Early diagnosis and clinical decision-making depend on accurate brain tumor classification using magnetic resonance imaging (MRI). However, traditional deep learning methods usually rely on centralized medical data, which raises privacy concerns and limits the use of distributed clinical data. This research proposes a privacy-preserving federated learning framework for MRI image-based binary brain tumor classification using a decentralized ResNet-18 architecture that enables collaborative training without sharing raw patient data. To reflect realistic clinical conditions, the framework integrates heterogeneous multi-source datasets in different image formats (PNG and JPG) and evaluates performance under both IID and non-IID settings. Experiments were conducted using the Kaggle Brain Tumor MRI dataset and Mendeley Data distributed across five simulated institutions. Within the evaluated experimental setup, the proposed framework achieved approximately 92% accuracy under IID conditions and 91.5% under non-IID settings, with an F1-score of approximately 0.90. Client-level evaluation demonstrated the model’s ability to handle data heterogeneity, while convergence analysis indicated stable training behavior across communication rounds. In addition, Grad-CAM visualization was employed to provide visual interpretability, showing that the model focuses on clinically relevant anatomical regions during prediction. Overall, the results demonstrate that combining federated learning with heterogeneous multi-source MRI data can preserve privacy, maintain robustness and interpretability, and achieve competitive classification performance, highlighting the potential of federated deep learning as a practical and scalable solution for privacy-aware medical image analysis in realistic clinical environments.
