A NOVEL BAYESIAN GRAPHICAL MODEL AND PERFECT MONTE CARLO EM ALGORITHM FOR AUTOMATED COLOCALIZATION ESTIMATION IN MULTICHANNEL FLUORESCENCE MICROSCOPY

In digital fluorescence microscopy, colocalization estimate between two biological entities within a specimen is often based on subjective visual inspection of images or ad hoc sequence of algorithms with several manually-tuned parameters, leading to irreproducible and unreliable estimates. We propose a novel Bayesian Markov random field (MRF) model for colocalization estimation from dual-channel images, encoding colocalization as a model parameter, to solve a unified data-driven optimization problem that, unlike existing methods, automatically deals with common-background removal, object labeling, parameter tuning, and noise. For model fitting, we propose Monte Carlo expectation maximization (EM) with perfect sampling extended from priors to posteriors, for our MRF model, to guarantee sampler convergence. We use consistent pseudo-likelihood estimators to deal with intractability in MRF parameter estimation. Results on simulated, benchmark, and real-world data show that our method estimates colocalization more reliably than the state of the art.