Spectral analysis of cell-graphs for automated cancer diagnosis

The cell-graph approach captures the information encoded in tissue samples by capturing the spatial distribution of cells and their cluster formations. In a cellgraph, nodes and edges represent the cell clusters and pairwise relationships between them, respectively. It is shown in [1] that the features of cell-graphs of cancerous tissues are significantly different from those of healthy tissues and benign reactive/inflammatory processes. Thus, cell-graphs can be used for the purpose of automated cancer diagnosis [1]. In this paper, we present a new set of features and show how effectively they aid to automated cancer diagnosis. In particular, we investigate the spectral decomposition of adjacency matrices and normalized Laplacian of the cell-graphs. To the best of our knowledge, this is the first use of the spectral graph theory in tissue-based automated cancer diagnosis. Working with 646 brain biopsy samples of 60 different human patients, we demonstrate that the spectra of the cell-graphs of cancerous tissues are unique and distinguishable from those of the healthy and inflamed tissues with accuracy >90% (with a sensitivity of 91.57% and specificities of 93.75% and 98.15% for the inflamed and healthy tissues, respectively).