WO2021071286A1

WO2021071286A1 - Generative adversarial network-based medical image learning method and device

Info

Publication number: WO2021071286A1
Application number: PCT/KR2020/013739
Authority: WO
Inventors: 김원태; 강신욱; 이명재; 김동민; 남동연
Original assignee: (주)제이엘케이
Priority date: 2019-10-08
Filing date: 2020-10-08
Publication date: 2021-04-15
Also published as: KR102119056B1

Abstract

A GAN-based medical image learning device can be provided according to the present invention. The GAN-based medical image learning device comprises: a learning data set management unit for managing a learning data set comprising a first medical image, labeled data and a filtered medical image; and a lesion learning unit having a generator learning part, for managing learning of a generator for generating a second medical image, and a discriminator learning part for managing learning of a discriminator, for constructing the labeled data, by means of the first medical image, the labeled data and the medical image filtered from the second image generated by means of the generator, wherein the learning data set management unit can have a second medical image filtering part for filtering the second medical image and selectively providing same as the filtered medical image.

Description

A method and apparatus for learning medical images based on generative adversarial neural networks

The present disclosure relates to a technology for learning a deep learning model, and more specifically, to a method and an apparatus for learning about a lesion through unsupervised learning.

Deep learning is to learn a very large amount of data, and when new data is input, the highest probability is selected based on the learning result. Such deep learning can operate adaptively according to an image, and since it automatically finds a characteristic factor in the process of learning a model based on data, there are increasing attempts to utilize this in the field of artificial intelligence in recent years.

However, in order for the learned model to derive accurate and accurate results, large-scale data learning is required.

In particular, in order to apply artificial intelligence technology to the medical field, a large amount of data confirmed by an expert (i.e., labeling data) is required, but a large amount of data confirmed by an expert is constructed due to time and cost issues. There is a problem that is not easy to do.

The technical task of the present disclosure is to provide a method and apparatus for learning a medical image based on a Generative Adversarial Network (GAN) capable of constructing a high-performance learning model using a small number of labeling data.

The technical problems to be achieved in the present disclosure are not limited to the technical problems mentioned above, and other technical problems that are not mentioned will be clearly understood by those of ordinary skill in the technical field to which the present disclosure belongs from the following description. I will be able to.

According to an aspect of the present disclosure, a GAN-based medical image learning apparatus may be provided. In the apparatus for learning a medical image learning model, the apparatus includes a learning data set management unit that manages a learning data set including a first medical image, labeling data, and a filtered medical image, and a generator that generates a second medical image. A generator learning unit that manages learning of (Generator), and a discreminator configuring the labeling data by using the medical image filtered from the first medical image, the labeling data, and the second image generated through the generator. A lesion learning unit having a discreminator learning unit that manages learning of (Discriminator), wherein the learning data set management unit filters the second medical image to selectively provide a second medical image as the filtered medical image. A filtering unit may be provided.

According to another aspect of the present disclosure, a GAN-based medical image learning method may be provided. The method is a method of learning a learning model for lesion detection, the process of managing and storing a learning data set including a first medical image, labeling data, and a filtered medical image, and generating a second medical image. And a discriminator constituting the labeling data by using a generator to perform the labeling data, and a medical image filtered from the first medical image, labeling data, and the second image generated through the generator. Including the process of performing the learning of the lesion learning model, the process of performing the learning of the lesion learning model includes a process of detecting and providing a second medical image generated by the generator, and managing the training data set And the storing process may include a process of filtering the second medical image and selectively storing and managing it as the filtered medical image.

Features briefly summarized above with respect to the present disclosure are only exemplary aspects of the detailed description of the present disclosure described below, and do not limit the scope of the present disclosure.

According to the present disclosure, a GAN-based medical image learning method and apparatus capable of constructing a high-performance learning model using a small number of labeling data may be provided.

The effects obtainable in the present disclosure are not limited to the above-mentioned effects, and other effects not mentioned may be clearly understood by those of ordinary skill in the art from the following description. will be.

