Abstract
Federated learning enables collaborative fundus diagnosis without centralizing sensitive data. However, non-IID data across clinical clients, caused by differences in imaging devices, patient populations, and acquisition protocols, often leads to inconsistent client updates and unstable aggregation. To address this issue, we propose Federated Gradient Consensus Alignment (FedGCA), a gradient-level optimization framework for heterogeneous federated diagnosis. FedGCA constructs Local Gradient Basis Memory (LGBM) to capture dominant local gradient-related directions and builds client-specific Cross-Client Gradient Consensus Matrices (CGCMs) on the server to summarize cross-client optimization trends. During local training, FedGCA uses directional consistency to detect conflicting components and adaptively suppresses them before aggregation. Experiments on five multi-source fundus datasets show that FedGCA achieves the best average AUC and F1-score among representative FL baselines. Ablation and convergence analyses further validate the effectiveness of gradient consensus alignment under multi-source data heterogeneity.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/OMIA_004.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=BE7mJmhxQx
BibTex
@InProceedings{PanYij_CrossClient_MICCAISAT2026,
author = { Pan, Yijie AND Cao, Yufei AND Lin, Huiyan AND Li, Haojin AND Shi, Yuanchun AND Yu, Chun AND Li, Heng},
title = { { Cross-Client Gradient Alignment for Federated Multi-Source Fundus Diagnosis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17270},
month = {pending},
page = {pending}
}