System for reducing noise of CBCT image and method therefor

The noise reduction system for CBCT images employs Cycle GAN and AI learning to address noise and posture differences, achieving high-quality image conversion and accurate segmentation, addressing the limitations of existing CBCT image analysis.

WO2024248517A1PCT designated stage expired Publication Date: 2024-12-05IND ACADEMIC COOP FOUND YONSEI UNIV
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
PCT/KR2024/007431
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-06-01
Filing Date
2024-05-31
Publication Date
2024-12-05

AI Technical Summary

Technical Problem

CBCT images suffer from noise and low image quality due to artifacts like scattering and beam hardening, making it difficult for medical analysts to accurately determine body shape and airflow, especially in areas like the nasal cavity and airway, and existing noise filtering techniques result in blurry images and long calculation times.

Method used

A noise reduction system using Cycle GAN technology to convert low-quality CBCT image data to the level of CT image data, incorporating artificial intelligence learning to account for posture differences between CBCT and CT imaging, and a regression analysis to correct shape deformation, thereby improving image quality and segmentation accuracy.

Benefits of technology

The system effectively reduces noise in CBCT images to the quality level of CT images, enhancing image quality and segmentation accuracy, allowing for more precise analysis and surgical planning, with improved SSIM and PSNR scores and airway shape segmentation accuracy comparable to CT images.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure KR2024007431_05122024_PF_FP_ABST
    Figure KR2024007431_05122024_PF_FP_ABST
Patent Text Reader

Abstract

The present invention relates to a system for reducing noise of a CBCT image and a method therefor, wherein the system can reduce noise of a CBCT image by applying the difference between the CBCT image and a CT image to artificial intelligence learning by using CYCLE GAN so as to convert CBCT image data such that it has the same standards as CT image data and provide same, the CBCT image data having low quality and more noise than the CT image data.
Need to check novelty before this filing date? Find Prior Art

Description

Noise reduction system and method for CBCT images

[0001] The present invention relates to a noise reduction system and method for CBCT images, and more particularly, to a noise reduction system and method for CBCT images capable of reducing noise in CBCT images by applying differences between CBCT images and CT images to artificial intelligence learning.

[0002] In addition, the present invention relates to a noise reduction system and method for CBCT images, which can perform analysis of a desired human body region based on CT image shooting conditions using CBCT images with reduced noise by applying differences in deformation of the shape of internal / external soft tissues of the human body due to differences in posture during the process of taking CBCT images and CT images to artificial intelligence learning.

[0003]

[0004] When a patient visits the hospital, medical staff, including doctors, prescribe imaging tests such as CT (Computed Tomography) to accurately understand the internal and external conditions of each patient's body and establish a surgical plan.

[0005] In particular, imaging is essential for internal structures of the human body, such as the nasal cavity and airway, which cannot be seen with the naked eye.

[0006] After this, the video recording results are analyzed to measure the shape of the internal structure of the human body, or the shape of the internal structure of the human body (such as the airway) is extracted, and used to compare and analyze the airflow state of a normal person and the airflow state of a patient through fluid simulation.

[0007]

[0008] Recently, not only CT but also CBCT (Cone Beam Computed Tomography) imaging is being performed a lot. CBCT has the advantage of reducing the burden on patients regarding radiation because the equipment cost is lower than CT and it can capture a wide area with a small amount of radiation.

[0009] However, since it captures a wide area with less radiation, there is a problem that the image has more noise and lower image quality compared to CT image data.

[0010] Examples of noise commonly observed in CBCT image data include scattering artifacts, which occur when the source radiated to the patient is not properly transmitted and is scattered, and beam hardening artifacts, which occur due to rapid intensity changes near areas with very high intensity, such as bones or teeth.

[0011] This noise makes it difficult for analysts (medical staff, observers, etc.) of the imaging results to properly understand the shape and area of ​​the shape, making it difficult to establish a medical plan.

[0012] To solve the problems of these CBCT image data, a noise filtering technique has been proposed. However, this technique produces blurry and hazy images and requires a long computational time for noise filtering, so it has limitations in applying it to all medical fields.

[0013]

[0014] Moreover, CT scans are typically performed with the patient lying on the medical equipment, with the bed itself being inserted into the medical equipment, while CBCT scans are performed with the patient standing or sitting, depending on the body part to be imaged. Therefore, when imaging areas such as the nasal cavity or airway, there is a problem in that it is difficult to determine the shape of the patient when lying down through CBCT image data.

[0015]

[0016] In this regard, Korean Patent Publication No. 10-2438068 (“Method and device for improving image quality using mutual information-based non-local global deformation reduction technique”) discloses a technology for improving the performance of low-dose CBCT images.

[0017]

[0018] Korean Patent Publication No. 10-24380686 (registration date: August 25, 2022)

[0019]

[0020] The present invention has been devised to solve the above-mentioned problems, and provides a CBCT image noise reduction system and method capable of reducing noise in a CBCT image by applying the difference between a CBCT image and a CT image to artificial intelligence learning using Cycle GAN so that CBCT image data containing a lot of noise and having low quality compared to CT image data can be converted to the level of CT image data and provided.

[0021] In addition, by applying the difference in shape deformation of internal / external soft tissues of the human body due to the difference in posture during the process of taking CBCT images and CT images to artificial intelligence learning, a noise reduction system and method for CBCT images are provided that can perform analysis of a desired human body area based on CT image shooting conditions using CBCT images with reduced noise.

[0022]

[0023] In order to solve the above-described problem, the noise reduction system of the CBCT image according to the present invention preferably includes an image input unit for receiving CBCT (Cone Beam Computed Tomography) image data photographing a body part of a subject, an image inference unit for inputting the CBCT image data by the image input unit into a stored first artificial intelligence model, converting and outputting the data into virtual image data with reduced noise, and an image analysis unit for analyzing the virtual image data output from the image inference unit and performing image segmentation corresponding to the body part.

[0024] Furthermore, it is preferable that the noise reduction system of the CBCT image further include a shape measurement unit that measures shape factor information of a corresponding body part from the segmentation result performed by the image analysis unit, taking into account shape factor information of preset items for each body part.

[0025] Furthermore, it is preferable that the noise reduction system of the CBCT image further includes a shape measurement correction unit that inputs preset items of the subject's biometric information input from the outside and shape factor information by the shape measurement unit into a stored second artificial intelligence model, and outputs the correction as shape factor information having the domain characteristics of CT (Computed Tomography) image data.

[0026] Furthermore, in order to store the first artificial intelligence model, the noise reduction system of the CBCT image preferably includes an image collection unit that collects CBCT image data and CT image data, each of which is photographed of the same body part for a plurality of subjects, an image preprocessing unit that processes the CBCT image data and CT image data by the image collection unit into pairs for each subject, and matches the photographed area of ​​the corresponding CBCT image data and CT image data with the coordinate center point of the image data based on each pair, and a first learning processing unit that generates a learning data set with the CBCT image data and CT image data by the image preprocessing unit and performs learning processing based on a pre-stored Cycle GAN (Generative Adversarial Network), and it is preferable that a generator network that has been trained by the first learning processing unit is stored as the first artificial intelligence model of the image inference unit.

