Lung disease classification from chest X-ray images remains challenging due to subtle pathological patterns, low contrast, and domain variability. Although transfer learning (TL) alleviates data scarcity, models pretrained on natural images often suffer from domain shift, limiting their effectiveness in medical imaging. To address these challenges, we propose ATL-ResNeXt50, an adaptive transfer learning framework that integrates a ResNeXt50 backbone with adaptive attention mechanisms, multi-head classification, and external adaptive modules to dynamically mitigate domain shift and enhance feature adaptation. A standardized preprocessing pipeline incorporating CLAHE, denoising, Gaussian blurring, and normalization further improves image quality and consistency. The proposed model was evaluated on non-augmented chest X-rays, augmented chest X-rays, and an independent lung CT-scan dataset for cross-modality validation. ATL-ResNeXt50 achieved accuracies of 94%, 97%, and 98%, respectively, outperforming conventional TL approaches while maintaining computational efficiency with 45 million parameters. The framework demonstrates balanced precision and recall across Normal, Lung Opacity, and Viral Pneumonia classes. Cross-modality evaluation confirms strong generalizability, and interpretability is ensured through Grad-CAM and LIME visualizations. Overall, this study presents a robust, explainable, and efficient adaptive transfer learning approach for lung disease classification under domain variability.