Abstract
Brain tumor detection from medical images, especially magnetic resonance imaging (MRI) scans, is a critical task in early diagnosis and treatment planning. Traditional machine learning approaches often rely on centralized data, raising concerns about data privacy, security, and the difficulty of obtaining large annotated datasets. Federated learning (FL) has emerged as a promising solution for training models across decentralized devices while maintaining data privacy. However, challenges remain in dealing with non-IID (independent and identically distributed) data, which is common in real-world scenarios. In this research, we used a client-server-based federated learning framework for brain tumor detection using MRI images, leveraging VGG19 as the backbone model. To improve clinical relevance and model interpretability, we have included explainability techniques, particularly Grad-CAM. We trained our model across four clients with non-IID data distribution to simulate real-world conditions. For performance evaluation, we used a centralized test dataset, consisting of 20% of the original data, with the test set used collectively for evaluating model performance after completing federated learning rounds. Using a separate test dataset ensures that all models are evaluated on the same data, making comparisons fair. Since the test dataset is not part of the FL training process, it does not violate the privacy-preserving nature of FL. The experimental results demonstrate that the VGG19 model achieves a high test accuracy of 97.18% (FedAVG), 98.24% (FedProx), and 98.45% (Scaffold) than other state-of-the-art models, showcasing the effectiveness of federated learning in handling distributed and non-IID data. Our findings highlight the potential of federated learning to address privacy concerns in medical image analysis while maintaining high performance even in non-IID settings. This approach provides a promising direction for future research in privacy-preserving AI for healthcare applications.