Optimizing Privacy and Efficiency in Brain Tumor Classification through Advanced Non-IID Federated Deep Learning

Recent advancements in Artificial Intelligence (AI) have significantly transformed various fields, especially medical diagnostics. However, centralized deep learning models face significant challenges due to the increasing data volumes from edge devices, including issues related to data privacy and heterogeneity. In this study, a federated deep learning model for brain tumor classification using MRI data is presented. The model enhances the MobileNetV3 architecture with Squeeze-and-Excitation (SE) and Convolutional Block Attention Module (CBAM) blocks. A novel non-IID (non-Independent and Identically Distributed) data partitioning strategy, proposed in this study, ensures robust data distribution across clients. Utilizing the Flower framework and the FedOpt algorithm, the model achieved a notable accuracy of 97.40% in eight training rounds, operating efficiently on 15GB of RAM and 10GB of GPU memory. This research underscores the potential of Federated Learning (FL) to address data privacy and resource constraints, demonstrating its efficacy in accurate brain tumor classification. These findings pave the way for future scalable and secure medical AI applications.