A Privacy-Preserving Customized CNN-Federated Learning Framework with XAI for Lung Cancer Detection and Classification Using CT Images

Early and precise diagnosis of lung cancer is vital for effective treatment and maintaining overall patient health. In recent years, computer-aided detection (CAD) systems, particularly those based on deep learning (DL), have emerged as powerful tools for accurate lung cancer diagnosis, even in lowresource clinical environments. This study proposes a customized convolutional neural network (CNN)-based federated learning architecture that provides high diagnostic accuracy while addressing the privacy concerns inherent in centralized systems. The deep CNN performs local feature extraction, where the federated approach facilitates collaborative model training while keeping sensitive data at the source. The proposed model was trained using multisource datasets, created by integrating two publicly available datasets to enhance generalization. The model performance was evaluated using three different optimizers, among which AdamW demonstrated the best performance. The model achieved 99.78% training accuracy, 99.34% validation accuracy, and 99.12% testing accuracy, with a minimal loss of 0.0392, demonstrating its robustness. Explainable Artificial Intelligence (XAI) approaches, such as LIME and Grad-cam. were used to visualize the particular lung regions that affect the decisions made by the proposed model, thus, improving interpretability and clinical trust. Comparative analyses and XAI incorporation ensured that the model was clinically dependable and generalized well to effectively detect lung cancer.