Lung cancer is widely acknowledged as one of the most prevalent cancers worldwide, in large part due to the lack of an initial stage of diagnosis. Early detection significantly improves prognosis and survival rates. Traditional CT scan diagnostics often suffer from inconsistency between radiologists and limited sensitivity in early stage detection. Although ma- chine learning offers potential improvements, privacy concerns remain a significant challenge in multiinstitutional medical data sharing. This study proposes a privacy-focused strategy to lung cancer prediction leveraging Federated Learning (FL) alongside DenseNet-201, using chest CT scan images. The model is trained across multiple clients using the Flower framework, eliminating the need to centralize sensitive patient data. A dataset of 1,000 CT images that include normal and cancer case used. Advanced preprocessing techniques such as resizing, normalization, and grayscale-to-RGB conversion are applied to prepare the data. The experimental results show that DenseNet-201 achieves high accuracy across both clients. For Client1, it recorded 90.47% accuracy, 91.35% precision, 90.47% recall, and 90.35% F1 score. For Client2, it achieved 90.72% accuracy, 91.15% precision score, 90.72% recall score, and 90.74% F1-Score. These results confirm the reliability and performance of the federated learning method in preserving data privacy while achieving strong predictive performance. EfficientNet was integrated as a high-performance alternative, improving the model by balancing accuracy and com- putational efficiency. Future work includes integrating radiomic features and expanding datasets between institutions. Hybrid deep learning models will also be explored to increase accuracy and clinical interpretability.