Abstract
Federated Learning (FL) enables privacy-preserving decentralized training, making it well-suited for healthcare. We study CT super-resolution (SR) in a heterogeneous FL setting, where each client holds CT scans of a different anatomical region. Generative SR models are typically large, leading to high communication overhead, while data heterogeneity across clients further degrades performance. We propose FedDSR, a two-stage lightweight federated framework for heterogeneous CT image super-resolution. In the first stage, we introduce a distribution-aware aggregation strategy that leverages client codebook embeddings to capture semantic similarity across clients, producing robust and representative global encoder updates. In the second stage, each client independently fine-tunes its local decoder on institutionspecific data, enabling personalization without modifying the shared encoder. Experiments on four public CT datasets, each assigned to a separate client, show that FedDSR outperforms state-of-the-art FL baselines. FedDSR achieves 3.2% higher PSNR, 0.33% higher SSIM, and 33.5% lower LPIPS, along with improved FID. The source code is available at https://github.com/xzarulin7/FedDSR
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DGM4MICCAI_006.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=~Suvadip_Chakraborty1
BibTex
@InProceedings{ChaSuv_FedDSR_MICCAISAT2026,
author = { Chakraborty, Suvadip AND Jamir, Obed AND Paul, Angshuman},
title = { { FedDSR: Codebook-Based Distribution Alignment for Heterogeneous Federated CT Super-Resolution } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17258},
month = {pending},
page = {pending}
}