Chapter 5 - Adaptive graph convolutional neural network and its biomedical applications

Abstract

Graph convolutional neural networks (GCN) are generalizations of classical CNNs to better work with graph-structured data that include biochemical molecular graph, 3D point cloud and social networks. Current convolutional kernels in GCNs were built upon fixed and shared graph structure. However, for most real-world data, graph structure varies in terms of both scale and topology. A generalizable convolutional operator on graph is supposed to be compatible with different graph topologies. In the article, authors introduced a generalized and flexible GCN framework along with a new spectral graph convolutional layer parameterizing distance metric and feature transform. Besides original graph structure, a residual graph is constructed and learned throughout training. Therefore, the introduced Adaptive graph convolutional network (AGCN) is adaptive to graphs of arbitrary topological structure and scale and is also adaptive to various learning tasks easily. With graph attention network, we enabled AGCN to learn graph representation from a bag of patches randomly sampled from large medical images such as Whole-slide-image for sophisticated understanding tasks, for example survival prediction. The DeepGraphSurv is an end-to-end framework that directly predict survival probability from patients’ WSI on tissues. Empowered by graph attention, an intuitive annotation of tumor cells is also learned and generated by the model.