WO2022177044A1 - Apparatus and method for generating high-resolution chest x-ray image by using attention-mechanism-based multi-scale conditional generative adversarial neural network - Google Patents

Abstract

An apparatus for generating a high-resolution chest X-ray image by using an attention-mechanism-based multi-scale conditional generative adversarial neural network may comprise: a model construction unit for constructing a multi-scale conditional generative adversarial neural network including a generator, which receives a potential variable and a conditional variable for the characteristic of a specific disease and performs up-sampling at every quarter to generate a plurality of chest X-ray images indicating the specific disease in different resolutions, and a discriminator for identifying whether the plurality of chest X-ray images is true; and an image generation unit, which inputs the potential variable and the conditional variable for the characteristic of the specific disease into the multi-scale conditional generative adversarial neural network to generate a chest X-ray image indicating the specific disease in a resolution greater than or equal to a preset value.

Description

Apparatus and method for generating high-resolution chest X-ray images using multi-scale conditional adversarial neural network based on attention mechanism

The present invention relates to an apparatus and method for generating a high-resolution chest X-ray image using a multi-scale conditional adversarial generating neural network based on a mechanism of interest.

With the recent development of artificial intelligence technology, it has been applied in various fields to achieve dramatic performance improvement. Among them, many AI-based researches in the medical domain are being actively conducted because the economic value of the medical industry has increased due to the expansion of medical demand due to the increase in human life expectancy due to the aging of the world. In particular, among various medical convergence studies, there are many studies on the heart and lungs. Chest-related diseases are common diseases that are frequently encountered by the general public due to environmental factors, but can be fatal if appropriate measures are not taken after early detection. Therefore, compared with various diseases occurring in other organs, the relative importance is high, and various studies are being conducted on the thoracic organ.

A common test for early detection of heart and lung diseases is chest X-ray, and the X-ray is transmitted through the chest area to determine the presence or absence of heart and lung-related diseases. Compared to other expensive tests such as Computed Tomography (CT) or Magnetic Resonance Imaging (MRI), chest X-ray examination is relatively inexpensive, can be taken in a short time, and has the advantage of low exposure dose.

However, it is a time-consuming task for clinicians to quantitatively and accurately read images of numerous patients in a busy medical field, and the labor cost is also high. To supplement this, recently, research results on a module that can read images quickly and accurately using artificial intelligence technology are coming out.

On the other hand, a balanced data set is very important for learning convergence in learning the current supervised learning-based AI model well. However, most of the medical data collected due to the difference in disease prevalence is numerically unbalanced, and when used as training data without pre-processing, it may overfit to a specific disease, resulting in undesirable convergence results. Similarly, in the case of chest X-ray images, learning progresses well for frequently occurring diseases (atelectasis, pleural effusion, infiltration, etc.), whereas other diseases (pneumonia, cardiac hypertrophy, etc.) have relatively poor learning performance.

In order to solve the problem of learning performance degradation due to such data numerical imbalance, various data augmentation techniques have been recently proposed and performance has been improved. Representative data augmentation methods of image data include up-down, left-right inversion, brightness adjustment, etc. However, since the reconstructed image data is generated based on standard data, the performance improvement may not be high for a small amount of data. Therefore, it is necessary to increase the number of standard data as a fundamental solution to create a much wider data set.

Recently, as a new solution to this problem, a data augmentation technique based on a generative adversarial network (GAN) that learns to estimate the distribution of standard data has been introduced. This neural network has been applied to various cases and proved that classification performance is improved by solving the problem of numerical imbalance in the data by creating false data that is almost similar to the real one. In this regard, Korean Patent Registration No. 10-2119056 discloses a method and apparatus for learning a medical image based on a generative adversarial neural network.

However, when there are multiple target classes, it is inefficient because the adversarial generative neural network must be trained for each class, and when the number of images is extremely limited, there is a problem that learning is almost impossible.

The present invention provides an apparatus and method for generating a chest X-ray image that can generate a high-resolution image that reflects the characteristics of each disease well with only one network for the purpose of solving the problem of numerical imbalance of chest X-ray image data sets for multiple diseases. would like to provide

However, the technical problems to be achieved by the present embodiment are not limited to the technical problems described above, and other technical problems may exist.

