A Comparative Study of Deep Learning Architectures for COVID-19 X-ray Classification

Abstract

The swift transmission of COVID-19 has required the creation of effective diagnostic instruments, with X-ray imaging becoming a crucial asset for early identification. This study involved a comprehensive examination of ten models, consisting of eight deep learning models and two classical machine learning models. The analysis was conducted using a dataset of X-ray pictures related to COVID-19. We conducted a comprehensive investigation that involved various methods for extracting features and using pre-trained models. This study concluded with a thorough evaluation of the performance of each model. The findings indicated that conventional machine learning models, particularly the Random Forest classifier, attained the best level of accuracy (100%). Out of the several deep learning models, VGG and Xception both earned an accuracy rate of 97.5%, whereas DenseNet201 and CNN attained a little lower accuracy of 95%. The accuracy of InceptionResNetV2 was 92.5%. Nevertheless, ResNet50 and EfficientNetB3 exhibited lower performance, with accuracies of 77.5% and 50%, respectively. These findings provide useful insights into the efficacy of different methods for classifying COVID-19 in X-ray pictures, hence helping to the progress of diagnostic techniques. Among the deep learning models, VGG and Xception both attained an accuracy of 97.5%, while DenseNet201 and CNN achieved 95% accuracy. InceptionResNetV2 followed closely with 92.5% accuracy. However, ResNet50 and EfficientNetB3 underperformed with accuracies of 77.5% and 50%, respectively. These findings offer valuable insights into the effectiveness of various approaches for classifying COVID-19 in X-ray images, contributing to the advancement of diagnostic methods.