1 is a block diagram showing the configuration of a GAN-based medical image learning apparatus according to an embodiment of the present disclosure.

2A and 2B are diagrams illustrating an operation of learning a lesion learning model by the GAN-based medical image learning apparatus according to an embodiment of the present disclosure.

3 is a diagram illustrating an operation of restoring an error image by the GAN-based medical image learning apparatus according to an embodiment of the present disclosure.

4A and 4B are diagrams illustrating a medical image in a normal state and a medical image in an abnormal state used in the GAN-based medical image learning apparatus according to an embodiment of the present disclosure.

5 is a flowchart illustrating a procedure of a GAN-based medical image learning method according to an embodiment of the present disclosure.

6 is a block diagram illustrating a computing system that executes a GAN-based medical image learning method and apparatus according to an embodiment of the present disclosure.

Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art may easily implement the embodiments. However, the present disclosure may be implemented in various different forms and is not limited to the embodiments described herein.

In describing an embodiment of the present disclosure, when it is determined that a detailed description of a known configuration or function may obscure the subject matter of the present disclosure, a detailed description thereof will be omitted. In addition, parts not related to the description of the present disclosure in the drawings are omitted, and similar reference numerals are attached to similar parts.

In the present disclosure, when a component is said to be "connected", "coupled" or "connected" with another component, it is not only a direct connection relationship, but also an indirect connection relationship in which another component exists in the middle. It can also include. In addition, when a certain component "includes" or "have" another component, it means that other components may be further included rather than excluding other components unless otherwise stated. .

In the present disclosure, terms such as first and second are used only for the purpose of distinguishing one component from other components, and do not limit the order or importance of the components unless otherwise noted. Accordingly, within the scope of the present disclosure, a first component in one embodiment may be referred to as a second component in another embodiment, and similarly, a second component in one embodiment is referred to as a first component in another embodiment. It can also be called.

In the present disclosure, components that are distinguished from each other are intended to clearly describe each feature, and do not necessarily mean that the components are separated. That is, a plurality of components may be integrated into one hardware or software unit, or one component may be distributed to form a plurality of hardware or software units. Therefore, even if not stated otherwise, such integrated or distributed embodiments are also included in the scope of the present disclosure.

In the present disclosure, components described in various embodiments do not necessarily mean essential components, and some may be optional components. Accordingly, an embodiment consisting of a subset of components described in an embodiment is also included in the scope of the present disclosure. In addition, embodiments including other elements in addition to the elements described in the various embodiments are included in the scope of the present disclosure.

Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings.

Referring to FIG. 1, the GAN-based medical image learning apparatus may include a lesion learning unit 10 and a learning data set management unit 15.

The lesion learning unit 10 is a component that processes learning about the lesion learning model 110, and processes learning about the lesion learning model 110 using the learning data set provided by the learning data set management unit 15. do.

In particular, the lesion learning model 110 may be a learning model built through GAN-based quasi-supervised learning, and may be composed of a combination of a generator and a discriminator. Correspondingly, the lesion learning unit 10 may include a generator learning unit 11 that manages learning of the generator and a discreminator learning unit 12 that manages the learning of the discreminator.

The generator learning unit 11 may perform learning so that the generator can generate a medical image. In addition, the discreminator learning unit 12 includes a medical image generated by the generator (hereinafter, referred to as a'second medical image') and a medical image provided by the learning data set management unit 15 (hereinafter, referred to as a'first medical image'). Medical images') and labeling data may be input, and the learning may be performed to determine labeling data among the received images.

Furthermore, the lesion learning model 110 may include a lesion region detection learning model 111 and a lesion diagnosis learning model 113. The lesion region detection learning model 111 may be a learning model that receives a medical image and detects and outputs a lesion region in the medical image. In addition, the lesion diagnosis learning model 113 may be a learning model that receives a region in which a lesion is generated provided from the lesion region detection learning model 111 and determines and outputs a disease type and disease severity.

