Lung cancer remains the deadliest cancer type worldwide, necessitating improved early detection and diagnostic methods to reduce the high mortality rate. Artificial intelligence, particularly deep learning, has shown remarkable success in analyzing medical images and classifying tumors. Despite their success, the majority of these systems have a lack of generalizability. In this work, we provide a novel generalizable method by employing transfer learning with the EfficientNetV2-L model using the LIDCI-IDRI dataset, with further validation using the IQ-OTH/NCCD dataset. The model was pre-trained on the massive ImageNet database, which enabled it to leverage its learning of numerous and valuable features. Our results display significant success, with the model achieving exceptional accuracy rates of 99.31% and a loss of 2.89% during the validation phases on the LIDCI-IDRI dataset. Furthermore, when tested on the external validation dataset, the model demonstrated a commendable accuracy of 96.46% and a loss of 12.3%. Notably, our approach yielded excellent discrimination of 100% using Area under the ROC Curve, F1-score, and specificity metrics, indicating its generalizability, robustness, and potential clinical utility. Additionally, it provides an effective solution to the issue of false negatives and false positives with 0 FP per scan. Looking ahead, future research could explore additional refinements to our methodology and investigate its application in diverse clinical settings, ultimately advancing the fight against this deadly disease.