Abstract
Pathology foundation models (FMs) hold promise for robust
cross-center generalization. Yet out-of-domain (OOD) performance
degradation remains an issue, and a common cause is shortcut learning,
where downstream models exploit center-specific cues confounded with
biological labels. Test-Time Training (TTT) o!ers a potential inferencetime
remedy: an adaptation strategy that requires no labeled targetdomain
data. TTT employs an auxiliary task for test-time (label-free)
supervision to update part of the model before prediction. Whether
it adds value on top of pathology-specific FMs, however, remains an
open question. We developed two pathology-specific TTT approaches:
1) AuxSeg, which uses nuclei segmentation maps derived from Histo-
PLUS, and 2) AuxMag, which utilises the image magnification level, as
auxiliary tasks. We evaluated six pathology FMs on H&E patch classification
under two complementary domain-shift settings: induced center
bias, where biological class was correlated with medical center of origin,
and cross-dataset transfer between independent colorectal cancer
cohorts. We compared classification-only baselines, auxiliary multitask
training, standard per-sample TTT, and online TTT, where test-time
updates accumulate across the test set. Baseline OOD robustness varied
substantially across FM backbones and datasets. Neither auxiliary task
consistently improved performance beyond baseline, and additional testtime
adaptation provided negligible to no benefit beyond multitask training.
Strong pathology-specific FMs, including H-optimus-1, Virchow2,
and Kaiko-Midnight, generalized well, independent of auxiliary adaptation.
These findings indicate that, for OOD H&E patch classification, FM
selection is more e!ective for robust generalization than the investigated
TTT strategies.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/COMPAYL_037.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/COMPAYL_037_supp.pdf
Link to Open Review
BibTex
@InProceedings{KlöPas_Investigating_MICCAISAT2026,
author = { Klöckner, Pascal AND Georgiou, Efthymios AND Nazarian, Javad AND Zlobec, Inti AND Brüningk, Sarah},
title = { { Investigating Test-Time Training for Patch Classification in Pathology } },
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}
}