Computational features of TIL architecture are differentially prognostic of uterine cancer between African and Caucasian American women.

Abstract

5585 Background: Although the vast majority of endometrial cancer (EC) is early-stage and thus, curable by surgery, chemotherapy, and radiotherapy (with at least 85% 5-year OS), a fraction of them are aggressive neoplasms such as high-grade or deeply invasive lesions and thus exhibit poor prognosis. African American (AA) women are disproportionately affected by high-grade EC and have 80% higher mortality rate compared with Caucasian American (CA) women. In this work, we evaluated the prognostic ability of computational measurements of architecture of tumor-infiltrating lymphocytes (ArcTIL) from H&E slide images for EC. We also investigated the presence of morphologic differences in terms of ArcTIL features between AA and CA women and whether ArcTIL based population-specific models were more prognostic of OS in AA women compared to a population-agnostic model. Methods: The study included digitized H&E tissue slides from 445 post-surgery EC patients from TCGA, with further chemotherapy, or radiotherapy, including only the AA and CA patients, patients without reported race or from other populations were excluded. The dataset was divided into discovery (D1, n = 300), and a validation set (D2, n = 145), while ensuring population balance between two splits (D1(AA) = 65, D1(CA) = 235, D2(AA) = 37, D2(CA) = 108). A machine learning approach was employed to identify tumor regions, and tumor-associated stroma on the diagnostic slides and then used to automatically identify TILs within these compartments. Graph network theory based computational algorithms were used to capture 85 quantitative descriptors of the architectural patterns of intratumoral and stromal TILs. A multivariable Cox regression model (MCRM) was used to create population specific-prognostic models (MAA, MCA) and a population-agnostic model (MAA+CA)) to predict OS. All 3 models were evaluated on D2(AA), D2(CA), and D2. Results: MAA identified 4 prognostic features relating to interaction of TIL clusters with cancer nuclei in stromal compartment and was prognostic of OS on D2(AA) (see Table) but not prognostic in D2(CA) nor D2(AA+CA). MCA and MAA+CA identified respectively 7 and 6 prognostic features relating to interaction of TIL clusters with cancer nuclei (both in the epithelial and stromal regions) and were prognostic of OS on D2(CA) and D2, but not prognostic in D2(AA). Conclusions: Our findings suggest an important role of stromal TIL architecture in prognosticating OS in AA women with EC, while epithelial TIL features were more prognostic in CA women. These findings need to be validated in larger, multi-site validation sets.[Table: see text]