learned generator (G) and the discriminator (D). As shown in FIG. 2, the method may be used in an image correction step in an actual clinical environment.

Meanwhile, the learning process of the generator (G) and the discriminator (D) of the present invention will be described in detail below.

In general, a generative adversarial neural network-based network has a generator (G) and discriminator (D) networks, and each network is a generator (G) and a discriminator (G) and discriminator ( D) to learn.

In the present invention, the loss function of the generator (G) network

를 개시한다.

Initiate.

은 제1CT영상이고,

은 생성자가 변환한 가상의 CT영상이며, λ는 조정되는 파라미터이고,

는 가상의 CT영상에 대한 판별자(D)의 판별결과이며,

는 상기 제1CT영상에 대한 확률분포이며, E는 주어진 확률분포에 대한 기댓값을 나타낸다.here

is the first CT image,

is a virtual CT image converted by the creator, λ is a parameter to be adjusted,

is the discrimination result of the discriminator (D) for the virtual CT image,

Is a probability distribution for the first CT image, and E represents an expected value for a given probability distribution.

는, 상기 함수의 손실값을 최소화하는 생성자(G)를 찾는 것을 나타낸다.and,

Indicates finding a generator (G) that minimizes the loss value of the function.

Specifically, the generator (G) function may be composed of two terms, and the generator (G) function obtains the entropy by taking the expected value in order to obtain the average amount of information for the two data from the two terms , and try to obtain an optimized generator (G) by calculating the total amount of information.

에서,

는 판별자(D)가 가상의 CT영상을 보고 판별한 0에서 1 사이의 예측 확률 값으로서, 상기 첫번째 항은

가 1에 가까운 값을 생성하도록 상기 생성자(G)를 학습시키는 부분이다.More specifically, the first term

at,

Is a predicted probability value between 0 and 1 determined by the discriminator (D) by looking at a virtual CT image, and the first term is

This is the part where the generator (G) is trained to generate a value close to 1.

에서,

은, 생성자(G)가 생성한 가상의 CT영상과 생성자(G)에 입력된 제1CT영상의 차이에 관한 것으로서, 상기 두 번째 항은, 딥러닝 생성자(G)를 통한 영상과 원 영상의 차이(distance)를 최소화하여 영상의 형태학적 정보를 최대한 유지하면서 잡음을 제거하기 위한 부분으로 제약조건 부분이다.Also, the second term

at,

relates to the difference between the virtual CT image generated by the creator (G) and the 1st CT image input to the creator (G), and the second term is the difference between the image through the deep learning creator (G) and the original image It is a constraint condition part to remove noise while maintaining the morphological information of the image as much as possible by minimizing the distance.

That is, the generator (G) learns in a direction in which the expected value of the sum of the two terms is minimized, so that the discriminator (D) cannot distinguish it from the 2nd CT image and at the same time maintains the morphological information of the original image as much as possible. It develops into an optimized generator (G) capable of generating a virtual CT image.

Meanwhile, for the optimization of the generator (G), the discriminator (D) is also optimized so that feedback can be provided to the generator (G) network.

In the present invention, the loss function of the discriminator (D) network

의 함수를 개시한다.

Initiate the function of

는 제2CT영상이며,

는 가상의 CT영상에 대한 판별자(D)의 판별결과이고,

는 제2CT영상에 대한 판별자(D)의 판별결과이며,

는 상기 제2CT영상에 대한 확률분포이고,

는 상기 제1CT영상에 대한 확률분포이며, E는 주어진 확률분포에 대한 기댓값을 나타낸다.here

is the second CT image,

Is the discrimination result of the discriminator (D) for the virtual CT image,

is the discrimination result of the discriminator (D) for the second CT image,

Is a probability distribution for the second CT image,

Is a probability distribution for the first CT image, and E represents an expected value for a given probability distribution.

Specifically, the discriminator (D) function may be composed of two terms, and the discriminator (D) function calculates the entropy by taking the expected value in order to obtain the average amount of information for the two data from the two terms. , and try to obtain an optimized discriminator (D) by calculating the total amount of information.

는, 제2CT영상에 대한 판별자(D)의 판별결과에 대한 기댓값을 최대화되도록 판별자(D)를 학습시키는 부분이다.More specifically, the discriminator (D) function may be composed of two terms, the first term

is a part for learning the discriminator (D) to maximize the expected value of the discrimination result of the discriminator (D) for the 2nd CT image.

