Abstract 5447: Artificial intelligence (AI)-based classification of stromal subtypes reveals associations between stromal composition and prognosis in NSCLC

Background: Many cancers are known to involve a desmoplastic stromal reaction. While cancer-associated stroma (CAS) has long been appreciated histologically, single-cell molecular analyses have revealed its heterogeneity, and the cellular composition of CAS has been linked to prognosis in several cancer types, including NSCLC. While pathologists have sought to manually classify CAS based on collagen architecture or nuclear density, this approach has not sufficiently captured the heterogeneity of CAS. To this end, we have developed an artificial intelligence (AI)-based model to predict stromal composition in NSCLC from hematoxylin and eosin (H&E)-stained tissue Methods: We developed a convolutional neural network-based model to classify CAS as immature, mature, densely inflamed, densely fibroblastic, or elastosis. This model was trained using manual pathologist-derived annotations (N=3019) of H&E-stained whole slide images (WSIs) of PDAC obtained from the TCGA (N=126). This stromal subdivision model was deployed on H&E-stained LUAD (N=468) and LUSC (N=430) WSIs. Model performance was assessed by qualitative review by expert pathologists. Human interpretable features (HIFs) were extracted from the stromal subdivision model (e.g., proportional area of mature relative to total stroma) and were assessed to identify associations with overall survival (OS) using univariate Cox regression analysis after adjusting for age, sex, and tumor stage. Results: The stromal subdivision model successfully predicted areas of immature, mature, densely inflammatory, and densely fibrotic stroma, as well as elastosis, in LUAD and LUSC. In LUAD, higher combined proportional areas of mature and fibroblastic stroma relative to total cancer stroma was associated with poor OS (p=0.007), while higher combined proportional areas of densely inflamed stroma and elastosis relative to total cancer stroma was associated with improved OS (p=0.007). These findings were validated by stratified tertile analysis based on the corresponding risk direction. Notably, while the average stromal compositions did not differ significantly between NSCLC subtypes, the stromal HIFs were only prognostic in LUAD but not in LUSC. Conclusions: We developed a first of its kind model to characterize CAS subtypes in NSCLC tissue. Features extracted from this model are related to prognosis in LUAD, but not in LUSC, further confirming the importance of CAS to tumor biology and the importance of considering histologic subtypes. Work is ongoing to identify relationships between stromal HIFs and treatment response in NSCLC as well as other cancer indications. Citation Format: Fedaa Najdawi, Sandhya Srinivasan, Neel Patel, Michael G. Drage, Christian Kirkup, Chintan Parmar, Jacqueline Brosnan-Cashman, Michael Montalto, Andrew H. Beck, Archit Khosla, Ilan Wapinski, Ben Glass, Murray Resnick, Matthew Bronnimann. Artificial intelligence (AI)-based classification of stromal subtypes reveals associations between stromal composition and prognosis in NSCLC. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5447.