BAYESIAN SEGMENTATION OF MULTISLICE BRAIN MAGNETIC-RESONANCE-IMAGING USING 3-DIMENSIONAL GIBBSIAN PRIORS

We propose a maximum a posteriori probability (MAP) method, which feat ures 3-D Gibbsian priors and the highest confidence first (HCF) optimi zation method, for segmentation of cross-sectional images through a 3- D volume. The proposed algorithm has been successfully applied to segm entation of clinical magnetic resonance imaging (MRI) data on the huma n brain. Modeling the a priori probability of the segmentation by a 3- D Gibbs random field (GRF) imposes connectivity and smoothness constra ints on the desired segmentation in ail three directions. HCF is a rec ently proposed optimization method, which proves superior to other exi sting methods in this application. We discuss several implementation i ssues, including the effects of varying parameter values on algorithm performance. Experimental results with both phantoms and clinical MR d ata show that our proposed approach improves on existing methods in de lineation of the cerebral cortex, deep nuclei (such as the caudate and hippocampi), and the cerebral white matter. (C) 1996 Society of Photo -Optical instrumentation Engineers.