As a technical means for achieving the above-described technical problem, an embodiment of the present invention receives a latent variable and a condition variable for a characteristic of a specific disease and up-sampling for each branch to perform up-sampling for each of the specific resolutions of the specific A model building unit for constructing a multi-scale conditional adversarial generation neural network including a generator generating a plurality of chest X-ray images indicating a disease and a discriminator for discriminating whether the plurality of chest X-ray images are authentic, and the multi-scale conditional adversarial generation The apparatus for generating a chest X-ray image, comprising an image generator for generating a chest X-ray image representing the specific disease with a resolution greater than or equal to a preset value by inputting the latent variable and the condition variable for the characteristic of a specific disease into a neural network can provide

In addition, another embodiment of the present invention receives a latent variable and a condition variable for the characteristic of a specific disease, and up-sampling for each quarter to obtain a plurality of chest X-ray images representing the specific disease with different resolutions. generating a multi-scale conditionally adversarial generating neural network including a generating generator and a discriminator for discriminating whether or not the plurality of chest X-ray images are authentic; It is possible to provide a method for generating a chest X-ray image, which includes generating a chest X-ray image indicating the specific disease by inputting a condition variable and having a resolution greater than or equal to a preset value.

The above-described problem solving means are merely exemplary, and should not be construed as limiting the present invention. In addition to the exemplary embodiments described above, there may be additional embodiments described in the drawings and detailed description.

According to any one of the above-described problem solving means of the present invention, it is possible to control the target disease of the chest X-ray image generated by adding a condition variable as an input, thereby eliminating the inefficiency of learning as well as learning when the learning is extremely unbalanced. This impossible problem can be solved.

In addition, through the application of an attention mechanism, it is possible to generate a more realistic image by preserving the coherence of the form as a whole, not just a local part.

1 is a diagram illustrating an apparatus for generating a chest X-ray image according to an embodiment of the present invention.

2 is a diagram illustrating a network structure of a multi-scale conditional adversarial generating neural network according to an embodiment of the present invention.

3 is a diagram illustrating a loss graph of a generator and a discriminator according to an embodiment of the present invention.

4 is a view showing the distance of the Frechet Densnet as a result of comparing the performance of the experimental example and the comparative example of the present invention.

5 is a view showing a chest X-ray image generated as a result of comparing the performance of the experimental example and the comparative example of the present invention.

6 is a flowchart illustrating a method for generating a chest X-ray image according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. However, the present invention may be embodied in several different forms and is not limited to the embodiments described herein. And in order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

Throughout the specification, when a part is "connected" with another part, this includes not only the case of being "directly connected" but also the case of being "electrically connected" with another element interposed therebetween. . Also, when a part "includes" a certain component, it means that other components may be further included, rather than excluding other components, unless otherwise stated, and one or more other features However, it is to be understood that the existence or addition of numbers, steps, operations, components, parts, or combinations thereof is not precluded in advance.

In this specification, a "part" includes a unit realized by hardware, a unit realized by software, and a unit realized using both. In addition, one unit may be implemented using two or more hardware, and two or more units may be implemented by one hardware.

Some of the operations or functions described as being performed by the terminal or device in this specification may be instead performed by a server connected to the terminal or device. Similarly, some of the operations or functions described as being performed by the server may also be performed in a terminal or device connected to the corresponding server.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

1 is a diagram illustrating an apparatus for generating a chest X-ray image according to an embodiment of the present invention, and FIG. 2 is a diagram illustrating a network structure of a multi-scale conditional adversarial generation neural network according to an embodiment of the present invention.

The X-ray image generating apparatus 1 may generate a high-resolution image by well reflecting detailed characteristics of the ribs, diaphragm, lung, heart, etc. in the chest X-ray image through a multi-scale conditional adversarial generating neural network.

That is, a high-resolution image is essential to reflect detailed characteristics of organs in a chest X-ray image. According to the present application, a high-resolution image can be generated by learning the multi-scale image distribution from a low-resolution image to a high-resolution image.

According to the present application, the multi-scale conditional adversarial generation neural network can generate images for multiple diseases with only one network by adding a condition variable, which is a disease condition control factor.

On the other hand, in the case of chest X-ray images, the shape of the skeleton or organs is different for each patient (patient-specific), and the distribution of the intensity level is unclear due to peripheral organs such as blood vessels and noise, so it is difficult to determine the disease by only considering regional characteristics. . In addition, the problem of long-distance dependence within the image is essential because the diseased region may span the entire image or several may be distributed far apart.

According to the present application, it is possible to improve the generating performance of a chest X-ray image by solving the problem of organ dependence in a chest X-ray image generated by applying an attention mechanism to a multi-scale conditional adversarial generating neural network.

