120P In development: Fully automatic treatment benefit prediction in advanced colorectal cancer patients

Le résumé fourni par la source

Background: Accurate and reproducible assessment of HER2 IHC is essential for the appropriate selection of targeted therapies, especially in HER2 low or ultralow tumors.Yet, interobserver and inter-scanner variability hinder diagnostic consistency.AI could enhance accuracy and reproducibility, but large-scale, real-world evidence across diverse scanners is scarce.Methods: We analyzed 2,015 HER2 IHC-stained breast cancer slides collected from 14 centers using seven widely deployed scanner systems (Roche, Leica, Olympus, Morphle, Philips, 3D Histech, KFBIO).Consensus ground-truth HER2 scores were established by expert pathologists without AI.Subsequently, two pathologists at each institution scored ∼150 cases manually and, after a washout period, with AI support.Scoring followed ASCO/CAP 2023 HER2 guidelines, with IHC 0 further subclassified into "null" and "ultralow.". Results:In an interim analysis, fully-autonomous scoring by AI demonstrated a 4.1% higher median accuracy compared to manual scoring.Median accuracy across centers and scanners was increased from 75.4% to 80.0% when pathologists used AI for decision support.Accuracy increased from 70.2% to 77.3% on the subset of data from Roche scanners, 79.6% to 83.2% on Leica, and 73.5% to 76.6% on other scanners.Similarly, median interobserver agreement improved from 73.4% at baseline to 86.05% with AI assistance (Roche: 59.4% to 82.5%; Leica: 74.4% to 89.5%; Others: 73.8% to 82.2%).Crucially, when HER2-ultralow was included as a fifth scoring category, median interobserver agreement increased from 68.0% at baseline to 82.0% with AI assistance (Roche: 47.1% to 75.6%, Leica: 69.8% to 86.9%, Others: 67.4% to 77.5%).Conclusions: This follow-up analysis demonstrates that AI assistance consistently improved accuracy and interobserver agreement in HER2 IHC interpretation across scanner platforms.Overall, our findings support the integration of AI into routine diagnostic workflows to reduce variability and enhance reproducibility of HER2 IHC scoring.Particularly in the clinically challenging HER2 low and ultralow categories, AI may improve personalized treatment decision making for HER2-targeted therapies.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.