KR102182048B1 - System and method for scatter correction of x-ray image - Google Patents

Abstract

An X-ray image scattering correction system according to an embodiment of the present invention includes: a training image generator configured to generate a simulation image with scattering lines and a target image without scattering lines through Monte Carlo simulation using a 3D volume image; And a scattering correction unit for correcting scattering of an X-ray image captured by an X-ray image acquisition apparatus using a deep learning algorithm trained using the simulation image and the target image.

Description

X-ray image scattering correction system and method {SYSTEM AND METHOD FOR SCATTER CORRECTION OF X-RAY IMAGE}

The present application relates to an X-ray image scattering correction system and method.

The X-ray imaging system is a technology for imaging by measuring the attenuation of photons passing through an object to be photographed by a detector, and is usefully used to diagnose patient lesions in the medical field. In the X-ray imaging system, the longer the X-rays pass through the object to be photographed, the higher the probability that photons cannot be transmitted straight ahead and scattered. Since the meaningful information here is the primary line (photons in a straight path) attenuated by the subject, scattered lines (photons out of the straight path, secondary lines) generated inside the subject are serious noise factors that deteriorate the quality of the X-ray image. Works. When a large number of photons are scattered, the contrast and sharpness of the image decrease, but the more photons are used, the better the image can be obtained. However, a problem arises that the more photons are used, the more the exposure dose to the subject is increased. For this reason, in the medical X-ray imaging system, the removal of scattered rays to improve the image quality without increasing the patient's exposure dose is recognized as a very important problem, and until now, research to remove or correct scattered rays has been continuously conducted. Has become.

The scattering correction methods proposed so far can be largely divided into a physical scatter removal/correction method that solves a problem using the physical characteristics of the equipment, and a computational scatter correction method that extracts and removes a scattering signal from the acquired signal.

Specifically, physical scattering removal/correction methods include a method of removing scattered rays by placing an anti-scatter grid between the imaging area and an X-ray detector, and a measurement-based scattering correction method that directly measures and corrects scattering components. The computational scattering correction method includes a convolution method that performs de-convolution by estimating a scattering kernel, and a Monte Carlo-based iterative scatter correction method that repeatedly corrects the scattering using Monte Carlo simulation. have.

However, the conventional physical scattering removal/correction method requires adding a physical device to the existing system, and since a certain amount of X-rays is blocked to estimate or remove scattering, increase the exposure dose or increase the amount of exposure to compensate for the blocked dose. There is a drawback of having to consider the deterioration of image quality.

In addition, in the case of the conventional computational scattering correction method, the convolution method guarantees performance when the subject is made of the same material, but has a disadvantage that performance is greatly degraded when the subject is made of various materials, such as in the case of the human body. The repetitive scattering correction method has disadvantages that it has a processing speed that cannot be used clinically due to a large amount of computation, and that it cannot be applied to a 2D imaging system.

Therefore, in the art, there is a need for a method for improving scattering correction performance of an X-ray image while overcoming the disadvantages of conventional scattering correction methods.

In order to solve the above problems, an embodiment of the present invention provides an X-ray image scattering correction system.

The X-ray image scattering correction system includes: a training image generator configured to generate a simulation image with scattered rays and a target image without scattering rays through Monte Carlo simulation using a 3D volume image; And a scattering correction unit for correcting scattering of an X-ray image captured by an X-ray image acquisition apparatus using a deep learning algorithm trained using the simulation image and the target image.

In addition, another embodiment of the present invention provides a method for correcting scattering of an X-ray image.

The X-ray image scattering correction method includes: generating a training image including a simulation image with scattering rays and a target image with no scattering rays through Monte Carlo simulation using a 3D volume image; And correcting the scattering of the input X-ray image using a deep learning algorithm trained using the simulation image and the target image.

In addition, the solution to the above-described problem does not enumerate all the features of the present invention. Various features of the present invention and advantages and effects thereof may be understood in more detail with reference to the following specific embodiments.

According to an embodiment of the present invention, it is possible to greatly improve scattering correction performance of an X-ray image while overcoming disadvantages of conventional scattering correction methods.