An example of the X-ray image generating apparatus 1 may include a mobile terminal capable of wired/wireless communication as well as a personal computer such as a desktop or a notebook computer. A mobile terminal is a wireless communication device that guarantees portability and mobility, and includes not only smartphones, tablet PCs, and wearable devices, but also Bluetooth (BLE, Bluetooth Low Energy), NFC, RFID, Ultrasonic, infrared, and Wi-Fi ( WiFi) and Li-Fi (LiFi) may include various devices equipped with a communication module. However, the X-ray image generating apparatus 1 is not limited to the shape shown in FIG. 1 or those exemplified above.

1 and 2 , the chest X-ray image generating apparatus 1 may include a model building unit 100 and an image generating unit 110 .

The model building unit 100 may build a multi-scale conditional adversarial generating neural network 200 including a generator 201 and discriminators 203 , 205 , and 207 .

The multi-scale conditional adversarial generating neural network 200 may be, for example, a network for learning multi-scale image distribution from a low-resolution image to a high-resolution image by extending StackGAN++ and LSGAN.

The multi-scale conditional adversarial generation neural network 200 is composed of one generator 201 and a plurality of discriminators 203, 205, and 207 to form a tree-structured network.

The generator 201 of the multi-scale conditional adversarial generative neural network 200 receives a latent variable and a condition variable for the characteristic of a specific disease, and up-sampling it for each branch to obtain a plurality of units representing a specific disease with different resolution. A chest X-ray image can be generated.

For example, the generator 201 may learn a multi-scale image distribution from a low-resolution image to a high-resolution image for each branch in the generator 201 to gradually generate a high-quality chest X-ray image.

In this case, the condition variable allows control of the chest X-ray image to be generated on behalf of a part of the latent variable for each disease.

The condition variable may be a one-hot encoding value for any one of a plurality of classes in order to generate a plurality of chest X-ray images indicating a plurality of chest diseases. For example, according to the present disclosure, one-hot encoding values for 8 classes may be assigned to the condition variable in order to generate chest X-ray images corresponding to 8 representative chest diseases.

That is, the initial distribution expressed by the standard normal distribution may be approximated to data within the standard data distribution of the corresponding scale for each quarter while passing through several hidden layers of the generator 201 together with the condition variable.

In addition, the discriminators 203 , 205 , and 207 of the multi-scale conditional adversarial generation neural network 200 may discriminate whether a plurality of chest X-ray images are authentic or not. In addition, the discriminators 203 , 205 , and 207 can further discriminate whether the plurality of chest X-ray images generated by the generator 201 satisfy the condition variable, unlike the discriminator of the conventional adversarial generative neural network. .

That is, the discriminator 203, 205, 207 determines whether the multi-scale image generated by the generator 201 is well made by the discriminator 203, 205, 207 corresponding to each branch of the generator 201. By evaluating, it guides the constructor 201 to be optimized.

The generator 201 may be divided into three branches to generate a chest X-ray image having a high resolution at a low resolution.

That is, the generator 201 includes a first branch 209 generating a chest X-ray image having a first resolution, a second branch 211 generating a chest X-ray image having a second resolution higher than the first resolution, and A third branch 213 for generating a chest X-ray image having a third resolution higher than the second resolution may be included.

For example, the generator 201 generates a chest X-ray image having the primary color and structure by approximating the image distribution of the first resolution in the first branch, and the chest X-ray image having the primary color and structure in the second branch A chest X-ray image expressing detailed information may be generated by approximating the image distribution of the second resolution higher than the first resolution by receiving the gradient of .

For example, the first branch 209 is a sub-generator for a low-resolution image 64x64, including four up-block layers, an attention layer, and a 3*3 convolution layer. ) may be included. Here, in the up-block layer, a nearest-neighbor method may be used for upsampling in order to mitigate the occurrence of checkerboard-artifact.

The sub-generators of the second branch 211 and the third branch 213 may include a joining layer and two residual layers and an up-block layer. In the second branch 211 and the third branch 213 , a 128×128 chest X-ray image and a 256×256 chest X-ray image may be output through 3×3 convolution after passing through the four layers.

Each quarter discriminator (203, 205, 207) consists of a down-sampling layer, and for calculating conditional and unconditional loss functions, the input part of the previous step of the last layer is divided into two parts and the condition variable is combined in only one place. can do it

The discriminators 203 , 205 and 207 are the first discriminator 203 , which determines whether the chest X-ray image generated in the first branch 209 is authentic or not, and the chest X-ray image generated in the second branch 211 . may include a second discriminator 205 for determining the authenticity of , and a third discriminator 207 for discriminating whether the chest X-ray image generated in the third branch 213 is authentic or not.

