Chest X-rays (CXRs) are widely used for diagnosing respiratory diseases, including the recent example of COVID-19. Supervised deep learning techniques can help detect cases faster and monitor disease progression. However, they are usually developed using coarser data annotations, which may insufficiently capture the heterogeneous disease portrait. We propose the pipeline called CIRCA ( https://circa.aei.polsl.pl ) for a CXR-based screening support system, developed using 6 diverse datasets. Our tool includes lung segmentation, quantitative assessment of data heterogeneity, and a hierarchical three-class decision system using a convolutional network and radiomic features. Lung segmentation showed an accuracy of ~ 94% in the validation and test sets, while classification accuracy was equal 86%, 83%, and 72% for normal, COVID-19, and other pneumonia classes in the independent test set. Three radiomically distinct subtypes were identified per class. In the hold-out set, the classification subtype-specific cross-dataset NPV ranged from 95 to 100%, with PPV from 86 to 100% for all subtypes except N3 (early stage or convalescent) and both C3 and P3 (probable co-occurrence of COVID-19). Using an independent test set gave similar results. The dataset-specific subtype proportions combined with various predictive qualities of subtypes partly explain the widely reported poor generalization of AI-based prediction systems.
Keywords: COVID-19; Chest X-ray; Classification; Deep learning; Heterogeneous disease; Lung segmentation.
© 2025. The Author(s).