[0027] Furthermore, the image preprocessing unit preferably further includes a first preprocessing unit that matches the feature points of the CBCT image data and the CT image data based on the CBCT image data or the CT image data by the image collecting unit by considering at least one feature point preset for each body part, a second preprocessing unit that extracts specific slice data matching the shooting area of ​​the CBCT image data using the CT image data whose feature points are matched to the CBCT image data by the first preprocessing unit, and a third preprocessing unit that performs masking processing of a preset size and shape on the specific slice data of the CT image data extracted by the second preprocessing unit, and the first learning processing unit preferably uses the CBCT image data by the first preprocessing unit and the CT image data by the third preprocessing unit.

[0028] Furthermore, in order to store the second artificial intelligence model, the noise reduction system for the CBCT image further includes an information input unit that receives biometric information of preset items for a plurality of subjects, shape factor information extracted by analyzing CBCT image data photographed of a specific body part, and shape factor information extracted by analyzing CT image data photographed of the same body part as the CBCT image data, and a second learning processing unit that processes the biometric information, shape factor information of the CBCT image data, and shape factor information of the CT image data by the information input unit in pairs for each subject to generate a learning data set, and performs a regression analysis on the relationship between the shape factor information of the CBCT image data and the shape factor information of the CT image data based on the biometric information using a pre-stored artificial intelligence network, and it is preferable that the regression analysis result model by the second learning processing unit is stored as the second artificial intelligence model of the shape measurement correction unit.

[0029] Furthermore, in order to solve the above-described problem, it is preferable to include a noise reduction method of a CBCT image using a noise reduction system of a CBCT image, in which each step is performed by an operation processing means, the method comprising an image input step of receiving CBCT (Cone Beam Computed Tomography) image data of a body part of a subject in an image input unit, an image inference step of converting and outputting virtual image data with reduced noise by inputting the CBCT image data by the image input step using a stored first artificial intelligence model in an image inference unit, an image analysis step of analyzing the virtual image data output by the image inference step in an image analysis unit to perform image segmentation corresponding to the body part, and a shape measurement step of measuring shape factor information of a corresponding body part from a segmentation result by the image analysis step by considering shape factor information of items preset for each body part in a shape measurement unit, and providing the same to the outside.

[0030] Furthermore, it is preferable that the noise reduction method of the CBCT image further includes a shape measurement correction step in which, after performing the shape measurement step, a shape measurement correction unit inputs preset items of the subject's biometric information input from the outside and shape factor information obtained by the shape measurement step using a stored second artificial intelligence model, and outputs the correction as shape factor information having the domain characteristics of CT (Computed Tomography) image data.

[0031] Furthermore, the method for reducing noise in the CBCT image includes, before performing the image inference step, an image collection step in which, in an image collection unit, CBCT image data and CT image data, each of which is photographed of the same body part for a plurality of subjects, are collected in order to store the first artificial intelligence model, an image preprocessing step in which, in an image preprocessing unit, the CBCT image data and CT image data obtained by the image collection step are paired for each subject and the corresponding CBCT image data and CT image data are matched with the shooting area and the coordinate center point of the image data based on each pair, and a first learning processing step in which, in a first learning processing unit, a learning data set is generated with the CBCT image data and CT image data obtained by the image preprocessing step and learning processing is performed based on a pre-stored Cycle GAN (Generative Adversarial Network), and it is preferable that a generator network obtained by the learning process in the first learning processing step is stored as the first artificial intelligence model obtained by the image inference step.

[0032] Furthermore, the image preprocessing step preferably further includes a first preprocessing step of matching the feature points of the CBCT image data and the CT image data based on the CBCT image data or the CT image data by the image collection step by considering at least one feature point preset for each body part, a second preprocessing step of extracting specific slice data matching the shooting area of ​​the CBCT image data using the CT image data whose feature points are matched to the CBCT image data by the first preprocessing step, and a third preprocessing step of performing masking processing of a preset size and shape on the specific slice data of the CT image data extracted by the second preprocessing step, and the first learning processing step preferably uses the CBCT image data by the first preprocessing step and the CT image data by the third preprocessing step.

[0033] Furthermore, the noise reduction method of the CBCT image further includes an information input step in which, before performing the shape measurement correction step, an information input unit receives, in order to store the second artificial intelligence model, biometric information of preset items for a plurality of subjects, shape factor information extracted by analyzing CBCT image data photographed of a specific body part, and shape factor information extracted by analyzing CT image data photographed of the same body part as the CBCT image data, and a second learning processing unit processes the biometric information, shape factor information of the CBCT image data, and shape factor information of the CT image data by the information input step in pairs for each subject to generate a learning data set, and performs a regression analysis on the relationship between the shape factor information of the CBCT image data and the shape factor information of the CT image data based on the biometric information using a pre-stored artificial intelligence network, and it is preferable that a regression analysis result model by the second learning processing step is stored as the second artificial intelligence model of the shape measurement correction step.

[0034]

[0035] According to the noise reduction system and method of the CBCT image according to the present invention as described above, since the equipment cost is lower than that of CT scanning and a wide area can be captured, the burden on patients regarding radiation is reduced, and there is an advantage in that the Cycle GAN technology is applied to input CBCT image data and output virtual image data corresponding to the image quality level of CT image data so as to resolve image quality issues (artifacts, noise, etc.) of widely used CBCT image data.

[0036] In addition, according to the noise reduction system and method of the CBCT image according to the present invention as described above, in order to correct for the occurrence of deformation of the shape of soft tissue inside / outside the body due to the difference in the physical condition of the subject under the CBCT image shooting condition and the CT image shooting condition, there is an advantage in that the degree of difference between the two image shooting data can be predicted by performing regression analysis of the two image shooting data based on the subject's biometric information.

[0037] Through this, the noise reduction system and method of the CBCT image according to the present invention as described above can output virtual image data with high image quality, more specifically, with noise reduced to the level of image quality of CT image data, through CBCT image data, and has the advantage of being able to correct shape factor information measured through the virtual image data according to the domain characteristics (soft tissue shape deformation characteristics) of CT image data.

[0038]

[0039] FIG. 1 is a schematic diagram illustrating a noise reduction system for a CBCT image according to one embodiment of the present invention.

[0040] FIG. 2 is a flowchart illustrating a method for reducing noise in a CBCT image according to one embodiment of the present invention.

[0041]

[0042] The most preferred embodiment according to the present invention is characterized by including: an image input unit for receiving CBCT (Cone Beam Computed Tomography) image data of a body part of a subject; an image inference unit for inputting the CBCT image data by the image input unit into a stored first artificial intelligence model, converting and outputting the CBCT image data into virtual image data with reduced noise; and an image analysis unit for analyzing the virtual image data output from the image inference unit and performing image segmentation corresponding to the body part.

