Low-dose CT Image Denoising Based On Generative Adversarial Network

Computer tomography(CT)imaging has received great attention since its inception,and has become an indispensable auxiliary method in medical diagnosis.The use of higher dose X-rays in the CT image acquisition process will increase the patient’s potential risk of cancer,so most of the current medical clinical diagnosis uses low-dose CT imaging technology.However,reducing the radiation dose introduces a lot of noise and artifacts to the CT image,which in turn leads to image quality degradation.On the basis of maintaining the integrity of the tissue structure and pathological information on the CT image,removing noise and artifacts on the low-dose CT image as much as possible has important theoretical research significance and clinical application value.The performance of traditional image processing denoising algorithm is poor in the field of low dose CT image denoising.After denoising,it will cause a series of problems such as image blur and loss of texture information.In recent years,Generating Adversarial Networks(GAN)based on deep learning has made some progress in CT image denoising,and it has become a new research hotspot in this field.This paper focuses on the design of the network framework for generating adversarial networks for CT image denoising.The main research work is as follows:(1)A high-frequency sensitive denoising network based on GAN is proposed.Considering that noise and artifacts mainly exist in the high-frequency domain of CT images,and the details of patient pathological information are also reflected in the high-frequency domain.Therefore,when removing noise and artifacts,some useful details may be lost or destroyed.In order to improve the network’s ability to process high-frequency information,The LDCT image was processed by frequency division.The U-Net channel in the high-frequency domain is specially designed to process high-frequency components,and use the high-frequency components of NDCT to supervise the channel.In addition,the Inception module is used in the discriminator to extractthe multi-scale features of the image to improve the discriminatory ability of the discriminator,thereby further constraining the output quality of the generator.Experimental results show that the proposed high-frequency sensitive denoising network can improve the network’s ability to process high-frequency information,remove artifacts and noise in low-dose CT images,and maintain higher fidelity to high-frequency useful details.(2)A multi-channel convolution U-Net noise reduction network is proposed.The problem of image noise reduction is to eliminate unknown noise from noisy images.In order to design a unified noise reduction model that can handle various low-dose CT images well,First,a noise level estimation subnet with codec structure is designed to improve the network’s ability to process LDCT images with different doses which have obvious differences in noise;Secondly,the backbone noise reduction network is designed as a GAN framework that can achieve internal self-optimization through game confrontation,and the generator is designed as a multi-coding U-Net structure to ensure LDCT noise reduction can meet the needs of medical diagnosis;Finally,the performance of the LDCT image noise reduction network is further guaranteed through the confrontation training between the generator and the discriminator.Experimental results show that,compared with the current popular algorithm,the noise reduction network proposed in this paper performs well,and can achieve good noise reduction effect on the basis of retaining the original important physiological information of the LDCT image.