Effective screening of patients presenting with severe respiratory symptoms remains a cornerstone of managing infectious disease outbreaks at scale. Chest radiography has emerged as one of the more practical screening tools, and deep learning models applied to chest X-ray analysis have demonstrated strong diagnostic accuracy in numerous studies.
However, a significant limitation of many published approaches lies in their lack of transparency: it is often unclear on what basis a model reaches its conclusions. Using explainable artificial intelligence methods, this study reveals that model decisions can be driven by confounding factors present in the image rather than clinically relevant pathology. Following a systematic analysis of such factors in chest X-ray data — including artefacts such as ECG leads — the authors propose a novel method to reduce their influence on model predictions.
The proposed approach proves more robust against these confounders than previously published solutions, while maintaining performance comparable to state-of-the-art methods. By addressing the gap between predictive accuracy and decision transparency, this work contributes to the development of more trustworthy and clinically reliable deep learning tools. Source code and pre-trained model weights are publicly available.
Autors: Tomasz Szczepański, Arkadiusz Sitek, Tomasz Trzciński, Szymon Płotka
Keywords: COVID-19, Deep learning, Explainable AI
Read the article