The mirror method of assessing segmentation quality in atlas label propagation

Abstract

Atlas-based brain image segmentation quality is difficult to assess in the absence of reference target segmentations. We propose a measure of segmentation success based on transforming the atlas label twice: once by registering the atlas to the target and a second time by registering the target to the atlas. Each registration represents transformations by free-form deformations. The overlap between the twice-transformed label and the original (‘mirror overlap’) correlates with the forward overlap (between the once-transformed label and a target reference), especially for subcortical structures. Using mirror overlap as an atlas selection criterion results in improved segmentations versus random selection.