Evaluation of AI-assisted HER2 scoring in breast carcinoma-stained whole-slide images.

e13029 Background: HER2 gene amplification occurs in 15% of invasive breast carcinomas. Given the exclusive effectiveness of HER2-targeted therapy in HER2-overexpressed/amplified carcinomas, accurate HER2 status assessment is crucial in clinical decision-making. Moreover, scores 1+ and 2+ not amplified have been shown to benefit from treatment with the new HER2-based antibody-drug conjugates. In this context, HER2 scoring needs to be highly accurate. We developed an artificial intelligence (AI) algorithm for precise, replicable quantification of HER2 protein overexpression, assisting pathologists in HER2 immunohistochemistry scoring. Methods: We developed a fully automated algorithm for assisting pathologists in HER2 scoring in breast carcinoma stained whole slides images (WSI). 5901mm2 of HER2 immunostained sections (standard protocol, with Roche Ultra, 4B5 Clone) of digitized (with several scanners) tumor tissue were annotated at magnification x5 by 10 senior pathologists and used to train a Deep Learning model for invasive region detection. Hence, 4793 square images (tiles) of size 256 pixels were annotated at the cell level, at magnification x40 and used to train another model. The proposed algorithm automatically detects invasive regions of the tumor and quantifies the tumor cells in these regions. It also characterizes each cell according to its membrane completeness and intensity and provides pathologists a readily interpretable score along with the details of the slide analysis (cells categories distribution and heatmap), in accordance with the 2018 ASCO/CAP guidelines. The proposed algorithm works in an end-to-end manner, without requiring any manual intervention from pathologists. Three breast expert pathologists scored an unseen dataset of 68 WSI. After a washout period of two months, they re-scored these WSI with AI assistance. The performances of the model are evaluated by comparing the inter-observer agreement rate with and without AI, and by comparing the model results with the pathologists ground truth. Results: The model demonstrated an overall balanced accuracy of 90. 0% on the validation dataset. Moreover, the overall inter-observer agreement (Fleiss Kappa) significantly improved, going from 0. 37 without A. I to 0. 64 with AI assistance. Conclusions: The proposed fully automated solution could perform exhaustive tumor cell analysis in invasive tissue, without requiring any pathologist intervention. Moreover, such a solution could help to standardize the routine results between pathologists which can suffer from an important variability, thus improving the targeted therapeutic decision. Confusion matrix on the test set. True Label Label 0 1+ 2+ 3+ 0 10 2 0 0 1+ 2 28 0 0 2+ 0 3 15 0 3+ 0 0 0 6