1 is a block diagram of an X-ray image scattering correction system according to an embodiment of the present invention.
2 is a diagram illustrating an example of generating a training image using Monte Carlo simulation according to an embodiment of the present invention.
3 is a diagram for describing a method of generating a target image for training a deep learning algorithm according to an embodiment of the present invention.
4 is a diagram illustrating a deep learning model used in a scattering correction unit according to an embodiment of the present invention.
FIG. 5 is a diagram illustrating a process of designing the deep learning model shown in FIG. 4 and performing training for each frequency component in the deep learning model.
6 is a graph showing the dilation rate of Ls-Net shown in FIG. 4.
7 is a graph showing the cumulative receiving area size of Ls-Net shown in FIG. 4.
8 is a diagram illustrating a simulation diagram of an X-ray image using a GATE code according to an embodiment of the present invention.
9 is a diagram illustrating an image for network training among 3D human body CT images used for X-ray image simulation according to an embodiment of the present invention.
10 is a diagram illustrating an image for a network test among 3D human body CT images used in an X-ray image simulation according to an embodiment of the present invention.
11 is a diagram illustrating an X-ray image simulation result according to an embodiment of the present invention.
12 is a graph showing SSIM and RMSE of a test image and loss according to network training according to an embodiment of the present invention.
13 is a diagram illustrating a result of predicting an X-ray scattering component according to an embodiment of the present invention.
14 is a diagram showing a scattering correction result by a Monte Carlo-based iterative scattering correction method according to the prior art.
15 is a graph showing RMSE according to the number of repetitions of a Monte Carlo-based iterative scattering correction method according to the prior art.
16 is a diagram showing a scattering correction result according to an embodiment of the present invention.
17 is a diagram illustrating a 3D reconstruction result and profile of a scattering corrected image by Ls-Net and Hs-Net according to an embodiment of the present invention.
18 is a diagram showing a 3D reconstruction result and profile of a scattering corrected image according to the prior art and an embodiment of the present invention.
19 is a flowchart of a method for correcting scattering of an X-ray image according to another embodiment of the present invention.

Hereinafter, preferred embodiments will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art may easily implement the present invention. However, in describing a preferred embodiment of the present invention in detail, if it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted. In addition, the same reference numerals are used throughout the drawings for portions having similar functions and functions.

In addition, throughout the specification, when a part is said to be'connected' to another part, it is not only'directly connected', but also'indirectly connected' with another element in the middle. Include. In addition, "including" a certain component means that other components may be further included rather than excluding other components unless specifically stated to the contrary.

1 is a block diagram of an X-ray image scattering correction system according to an embodiment of the present invention.

Referring to FIG. 1, an X-ray image scattering correction system 200 according to an embodiment of the present invention may include a training image generation unit 210 and a scattering correction unit 220.

The training image generator 210 is for generating a simulation image and a target image for training a deep learning algorithm used in the scattering correction unit 220 to be described later.

In order for the scattering correction unit 220 to correct scattering of an X-ray image using a deep learning algorithm, an input image with scattering lines and a target image without scattering lines are required under the same conditions to train a deep learning algorithm. . In addition, a large amount of training images are required to perform X-ray image scattering correction using a deep learning algorithm. However, due to the nature of the X-ray image, it is difficult to acquire an image in which scattered rays do not exist, and it is difficult to acquire a large amount of training images with actual equipment.

Therefore, according to an embodiment of the present invention, the training image generation unit 210 may generate a simulation image with scattering lines and a target image without scattering lines through Monte Carlo simulation using a 3D volume image. have. Here, the 3D volume image may be, for example, a 3D volume image obtained by equipment such as MDCT (Multidetector computed tomography) or a volume image generated by a 3D design tool and having a density value of an actual object.

2 is a diagram illustrating an example of generating a training image using Monte Carlo simulation according to an embodiment of the present invention.

The training image generator 210 acquires a simulation image 23 in which the scattered rays exist through the Monte Carlo simulation 22 using the 3D volume image 21, and the main X-ray components and scattering from the simulation image 23 are obtained. By separating the ray component, the main X-ray image 24 and the scattering ray image 25 may be obtained, respectively. Here, the simulation image 23 in which the scattered ray exists is an input image for training the deep learning algorithm, and the main X-ray image 24 in which the scattering ray does not exist is the target image.

In this case, by setting the X-ray dose of the target image higher than the X-ray dose of the simulation image by a preset ratio (for example, about 20%), a system that is robust against image noise due to low dose can be configured.

3 is a diagram for describing a method of generating a target image for training a deep learning algorithm according to an embodiment of the present invention.