[0043]

[0044] The above-described objects, features, and advantages of the present invention will become more apparent through the following examples taken in conjunction with the accompanying drawings. The specific structural and functional descriptions below are merely illustrative for the purpose of explaining embodiments according to the concept of the present invention, and embodiments according to the concept of the present invention may be implemented in various forms and should not be construed as limited to the embodiments described in this specification or application. Since embodiments according to the concept of the present invention may have various modifications and may take various forms, specific embodiments are illustrated in the drawings and described in detail in this specification or application. However, this is not intended to limit embodiments according to the concept of the present invention to specific disclosed forms, but should be understood to include all modifications, equivalents, and substitutes included in the spirit and technical scope of the present invention. Terms such as first and / or second may be used to describe various components, but the components are not limited to the terms. The above terms are used solely for the purpose of distinguishing one component from another, for example, without departing from the scope of the rights according to the concept of the present invention, a first component may be referred to as a second component, and similarly, a second component may also be referred to as a first component. When a component is referred to as being connected or coupled to another component, it should be understood that it may be directly connected or coupled to the other component, but there may also be other components in between. Conversely, when a component is referred to as being directly connected or coupled to another component, it should be understood that there are no other components in between. Other expressions used to describe the relationship between components, such as "between" and "directly between" or "adjacent to" and "directly adjacent to", should be interpreted similarly.The terminology used herein is used solely to describe specific embodiments and is not intended to limit the present invention. The singular expression includes the plural expression unless the context clearly dictates otherwise. It should be understood that the terms "comprise" and "have" used herein indicate the presence of a described feature, number, step, operation, component, part, or combination thereof, and do not preclude the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof. Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention pertains. Terms defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning they have in the context of the relevant technology, and shall not be interpreted in an idealized or overly formal sense unless explicitly defined herein. Hereinafter, the present invention will be described in detail by describing a preferred embodiment of the present invention with reference to the accompanying drawings. The same reference numerals in each drawing represent the same parts.

[0045] In addition, a system refers to a set of components including devices, mechanisms, and means that are organized and interact regularly to perform necessary functions.

[0046]

[0047] A noise reduction system and method for a CBCT image according to one embodiment of the present invention have a technical purpose of reducing noise included in CBCT image data that has been widely used recently.

[0048] At this time, it is desirable to use an artificial intelligence network to learn the degree of noise reduction included in CBCT image data based on CT image data.

[0049] In addition, considering that the posture of a patient taking a CBCT image and the posture of a patient taking a CT image may be different, it is preferable to learn the difference between the two shooting postures using an artificial intelligence network according to an embodiment of the present invention and the method for reducing noise in a CBCT image.

[0050]

[0051] Comparing the inherent differences between CBCT and CT imaging data, first, CT and CBCT imaging devices have different configurations, making simultaneous imaging impossible. Typically, CBCT images are taken with the subject (e.g., a patient) standing or sitting upright, while CT images are taken with the subject lying on a bed.

[0052] Additionally, since CBCT images are the result of capturing a wide area with less radiation than CT images, they contain more artifacts and noise than CT images.

[0053] Accordingly, although it has the advantage of being able to reduce the burden on the subject of radiation by being able to capture a wide area and having lower costs for imaging devices (equipment, etc.) than CT scanning, there is a problem that its use is limited due to many artifacts and noise in the image.

[0054] Accordingly, the technical purpose of the present invention is to reduce noise contained in CBCT image data, which has been widely used recently, to a level comparable to the quality of CT image data, so that CBCT image data can be sufficiently utilized in various medical fields.

[0055] In addition, the noise reduction system and method of CBCT images according to one embodiment of the present invention can be applied to numerous and diverse body parts without limiting the area of ​​the patient's body that the medical staff wishes to analyze for smooth explanation.

[0056]

[0057] FIG. 1 is a diagram illustrating a configuration of a noise reduction system for a CBCT image according to one embodiment of the present invention.

[0058] As illustrated in FIG. 1, a noise reduction system for a CBCT image according to one embodiment of the present invention includes an image input unit (100), an image inference unit (200), and an image analysis unit (300). It is preferable that each component performs operations through an operation processing means including a computer.

[0059]

[0060] Let's take a closer look at each component:

[0061] It is preferable that the above image input unit (100) receives CBCT (Cone Beam Computed Tomography) image data that has captured a body part of the subject.

[0062] At this time, the CBCT image data is image data that is likely to contain many artifacts, noise, etc., as described above.

[0063]

[0064] The above image inference unit (200) is configured to perform noise reduction of the CBCT image data input through the image input unit (100).

[0065] In detail, it is preferable that the image inference unit (200) inputs CBCT image data by the image input unit (100) using the stored first artificial intelligence model and converts and outputs virtual image data with reduced noise.

[0066] At this time, the image inference unit (200) does not remove noise included in the CBCT image data input through the image input unit (100), but rather generates virtual image data with reduced noise by reflecting the CBCT image data.

[0067]

[0068] In order to store the above first artificial intelligence model, it is preferable that the noise reduction system of a CBCT image according to one embodiment of the present invention further includes an image collection unit (10), an image preprocessing unit (20), and a first learning processing unit (30), as illustrated in FIG. 1.

[0069] Through the operations of the image collection unit (10), image preprocessing unit (20), and first learning processing unit (30), the first artificial intelligence model is created.

[0070] Considering that the image quality of the CBCT image data cannot be improved infinitely in reality and that the CT image data has a higher image quality than the CBCT image data, the first artificial intelligence model receives the CBCT image data as input and performs learning processing to output virtual image data with reduced noise to the image quality level of the CT image data.

[0071]

[0072] In detail, it is preferable that the image collection unit (10) collects CBCT image data and CT image data for each of the same body parts taken for a plurality of subjects.

[0073] At this time, since the CBCT image data and CT image data are medical image data, they are naturally input and collected through separate linked organizations, servers, means, etc.

[0074] In addition, it is desirable for the image collection unit (10) to collect CBCT image data and CT image data of the same body part for as many subjects as possible.

[0075] Since a unique imaging prescription is usually made for each specific body part of the human body, it is preferable that the image collection unit (10) collects image data of a subject who has taken both the CBCT image data and the CT image data among the imaging result data of a subject who has been prescribed imaging due to the same or similar disease or injury.

[0076] Of course, even if it is not a patient group for which imaging prescription is to be made, it is possible to collect CBCT image data and CT image data by collecting a random group of subjects and taking images of each body part for which imaging prescription is to be made.

[0077]

[0078] The image preprocessing unit (20) processes the CBCT image data and CT image data by the image collection unit (10) in pairs for each subject, and it is preferable to match the shooting area of ​​the corresponding CBCT image data and CT image data with the coordinate center point of the image data based on each pair.

[0079] This is because the CBCT image data and CT image data for each subject collected by the image collection unit (10) are captured at different times using different imaging devices (equipment, etc.), so naturally the coordinate center point of each image data is different and the shooting area is also different. In order to solve this problem, the image preprocessing unit (20) performs an operation to match the ROI (Region Of Interest) of each image data.

[0080] In other words, while the body area being imaged is essentially the same, each image is captured using different imaging devices. Consequently, the coordinate centers for the CBCT and CT image data differ, and the imaged areas do not match. Applying this issue to AI learning without addressing it can lead to issues such as reduced accuracy and reliability.

[0081] To solve this, the image preprocessing unit (20) includes a first preprocessing unit (21), a second preprocessing unit (22), and a third preprocessing unit (23), as shown in FIG. 1.

