Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (2): 225-234.DOI: 10.3778/j.issn.1002-8331.2106-0289

• Graphics and Image Processing • Previous Articles     Next Articles

Research on COVID-19 CT Image Classification Method Based on Improved Convolutional Neural Network

WU Chenwen, LIANG Yuxin, TIAN Hongyan   

  1. School of Electronics and Information Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China
  • Online:2022-01-15 Published:2022-01-18



  1. 兰州交通大学 电子与信息工程学院,兰州 730070

Abstract: In response to the COVID-19 discovered in Wuhan, China in December 2019, due to the high false-negative rate of RT-PCR testing and the fact that it takes a lot of time to get the results, research has proved that computer tomography(CT) has become one of the important methods to assist in the diagnosis and treatment of COVID-19. Since there are currently fewer COVID-19 CT data sets publicly available, this paper proposes to use conditional generative adversarial networks for data enhancement to obtain more sample CT data sets, so as to reduce the risk of overfitting. In addition, an improved U-Net network based on BIN residual blocks is proposed to perform image segmentation, and then combined with multi-layer perceptrons for classification prediction. By comparing with network models such as AlexNet and GoogleNet, it is concluded that the BUF-Net network model proposed in this paper has the best performance, reaching an accuracy of 93%. Finally, the Grad-CAM technology is used to visualize the output of the system, which can more intuitively explain the important role of CT images in the diagnosis of COVID-19. The application of deep learning technology to medical images helps radiologists to obtain more effective diagnosis.

Key words: COVID-19, deep learning, CT images, conditional generative adversarial networks(CGAN), U-Net

摘要: 针对2019年12月在中国武汉发现的新型冠状病毒,由于RT-PCR检测具有假阴性率过高且得出结果会花费大量时间等问题,研究证明计算机断层扫描(CT)已经成为了辅助诊断和治疗新型冠状病毒肺炎的重要手段之一。由于目前公开的COVID-19 CT数据集较少,提出利用条件生成对抗网络进行数据增强以获得更多样本的CT数据集,以此降低发生过拟合风险;另外提出一种基于BIN残差块的改进U-Net网络来进行图像分割,再结合多层感知器进行分类预测。通过与AlexNet、GoogleNet等网络模型进行比较,得出提出的BUF-Net网络模型性能最优,达到了93%的准确率。利用Grad-CAM技术对系统的输出进行可视化,能够更加直观地说明CT影像对于诊断COVID-19的重要作用。将深度学习技术应用到医学影像中有助于协助放射科医生获得更为有效的诊断。

关键词: 新型冠状病毒, 深度学习, CT影像, 条件生成对抗网络, U-Net