FedTED: Federated Learning for Robust Thyroid Eye Disease Detection with Masked Autoencoders

Wai Tak Lau, Angela McCarthy, Ye Tian, Christopher Nielson, Rashmi Chelliah, Sina Gholami, Andrea Kossler, MINHAJ NUR ALAM, Lora Dagi Glass, Kaveri A. Thakoor

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Federated Learning, Self-Supervised Training, Thyroid Eye Disease

TL;DR: We showcase the power of federated learning, contrastive pretraining, and masked autoencoders to enable robust thyroid eye disease detection by combining data across different sites while retaining data privacy.

Abstract: Thyroid eye disease (TED) detection presents diagnostic challenges due to its heterogenoous clinical presentation and limited data availability across institutions. In this paper, we propose FedTED, a privacy-preserving framework that integrates Federated Learning and self-supervised pretraining to detect TED from external facial images without sharing sensitive patient data. Our method integrates masked autoencoders to capture rich representations, followed by fine-tuning under a federated setting. We evaluate FedTED across different training regimes and show that federated MAE-based models outperform supervised baselines, achieving highest performance - AUC up to 98.70% - across validation folds. These results demonstrate the feasibility and utility of combining federated learning with self-supervised training for sensitive medical applications, particularly in settings with limited data and privacy constraints. In clinical settings, this translates to potential for deploying robust models in the real world for early disease detection.

Track: 3. Imaging Informatics

Registration Id: J5N3MXMVRR4

Submission Number: 127