OncoPredikt: A deep-learning framework for tumor detection and biomarker quantification in breast cancer IHC whole-slide images

Background: Biomarker quantification in breast cancer remains challenging despite standardized protocols. Manual pathologist assessment introduces variability and misses subtle HER2 patterns, with inter-pathologist concordance for HER2 0 vs. 1+ distinction at only 26%. Second-generation antibody-drug conjugates now target HER2-low and HER2-ultralow tumors, but traditional visual assessment fails to reliably identify patients with barely detectable HER2 (faint staining in ≤10% of cells). We present an AI-driven approach for objective tumor detection and automated biomarker quantification of ER, PR, HER2 and Ki67 in Breast Cancer. Methods: AI-based OncoPredikt model was designed and trained on H&E images (130 training, 53 validations from TCGA/in-house cohorts), tested on IHC WSIs for cross-stain generalization for automated Tumor detection. Further, automated biomarker quantification was performed within the detected tumor regions. The workflows were all validated against pathologist annotations. Results: Tumor masks achieved Dice Similarity Coefficient >0.8 on H&E and better results with IHC WSIs. The ER/PR prediction using our AI-based approach shows a weak positive for a negative sample. For 2 pathologist-analyzed IHC 0 (Neg) HER2 samples tested through the AI-based approach yielded Ultra-Low (0+) and (1+) predictions indicating subtle positivity. The algorithm demonstrated promising performance in discriminating 0+ ultra-low (faint staining ≤10% cells) from 0 (no staining), addressing the critical diagnostic gap where HER2-ultralow patients were previously inaccessible for ADC selection. Automated assessments showed strong concordance with pathologist scoring across ER/PR Allred classification and Ki67 proliferation indices. Conclusion: The HER2 ultra-low detection capability directly addresses a recognized bottleneck: visual inspection of faint, incomplete membrane staining in <10% of cells fall below reliable human detection, yet clinical trials confirm ADC efficacy in HER2-low/ultralow disease. The 26% inter-pathologist concordance for 0 vs. 1+ distinction underscores the clinical value of objective quantification. The shift from manual annotation to algorithm-derived tumor masks eliminates observer variability and saves the time of pathologists with heavy workload and reduces subjectivity in reporting. Thus, AI-driven tumor detection coupled with objective biomarker quantification constitutes a meaningful advancement for precision oncology in breast cancer, enabling identification of biomarkers previously invisible to standard pathology assessment. Further validation on large sample sets is still warranted.