Healthcare data is highly sensitive and has challenges in privacy, security, and scalability when shared across medical institutions. To address these challenges, collaborative learning approaches such as Federated Learning (FL) and Gossip Learning (GL) have emerged in recent years. FL enables decentralized model training without sharing raw data but can suffer from communication overhead and inconsistent client participation. GL allows direct client-to-client updates, reducing communication overhead but the model parameters are not shared across all clients, limiting overall improvement. This paper proposed an integrated approach that combines the strengths of both FL and GL for efficient brain tumor detection using CT and MRI image datasets. Experimental evaluation shows that the proposed approach achieved 88.16% accuracy with just three communication rounds across four clients, outperforming traditional FL and GL which require six and fifteen rounds respectively. These results indicate that the proposed approach ensures faster convergence with reduced communication overhead.