Pathology model biases from image centers affect slide analysis accuracy
Do Center Biases Propagate? Robustness of Pathology Foundation Models in Whole-Slide Image Classification
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Computational models used to analyze medical images of tissue slides sometimes learn patterns related to where the images were taken rather than the biology itself. This can cause mistakes when the model is used on new data from different centers. The authors studied how much these center-related biases affect classification performance and tested methods to reduce this problem. They found that the bias does indeed carry through to predictions, varies with the model and method used, and that a common technique to reduce such bias was not consistently helpful.
What this means in practice
- •For medical imaging developers: Improve slide classification tools by selecting pathology models and aggregation methods less prone to center-related biases in whole-slide images.
- •For healthcare data engineers: Assess and monitor the impact of image acquisition center artifacts on pathology AI systems to reduce unreliable diagnostic outputs across hospitals.
Authors
Ilán Carretero, Pablo Meseguer, Rocío del Amor, Valery Naranjo
Abstract
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 Cramér'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 Cramér'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.