[0082] Since it is clear that the subject cannot assume the same posture in relation to the CBCT imaging device and the CT imaging device, the position, size, angle, etc. are matched through the first preprocessing unit (21).

[0083] In detail, it is preferable that the first preprocessing unit (21) matches the feature points of the CBCT image data and the CT image data based on the CBCT image data or CT image data by the image collection unit (10) by taking into account at least one feature point preset for each body part.

[0084] As described above, when the CBCT image data and CT image data are taken of a body part corresponding to the head and neck, it is preferable that the preset characteristic points include N (Nasion, the anteriormost point of the frontonasal suture), PNS (Posterior Nasal Spine, the posteriormost point of the hard palate), and Pog (Pogonion, the anteriormost point of the mandibular prominence).

[0085] It is natural that these characteristic points are set differently for each body part, and the location of the characteristic points themselves is not limited.

[0086] However, the first preprocessing unit (21) extracts the feature points appearing in the CBCT image data by the image collection unit (10) and the feature points appearing in the CT image data by utilizing a separate medical image analysis algorithm or from an external medical expert (medical staff, etc.).

[0087] After this, it is preferable that the first preprocessing unit (21) matches the feature points of the extracted CBCT image data with the feature points of the extracted CT image data based on the CBCT image data or the CT image data.

[0088] That is, it is desirable to align the feature points of the extracted CT image data so that they match at the same location based on the feature points of the extracted CBCT image data, or to align the feature points of the extracted CBCT image data so that they match at the same location based on the feature points of the extracted CT image data.

[0089] At this time, it is preferable that the first preprocessing unit (21) performs an operation using the CBCT image data and CT image data that are paired for each subject.

[0090]

[0091] It is preferable that the second preprocessing unit (22) extracts specific slice data matching the shooting area of ​​the CBCT image data by using the CT image data whose feature points are matched with the CBCT image data by the first preprocessing unit (21).

[0092] That is, it is desirable to use only specific slice data of the CT image data by cutting it out so that the photographed areas of the CBCT image data and the CT image data are the same.

[0093] Through this, for each subject, the paired CBCT image data and CT image data can be matched in terms of location, size, angle, and even the photographed area.

[0094]

[0095] In addition, since CBCT images typically have a different type of ROI from CT images, it is preferable to perform masking processing, such as the ROI of the CBCT image data, on specific slice data of the image data extracted by the second preprocessing unit (22) through the third preprocessing unit (23).

[0096] That is, it is preferable that the third preprocessing unit (23) performs circular masking processing with a preset size and shape, for example, a diameter equal to the number of pixels of the image, on specific slice data of the image data extracted by the second preprocessing unit (22).

[0097]

[0098] It is preferable that the above first learning processing unit (30) generates a learning data set using the CBCT image data and CT image data by the image preprocessing unit (20) and performs learning processing based on a pre-stored Cycle GAN (Generative Adversarial Network).

[0099] More specifically, the first learning processing unit (30) generates a learning data set using the CBCT image data by the first preprocessing unit (21) and the CT image data by the third preprocessing unit (23), and performs learning processing based on a pre-stored Cycle GAN.

[0100]

[0101] In addition, to explain 'Cycle GAN', Cycle GAN is a method of training a GAN model by circulating a generator created from an image data set that does not correspond one-to-one.

[0102] Here, GAN is a generative adversarial network, a model in which a generator and an identifier compete with each other to generate data.

[0103] That is, the generator creates a virtual image, and the discriminator evaluates whether the input image is virtual or real. The generator and discriminator learn by competing with each other and gradually improving each other's performance.

[0104] At this time, Cycle GAN learns a method to convert images between two domains by using an X domain data set (a domain corresponding to CBCT image data in the present invention) and a Y domain data set (a domain corresponding to CT image data in the present invention) rather than using a one-to-one corresponding image data set.

[0105] For two domain images, generator G generates image Y from image X, and generator F generates image X from image Y.

[0106] In the process, discriminators D1 and D2 learn whether each generated image is true or false.

[0107] In addition, by connecting the generator G and the generator F, it generates X image -> Y image (transformation) -> X' image (inverse transformation), and finally learns with the consistency loss values ​​of the X image and X' image.

[0108] Loss functions applied to learning include Cycle consistency loss and Adversarial loss. Cycle consistency loss is a loss function for generators G and F, and Adversarial loss is a loss function for discriminators D1 and D2.

[0109] In the case of the discriminator's loss function, it is desirable to train so that the discriminator loss does not decrease, because a small value means that learning is not possible.

[0110] In addition, the generator G computed a pixel-level loss function using the Mean Absolute Error (MAE) to compare the degree of misprediction between the generated image and the corresponding correct image, and computed a feature-level loss function using the Visual Geometry Group (VGG).

[0111] When applying this Cycle GAN to the present invention, the first generator learns to convert CBCT image data into CT image data, and the second generator performs reverse learning to inversely convert CT image data into CBCT image data.

[0112] At this time, the discriminators connected to each generator determine whether the image is an actual image or a transformed image.

[0113] Through this learning process, the first generator that has undergone the final learning process is the first artificial intelligence model, and is stored in the image inference unit (200) of the present invention to perform operations.

[0114] By applying Cycle GAN having the above-described technical features, the first artificial intelligence model is trained by receiving CBCT image data by the first preprocessing unit (21) and CT image data by the third preprocessing unit (23) for each subject as a learning data set by the first learning processing unit (30).

[0115]

[0116] After learning is completed, in other words, after final learning is performed, the image inference unit (200) receives CBCT image data by the image input unit (100) and inputs it into the first artificial intelligence model, and the first artificial intelligence model uses this to convert and output virtual CT image data, that is, virtual image data with reduced noise to the image quality level of the CT image data.

[0117] When comparing the virtual image data output through the first artificial intelligence model, that is, the virtual image data with reduced noise to the image quality level of the CT image data, using SSIM (Structural Similarity Index Map) and PSNR (Peak Signal-to-Noise Ratio), which are quantitative indicators of image quality, the CBCT image data, which is the input data, had SSIM: 0.6446, PSNR: 22.4351, but the output data, the virtual image data with reduced noise, had SSIM: 0.7179, PSNR: 23.1368, showing an improvement in quality.

[0118] In addition, it was confirmed that the virtual image data with reduced noise output through the image inference unit (200) improved the airway shape segmentation accuracy (Dice score) from 0.838 of the input data (CBCT image data) to 0.943, and through this, when applied to airway shape segmentation, complex and fine duct shapes such as the nasal cavity can be better identified.

[0119] The airway shape segmentation accuracy shown by this noise-reduced virtual image data is judged to be comparable to the airway shape segmentation accuracy of 0.955 shown when the airway shape is segmented by CT image data taken with a general dose.

[0120]

[0121] Considering this, it is preferable that the image analysis unit (300) analyzes the virtual image data output from the image inference unit (200) and performs image segmentation corresponding to the body part.

