Adaptive Normalization Enhances the Generalization of Deep Learning Model in Chest X-Ray Classification

Image normalization plays a critical role in improving the robustness of deep learning models for chest X-ray (CXR) classification, particularly under cross-dataset variability and domain shift. This study evaluates three normalization strategies: min–max scal-ing, Z-score normalization, and an adaptive approach combining percentile-based ROI cropping with histogram standardization—across four benchmark CXR datasets (ChestX-ray14, CheXpert, MIMIC-CXR, and Chest X-ray Pneumonia) and three CNN architectures (a lightweight CNN, EfficientNet-B0, and MobileNetV2). The adaptive method consistently enhances validation accuracy, F1-score, and training stability compared to conventional normalization. MobileNetV2, in particular, achieves the highest F1-score of 0.89 on the Chest-Xray-Pneumonia dataset under domain shift. Sta-tistical analyses using Friedman–Nemenyi and Wilcoxon signed-rank tests confirm that these performance gains are significant. The results also indicate improved cali-bration and reduced overfitting when adaptive normalization is applied. By address-ing variability in both image quality and ROI localization, this study demonstrates that adaptive preprocessing is an important design choice for developing reliable and gen-eralizable AI models in radiological applications.