Quantifying and Mitigating Hospital Domain Bias in Pathology Foundation Models using Adversarial Feature Disentanglement

Mengliang Zhang

Keywords: pathology image, foundation model, domain bias

Abstract: Pathology foundation models (PFMs) have demonstrated remarkable potential in whole-slide image (WSI) diagnosis. However, pathology images from different hospitals exhibit domain shifts due to variations in scanning hardware and preprocessing. These differences cause PFMs to learn spurious hospital-specific features, severely compromising their robustness and generalizability in clinical settings. We present the first systematic study of this hospital-source domain bias in PFMs. To address the critical trade-off between diagnostic utility and domain predictability, we establish a quantification pipeline and introduce the Robustness Index (RI). Furthermore, we propose a lightweight adversarial framework for feature disentanglement. This framework employs a trainable adapter and a domain classifier connected via a Gradient Reversal Layer (GRL) to remove latent hospital-specific information from frozen PFM representations without modifying the encoder itself. Experiments on multi-center histopathology datasets demonstrate that our approach substantially suppresses domain predictability and achieves significant gains in feature robustness. Crucially, the method maintains or improves disease classification performance, proving its efficacy particularly in out-of-domain scenarios. Our code is provided at: [https://github.com/ana-com-00/anonymous_repo](https://github.com/ana-com-00/anonymous_repo)

Primary Subject Area: Fairness and Bias

Secondary Subject Area: Application: Histopathology

Registration Requirement: Yes

Reproducibility: https://github.com/ana-com-00/anonymous_repo

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 366