A Novel Compressed and Accelerated Convolution Neural Network for COVID-19 Disease Classification: A Genetic Algorithm Based Approach

Abstract

Covid 19 is an infectious disease caused by SARS-Cov-2 virus. It generally affects respiratory system of human and can be fatal if not treated early. It can be caused by coming in contact with an infected person through his/her mouth or nose due to transmission of small liquid particles by way or sneezing or coughing. Since the doctors generally depend on CT scan of suspected patients to confirm if he or she is infected. Proposed research focuses on using CT scan images for Covid-19 diagnosis. In proposed Convolution Neural Network (CNN), there are three convolution layer with 32, 16 and 8 filters in respective layers. The training accuracy of proposed model is 96.71% and testing accuracy is 84.21%. The model was also trained and tested using transfer learning and best test accuracy of 94.73% was obtained using VGG19 pre-trained network. Similarly machine learning methods were also used to classify the images and Random Forest classifier gave best accuracy of 93.33%. Since storage size of pre-trained models was very large hence they were compressed using Genetic Algorithm (GA) without much loss in performance. The VGG16 model could be compressed by 81%, AlexNet by 77.8% and VGG19 by 65.74% without drop in the F1-score. The inference time was also reduced considerably by around 79% for VGG16, 78% for VGG19 and 38% for AlexNet. © 2022, Springer Nature Switzerland AG.