This research examines how image normalization enhances generalization of deep learning models in chest X-ray (CXR) classification. Two key challenges are addressed: inaccurate localization of the region of interest (ROI) and variability in image quality across datasets. Three normalization methods scaling, Z-score, and an adaptive ap-proach are compared using four benchmark datasets (ChestX-ray14, CheXpert, MIM-IC-CXR, and Chest X-ray Pneumonia) and three architectures: lightweight CNN, Effi-cientNet-B0, and MobileNetV2. Results show that adaptive normalization consistently improves validation accuracy, convergence stability, and F1-score, especially with Mo-bileNetV2. This configuration achieves the highest F1-score of 0.89 under domain shift. Statistical analyses using Friedman-Nemenyi and Wilcoxon signed-rank tests confirm the significance of these gains. Compared to conventional methods, adaptive normali-zation offers better calibration and reduced overfitting. These findings support its role as a critical design choice in medical imaging pipelines. Future work includes extending to federated and self-supervised settings to improve scalability and priva-cy. By addressing dynamic, context-aware preprocessing, this study contributes to building more efficient, robust, and deployable AI systems for clinical decision-making in radiology.