Tissue segmentation and classification using graph-based unsupervised clustering

Automated segmentation and quantification of cellular and subcellular components in multiplexed images has allowed for a combination of both spatial and protein expression information to become available for analysis. However, performing analyses across multiple patients and tissue types continues to be a challenge, as well as the greater challenge of tissue classification itself. We propose a model of tissues as interconnected networks of epithelial cells whose connectivity is determined by their size, specific expression levels, and proximity to other cells. These Biomarker Enhanced Tissue Networks (BETN) reflect both the individual nature of the cells and the complex cell to cell relationships within the tissue. Performing a simple analysis of such tissue networks managed to successfully classify epithelial cells from stromal cells across multiple patients and tissue types. Further experiments show that significant information about the structure and nature of tissues can also be extracted through analysis of the networks, which will hopefully move towards the eventual goal of true tissue classification.