Abstract
Federated learning makes it possible to train diagnostic models across several hospitals without moving patient data, which is especially valuable where data is sensitive and hard to share. Real hospital networks, however, are heterogeneous in many ways at once, and most federated benchmarks reduce a method to a single average score that is dominated by the largest sites and silent about the smallest. In this paper we introduce FedMedSim, a controlled and tunable benchmark that combines four sources of clinical heterogeneity, label skew, quantity imbalance, scanner appearance, and temporal drift, on a shared chest radiograph dataset. We compare five federated algorithms across three heterogeneity regimes of varying imbalance severity, and we evaluate not only aggregate performance but also how each method treats its worst-served site. Every evaluated method overfits, and the apparent ranking depends on how performance is measured. More importantly, the site that fares worst tends to be a small, under-resourced center, and among the methods we test FedBN protects it best. These differences are largely hidden by aggregate evaluation and surface only when each site is measured on its own. Our results suggest that equity metrics should complement aggregate evaluation when multi-site medical models are chosen.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AFRICAI_036.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=UMTYncmy2l
BibTex
@InProceedings{TalMoh_Federated_MICCAISAT2026,
author = { Talhaoui, Mohamed Achraf AND Merrouni, Zakariae Alami AND Frikh, Bouchra AND Ouhbi, Brahim},
title = { { Federated Learning under Combined Clinical Heterogeneity: Evaluation Protocol and Inter-Site Equity } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17264},
month = {pending},
page = {pending}
}