Whole Slide Images are 2D Point Clouds: Context-Aware Survival Prediction using Patch-based Graph Convolutional Networks

Cancer prognostication is a challenging task in computational pathology that\nrequires context-aware representations of histology features to adequately\ninfer patient survival. Despite the advancements made in weakly-supervised deep\nlearning, many approaches are not context-aware and are unable to model\nimportant morphological feature interactions between cell identities and tissue\ntypes that are prognostic for patient survival. In this work, we present\nPatch-GCN, a context-aware, spatially-resolved patch-based graph convolutional\nnetwork that hierarchically aggregates instance-level histology features to\nmodel local- and global-level topological structures in the tumor\nmicroenvironment. We validate Patch-GCN with 4,370 gigapixel WSIs across five\ndifferent cancer types from the Cancer Genome Atlas (TCGA), and demonstrate\nthat Patch-GCN outperforms all prior weakly-supervised approaches by\n3.58-9.46%. Our code and corresponding models are publicly available at\nhttps://github.com/mahmoodlab/Patch-GCN.\n