Scholarlink

Abstract <p>Molecular classification guides breast cancer treatment, but PAM50 and immunohistochemistry (IHC) remain costly and unavailable in many settings. Foundation models (FMs) combined with multiple instance learning (MIL) show promise for predicting molecular subtypes from haematoxylin-and-eosin-stained slides, yet most studies report only internal validation. This study evaluates FMs with MIL across cohorts and identifies factors associated with domain-induced performance degradation. We evaluate 13 FMs and 3 complementary MIL architectures for PAM50 subtyping and IHC biomarker prediction using cross-validation on TCGA-BRCA (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\varvec{n=1,079}\)</EquationSource> </InlineEquation>) and external validation on CPTAC-BRCA (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\varvec{n=120}\)</EquationSource> </InlineEquation>). Virchow v2 achieves the best overall performance but exhibits severe degradation upon external validation, consistent across all three MIL architectures especially for HER2-enriched and Normal-like PAM50 subtypes and HER2-positive IHC prediction. Four hypothesised domain shift factors are quantified through exploratory regression analysis to explain relative performance drop (RPD). Staining variability, feature space divergence and morphological separability reach significance in univariate analysis, whilst prevalence shift does not. Staining variability and feature space divergence as covariate-level factors jointly account for 80.0% of RPD variance in the most parsimonious multivariate model (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\varvec{R^2=0.800}\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\varvec{R^2_{\text {adj}}=0.750}\)</EquationSource> </InlineEquation>). Although based on a limited number of class-level observations and therefore exploratory in nature, these findings highlight the need for domain generalisation strategies targeting covariate shift, even when specialised FMs are used as feature encoders.</p> Graphical abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

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

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月11日

查看全文
  • Jesus Fernandez-Romero,
  • Pablo Ramos-Berciano,
  • Manuel Perez-Perez,
  • David Benavides,
  • Antonio Robles-Frias,
  • Jorge Garcia-Gutierrez,
  • Laura Macias-Garcia

摘要

Abstract

Molecular classification guides breast cancer treatment, but PAM50 and immunohistochemistry (IHC) remain costly and unavailable in many settings. Foundation models (FMs) combined with multiple instance learning (MIL) show promise for predicting molecular subtypes from haematoxylin-and-eosin-stained slides, yet most studies report only internal validation. This study evaluates FMs with MIL across cohorts and identifies factors associated with domain-induced performance degradation. We evaluate 13 FMs and 3 complementary MIL architectures for PAM50 subtyping and IHC biomarker prediction using cross-validation on TCGA-BRCA ( \(\varvec{n=1,079}\) ) and external validation on CPTAC-BRCA ( \(\varvec{n=120}\) ). Virchow v2 achieves the best overall performance but exhibits severe degradation upon external validation, consistent across all three MIL architectures especially for HER2-enriched and Normal-like PAM50 subtypes and HER2-positive IHC prediction. Four hypothesised domain shift factors are quantified through exploratory regression analysis to explain relative performance drop (RPD). Staining variability, feature space divergence and morphological separability reach significance in univariate analysis, whilst prevalence shift does not. Staining variability and feature space divergence as covariate-level factors jointly account for 80.0% of RPD variance in the most parsimonious multivariate model ( \(\varvec{R^2=0.800}\) , \(\varvec{R^2_{\text {adj}}=0.750}\) ). Although based on a limited number of class-level observations and therefore exploratory in nature, these findings highlight the need for domain generalisation strategies targeting covariate shift, even when specialised FMs are used as feature encoders.

Graphical abstract