As will be described later, in order to generate a target image for training two networks of Ds-DRCNN (Dual Scale Learing of Dilated Residual CNN) in the scattering correction unit 220, the training image generation unit 210 By applying a Gaussian Blur to the existing scattering target image 31, a low-frequency target image 32 including a low-frequency component is obtained, and the low-frequency target image 32 is subtracted from the scattering target image 31. A high-frequency target image 33 including a high-frequency component may be obtained. Here, the low-frequency component is a scattering component generated by surrounding pixels in a wide area, and the high-frequency component may be a noise and scattering component mainly generated by surrounding pixels in a non-wide area.

The above-described training image generator 210 needs to acquire a large amount of simulation images using a 3D volume image having a relatively large capacity, so that the simulation speed is important. According to an embodiment, the training image generator 210 may use a GATE that supports GPU calculation among simulation tools, but is not limited thereto.

The scattering correction unit 220 is an X-ray image captured by the X-ray image acquisition apparatus 100 using a training image transmitted from the training image generation unit 210, that is, a deep learning algorithm trained using a simulation image and a target image. It is to correct the scattering of.

Here, the X-ray image acquisition apparatus 100 may be a device that acquires an X-ray image of a subject to be photographed, such as a 2D DR device or a 3D CT device, and the X-ray image obtained thereby includes scattered rays.

Even when X-rays pass through an object with the same component and thickness, the distribution of the scattered rays appears different depending on the location and distance at which the scattering occurs.Thus, a deep learning algorithm mainly used for image noise removal (e.g., CNN (Convolution Neural Network)) is difficult to apply as it is.

Accordingly, according to an embodiment of the present invention, the scattering correction unit 220 separates scattering and noise of an X-ray image into a high-frequency component and a low-frequency component, and independently constructs and applies two networks for each component. It adopts DRCNN (Dual Scale Learing of Dilated Residual CNN).

4 is a diagram illustrating a deep learning model used in a scattering correction unit according to an embodiment of the present invention.

Referring to FIG. 4, an input image of Ds-DRCNN is an X-ray image 41 including scattering rays, and an output image is a main X-ray image 45 in which scattering rays are corrected and no scattering rays exist.

In addition, Ds-DRCNN is Hs-Net (Residual CNN for Estimation of High Frequency Scatter) (42) for correction of high-frequency components among scattering components and Ls-Net (Dilated Residual CNN for Estimation of Low Frequency) (Ls-Net) for correction of low-frequency components. Scatter)(43), and when the outputs of the two networks are summed, the scattering ray estimation image 44 is obtained by estimating the scattering ray component, and the scattering ray estimation image 44 is subtracted from the input X-ray image 41 to scatter. It is configured to obtain a corrected X-ray image 45 in which no ray exists.

Here, the Hs-Net 42 may be designed to enable effective parameter training within a narrow receiving area, and the Ls-Net 43 may be designed to take a wide receiving area without significantly increasing the number of parameters. .

FIG. 5 is a diagram illustrating a process of designing the deep learning model shown in FIG. 4 and performing training for each frequency component in the deep learning model.

Referring to FIG. 5, Hs-Net and Ls-Net each have a simulation image 51 including a scattering line obtained by the training image generator 210 as an input image, and a high frequency including a scattering line high frequency component. The target image 53 and the low-frequency target image 55 including the scattering line low-frequency component are each target image, and the scattering line component of the corresponding frequency component is trained through a plurality of layers 52 and 54 in each network. I can make it.

In addition, by applying residual learning to each network, it is possible to effectively perform training even in a deep CNN.

In a general deep learning model, as multilayers are stacked, the gradient vanishing problem is intensified as the weight product of 0 and 1 approaches 0, making it difficult to properly learn. The residual learning of CNN is to solve this problem.

In the case of scattering of high-frequency components, it has local characteristics and contains noise. Therefore, it is preferable that the Hs-Net 42 for correcting the high-frequency component is configured to extract a combination of various filters by deepening the network layer rather than increasing the size of the receiving region.

In an embodiment of the present invention, the Hs-Net 42 may be configured to modify and apply the noise reduction model DnCNN.

3으로 설정하고 모든 풀링 레이어들(pooling layers)을 제거한다. 따라서, DnCNN의 수용영역 크기는 네트워크 깊이가 d일 때 (2d+1)

(2d+1)이 된다.For network architecture design, DnCNN can modify the VGG network for image noise reduction and set the depth of the network based on the effective patch size used in recent noise reduction methods. For model training, we can adopt a residual learning strategy and apply batch regularization for fast regular training and improved noise reduction performance. Filter size 3 in VGG network

Set to 3 and remove all pooling layers. Therefore, the receiving area size of DnCNN is (2d+1) when the network depth is d

It becomes (2d+1).

