Deep learning-based analysis and classification of COVID patient through CT images

Abstract

The world is facing a prodigious pandemic initiated through SARS-CoV-2. Corona Infection was first emerged and recognized in Wuhan China which became a global pandemic. Industries all around the world are crippled and many people are killed by this deadly virus. It is not viable to control the spread when this variant of the virus emerged. Therefore, effectively identifying and isolating the people who have symptoms of the disease plays a great role in preventing it. To test COVID 19, the Reverse transcription-polymerase chain reaction (RT-PCR) is the current methodology. Researchers are trying to find different other methods, all over the world, to diagnose coronavirus in affected people. Radiological equipment such as Radiographical images came up as potential alternatives for COVID-19 diagnosis. New coronavirus caused pneumonia in the patients of COVID-19 and it is analyzed by the CT scans. So, CT scan is considered as an effective approach for screening and diagnosing COVID-19. An important rolled could be played by Artificial intelligence and machine learning, in this time of need, only if we can accumulate the available data that can help us pick out the infected patients from the healthy ones. In this research, we have presented technique for the analysis of lungs CT-scan images to classify and detect the infected patient. The proposed methodology is mainly consisted of two-part, classification on original dataset with different loss functions, followed by testing on other datasets and classification on color-mapped dataset followed by fine-tuned transfer learning on other datasets. The lightweight neural network, Efficient-Net B0 is utilized in proposed technique, for classification on publicly available dataset which has largest number of samples. It gives the accuracy of 90.35%. Efficient Net model is also trained using different loss functions, kullback leibier divergence and sparse cross entropy, which gives the accuracy of 99.97% and 99.67%. Different other datasets have also been tested on trained weights of Efficient-Net architecture and accuracy of 99.89% and 88.98% is achieved. We also have color-mapped the dataset into JET color-map and trained the Efficient-Net B0 model, which gives the accuracy of 99.90%, and fine-tuned the trained weights of color-mapped dataset on two different datasets and achieved accuracy of 99.18% and 86.62%, respectively. We have done a comparative analysis through specialized pre-processing techniques.