Abstract 7399: Development of a reproducible AI-based spatial biomarker of the tumor immune infiltrate on H&E slides

Abstract

Abstract Background: The immune infiltrate in the tumor microenvironment (TME) has been shown to be associated with prognosis in several studies. However, a comprehensive assessment of the TME composition is not performed in routine practice due to the lack of standardized methods and tools for quantification. The aim of our study is to develop a reproducible, artificial intelligence (AI)-based, spatial biomarker of the tumor related immune response to predict patients outcome on routine Hematoxylin & Eosin (H&E) slides. To the best of our knowledge, this is the first study that implements a robust H&E pipeline to assess, across multiple cohorts, the prognostic power and the added-value of a spatial marker in comparison to non-spatial measures such as the overall lymphocytes density on the slide. Methods: We trained AI models on both publicly available (NuCLS and Lizard) and proprietary datasets to automatically detect cells and tissue types on H&E stained Whole Slide Images of cancers from different origins. We used these models on TCGA COAD and TCGA BRCA and computed spatial measures related to the lymphocyte densities in the tumor core versus the invasive margin as well as the immune infiltration in the tumor stroma (in between tumor islets) versus in the tumor islets themselves. Univariate and multivariate survival analysis is performed through repeated cross-validation. A log-rank test on the stratified groups is performed for each evaluated marker, using the median value as cutoff. Preliminary results on TCGA-COAD and TCGA-BRCA show that spatial markers of the tumor-related immune response are individually associated with overall survival, reinforcing findings from prior studies emphasizing the importance of the spatial relationship between tumor and immune cells. Multivariate analysis with both linear and non-linear models is currently ongoing across selected cohorts. Citation Format: Valentina Di Proietto, Jean El-Khoury, Benjamin Adjadj, Lucie Gillet, Ulysse Marteau, Kathryn Schutte, Katharina Von Loga. Development of a reproducible AI-based spatial biomarker of the tumor immune infiltrate on H&E slides [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 7399.