There are three types of layers in DnCNN with network depth d as follows.

3

c 크기의 필터 64개가 사용된다. 활성화 함수 ReLU(Rectified Linear Units)를 적용하여 비선형성을 추가한다. 여기서, c는 영상의 채널을 의미하는데, 예를 들어 그레이 영상일 경우 c=1이고, 컬러 영상일 경우 c=3이다. (1) Conv+ReLU is applied to the first layer, and 3 to generate 64 feature maps.

3

64 filters of size c are used. Nonlinearity is added by applying the activation function ReLU (Rectified Linear Units). Here, c denotes a channel of an image. For example, c = 1 for a gray image and c = 3 for a color image.

3

64 크기의 필터 64개가 사용된다. 그리고 배치 정규화(Batch Normalization, BN)가 합성곱(Convolution)과 ReLU 사이에 추가된다. (2) Conv+BN+ReLU is applied from the second layer to the (d-1) layer, and 3

3

64 filters of size 64 are used. And batch normalization (BN) is added between convolution and ReLU.

3

64 크기의 필터 c개가 사용된다. 또한 중간 레이어의 각 feature map이 입력 영상과 동일한 크기를 가지도록 각 레이어의 합성곱 전에 제로패딩(zero padding)을 적용한다.(3) Conv is applied to the last layer, and 3

3

C filters of size 64 are used. In addition, zero padding is applied before convolution of each layer so that each feature map of the intermediate layer has the same size as the input image.

As described above, the Hs-Net 42 according to the present embodiment configures a network using the structure of DnCNN, but the network depth may be set equal to that of the Ls-Net 43 to be described later. Through this, it is possible to increase the computational efficiency of networks arranged in parallel by maintaining the same computational amount of the two networks.

On the other hand, in the case of scattering of low-frequency components, since the scattering components generated in the object to be photographed can be detected in all regions of the image, Ls-Net (43) for correction of low-frequency components expands the size of the receptive region of the network to the entire image You should be able to.

In one embodiment of the present invention, an Ls-Net 43 capable of extending the size of the receiving area without increasing the amount of computation without increasing the parameters and network depth of the Hs-Net 42 is proposed.

3으로 동일하게 설정하지만, dilated filter를 적용함으로써 네트워크 깊이를 늘리지 않고 수용영역의 확장이 가능하다. 여기서, dilation rate는 수용영역의 크기가 입력영상의 크기에 대응될 수 있도록 입력영상의 크기에 따라 상이하게 설정할 수 있다. 입력영상의 전체 크기로 수용영역의 크기를 확장하기 위해 네트워크를 다음과 같이 설정할 수 있다.Specifically, the Ls-Net 43 may be a model in which a dilated filter is applied to the network structure of the Hs-Net 42 described above in order to expand the size of the receiving area. Ls-Net(43) has the same structure as Hs-Net(42) and has a filter size of 3

The same setting as 3, but by applying a dilated filter, it is possible to expand the receiving area without increasing the network depth. Here, the dilation rate may be set differently according to the size of the input image so that the size of the receiving area may correspond to the size of the input image. In order to expand the size of the receiving area to the total size of the input image, the network can be set as follows.

k이고 dilation rate가 r일 때, 수용영역 크기는 {r*(k-1)+1}

{r*(k-1)+1}이 된다. 깊이가 d인 Ls-Net(43)의 n번째 레이어의 dilation rate r_n은 하기의 수학식 1과 같이 정의할 수 있다. 여기서, n번째 레이어까지의 누적 수용영역의 크기를 S_n

S_n이라고 하면 여기서 S_n은 하기의 수학식 2와 같다. The filter size of the network is k

When k and dilation rate r, the receiving area size is {r*(k-1)+1}

It becomes {r*(k-1)+1}. The dilation rate r _n of the n-th layer of the Ls-Net 43 having a depth of d may be defined as Equation 1 below. Here, the size of the accumulated receiving area up to the nth layer is S _n

Speaking of S _n , where S _n is the same as in Equation 2 below.

[Equation 1]

[Equation 2]

The reason for defining the dilation rate as in Equation 1 is to remove gridding artifacts that appear when a dilated filter is applied by making the dilation rate have a symmetrical structure according to the depth of the network. For a symmetrical structure, the network depth d can be set to a multiple of 2.

