Automatic segmentation of pediatric brain MRIs using a maximum probability pediatric atlas

Abstract

Automatic anatomical segmentation of pediatric brain MR data sets can be pursued with the use of registration algorithms when segmentation priors (atlases) are in hand. We investigated the performance of a maximum probability pediatric atlas (MPPA), template based registration and label propagation. The MPPA was created from the 33 pediatric data sets, available through www.brain-development.org. We evaluated the performance of the MPPA comparing with manual segmentations by means of the Dice overlap coefficient. Dice values, averaged across representative regions, were 0.90 ± 0.03 for the hippocampus, 0.92 ± 0.01 for the caudate nucleus and 0.92 ± 0.02 for the pre-central gyrus. Segmentations of 36 further unsegmented target 3T images (1-year-olds and 2-year-olds) yielded visibly high-quality results. This registration approach allows the rapid construction of automatically labeled pediatric brain atlases in a single registration step.