93P An AI-driven computational biomarker from H&E slides recovers cases with low levels of HER2 from immunohistochemically HER2-negative breast cancers

摘要: BackgroundNovel anti-HER2 antibody drug conjugates (ADCs) have shown efficacy in breast cancers (BCs) expressing low levels of HER2 (ie, IHC 1+/2+). A subset of BCs classified as HER2 0 by current IHC methods may express HER2 protein. We sought to define whether BCs expressing HER2 could be accurately detected by deep learning (DL) methods applied to H&E whole slide images (WSIs), using a combination of IHC and HER2 mRNA expression as ‘gold standard’.Methods1479 H&E-stained WSIs from 417 primary BCs were categorized according to HER2 IHC, FISH and HER2 copy number amplification from a cohort of 2188 H&E-stained WSI. All HER2 0 and HER2-low (ie, 1+ and 2+) samples were also tested for HER2 mRNA expression. A SE-ResNet-50 CNN and aggregator were trained from WSIs of H&E sections at 20x. Slide-level predictions were evaluated with 8-fold cross-validation.ResultsA …