Abstract
Reliable medical-image classification requires assessment of accuracy, calibration, and uncertainty under image degradation. We analyze the unit-prior Evidential Mean Squared Error (EMSE) objective with masked Kullback–Leibler (KL) regularization and evaluate four ResNet-18 classifiers for four-class blood-smear classification. We derive confidence–uncertainty bounds and a Brier-loss lower bound, and we prove that masked KL leaves a target-evidence direction along which the objective approaches an unattained zero infimum. Using five optimization seeds, we compare regularized evidential learning, its no-KL ablation, softmax, and 20-pass Monte Carlo dropout on 3242 external images across 12 corruption conditions. Mean clean accuracy exceeds 0.998 for all methods; softmax has the best mean clean Brier score, expected calibration error (ECE), area under the risk–coverage curve (AURC), and error-detection AUROC. Across corruptions, the no-KL ablation has the highest equal-weighted mean accuracy, while reliability rankings vary by corruption and metric. None of 390 paired comparisons is significant after Holm correction. These findings do not establish a general reliability benefit from masked KL and clarify the distinct roles of predictive confidence and evidential uncertainty.