That is, for a total of m branches of the generator 201, there are sequentially a low-resolution image discriminator (i=1) and a high-resolution image discriminator (i=m), and the loss function of the discriminator of the i-th generator 201 branch is It can be defined as Equation 1. Here, the values of a and b are set to 0 and 1, respectively.

In Equation 1, s _i denotes samples for each size generated by the generator. That is, s _i = G _i (h _i ) and i has the range 0, 1, ..., m-1. Also, h _i = F(h _i-1 , z), where h _i may be a hidden feature of G _i .

The generator 201 is expressed as the sum of the discriminator loss functions for each branch, and the loss function may be defined as in Equation (2). Here, the value of d is set to 1.

That is, the generator 201 is trained in a direction of approximating the distribution of the non-conditional image and the conditional image with respect to images of high resolution from low resolution.

In this way, modeling the image distribution in multi-scale has the advantage of transferring the gradient to the initial layer well because the gradient for each scale is generated and transmitted. Ultimately, these features play a key role in stabilizing network learning, making it possible to generate high-resolution images.

3 is a diagram illustrating a loss graph of a generator and a discriminator according to an embodiment of the present invention.

Figure 3 (a) shows the loss graph of the generator 201 in the learning phase, Figure 3 (b) shows the loss graph of the discriminators (203, 205, 207). Here, the x-axis means iteration and the y-axis means loss.

Since the two neural networks compete with each other, it means that each neural network is properly trained only when the loss value is large.

4 is a view showing the Frechet Densnet distance as a performance comparison result of the Experimental Example and Comparative Example of the present invention, and FIG. 5 is a view showing a chest X-ray image generated as a performance comparison result of the Experimental Example and Comparative Example of the present invention. to be.

The present applicant used 112,120 chest X-ray image data for 14 heart and lung diseases provided by the NIH Clinical Center for the chest X-ray image data augmentation experiment. Among them, only 8 representative diseases (atelectasis, cardiac hypertrophy, pleural effusion, infiltration, mass, nodule, pneumonia, and pneumothorax) were used.

The present applicant conducted two experiments for 8 representative diseases of the chest for performance comparison.

The first experiment consists of 1) a conventional deep convolutional adversarial generating neural network and 2) generating high-resolution images (256x256) for each disease with our multi-scale conditional adversarial generating neural network. The performance of the network is compared with the Fre'chet DenseNet distance (FDD) (FIG. 4).

The number of images used in the experiment is about 100,000, and when the batch size is 16, 7,000 repetitions per epoch are made. Referring to the graph of FIG. 3 , the discriminator was gradually stabilized after 5 epochs, and the function fluctuation was stabilized in the generator around 20 epochs (repeat 140,000). When the training was conducted beyond that, both models lost their balance and both models collapsed and the learning was stopped.

The Fre'chet Inception Distance (FID) is a measure of the difference between two normal distributions, and it was proposed to compensate for the disadvantage that the Inception Score (IS) does not use the distribution of actual data. The FID is calculated as in Equation 3, and a smaller value means better quality. where (m, C) and (m _w , C _w ) is the mean and covariance of the generated data and the actual data.

However, if each normal distribution is obtained based on the inception model pre-trained on ImageNet, which is a natural image consisting of 1,000 classes and 1.2 million, the probability that the image used to find the distribution belongs to 1,000 classes through the pre-trained model. This is problematic in that it outputs a vector. The reason is that because the inception model is trained based on natural images, it does not include medical images such as chest X-rays as a class. In other words, the chest X-ray image is not a natural image, so it is not suitable for interpreting these specialized features. Therefore, in order to obtain the distribution of medical images and to evaluate the quality through the distance between distributions, a model optimized for the medical domain is needed. Therefore, after training DenseNet using medical images as training data, Fre'chet DenseNet Distance (FDD), which measures the distance by finding the normal distribution of two medical images based on this model, is proposed as a new index. to conduct the experiment.

First, for the first experiment, DenseNet-121 was trained with 78,468 training data and 11,219 validation data. And based on the learned model, the performance of the model was quantitatively evaluated by obtaining the FDD values for the real image, cDCGAN, and the real image and the proposed model. Here, the actual video means the remaining part that is not used for DenseNet-121 training. However, due to the high data imbalance due to the characteristics of medical imaging, the number of image data for each class is uneven. Therefore, the number of actual images and generated images by class is 1,000 attelectasis, 575 cardiomegaly, 1,000 effusion, 1,000 infiltration, 729 mass, 774 nodule, 242 pneumonia, and 539 pneumothorax, for a total of 5,859 images. The test was performed in the same way.

