Characterizing the Histology Spatial Intersections Between Tumor-Infiltrating Lymphocytes and Tumors for Survival Prediction of Cancers Via Graph Contrastive Learning

Tumor-infiltrating lymphocytes (TILs) and their spatial interactions with tumors on whole-slide images (WSIs) of histopathology sections can provide valuable information about the immune response within the tumor micro-environment that is closely associated with the progression of human cancers. To effectively exploit the interactions between TILs and Tumors from WSIs, spatially informed analysis tools are required. Here, we present GCTIL, a simple but effective graph contrastive learning framework to learn meaningful representations for the TILs and tumor nodes extracted from the WSIs. Specifically, GCTIL considers the graph permutation of different strength to help learn robust node representations that can not only capture the structural characteristics of the graph but also preserve the correct distance orders among different permutations. Moreover, GCTIL also imposes distance constraints to distinguish the node embeddings of different types (i.e., TILs and Tumor). Then, based on the patch representation derived from GCTILs, we apply the graph attention networks (GATs) to describe the spatial interactions between TILs and tumor regions in WSIs for survival analysis of human cancers. We evaluate the performance of our method on the Breast Invasive Carcinoma (i.e., BRCA) cohort derived from The Cancer Genome Atlas (TCGA), and the experimental results indicate that our method is superior to the comparing methods.