基于编解码器生成对抗网络的CT去噪  

CT image denoising using generative adversarial network with residual encoder-decoder generator

作　　者：周广宇 张鹏程[1] 刘生富 刘祎[1] 桂志国[1] ZHOU Guang-yu;ZHANG Peng-cheng;LIU Sheng-fu;LIU Yi;GUI Zhi-guo(Shanxi Provincial Key Laboratory for Biomedical Imaging and Big Data,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学生物医学成像与影像大数据重点实验室,山西太原030051

出　　处：《计算机工程与设计》2022年第4期1059-1065,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61671413、61801438);高等学校科技创新基金项目(2020L0282);国家重大科学仪器设备开发专项基金项目(2014YQ24044508);山西省应用基础研究基金项目(201901D211246);山西省面上青年基金项目(201801D221196)。

摘　　要：针对生成对抗网络在训练中损失函数收敛慢,难以恢复图像细节的问题,提出一种基于编解码器与多尺度损失函数的生成对抗网络模型。使用含残差连接的编解码器作为生成器,该网络易于训练,能够加快对抗损失函数的收敛;引入噪声损失,与使用VGG19模型的感知损失构成多尺度损失函数,使图像在视觉上的纹理细节达到更细致的恢复效果。实验结果表明,与低剂量CT相比,去噪后图像的峰值信噪比提升了8.1%,结构相似性指数提升了4.8%,改进后的网络加快了损失函数收敛,有效改善了生成对抗网络训练困难、损失函数收敛慢、图像细节难以恢复等问题。To solve the problem of low convergence speed of loss function and difficult restoration of image detail in the training process of generative adversarial network(GAN),a GAN model based on encoder-decoder and multi-scale loss function was proposed.The encoder-decoder network with residual connection was used as generator in the model,which was easy to train and could effectively accelerate the convergence of adversarial loss function.Noise loss was introduced into the network model,which formed a multi-scale loss function together with the perceptual loss using VGG19 model,so that the visual texture details of the image achieved more detailed restoration effects.Experimental results show that,compared with low-dose CT,the peak signal-to-noise ratio(PSNR)of the denoised image is increased by 8.1%,and the structural similarity index measure(SSIM)is increased by 4.8%.The improved network speeds up the convergence of the loss function,and effectively solves the difficulties of GAN training,slow convergence of the loss function,and difficult restoration of image details.

关 键 词：CT图像去噪 生成对抗网络 编解码器 残差连接 损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]