Abstract
Federated learning (FL) enables collaborative training of medical image segmentation models without sharing sensitive patient data, making it well-suited for multi-center applications. However, severe data heterogeneity due to differences in imaging devices, acquisition, and protocols leads to domain shifts that degrade the generalization ability of methods. In this paper, we propose \textit{Fed-FGS}, a frequency-aware federated learning framework for robust colorectal polyp segmentation. Fed-FGS combines three complementary strategies: (1) difficulty-aware federated gradient scaling emphasizes clinically challenging small polyps to the global optimization; (2) a Discrete Fourier Transform Hard Thresholding (DFT-HT) augmentation to suppress domain-specific noise while preserving structural semantics; and (3) a composite Dice-BCE-Focal loss to address foreground-background imbalance. We evaluate Fed-FGS on the multi-center PolypGen benchmark using five-fold cross-validation across Centers ~1-5 with Center ~6 reserved for unseen-center evaluation. The proposed Fed-FGS achieves a Dice score of 0.75+-0.02, outperforming the FedGS baseline (0.73 +- 0.01), while demonstrating greater stability under the Small Best selection criterion (0.73+-0.01 vs 0.66+-0.16). These results suggest that Fed-FGS provides a practical and privacy-preserving framework for deploying robust medical image segmentation models across heterogeneous clinical centers.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CaPTion_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/profile?id=~Noha_Ghatwary1
BibTex
@InProceedings{GabAhm_FedFGS_MICCAISAT2026,
author = { Gaber, Ahmed M. AND Tawfik, Noha S. AND Fayed, Salema AND Ghatwary, Noha},
title = { { Fed-FGS: Domain-Generalized Federated Polyp Segmentation via Fourier Hard-Thresholding and Gradient Scaling } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17278},
month = {pending},
page = {pending}
}