Benchmarking of pathology foundation models for assessing HER2 amplification and CEP17 polysomy in breast cancer

Abstract

Abstract Accurate assessment of Human Epidermal Growth Factor Receptor 2 (HER2) amplification and Chromosome 17 (CEP17) copy number is critical for breast cancer treatment, yet diagnostically equivocal (IHC 2+) cases remain a clinical bottleneck. This study systematically benchmarks 14 feature encoders, including 12 Pathology Foundation Models (PFMs), using 2393 whole slide images across four diverse cohorts. We specifically targeted the challenging task of resolving FISH status within equivocal cases. PFM's features generally supported higher downstream predictive performance than features from ImageNet-pretrained baselines. Notably, CLAM classifiers using H-optimus-1 and UNI2-H features achieved Area Under the Curve (AUC) exceeding 0.80 in the equivocal cohorts, compared with approximately 0.60 for CLAM classifiers using ResNet-50 features. Furthermore, we validated the feasibility of predicting CEP17 polysomy directly from H&E images (AUC 0.756). These findings establish that advanced PFMs can effectively capture subtle morphological features, offering a scalable, cost-effective auxiliary tool to refine patient triage and reduce reliance on expensive confirmatory testing.