는, 가상의 CT영상에 대한 판별자(D)의 판별결과에 대한 기댓값을 최대화 되도록 판별자(D)를 학습시키는 부분으로서, 여기서

는 판별자(D)가 가상의 CT영상을 보고 판별한 0에서 1 사이의 예측 확률 값으로,

가 0에 가까운 값을 생성하도록 상기 판별자(D)를 학습시키는 부분이다.Also, the second term

is a part that learns the discriminator (D) to maximize the expected value of the discriminator (D)'s discrimination result for the virtual CT image, where

is a predicted probability value between 0 and 1 determined by the discriminator (D) by looking at a virtual CT image,

This is the part where the discriminator (D) is trained to generate a value close to 0.

That is, the discriminator (D) more accurately discriminates the virtual CT image generated by the generator (G) from the 2nd CT image actually taken by the CT scan through the sum of the two terms of the discriminator (D) function. Discrimination ability can be improved by learning to do it.

Here, the generator (G) or discriminator (D) repeats the parameters of the generator (G) or discriminator (D) according to the result values output from the generator (G) and the discriminator (D). By adjusting, it goes through an optimization process.

In summary, a loss is calculated according to the authenticity output from the loss function of the discriminator (D) and transmitted to the loss function of the generator (G) to propagate the error, and through the error, the 2nd CT image and the The generator (G) and the discriminator (D) are trained by updating the weights or parameters of the neural network of the generator (G) or discriminator (D) to minimize the difference between the virtual CT images.

That is, the generator (G) and the discriminator (D) are learned while updating to minimize the loss, and these updates are repeatedly performed, but the discriminator (D) is the virtual value generated by the generator (G). It is preferable to perform iteratively until the CT image and the second CT image cannot be distinguished.

Meanwhile, in the present invention, tanh(kx) may be placed in the last layer of the generator (G) network so that denoising learning of the generator (G) can be performed more effectively.

가,

의 함수로 설정되게 하여 상기 생성자(G)를 통하여 생성된 영상

도 같은 범위에 있게 할 수 있다.That is, the above

go,

An image generated through the generator (G) by being set as a function of

may also be in the same range.

는 함수의 합성을 나타내며, 상기

는 k번째 레이어를 나타내고, 상기 k는 조정되는 파라미터, 즉 상수이며, k>0 이다.here, above

Represents the composition of the function, wherein

denotes a k-th layer, where k is a parameter to be adjusted, that is, a constant, and k>0.

와 같은 네트워크 구조를 이용하는 경우, 상수 k를 증가시킬수록 tanh(kx)의 그래프가 0 근방에서 기울기가 증가 될 수 있다.Specifically,

In the case of using a network structure such as , as the constant k is increased, the graph of tanh(kx) may have an increased slope near 0.

값이

로 스케일링 되므로, 잡음을 제거하고자 하는 tissue의

값이 0 근방에 위치되고, 상기

네트워크에서 상수 k를 증가시켜 0 근방의 기울기를 증가시킬 수 있으므로, 생성자(G)에 대한 잡음제거 훈련의 효율성을 향상시킬 수 있는 것이다. At this time, through the above-described rescaling

value

Since it is scaled by , the number of tissues to remove noise

value is located around 0, and the

Since the gradient near 0 can be increased by increasing the constant k in the network, the efficiency of denoising training for the generator (G) can be improved.

This effect can be seen more clearly through the graph of FIG. 4 .

)으로 스케일링 되어진

가 tanh(x)를 이용한 네트워크를 통해 생성된 tissue이며, tanh(3x)위의 빨간색 점들은

가 본 발명에 따른 리스케일링방식(즉,

)을 통해 0근방으로 스케일링되되, tanh(3x)를 이용한 네트워크를 통해 생성된 tissue이다.The graph of Figure 4 is a graph showing tissue values converted according to the network structure of the present invention, and the blue dots on tanh (x) are the prior art data normalization method (ie,

) scaled by

is a tissue generated through a network using tanh(x), and the red dots on tanh(3x) are

Is the rescaling method according to the present invention (ie,

), it is scaled to around 0, but it is a tissue created through a network using tanh (3x).

값이 0 근방에 위치되는 동시에, tanh(x) 그래프의 기울기보다 tanh(3x) 그래프의 기울기의 변화량이 크므로, 변환된 tissue값의 변화량이 커져 잡음제거훈련이 효율적으로 수행됨을 알 수 있다.That is, comparing the graph of tanh(3x) with the graph of tanh(x), in tanh(3x)

Since the value is located near 0 and the change in the slope of the tanh(3x) graph is greater than the slope of the tanh(x) graph, the change in the converted tissue value increases, which means that noise removal training is performed efficiently.

Here, in the generator (G) network, various deep learning algorithms can be used, for example, convolution, Rectified Linear Unit (ReLU), pooling, unplooling, etc. can be used, of course.

On the other hand, in order to show the effect of the image processing method of the present invention more clearly, reference is made to a plurality of corrected image drawings obtained using different image processing methods of FIG. 5B.

