Federated Deep Learning for Cancer Diagnosis: Combining Heterogeneous Multi-Institutional Data Using GAN-Based Augmentation
Mangalagiri Pallavi, Koushik Reddy Chaganti, Kadali Ravi Kumar, Mothe Rakesh, K Sowjanya Bharathi, Ramula Uttham Sai
Abstract
Cancer diagnosis using deep learning requires large, diverse datasets to achieve high accuracy and generalizability. However, data-sharing restrictions and privacy regulations hinder the centralized collection of medical images from multiple institutions. Federated Learning (FL) enables collaborative model training across decentralized institutions while preserving patient privacy. However, FL suffers from data heterogeneity, where variations in imaging protocols and patient demographics degrade model performance. To address this challenge, we propose an FL framework enhanced with Generative Adversarial Networks (GANs) to generate synthetic medical images, improving model generalization. The GAN-based augmentation reduces class imbalances and enhances the diversity of training data across institutions. Our experimental results demonstrate that the proposed FL + GAN augmentation approach achieves 88.7% accuracy, outperforming standard FL (82.3%) while approaching the performance of centralized learning (91.5%). The results highlight that GAN-enhanced FL significantly improves model robustness and diagnostic accuracy in privacy-preserving, multi-institutional cancer diagnosis.