ORSAN: A ResNet-based Model Optimized for COVID-19 Diagnosis using X-ray Images

Abstract

The use of deep learning (DL) models for analysing X-ray pictures in the context of novel coronavirus illness (COVID-19) is the main topic of this research. By simply replacing spatial convolutions with global self-attention in the final three bottleneck blocks of a ResNet, a unique DL technique known as ORSAN is created. A dataset of 5,935 X-ray images gathered from publically available datasets is used to train and assess the ORSAN. Different assessment metrics are utilised to evaluate the model's performance, including classification accuracy, F1 score, recall, precision, and area under the receiver operating characteristics curve (AUC). Preprocessing techniques including resizing, augmentation, and normalisation are used. The experimental findings show that the ORSAN consistently performs the best when classifying X-ray pictures into three groups: healthy lungs, lungs affected by pneumonia, and lungs affected by COVID-19. When utilising ResNet101 as the foundation, the detection accuracy for COVID-19 is 98.74%, and for healthy lungs and lungs damaged by pneumonia, it is 92.08% and 91.32%, respectively. Additionally, using the ResNet152 backbone, accuracy values of 83.68% and 82%, respectively, are obtained for healthy lungs and lungs afflicted by pneumonia. These results demonstrate the possibility of the proposed ORSAN model for precise COVID-19 and pneumonia case classification using X-ray images.