Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hospital-Specific Bias in Patch-Based Pathology Models

Mengliang Zhang

TL;DR

PFMs in histopathology can encode hospital-specific information, hindering cross-institution generalization. The authors propose a plug-in adversarial adaptor based on a gradient reversal layer to suppress hospital cues in latent features while preserving disease discriminability on frozen backbones. On TCGA-BRCA patches from multiple hospitals, the adaptor reduces hospital-domain bias while maintaining high disease accuracy, with t-SNE visualizations confirming reduced hospital clustering. This work provides a practical methodology and benchmark for achieving domain-invariant representations in computational pathology, facilitating robust deployment across heterogeneous clinical settings.

Abstract

Pathology foundation models (PFMs) achieve strong performance on diverse histopathology tasks, but their sensitivity to hospital-specific domain shifts remains underexplored. We systematically evaluate state-of-the-art PFMs on TCGA patch-level datasets and introduce a lightweight adversarial adaptor to remove hospital-related domain information from latent representations. Experiments show that, while disease classification accuracy is largely maintained, the adaptor effectively reduces hospital-specific bias, as confirmed by t-SNE visualizations. Our study establishes a benchmark for assessing cross-hospital robustness in PFMs and provides a practical strategy for enhancing generalization under heterogeneous clinical settings. Our code is available at https://github.com/MengRes/pfm_domain_bias.

Hospital-Specific Bias in Patch-Based Pathology Models

TL;DR

PFMs in histopathology can encode hospital-specific information, hindering cross-institution generalization. The authors propose a plug-in adversarial adaptor based on a gradient reversal layer to suppress hospital cues in latent features while preserving disease discriminability on frozen backbones. On TCGA-BRCA patches from multiple hospitals, the adaptor reduces hospital-domain bias while maintaining high disease accuracy, with t-SNE visualizations confirming reduced hospital clustering. This work provides a practical methodology and benchmark for achieving domain-invariant representations in computational pathology, facilitating robust deployment across heterogeneous clinical settings.

Abstract

Pathology foundation models (PFMs) achieve strong performance on diverse histopathology tasks, but their sensitivity to hospital-specific domain shifts remains underexplored. We systematically evaluate state-of-the-art PFMs on TCGA patch-level datasets and introduce a lightweight adversarial adaptor to remove hospital-related domain information from latent representations. Experiments show that, while disease classification accuracy is largely maintained, the adaptor effectively reduces hospital-specific bias, as confirmed by t-SNE visualizations. Our study establishes a benchmark for assessing cross-hospital robustness in PFMs and provides a practical strategy for enhancing generalization under heterogeneous clinical settings. Our code is available at https://github.com/MengRes/pfm_domain_bias.

Paper Structure

This paper contains 7 sections, 6 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Patch Selection
  4. Adversarial Method
  5. Experiment
  6. Results
  7. Conclusion

Figures (3)

  • Figure 1: t-SNE of patch features for IDC in TCGA-BRCA, extracted by the UNI model. Different color represent different hospital source, showing hospital-specific clustering.
  • Figure 2: Adversarial framework for pathology classification: features from a shared foundation model are optimized for disease prediction while suppressing hospital-specific cues via a gradient reversal layer (GRL).
  • Figure 3: t-SNE clustering results of the original patch features using the UNI model (left) and the t-SNE clustering results of the features after adversarial training (right). All patches share the same label, while points in different colors correspond to different hospitals.