The scarcity of labeled medical data often drives the adoption of complex architectures, potentially overlooking the optimization of foundational training strategies. This study systematically investigates the synergistic impact of data augmentation techniques and rigorous training protocols on lung nodule classification. While geometric augmentation is standard in medical imaging, the role of photometric transformation on Computed Tomography (CT)—which relies on standardized Hounsfield Units—remains under-explored. Using a standard ResNet50 backbone on the LUNA16 dataset to control for architectural variance, we demonstrate that naive training yields poor generalization. By implementing a tuned protocol (incorporating cyclical learning rates, class weighting, and early stopping), we uncover a critical insight: photometric augmentation significantly outperforms geometric strategies in clinical feature learning. Our tuned photometric model achieves a superior F1-Score of 0.5267, correcting the high-precision/low-recall imbalance common in existing baselines. Crucially, Explainable AI (Grad-CAM) analysis confirms that photometric noise forces the model to abandon superficial background correlations and focus on nodule morphology. These findings suggest that robust training protocols and photometric invariance are as critical as architectural innovation for reliable medical AI.