COViT-GAN: Vision Transformer forCOVID-19 Detection in CT Scan Imageswith Self-Attention GAN forDataAugmentation

Ambita, A. A. E.; Boquio, E. N. V.; Naval, P. C.; Jr.,

The Vision Transformer (ViT) is currently gaining popularity in computer vision circles due to its record-breaking performance and faster training time achieved without relying on convolution operations found in CNN architectures. In this study, the Vision Transformer is applied to the task of COVID-19 detection from computed tomography (CT) scan images, specifically on the COVID-CT and Sars-CoV-2 datasets. Using a model pretrained on the mid-sized ImageNet-21k dataset, results show that even the smallest ViT variant that uses small input patch sizes outperformed cutting-edge CNNs especially on the smaller COVID-CT dataset with only a few hundred training images. Furthermore, generation of synthetic images using a ResNet-based Self-Attention Generative Adversarial Network (SAGAN-ResNet) was employed as a data augmentation method to alleviate the problem of limited data and was found to further improve accuracy by approximately 3% and 2% on the COVID-CT and Sars-CoV-2 datasets, respectively. In addition to being more computationally efficient and scalable than CNNs, ViT also provides representations that allow visualization of areas that are semantically relevant for detection. © 2021, Springer Nature Switzerland AG.