Based on the structure of the lesion learning model 110 described above, the generator learning unit 11 may perform learning so that the generator provided in the lesion region detection learning model 111 can generate a second medical image. , The disc limiter learning unit 12 may perform learning so that the disc limiter receives the first medical image, the second medical image, and labeling data, and identifies the labeling data. In this case, the labeling data may be an image in which a region in which a lesion occurs is detected in the first medical image.

In addition, the generator learning unit 11 may perform learning so that the generator provided in the lesion diagnosis learning model 113 can generate an image of an area where a lesion occurs, and the discreminator learning unit 12 The reminator receives the image of the area where the lesion occurs (e.g., the image of the area where the lesion is generated provided by the generator, the image of the area where the lesion is generated provided from the training data set, etc.), and labeling data. Learning can be performed to identify the data. In this case, the labeling data may be an image in which the type of disease and the severity of the disease are detected based on the image of the area where the lesion occurs.

Furthermore, the generator learning unit 11 resizes the image of the lesion-prone area provided by the generator according to the size and resolution of the first and second medical images, and then provides the resized image to the discreminator for learning. You can handle it.

In the above-described embodiment, while performing the learning of the lesion learning model 110, the generator constructs a predetermined image and uses it for learning of the discreminator. However, when all the images generated by the generator are used in the discreminator, a problem may occur in that the performance of the discreminator is affected by the performance level of the generator. Accordingly, it is preferable that the lesion learning apparatus according to an embodiment of the present disclosure is configured to selectively use the image generated by the generator for learning of the discreminator. To this end, the generator learning unit 11 does not deliver all of the second medical images generated by the generator to the discreminator, but delivers them to the learning data set management unit 15, and then the learning data set management unit 15 provides It may be configured to select an image and provide it to the discreminator.

Specifically, the generator learning unit 11 provides the second medical image generated in the learning process of the lesion learning model 110 to the learning data set management unit 15. In response to this, the learning data set management unit 15 may include a second medical image filtering unit 16 that selects a second medical image.

Preferably, the training data set management unit 15 may include a histogram verification unit 16 that checks the histogram of the second medical image, and the histogram verification unit 16 is a second generation unit provided by the generator learning unit 11. The histogram of the medical image may be checked and provided to the second medical image filtering unit 16. In addition, the histogram checking unit 16 may check the histogram of the first medical image 151 included in the training data set 150 and provide the result to the second medical image filtering unit 16.

Based on the foregoing, the second medical image filtering unit 16 compares the histogram for the first medical image 151 and the histogram for the second medical image to filter only the image corresponding to the first medical image. I can. For example, the second medical image filtering unit 16 constructs at least one reference histogram information using the histogram for the first medical image 151, and includes at least one reference histogram information and the second medical image. By comparing the histograms, a second medical image exceeding at least one reference histogram information may be determined as the filtered image. Thereafter, the second medical image filtering unit 16 may provide the determined image, that is, the filtered image, and the learning data set management unit 15 may store this as the filtered image 153 in the learning data set 150. have.

Based on the above-described structure, the learning data set 150 may include the first medical image 151, the labeling data 152, and the filtered image 153, and the discreminator learning unit 12 May perform discreminator learning of the lesion learning model 110 using the first medical image 151, the labeling data 152, and the filtered image 153 included in the training data set 150. have.

Meanwhile, an error may occur in the first medical image due to data loss in a process of collecting the first medical image, for example, in the process of photographing, transmitting, and storing the medical image. Since the first medical image is important data used for learning, when an error occurs, there may be a problem in that the learning of the lesion learning model 110 is not accurately performed. In consideration of this, the learning data set management unit 15 may further include an error image management unit 18 that checks and manages whether or not an image of the first medical image has an error.

The error image management unit 18 may include an error image check unit 18a for checking whether an error exists in the first medical image, and an error image restoration unit 18b for restoring the error image.

The error image verification unit 18a analyzes the first medical image and checks whether or not there is an error. For example, the error image verification unit 18a may check the histogram of the first medical image 151 and determine whether or not there is an error in the first medical image 151 by comparing it with a predetermined criterion.

