Key Takeaways & Clinical Implications
- Significant Diagnostic Accuracy Improvement: Pathologist diagnostic accuracy for differentiating HER2 0 -null versus HER2 0 UL/1+/2+/3+ increased significantly from 81.8% (95% CI: 79.8%–83.7%) with manual review to 88.4% (95% CI: 86.6%–89.9%) with AI assistance (p < 0.001) across 1,487 total readings.
- 35% Reduction in False Negatives: AI support reduced false negative HER2-ultralow classifications by 35% among ground-truth HER2 0 cases.
- 16% Increase in Accurate Patient Identification: AI assistance increased the accurate identification of HER2-ultralow patients eligible for targeted therapy by 16%.
- Enhanced Inter-Pathologist Reproducibility: Inter-observer agreement among pathologists improved significantly from 65.7% with manual review to 76.4% with AI support (p = 0.001).
- High Standalone AI Accuracy: The Ibex Breast HER2 AI solution achieved an overall accuracy of 87.1% (95% CI: 83.3%–90.1%) for HER2 0-UL/null scoring.
- Workflow Optimization via Region Highlighting: Automated Points of Interest (POIs) highlighting enabled pathologists to focus efficiently on the most relevant areas with likelihood for stained tissue during diagnostic review.
Executive Summary
The clinical efficacy of trastuzumab deruxtecan (T-DXd) in HER2-ultralow breast cancer requires accurate differentiation between HER2-ultralow (IHC 0 with membrane staining) and HER2-null (IHC 0 absent membrane staining) specimens. This international multi-center study evaluated 800 breast biopsies and excisions across four European and US centers with 11 reviewing pathologists. The results demonstrate that the Ibex Breast HER2 AI support tool significantly improved diagnostic accuracy (from 81.8% to 88.4%, p < 0.001), reduced false negatives by 35%, and increased inter-pathologist reproducibility to 76.4%, ensuring eligible patients are accurately identified for targeted HER2 therapies.
Background & Objective
The demonstrated effectiveness of trastuzumab deruxtecan (T-DXd) in HER2 0-ultralow (UL; defined as IHC 0 with membrane staining) breast cancer now necessitates clear and accurate differentiation from HER2 0-null (defined as IHC 0 absent membrane staining). Visual differentiation remains challenging in routine practice. This multi-center study evaluated whether the Ibex Breast HER2 AI decision-support tool could enhance pathologists’ diagnostic accuracy and inter-observer reproducibility when distinguishing UL from Null cases, directly supporting accurate patient identification for targeted therapy.
Methods
- Study Design: A two-arm, multi-reader study comparing HER2 IHC scoring performance with and without AI decision support.
- Specimen Cohort: 800 breast tissue specimens (biopsies and excisions) enrolled from four centers in Europe and the United States.
- Readers & Reviews: 11 pathologists (“readers”) from these centers participated, each reviewing 100 to 200 slides with and without AI across 1,487 total slide readings.
- Ground Truth Establishment: Ground truth (GT) was established by majority vote or consensus among three expert breast subspecialists.
- Evaluations: Inter-reader agreement and reader accuracy against ground truth were calculated for manual versus AI-assisted readings.
Results
- Overall Diagnostic Accuracy: Across 1,487 readings, AI assistance significantly increased pathologists’ overall diagnostic accuracy for HER2-null versus HER2 UL/1+/2+/3+ from 81.8% (95% CI: 79.8%–83.7%) during manual review to 88.4% (95% CI: 86.6%–89.9%) with AI assistance (p < 0.001).
- Impact on Ground-Truth HER2 0 Cases:
- Increased UL Detection: Accurate identification of HER2-ultralow patients increased by 16%.
- Reduced Diagnostic Error: False negative classifications decreased by 35%.
- Improved Agreement: Inter-pathologist concordance increased significantly from 65.7% (manual) to 76.4% (AI-assisted, p = 0.001).
- Standalone AI Accuracy: The AI solution alone demonstrated 87.1% accuracy (95% CI: 83.3%–90.1%) for HER2-UL/null scoring.
- Pathologist Workflow Efficiency: Survey feedback confirmed that automated ‘Regions of Interest’ highlighted by the AI enabled pathologists to focus efficiently on the most relevant slide areas for accurate ultralow identification.
Conclusions / Clinical Implications
Multi-center validation confirms that integrating Ibex Breast HER2 into pathology workflows significantly improves diagnostic accuracy and scoring reproducibility when distinguishing HER2-ultralow from HER2-null breast cancer. By reducing false negative diagnoses by 35% and increasing inter-observer agreement, AI decision support enhances accurate patient identification for T-DXd targeted therapy while streamlining routine diagnostic review through automated region highlighting.