删除或更新信息,请邮件至freekaoyan#163.com(#换成@)

基于生成对抗网络和噪声水平估计的低剂量CT图像降噪方法

本站小编 Free考研考试/2022-01-03

张雄,,
杨琳琳,
上官宏,
韩泽芳,
韩兴隆,
王安红,
崔学英
太原科技大学电子信息工程学院 太原 030024
基金项目:国家青年科学基金(62001321),山西省高等学校科技创新项目(2019L0642),山西省自然科学基金(201901D111261)

详细信息
作者简介:张雄:男,1973年生,教授,硕士生导师,研究方向为模式识别、医学图像处理和视频目标跟踪
杨琳琳:女,1992年生,硕士生,研究方向为医学图像处理
上官宏:女,1988年生,副教授,硕士生导师,研究方向为模式识别、医学图像处理
韩泽芳:女,1996年生,硕士生,研究方向为医学图像处理
韩兴隆:男,1995年生,硕士生,研究方向为医学图像处理
王安红:女,1972年生,教授,博士生导师,研究方向为图像视频编码
崔学英:女,1978年生,副教授,硕士生导师,研究方向为图像处理与重建
通讯作者:张雄 zx@tyust.edu.cn
中图分类号:TN911.73; TP391

计量

文章访问数:716
HTML全文浏览量:282
PDF下载量:112
被引次数:0
出版历程

收稿日期:2020-07-17
修回日期:2021-02-03
网络出版日期:2021-03-01
刊出日期:2021-08-10

A Low-Dose CT Image Denoising Method Based on Generative Adversarial Network and Noise Level Estimation

Xiong ZHANG,,
Linlin YANG,
Hong SHANGGUAN,
Zefang HAN,
Xinglong HAN,
Anhong WANG,
Xueying CUI
School of Electronic Information Engineering, Taiyuan University of Science and Technology, Taiyuan 030024, China
Funds:The Natural Science for Youth Foundation (62001321), The Scientific and Technological Innovation Programs of Higher Educations Institutions in Shanxi (2019L0642), The Natural Science Foundation of Shanxi Province (201901D111261)


摘要
摘要:生成对抗网络(GAN)用于低剂量CT(LDCT)图像降噪具有一定的性能优势,成为近年CT图像降噪领域新的研究热点。不同剂量的LDCT图像中噪声和伪影分布的强度发生变化时,GAN网络降噪性能不稳定,网络泛化能力较差。为了克服这一缺陷,该文首先设计了一个编解码结构的噪声水平估计子网,用于生成不同剂量LDCT图像对应的噪声图,并用原始输入图像与之相减来初步抑制噪声;其次,在主干降噪网络中,采用GAN框架,并将生成器设计为多路编码的U-Net结构,通过博弈对抗实现网络结构优化,进一步抑制CT图像噪声;最后,设计了多种损失函数来约束不同功能模块的参数优化,进一步保障了LDCT图像降噪网络的性能。实验结果表明,与目前流行算法相比,所提出的降噪网络能够在保留LDCT图像原有重要信息的基础上,取得较好的降噪效果。
关键词:图像降噪/
生成对抗网络/
低剂量CT/
U-Net/
噪声水平
Abstract:Generative Adversarial Network (GAN) for Low-Dose CT (LDCT) image noise reduction has certain performance advantages, and has become a new research hot field of CT image noise reduction in recent years. When the intensity of noise and artifact distribution changes in LDCT images of different doses, the noise reduction performance of GAN network is unstable, and the generalization ability of the network is low. In order to overcome these shortcomings, this paper first designs a noise level estimation subnet with a encoder-decoder structure to generate the noise maps corresponding to LDCT images with different doses, which is subtracted from the original input image to initially suppress the noise; Secondly, the backbone of the noise reduction network is designed as a multi-coded U-Net structure that is optimized through game confrontation to suppress further CT image noise; Finally, a variety of loss functions are designed to constrain the parameter optimization of each function modules, thus to guarantee further the performance of the LDCT image noise reduction network. Experimental results show that compared with current popular algorithms, the noise reduction network proposed in this paper can achieve a better noise reduction on the basis of retaining the original important information of LDCT images.
Key words:Image denoising/
Generative Adversarial Network (GAN)/
Low-Dose CT (LDCT)/
U-Net/
Noise level



PDF全文下载地址:

https://jeit.ac.cn/article/exportPdf?id=6c59cfdf-4576-4d2a-b60b-4815a53249a1
相关话题/图像 网络 医学 硕士生导师 设计