The impact of atlas formation methods on atlas-guided brain segmentation
Martha E. Shenton, William Wells, Kilian M. Pohl · 2007
Abstract. We analyze the impact of atlas construction within the context of an atlas-guided segmenter applied to a morphometry study in neuroanatomy. Auto-matic segmenters often rely on anatomical information encoded via probabilis-tic atlases. These atlases are frequently constructed by registering collections of training data. In this paper, we study the impact of registration methods as well as the training data on automatic segmentation results. With respect to registra-tion, we focus our comparison on pairwise vs. group-wise methods and fixed vs. online coordinate systems. For the training data, we consider collections of pop-ulation specific and general population data. To study the impact of these factors, we revisit a previously published statistical group comparison that was based on manual segmentations. For each atlas type, we record the group differences based on automatic segmentations and compare these findings to the original ones. Fur-thermore, we measure the Dice overlap between manual and automatic segmen-tations. Our results indicate some advantages for coordinate systems that are de-veloped in an online fashion. 1