Applying spatial attention-based autoencoder learning of latent representation for unsupervised characterization of tumor microenvironmentDownload PDF

Published: 28 Apr 2023, Last Modified: 28 Apr 2023MIDL 2023 Short paper track PosterReaders: Everyone
Keywords: Autoencodeurs, spatial attention, computational pathology
TL;DR: Unsupervised discovery of tissue architecture in spatial tissue imaging
Abstract: Spatial tissue imaging technologies enable highly resolved spatial characterization of cellular phenotypes. Today this spatial mapping still largely depends on laborious manual annotation and molecular labels to understand tissue organization. As a result, we are not optimally leveraging higher-order patterns of cell organization potentially connected to disease pathology or clinical outcomes. To address this gap, we propose a novel approach how autoencoders with spatial attention mechanism can be trained to enrich cell phenotyping. Our approach combines information on cellular phenotypes with the physical proximity of cells to accurately identify organ-specific microanatomical in the tumor microenvironment. We apply our method to lung tumor tissues imaging mass cytometry data to show how it can detect higher-level cell organizations and information on structural differences.
2 Replies