Source-linked AI summary
Do Center Biases Propagate? Robustness of Pathology Foundation Models in Whole-Slide Image Classification
Ilán Carretero, Pablo Meseguer, Rocío del Amor, Valery Naranjo
TL;DR
The paper examines whether pathology foundation models encode acquisition-center signals that affect whole-slide classification. It tests this under controlled center–label correlations across models, datasets, and aggregators, finding that center information can propagate to predictions and that robustness and ComBat effects vary by setting.
Problem
PFMs may encode acquisition-center signals alongside biological patterns, creating shortcuts that can limit generalization across centers.
Source: §1 · p. 1Method
The study varies training-set class–center correlations in weakly supervised WSI classification, uses AUCC to summarize performance and robustness, and evaluates ComBat.
Source: §1 · p. 1, §2.2 · p. 2, §2.3 · p. 2Results
Center information encoded by PFMs can propagate to WSI predictions, robustness varies by representation and aggregator, and ComBat does not provide consistent mitigation.
Source: §5 · p. 4, §4.2 · p. 4, §4.3 · p. 4Takeaways & Limitations
AUCC provides a unified assessment of predictive performance and robustness, while ComBat’s mitigation of center effects is limited and inconsistent.
Source: §5 · p. 4The evaluation is constrained to controlled Cramér’s V settings, binary tasks with two acquisition centers, two MIL aggregators, and patch-level PFMs.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Pathology foundation models (PFMs) have transformed computational pathology through powerful representation learning from histopathological images. PFMs provide rich, discriminative representations for whole slide image (WSI) analysis, enabling tasks such as slide-level classification under multiple instance learning (MIL). However, these representations may also encode non-biological signals associated with acquisition centers, potentially introducing spurious shortcuts into downstream predictions. In this work, we evaluate center-associated robustness in WSI classification using a controlled training setting with increasing class-center correlations quantified by Cram\'er's V. We benchmark six PFMs across four datasets and two MIL aggregators, while evaluating ComBat as a robustification strategy. We further introduce the Area Under the Cram\'er's V Curve (AUCC) to jointly capture absolute classification performance and its degradation as spurious correlation increases. Results show that center-related information encoded by PFMs propagates to WSI-level predictions, with robustness depending on both the PFM representation and MIL aggregation strategy. Additionally, ComBat harmonization does not provide consistent robustness gains across datasets.
1. Introduction
Pathology foundation models provide representations used for whole-slide image classification, but they may also encode acquisition-center signals. The paper evaluates whether those signals affect downstream predictions under controlled center–label correlations.
1. Introduction
- PFMs provide representations for WSI classification, commonly using MIL because slides are gigapixel-scale and require only slide-level annotations.The introduction describes PFMs as extracting rich representations from histopathological images.Source: §1 · p. 1
- PFMs can encode center-specific technical signatures that may act as shortcuts, displacing biological signal and limiting generalization across centers.Processing, staining, and scanner differences can alter image color, intensity, and other characteristics.Source: §1 · p. 1
- The study tests center-associated robustness under increasing class–center correlation, introduces AUCC, and evaluates ComBat as a mitigation strategy.AUCC jointly captures absolute classification performance and its degradation as correlation increases.Source: §1 · p. 1
2. Methodology
The methodology evaluates how center–label correlations affect WSI classification using a weakly supervised MIL framework. It varies training-set association strength and summarizes performance and robustness across correlation levels with AUCC.
2. Methodology
- The framework combines MIL-based WSI classification, controlled center–label spurious correlations, and robustness quantification.Figure 1 summarizes the three components of the evaluation approach.Source: §2 · p. 1, Figure 1 · p. 2
2.1. Multiple Instance Learning (MIL)
- Each WSI is a bag of image patches with a slide-level label; a PFM encodes patches, MIL aggregates them, and an MLP predicts the slide label.The downstream model is optimized using cross-entropy loss.Source: §2.1 · p. 1
2.2. Center-Label Spurious Correlations
- Training sets vary class–center association while keeping the number of WSIs fixed; the test set remains uncorrelated so degradation reflects reliance on that association.Cramér’s V ranges from 0 for independence to 1 for maximal association.Source: §2.2 · p. 2
2.3. Robustness Quantification with AUCC
- AUCC integrates balanced accuracy across ordered correlation settings, jointly capturing absolute predictive performance and stability as spurious association increases.Higher AUCC indicates a stronger performance–robustness trade-off across the correlation range.Source: §2.3 · p. 2
3. Experimental design
The experiments cover four binary WSI classification tasks across two acquisition centers, using six frozen PFMs and two MIL aggregators. Evaluation varies center–label correlation while keeping test folds uncorrelated.
3.1. Datasets and Preprocessing
- Four binary WSI classification tasks span skin, breast, and lung cancer, with WSIs retained from two acquisition centers for each dataset.The datasets are AI4SKIN, CAMELYON16, TCGA-BRCA, and TCGA-NSCLC.Source: §3.1 · p. 2
- The preprocessing pipeline detects tissue, downsamples slides to 10× magnification, and tiles them into 512 × 512 patches without stain normalization.Input transformations prescribed by each PFM are applied to preserve center-associated appearance.Source: §3.1 · p. 2
3.2. Models and Evaluation Protocol
- The benchmark uses six frozen PFMs—two vision-language and four vision-only models—with ABMIL and TransMIL for slide-level aggregation.The aggregators are trained for 20 epochs using AdamW.Source: §3.2 · p. 2
- Evaluation uses five-fold cross-validation at five Cramér’s V levels, repeats each configuration over ten training-set samplings, and keeps each test fold uncorrelated.Results are averaged across folds and samplings; Table 2 covers AUCC across datasets and MIL aggregators.Source: §3.2 · pp. 2–3, Table 2 · p. 3
4. Results
Center-related information in PFM representations can propagate to slide-level predictions, with robustness varying across representations and MIL aggregators. ComBat does not reliably improve robustness across datasets.
4.1. Representation-Level Center Encoding
- CONCH and KEEP are most robust, while KAIKO shows the strongest center encoding, followed by H-Optimus-1; both vision-language PFMs outperform vision-only models on FM-SI and RI.Lower FM-SI is consistently associated with higher RI; FM-SI measures center-wise clustering and RI contrasts biological- and center-driven neighborhood structure.Source: Figure 2 · p. 3, §4.1 · p. 3
4.2. Robustness to Spurious Correlations
- 1.18 percentage points higher average AUCC for ABMIL than TransMIL, with ABMIL ahead on five of six PFMs.BACC degrades as center–label correlation rises, although CONCH and KEEP remain notably robust on TCGA-BRCA at high V.Source: Figure 3 · p. 3, §4.2 · p. 4
4.3. Effect of ComBat Harmonization
- 34 of 48 configurations show lower AUCC under ComBat and 14 improve, for an average change of −0.63 percentage points.Effects vary by dataset, with frequent improvements on CAMELYON16 but almost systematic degradation on TCGA-BRCA.Source: §4.3 · p. 4
5. Conclusion
Center information encoded in PFM representations can propagate to downstream WSI predictions under spurious center–label correlations. The evaluation is constrained to controlled binary tasks with two acquisition centers, two MIL aggregators, and patch-level PFMs.
5. Conclusion
- Center information encoded at the representation level can propagate to WSI classification under spurious center–label correlations; the evaluation is limited to binary tasks with two centers, two MIL aggregators, and patch-level PFMs.The study evaluates six PFMs across four datasets; broader clinical settings and slide-level foundation models are identified as future extensions.Source: §5 · p. 4