Representation Learning of Histopathology Images using Graph Neural Networks

Representation learning for Whole Slide Images (WSIs) is pivotal in\ndeveloping image-based systems to achieve higher precision in diagnostic\npathology. We propose a two-stage framework for WSI representation learning. We\nsample relevant patches using a color-based method and use graph neural\nnetworks to learn relations among sampled patches to aggregate the image\ninformation into a single vector representation. We introduce attention via\ngraph pooling to automatically infer patches with higher relevance. We\ndemonstrate the performance of our approach for discriminating two sub-types of\nlung cancers, Lung Adenocarcinoma (LUAD) & Lung Squamous Cell Carcinoma (LUSC).\nWe collected 1,026 lung cancer WSIs with the 40$\\times$ magnification from The\nCancer Genome Atlas (TCGA) dataset, the largest public repository of\nhistopathology images and achieved state-of-the-art accuracy of 88.8% and AUC\nof 0.89 on lung cancer sub-type classification by extracting features from a\npre-trained DenseNet\n