Ensemble-based medical image labelling via sampling morphological appearance manifolds

Atlas-based image labeling is a fundamental tool in medical image segmentation. Recent years have seen extension of single-atlas warping to multi-atlas warping and fusion, which has clearly demonstrated the advantage of consensus-based segmentation. Herein, we further extend this concept, by leveraging upon morphological appearance manifolds (MAMs), which have been previously used to represent a single anatomy with a multitude of equivalent representations derived by varying the atlas or the deformation parameters. We sample the MAM by using different atlases, deformation algorithms, and regularization parameters, thereby generating an ensemble of transformations and residuals. The former are used to generate tentative labels for the individual image, and the latter are used to locally rank/weight each member of the ensemble according to its similarity with the individual. The method obtained a high accuracy both on training and testing datasets of the MICCAI’13 mid-brain segmentation challenge (average Dice score d = 0.8686).