Whole Slide Images (WSIs) contain multiple morphological structures, each corresponding to a distinct style. Performing style transfer may potentially shift the region of interests (RoIs) in the augmented WSIs. To address these challenges, we propose HistoFS, a federated learning framework for computational pathology on non-i.i.d. feature shifts in WSI classification.
Specifically, we introduce pseudo bag styles that capture multiple style variations within a single WSI. In addition, an authenticity module is introduced to ensure that RoIs are preserved, allowing local models to learn WSIs with diverse styles while maintaining essential RoIs.
Extensive experiments validate the superiority of HistoFS over state-of-the-art methods on three clinical datasets.
In federated learning for WSIs, each institution is treated as a client, where WSIs exhibit distinct style distributions. The patch features extracted by the SSL-ViT encoder from four different institutions are visualized via t-SNE. It is evident that WSIs from different institutions form well-separated clusters, indicating significant domain shifts.
Previous works have often overlooked these diverse distributions, leading to suboptimal performance. The challenge lies in handling the non-i.i.d. nature of WSIs while preserving key regions of interest (RoIs) during federated training.
HistoFS: Classifying WSIs in a federated setting. In an FL process ①, we construct pseudo bag styles of each WSI and transmit these styles and W (weight matrix of the MIL model) to the server. Then, the server sends back these styles and W to all institutions. In the local update ②, we augment local WSIs with pseudo bag styles transfer and employ the authenticity module to preserve RoIs.
This work was supported by the National Science and Technology Council (NSTC), Taiwan, ROC, under Grant NSTC112-2634-F-006-002.
@article{raswa2025,
author = {Farchan Hakim Raswa and Chun-Shien Lu and Jia-Ching Wang},
title = {HistoFS: Non-IID Histopathologic Whole Slide Image Classification via Federated Style Transfer with RoI-Preserving},
journal = {CVPR},
year = {2025}
}