COVID-19 Diagnosis Based on CT- Scan of Lungs

Abstract

Novel Virus COVID-19 is a new kind of virus from coronavirus family which causes sickness from common cold to Middle East Respiratory Syndrome and Severe Acute Respiratory Syndrome. Ever since the Novel Virus COVID-19 began back in December 2019 it has cause massive destruction across world. As of May 1, 2021, it has claimed 3.17 million lives. One of the major strategies against this virus is to conduct tests at larger scale. Testing not only helps in treatment of infected patients but also helps states to formulate policies to prevent its spread. CT-scan base diagnosis is more preferred than PCR testing as it has higher sensitivity as compared to PCR testing. Furthermore, it also helps to measure Morbidity of disease. Only problem with CT-scan base testing is that it requires medical professional to generate report which could take hours. The objective of this project is to build web application which will computerize the process for CT-scan based diagnosis of Novel Virus COVID-19 . A Convolutional neural network was trained and deployed at server side which takes CT-Scan image from client side and classify them into Control (negative), Regular and Severe class. During model Training phase in addition to original dataset we also composed a patient wise dataset and train well know architectures such AlexNet, VGG16 and ResNet50 on this dataset. We further modified and tune these architectures to extract best out of them and to improve their accuracies. Out of all architectures we tried modified form of Resnet50 with increased dropout rate and batch normalization in dense layer gave us the highest accuracy. This model showed an accuracy of 90% on testing set .