Publication

Abstract : Accurate and trustworthy diagnosis from medical images remains a critical challenge in clinical decision support, particularly due to data noise, label ambiguity, and inter-class overlaps. In order to solve this, we introduce an Uncertainty Aware Deep Ensemble Learning Framework, a combination of multi-level feature extraction, Bayesian quantification of uncertainty and fuzzy ensemble fusion to classify thoracic diseases with high accuracy using the NIH ChestX-ray14 dataset. It uses DenseNet, CapsuleNet and EfficientNet backbones with frequency-aware attention and Monte Carlo Dropout-based probabilistic reasoning, as well as with Evidential Deep Learning. The ensemble results are combined through Choquet fuzzy integrals where weighting is done based on confidence. The proposed model has an overall accuracy of 91.4% which is higher than traditional CNN ensembles and uncertainty-agnostic models. Calibration measures like Expected Calibration Error (ECE=0.065) and Brier score (0.078) show that there is a well-matched predictive confidence. In addition, the resistance to noisy inputs and the out-of-distribution rejection is tested on Mahalanobis based distance thresholds. We have found that uncertainty estimation and ensemble diversity can be used to significantly enhance safety, reliability, and clinical trust of automated image-based diagnostic systems.