Automated Diagnosis of COVID-19 from CT Scans Based on Concatenation of Mobilenetv2 and ResNet50 Features

Abstract

Timely and precise identification of COVID19 is an arduous task due to the shortage and the inefficiency of the medical test kits. As a result of which medical professionals have turned their attention towards radiological images like Computed Tomography (CT) scans. There have been continued attempts on creating deep learning models to detect COVID-19 using CT scans. This has certainly reduced the manual intervention in disease detection but the reported detection accuracy is limited. Motivated by this, in the present work, an automatic system for COVID-19 diagnosis is proposed using a concatenation of the Mobilenetv2 and ResNet50 features. Typically, the features from the last convolution layer of the transfer learned Mobilenetv2, and the last average pooling layer of the learned ResNet50 are fused to improve the classification accuracy. The fused feature vector along with the corresponding labels is used to train an SVM classifier to give the output. The proposed technique is validated on the benchmark COVID CT dataset comprising of a total of 2482 images with 1252 positive and 1230 negative cases. The experimental results reveal that the proposed feature fusion strategy achieves a validation accuracy of 98.35%, F1-score of 98.39%, the precision of 99.19%, and a recall of 97.60% for detecting COVID-19 cases with 80% training and 20% validation scheme. The obtained results are better than the comparison models and the existing state of artworks reported in the literature.