Silent Failures in Degraded Chest X-Rays: Benchmarking Uncertainty Methods for Pneumonia Triage in Resource-Constrained Settings

Automated pneumonia screening from chest X-rays can reduce diagnostic delays where radiologists are scarce. Safe deployment requires that model uncertainty reliably signals imminent misclassification when image quality degrades due to hardware limitations. We benchmark four uncertainty quantification (UQ) methods (softmax confidence, ensemble entropy, MC Dropout, temperature scaling) across three compact backbones (MobileNetV3-Small, EfficientNet-B0, ConvNeXt-Tiny) fine-tuned on PneumoniaMNIST. We evaluate all 13 MedMNIST-C corruptions at five severity levels. While all backbones achieve 83–86% clean accuracy, EfficientNet-B0 accuracy collapses to ≈71% at corruption severity 1, a drop driven by a shifted decision threshold rather than lost discriminative ability: its classification AUROC only falls from 0.915 to 0.76–0.81 over the same corruptions. MobileNetV3-Small and ConvNeXt-Tiny maintain clean-set performance, though ConvNeXt-Tiny exhibits poor calibration (ECE 16.5%). Temperature scaling consistently reduces ECE; ensemble entropy provides the strongest failure detection (AUROC 0.804 clean). Failure detection AUROC exceeds 0.75 at mild corruption for MobileNetV3-Small, ConvNeXt-Tiny, and the ensemble, but not for EfficientNet-B0 (≈0.69 at every severity); MobileNetV3-Small's failure detection also falls below 0.75 by severity 4, demonstrating that image quality monitoring remains essential for safe triage.