Deep Learning Architecture Comparison for Pneumonia Detection in Pediatric Chest X-Rays: Ensemble Learning and Cross-Population External Evaluation

Two assumptions dominate deep learning research on pneumonia detection from chest radiographs: that newer architectures outperform older ones, and that intra-dataset evaluation suffices to establish clinical utility. This study examined both on the pediatric Chest X-Ray Images dataset. Nine architectures, spanning classical, efficient-scaling, and modern convolutional networks and Vision Transformers, plus a weighted soft-voting ensemble, were compared under a unified protocol. No post-2020 architecture provided a detectable advantage: InceptionV3 (2016) achieved the best accuracy in the main experiment (98.18%), the order of the top five models changed in an independent retraining, and Bonferroni-corrected McNemar tests found no significant differences among them. With components and weights selected on validation data, the ensemble reached 97.95% accuracy (ROC-AUC 0.997), versus 97.84% for its best component. An external evaluation on an adult RSNA subset, which combines age, institutional, acquisition, and label shifts, revealed markedly different behavior: external accuracy ranged from 71.30% to 86.20%, and VGG16, among the most accurate models internally, was the least robust. Calibration, ablation, quantitative Grad-CAM, and efficiency analyses complemented these findings. In this retrospective setting, architectural recency did not predict performance and intra-dataset accuracy did not anticipate cross-population robustness; external evaluation on independent cohorts remains necessary before clinical use.