COViT: Convolutions and ViT based Deep Learning Model for Covid19 and Viral Pneumonia Classification using X-ray Datasets

Artificial Intelligence based Covid19 through X-ray scans has revolutionized early diagnosis and treatment since the outbreak. There have been remarkable achievements in the research of Covid19 from Normal or other Pneumonia X-ray image classification using a convolutional neural network (CNN). CNN alone face problems in describing low-level features and can miss important information. Moreover, accurate diagnosis is important in the medical field with minimum false alarms. To answer the issue, the researchers of this paper have turned to self-attention mechanism inspired by the ViT, which has displayed state-of-the-art performance in the classification task. The proposed COViT method uses convolutions of 3 × 3 instead of patch embedding as in ViT, then alternate self-attention and MLP with hardswish function are added, and finally, MLP head is proposed which has average pooling, fully connected (FC) layer with ReLU function and kernel L2 as a classifier which improves the accuracy. Exhaustive experiments are carried out on three datasets. We have only considered Covid19 and Viral Pneumonia classes for our problem. The proposed model has achieved 98.98% classification accuracy on dataset1, 99.50% on dataset2 and 99.18% on dataset3, which validates the efficiency of COViT The proposed COViT which uses self-attention and CNN shows superiority over other SOTA models and has better accuracy than the methods in the literature.