Multicenter study evaluating the utility of AI-aided HER2 scoring in the clinically relevant categories.

Abstract

e15150 Background: The availability of multiple therapeutic options for breast cancer patients with varying levels of protein expression necessitates precise and standardized HER2 scoring. This study explored the clinical utility of AI-aided HER2 scoring of whole-slide digital images by pathologists in breast core needle biopsies and excisions. Methods: The study included patients (n=764) with multiple subtypes of invasive breast cancer from 3 US reference laboratories and 1 academic medical center. The non-interventional two-arm multi-reader study retrospectively compared manual digital vs. AI-aided HER2 scoring of 8 pathologists. The fully automated HER2 AI solution detects the invasive tumor area and then classifies the tumor cells based on staining pattern based and the 2023 ASCO/CAP guidelines to output slide-level HER2 score. Both study arms were compared to reference ground truth (GT) established by the consensus of two breast subspecialists, ground truthers (GTer), who reviewed the slides digitally. One site recorded the time taken for HER2 scoring of both arms to determine the efficacy of the HER2 AI solution. Results: The initial interobserver agreement among expert GT pathologists for each HER2 score, 0, 1+, 2+, 3+, were 83%, 63.8%, 46%, and 92.6, respectively. At the clinically meaningful cut-offs, the reader pathologists’ performance improved with the AI support. At the 0 vs. 1+/2+/3+ and 0/1+ vs 2+/3+ clinical cutoffs, the readers' accuracy was 86.3% vs 88.6% and 93.7%, 93.8% with and without AI support, respectively. The AI use contributed to approximately 12%-time savings for the readers at the medical center. The AI solution demonstrated high accuracy for HER2 scoring, with 89.2% agreement with GT for 0 vs 1+/2+/3+ scores and 93.1% for 0/1+ vs 2+/3+ cutoff. The results will be updated in the final submission with the additional site data. The reader feedback survey regarding the solution’s usability will also be presented. Conclusions: The pathologists demonstrated improvements in consistency, evidenced by inter-reader agreement and accuracy with the aid of the AI. The AI solution can improve efficiency toward accurate HER2 scoring. [Table: see text]

Ähnliche Arbeiten