AI-based HER2-low IHC scoring in breast cancer across multiple sites, clones, and scanners.

• DOI: 10.1200/jco.2023.41.16_suppl.516
• Publication Date: 2023-06-04T14:02:11Z
• Authors: Patrick Frey; Andreas Mamilos; Evgeny Minin; Ralf Banisch; Stefan Günther; Carolin Schmidt; Niklas Abele; Arndt Hartmann; Peter A. Fasching; Matthias Rübner; Hanna Huebner; Tobias Lang; Ramona Erber
• Keywords: Concordance

516 Background: Assessment of immunohistochemical (IHC) HER2 expression plays a pivotal role in breast cancer diagnostics. In the era HER2-low and HER2-targeted antibody drug conjugates, accurate discrimination defined IHC scores is essential. At same time, scoring suffers from poor interobserver concordance. Artificial intelligence (AI) may optimize this regard to standardization, accuracy efficiency, but previous approaches fail show required consistency across samples different sites, clones scanning hardware. Methods: We have investigated use an AI-based quantifier software support pathologists standardized assessment cancer. Validation specimens were derived four institutions five scanners. Using “region interest” (ROI) version (part I), choose ROI be assessed within whole slide image (WSI). contrast, fully automatic II) analyzes complete WSI. Part I: Three selected one per cohort n = 150 specimens. They scored these ROIs (path-only) according ASCO/CAP 2018 guidelines (each pathologist 50). After 2-week washout period, presented with corresponding AI-suggested results (AI-only), then decided on final (AI-assisted). Scoring times recorded. II: Fully AI without human intervention was analyzed using WSI part I additional 94 WSIs. For both parts, compared clinical workflow ground truth as manually score. Results: discriminating HER2-neg HER2-low/pos cases, AI-assisted AI-only showed agreement rates 91.3% 86.7%, respectively, path-only decisions all (0, 1+, 2+, 3+) individually, interrater-agreement vs. 78.7%, exceeding literature < 70%, mean time being 29 sec 50 sec, respectively; 85.3%. 89.1%/86.2% for cohorts, respectively. Conclusions: Across challenging validation data scanners, very high when cases general scoring. When AI-assistance, reduced by almost 50%. Altogether, demonstrate potential solutions increase efficiency ultimately improve patient outcome.