Abstract

The world at present is still suffering from COVID-19 pandemic. COVID-19 is caused by SARS-CoV-2, which damages human lungs and results in pneumonia. Pneumonia is a disease in which the lungs become filled with fluid and inflame, leading to breathing difficulties. Sometimes, breathing problems become very severe, requiring proper treatment, including oxygen or a ventilator.
Though RT-PCR tests are the commonly used method to detect COVID-19 virus, radiological tests are often used by doctors to diagnose the disease based on severity level and risk factors. This thesis concentrates on two major issues, automatic detection of the COVID-19 infection using deep learning techniques and the determination of the severity level of the infection to help reduce the manual tasks and burden of the radiologist.
In the first part, different deep neural network architectures including convolutional neural network-based ResNet50, DarkNet19, GoogLeNet, and VGG16 methods along with a self-attention based vision transformer (ViT) approach called COViT-CT are implemented to detect COVID-19 CT-Scan images. The performance of the various architectures are compared using various metrics, such as accuracy, precision, recall (sensitivity), specificity, F1-score and AUC, as well the confusion matrix, and the best architecture with the highest accuracy, which is COViT-CT, is selected for automatic COVID-19 detection.
In the second part, if the CT-Scan image is COVID-19 positive during the first part, then the image segmentation method is used to extract the COVID-19 infected region from the Lung CT-Scan images. The infected region is useful in determining the severity level of the COVID-19 infection, which helps in the diagnosis of the disease.
All the experiments are performed using the SARS-CoV-2 CT-Scan dataset. It is shown that the self-attention based COViT-CT method provides the best performance on the test sets of the above-mentioned dataset.