Locally weighted Markov random fields for cortical segmentation

Abstract

Segmenting the human brain from magnetic resonance images is a challenging task due to the convoluted shape of the cortex, noise, intensity non-uniformity and partial volume effects. We propose a new way to overcome part of the bias-variance tradeoff existent in any segmentation technique by locally varying the behaviour of the model. We developed a novel metric based on the Laplacian of the geodesic distance to localise and iteratively modify the prior information and Markov random field weights, leading to a better delineation of deep sulci and narrow gyri. Experiments performed on 20 Brainweb datasets show statistically significant improvements in Dice scores and partial volume estimation when compared to two well established techniques.