文章摘要
张硕,余世明.基于多尺度并行深度可拆分的CNN新冠肺炎CT图像去噪方法[J].高技术通讯(中文),2021,31(11):1145~1153
基于多尺度并行深度可拆分的CNN新冠肺炎CT图像去噪方法
COVID-19 CT images denoising method based on multi-scale parallel deep split CNN
  
DOI:10.3772/j.issn.1002-0470.2021.11.004
中文关键词: 新冠肺炎(COVID-19)电子计算机断层扫描(CT)图像, 图像去噪, 多尺度特征, 深浅通道并行, 拆分卷积
英文关键词: coronavirus disease 2019 (COVID-19) computed tomography (CT) image, image denoising, multi-scale feature, deep and shallow parallel channel, split convolution
基金项目:
作者单位
张硕  
余世明  
摘要点击次数: 862
全文下载次数: 571
中文摘要:
      目前新冠肺炎(COVID-19)在全球蔓延,为了对新冠肺炎进行早期诊断,同时减轻医护人员的工作压力,使用深度学习对患者胸部电子计算机断层扫描(CT)图像进行分析变得越来越重要。针对肺炎图像中纹理细节较为丰富、边缘结构模糊、极易干扰机器及医生诊断的问题,本文提出一种基于多尺度并行深度可拆分卷积神经网络(MSP-ReCNN),对新冠肺炎CT图像进行去噪处理,提升肺炎图像质量。多尺度特征提取模块从不同尺度提取肺炎图像中的纹理特征细节,采用深浅通道并行方式,分别提取肺炎图像中的高维度以及低维度的特征。为进一步优化网络模型,提出一种拆分卷积方式,可将特征图拆分为两类,一类为主要关注特征,另一类为次要关注特征。使用复杂度高的计算方式从主要关注特征中提取关键信息,对于次要关注特征,则采取复杂度低的计算方式提取补偿信息。通过与非局部均值(NLM)去噪算法、收缩卷积神经网络(SCNN)深度模型、去噪卷积神经网络(DnCNN)深度模型对比,以及网络消融实验,可以看出本文提出的模型能有效去除肺炎图像中的噪声,并且可以更好地保留原始图像中的纹理结构细节,为机器以及医生提供更可靠的辅助诊断。
英文摘要:
      The coronavirus disease 2019 (COVID-19) has spread worldwide. To early diagnose COVID-19 and reduce the pressure of medical staff, using deep learning methods to analyze chest computed tomography (CT) images of patients becomes more and more important. The images of pneumonia have rich texture details and fuzzy edge structure, which are easy to interfere with the diagnosis of machine and doctor. COVID-19 CT images denoising method based on multi-scale parallel deep split convolution neural network (MSP-ReCNN) is proposed in this paper to enhance the quality of pneumonia images. Multi scale feature extraction module can extract the details of texture features in pneumonia images from different scales. The parallel method of deep and shallow channels are utilized to extract the high-dimensional and low-dimensional features of pneumonia images. To further optimize the network model, the split convolution method is proposed. The feature graph can be divided into two categories, one is the primary concern feature, the other is the secondary concern feature. High complexity computing method is used to extract the core information from the primary concern features, and the low complexity calculation method is used to extract the compensation information for others. Compared with non-local mean (NLM) denoising algorithm, shrinkage convdutional neural network (SCNN) model, denosing convolutional neural network (DnCNN) model, and through network ablation experiments, it can be drawn that the proposed model can effectively remove the noise in COVID-19 CT images, and can retain the texture structure details of the original image, as well as provide more reliable auxiliary diagnosis for machines and doctors.
查看全文   查看/发表评论  下载PDF阅读器
关闭

分享按钮