<![CDATA[This study investigates how preprocessing selection influences the cross-domain generalization of convolutional neural networks (CNNs) for pneumonia detection and whether it changes performance rankings among architectures. Five preprocessing conditions—Baseline, Standard Deviation Normalization, Contrast Limited Adaptive Histogram Equalization (CLAHE), Histogram Equalization, and StdDev+CLAHE—were combined with three architectures: a simple CNN, ResNet-50, and EfficientNet-B0. The resulting 15 models were trained on the public Kaggle dataset containing 5,863 chest X-ray images and externally evaluated on 1,800 clinical chest X-rays collected in Central Java, Indonesia. Results showed that preprocessing rankings were architecture-dependent. The simple CNN achieved its highest external Recall with Histogram Equalization (0.9856), although Baseline ranked first internally. ResNet-50 was most stable with Standard Deviation Normalization, which ranked first internally (F1 = 0.9821) and externally (Recall = 0.9944). EfficientNet-B0 achieved the highest external Recall with CLAHE (0.9994), but its lower internal precision (0.77–0.87) indicated bias toward the pneumonia class. These findings demonstrate that high benchmark performance may not transfer directly to regional clinical settings because preprocessing interacts with scanner characteristics, data distribution, and network architecture. Therefore, each preprocessing–architecture pair should be validated on locally acquired images before clinical deployment.]]>
Copyrights © 2026
