Evaluating Pneumonia Classification Robustness across Architecture under Variable-Quality Chest Radiographs

In real-world and resource-constrained settings, chest radiographs may be acquired under variable technical conditions, raising concern that image degradation can weaken classification robustness and reduce the reliability of downstream AI systems. This study examined how controlled image degradation affects pneumonia classification and whether four preprocessing branches mitigate that effect in an architecturedependent manner. A balanced binary cohort was derived from CheXpert by selecting frontal chest radiographs with definite pneumonia labels and constructing a 1:1 pneumonia versus non-pneumonia subset. Data were partitioned at patient level into approximately 70% training, 20% validation, and 10% held-out testing to avoid leakage across splits. Three model families, Vision Transformer (ViT), EfficientNet-B0, and ResNet101V2,
were trained on four branch variants: raw, enhanced, lung-masked, and enhanced-then-lung-masked. The held-out test split was transformed into seven image-degradation variants: Gaussian noise, Gaussian blur, low contrast, underexposure, rotation, clipped anatomy, and JPEG compression. ViT achieved the strongest degraded-average discrimination overall, with mean AUROC 0.688 and mean AUPRC 0.645. The most robust branch depended on architecture: enhancement favoured ViT, whereas lung masking favoured EfficientNet-B0 and ResNet101V2. The highest clean validation AUROC within each family was achieved by the raw branch, but this did not remain the most robust branch under
degradation. Gaussian noise was the most damaging degradation overall. These findings indicate that no single preprocessing pathway is uniformly optimal under variable image quality and support more trustworthy, quality-aware selection or routing of architecture–preprocessing configurations at deployment, particularly where repeat acquisition or extensive manual quality control may be limited.