Abstract
Prototype-based federated learning represents each class with compact feature summaries, but alignment to a single mean may suppress within-class variation in heterogeneous images. We test this mechanism in federated ChestX-ray14 classification and ask whether federation is necessary. Across three seeds, two patient-level client allocations, and weight-averaging intervals K∈2,5,20, we compare class-mean alignment with matched averaging-only controls. Alignment lowers the mean representation effective rank by 84–157 dimensions and the mean covariance effective rank by 114–129 dimensions across the six federated designs. Predictive effects vary across settings. Mean paired differences range from -0.0165 to +0.0073 in area under the receiver operating characteristic curve, or AUROC, and from -0.0022 to +0.0114 in average precision. Three pooled training pairs reproduce the concentration. Their mean log ratio of aligned to control covariance effective rank is -4.4241, and their mean AUROC difference is -0.0444. On 3,000 VinDr-CXR images, all 21 paired comparisons show lower effective ranks and a larger leading variance fraction after alignment. Class-mean alignment concentrates representations across training designs, but its geometry does not consistently track predictive benefit.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DeCaF_001.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=nkCer6ntgE
BibTex
@InProceedings{YanJin_Class_MICCAISAT2026,
author = { Yang, Jingjing AND Zhao, Xia},
title = { { Class Prototype Alignment Concentrates Chest X-ray Representations in Federated and Pooled Training } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17252},
month = {pending},
page = {pending}
}