The error image restoration unit 18b may restore the second medical image to an image that replaces the first medical image 151 in which an error has occurred. For example, the error image restoration unit 18b may detect a second medical image having high similarity by comparing the histogram of the second medical image with the histogram of the first medical image 151 in which an error has occurred. The error image restoration unit 18b may restore the detected second medical image to the first medical image 151 in which an error has occurred.

On the other hand, when the labeling data is remarkably small, since the lesion learning model 110 may not be properly trained, a certain number of labeling data or more is required. To this end, the learning data set management unit 15 may check the number of labeling data included in the learning data set 150 and, if less than a predetermined number or ratio, configure and store the labeling data. For example, the learning data set management unit 15 may include a learning model (not shown) for learning labeling data 152 corresponding to the first medical image 151. Here, the learning model (not shown) may be configured based on supervised learning. In addition, the learning data set management unit 15 can generate labeling data by inputting the first medical image 151 into the learning model (not shown), and the labeling data generated in this way has a relatively high accuracy. It can fall. In consideration of this, the learning data set management unit 15 checks the probability value of the labeling data output by the learning model (not shown), selects labeling data having a relatively high probability value, and labels the learning data set 150 It can be stored as data 152.

First, FIG. 2A shows an operation of learning a lesion region detection learning model 111 by a GAN-based medical image learning apparatus.

First, the discreminator 21 learns the characteristics of the lesion area by proactively learning the lesion area. In a state in which the discreminator 21 has previously learned about the lesion area, the generator 22 may generate the second medical image 202. The second medical image 202 generated in this way may be provided to the learning data set management unit 23, and the learning data set management unit 23 checks the histogram of the second medical image 202, and based on this, the second medical image 202 Filtering on the medical image 202 may be performed. In addition, the learning data set management unit 23 may store and manage the filtered second medical image 203 in the learning data set. Thereafter, the learning data set management unit 23 may provide the first medical image 201, the second medical image 203, and the first labeling data 205 to the disc limiter 21, and the disc limiter The navigator 21 may be trained to classify the first labeling data 205 among the provided data 201, 203, and 205. In this case, the first labeling data 205 may be data obtained by extracting a lesion area from a medical image.

In this way, the lesion region detection learning model 111 can construct a GAN-based unsupervised learning model through the discreminator 21 and the generator 22, and detects a lesion region corresponding thereto when a medical image is input. Can be printed.

Furthermore, the lesion learning unit 10 provides the first labeling data 205, that is, data extracted from the medical image, to the lesion diagnosis learning model 113 when learning the lesion area detection learning model 111. , It can be configured to link the learning of the lesion region detection learning model 111 and the lesion diagnosis learning model 113.

Meanwhile, referring to FIG. 2B, the discreminator 21 learns characteristics of the disease type and disease severity by proactively learning about the type of disease and the severity of the disease. In a state in which the discreminator 21 has previously learned about the type of disease, the severity of the disease, and the like, the generator 22 may generate the fourth medical image 212. In this case, the fourth medical image 212 generated by the generator 22 may be an image from which a lesion area is extracted.

The fourth medical image 212 generated through the above-described operation may be provided to the learning data set management unit 23, and the learning data set management unit 23 checks the histogram of the fourth medical image 212, and Based on the filtering, the fourth medical image 212 may be filtered. In addition, the learning data set management unit 23 may store and manage the filtered fourth medical image 213 in the learning data set. After that, the learning data set management unit 23 may provide the third medical image 211, the fourth medical image 213, and the second labeling data 215 to the disc limiter 21, and The navigator 21 may be trained to classify the second labeling data 215 among the provided data 211, 213, and 215. In this case, the second labeling data 215 may be data obtained by extracting a disease type and a disease severity from an image of a lesion area.

Further, in order to link the learning of the lesion region detection learning model 111 and the lesion diagnosis learning model 113, instead of the third medical image 211, the lesion region provided by the lesion region detection learning model 111 ( 217) may be used as an input to the lesion diagnosis learning model 113.

