Performance benchmarking of different convolutional neural network architectures on COVID-19 dataset

Abstract

The utilization of chest X-rays could offer valuable assistance in the initial screening of patients before undergoing RT-PCR testing. This potential approach holds promise within hospital environments grappling with the challenge of categorizing patients for either general ward placement or isolation within designated COVID-19 zones. This study investigates the use of chest X-rays as a preliminary screening technique for suspected COVID-19 cases in hospital settings, given the limited testing capacity and probable delays for RT-PCR testing. We assess how well several neural network architectures perform in automated COVID-19 identification in X-rays with the goal of locating a model that has the highest levels of sensitivity, low latency, and accuracy. The results reveal that InceptionV3 exhibits better robustness while MobileNet obtains the maximum accuracy. This strategy may help healthcare organisations better manage patients and allocate resources optimally, especially when radiologists are hard to come by. This will help in choosing an architecture that has better accuracy, sensitivity, and lower latency. The chosen models are pre-trained using the technique of transfer learning to save computation power and time. After the training and testing of the model, we observed that while MobileNet gave the best accuracy among all the models (VGG16, VGG19, MobileNet and InceptionV3), IncpetionV3 was still better when it comes to robustness. © 2024, Bentham Books imprint. All rights reserved.