COVID-Net

COVID-Net predicts airspace severity from chest x-ray (CXR) images to quantify lung involvement in patients with confirmed SARS-CoV-2 infection.

Key Features:

• Deep convolutional neural network: Implements a deep convolutional neural network (CNN) architecture tailored for chest x-ray analysis.
• Severity prediction: Predicts airspace severity and extent of lung involvement from CXR images in patients with SARS-CoV-2 infection.
• Transfer learning: Employs transfer learning leveraging over 16,000 CXR images sourced from a multinational cohort of more than 15,000 patients.
• Custom network architecture: Integrates representational knowledge into a custom network architecture optimized for severity assessment.
• Dataset curation: Trained and validated on a curated multi-national patient cohort assembled through the Radiological Society of North America’s RICORD initiative.
• Radiologist validation: Validated against interpretations from two board-certified radiologists with 10 and 19 years of experience.

Scientific Applications:

• Clinical severity assessment: Provides quantitative severity estimates of COVID-19 lung disease from CXR to inform clinical evaluation.
• Patient triage and prioritization: Aids prioritization of patient care and resource allocation by estimating airspace severity.
• Benchmarking against expert interpretation: Enables comparison of automated severity assessments with experienced radiologist interpretations for validation studies.
• Research and method development: Serves as a foundation for research in clinical decision-support systems and machine learning for medical imaging.

Methodology:

Deep convolutional neural network trained via transfer learning on a curated multinational CXR cohort (~16,000 images from >15,000 patients from the RSNA RICORD initiative), using a custom architecture that integrates representational knowledge and validated against two board-certified radiologists.