COVIDGR Dataset and COVID-SDNet Methodology for Predicting COVID-19 Based on Chest X-Ray Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Diseases

Background:

• Coronavirus disease (COVID-19) diagnosis relies on RT-PCR, CT scans, and Chest X-Ray (CXR) images.
• CXR is a cost-effective tool for COVID-19 detection when advanced diagnostics are unavailable.
• Existing COVID-19 datasets are heterogeneous and biased towards severe cases, hindering deep learning model development.

Purpose of the Study:

• To demystify high sensitivities in current COVID-19 classification models.
• To introduce COVIDGR-1.0, a homogeneous and balanced CXR database with all COVID-19 severity levels.
• To propose the COVID Smart Data based Network (COVID-SDNet) for enhanced COVID-19 classification model generalization.

Main Methods:

• Development of the COVIDGR-1.0 database comprising 426 positive and 426 negative PA CXR views, balanced across severity levels (Normal, Mild, Moderate, Severe).
• Implementation of the COVID Smart Data based Network (COVID-SDNet) methodology.
• Validation of the model's performance on different COVID-19 severity levels.

Main Results:

• COVIDGR-1.0 provides a balanced dataset for training robust COVID-19 detection models.
• COVID-SDNet achieves good and stable accuracy across severe, moderate, and mild COVID-19 severity levels.
• The proposed methodology demonstrates potential for early COVID-19 detection using CXR images.

Conclusions:

• The COVIDGR-1.0 database and COVID-SDNet methodology offer a valuable resource for advancing AI-driven COVID-19 diagnostics.
• This approach can aid in the early detection and classification of COVID-19 patients, particularly in resource-limited settings.
• The study emphasizes the importance of balanced datasets for developing reliable deep learning models in medical imaging.