Optimal ML model for Covid-19 chest X-Ray classification: GAN based augmentation vs Few-Shot Learning

Abstract

"The global impact of the COVID-19 pandemic has become a huge problem all over the world as it mutates itself and spreads the virus rapidly. So there are several studies and experiments in the healthcare sector using Convolutional Neural Network(CNN) models .To support this, there is a requirement to find out the best optimal Machine Learning (ML) model in identifying Covid-19 cases. The study explores on identifying the optimal model to detect the Covid-19 cases when there’s an imbalanced dataset in-placed. Suitable dataset were taken from kaggle with the imbalance counts having the subclasses of Covid-19,Pneumonia and Healthy x-Ray images to facilitate the study. The study employs popular Convolutional Neural Network(CNN) models, such as InceptionV3,ResNet50,VGG16,VGG19 to find the accuracy levels with non-augmented datasets using the traditional Transfer Learning approach and augmented dataset artificially generated through Generative Adversarial Network (GAN) due to the absence of extensive dataset. Then, the Few-shot learning(FSL) model, Prototypical Network, was introduced with pre-trained CNN models(InceptionV3,ResNet50,VGG19) as backbone to check the accuracy level in detecting Covid-19 cases using an original dataset. Experimental results demonstrate that the Few-shot learning model has high overall accuracy levels in all models and it’s almost 97% compared to the overall accuracy levels of GAN based augmented CNN model."