Segmenting and Classification of Covid-19 in Lung CT Scan Images Using Various Transfer Learning Algorithms and Performance Enhancement by Ensemble Based Approaches

Abstract

Covid-19 Pandemic has affected India’s economy and way of life. Early detection is the key to avoid further spread. For accurate detection of Covid-19 on different imaging modalities, many Deep Learning based image processing algorithms have been proposed. This paper proposes a methodology to detect Covid through CT scans of the Lungs. They are segmented to remove the noise and then classified using various transfer learning-based approaches. U-Net is used for segmenting the CT scans. The ground truths for segmented CT scans required for U-Net were created using Otsu’s global thresholding algorithm which produced higher PSNR value of 2.3898 db and lower MSE value of 37800.5621 when compared to other algorithms such as K-Means, Watershed, Fuzzy C Means and local thresholding. U-Net produced 97.50% accuracy. IOU for Covid and Non Covid images were found to be 94.22% and 94.87% respectively. Dice Coefficients for Covid and Non Covid images were found to be 97.02% and 97.37% respectively. These segmented images were given as input to the classification algorithms like Convolution Neural Network (CNN) and transfer learning-based algorithms like VGG-16, VGG -19, DenseNet-169 and DenseNet-201.The accuracies for these models were found to be 85.45%, 88.05%, 88.48%, 94.74%, 98.78% respectively. The DenseNet-201 outperformed all the algorithms. CNN, VGG16, VGG19 showed average performance. In order to make these average learners into strong learners the ensemble models like model averaging, weighted average ensemble, majority voting were considered. The respective accuracies were found to be 91.11%, 91.71% and 89.89%. Weighted Average ensemble performed better than other ensemble-based approaches.