As a result, as shown in FIG. 4, the FDD value calculated by calculating the distance between the real image and the image generated through the multi-scale conditional adversarial generating neural network of the present application is the distance between the real image and the image generated by cDCGAN, a conventional deep convolutional adversarial generating neural network. Since it is generally lower than the calculated FDD value, it can be confirmed that the multi-scale conditional adversarial generating neural network of the present application is superior.

The second experiment is a qualitative comparison of the experimental results from the conventional adversarial generative neural network, DCGAN, and our multi-scale conditionally adversarial neural network, with actual images. The location was found and marked.

FIG. 5 shows the results of real data and the conventional deep convolutional adversarial neural network, and the multi-scale conditional adversarial neural network of the present application. The specific method for distinguishing the disease is as follows.

Atelectasis causes air loss of all or part of the lungs, usually accompanied by a decrease in lung volume, biased respiratory tract (airway), or whitening of the lungs. Cardiomegaly is defined as a condition in which the ventricular wall (muscle) thickens and the weight of the myocardium increases, and the length of the cardiac shadow occupies more than half of the internal length of the thoracic shadow. Effusion is a stagnant fluid in the pleural cavity, which shows an unbalanced increase in shading in the left and right thoracic cavities. Infiltration is a form of inflammatory cells gathered in normal tissue, and it shows an increase in alveolar shade and is well seen around the lungs. A mass is a large mass with a diameter of 3 cm or more and increased shading. It can be described anywhere in the chest, including the lungs, pleura, mediastinum (sinus), and chest wall. A nodule is a round lung shade with a border of 3 cm or less. Pneumonia is an inflammatory reaction in the lungs, mainly caused by bacterial infection, and shadows such as mesh or honeycomb are found in the image. Finally, in pneumothorax, a hole in the pleural cavity is formed and a large air sac is visible.

Based on the above disease description, it can be confirmed that the results of the conventional deep convolutional adversarial neural network have significantly lowered image resolution compared to the actual data. That is, the shape of the ribs or the shape of the lung bronchus is not clearly visible, so it is difficult to say that the main features have been properly learned. On the other hand, the results of our multi-scale conditional adversarial neural network showed image features and resolution that were indistinguishable from the actual images, and the characteristics of each disease were also learned so well that there was no significant difference from the actual images. For example, in the case of cardiac hypertrophy, compared to other diseases, the shape of the heart is significantly increased, which is a result of well learning the characteristic of the disease, 'increase in heart size'.

6 is a flowchart illustrating a method for generating a chest X-ray image according to an embodiment of the present invention. The method for generating a chest X-ray image according to an embodiment illustrated in FIG. 6 includes operations that are time-series processed by the apparatus 1 for generating a chest X-ray image illustrated in FIG. 1 . Therefore, even if omitted below, it is also applied to the method for generating a chest X-ray image performed according to the embodiment shown in FIG. 6 .

Referring to FIG. 6 , in step S600 , a multi-scale conditional adversarial generation neural network including a generator and a discriminator may be generated. Here, the generator may generate a plurality of chest X-ray images representing specific diseases having different resolutions by receiving the latent variable and the condition variable for the characteristic of a specific disease and performing up-sampling for each branch. In addition, the discriminator may discriminate whether the plurality of chest X-ray images are authentic or not.

In operation S610, a chest X-ray image may be generated through the multi-scale conditional adversarial generation neural network. For example, by inputting a latent variable and a condition variable for a characteristic of a specific disease into a multi-scale conditional adversarial generating neural network, a chest X-ray image representing a specific disease with a resolution greater than or equal to a preset value may be generated.

The chest X-ray image generating method described with reference to FIG. 6 may be implemented in the form of a computer program stored in a medium, or in the form of a recording medium including instructions executable by a computer, such as a program module executed by a computer. can Computer-readable media can be any available media that can be accessed by a computer and includes both volatile and nonvolatile media, removable and non-removable media. Also, computer-readable media may include computer storage media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.

The foregoing description of the present invention is for illustration, and those of ordinary skill in the art to which the present invention pertains can understand that it can be easily modified into other specific forms without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. For example, each component described as a single type may be implemented in a distributed manner, and likewise components described as distributed may also be implemented in a combined form.

