AURA-CXR: Development-locked Hybrid CNN–ViT and Adaptive Radiomics Stacking with Quantitative XAI for Pneumonia Detection
Pneumonia detection on chest X-rays is difficult because findings are subtle and images vary across datasets.Combining representations can help, but components must be selected without reusing test data.AURA-CXR stacks a convolutional network (DenseNet121-XRV), a Vision Transformer (EVA-X-S), and entropy-guided Adaptive Radiomics via Logistic Regression.Of 26,684 RSNA patientIds, 21,346 formed the development cohort and 5,338 the locked test.Every component, coefficient, and threshold was fixed from out-of-fold development predictions before testing.The frozen stack reached a locked-test ROC-AUC of 0.8988, above either encoder alone and slightly above a same-split ConvNeXt V2 Tiny baseline (0.8922; 0.8960 in an initial run).Without adaptation, it achieved 0.9309 on the pediatric Kermany cohort.Ablations attributed the RSNA gain to combining both encoders, not to radiomics or adaptive wavelet selection.Deep-branch attribution maps localized better than simple spatial priors but were methoddependent.AURA-CXR offers a selection-locked, reproducible protocol measuring each component's contribution.
Authors
- Chastine Fatichah (ORCID: https://orcid.org/0000-0002-7348-9762)
- Dwi Sunaryono (ORCID: https://orcid.org/0000-0001-9925-9263)
- Maria Marcelina Widyastuti
Institutions
- Sepuluh Nopember Institute of Technology (ID)
Publication Details
- Journal
- International journal of intelligent engineering and systems
- Published
- 2026-11-17
- DOI
- https://doi.org/10.22266/ijies2026.1231.59
- Primary Topic
- COVID-19 diagnosis using AI
- Type
- article
- Field-Weighted Citation Impact
- 0.00