[0122] In summary, the noise reduction system of a CBCT image according to one embodiment of the present invention receives the CBCT image data of a subject, converts it into virtual image data with reduced noise, outputs it, and performs image segmentation corresponding to a body part using the same, but has the advantage of being able to perform segmentation of a desired body part with an accuracy similar to that of CT image data taken with a general dose.

[0123]

[0124] A noise reduction system for a CBCT image according to one embodiment of the present invention preferably further includes a shape measurement unit (400) and a shape measurement correction unit (500), as illustrated in FIG. 1, in order to predict and correct the degree of deformation caused between the CBCT image data and the photographing of the subject of the CT image data.

[0125] As described above, CBCT image data is generally captured while the subject is standing or sitting, but CT image data is captured while lying down, so shape deformation of soft tissues inside and outside the body occurs depending on the direction of gravity and stiffness.

[0126] Therefore, even if the virtual image data with reduced noise of the CBCT image data is output through the image inference unit (200) described above, this is only the reduced noise of the image data of the standing subject, and there will be a difference from the shape of the soft tissue inside / outside the body included in the image data of the lying subject.

[0127] Accordingly, it is desirable to perform an operation of predicting and correcting body shape deformation between CBCT image data and CT image data through the shape measurement unit (400) and shape measurement correction unit (500).

[0128]

[0129] It is preferable that the shape measurement unit (400) measures the shape factor information of the corresponding body part from the segmentation result performed by the image analysis unit (300) by considering the shape factor information of the items preset for each body part.

[0130] When the body part is set to the airway, the shape factor information of the preset items includes AN (anteroposterior width at the nasopharynx), LM (lateral width at the minimum area), AM (anteroposterior width at the minimum area), and LNM (length from the nasopharynx to the minimum area).

[0131] Typically, medical professionals establish surgical plans based on measured shape factor information rather than segmented images.

[0132] The shape factor information measured by the above shape measurement unit (400) is measured through virtual image data that has accuracy similar to that of CT image data when judged only in terms of the image quality level, but corresponds to the shape measurement value of soft tissue photographed under the shooting conditions of CBCT image data in a standing or sitting state, rather than the shape measurement value of soft tissue photographed in a lying state like CT image data.

[0133]

[0134] Of course, the shape factor information measured by the shape measurement unit (400) can also be sufficiently used by medical staff to establish a surgical plan.

[0135] However, in the case of image data captured before the noise reduction system of the CBCT image according to one embodiment of the present invention is applied, CBCT image data and CT image data are mixed, or in some cases, additional shooting data is required in the process of establishing a surgical plan based on CT image data, so CBCT image data is captured, or in cases where CT shooting itself cannot be easily performed due to the physical condition of the subject, the shape measurement values ​​of the captured soft tissue must be inferred under the shooting conditions of the CT image data using only the CBCT image data.

[0136]

[0137] To this end, it is preferable that the shape measurement correction unit (500) inputs the biometric information of the preset items of the subject input from the outside and the shape factor information by the shape measurement unit (400) into the stored second artificial intelligence model, and outputs the correction as shape factor information having the domain characteristics of CT image data.

[0138] That is, since the shape factor information by the shape measurement unit (400) is shape factor information measured from image data captured in a standing or sitting state, there may be a difference from the shape factor information measured from CT image data, which is image data captured in a lying state.

[0139] At this time, the difference that occurs is caused by the deformation of the shape of the soft tissue inside / outside the body depending on the direction of gravity and stiffness, and since this is greatly influenced by gender / age, etc., it is preferable that the shape measurement correction unit (500) include gender and age as the biometric information of the preset items of the subject.

[0140]

[0141] In order to store the second artificial intelligence model, the noise reduction system for a CBCT image according to one embodiment of the present invention further includes an information input unit (1000) and a second learning processing unit (2000), as illustrated in FIG. 1.

[0142] It is preferable that the above information input unit (1000) receives biometric information of preset items for a plurality of subjects, shape factor information extracted by analyzing CBCT image data photographed of a specific body part, and shape factor information extracted by analyzing CT image data photographed of the same body part as the CBCT image data.

[0143] That is, for each subject who has both CBCT image data and CT image data, the age, gender, shape factor information extracted by analyzing CBCT image data, and shape factor information extracted by analyzing CT image data are input.

[0144] For each subject, the CBCT image data and CT image data are taken of the same body part, and considering the shape factor information of the items set in advance for each body part, the CBCT image data and CT image data are analyzed to output the shape factor information of each.

[0145] In detail, after performing image shape segmentation of each CBCT image data and CT image data, shape factor information of the corresponding body part is measured.

[0146] The above information input unit (1000) may measure shape factor information by performing the above-described process by inputting the CBCT image data or CT image data, or may only receive shape measurement values ​​from an external configuration.

[0147]

[0148] The second learning processing unit (2000) processes the biometric information, shape factor information of CBCT image data, and shape factor information of CT image data by the information input unit (1000) in pairs for each subject to generate a learning data set.

[0149] After this, it is preferable that the second learning processing unit (2000) performs a regression analysis on the relationship between the shape factor information of the CBCT image data and the shape factor information of the CT image data based on the biometric information using a pre-stored artificial intelligence network.

[0150] It is preferable that the regression analysis result model by the second learning processing unit (2000) is stored as the second artificial intelligence model of the shape measurement correction unit (500).

[0151] That is, the second artificial intelligence model performs a regression analysis on the relationship between the shape factor information of the CBCT image data and the shape factor information of the CT image data for each gender and age.

[0152] Through this, when the shape measurement correction unit (500) inputs the gender and age of the subject and the shape factor information (shape factor information analyzed by converting and outputting virtual image data using CBCT image data by the image input unit (100)) by the shape measurement unit (400) into the second artificial intelligence model, the shape factor information is corrected and output with the domain characteristics of the CT image data, that is, the soft tissue characteristics of the shooting conditions (shooting in a lying down state) of the CT image data.

[0153]

[0154] FIG. 2 is a flowchart illustrating a method for reducing noise in a CBCT image according to one embodiment of the present invention.

[0155] As illustrated in FIG. 2, a noise reduction method for a CBCT image according to one embodiment of the present invention includes an image input step (S100), an image inference step (S200), an image analysis step (S300), and a shape measurement step (S400). It is preferable that each step utilize a noise reduction system for a CBCT image, which is operated by an operation processing means.

[0156]

[0157] Let's take a closer look at each step:

[0158] The above image input step (S100) receives CBCT (Cone Beam Computed Tomography) image data of a body part of the subject being photographed from the image input unit (100).

[0159] At this time, the CBCT image data is image data that is likely to contain many artifacts, noise, etc., as described above.

[0160]

[0161] The above image inference step (S200) uses the stored first artificial intelligence model in the image inference unit (200) to input CBCT image data from the image input step (S100) and convert it into virtual image data with reduced noise and output it.

[0162] At this time, the image inference step (S200) does not remove noise included in the CBCT image data input by the image input step (S100), but rather generates virtual image data with reduced noise by reflecting the CBCT image data.

[0163]

[0164] In order to perform such operations, the noise reduction method of a CBCT image according to one embodiment of the present invention further includes an image collection step (S10), an image preprocessing step (S20), and a first learning processing step (S30), as illustrated in FIG. 2, to store the first artificial intelligence model before performing the image inference step (S200).

