COVID-19 lung infection detection using deep learning with transfer learning and ResNet101 features extraction and selection

Abstract

Deep learning artificial intelligent (AI) methods have potential to help improve diagnostic efficiency and accuracy for reading portable CXRs. Public datasets taken from SIRM and Kaggle repositories comprised of COVID-19 (N = 130, 975), normal (N = 138, 1525), bacterial pneumonia (N = 145, 2521), non-COVID-19 viral pneumonia (N = 145, 1342) respectively CXRs were analyzed. On the first dataset, we first extracted 2048 features from last pooling layer of Residual Network 101 (ResNet101) which were fed into selected classifiers. The three-class (Covid-19, normal, viral) yielded highest accuracy of 97.30% using support vector machine linear (SVM-L). This accuracy was further improved to 98.20% by applying the chi-square feature selection method. The four-class using original ResNet101 features yielded highest accuracy of 85.06% which was further improved to 87.01% using chi-square and recursive feature elimination (RFE) feature selection methods. Moreover, using the second dataset, we utilized and optimized robust deep learning methods including densenet201, inception-V3, ResNet101, GoogleNet and VGG-19 using transfer learning approach. The densenet201 yielded the highest performance for three-class (Covid-19, normal, pneumonia) to detect Covid-19 with accuracy (99.92%).The results revealed that feature selection methods improved multiclass classification as dynamic deep feature may contains redundant information. Thus, proposed methods have the potential to improve diagnostic efficiency and accuracy for portable CXRs.