65P Multi-site European study of a fully automated artificial intelligence solution for HER2 IHC scoring in breast cancer

HER2 is a major predictive biomarker routinely assessed for all invasive breast carcinoma (BC). We aimed to assess if an artificial intelligence (AI) solution improves pathologists' concordance and accuracy of HER2 scoring in BC. The cohort included 577 biopsies and excisions with different BC subtypes (IDC, ILC, ± DCIS) from 4 different European sites. HER2 slides were stained with anti-HER2 antibody (4B5, VENTANA) and scanned with different scanners (Leica GT450DX, Philips UFS). This two-arm multi-reader study compared the performance of 10 pathologists ("readers") on HER2 scoring (each reviewed 50-200 slides) unassisted vs. supported by an AI HER2 solution (Galen Breast HER2®), which detects the invasive tumor area, classifies tumor cells based on their staining pattern and derives a slide-level HER2 score by applying 2018 ASCO/CAP guidelines. Both study arms were compared to ground truth established by three breast pathologists ("experts"). Experts' overall inter-observer agreement was 77.5% and for 0/1+/2+/3+ scores it was 84.4%/76%/61.7%/93.3%, respectively. Readers' agreement was significantly higher when assisted by AI (90.1% vs. 77.2% without AI). For the HER2 Low relevant cut-off (0 vs. 1+/2+/3+), readers with AI showed significantly higher sensitivity (98.4% vs. 93.7% without AI), and higher inter-reader agreement (96.9% vs. 89.9%)(p < 0.05). For the 0/1+ vs. 2+/3+ classic regulatory cut-off, a trend towards higher accuracy was observed, with significant improvement in specificity (98.1% vs. 92.1%) and in inter-observer agreement (94.3% vs. 88.2%) (p < 0.05). The AI solution demonstrated high accuracy for HER2 scoring (93.4%) for 0 vs. 1+/2+/3+ and overall (82.7%). A comparison of the AI accuracy on slide subset stained with 4B5 vs A0485 antibodies will be also presented. This study reports an independent multi-site validation of a fully automated AI solution for HER2 scoring in BC. Pathologists supported by AI showed improvements in HER2 scoring consistency, and a trend for better accuracy overall and for the analyzed clinical cut-offs. These results suggest AI may improve reproducibility and standardization of HER2 scoring and will be validated in additional ongoing studies.