Computer-aided COVID-19 diagnosis: a possibility?

Abstract

Coronavirus Disease 2019 (COVID-19) is extremely contagious with a very high mortality rate. Effective and early diagnosis of COVID-19 is therefore crucial when treating patients and limiting its spread. The currently available methods for reliably identifying COVID are time-consuming. Infected people display various symptoms, some of which can be manifested by radiographic imaging such as chest X-rays and CT scans. Recently, many advanced machine learning and deep learning models have been proposed for predicting COVID using chest X-rays and CT scans that have paved the way for computer-aided COVID-19 diagnosis (CACD) systems. Unfortunately, most of these studies employ specific model(s) using a specific dataset making comparison difficult and inconclusive. We still lack a clear picture on which technique is best for a reliable CACD system. In this study, we provide a comprehensive analysis to determine if a CACD system can be developed that can reliably and automatically predict COVID-19 with zero human intervention using currently available tools and techniques? For this purpose, we explore and implement five machine learning models (SVM, LR, RF, KNN and ANN) and three pre-trained deep learning classifiers (VGG-16, Xception and ResNet-50) to compare their performance using 17 benchmark chest X-rays and CT-scans datasets to predict normal and infected samples. Using different classifiers and different datasets, we show that VGG16 with a superior average accuracy (99.10%) is the most suited classifier for CACD when chest X-rays are used. For CT scans, RF can also be used in addition to VGG16 as both records an average accuracy of 93% overall CT scan datasets. Based on the number of experiments, and an average accuracy of 99.10% for the chest X-rays datasets, we conclude that a reliable CACD system is possible.

Ähnliche Arbeiten