Topology-preserving STAPLE

Methodology for fusing multiple segmentations to produce an improved result has been useful in computational anatomical studies. Although obtaining segmentations of anatomy having a particular topology are essential to studies using diffeomorphic deformation based analyses, no methods of label fusion presented to date have incorporated information regarding the topology of the anatomy. In this paper, we introduce “Topology STAPLE”, a novel method that statistically fuses multiple rater segmentations into a topologically correct segmentation. We evaluate the method on both simulated data and real delineations of the cerebellum produced by human raters.