[0165] Considering that the image quality of the CBCT image data cannot be improved infinitely in reality and that the CT image data has a higher image quality than the CBCT image data, the first artificial intelligence model receives the CBCT image data as input and performs learning processing to output virtual image data with reduced noise to the image quality level of the CT image data.

[0166]

[0167] In the above image collection step (S10), the image collection unit (10) collects CBCT image data and CT image data of the same body part for each of a plurality of subjects.

[0168] At this time, since the CBCT image data and CT image data are medical image data, they are naturally input and collected through separate linked organizations, servers, means, etc.

[0169] In addition, it is preferable that the image collection step (S10) collects CBCT image data and CT image data of the same body part for as many subjects as possible.

[0170] Since a unique imaging prescription is usually made for each specific body part of the human body, it is preferable that the image collection unit (10) collects image data of a subject who has taken both the CBCT image data and the CT image data among the imaging result data of a subject who has been prescribed imaging due to the same or similar disease or injury.

[0171] Of course, even if it is not a patient group for which imaging prescription is to be made, it is possible to collect CBCT image data and CT image data by collecting a random group of subjects and taking images of each body part for which imaging prescription is to be made.

[0172]

[0173] In the image preprocessing step (S20), the image preprocessing unit (20) processes the CBCT image data and CT image data obtained in the image collection step (S10) in pairs for each subject, and it is preferable to match the shooting area of ​​the corresponding CBCT image data and CT image data with the coordinate center point of the image data based on each pair.

[0174] This is because the CBCT image data and CT image data for each subject are captured at different times using different imaging devices (equipment, etc.), so naturally the coordinate center points of each image data are different, and the captured areas are also different. To solve this problem, the image preprocessing step (S20) performs an operation to match the ROI (Region Of Interest) of each image data.

[0175] In other words, although the body area being imaged is essentially the same, each image is captured using different imaging devices. Consequently, the coordinate centers for the CBCT and CT image data differ, and the imaged areas do not match. Applying this issue to AI learning without addressing it can lead to problems such as reduced accuracy and reliability.

[0176] To solve this, the image preprocessing step (S20) includes a first preprocessing step (S21), a second preprocessing step (S22), and a third preprocessing step (S23), as illustrated in FIG. 2.

[0177]

[0178] It is preferable that the first preprocessing step (S21) matches the feature points of the CBCT image data and the CT image data based on the CBCT image data or CT image data obtained through the image collection step (S10) by considering at least one feature point preset for each body part.

[0179] That is, since it is clear that the subject cannot assume the same posture with respect to the CBCT imaging device and the CT imaging device, the position, size, angle, etc. are matched through the first preprocessing step (S21).

[0180] As described above, when the CBCT image data and CT image data are taken of a body part corresponding to the head and neck, it is preferable that the preset characteristic points include N (Nasion, the anteriormost point of the frontonasal suture), PNS (Posterior Nasal Spine, the posteriormost point of the hard palate), and Pog (Pogonion, the anteriormost point of the mandibular prominence).

[0181] It is natural that these characteristic points are set differently for each body part, and the location of the characteristic points themselves is not limited.

[0182] However, the first preprocessing step (S21) utilizes a separate medical image analysis algorithm or extracts feature points appearing in the CBCT image data through the image collection step (S10) and feature points appearing in the CT image data from an external medical expert (medical staff, etc.).

[0183] After this, the first preprocessing step (S21) matches the extracted feature points of the CBCT image data with the extracted feature points of the CT image data based on the CBCT image data or the CT image data.

[0184] That is, it is desirable to align the feature points of the extracted CT image data so that they match at the same location based on the feature points of the extracted CBCT image data, or to align the feature points of the extracted CBCT image data so that they match at the same location based on the feature points of the extracted CT image data.

[0185] At this time, the first preprocessing step (S21) is performed using the CBCT image data and CT image data that are paired for each subject.

[0186]

[0187] The second preprocessing step (S22) extracts specific slice data matching the shooting area of ​​the CBCT image data by using the CT image data whose feature points are matched with the CBCT image data by the first preprocessing step (S21).

[0188] That is, it is desirable to use only specific slice data of the CT image data so that the captured areas of the CBCT image data and the CT image data are identical. Through this, the CBCT image data and CT image data that form a pair for each subject can be matched in terms of location, size, angle, and captured area.

[0189]

[0190] The third preprocessing step (S23) performs masking processing of a preset size and shape on specific slice data of the CT image data extracted by the second preprocessing step (S22).

[0191] Typically, CBCT images have a different ROI shape from CT images, so the third preprocessing step (S23) performs masking processing on specific slice data of the CT image data extracted by the second preprocessing step (S22) in the same manner as the ROI of the CBCT image data. For example, it is preferable to perform circular masking processing with a diameter equal to the number of pixels of the image.

[0192]

[0193] The first learning processing step (S30) generates a learning data set using the CBCT image data and CT image data obtained through the image preprocessing step (S20), more precisely, the CBCT image data obtained through the first preprocessing step (S21) and the CT image data obtained through the third preprocessing step (S23), in the first learning processing unit (30).

[0194] After this, learning processing is performed on the learning data set generated based on the pre-saved Cycle GAN (Generative Adversarial Network).

[0195] The first generator that has undergone the final learning process through the learning process is the first artificial intelligence model, and is stored as the first artificial intelligence model by the image inference step (S200) of the present invention to perform operations.

[0196]

[0197] That is, by applying Cycle GAN having the above-described technical characteristics, the first artificial intelligence model is trained by receiving CBCT image data from the first preprocessing step (S21) and CT image data from the third preprocessing step (S23) as input for each subject as a learning data set in the first learning processing step (S30).

[0198]

[0199] After learning is completed, in other words, after final learning is performed, the image inference step (S200) receives the CBCT image data from the image input step (S100) and inputs it into the first artificial intelligence model, and the first artificial intelligence model uses it to convert and output virtual CT image data, that is, virtual image data with reduced noise to the image quality level of the CT image data.

[0200] When comparing the virtual image data output through the first artificial intelligence model, that is, the virtual image data with reduced noise to the image quality level of the CT image data, using SSIM (Structural Similarity Index Map) and PSNR (Peak Signal-to-Noise Ratio), which are quantitative indicators of image quality, the CBCT image data, which is the input data, had SSIM: 0.6446, PSNR: 22.4351, but the output data, the virtual image data with reduced noise, had SSIM: 0.7179, PSNR: 23.1368, showing an improvement in quality.

[0201] In addition, it was confirmed that the virtual image data with reduced noise output through the image inference step (S200) improved the airway shape segmentation accuracy (Dice score) from 0.838 of the input data (CBCT image data) to 0.943, and through this, when applied to airway shape segmentation, complex and fine duct shapes such as the nasal cavity can be better identified.

[0202] The airway shape segmentation accuracy shown by this noise-reduced virtual image data is judged to be comparable to the airway shape segmentation accuracy of 0.955 shown when the airway shape is segmented by CT image data taken with a general dose.

