Table 1 Foundation models applied to breast cancer whole-slide images

Model	Breast applications	Performance (AUC / accuracy)	Endpoints predicted
Prov-GigaPath¹²	Breast cancer subtyping & mutation prediction	AUC ≥ 0.90 for breast cancer subtyping; pan-cancer biomarker detection: +3.3% AUC improvement	ER, PR, HER2 subtypes; genomic mutations (pan-cancer)
CONCH⁴⁹	Weakly-supervised classification on TCGA-BRCA slides	~91.4% accuracy (4 slide-level tasks)	ER/PR/HER2 status; Ki-67; morphology
UNI⁵⁰	Breast cancer classification (e.g., BreakHis); recurrence risk modeling	AUC 0.999 (binary cancer); accuracy 95.5% (8-way); recurrence-risk AUC up to ~0.86	ER/PR; general classification; recurrence risk
Virchow⁵¹	Pan-cancer biomarker detection including breast	Sensitivity 95% / Specificity ~72.5% at threshold	ER/PR/HER2; general cancer detection
Phikon / Kaiko⁵² (TCGA)	Biomarker/mutation prediction on TCGA (including breast)	Similar AUCs to UNI & Prov-GigaPath in mutation tasks (no exact values public)	ESR1, PIK3CA pathway mutations
CTransPath⁵³	Recurrence risk modeling & subtyping	AUC ~ 0.54 for risk (weaker); general ~0.75–0.84 depending on encoder	ER/PR; HER2; recurrence risk

The table summarizes the main clinical endpoints assessed by each model (e.g., hormone receptors, HER2, Ki-67, molecular mutations), the reported performance metrics, and the corresponding peer-reviewed references. To ensure consistency, performance metrics reported in the original studies, most commonly area under the curve (AUC), are included. Both accuracy and AUC, however, have inherent limitations: accuracy is sensitive to class imbalance, while AUC may obscure clinically relevant differences by averaging across thresholds and disregarding model calibration. Only models explicitly applied to breast tissue, or trained on datasets including breast cancer cases, are considered. The listed biomarkers illustrate the capacity of these models to predict clinically meaningful features directly from whole-slide images.

Quick links

Search