Decision Support System to Diagnose COVID-19 through Chest X-Ray Analysis

Abstract

The emergence of COVID-19 in early December 2019 has caused enormous damage to health and global well-being. It is still spreading worldwide, with some countries/ areas under lockdown. Although vaccinations for emergency use are available, no proven and tested vaccination is available in the market today to make a person entirely immune to COVID-19. Hence, continuous testing and monitoring are vital to hinder the spread of COVID-19. Today, PCR with reverse transcription (RT–PCR) is the choice test for diagnosing COVID-19. As an alternative, Rapid antigen test kits are also used. However, hospitals, especially in rural areas in countries like Sri Lanka, are deprived of these test kits. Also, except for a few countries, most people now tend to refrain from testing since they are more accustomed to COVID-19, and their fear of it has gradually decreased. Therefore, it is imperative to design an automated decision support system through different means, which can assist in providing fast decision-making with a low diagnosis error. Chest X-ray images and Deep Learning algorithms have recently become a worthy choice for COVID-19 screening. This thesis proposes a model-based decision support system to diagnose COVID-19. It is a multi-class classification system that can classify an X-Ray DICOM (Digital Imaging and Communications in Medicine) object into one of the four COVID-19 Pneumonia classes: ‘Negative for Pneumonia’, ‘Typical Appearance’, ‘Indeterminate Appearance’ and ‘Atypical Appearance’. Furthermore, this decision support system can interpret using a heat map why a specific classification or a decision has been made. This thesis further discusses the three main modules, or the building blocks used to develop this decision support system: Data Pre-Processing Module, Model Training Module and Model Inference Module. The Data Pre-Processing Module describes the pre-processing steps that must be applied to a DICOM object. The Model Training Module focuses on developing the best-performing model. Here the effectiveness of five pre-trained Convolutional Neural Network (CNN) models, namely Densenet121, VGG-16, ResNet-50, Inception-V3 and CheXNet, have been evaluated. The evaluation is done through comparative analysis considering several important factors such as batch size, learning rate, number of epochs, types of loss functions and optimizers. A publicly available DICOM chest X-ray dataset from Kaggle is used to validate the models, and CheXNet obtains the best performance. Finally, the Model Inference Module focuses on the Web Application developed to make inferences from Chest X-Ray images.