Investigating Spatial Patterns of Tumor and Stroma in Gastric and Colorectal Cancer for Survival Prediction

Abstract

The spatial patterns of tumor, stroma, and tumor-infiltrating lymphocytes (TILs) within the tumor microenvironment significantly affect cancer progression and are associated with clinical outcomes. Understanding the exact significance and statistical implications of these spatial patterns is, however, challenging due to the complexity of spatial interactions. In this paper, we investigate the spatial patterns related to patient survival in gastric and colorectal cancer using four classifiers to predict tumor, stroma, TILs, and Microsatellite Instability (MSI) status, along with the Getic-Ord-Gi* statistic as the spatial image analysis, analyzing four large patient cohorts. U-Net, a deep learning model for semantic image segmentation, was used to predict tumor, stroma, and TILs in digitized Hematoxylin and Eosin-stained formalin-fixed paraffin-embedded (FFPE) sections, and ResNet-18 was employed to predict the MSI status. The Getis-Ord-Gi* statistic was applied to determine statistically significant tumor hotspot regions by examining their relationship to surrounding areas. Kaplan-Meier analyses and log-rank tests were performed to assess the association between the spatial patterns of tumor and stroma and overall survival in the patient cohorts. Results indicate that the stroma composition around tumor hotspot regions, identified by the Getis-Ord-Gi* statistic, shows significant differences in overall survival among patients with gastric and colorectal cancer. The log-rank test was included to examine the relationship between stroma composition and MSI and ACTA2 expression levels.