In this way, the lesion diagnosis learning model 113 can construct a GAN-based unsupervised learning model through the discreminator 21 and the generator 22, and the disease type corresponding thereto when an image of the lesion area is input, Disease severity, etc. can be detected and output.

When configuring the training data set, the learning data set management unit 15 requests the error image verification unit 18a to check the error of the first medical image 151 in order to check whether the first medical image 151 has an error. I can. Correspondingly, the error image verification unit 18a may check the histogram information of the first medical image 151 (301), and check whether the first medical image 151 is in error based on the confirmed histogram information. (302).

As illustrated in FIG. 4A, the histogram of the medical image 410 in a normal state may be evenly distributed, but when an error occurs in the generation, storage, or transmission of the medical image, image information for a part of the area is lost. Can be. Accordingly, when the medical image 420 is configured using the lost information, the lost information cannot be restored to the image information, so the corresponding area 425 may be configured as an abnormal area. In this way, when a medical image including the abnormal region 425 is configured, the histogram of the medical image 420 including the abnormal region 425 is not evenly distributed, and the distribution for a specific color is concentrated. Can be.

In consideration of the above, the error image verification unit 18a checks the histogram information of the first medical image 151, and when the corresponding histogram information indicates a distribution concentrated on a specific color, that is, the histogram of a specific color is If it exceeds the predetermined threshold, it may be determined that an error exists in the medical image.

The error image restoration unit 18b may be configured to replace the second medical image generated in the learning process of the lesion learning unit 10 with a first medical image in which an error exists. Specifically, the similarity of the histogram of the first medical image and the second medical image is calculated (303), and when the calculated similarity exceeds a predetermined threshold, the second medical image is restored to be replaced with the first medical image. It can be determined by the image (304).

As another example, the error image restoration unit 18b calculates a histogram in the horizontal and vertical directions in units of a predetermined line area of the first medical image, and detects a line area in which the histogram of a specific color exceeds a predetermined threshold. can do. Based on this, the error image restoration unit 18b may detect an error area in a horizontal direction and a vertical direction, and may distinguish an error area from a non-error area. Thereafter, the error image restoration unit 18b checks the histogram of the first medical image and the second medical image based on the non-error region, calculates the similarity of the histogram, and the calculated similarity exceeds a predetermined threshold. After extracting the second medical image, it may be replaced with the first medical image.

Although, in the embodiment of the present disclosure, the operation of confirming an error image and determining a reconstructed image using the first and second medical images is exemplified, the present disclosure is not limited thereto, and may be variously changed or applied. For example, the operation of checking the error image and determining the reconstructed image may be applied to the third and fourth medical images.

The GAN-based medical image learning method may be performed by the GAN-based medical image learning apparatus described above.

First, in step S510, the apparatus for learning a medical image may configure and store a learning data set. The training data set is a data set for learning a lesion learning model, and may include a medical image and labeling data.

Furthermore, the lesion learning model may include a lesion region detection learning model and a lesion diagnosis learning model. The lesion region detection learning model is a learning model that receives a medical image and detects and outputs a lesion area in the medical image. I can. In addition, the lesion diagnosis learning model may be a learning model that receives a region in which a lesion is generated provided from a lesion region detection learning model, determines a type of disease, a severity of a disease, and the like, and outputs it. Based on this, the medical image and the labeling data may be configured such that data provided to the lesion region detection learning model and the lesion diagnosis learning model are classified.

For example, the first medical image provided to the lesion region detection learning model may be a medical image photographing a patient's body, and the first labeling data may be an image that detects an area where a lesion occurs in the first medical image. have. In addition, the medical image provided to the lesion diagnosis learning model may be an image (hereinafter referred to as a third medical image) in which the image of the lesion area is resized according to the size and resolution of the first medical image, and the second labeling The data may be an image in which the type of disease and the severity of the disease are detected based on the image of the area where the lesion occurs.

