ANALYSIS OF NOVEL MACHINE LEARNING MODELS FOR DIAGNOSIS OF COVID-19 THROUGH RADIOGRAPHY IMAGES

Abstract

COVID-19 had a devastating effect on human lives. Not only it infected millions of people and took precious lives but it also became quite a task to control its rising trajectory. Different researchers found that radiography images of infected patients were quite dissimilar in comparison to normal or viral pneumonia patients. This led scientists think differently in terms of classifying COVID-19 patients rather than relying on expensive and time-consuming PCR tests. Deep learning algorithms provides a completely different dimension to solve critical issues related to human lives. That dimension is different in terms of smartness and efficiency that drives us to a completely different solution is in very lesser time. Also, this solution is data-driven so there is a lesser chance to get the proposed solution wrong. Also, one of the pros of these algorithms is the classification of multiple groups in data which helps to understand the problem better. Deep learning solutions have shown excellency in different fields of life and one of its achievements is in the field of bio-medical image analysis. Using this concept in mind, we have proposed a series of COVID-19 models to classify positive COVID-19 patients with the help of dataset consisting of radiography images. For this purpose, we have used 4 different CNN Architectures i.e., AlexNet, ResNet-50, VGG-16, and Inception V3. We have also used 4 different datasets. So, in total, we have developed 16 models. Another thing that is important to state here is that we have used a new technique called “Transfer Learning”. Second, third and fourth model were developed with the help of first model on each of the architectures. The reason here was to make use of the knowledge learnt with first model. The results received suggested better numbers compared to predecessor models and proved a point that these models can be relied in reaching to conclusion.