Lung Infection Severity Prediction with Parallel Transformers and Score-Correlated Anatomical CutMix Augmentation

Accurate prediction of lung infection severity from chest X-ray (CXR) images is essential for timely and effective intervention. In this article, we propose a transformer-based approach enhanced by a novel data augmentation strategy, Score-correlated Anatomical CutMix (SACM), which incorporates anatomical structures into region-based CutMix operations. By segmenting each lung into six sub-regions and applying a CutMix operation using these regions, our method provides semantically informed image augmentations that improve model generalization. We evaluate our approach on the RALO dataset, comparing various augmentation strategies, and demonstrate superior performance in both Mean Absolute Error (MAE) and Pearson Correlation (PC) metrics for Geographic Extent (GE) and Lung Opacity (LO) scores. Further validation on the external BrixIA dataset confirms enhanced robustness and transferability of our method. Our findings support the utility of anatomically guided augmentations in AI-based image interpretation for clinical decision support. The source code will be publicly available at: https://github.com/Adn-an/LISP-SACM .