6 is a graph showing the dilation rate of Ls-Net shown in FIG. 4, and FIG. 7 is a graph showing the cumulative receiving area size of Ls-Net shown in FIG. 4.

61이지만, 네트워크 깊이를 16으로 늘릴 경우 수용영역의 크기는 1021

1021로 크게 확장될 수 있다.As described above, the Ls-Net 43 according to an embodiment of the present invention has excellent scalability because it can increase the size of the receiving area without greatly increasing the network depth. For example, if the network depth is 8, the size of the receiving area is 61

61, but if the network depth is increased to 16, the size of the receiving area is 1021

It can be greatly extended to 1021.

128이므로, 이에 가까운 수용영역을 가질 수 있도록 네트워크 깊이를 12로 설정할 수 있으며, 이 경우 Ds-DRCNN는 하기의 표 1과 같은 세부 구조를 가질 수 있다.The network depth of the Ds-DRCNN described above may be set in consideration of the size of the receiving area of the Ls-Net 43 according to the size of the input image. The size of the input image used in the experiment described below is 256

Since it is 128, the network depth can be set to 12 to have an accommodation area close to this, and in this case, Ds-DRCNN can have a detailed structure as shown in Table 1 below.

[Table 1]

를 채택하여 산란이 보정된 주 엑스선 예측영상 p'는

를 통해 얻을 수 있다. 네트워크 파라미터

는 하기의 수학식 3에 나타낸 손실함수가 최소화되도록 훈련시킬 수 있다.The input image y of the Ds-DRCNN described above is an image obtained by adding the scattering image s to the main X-ray image p, and is y=p+s. Here, the input image y may be an image obtained by normalizing the maximum value to 1. Also, residual mapping for residual learning training

The main X-ray prediction image p'for which scattering is corrected by adopting

You can get it through. Network parameters

Can be trained to minimize the loss function shown in Equation 3 below.

[Equation 3]

및

은 각각 Hs-Net 및 Ls-Net의 손실함수로 하기의 수학식 4 및 수학식 5와 같고,

는 각 손실함수의 중요도를 설정하는 가중치이다.here,

And

Is the same as Equation 4 and Equation 5 below as loss functions of Hs-Net and Ls-Net, respectively,

Is a weight that sets the importance of each loss function.

[Equation 4]

[Equation 5]

는 Hs-Net의 파라미터

를 통해 예측된 고주파 산란 영상

이고,

은 Ls-Net의 파라미터

을 통해 예측된 저주파 산란 영상

이다. 여기서 훈련 집합

은 산란 포함 입력 영상 y_i와 산란 영상 s_i로 몬테카를로 시뮬레이션을 통해 획득한 N개의 영상 쌍으로 이루어질 수 있다.The low-frequency scattering target image l _i is the scattering image s _i with a Gaussian filter G, l _i =G(s _i ), and the high-frequency scattering target image h _i is the scattering image s _i minus the low-frequency scattering image l _i The image is h _i = s _i -l _i . Also, when the input image is y _i ,

Is the parameter of Hs-Net

High-frequency scattered image predicted through

ego,

Is the parameter of Ls-Net

Low-frequency scattering image predicted through

to be. Training set here

May be composed of N image pairs obtained through Monte Carlo simulation with the scattering input image y _i and the scattering image s _i .

를 통해 산출할 수 있다.The Adam optimizer is used to minimize the loss function during training, and the final scattering image of the input image y _i after training is completed is

It can be calculated through

Next, experiments and results for validating the above-described X-ray image scattering correction system will be described in detail with reference to FIGS. 8 to 18.

To train and verify the deep learning network model of the above-described X-ray image scattering correction system, a training image was generated through Monte Carlo simulation. By setting the image resolution of the X-ray image acquisition device to a low setting so that the image can be simulated at a high speed, it is possible to acquire as many simulation images as possible within the effective time. The size of the X-ray acquisition device is set as large as the actual system size. It was made to be able to check the effectiveness of the invention. Detailed specifications of the X-ray image acquisition device for Monte Carlo simulation are shown in Table 2 below.

[Table 2]

The X-ray image acquisition apparatus shown in Table 2 was implemented with a GATE code to simulate an X-ray image, and FIG. 8 is a diagram illustrating an X-ray image simulation diagram using a GATE code according to an embodiment of the present invention.

