Uncertainty-Aware adaptive neuro-fuzzy transformer framework for robust multi-center lung disease classification

The accurate multi-class classification of lung disease from chest radiographs remains challenging because of the diagnostic uncertainty, domain shift, class imbalance and over-reliance on the confidence-based thresholds. The existing CNN–fuzzy based hybrid frameworks frequently employ the static membership functions and manually designed rule bases, which limits the adaptability and cross-dataset generalization. In this paper, we propose the Uncertainty-Aware Adaptive Neuro-Fuzzy Transformer (UANFT) framework which jointly learns the feature representations, uncertainty estimation and fuzzy inference parameters in the end-to-end architecture. Proposed method combines the lung-region-aware Swin Transformer backbone, Bayesian predictive uncertainty modeling via Monte Carlo sampling, adaptive Gaussian neuro-fuzzy inference with the learnable membership parameters and domain generalization via correlation alignment regularization. The experimental evaluation on NIH ChestX-ray14, CheXpert and COVID-19 Radiography datasets demonstrates improved cross-dataset robustness (accuracy drop reduced by 3.8%), enhanced minority-class recall (+ 9.6%) and significantly improved the calibration (ECE reduced by 41%) as compared to the static CNN–fuzzy baselines. The proposed framework incorporates uncertainty estimation to enhance prediction reliability, while maintaining computational efficiency for resource-constrained environments.