KR-20260064998-A - LDCT DENOISING NEURAL NETWORK BASED ON SGM(SCORE-BASED GENERATIVE MODEL) APPARATUS AND METHOD

Abstract

The present invention relates to an SGM-based LDCT denoising neural network device, comprising: an input unit that receives an LDCT (Low-dose computed tomography) image; a score model unit that processes the LDCT image in time steps by constructing a score-based generative model that is pre-trained to remove noise from the LDCT training image; an image adjustment unit that generates a noise-fitted LDCT image by adjusting the noise characteristics of the LDCT image using the LDCT image at a reference time step through the score-based generative model; a denoising unit that gradually reduces the noise of the noise-fitted LDCT image through repetitive input and output in reverse time steps of the score-based generative model; and a noise-restored image output unit that outputs a noise-restored image generated based on the noise-fitted LDCT image.

Inventors

• 백종덕
• 정호진

Assignees

• 연세대학교 산학협력단

Dates

Publication Date

20260508

Application Date

20241030

Claims (10)

1. Input unit for receiving LDCT (Low-dose computed tomography) images; A score model unit that constructs a pre-trained score-based generative model to remove noise from LDCT training images and processes the LDCT images at time steps; An image adjustment unit that generates a noise-fitted LDCT image by adjusting the noise characteristics of the LDCT image using the LDCT image of the reference time step through the above score-based generation model; A denoising unit that gradually reduces noise in the noise-fitted LDCT image through iterative input and output at inverse time steps of the score-based generative model; and An SGM (Score-based Generative Model) based LDCT denoising neural network device comprising a noise restoration image output unit that outputs a noise restoration image generated based on the noise-fitted LDCT image above.
2. In paragraph 1, the score model part An SGM-based LDCT denoising neural network device characterized by training the score-based generative model with data to which noise has been added at various time steps of the above LDCT training image.
3. In paragraph 2, the score model part An SGM-based LDCT denoising neural network device characterized by removing noise from the LDCT image at the reference time step (t = t_fit) and providing it as an input image to the image adjustment unit.
4. In paragraph 2, the score model part An SGM-based LDCT denoising neural network device characterized by processing the noise of the LDCT image in reverse time steps from the reference time step (t = t_fit) to the first time step (t=0) and sequentially providing it as the input image of the denoising unit.
5. In paragraph 1, the image adjustment unit An SGM-based LDCT denoising neural network device characterized by adjusting the noise characteristics through a Langevin MCMC (Markov Chain Monte Carlo) process.
6. In paragraph 1, the image adjustment unit An SGM-based LDCT denoising neural network device characterized by performing a noise filtering process on the LDCT image through a noise characteristic adjustment process.
7. In claim 1, the denoising part An SGM-based LDCT denoising neural network device characterized by generating the noise restoration image for the noise-fitted LDCT image by inversely utilizing the LDCT image from the reference time step to the LDCT image from the first time step through the score-based generative model.
8. In paragraph 7, the above denoising part An SGM-based LDCT denoising neural network device characterized by generating the noise-restored image by applying a Stochastic Differential Equation (SDE).
9. In claim 1, the denoising part An SGM-based LDCT denoising neural network device characterized by performing an inverse transform restoration process for the LDCT image through a noise temporal inverse transformation process.
10. In an SGM (Score-based Generative Model) based LDCT denoising neural network method performed in an SGM-based LDCT denoising neural network device, Input step for receiving LDCT (Low-dose computed tomography) images; A score model step for processing the LDCT images at time steps by constructing a pre-trained score-based generative model to remove noise from the LDCT training images; An image adjustment step for generating a noise-fitted LDCT image by adjusting the noise characteristics of the LDCT image using the LDCT image of the reference time step through the score-based generative model; A denoising step for gradually reducing noise in the noise-fitted LDCT image through iterative input and output at the inverse time steps of the score-based generative model; and An SGM-based LDCT denoising neural network method comprising a noise restoration image output step that outputs a noise restoration image generated based on the noise-fitted LDCT image.

