Quantitative profiling of microglia populations using harmonic co-clustering of arbor morphology measurements

Microglia are the resident immune cell population in the mammalian central nervous system (CNS). These highly plastic cells exhibit ramified arbors in their resting state, and progressively less-complex arbors when activated. Our goal is to compare the spatial distributions of resting and activated microglia in normal brain tissue against tissue that is perturbed by insertion of a neural recording device. For this, microglia were imaged using multiplex immunostaining and confocal microscopy. The cell arbors were traced automatically, and 127 quantitative measurements based on the L-measure [8] were computed for each cell. A hierarchical extension of Coifman's [1,2] unsupervised harmonic analysis method was used to profile these multivariate data and identify groups of similar cells and the underlying features. This iterative procedure induces an orthogonal basis by constructing a coupled geometry over the row and column spaces of the feature matrix. Smoothing of the dataset, and the row and column clusters is achieved simultaneously when the algorithm converges. Experiments on real image datasets demonstrate the ability of this method to generate qualitative and quantitative groups that are biologically meaningful despite the existence of noise and missing values.