An efficient transfer learning approach for prediction and classification of SARS - COVID-19

COVID-19 or Corona Virus Disease is a dangerous disease that spreads quickly and affects human life brutally, becoming the cause of death of millions of people. COVID mainly infects the lungs of human beings and that's why lung images are mostly used for Covid detection. In this research paper, an innovative efficient Transfer Learning approach for Covid -19 prediction is described. The dataset contains Computer Tomography images of the Chest and has been collected from two sources: First SARS-CoV-2 CT has 2482 images of Covid and healthy persons. Non-Covid CT scan images of both SARS-CoV-2 positive and negative patients and second UCSD-AI4H COVID CT scan dataset containing 745 CT scan images of both SARS-CoV-2 positive and negative patients. As datasets containing images are collected from various sources, machine learning pre-processing techniques such as CLAHE, and data augmentation, are applied to enhance contrast, quality, and quantity. Then, a popular pre-trained CNN, called VGG16, is used as the base model, as it has shown its ability on the image dataset in terms of very good accuracy in the prediction and classification of the images related to medical diagnosis. Then we build our sequential model by applying ReLU and Softmax functions and trained it with 21,137,986 parameters. The dataset is divided into three parts in the ratio of 80%, 10%, and 10%, for Training, Testing, and Validation, respectively. The model shows a very good classification of Covid-positive and Covid Negative images as it was able to provide an accuracy of 95%. The Precision was 95%, Recall was 95%, and the F-1 Score was 95%.