45P Improved accuracy and consistency in breast cancer HER2 scoring using an AI-based diagnostic solution: Multi-site validation supporting CE-IVDR certification

Background:The Residual Cancer Burden (RCB) index is widely used to quantify remaining disease and offer prognostic information in breast cancer (BC).Although adjuvant treatment options for patients with residual tumor have expanded, optimal risk stratification is still challenging.In this study, we evaluated whether deep learning-based quantification of residual tumor cellularity from routine histology can improve prognostic assessment. Methods:We applied an automated, interpretable deep learning workflow to post-NAT whole-slide images.Three quantitative metrics were computed: tumor-tissue ratio (TTR), tumor-stroma ratio (TSR), and invasive cancer cell density (ICCD).An AIbased residual cancer burden (AI-RCB) score was derived from these cellularity measures.Associations with disease-free survival (DFS) and overall survival (OS) were evaluated using uni-and multivariable Cox models in a multicentre Danish cohort comprising ER+/HER2-, TNBC, and HER2+ breast cancers with residual invasive disease.Results: Overall, 1,318 whole-slide images from 527 patients were evaluated.AIbased cellularity estimates showed moderate-to-strong concordance with pathologist assessments (Spearman = 0.53-0.75;all p < 0.001).Increasing AI-quantified residual tumor burden was consistently associated with poorer DFS and OS (all p < 0.02).In adjusted analyses, TTR (HR 1.27, 95% CI 1.09-1.49)and ICCD (HR 1.35, 95% CI 1.11-1.65)remained independent prognostic factors.Substitution of manual cellularity with AI-TTR within the RCB algorithm (AI-RCB) resulted in comparable, and in some models superior, prognostic stratification.