一种基于伪影估计的低剂量CT图像降噪方法

A low-dose CT image denoising method based on artifact estimation

低剂量CT(LDCT)包含丰富组织结构、病理信息和分布极其不规律的噪声伪影,这2种信息的幅度值分布规律相似。因此,LDCT降噪任务易出现特征提取不充分、网络对噪声伪影方向特性敏感度不足及降噪结果过度平滑等问题。为此,应用U-Net网络作为去噪网络的基本模型,设计了一种基于伪影估计的LDCT降噪网络。所提网络模型主要包括主特征提取网络和方向敏感注意力子网络2部分。为充分利用不同尺度特征之间的差异性,提高特征提取有效性,在编解码U-Net结构基础上增加了一个稠密特征增强模块;为提高降噪网络对噪声伪影方向特征的敏感度,设计了一个方向敏感注意力子网络;为保障网络训练稳定性,设计了多种损失函数来共同优化网络训练过程。实验结果表明:与目前主流的LDCT降噪方法相比,所提方法降噪结果的视觉效果与量化指标均表现最佳。

Low-dose CT(LDCT)contains abundant tissue structure,pathological information and noise artifacts with extremely irregular distribution.These two different types of information have comparable amplitude distributions.Therefore,the LDCT denoising task is prone to some problems,such as insufficient feature extraction,insufficient network sensitivity to the directional characteristics of noise artifacts,and excessive smoothing of the denoising results.In response to the above problems,this work uses the U-Net network as the basic model of the denoising network,and designs a LDCT denoising network based on artifact estimation.The proposed network mainly includes two parts:the main feature extraction network and the direction-sensitive attention subnetwork.Firstly,to better use the differences between various scale features and increase the efficiency of feature extraction,we add a dense feature improvement module to the codec U-Net structure.Secondly,we design a direction-sensitive attention subnetwork to improve the sensitivity of the denoising network to the direction characteristics of the noise artifacts.Finally,to ensure the stability of network training,we utilize a variety of loss functions to optimize the network training process.The experimental results show that the proposed algorithm is superior to other mainstream LDCT denoising algorithms in terms of visual effects and quantitative indicators.