Here, Group1 is a conventional low-dose image, Group2 is an image using an image with increased tube voltage as the second CT image of the present invention, and Group3 shows an image obtained by a conventional training method using data normalization.

That is, the image of Group 1 is a low-dose image, and the brightness level is high, but there is a problem with a lot of noise. There is a problem, and Group 3 has a problem in that noise is not effectively reduced by correcting using a conventional data normalization method.

However, the image of Group 4 is a corrected image obtained by the image processing method learned using the rescaling and improved function network of the present invention, and the image noise of Group 1 is effectively reduced, and the brightness level is higher than that of the image of Group 2. is maintained, and it can be confirmed that the noise is further reduced compared to the image of Group3.

That is, the image processing method learned using the rescaling and improved generator (G) and discriminator (D) network of the present invention can remove noise more effectively while maintaining the same brightness level as a low-dose image.

Meanwhile, the deep learning network learned through each step described above can be applied to existing CT equipment.

That is, as shown in FIG. 2, by converting the low-dose CT image obtained by exposing the patient to a smaller radiation dose through the learned image processing method, a virtual standard-dose image from which noise is removed while maintaining the brightness level is output, A user of a CT device can more easily read a CT image.

In addition, the present invention discloses an image processing device including a memory and a processor for storing the above-described programs, wherein the image processing method is performed when the programs are executed by the processor. can

The processor is not limited to its type, and may be, for example, a general-purpose processor, a controller, a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), and the like.

In addition, each step of the image processing method may be performed by a computing device, and each step of the method may be implemented as one or more instructions executed by a processor of the computing device.

Of course, all steps included in the method may be distributed and executed by one physical computing device or a plurality of physical computing devices.

In addition, the computing device may further provide a computer readable storage medium for storing computer programs.

The memory and the computer readable storage medium are not limited in their types, and may be, for example, RAM, ROPM, PROM, EPROM, magnetic disk, tape, or non-transitory storage medium.

Since the above has only been described with respect to some of the preferred embodiments that can be implemented by the present invention, as noted, the scope of the present invention should not be construed as being limited to the above embodiments, and the scope of the present invention described above It will be said that the technical idea and the technical idea accompanying the root are all included in the scope of the present invention.

S10: Image conversion step S20: Image discrimination step
S30: learning step

Claims (10)

As an image processing method based on generative adversarial networks (GAN),
An image conversion step (S10) of converting and outputting a virtual CT image by removing noise from the first CT image as the first CT image captured by the CT device is input to the generator (G);
An image discrimination step (S20) of determining authenticity as the virtual CT image or the 2nd CT image captured by the CT device is input to the discriminator (D);
And a learning step (S30) of adjusting parameters of the generator (G) or the discriminator (D) according to the authenticity output from the discriminator (D),
The constructor (G),

has a network of
here, above

is the first CT image, and

is the virtual CT image converted from the 1st CT image by the creator (G), λ is a parameter to be adjusted,

Is the discrimination result of the discriminator (D) for the virtual CT image,

Represents a probability distribution for the 1st CT image, E represents an expected value for a given probability distribution,
remind

Is,
remind

so that the tissue value of is 0

is rescaled to
here, above

is the rescaled 1 CT image,

Is an average Hounsfield Unit (HU) value, and c is an image processing method characterized in that the reciprocal of the maximum value of CT. The method of claim 1,
The image processing method, characterized in that the first CT image and the second CT image do not match. The method of claim 1,
A tube current (TC2) in the CT scan of the second CT image is greater than a tube current (TC1) in the CT scan of the first CT image. The method of claim 3,
The tube voltage (TV2) in the CT scan of the second CT image is the same as the tube voltage (TV1) in the CT scan of the first CT image. delete delete The method of claim 1,
The generator (G) is the rescaled 1CT image

virtual CT image from

convert,
remind

Is,

With a function of
here, above

Represents the composition of the function, wherein

represents a k-th layer, k is a parameter to be adjusted, and k>0. The method of claim 1,
The discriminator (D),

has a function of
here, above

is the second CT image, and

Is the discrimination result of the discriminator (D) for the virtual CT image,

Is the discrimination result of the discriminator (D) for the 2nd CT image,

Represents a probability distribution for the second CT image,

is a probability distribution for the first CT image, and E is an image processing method characterized in that it represents an expected value for a given probability distribution. The method of claim 8,
remind

Is,
remind

so that the value of tissue in is 0

is rescaled to
here, above

is the rescaled 2 CT image,

Is an average Hounsfield Unit (HU) value, and c is an image processing method characterized in that the reciprocal of the maximum value of CT. An image processing apparatus including a memory and a processor for storing a program,
An image processing device characterized in that performing the image processing method of any one of claims 1 to 4 and claims 7 to 9 when the programs are executed by the processor.

Images