Meanwhile, an error may occur in the first medical image (or the third medical image) due to data loss in the process of collecting the first medical image, for example, in the process of capturing, transmitting, and storing the medical image. Since the first medical image is important data used for learning, when an error occurs, there may be a problem in that the learning of the lesion learning model is not accurately performed. In consideration of this, the apparatus for learning a medical image may check and manage whether the first medical image (or the third medical image) is an error. Specifically, the apparatus for learning a medical image checks whether or not there is an error by analyzing the first medical image (or the third medical image). For example, the medical image learning apparatus may check the histogram of the first medical image (or the third medical image) and determine whether or not there is an error in the first medical image (or the third medical image) by comparing it with a predetermined criterion. .

The apparatus for learning a medical image may restore the second medical image to an image that replaces the first medical image (or the third medical image) in which an error has occurred. For example, the error image medical image learning apparatus compares the histogram of the second medical image (or the fourth medical image) with the histogram of the first medical image (or the third medical image) in which the error has occurred, and has a high similarity. 2 A medical image (or a fourth medical image) can be detected. The apparatus for learning a medical image may restore the detected second medical image (or fourth medical image) to a first medical image (or a third medical image) in which an error has occurred.

On the other hand, if the labeling data is remarkably small, since the lesion learning model may not be properly trained, a certain number of labeling data or more is required. To this end, the medical image learning apparatus may check the number of labeling data included in the training data set, and if it is included in less than a predetermined number or ratio, configure and store the labeling data. For example, the apparatus for learning medical images may include a learning model (not shown) for learning labeling data corresponding to a first medical image. Here, the learning model (not shown) may be configured based on supervised learning. In addition, the apparatus for learning medical images may generate labeling data by inputting the first medical image to the learning model (not shown), and the labeling data generated in this manner may be relatively inferior in accuracy. In consideration of this, the apparatus for learning a medical image may check a probability value of labeling data output by a learning model (not shown), select labeling data having a relatively high probability value, and store it as labeling data of a training data set.

In step S520, the apparatus for learning a medical image processes learning about a lesion learning model by using the learning data set.

In particular, the lesion learning model may be a learning model built through GAN-based quasi-supervised learning, and may be composed of a combination of a generator and a discriminator. In response to this, the apparatus for learning a medical image may perform learning of the generator and the discreminator, respectively. For example, the apparatus for learning a medical image may perform learning so that a generator can generate a medical image. In addition, the apparatus for learning a medical image may perform learning of the discreminator. Specifically, the apparatus for learning a medical image may be trained to receive a medical image generated by a generator, a medical image of a learning data set, and labeling data, and to determine labeling data from the received images.

Further, based on the structure of the lesion learning model described above, in step S521, the medical image learning apparatus may perform learning so that the generator provided in the lesion region detection learning model generates a second medical image, and in step S522 At, the discreminator may receive the first medical image, the second medical image, and the first labeling data, and perform learning so that the first labeling data can be identified. In this case, the first labeling data may be an image obtained by detecting a region in which a lesion occurs in the first medical image.

In addition, in step S523, the medical image learning apparatus may perform learning so that the generator provided in the lesion diagnosis learning model can generate an image (hereinafter, referred to as a'fourth medical image') of the area where the lesion occurs. , In step S524, the discreminator receives the fourth medical image, the image (ie, the third medical image) of the lesion-prone area provided from the training data set, and the second labeling data, and receives the second labeling data. Learning can be performed so that you can identify it. In this case, the second labeling data may be an image in which the type of disease and the severity of the disease are detected based on the image of the area where the lesion occurs.

Further, the medical image learning apparatus resizes the image of the lesion-prone area provided by the generator according to the size and resolution of the first and second medical images, and then provides the resized image to the discreminator to process the learning. I can.

In the above-described embodiment, while performing the learning of the lesion learning model, it is exemplified that the generator constructs a predetermined image and uses it for learning of the discreminator. However, when all the images generated by the generator are used in the discreminator, a problem may occur in that the performance of the discreminator is affected by the performance level of the generator. Accordingly, it is preferable that the apparatus for learning a medical image according to an embodiment of the present disclosure is configured to selectively use the image generated by the generator for learning of the discreminator. To this end, the medical image learning apparatus does not deliver all the second medical images (or fourth medical images) generated by the generator to the discreminator, but selectively transmits the second medical images (or fourth medical images) to the disc limiter. It can be configured to provide to the nator.

