Background: Contemporary breast cancer treatment relies on precise diagnostic categorization for tailored therapy. HER2 status, assessed using immunohistochemistry (IHC), is a key biomarker guiding the choice for the antibody-drug conjugate trastuzumab deruxtecan, which is effective even in HER2-low/ ultralow breast cancer (defined as HER2 IHC score 0+, 1+ or 2+ without ISH amplification). While manual slide interpretation of IHC is subjective and prone to inter- and intraobserver variability, digitization potentially enables computer-assisted image analysis to improve accuracy. However, implementation and infrastructure barriers persist, especially in low- to middleincome countries such as the Philippines. Objective: This study evaluated the concordance between manual and AI-assisted HER2 IHC scoring at St. Luke’s Medical Center – Quezon City (SLMC QC) and identified areas of discrepancy and their possible causes. Methodology: This retrospective study included all Invasive Breast Carcinoma, No Special Type (IBC, NST) tumors diagnosed on core needle biopsy and with HER2 IHC performed in SLMC QC from January 1 to December 31, 2024. Original HER2 IHC slides were independently scored by two breast pathology consultants; consensus results served as the ground truth. Slides were digitized using Microvisioneer, and three regions of interest (with ≥ 100 tumor cells) selected by a senior breast pathologist were analyzed using an AI-based image deconvolution algorithm to produce computer image-enhanced HER2 scores. After a minimum of two-week washout, the senior breast pathologist re-evaluated the digitized slides with AI-generated outputs to derive the Computer Image-Enhanced (AI-Assisted) Scores. Concordance between ground truth and AI scores was then assessed. Results: A total of 161 HER2 IHC slides from patients with invasive breast carcinoma were evaluated. Manual and AI-assisted scoring showed an overall percent agreement of 65.22% and a Cohen’s kappa of 0.53, indicating moderate concordance. Weighted analysis demonstrated markedly higher alignment (95.10% agreement; kappa 0.85, almost perfect). Discrepancies were most common among manual scores of 0 to 2+, largely driven by background staining, whereas manual score 3+ classifications remained highly consistent with AI output. Conclusion The study supports the emerging role of AI as a valuable complementary tool in HER2 interpretation. With proper validation and integration, AI-assisted image analysis can strengthen standardization in HER2 assessment, improving precision in patient care especially in resourcelimited settings.
Artificial Intelligence