KR-20260064995-A - LOW-DOSE COMPUTED TOMOGRAPHY DENOISING NEURAL NETWORK APPARATUS AND METHOD

Abstract

The present invention relates to a low-dose computed tomography denoising neural network device, wherein the LDCT denoising neural network device comprises a dose recognition network unit that provides an LDCT (Low-dose computed tomography) image as an input image to a CNN (convolutional neural network) and outputs an NDCT (Normal-Dose Computed Tomography) similar image as an output image, a noise variance correction unit that performs noise-variance correction on the input image for the CNN to resolve oversmoothing of the CNN, and an adaptive noise reduction unit that controls the CNN to gradually remove noise step by step according to the dose level of a given LDCT image.

Inventors

• 백종덕
• 김성준

Assignees

• 연세대학교 산학협력단

Dates

Publication Date

20260508

Application Date

20241030

Claims (9)

1. A dose recognition network unit that provides an LDCT (Low-dose computed tomography) image as an input image to a CNN (convolutional neural network) and outputs an NDCT (Normal-Dose Computed Tomography) similar image as an output image; A noise-variance correction unit that performs noise-variance correction on an input image for the above CNN to resolve oversmoothing of the above CNN; and An LDCT denoising neural network device comprising an adaptive noise reduction unit that controls the CNN to progressively remove noise step by step according to the dose level of a given LDCT image.
2. In paragraph 1, the above dose recognition network unit An LDCT denoising neural network device characterized by pre-generating the LDCT image composed of various dose levels by adding noise to the NDCT image.
3. In paragraph 2, the above Dose recognition network unit Alpha for the above NDCT image An LDCT denoising neural network device characterized by generating various dose levels by controlling parameters to adjust the intensity of the noise.
4. In paragraph 2, the above Dose recognition network unit LDCT denoising neural network device characterized by training the CNN to be the minimum between the NDCT image and the NDCT similar image.
5. In claim 1, the noise dispersion correction unit An LDCT denoising neural network device characterized by correcting the noise based on the difference between the output image and the input image of the above CNN so that the detailed structure of the input image is not lost.
6. In paragraph 5, the noise dispersion correction unit An LDCT denoising neural network device characterized by correcting the noise so that the output image has a noise-variance level greater than or equal to a threshold with respect to the NDCT image.
7. In paragraph 1, the adaptive noise reduction part An LDCT denoising neural network device characterized by recognizing the dose level of the given LDCT image through the above CNN and performing iterative input/output through the above CNN to gradually remove the noise.
8. In paragraph 7, the adaptive noise reduction unit An LDCT denoising neural network device characterized by inputting the Nth (where N is a natural number) intermediate image into the CNN to generate the (N+1) intermediate image, and inputting the (N+1) intermediate image back into the CNN to generate the (N+2) intermediate image.
9. In an LDCT (Low-dose computed tomography) denoising neural network method performed in an LDCT denoising neural network device, A dose recognition network step that provides an LDCT image as an input image to a CNN (convolutional neural network) and outputs an NDCT (Normal-Dose Computed Tomography) similar image as an output image; A noise-variance correction step for performing noise-variance correction on an input image for the above CNN to resolve oversmoothing of the above CNN; and An LDCT denoising neural network method comprising an adaptive noise reduction step that controls the CNN to progressively remove noise stepwise according to the dose level of a given LDCT image.

Description

Low-Dose Computed Tomography Denoising Neural Network Apparatus and Method The present invention relates to a low-dose computed tomography denoising technology, and more specifically, to a low-dose computed tomography denoising neural network device and method capable of effectively removing noise to improve the quality of low-dose computed tomography (LDCT) images and providing image quality at the level of normal-dose computed tomography (NDCT). Computed Tomography (CT) is an important medical imaging technique widely used to diagnose diseases in patients. However, patients are exposed to radiation during CT scans, and there is a potential risk associated with this. Low-Dose Computed Tomography (LDCT) has been introduced as a representative method to reduce this burden, which is a method of imaging with reduced X-ray dose. However, while low-dose computed tomography (LDCT) reduces radiation exposure, it can cause image noise to be amplified, which may hinder accurate diagnosis by medical professionals. Compared to normal-dose computed tomography (NDCT), it can be difficult to identify anatomical structures in LDCT images due to noise. To address these issues, denoising techniques utilizing deep learning-based Convolutional Neural Networks (CNNs) are being developed. While CNN-based denoising enables fast image processing, it requires large amounts of data from both low-dose computed tomography (LDCT) and standard-dose computed tomography (NDCT) for effective training. Since it is difficult to capture both types of images on actual patients, a method of training by simulating low-dose CT based on standard-dose CT is widely used. However, a generalization problem may occur where performance deteriorates on data with noise levels different from those used for training. While additional training (Fine-Tuning) is required to overcome this, it has limitations due to the large amount of data and time required. Korean Published Patent No. 10-2022-0135683 (October 7, 2022) provides a low-dose CT image noise reduction device capable of reducing noise without causing blurring or changing pixel values in low-dose CT images, as well as a learning device and method for the same. A low-dose CT image noise reduction device is implemented with a pre-trained artificial neural network and includes a denoising neural network that outputs a denoising CT image by removing noise from an applied low-dose CT image according to a learned method. The denoising neural network is trained by applying a denoising CT image and a normal-dose CT image output from a low-dose CT image to a lesion identification neural network, which is a pre-trained artificial neural network that identifies images containing lesions during the learning process, respectively. The lesion identification neural network is trained to reduce the observation loss calculated as the difference between a first feature map and a second feature map obtained during the lesion identification process from the denoising CT image and the normal-dose CT image, thereby effectively reducing noise in the low-dose CT image without causing blurring or changing pixel values. FIG. 1 is a diagram illustrating the configuration of a low-dose computed tomography denoising neural network device according to one embodiment of the present invention. FIG. 2 is a diagram illustrating the system configuration of the low-dose computed tomography denoising neural network device (100) of FIG. 1. Figure 3 is a diagram illustrating the learning and testing process of a network for removing noise from low-dose CT (LDCT) images of the low-dose computed tomography denoising neural network device of Figure 1. Figure 4 is a diagram illustrating a noise removal method for low-dose CT (LDCT) images by comparing existing technology with the technology proposed in the present invention. Figure 5 is a flowchart illustrating the operation of the low-dose computed tomography denoising neural network device of Figure 1. Figure 6 shows CT images reconstructed at various dose conditions (25%, 10%, 5%) of the Mayo dataset in the low-dose computed tomography denoising neural network device of Figure 1, and each column is a diagram showing a comparison of different networks and reconstruction methods. Figure 7 is a diagram showing the results of an ablation study conducted to evaluate the effect of the Noise-Variance Calibration Module (NCM) in the low-dose computed tomography denoising neural network device of Figure 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 i