<![CDATA[Pneumonia remains a major cause of morbidity and mortality, and chest X-ray screening in resource-constrained settings could benefit from lightweight deep learning models. This study provides an integrated evaluation of established techniques, adaptive histogram standardization (AHS), automatic mixed-precision (AMP) training, and FP16 post-training quantization, applied to a MobileNetV2-based pneumonia classifier, using a four-scenario ablation design with repeated hold-out validation across three seeds. The selected model of each scenario reached 91.03–91.42% test accuracy, and peak validation accuracy across seeds did not differ significantly between scenarios (p > 0.05, n = 3, which does not prove equivalence). AMP reduced training VRAM by 47.9%; FP16 quantization reduced model size by 49.6% and increased single-image throughput by 29.9% (1.30× speedup) over its FP32 counterpart, with a test confusion matrix identical to that of the autocast-evaluated FP32 model. False negatives decreased from 14 to 8 (42.9%) in the selected models. In three high-confidence true-positive cases, Grad-CAM maps remained qualitatively similar in location across all stages. In this dataset and configuration, staged optimization improved efficiency without measurable loss of accuracy; external validation and tests on physical edge devices are needed before deployment claims.]]>
Copyrights © 2026
