Exploring the Efficacy of Data-Decoupled Federated Learning for Image Classification and Medical Imaging Analysis

Muhammad Jahanzeb Khan, Olamide Timothy Tawose, Rui Hu, Dongfang Zhao

Keywords: Federated Learning, Image Classification, Medical Imaging, Data Decoupling

Abstract: The extensive use of medical imaging datasets, like Magnetic Resonance Imaging (MRI), in healthcare research and diagnosis, is often impeded by privacy concerns and computational costs. We address these challenges with a novel solution that integrates federated learning and data decoupling techniques, enabling efficient utilization of medical imaging datasets on high-performance computing (HPC) systems while ensuring data privacy and integrity. Our data decoupling federated learning framework allows hospitals to train a shared model on local MRI datasets without exposing raw data. By separating data management functionalities from the federated learning system, hospitals can utilize their existing HPC resources and maintain control over sensitive data. We extensively evaluated our approach using various image classification datasets, including SVHN, CIFAR10, and a specific medical imaging domain – Brain MRI datasets. Our results indicate improved model accuracy, reduced computational costs, and enhanced scalability while maintaining data privacy. Our work presents federated learning as a promising tool for healthcare, emphasizing the importance of data decoupling techniques in ensuring secure and cost-effective medical imaging data analysis.