Abstract

Shortcut learning and reliance on confounding features have harmed the ability of COVID-19 chest X-ray (CXR) artificial intelligence (AI) models to generalise. I describe an automated quality control (Auto-QC) pipeline developed using the largest COVID-19 CXR dataset curated to date. The aim is to rapidly clean CXR data by automatically standardising or rejecting images, whilst providing labels to identify con- founding features, such as pacemakers and radiographic projection.

Presentation

Presentation unavailable