A systematic comparison of transfer learning models for COVID-19 prediction

Abstract

The pandemic COVID-19 is already in its third year and there is no sign of ebbing. The world continues to be in a never-ending cycle of disease outbreaks. Since the introduction of Omicron-the most mutated and transmissible of the five variants of COVID-19-fear and instability have grown. Many papers have been written on this topic, as early detection of COVID-19 infection is crucial. Most studies have used X-rays and CT images as these are highly sensitive to detect early lung changes. However, for privacy reasons, large databases of these images are not publicly available, making it difficult to obtain very accurate AI Deep Learning models. To address this shortcoming, transfer learning (pre-trained) models are used. The current study aims to provide a thorough comparison of known AI Deep Transfer Learning models for classifying lung radiographs into COVID-19, non COVID pneumonia and normal (healthy). The VGG-19, Inception-ResNet, EfficientNet-B0, ResNet-50, Xception and Inception models were trained and tested on 3568 radiographs. The performance of the models was evaluated using accuracy, sensitivity, precision and F1 score. High detection accuracy scores of 98% and 97% were found for the VGG-19 and Inception-ResNet models, respectively.