Biologically-driven cell-graphs for breast tissue grading

We present biologically-driven cell-graphs for the computer-aided grading of cancer in 3D breast tissue slides. Cell-graphs were previously proposed to capture structure-function relationships within tissues by representing the tissues with undirected and unweighted graphs wherein the cell nuclei constitute the graph nodes and the approximate adjacencies of the nuclei are represented with edges. Using advanced immunohistochemistry staining, levels of interactions between the cells can be captured via the expressions of particular proteins that yield more biological attributes for constructing the graph edges. With this motivation, we fluorescently labeled proteins that participate in cell-to-cell communications and analyze the structural organization of the cells in three grades of breast cancer using biologically-driven cell-graphs. Using the features extracted from the graphs in a supervised machine learning setting, computer-aided grading is achieved. Biologically-driven cell-graphs achieve 98.5% accuracy in grading the cancer into non-malignant, non-invasive, and invasive grades that significantly outperforms the conventional proximity-based cell-graphs and a Delaunay triangulation based method by 9.0% and 13.5%, respectively.