Federated Learning for Lung Cancer Detection: Comparative Analysis and Visual Interpretability

Artificial Intelligence (AI) has become a crucial tool in the detection of lung can-cer through medical image segmentation. However, traditional AI approaches, which require centralizing sensitive patient data for model training, raise sig-nificant privacy concerns. This project investigates the efficiency of Federated Learning (FL) frameworks in comparison to a conventional centralized AI model. We evaluated seven different state-of-the art Federated Learning frameworks to assess their performance in maintaining model accuracy and scalability. Among these, Per-FedAvg and FedOpt demonstrated efficiency compared to the cen-tralized framework. To further understand the performance of these models, we utilized Gradient-weighted Class Activation Mapping (Grad-CAM) to visually interpret their predictions, ensuring that they focus on medically relevant fea-tures. This project highlights that Federated Learning frameworks, specifically Per-FedAvg and FedOpt, offer promising alternatives to traditional AI methods by providing enhanced performance in lung cancer detection.