Abstract 93P
Background
Novel anti-HER2 antibody drug conjugates (ADCs) have shown efficacy in breast cancers (BCs) expressing low levels of HER2 (i.e., 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’.
Methods
1479 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 (i.e., 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.
Results
A total of 1098, 820, 122, and 148 WSIs were categorized as HER2 0, 1+, 2+ and 3+ by IHC staining, respectively. mRNA expression data revealed a range of <7.6 to 11.6 (SD=0.63) for HER2 mRNA expression level. When stratified by HER2 mRNA expression, a cut-off of <7.6 mRNA was selected to represent HER2-null, which included IHC 0 (n=32), IHC 1+ (n=3) and IHC 2+ (n=1). HER2 IHC 0 to 2+ with HER2 mRNA expression >9 and HER2 FISH not amplified were considered as being HER2-low (IHC 0, n=494; IHC 1+, n=562; IHC 2+, n=103). Cases with HER2 IHC 3+ and/or FISH amplification were considered HER2-positive. Model development was based on 417 cases (1479 WSIs) including 32 HER2-null, 292 HER2-low and 93 HER2-positive cases. When distinguishing HER2-low and amplified cases from HER2-null, the model’s performance had an AUC 0.78, a sensitivity of 76%, a specificity of 73%, PPV of 97%, NPV of 21%, and F1=0.85. The model identified 21/25 (84%) HER2 IHC 0 and mRNA<7.6, classified as HER2-null, and 136/167 (81%) HER2 IHC 1+/2+ and mRNA>9, classified as HER2-low.
Conclusions
Our AI system applied to H&E-stained WSIs can distinguish between BCs lacking any HER2 protein and mRNA (HER2-null) and HER2-low tumors, warranting further validation in cohorts of patients treated by new HER2 ADCs to support its use in trials and future clinical decision-making.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
Paige.AI, Inc.
Funding
Paige.AI, Inc.
Disclosure
M. Goldfinger: Financial Interests, Institutional, Full or part-time Employment: Paige. E. Millar: Financial Interests, Institutional, Advisory Role: Paige. M. Hanna: Financial Interests, Personal, Advisory Role: Paige. B. Rothrock, M. Lee, Y. Wang, A. van Eck, L. Trlifaj Tydlitatova, J. Sue, J. Stefanelli, J. Retamero, P. Hamilton, T. Fuchs, D. Klimstra: Financial Interests, Personal, Full or part-time Employment: Paige. Financial Interests, Personal, Full or part-time Employment: Paige. J.S. Reis-Filho: Financial Interests, Personal, Advisory Board: Paige. All other authors have declared no conflicts of interest.