کنفرانس بین المللی فناوری اطلاعات و دانش

صفحه اصلی / شانزدهمین کنفرانس بین المللی فناوری اطلاعات و دانش

Handling Data Heterogeneity in Federated Medical Images Classification

نویسندگان :

Alireza Maleki¹ Hassan Khotanlou²

1- دانشگاه بوعلی سینا 2- دانشگاه بوعلی سینا

کلمات کلیدی :

Federated Learning،Data Heterogeneity،Medical Image Classification،Vision Transformer،SCAFFOLD

چکیده :

Deep learning-based medical image classification has significant problems with heterogeneity in the data generated by the variability of imaging equipment, protocols, and patient populations within institutions. Federated Learning (FL) suggests a solution by allowing collaborative model training across institutions while not actually sharing sensitive patient information, thus preserving privacy. However, the decentralized data's Non-Independent and Identically Distributed (Non-IID) nature presents fundamental challenges: data heterogeneity and client drift that lower model convergence and performance. To address these challenges, we propose a novel FL framework that integrates appropriate data augmentation, Vision Transformers (ViT), and the SCAFFOLD algorithm to neutralize client drift and enhance convergence in heterogeneous settings. Our approach supports federated training across decentralized medical facilities without raw data exchange, while preserving privacy and label skew and domain adaptation robustness. With testing on the FED-ISIC2019 dataset, we achieve improved performance, such as 86.02% global accuracy and 0.9759 AUC, over baselines like FedAvg and other state-of-the-art FL algorithms. Experiments confirm the key benefits of SCAFFOLD's control variates and conservative augmentation in stabilizing training and improving minority class handling. The work extends privacy-preserving collaborative learning in healthcare, demonstrating practical utility for real-world multi-institutional deployments. Code available at https://github.com/allirezamaleki/Federated-Medical-Image-Classification