<![CDATA[Alzheimer’s disease (AD) is a complicated neurodegenerative disease marked by the slow worsening of memory and cognitive functions. Early detection is a key to effective disease management, but the creation of reliable machine learning models is not an easy task because medical data are limited, and there are strict privacy requirements, such as GDPR and HIPAA, that inhibit the creation of central databases. In this paper, a new privacy-saving model that combines personalized federated learning (PFL) with generative adversarial networks (GANs) is proposed to diagnose Alzheimer’s disease. Under the proposed method, the synthesis of data using GAN is conducted on the client to expand small datasets and counteract the performance loss due to non-identically distributed (non-IID) data across institutions. Massive testing was performed on the OASIS MRI dataset and compared with known federated learning baselines, such as FedAvg, FedSGD, and FedPer. The perFedGAN framework was proposed, and in IID settings, the proposed method achieved an accuracy of 96.41%. In non-IID difficult settings, the proposed method had an accuracy of 92.5%, which is a significant improvement over traditional federated averaging methodologies and also better robustness and personalization among dissimilar clients. This finding demonstrates the importance of using a combination of generative modeling and personalized federated learning to enable the precise, scalable, and privacy-preserving early detection of AD.]]>