The simulation target of the X-ray image used a 3D CT image of a real patient in a medical MDCT system where scattering rarely occurs compared to CBCT in order to verify the effectiveness of the anatomical diversity of complex human structures. The 3D human body image used is a CT image of an actual patient's chest provided by the National Biomedical Imaging Archive (NBIA), and 10 3D volume images of a total of 10 patients. Among them, the simulation images of nine volumes shown in FIG. 9 were used for network training, and the simulation images of the remaining one volume shown in FIG. 10 were used for network testing and verification.

For each of the 10 volume images, the gantry structure including the X-ray generator and the detector was rotated 360 degrees in the CBCT system, and the two-dimensional projection images were simulated for each angle (with or without scattering rays). The number of images was expanded by 4 times by adding the vertically inverted image and the 180 degree rotation image. The number of simulated images thus acquired is 14,400 pairs (=28,800[sheets]=10[volume]*360[1sheet/degree*360degrees]*2[with or without scattering lines] *4[original + left and right flip + top and bottom Reverse + 180 degrees rotation]).

At this time, the simulation image, which is an image including scattering, simulates 163,840,000 photons per projection image so that an average of 5,000 photons can be obtained for each pixel, and the target image of the corresponding network is for each pixel 20% more than this. 196,608,000 photons per projection image were simulated so that 6,000 photons could be acquired. In this way, the reason why the X-ray dose of the target image is set higher than that of the simulation image when generating the simulation image is to build a system that is robust against image noise due to low dose.

11 is a diagram illustrating an X-ray image simulation result according to an embodiment of the present invention, and shows two pairs of Monte Carlo simulation results. As shown in FIG. 11, each simulation image may be divided into a main X-ray image including only a main X-ray component and a scattering image including only a scattering component.

In this experiment, the deep learning network was implemented and trained through TensorFlow, and parameters applied during training are as described later. First, the learning rate of the network was initially set to 0.001, and this value was decreased as the learning progressed so that it was fixed at 0.000001 after the 100th generation. The weights of the convolution kernel were initialized using a known technique. In addition, the batch size was set to 5, and the number of repetitions for each generation was set to 2592, so that all 12960 pairs of simulation images per generation were trained. In addition, network training was performed up to a total of 500 generations, and the training order of images was set randomly for each generation.

12 is a graph showing a Structural Similarity Index Measure (SSIM) and Root Mean Square Error (RMSE) of a test image and loss according to network training according to an embodiment of the present invention.

의 감소 추세에 따라 SSIM이 높아지며, Ls-Net의 손실 값인

의 감소 추세에 따라 RMSE가 낮아짐을 확인할 수 있다. 이는 Hs-Net이 고주파 성분의 보정을 통해 구조적인 유사도를 증가 시키고 Ls-Net이 저주파 성분의 보정을 통해 영상의 전체적인 유사도를 증가시키고 있음을 보여준다. 12, the loss value of Hs-Net

SSIM increases according to the decreasing trend of Ls-Net,

It can be seen that RMSE decreases according to the decreasing trend of. This shows that Hs-Net increases structural similarity through correction of high-frequency components and Ls-Net increases overall similarity of image through correction of low-frequency components.

13 is a diagram illustrating a result of predicting an X-ray scattering component according to an embodiment of the present invention.

Referring to FIG. 13, it can be seen that the network proposed in the present invention effectively predicts the scattering component present in the input image.

Since X-ray images are mainly used for diagnosis, even if the similarity between the target image and the result image increases after image improvement, it may be a big problem if information necessary for diagnosis is lost. In other words, in the case of an X-ray image, it is difficult to verify the performance through a simple similarity comparison with the result of the image enhancement algorithm.

In particular, in the case of the scattering correction method proposed in the present invention, since the scattering component of the entire image including the noise component is not simply removed, it is difficult to evaluate the performance by comparing the similarity of the image before and after scattering correction. it's difficult.

Therefore, in this experiment, we adopted a noise-sensitive 3D CBCT system as an experiment object and evaluated the performance of the 2D projection image before and after scattering correction, as well as the 3D reconstructed image reconstructed from the corrected 2D image. .

The evaluation index used for the quantitative quality evaluation of the scattering correction result of the network proposed in the present invention is Root Mean Square Error (RMSE), Peak Signal-to-noise ratio (PSNR), and structural similarity. Three indexes (Structural Similarity Index Measure, SSIM) were used to evaluate the similarity between the 2D target image and the predicted 2D image, and the similarity between the 3D target image and the 3D reconstructed image of the predicted image, respectively.

