Benchmarking Open-Source Pathology Foundation Models for Breast Cancer Biomarker Prediction from H&E Whole-Slide Images

Abstract

Simple Summary Breast cancer treatment is guided by three molecular biomarkers—estrogen receptor (ER), progesterone receptor (PR), and HER2—typically measured by immunohistochemistry (IHC), a process that requires additional staining, time, and specialist review. Whole-slide image (WSI) foundation models are large pre-trained neural networks that summarize a digital biopsy into a compact representation, raising the possibility of inferring biomarker status directly from routinely stained H&E images. In this study, we compare two open-source pathology foundation models—TITAN and CHIEF—for ER, PR, and HER2 prediction on the publicly available TCGA-BRCA cohort, under a strict patient-level evaluation protocol with 10 random partitions and 95% confidence intervals. ER and PR predictions show robust discriminative performance under retrospective evaluation; HER2 prediction at the default decision threshold remains limited and motivates threshold-calibration and multimodal extensions. The findings are hypothesis-generating and motivate prospective external validation before any clinical use.