Search

KR-20260065632-A - APPARATUS AND METHOD FOR NEURAL NETWORK TRAINING OF DENTAL IMAGES THROUGH PATIENT-SPECIFIC DATA AUGMENTATION

KR20260065632AKR 20260065632 AKR20260065632 AKR 20260065632AKR-20260065632-A

Abstract

The present invention relates to a neural network learning device for dental images through patient-customized data augmentation, comprising a bone segmentation unit that generates a bone mask by segmenting a bone in a dental image, a tooth labeling unit that generates a labeled dental image by labeling a tooth to be inserted with metal in the bone mask, a metal mask generation unit that generates a metal mask in the labeled dental image and performs data augmentation with the metal mask to generate a plurality of augmented metal masks, and a metal influence image processing unit that generates a metal influence dental image by simulating a multicolor spectrum-based metal artifact on the plurality of augmented metal masks and calculating a metal attenuation coefficient.

Inventors

  • 백종덕
  • 안준현

Assignees

  • 연세대학교 산학협력단

Dates

Publication Date
20260511
Application Date
20241029

Claims (11)

  1. A bone segmentation unit that segments the bone in a dental image to generate a bone mask; A tooth labeling unit that labels the teeth to be inserted with metal in the bone mask above to generate a labeled dental image; A metal mask generation unit that generates a metal mask from the above-described labeled dental image and performs data augmentation with the metal mask to generate a plurality of augmented metal masks; and A neural network learning device for dental images through patient-specific data augmentation, comprising a metal influence image processing unit that simulates multicolor spectrum-based metal artifacts on the plurality of augmented metal masks and calculates metal attenuation coefficients to generate metal influence dental images.
  2. In claim 1, the bone splitting part A neural network learning device for dental images through patient-specific data augmentation, characterized by applying a Gaussian mixture model (GMM) to the dental image to calculate the mean and variance of the tissue and calculating an adaptive bone threshold through the mean and variance.
  3. In paragraph 2, the bone splitting part A neural network learning device for dental images through patient-specific data augmentation, characterized by generating the bone mask to divide the bone from the tissue by considering the difference in damping coefficients between the tissue and the bone through the mean and variance.
  4. In claim 1, the tooth labeling part A neural network learning device for dental images through patient-specific data augmentation, characterized by performing initial labeling through connected component labeling (CCL) in the bone mask, separating the tooth through morphological erosion, and then completing the labeling.
  5. In paragraph 4, the tooth labeling part A neural network learning device for dental images through patient-specific data augmentation, characterized by performing actual tooth labeling by applying threshold filtering to the initial labels obtained through the above-mentioned connected component labeling (CCL) to remove noise below a certain threshold.
  6. In paragraph 5, the tooth labeling part A neural network learning device for dental images through patient-customized data augmentation, characterized by performing clear tooth labeling by applying the morphological erosion to the actual tooth labeling to sever the connection between adjacent teeth and clearly distinguish between adjacent teeth.
  7. In paragraph 6, the tooth labeling part A neural network learning device for dental images through patient-customized data augmentation, characterized by applying Shrunk Region Labeling to the clear tooth labeling to supplement the tooth region lost due to morphological erosion and generating the labeled dental image through tooth selection.
  8. In claim 1, the metal mask generating part A neural network learning device for dental images through patient-specific data augmentation, characterized by generating a plurality of augmented metal masks by determining the number, shape, and size of metals to be inserted into the metal mask through the data augmentation.
  9. In claim 1, the metal influence image processing unit A neural network learning device for dental images through patient-specific data augmentation, characterized by performing polychromatic sinogram simulation on the plurality of augmented metal masks, generating a metal insertion image, and applying beam hardening correction and filtered back projection to the metal insertion image to generate the metal-influenced dental image.
  10. In claim 9, the metal influence image processing unit A neural network learning device for dental images through patient-specific data augmentation, characterized by controlling the intensity of the metal artifact by adjusting the metal damping coefficient during the multicolor sinogram simulation process.
  11. In a method for neural network learning of dental images through patient-specific data augmentation performed in a neural network learning device for dental images through patient-specific data augmentation, A bone segmentation step for segmenting the bone in a dental image to generate a bone mask; A tooth labeling step of labeling the teeth to be inserted with metal in the bone mask to generate a labeled dental image; A metal mask generation step of generating a metal mask from the above-described labeled dental image and performing data augmentation with the metal mask to generate a plurality of augmented metal masks; and A neural network learning method for dental images through patient-specific data augmentation, comprising a metal influence image processing step of simulating multicolor spectrum-based metal artifacts on the plurality of augmented metal masks and calculating a metal attenuation coefficient to generate a metal influence dental image.