The scope of the present invention is indicated by the following claims rather than the above detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention. do.

Claims (18)

A high-resolution chest X-ray image generating apparatus using a multi-scale conditional adversarial generating neural network based on a mechanism of interest, the apparatus comprising: A generator for generating a plurality of chest X-ray images representing the specific disease having different resolutions by receiving input of latent variables and condition variables for characteristics of specific diseases and up-sampling for each branch, and the plurality of chest Xs a model building unit for constructing a multi-scale conditional adversarial generating neural network including a discriminator for discriminating whether a line image is authentic or not; and An image generator for generating a chest X-ray image representing the specific disease with a resolution greater than or equal to a preset value by inputting the latent variable and the conditional variable for the characteristic of a specific disease into the multi-scale conditional adversarial generation neural network That comprising a, chest X-ray image generating device. The method of claim 1, The generator generates a chest X-ray image having a basic color and structure by approximating an image distribution of a first resolution in a first branch, and a gradient of a chest X-ray image having the basic color and structure in a second branch to generate a chest X-ray image expressing detailed information by approximating an image distribution of a second resolution higher than the first resolution by receiving the information. 3. The method of claim 2, The discriminator is a first discriminator for determining the authenticity of the chest X-ray image generated in the first branch, a second discriminator for determining the authenticity of the chest X-ray image generated in the second branch, and the third discriminator The chest X-ray image generating apparatus comprising a third discriminator for determining whether the chest X-ray image generated in the branch is authentic. 4. The method of claim 3, The loss function of the first discriminator, the second discriminator, and the third discriminator is expressed by the following equation, a chest X-ray image generating apparatus.

5. The method of claim 4, The loss function of the generator is expressed by the following equation, chest X-ray image generating apparatus.

The method of claim 1, The discriminator will further discriminate whether the plurality of chest X-ray images satisfy the condition variable. 3. The method of claim 2, The first branch is an up-block layer (up-block layer), attention layer (attention layer) and the convolution layer (convolution layer) comprising a, chest X-ray image generating apparatus. 3. The method of claim 2, Wherein the second branch and the third branch include a residual layer and an up-block layer. The method of claim 1, The condition variable is a one-hot encoding value for any one of a plurality of classes to generate the plurality of chest X-ray images indicating a plurality of chest diseases, chest X-ray image generation Device. In a method for generating a high-resolution chest X-ray image using a multi-scale conditional adversarial generation neural network based on a mechanism of interest, A generator for generating a plurality of chest X-ray images representing the specific disease having different resolutions by receiving input of latent variables and condition variables for characteristics of specific diseases and up-sampling for each branch, and the plurality of chest Xs constructing a multi-scale conditional adversarial generative neural network including a discriminator for discriminating whether a line image is authentic or not; and generating a chest X-ray image representing the specific disease with a resolution greater than or equal to a preset value by inputting the latent variable and the condition variable for the characteristic of a specific disease into the multi-scale conditional adversarial generation neural network A method of generating a chest X-ray image comprising a. 11. The method of claim 10, The generator generates a chest X-ray image having a basic color and structure by approximating an image distribution of a first resolution in a first branch, and a gradient of a chest X-ray image having the basic color and structure in a second branch The method of generating a chest X-ray image that expresses detailed information by approximating an image distribution of a second resolution higher than the first resolution by receiving the information. 12. The method of claim 11, The discriminator is a first discriminator for determining the authenticity of the chest X-ray image generated in the first branch, a second discriminator for determining the authenticity of the chest X-ray image generated in the second branch, and the third discriminator A method of generating a chest X-ray image, comprising a third discriminator for determining the authenticity of the chest X-ray image generated in the branch. 13. The method of claim 12, The loss function of the first discriminator, the second discriminator and the third discriminator is expressed by the following equation, the chest X-ray image generating method.

14. The method of claim 13, The loss function of the generator is expressed by the following equation, chest X-ray image generation method.

11. The method of claim 10, The discriminator will further discriminate whether the plurality of chest X-ray images satisfy the condition variable. 12. The method of claim 11, Wherein the first branch comprises an up-block layer, an attention layer and a convolution layer. 12. The method of claim 11, Wherein the second branch and the third branch include a residual layer and an up-block layer. 11. The method of claim 10, The condition variable is a one-hot encoding value for any one of a plurality of classes to generate the plurality of chest X-ray images indicating a plurality of chest diseases, chest X-ray image generation Way.

Images