Pathological analysis of tumor microenvironmental features using deep learning to predict survival of colon carcinomas.

Abstract

3527 Background: Advances in deep learning may improve the ability to evaluate and quantify pathological features in solid tumors to improve prediction of patient disease-free survival (DFS). We used an artificial intelligence (AI) algorithm to quantify morphological features in the tumor microenvironment of stage III colon cancers. Methods: We analyzed digitized stage III colon carcinomas (N = 402; 382 met QC) from participants in a phase III trial of FOLFOX-based adjuvant chemotherapy that included all available tumors with dMMR and a randomly selected cohort of pMMR tumors (median follow-up 60 months) [NCCTG N0147; Alliance for Clinical Trials in Oncology]. Fifteen morphological features were extracted using the QuantCRC algorithm that segments images into carcinoma (low-grade, high-grade, signet ring cell), stroma (immature, mature, and inflammatory), mucin, tumor/stroma ratio, tumor budding/poorly differentiated clusters (TB/PDC), necrosis, smooth muscle, fat, and tumor-infiltrating lymphocytes (TILs) and tumor area (mucin, epithelium and TB/PDC). Analysis of AI-derived features with clinical variables, molecular alterations ( BRAF/KRAS), and DFS were examined using Kaplan-Meier methodology and multivariable Cox regression. Results: Among the 15 AI-derived morphological features, the following differed significantly between dMMR (n = 191) and pMMR (n = 189) whereby dMMR tumors had lower mature stroma, but higher inflammatory stroma and tumor, higher tumor grade and increased mucin, TB/PDCs, signet ring cells, and TILs (all p< 0.05). Among dMMR tumors, multivariable analysis revealed that tumor area [mucin, epithelium and TB/PDC] (HR adj 0.45, 95%CI (0.24, 0.84), p:0.01, [Q2/3 vs Q1]) and N stage were the only significant and independent prognostic variables. Among pMMR tumors, multivariable analysis identified that tumor budding/poorly differentiated clusters (TB/PDCs) (HR adj 0.20, 95%CI (0.06, 0.61), p:0.01, [Q1 vs Q4]) was the strongest prognostic variable and the only morphological feature that was significantly associated with DFS along with age, N stage and T stage. Conclusions: Using AI, we can extract and quantify distinct morphological features in tumor sections that differ between dMMR and pMMR and multivariately, can significantly and robustly enhance prognostication within each MMR group. Among pMMR tumors, tumor budding/poorly differentiated clusters was the strongest predictor of DFS. Support: NIH U10CA180821, U10CA180882, U24CA196171, R01 CA210509 (to FAS). Study NCCTG N0147 received funds from Sanofi https://acknowledgments.alliancefound.org . ClinicalTrials.gov Identifier: NCT00079274 Key words Artificial Intelligence; Tumor Microenvironment; Colonic Neoplasms; Disease-free survival.