A0876
Title: Improved spectral unmixing of highly complex biological fluorescence images using a priori knowledge
Authors: Alex Valm - State University of New York at Albany (United States) [presenting]
Abstract: Many biological systems are composed of complex consortia of interacting components. Human dental plaque is known to comprise a community of over 700 different species of bacteria. Knowledge of these communities' spatial structure is critical to understanding disease processes in the oral cavity, including periodontal disease. While it is possible, in principle, to specifically label every species of bacteria in a community with DNA probes, the broad emission spectra of compatible fluorescent dyes prevent the routine use of more than a handful of fluorescent reporters in any single imaging experiment. A constrained unmixing algorithm that incorporates prior knowledge to greatly improve the accuracy in classifying microbial cells according to their taxonomy in images of fluorescently labelled microbial communities is presented. The Sparse and Low-Rank Poisson Regression Unmixing (SL-PRU) approach incorporates multi-penalty terms for rewarding sparseness and spatial correlation of the estimated abundances among spatially correlated pixels. SL-PRU further uses Poisson regression for unmixing instead of least squares regression to better deal with photon shot noise. A general method is proposed to tune the SL-PRU parameter weights and demonstrate improved pixel-wise abundance estimation and endmember classification in complex images of human dental plaque with quantitative morphological analyses.