147P Blind validation of an AI-based tool for predicting distant relapse from breast cancer HES stained slides

BackgroundCorrectly classifying early invasive breast cancer (eiBC) cases into high-risk and low-risk is considered a key issue in treatment. This classification notably crucial determining the treatment regimen: chemo-endocrine therapy versus, endocrine alone. Last year, we presented RACE, an AI-based tool for assessing risk of distant relapse at 5 years ER+HER2- eiBC patients from HES (hematoxylin-eosin-safran) whole slide images (WSI). In present study, performed one-shot blind validation RACE on independent cohort 676 WSI.MethodsHES WSI ER+/HER2- diagnosed Gustave Roussy 2012 to 2017 included CANTO cohort, constituted dataset (19 relapsed years). We compared performance two most relevant clinical scores: Predict Breast CTS0. To assess performance, proceeded as follows. The scores were first terms both cumulative sensitivity dynamic specificity accuracy identify relapses. For this purpose, each score has been binarized (low risk/high risk) with respect threshold that set beforehand.ResultsThe (resp. specificity) 64% 78%) Race, 61% 77%) CTS0 43% 80%) Breast. Further analyses showed among low population treated alone, rate was 0.3% (1 out 324).ConclusionsWe fully relapse. First, obtained results ability generalize data thus endorse soundness method. Furthermore, additional analysis brings light value Race it could be used therapeutic de-escalation purposes. will extended multi-site multi-scanner under completion.Legal entity responsible studyGustave Roussy.FundingOwkin.DisclosureV. Gaury, V. Aubert, C. Saillard, K. Elgui: Financial Interests, Personal, Stocks/Shares: Owkin. F. André: Advisory Role: All other authors have declared no conflicts interest. WSI. Correctly MethodsHES beforehand. ResultsThe 324). ConclusionsWe completion.