Automated Detection of Covid-19 from HRCT Images using Deep Learning

Abstract

Accurate and early diagnosis of coronavirus disease is essential for quick decision-making and patient management. Polymerase chain reactions (PCR) confirm Covid-19 but are limited due to prolonged execution time for analysis. Early disease detection is possible using High Resonance Computerized Tomography (HRCT) images as input in an automated disease prediction model. This is helpful for patients to take early precautionary measures even before consulting a radiologist. Different pre-trained deep neural networks have been used and higher accuracy has been achieved in this model. This study aims to detect disease more accurately and rapidly by differentiating between normal and infected images. The dataset available on Kaggle is divided into 70:15:15 ratios in which training, validation, and test dataset contain 18676, 4000, and 4000 images, respectively. Five different pre-trained deep neural networks ResNet50V2, MobileNetV2, EffecientNetB0, InceptionV3, and Xception Net are compared using an open-source data set to train and validate our proposed model. The proposed model achieves an overall accuracy of 99.9% on ResNet50V2 followed by 99.1% on MobileNetV2, 98.5% on Xception, 98.4% on Inception, and 94.1% on Efficient NetB0. Mobile NetV2 achieves the highest covid-19 sensitivity of 99.5% and ResNet50V2 gives 99.9% covid specificity among all other networks. In future, such prediction models can be used in clinical settings to develop computer-aided diagnosis (CAD) models for clinicians.