Data-set class-balancing and the Convolutional Vision Transformer: An analysis of chest radiographs for SARS-CoV-2 detection

Abstract

Introduction:The COVID-19 pandemic has had impact on global public health and economies. Chest X-rays (CXRs) play a critical role in screening, especially in resource-constrained regions. This work focuses on the development of a deep learning model for detecting COVID-19 in CXRs images, aiming to address the challenges associated with manual interpretation. Methods:Our approach integrates Convolutional Vision Transformers with traditional Convolutional Neural Networks for CXRs analysis. Using a dataset comprising 5572 CXRs, including various COVID-19 severities and other pneumonia types, the model underwent training across multiple configurations and evaluation using various metrics. Each Dataset Configuration (DSC) involved 10 training iterations. We employed an enhanced attention visualization technique to identify key areas in X-rays. Results:This work presents the top five models based on training and validation outcomes across different radiographic classes. The model’s performance varied with dataset configurations, particularly in distinguishing COVID-19 from other pneumonia types. DSC 6 emerged as the most effective, demonstrating 97.1% sensitivity and a 2.9% false negative rate during training, and maintaining high performance during validation with 87.2% sensitivity and a 12.8% false negative rate. Conclusion:This fusion of engineering and medical expertise yields an efficient screening tool. The COVID-19 crisis underscores the importance of enhanced healthcare preparedness. Our proposed model achieves over 90% average recall, accuracy, and sensitivity, coupled with a low false negative rate. It reliably identifies COVID-19 across different severities and effectively distinguishes it from other pneumonia types, establishing its utility as a robust COVID-19 detection tool.