[0203]

[0204] Taking these points into consideration, the image analysis step (S300) analyzes the virtual image data output by the image inference step (S200) in the image analysis unit (300) to perform image segmentation corresponding to the body part.

[0205] That is, even if the CBCT image data of the subject is input, converted into virtual image data with reduced noise, and image segmentation corresponding to a body part is performed using this, there is an advantage in that segmentation of the desired body part can be performed with an accuracy similar to that of CT image data taken with a general dose.

[0206]

[0207] The above shape measurement step (S400) measures the shape factor information of the corresponding body part from the segmentation result performed by the image analysis step (S300) by considering the shape factor information of the items preset for each body part in the shape measurement unit (400) and provides the information to an external party (medical staff, etc.).

[0208] When the body part is set to the airway, the shape factor information of the preset items includes AN (anteroposterior width at the nasopharynx), LM (lateral width at the minimum area), AM (anteroposterior width at the minimum area), and LNM (length from the nasopharynx to the minimum area).

[0209] The shape factor information measured by the above shape measurement step (S400) is measured through virtual image data that has accuracy similar to that of CT image data when judged only in terms of the image quality level, but corresponds to the shape measurement value of soft tissue captured under the shooting conditions of CBCT image data in a standing or sitting state, rather than the shape measurement value of soft tissue captured in a lying state like CT image data.

[0210] Typically, medical professionals establish surgical plans based on measured shape factor information rather than segmented images.

[0211] The shape factor information measured by the above shape measurement step (S400) can also be sufficiently used by medical staff to establish a surgical plan.

[0212] Typically, CBCT image data is captured while the subject is standing or sitting, but CT image data is captured while lying down, so depending on the direction of gravity and stiffness, shape deformation of the soft tissues inside and outside the body occurs.

[0213] Therefore, even if the virtual image data with reduced noise of the CBCT image data is output through the image inference step (S200) described above, this is only the reduced noise of the image data of the standing subject, and there will be a difference from the shape of the soft tissue inside / outside the body included in the image data of the lying subject.

[0214] When the noise reduction method of a CBCT image according to an embodiment of the present invention is actually applied, if CBCT image data and CT image data are mixed, or if additional shooting data is required in the process of establishing a surgical plan based on CT image data, and CBCT image data is shot, or if CT shooting itself cannot be easily performed due to the physical condition of the subject, the shape measurement values ​​of the soft tissue shot under the shooting conditions of the CT image data must be inferred using only CBCT image data.

[0215]

[0216] To this end, a process of predicting and correcting body shape deformation between CBCT image data and CT image data is required, and a noise reduction method of a CBCT image according to one embodiment of the present invention is performed through a shape measurement correction step (S500) illustrated in FIG. 2.

[0217] The shape measurement correction step (S500) is performed after the shape measurement step (S400) is performed, and then, in the shape measurement correction unit (500), the stored second artificial intelligence model is used to input the biometric information of the preset items of the subject input from the outside and the shape factor information by the shape measurement step (S400), and the shape factor information having the domain characteristics of CT image data is corrected and output.

[0218] That is, since the shape factor information by the shape measurement step (S400) is shape factor information measured from image data captured in a standing or sitting state, there may be a difference from the shape factor information measured from CT image data, which is image data captured in a lying state.

[0219] At this time, the difference that occurs is caused by the deformation of the shape of the soft tissue inside / outside the body depending on the direction of gravity and stiffness, and since this is greatly affected by gender / age, etc., the shape measurement correction step (S500) includes gender and age as the biometric information of the preset items of the subject.

[0220]

[0221] In a method for reducing noise in a CBCT image according to one embodiment of the present invention, before performing the shape measurement correction step (S500), an information input step (S1000) and a second learning processing step (S2000) are further performed, as illustrated in FIG. 2, to store the second artificial intelligence model.

[0222] The above information input step (S1000) receives, from the information input unit (1000), biometric information of preset items for a plurality of subjects, shape factor information extracted by analyzing CBCT image data photographed of a specific body part, and shape factor information extracted by analyzing CT image data photographed of the same body part as the CBCT image data.

[0223] That is, for each subject who has both CBCT image data and CT image data, the age, gender, shape factor information extracted by analyzing CBCT image data, and shape factor information extracted by analyzing CT image data are input.

[0224] For each subject, the CBCT image data and CT image data are taken of the same body part, and considering the shape factor information of the items set in advance for each body part, the CBCT image data and CT image data are analyzed to output the shape factor information of each.

[0225] In detail, after performing image shape segmentation of each CBCT image data and CT image data, shape factor information of the corresponding body part is measured.

[0226] The above information input step (S1000) may be performed by inputting the CBCT image data or CT image data and performing the above-described process to measure shape factor information, or may only receive shape measurement values ​​from an external configuration.

[0227]

[0228] In the second learning processing step (S2000), the second learning processing unit (2000) processes the biometric information, shape factor information of CBCT image data, and shape factor information of CT image data by the information input step (S1000) in pairs for each subject to generate a learning data set.

[0229] After this, using the pre-stored artificial intelligence network, the learning data set is input and a regression analysis is performed on the relationship between the shape factor information of the CBCT image data and the shape factor information of the CT image data based on the biometric information.

[0230] The regression analysis result model by the second learning processing step (S2000) is saved as the second artificial intelligence model of the shape measurement correction step (S500).

[0231] That is, the second artificial intelligence model performs a regression analysis on the relationship between the shape factor information of the CBCT image data and the shape factor information of the CT image data for each gender and age.

[0232] Through this, in the shape measurement correction step (S500), when the gender and age of the subject and the shape factor information by the shape measurement step (S400) (shape factor information analyzed from the virtual image data converted and output using the CBCT image data by the image input step (S100)) are input into the second artificial intelligence model, the shape factor information having the domain characteristics of the CT image data, in other words, the soft tissue characteristics of the shooting conditions (shooting in a lying down state) of the CT image data is corrected and output.

[0233]

[0234] Although the preferred embodiments of the present invention have been described above, the embodiments disclosed in the present invention are not intended to limit the technical spirit of the present invention, but rather to explain it. Therefore, the technical spirit of the present invention includes not only each disclosed embodiment but also a combination of disclosed embodiments, and further, the scope of the technical spirit of the present invention is not limited by these embodiments. In addition, those skilled in the art to which the present invention pertains can make numerous changes and modifications to the present invention without departing from the spirit and scope of the appended claims, and all such appropriate changes and modifications should be considered as equivalents and falling within the scope of the present invention.

