Abstract:
The detection of lung cancer at an early stage requires a lot of proper and timely anal- ysis of the computed tomography (CT) images. Nevertheless, the development of high- performance diagnostic models often requires access to large data volumes, gathered in different healthcare centers. This data sharing cannot always take place in the actual clinical context due to the strict patient privacy regulations. This study will introduce a two-phase federated learning-based system that takes privacy into account in the as- sessment of lung cancer in a real hospital. The suggested system enables the involved hospitals to train the deep learning models together without having to share the uncoded CT scans. In this regard, sensitive data about patients are retained in-house within lo- cal servers, and model performance is optimized together. The framework is designed in accordance with the currently available data protection regulations like HIPAA, GDPR, and overall practices of hospital data governance. Stage 1 adopts a federated 2D U-Net trained across two simulated hospital sites using the LIDC-IDRI dataset. The global seg- mentation model yields a Dice score of 0.8568%, precisely delineating lung nodules under heterogeneous, non-IID data distributions. Stage 2 adopts the segmentation-guided lunregions to train a hybrid ResNet50–Vision Transformer classifier for normal, benign, and malignant cases. This results in a federated classifier achieving 98. This work verifies that multi-client FL can preserve diagnostic accuracy compared to centralised training while avoiding inter-hospital data transfer. The proposed framework provides a clinically viable direction for secure AI-assisted lung cancer screening and serves as a scalable foundation for future privacy-preserving medical imaging applications
