Abstract
The Chemotherapy Response Score (CRS) is the standard
histopathological metric for grading neoadjuvant chemotherapy response
in high-grade serous ovarian cancer (HGSOC), yet its assessment is limited
by inter-observer variability and requires specialized expert review.
We present a systematic benchmark of histopathology foundation models
for automated CRS classification from routine H&E whole-slide images
(WSIs). On an internal cohort of 368 omental WSIs, we evaluate
seven encoders (TITAN, PRISM, Feather, CONCH v1.5, Virchow2,
H-optimus-1 and UNI v2) using embeddings extracted via TRIDENT.
Four patch-level encoders are combined with four parameter-free, nonlearned
statistical aggregation strategies (mean, max, mean+std and moments
pooling) and task-specific attention-based multiple-instance learning
(ABMIL). All configurations are assessed using stratified 5-fold crossvalidation
and validated on an independent multi-site external cohort of
94 cases. Non-learned aggregation strongly affects performance, and its
internal optimum does not transfer: max pooling achieves the highest internal
AUC but degrades sharply out of distribution, whereas moments
pooling is the most robust non-learned strategy across encoders. ABMIL
improves external discrimination for selected encoders, with CONCH
v1.5 ABMIL reaching the highest external AUC of 0.982. CONCH v1.5
with moments pooling nevertheless provides the best-balanced operating
point at an internally selected threshold frozen for external validation,
achieving external macro-F1 of 0.887 and Cohen’s κ of 0.774, with internal
and external AUCs of 0.981 and 0.959. A magnification ablation
further shows that CRS classification is largely insensitive to input magnification
with coarse inputs matching 40×. These results have important
implications for practical real-time inference for CRS classification
of WSIs.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/COMPAYL_044.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
BibTex
@InProceedings{AbbMar_ABenchmark_MICCAISAT2026,
author = { Abbas, Marwan AND Shahi, Maryam AND Yadav, Siddhartha AND Kaufmann, Scott AND Hart, Steven},
title = { { A Benchmark of Foundation Models for Chemotherapy Response Scoring in High-Grade Serous Ovarian Cancer } },
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}
}