[0235] <Explanation of symbols>

[0236] 100: Image input section

[0237] 200: Image Inference Unit

[0238] 300: Image Analysis Department

[0239] 400: Shape measurement unit

[0240] 500: Shape measurement correction unit

[0241] 10: Image Collection Department

[0242] 20: Image preprocessing unit

[0243] 21: First preprocessing unit

[0244] 22: Second preprocessing unit

[0245] 23: Third Preprocessing Unit

[0246] 30: First Learning Processing Unit

[0247] 1000: Information input section

[0248] 2000: Second Learning Processing Unit

Claims

1. Image input unit that receives CBCT (Cone Beam Computed Tomography) image data of a body part of the subject being photographed; An image inference unit that inputs CBCT image data by the image input unit into the stored first artificial intelligence model and converts and outputs virtual image data with reduced noise; and An image analysis unit that analyzes the virtual image data output from the image inference unit and performs image segmentation corresponding to the body part; A noise reduction system for CBCT images, including:

2. In paragraph 1, The noise reduction system of the above CBCT image is A shape measurement unit that measures shape factor information of a corresponding body part from the segmentation result performed by the image analysis unit, taking into account shape factor information of preset items for each body part; A noise reduction system for CBCT images, further comprising:

3. In paragraph 2, The noise reduction system of the above CBCT image is A shape measurement correction unit that inputs the biometric information of the preset items of the subject input from the outside and the shape factor information by the shape measurement unit into the stored second artificial intelligence model, and outputs the correction as shape factor information having the domain characteristics of CT (Computed Tomography) image data; A noise reduction system for CBCT images, further comprising:

4. In paragraph 1, In order to store the above first artificial intelligence model, the noise reduction system of the CBCT image An image collection unit that collects CBCT image data and CT image data for each body part photographed for a plurality of subjects; An image preprocessing unit that processes the CBCT image data and CT image data by the image collection unit into pairs for each subject, and matches the shooting area of ​​the corresponding CBCT image data and CT image data with the coordinate center point of the image data based on each pair; and A first learning processing unit that generates a learning data set using the CBCT image data and CT image data obtained by the image preprocessing unit and performs learning processing based on a pre-stored Cycle GAN (Generative Adversarial Network); Including, A noise reduction system for CBCT images, wherein a generator network trained by the first learning processing unit is stored as a first artificial intelligence model of the image inference unit.

5. In paragraph 4, The above image preprocessing unit A first preprocessing unit that matches the feature points of the CBCT image data and the CT image data based on the CBCT image data or the CT image data by the image collecting unit, taking into account at least one feature point preset for each body part; A second preprocessing unit that extracts specific slice data matching the shooting area of ​​the CBCT image data by using the CT image data in which the CBCT image data and the feature points are matched by the first preprocessing unit; and A third preprocessing unit that performs masking processing of a preset size and shape on specific slice data of the CT image data extracted by the second preprocessing unit; Including more, The above first learning processing unit A noise reduction system for CBCT images using CBCT image data obtained by the first preprocessing unit and CT image data obtained by the third preprocessing unit.

6. In paragraph 3, In order to store the above second artificial intelligence model, the noise reduction system of the CBCT image An information input unit that receives biometric information of preset items for a plurality of subjects, shape factor information extracted by analyzing CBCT image data photographed of a specific body part, and shape factor information extracted by analyzing CT image data photographed of the same body part as the CBCT image data; and A second learning processing unit that processes the biometric information, shape factor information of CBCT image data, and shape factor information of CT image data by the information input unit in pairs for each subject to generate a learning data set, and performs a regression analysis on the relationship between the shape factor information of the CBCT image data and the shape factor information of the CT image data based on the biometric information using a pre-stored artificial intelligence network; Including more, A noise reduction system for CBCT images, wherein the regression analysis result model by the second learning processing unit is stored as a second artificial intelligence model of the shape measurement correction unit.

7. A method for reducing noise in a CBCT image using a noise reduction system in which each step is performed by an operation processing means, An image input step for receiving CBCT (Cone Beam Computed Tomography) image data of a body part of a subject in an image input section; In the image inference unit, an image inference step for inputting CBCT image data by the image input step using the stored first artificial intelligence model and converting and outputting the converted data into virtual image data with reduced noise; In the image analysis unit, an image analysis step of analyzing the virtual image data output by the image inference step and performing image segmentation corresponding to the body part; and In the shape measurement unit, a shape measurement step of measuring shape factor information of a preset item for each body part and providing the shape factor information of the corresponding body part to the outside from the segmentation result performed by the image analysis step; A method for reducing noise in a CBCT image including:

8. In paragraph 7, The above method for reducing noise in CBCT images is After performing the above shape measurement steps, In a shape measurement correction unit, a shape measurement correction step of inputting the biometric information of the preset items of the subject input from the outside and the shape factor information by the shape measurement step, and outputting the correction as shape factor information having the domain characteristics of CT (Computed Tomography) image data using the stored second artificial intelligence model; A method for reducing noise in a CBCT image, further comprising:

9. In paragraph 7, The above method for reducing noise in CBCT images is Before performing the above image inference step, to save the first artificial intelligence model, In the image collection unit, an image collection step of collecting CBCT image data and CT image data each taken of the same body part for a plurality of subjects; In the image preprocessing unit, the CBCT image data and CT image data obtained by the image collection step are processed in pairs for each subject, and an image preprocessing step is performed to match the shooting area of ​​the corresponding CBCT image data and CT image data with the coordinate center point of the image data based on each pair; and In the first learning processing unit, a first learning processing step of generating a learning data set using the CBCT image data and CT image data obtained through the image preprocessing step and performing learning processing based on a pre-stored Cycle GAN (Generative Adversarial Network); Including, A method for reducing noise in a CBCT image, wherein a generator network trained by the first learning processing step is stored as a first artificial intelligence model by the image inference step.

10. In paragraph 9, The above image preprocessing step is A first preprocessing step of matching the feature points of the CBCT image data and the CT image data based on the CBCT image data or the CT image data by the image collection step, taking into account at least one feature point preset for each body part; A second preprocessing step of extracting specific slice data matching the shooting area of ​​the CBCT image data by using the CT image data in which the CBCT image data and the feature points are matched through the first preprocessing step; and A third preprocessing step of performing masking processing of a preset size and shape on specific slice data of the CT image data extracted by the second preprocessing step; Including more, The above first learning processing step is A method for reducing noise in a CBCT image using CBCT image data obtained through the first preprocessing step and CT image data obtained through the third preprocessing step.

11. In paragraph 8, The above method for reducing noise in CBCT images is Before performing the above shape measurement correction step, to save the second artificial intelligence model, An information input step for receiving, in an information input section, biometric information of preset items for a plurality of subjects, shape factor information extracted by analyzing CBCT image data photographed of a specific body part, and shape factor information extracted by analyzing CT image data photographed of the same body part as the CBCT image data; and In a second learning processing unit, the biometric information, shape factor information of CBCT image data, and shape factor information of CT image data are processed in pairs for each subject by the information input step to generate a learning data set, and a second learning processing step performs a regression analysis on the relationship between the shape factor information of the CBCT image data and the shape factor information of the CT image data based on the biometric information using a pre-stored artificial intelligence network; Including more, A method for reducing noise in a CBCT image, wherein the regression analysis result model obtained by the second learning processing step is stored as a second artificial intelligence model of the shape measurement correction step.

Citation Information

Patent Citations

  • sCT Image Generation Using CycleGAN with Deformable Layers

    JP2022536107A

  • Tomography apparatus and method for reconstructing a tomography image thereof

    KR1020170030308A

  • Medical image processing method and apparatus

    KR102477991B1

  • Method and system for slow charging based on inverter and single phase system

    KR102568218B1

  • Arc evaporator having filtering net for macroparticle

    KR102602127B1