A Comparative Study on Graph Construction Methods for Survival Prediction using Histopathology Images

Survival analysis from whole slide histopathology images (WSIs) is challenging due to the high dimensionality of WSIs. Graph-based approaches thus have been extensively studied to analyze WSIs. Recent success of graph neural networks further promotes graph-based approaches for survival analysis. However, the performance of graph-based methods is heavily dependent on graph construction methods. In this paper, we compare and analyze different graph construction methods for survival prediction from WSIs on the TCGA bladder and brain cancer pathology image datasets