An Expectation Maximization-Like Algorithm for Multi-Atlas Multi-Label Segmentation

We present in this paper a novel interpretation of the concept of an "expert" in image segmentation as the pairing of an atlas image and a non-rigid registration algorithm. We introduce an extension to a recently presented expectation maximization (EM) algorithm for ground truth recovery, which allows us to integrate the segmentations obtained from multiple experts (i.e., from multiple atlases and/or using multiple image registration algorithms) and combine them into a final segmen- tation. In a validation study with randomly deformed segmentations we demonstrate the superiority of our method over simple label voting.