Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (5): 321-327.DOI: 10.3778/j.issn.1002-8331.2111-0345

• Engineering and Applications • Previous Articles     Next Articles

Dual-Path Network Model Fusing Frequency-Domain Features to Diagnose COVID-19

YANG Yuhang, LIN Min, WANG Changying, ZHONG Yiwen   

  1. College of Computer and Information, Fujian Agriculture and Forestry University, Fuzhou 350002, China
  • Online:2023-03-01 Published:2023-03-01



  1. 福建农林大学 计算机与信息学院,福州 350002

Abstract: CT examination plays an important role in the diagnosis of COVID-19. In order to obtain more characteristic information about COVID-19 from the limited CT chest image set and establish a more sensitive and general diagnostic model, this paper proposes a dual-path network model(Dp-Net) that integrates CT image frequency domain features. The main part of this model adopts ResNet network model. The training process of convolutional neural network is divided into two parts:one is to extract the spatial domain features of CT images, the other is to extract the frequency domain features through Fourier transform, and then the results of the two parts are fused with a certain weight, and then one more time feature extraction is carried out by Layer4 module. Finally, compared with traditional CNN models such as ResNet and VGG, as well as some improved CNN models such as Self-Trans and LA-DNN, on the open COVID-19 CT data set, and compared with fusion schemes with different weights. Experimental results show that the Dp-Net model achieves better results on various evaluation indexes.

Key words: COVID-19, CT images, frequency domain, feature fusion, convolutional neural network

摘要: CT检查在新冠肺炎诊断中起着重要作用,为了能够在有限的CT胸部图像集中获得更多有关新冠肺炎的特征信息、建立更加敏感通用的诊断模型,提出了融合CT图像频域特征的双路网络模型(Dp-Net),该模型主干部分采用ResNet网络模型,并将卷积神经网络的训练过程分为两个部分,一部分提取CT图像空间域的特征,另一部分通过傅里叶变换提取频率域上的特征,将两者训练的结果按照一定的权重进行融合,融合后再由Layer4模块进行一次特征提取。在公开的COVID-CT数据集上与ResNet、VGG等传统的CNN模型进行了比较,也与Self-Trans和LA-DNN等一些改进的CNN模型进行了比较,并对不同权重的融合方案进行了比较,实验结果表明提出的Dp-Net模型在各种评价指标上取得了更好的结果。

关键词: 新冠肺炎, CT图像, 频率域, 特征融合, 卷积神经网络