Federated Learning (FL) is a crucial technology in Healthcare 5.0, facilitating collaborative model training while preserving the confidentiality of raw data among institutions. This chapter examines two essential architectural concepts in federated learning: centralized and decentralized models. Centralized federated learning depends on a coordinating server to consolidate model changes, whereas decentralized federated learning eliminates this central authority, utilizing peer-to-peer communication or blockchain for coordination. This chapter begins by presenting basic principles, advantages and disadvantages regarding each architecture, and especially in healthcare environments when data privacy, model robustness, scalability, and regulations play a crucial role. The centralized model is praised for its simplicity and efficiency, but introduces concerns around single-point failure, server trust, and communication bottlenecks. On the other hand, decentralized approaches especially those using blockchain, gossip protocols, or multi-agent consensus mechanisms offer resilience, fault tolerance, and improved transparency, albeit at a higher computational cost. Example real world case studies are presented including examples of implementations from both architectures in disease detection and remote patient monitoring. We elaborate to view the choices of architectures, and in which they provide impact on security (e.g., robustness against adversary attacks), systems latency, energy consumption, or compliance to data protection frameworks (e.g., HIPAA and GDPR). The chapter concludes with a comparative evaluation and design recommendations to guide practitioners and researchers in selecting suitable architectures for different healthcare scenarios.