摘要
现有的低剂量CT图像去噪算法大多依赖于大样本的配对数据进行训练,而在实际中,很难同时获得同一患者的低剂量CT图像和常规剂量CT图像,从而导致训练样本量的不足.针对这一问题,本文在配对图像不足的条件下,提出了一种基于模块化神经网络的低剂量CT图像去噪算法.该方法采用模块化子网络串联,在子网络内部应用跨层连接增加特征图利用率,并且引入了一种新型的二次卷积提高去噪效果.实验表明,在缺少配对数据的弱监督条件下,该网络可以有效降低低剂量CT图像噪声,显著提升低剂量CT图像的视觉质量和客观评价指标.与目前的方法相比,本文所提出的网络可以更好地在弱监督条件下减少低剂量CT图像噪声.
Most of the existing low-dose CT image denoising algorithms rely on the paired data of large samples for training.However,in practice,it is difficult to obtain low-dose CT images and conventional-dose CT images of the same patient at the same time,which leads to the shortage of training sample size.This paper proposes a low dose CT image denoising algorithm based on modular neural network at the condition of insufficient paired images.In this method,modular subnetwork is used in series,cross-layer connections are applied in the subnetwork to increase the utilization ratio of feature map,and a new secondary convolution is introduced to improve the denoising effect.The experimental results show that the network can effectively reduce the noise of low-dose CT images and significantly improve the visual quality and objective evaluation indexes of low-dose CT images under the condition of weak supervision without paired data.Compared with current methods,the proposed network can better reduce low-dose CT image noise under weak supervision.
作者
赵桂宸
陈平
ZHAO Guichen;CHEN Ping(Shanxi Provincial Key Laboratory of Signal Capturing and Processing, College of Information and Communication Engineering, North of China University, Taiyuan 201800, China)
出处
《测试技术学报》
2021年第3期229-236,共8页
Journal of Test and Measurement Technology
关键词
模块化网络
二次卷积
低剂量CT
深度学习
弱监督
modular network
secondary convolution
low-dose CT
deep learning
weak supervision