Description

Apparatus and Method for Neural Network Training of Dental Images Through Patient-Specific Data Augmentation The present invention relates to a neural network learning technology for dental images, and more specifically, to a device and method for neural network learning of dental images through patient-specific data augmentation that can reduce image distortion caused by metal artifacts and improve diagnostic accuracy by labeling the location of metal insertion in dental images and generating augmented data that reflects the influence of metal. Computed Tomography (CT) is an essential tool in various medical diagnoses, providing detailed images of the human body to aid in diagnosis and treatment planning. However, when a patient has metal implants, the high attenuation coefficient of the metal can cause localized streak-like distortion—known as metal artifacts—around the objects in CT images. These metal artifacts are problematic because they degrade image resolution, hinder accurate diagnosis, and can distort critical information. Consequently, various technologies have been developed to reduce metal artifacts. Traditional Metal Artifact Reduction (MAR) methods include techniques such as Linearly Interpolated MAR (LMAR) and Normalized MAR (NMAR), which operate by interpolating regions containing metal traces in sinogram data to restore information from those areas. However, these techniques have limitations in that they cannot completely remove the complex patterns of metal artifacts, which may restrict their use in clinical settings. Recently, with the advancement of deep learning technology, various deep learning-based MAR methods have been proposed in image domains, sinogram domains, and hybrid domains combining these. Deep learning-based approaches utilize simulation data to demonstrate superior metal artifact reduction performance compared to traditional methods, contributing to improved image quality. However, most deep learning-based MAR methods require supervised learning, which necessitates data pairs where images containing anatomically identical metals correspond one-to-one with images without metals. There are limitations as collecting such data pairs is practically difficult and requires significant time and cost. Korean Published Patent No. 10-2022-0022328 (February 25, 2022) provides a method and apparatus for correcting artifacts in a CT image, comprising the steps of reconstructing a first back-projection image and a second back-projection image of each of the first projection data and the second projection data, and generating an image in which artifacts are attenuated from the reconstructed CT image, with the aim of reducing the amount of computation by avoiding an inefficient iterative reconstruction structure. A method for correcting metal artifacts in a CT image comprises the steps of: a processor reconstructing a CT image containing artifacts; segmenting a metal region in the reconstructed CT image; generating first projection data according to the metal region by calculating a light transmission length along the metal region; generating second projection data related to artifacts around the metal region using the light transmission length; reconstructing a first back-projection image and a second back-projection image of each of the first projection data and the second projection data, respectively; and generating an image in which artifacts included in the reconstructed CT image are attenuated using the first back-projection image and the second back-projection image. FIG. 1 is a diagram illustrating the configuration of a neural network learning device for dental images through patient-customized data augmentation according to one embodiment of the present invention. Figure 2 is a flowchart illustrating the operation of a neural network learning device for dental images through patient-specific data augmentation of Figure 1. Figure 3 is a diagram visually illustrating each step of the dental image processing process of the neural network learning device for dental images through patient-customized data augmentation of Figure 1. Figure 4 is a diagram illustrating the process of simulating metal artifact and non-artifact images of a neural network learning device for dental images through patient-specific data augmentation of Figure 1. Figure 5 is a diagram illustrating the process of setting the damping coefficient of a metal in the metal artifact simulation of a dental image neural network learning device through patient-specific data augmentation of Figure 1. Figure 6 is a diagram visually illustrating the proposed method and a comparison method through the flow of the neural network learning device for dental images through patient-specific data augmentation of Figure 1, the data augmentation and metal artifact simulation method, the learning dataset, the neural network learning, and the test results. Figure 7 is a diagram showing a visual comparison of a simulated metal artifa