Description

SGM-based LDCT Denoising Neural Network Apparatus and Method The present invention relates to a low-dose computed tomography denoising technology, and more specifically, to an SGM-based LDCT denoising neural network device and method capable of providing high-quality images at the level of normal-dose computed tomography (NDCT) by removing noise from LDCT images through a score-based generative model (SGM) to improve the quality of low-dose computed tomography (LDCT) images. Computed Tomography (CT) is a technology that uses radiation to reconstruct the inside of the human body into cross-sectional images. This technology plays an essential role in the medical field for diagnosing diseases and establishing treatment plans. However, low-dose scanning technology is necessary to reduce the risk to patients caused by radiation exposure during CT scans. Low-dose CT (LDCT) scans are a method to reduce patient radiation exposure by lowering the radiation dose. However, low-dose CT has a lower signal-to-noise ratio (SNR) compared to standard-dose CT, which leads to a problem of degraded image quality. This degradation in image quality causes a decrease in diagnostic accuracy, and various image restoration and noise reduction technologies have been studied to address this issue. Previously, noise reduction techniques utilizing Iterative Reconstruction (IR) and Convolutional Neural Networks (CNNs) were primarily used. However, these methods have limitations in image quality improvement and face constraints in practical application due to the difficulty of securing training data. Recently, diffusion models have been attracting attention for their ability to generate high-quality images by leveraging the characteristics of the data they have learned. Diffusion models are effective for reconstructing contaminated images or removing artifacts, and they can also be utilized to remove rarefaction artifacts from CT images. However, there are limitations to directly applying diffusion models used in natural images to CT images due to the difficulty of training images combined with text and securing large-scale training data. Korean Registered Patent No. 10-2708543 (September 13, 2024) provides a diffusion model-based computed tomography high-resolution conversion device and method that generates high-resolution CT images by removing noise generated in CT images and improves the quality of low-resolution CT images, thereby enabling more accurate diagnosis in the medical field. A method for converting a computed tomography image to a high resolution based on a diffusion model is performed by a computing device including at least a processor, and comprises the steps of: generating training data from an original image using a diffusion model; training a Convolutional Neural Network Model (CNN model) including an encoder and a decoder using the training data; generating an enlarged image by linearly interpolating an input image; generating a high-resolution residual using an arbitrary vector extracted from a Gaussian distribution and the enlarged image; and generating a high-resolution image by combining the enlarged image and the high-resolution residual. FIG. 1 is a drawing illustrating an SGM-based LDCT denoising neural network device (100) according to one embodiment of the present invention. FIG. 2 is a diagram illustrating the functional configuration of the SGM-based LDCT denoising neural network device (100) of FIG. 1. FIG. 3 is a diagram illustrating the system configuration of the SGM-based LDCT denoising neural network device (100) of FIG. 1. FIG. 4 is a flowchart illustrating the operation of the SGM-based LDCT denoising neural network device (100) of FIG. 1. FIG. 5 is a diagram showing a comparison of an LDCT (low-dose computed tomography) image, an FBPConvNet processing result, a result of the proposed method (where neither of the two uses Langevin MCMC), and a normal-dose CT (NDCT) image at 10% and 5% radiation doses in the SGM-based LDCT denoising neural network device (100) of FIG. 1. The description of the present invention is merely an example for structural or functional explanation, and therefore the scope of the present invention should not be interpreted as being limited by the examples described in the text. That is, since the examples are subject to various modifications and may take various forms, the scope of the present invention should be understood to include equivalents capable of realizing the technical concept. Furthermore, the objectives or effects presented in the present invention do not imply that a specific example must include all of them or only such effects; therefore, the scope of the present invention should not be understood as being limited by them. Meanwhile, the meaning of the terms described in this application should be understood as follows. Terms such as "first," "second," etc., are intended to distinguish one component from another, and the scope of rights shall not be limited b