For objective performance evaluation, the scattering correction result using the Monte Carlo-based iterative scattering correction method, which is considered to be the most rational and accurate method among the conventional scattering correction methods, and the scattering correction result according to an embodiment of the present invention were compared and evaluated.

Detailed steps of the Monte Carlo-based iterative scattering correction method used in the evaluation are as follows.

Step 0: Reconstruct the original projection image acquired in the CBCT system into 3D without scattering correction

Step 1: Convert the reconstructed 3D volume image to voxel-specific density values

Step 2: Generate main X-ray images and scattering images for each projection angle for the density phantom transformed in step 1 through Monte Carlo simulation

Step 3: Scatter correction of the original projection image with the scattering component generated in step 2

Step 4: Reconstruct the projected image with 3 levels of scattering corrected into 3D

Step 5: Repeat from step 1 until convergence

128로 설정하였고, 각 픽셀마다 평균적으로 1,000개의 광자가 획득될 수 있도록 투영 영상 한 장당 32,768,000개의 광자를 시뮬레이션 하였다. 5단계의 수렴 조건은 이전 반복시의 3차원 재구성 영상과 현재 반복 단계의 3차원 재구성 영상의 가운데 단면 영상간의 RMSE값을 이용하였으며, 이 값이 0.001(cm-1) 이하일 경우 수렴된 것으로 판단하였다.For comparison with the scattering correction method proposed in the present invention, the test image with scattering was used as the original projection image of step 0 of the Monte Carlo-based iterative scattering correction method, and the image size was the same as the input image during the Monte Carlo simulation of the second step. 256

It was set to 128, and 32,768,000 photons per projection image were simulated so that an average of 1,000 photons for each pixel could be obtained. The convergence condition of step 5 used the RMSE value between the 3D reconstructed image at the time of the previous iteration and the middle section image of the 3D reconstructed image at the current iteration step, and if this value is less than 0.001(cm-1), it was determined that convergence was achieved. .

As a result of correcting the scattering by the Monte Carlo-based iterative scattering correction method using the same input image used in the evaluation of the method proposed in the present invention, the result is converged at 5 repetitions, and the result is shown in FIG. 14. Here, the RMSE value for each repetition count, which is a convergence condition of the algorithm, is shown in FIG. 15.

16 is a diagram showing a scattering correction result according to an embodiment of the present invention.

Referring to FIG. 16, it can be seen that the scattering component present in the input image is effectively removed as a result of performing scattering correction by the method proposed in the present invention.

The quantitative evaluation results of the scattering correction results according to the present invention and the prior art are shown in Table 3, and it can be seen that if the scattering is corrected by the method proposed in the present invention, the similarity to the target image is greatly improved in all measurement items. In particular, the SSIM item was improved from 0.77 to 0.992, indicating that the scattering corrected image was corrected very similar to the target image.

[Table 3]

The above result was a comparison of the similarity between the 2D images before and after scattering correction, and it was confirmed that there was excellent improvement in all of the measurement items, but as described above, simply comparing the similarity between the 2D projection images was useful for the scatter removal result. It is difficult to verify. This is because the usefulness of the scattering correction method is deteriorated if artifacts that do not exist in the target image are additionally generated in the 3D reconstructed image of the image after scattering correction or information that existed in the target image disappears.

Therefore, in this experiment, evaluation was also performed on the 3D reconstructed image of the scattering corrected image.

17 is a diagram illustrating a 3D reconstruction result and profile of a scattering corrected image by Ls-Net and Hs-Net according to an embodiment of the present invention.

Referring to FIG. 17, as a result of scattering correction using only Ls-Net, as a result of correcting the scattering component of the low-frequency component, the overall contrast of the image was improved, and as a result of scattering correction using only Hs-Net, it was confirmed that the scattering component and noise of the high-frequency component were corrected. I can. Finally, from the result of Ds-DRCNN integrating Ls-Net and Hs-Net, it can be seen that the scattering was corrected and the contrast and noise of the image were greatly improved.

18 is a diagram showing a 3D reconstruction result and profile of a scattering corrected image according to the prior art and an embodiment of the present invention.

Referring to FIG. 18, as a result of scattering correction of an input image by the method proposed in the present invention, a profile similar to that of the target image is displayed, and it can be confirmed that the scattering correction is normally performed. In addition, it can be seen that the method proposed in the present invention shows a result closer to the target image than the conventional Monte Carlo-based iterative scatter correction method.

