Using consensus measures for atlas construction

Abstract

Atlas-based segmentation has been shown to provide promising results to delineate critical structures for radiotherapy planning. However, it requires to have a reference image with its reference segmentation available. Classical methods used to build an average segmentation can lead to over-segmentation in case of high variability among the manual segmentations. We propose in this paper a consensus-based approach to construct a reference segmentation from a database of manually delineated images. We first compute local consensus measures to estimate a variability map, and then deduct from it a consensus segmentation. Finally, the proposed method is evaluated using a dataset of 64 manually delineated images of the head and neck region.