Related Experiment Video
Updated: Jun 25, 2026
Building Up a High-throughput Screening Platform to Assess the Heterogeneity of HER2 Gene Amplification in Breast Cancers
Published on: December 5, 2017
Artificial Intelligence (AI)-Assisted Human Epidermal Growth Factor Receptor 2 (HER2) Scoring in Breast Cancer:
Frederik Deman1, Dusan Rasic2, Glenn Broeckx3
1Department of Pathology, Ziekenhuis aan de Stroom (ZAS), Antwerp, Belgium; Department of Diagnostic Sciences, Faculty of Medicine and Health Sciences, University of Ghent, Belgium.
Purpose:
With trastuzumab deruxtecan demonstrating clinical benefit in human epidermal growth factor receptor 2 (HER2)-low and -ultralow breast cancer, precise discrimination at the lowest end of the HER2 immunohistochemistry (IHC) spectrum has become essential. In this 1-year retrospective study, we evaluated whether a commercially available, in vitro diagnostic-approved artificial intelligence (AI) system could improve the consistency of HER2 IHC scoring tested on all breast cancers from a single center.
Materials And Methods:
In total, 853 HER2 IHC whole-slide images from 581 patients were analyzed and independently scored according to the American Society of Clinical Oncology/College of American Pathologists guidelines by 3 expert pathologists and 1 nonexpert reader, both without and with AI assistance. The AI provided categorical HER2 scores (0-3+) and quantitative cell-level staining metrics. Interobserver agreement was assessed using overall percent agreement (OPA), Conger's kappa (κ), and pairwise Cohen's κ. Logistic regression models were used to investigate factors associated with discrepancies between pathologists and the AI system.
Results:
Before AI assistance, overall multirater agreement among human readers was substantial (OPA, 79.8%; 95% CI, 77.9-81.7; Conger's κ = 0.72, 95% CI, 0.69-0.74) and increased to near-perfect levels following AI assistance (OPA, 88.6%; 95% CI, 86.9-90.1; κ = 0.84; 95% CI, 0.82-0.86). Most discordances clustered around the 10% American Society of Clinical Oncology/College of American Pathologists cutoff and were driven mainly by undercalling of 2+ cases. Discordance was also associated with lower tumor cell counts, biopsy samples, and heterogeneous staining patterns. Among AI-classified HER2-0 tumors, true HER2-null cases without any detectable positive staining were rare, supporting a biological continuum of HER2 expression.
Conclusions:
AI-assisted HER2 IHC scoring significantly improves interobserver consistency and provides quantitative support that helps address key limitations of conventional categorical HER2 assessment.
