Nonparametric Markov priors for tissue segmentation
Zhuang Song, Suyash P. Awate, James C. Gee · 2008
This paper presents a novel method to construct a probabilistic tissue prior, for Bayesian tissue segmentation, which is based on nonparametric Markov statistics of tissue intensities learned from training data. The proposed nonparametric Markov (NPM) prior is in contrast to the conventional tissue-probability-map (TPM) prior that is based on the voxel location in a common anatomical template space. Given a set of manually labeled voxels as the training set, the NPM prior is constructed by learning a fuzzy classification function that distinguishes the Markov statistics of tissue intensities in a statistical supervised-learning framework. The validation experiments in this paper compare the efficacy of the NPM prior to that of the TPM prior in producing tissue segmentations, and demonstrate the advantages of the NPM prior, qualitatively and quantitatively, over the TPM prior, especially in cortical regions.