摘要
当前COVID-19病毒流行,肺部CT图像已成为医生对COVID-19患者进行准确诊断和跟踪治疗的重要辅助工具之一,但公开的患者数据集较少及数据获取较为困难的问题,导致不易获得好的深度学习模型对肺部CT图像精准筛查和诊断。针对于此,本文提出了一种基于pix2pixHD的深度网络图像生成模型,用以提高COVID-19 CT图像的生成质量。采用pix2pixHD框架,通过对模型上采样的过程中加入SPADE残差块,并对模型的多尺度鉴别器增加1/8尺度鉴别器,使模型可以生成病灶边界更为清晰的COVID-19患者肺部CT图像。在公开数据集上与现有模型进行的实验比较结果表明,本文提出的模型在COVID-19 CT图像上峰值信噪比与结构相似性分别为12.46和0.327,Fr′echet Inception距离(FID)为57.4,较现有模型具有生成质量更高、生成细节更好、收敛速度更快的优势。
At present,COVID-19 virus is pandemic all over the world.Lung CT images have become one of the important auxiliary tools for doctors to accurately diagnose and follow up the treatment of COVID-19 patients.However,it is difficult to obtain a good deep learning model for accurate screening and diagnosis of lung CT images due to the small number of published patient data sets and difficult data acquisition.We propose a depth network image generation model based on pix2pixHD to improve the generation quality of COVID-19 CT images.By adding spade residual block in the process of upsampling the model and adding 1/8 scale discriminator to the multi-scale discriminator of the model,the improved pix2pixhd model can generate CT images of COVID-19 patients with clearer lesion boundaries.We performed experimental comparisons with existing models on open datasets.The peak signal-to-noise ratio was 12.46.The structural similarity was 0.327.The Fr′echet Inception Distance(FID)is 57.4.The experimental results show that this model has the advantages of higher generation quality,better generation details and faster convergence than the existing models.
作者
高志军
冀远明
史二美
GAO Zhijun;JI Yuanming;SHI Ermei(School of Computer and Information Engineering,Heilongjiang University of Science and Technology,Harbin 150022,China)
出处
《智能计算机与应用》
2023年第5期82-89,97,共9页
Intelligent Computer and Applications
基金
黑龙江省省属高等学校基本科研业务费科研项目(Hkdqg201911)。