Specifically, in step S530, the apparatus for learning a medical image may filter the medical image generated in step S520. Specifically, an operation of checking the second medical image generated in the learning process of the lesion region detection learning model and selecting the second medical image may be performed. Preferably, the apparatus for learning a medical image may check the histogram of the first medical image and the histogram of the second medical image, compare them, and filter only the image corresponding to the first medical image. For example, the medical image learning apparatus configures at least one reference histogram information by using the histogram for the first medical image, compares the at least one reference histogram information with the histogram for the second medical image, The second medical image exceeding the reference histogram information may be determined as the filtered image. Similarly, the apparatus for learning a medical image may check the fourth medical image generated during the learning process of the lesion diagnosis learning model, and may perform an operation of selecting the fourth medical image. Preferably, the apparatus for learning a medical image may check the histogram of the third medical image and the histogram of the fourth medical image, compare them, and filter only the image corresponding to the third medical image. For example, the medical image learning apparatus configures at least one reference histogram information using the histogram for the third medical image, compares the at least one reference histogram information with the histogram for the fourth medical image, The fourth medical image exceeding the reference histogram information may be determined as the filtered image.

Thereafter, the apparatus for learning a medical image may configure and provide the determined image, that is, the filtered image as learning data (S540). In response to this, the apparatus for learning a medical image may proceed to step S510 to store the filtered image in the learning data set.

Based on the above-described structure, the learning data set may include the first medical image, the third medical image, the first labeling data, the second labeling data, the filtered second medical image, and the filtered fourth medical image. In addition, the apparatus for learning a medical image may perform discreminator learning of a lesion learning model using data included in the training data set.

Referring to FIG. 6, the computing system 1000 includes at least one processor 1100, a memory 1300, a user interface input device 1400, a user interface output device 1500, and storage connected through a bus 1200. (1600), and a network interface (1700).

The processor 1100 may be a central processing unit (CPU) or a semiconductor device that processes instructions stored in the memory 1300 and/or the storage 1600. The memory 1300 and the storage 1600 may include various types of volatile or nonvolatile storage media. For example, the memory 1300 may include read only memory (ROM) and random access memory (RAM).

Accordingly, the steps of the method or algorithm described in connection with the embodiments disclosed herein may be directly implemented in hardware executed by the processor 1100, a software module, or a combination of the two. The software module resides in a storage medium such as RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM (i.e., memory 1300 and/or storage 1600). You may. An exemplary storage medium is coupled to the processor 1100, which can read information from and write information to the storage medium. Alternatively, the storage medium may be integral with the processor 1100. The processor and storage media may reside within an application specific integrated circuit (ASIC). The ASIC may reside within the user terminal. Alternatively, the processor and storage medium may reside as separate components within the user terminal.

Although the exemplary methods of the present disclosure are expressed as a series of operations for clarity of description, this is not intended to limit the order in which steps are performed, and each step may be performed simultaneously or in a different order if necessary. In order to implement the method according to the present disclosure, the exemplary steps may include additional steps, other steps may be included excluding some steps, or may include additional other steps excluding some steps.

Various embodiments of the present disclosure are not intended to list all possible combinations, but to describe representative aspects of the present disclosure, and matters described in the various embodiments may be applied independently or may be applied in combination of two or more.

In addition, various embodiments of the present disclosure may be implemented by hardware, firmware, software, or a combination thereof. For implementation by hardware, one or more ASICs (Application Specific Integrated Circuits), DSPs (Digital Signal Processors), DSPDs (Digital Signal Processing Devices), PLDs (Programmable Logic Devices), FPGAs (Field Programmable Gate Arrays), general purpose It may be implemented by a processor (general processor), a controller, a microcontroller, a microprocessor, or the like.

