Abstract
Histopathology vision foundation models provide rich semantic
embeddings that encode biologically relevant information as well
as acquisition-related variation, such as scanner and staining effects. This
can be problematic for retrieval and other distance-based applications,
where technical variation contributes to embedding distances despite
being irrelevant to the task. We study acquisition variation in frozen
histopathology foundation model embeddings as a structured nuisance
signal that can be attenuated post hoc by affine corrections without
target-domain statistics. Using paired multi-scanner acquisitions from
SCORPION and controlled stain perturbations, we estimate acquisitioninduced
displacements and introduce a soft shrinkage operator that attenuates
them while penalizing changes to the original representation.
We compare this approach with label-based concept erasure and hard
PCA projection. Correction hyperparameters are selected on SCORPION
and transferred without adaptation to two external benchmarks:
PLISM for scanner and stain robustness, and HEST for downstream
preservation. On PLISM, we additionally report a mean-alignment oracle
using target-domain statistics as a target-aware reference. The proposed
shrinkage operator achieves the strongest robustness gains among
source-estimated corrections, increasing joint scanner-and-staining top-
10 retrieval accuracy from 0.183 to 0.407 for H0-mini and from 0.120
to 0.154 for H-optimus-1. Although the oracle remains stronger overall,
soft shrinkage approaches its performance for H0-mini despite using
no target-domain statistics, suggesting that source-estimated acquisition
geometry can transfer across datasets. The remaining gap motivates improved
source-domain criteria for selecting correction strength.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/COMPAYL_049.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
BibTex
@InProceedings{OreVal_Posthoc_MICCAISAT2026,
author = { Oreiller, Valentin AND Ardan, Darya AND Müller, Henning},
title = { { Post-hoc Shrinkage of Acquisition-Induced Variability in Histopathology Vision Foundation Model Embeddings } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17251},
month = {pending},
page = {pending}
}
