Domain generalisation challenges in breast cancer molecular classification using foundation models: a cross-cohort...

Area of Science:

• Computational pathology
• Digital pathology
• Artificial intelligence in oncology

Background:

• Molecular classification (e.g., PAM50, IHC) is crucial for breast cancer treatment but often limited by cost and accessibility.
• Foundation models (FMs) with multiple instance learning (MIL) offer a potential solution for predicting molecular subtypes from H&E-stained slides.

Purpose of the Study:

• To evaluate the performance of FMs with MIL for PAM50 subtyping and IHC biomarker prediction across different cohorts.
• To identify factors contributing to performance degradation in external validation settings.

Main Methods:

• Evaluated 13 FMs and 3 MIL architectures using cross-validation on TCGA-BRCA and external validation on CPTAC-BRCA.
• Quantified four hypothesized domain shift factors (staining variability, feature space divergence, morphological separability, prevalence shift) to explain performance drops.
• Utilized exploratory regression analysis to determine factors influencing relative performance drop (RPD).

Main Results:

• Virchow v2 demonstrated strong performance but severe degradation on external validation, particularly for HER2-enriched/Normal-like PAM50 subtypes and HER2-positive IHC.
• Staining variability and feature space divergence were significant univariate predictors of RPD.
• These two factors jointly explained 80.0% of RPD variance in the multivariate model.

Conclusions:

• While FMs with MIL show potential for breast cancer molecular subtyping, domain shift poses a significant challenge, impacting generalizability.
• Staining variability and feature space divergence are key drivers of performance degradation.
• Future research should focus on domain generalization strategies to improve the robustness of AI models in diverse clinical settings.