The scope of the present disclosure is software or machine-executable instructions (e.g., operating systems, applications, firmware, programs, etc.) that cause an operation according to the method of various embodiments to be executed on a device or computer, and such software or It includes a non-transitory computer-readable medium (non-transitory computer-readable medium) which stores instructions and the like and is executable on a device or a computer.

Claims

In the apparatus for learning a medical image learning model,

A learning data set management unit that manages a learning data set including a first medical image, labeling data, and a filtered medical image,

Using a generator learning unit that manages learning of a generator that generates a second medical image, and a medical image filtered from the first medical image, labeling data, and the second image generated through the generator, the Including a lesion learning unit having a discreminator learning unit for managing learning of a discriminator constituting labeling data,

The learning data set management unit,

And a second medical image filtering unit that filters the second medical image and selectively provides the filtered medical image as the filtered medical image.
The method of claim 1

The learning data set management unit,

GAN-based medical image learning apparatus, characterized in that it further comprises a histogram check unit for checking histogram information on the first and second medical images.
The method of claim 2,

The second medical image filtering unit,

A GAN-based medical image learning apparatus, characterized in that the filtered medical image is determined from among the second medical images based on histogram information on the first and second medical images.
The method of claim 2,

The second medical image filtering unit,

Using the histogram information for the first medical image, construct reference histogram information,

The GAN-based medical image learning apparatus, characterized in that the filtered medical image is determined by comparing histogram information on the second medical image with the reference histogram information.
The method of claim 2,

The learning data set management unit,

GAN-based medical image learning apparatus, characterized in that it further comprises an error image check unit to check the error of the first medical image.
The method of claim 5,

The learning data set management unit,

GAN-based medical image learning apparatus, characterized in that it further comprises an error image restoration unit for restoring an error image using the second medical image, based on histogram information on the first and second medical images.
The method of claim 1,

The learning data set management unit,

Construct a supervised learning-based learning model using the first medical image and labeling data corresponding to the first medical image,

Generate labeling data corresponding to the first medical image using the supervised learning-based learning model,

The GAN-based medical image learning apparatus, characterized in that the generated labeling data having a probability score equal to or greater than a predetermined threshold is set and stored as the first medical image.
In the method of learning a medical image learning model,

A process of managing and storing a learning data set including a first medical image, labeling data, and a filtered medical image,

A generator that generates a second medical image, and a discreminator that configures the labeling data by using the medical image filtered from the first medical image, the labeling data, and the second image generated through the generator. Including the process of performing the learning of the lesion learning model including (Discriminator),

The process of performing the learning of the lesion learning model,

Including the process of detecting and providing a second medical image generated by the generator,

The process of managing and storing the training data set,

And a process of filtering the second medical image and selectively storing and managing it as the filtered medical image.
According to claim 8

The process of managing and storing the training data set,

And checking histogram information on the first and second medical images.
The method of claim 9,

The process of filtering the second medical image and selectively storing and managing it as the filtered medical image,

And determining the filtered medical image from among the second medical images based on histogram information on the first and second medical images.
The method of claim 9,

The process of filtering the second medical image and selectively storing and managing it as the filtered medical image,

A process of constructing reference histogram information by using the histogram information for the first medical image, and

And determining the filtered medical image by comparing histogram information on the second medical image with the reference histogram information.
The method of claim 9,

The process of managing and storing the training data set,

GAN-based medical image learning method, characterized in that it further comprises a process of checking the error of the first medical image.
The method of claim 12,

The process of managing and storing the training data set,

The GAN-based medical image learning method, further comprising a process of restoring an error image using the second medical image, based on histogram information on the first and second medical images.
The method of claim 8,

The process of managing and storing the training data set,

A process of constructing a supervised learning-based learning model using the first medical image and labeling data corresponding to the first medical image,

A process of generating labeling data corresponding to the first medical image using the supervised learning-based learning model, and

And setting and storing the generated labeling data having a probability score equal to or greater than a predetermined threshold value as the first medical image.