The quantitative evaluation results for the 3D reconstruction result image of the scattering correction method proposed in the present invention are shown in Table 4, and the similarity to the target image is measured as a result of performing 3D reconstruction after scattering by the method proposed in the present invention. You can see that all of the items have been improved. In particular, as a result of the scattering correction of the method proposed in the present invention, the SSIM with the target image was greatly improved from 0.823 to 0.96, and it can be seen that it shows superior performance in all of the measurement items compared to the existing Monte Carlo-based iterative scattering correction method.

[Table 4]

19 is a flowchart of a method for correcting scattering of an X-ray image according to another embodiment of the present invention.

Referring to FIG. 19, an X-ray image scattering correction method according to another embodiment of the present invention includes a training image generation step (S1910) using a Monte Carlo simulation and a scatter correction step (S1920) of an X-ray image using a deep learning technique. I can.

In the step of generating a training image using Monte Carlo simulation (S1910), a training image including a simulation image with a scattering line and a target image without a scattering line may be generated through Monte Carlo simulation using a 3D volume image. .

Thereafter, in the scattering correction step (S1920) of the X-ray image using the deep learning technique, the scattering of the input X-ray image may be corrected using a deep learning algorithm trained using the simulation image and the target image.

Here, the deep learning model uses Ds-DRCNN (Dual Scale Learing of Dilated Residual CNN), which separates scattering and noise of X-ray images into high-frequency components and low-frequency components, and independently constructs and applies two networks for each component. Can be adopted.

Ds-DRCNN includes Hs-Net (Residual CNN for Estimation of High Frequency Scatter) for correction of high frequency components and Ls-Net (Dilated Residual CNN for Estimation of Low Frequency Scatter) for correction of low frequency components. , A scattering-ray estimation image obtained by estimating a scattering ray component is obtained by summing the outputs of the two networks, and a scattering-corrected X-ray image in which the scattering ray does not exist may be obtained by subtracting the scattering ray estimation image from the input X-ray image.

Each step of the above-described X-ray image scattering correction method may be performed by a simulation tool and a deep learning engine supporting GPU calculation, and specific details thereof are the same as described above with reference to FIGS. 1 to 7 A typical explanation is omitted.

In addition, in the above-described embodiment, it has been described that a large amount of scattering occurs and a CBCT system that is sensitive to scattering is adopted to confirm the effectiveness of the present invention, but the present invention is not necessarily limited thereto. For example, the scattering correction method according to the present invention can be applied not only to a CBCT system but also to most X-ray imaging systems including 2D DR equipment.

The present invention is not limited by the above-described embodiments and the accompanying drawings. It will be apparent to those of ordinary skill in the art to which the present invention pertains, that components according to the present invention can be substituted, modified, and changed within the scope of the technical spirit of the present invention.

100: X-ray image acquisition device
200: X-ray image scattering correction system
210: training image generation unit
220: scattering correction unit

Claims (6)

A training image generator for generating a simulation image with scattering lines and a target image without scattering lines through Monte Carlo simulation using a 3D volume image; And
A scattering correction unit for correcting scattering of an X-ray image captured by an X-ray image acquisition device using a deep learning algorithm trained using the simulation image and the target image,
The scattering correction unit corrects scattering of the X-ray image using a deep learning network independently configured for a high frequency component and a low frequency component of the scattering components of the X-ray image.
The method of claim 1, wherein the training image generator,
A scattering correction system for an X-ray image, comprising obtaining the simulated image and separating a main X-ray component and a scattering ray component from the simulation image to obtain a main X-ray image and a scattering ray image, respectively.
The method of claim 1,
The X-ray image acquisition device is a 2D DR device or a 3D CT device.
delete Generating a training image including a simulation image with a scattering line and a target image without a scattering line through Monte Carlo simulation using a 3D volume image; And
Compensating the scattering of the input X-ray image using a deep learning algorithm trained using the simulation image and the target image,
In the step of correcting, scattering of the X-ray image is corrected using a deep learning network independently configured for a high frequency component and a low frequency component of the scattering components of the X-ray image.
The method of claim 5, wherein generating the training image comprises:
A method for compensating scattering of an X-ray image, comprising acquiring the simulation image and separating a main X-ray component and a scattering ray component from the simulation image to obtain a main X-ray image and a scattering ray image, respectively.

Images