Combining Markov random field and marked point process for microscopy image modeling
In many microscopy image analysis applications, it is of critical importance to address both pixel-level segmentation and object-level feature extraction. While Markov random field (MRF) models are useful for imposing local constraints at the pixel level, they have limited capability for imposing global constraints. Marked Point Process (MPP) models incorporate global information, such as shape, as a prior, but local constraints, such as pixel-wise interaction, are not easily modeled. To address the problem, we first propose a hybrid MRF/MPP model to incorporate both local and global constraints within one single energy function for image analysis. Optimization using this model is performed using simulation schemes, including reversible jump Markov chain Monte Carlo (RJ MCMC) and multiple birth and death algorithms. Secondly, we propose a two-pass multiple birth and death algorithm. In the death step of the original multiple birth and death algorithm, objects which are killed in later stages might affect the accuracy of the death rate of objects processed earlier, especially for adjacent object pairs, where two objects have close interaction. In our algorithm, we add a rebirth step after the death step to solve this problem. Finally, we propose a joint MRF/MPP model. Unlike the hybrid model, where the MRF is interpreted as an energy term within the MPP framework, this model combines the MRF and the MPP into a joint probability distribution. We show experiments to demonstrate the comparison of this model and the hybrid model.
Comer, Purdue University.
Off-Campus Purdue Users:
To access this dissertation, please log in to our