Improving multi-atlas segmentation accuracy by leveraging local neighborhood information during label-fusion

Abstract

Multi-atlas segmentation techniques typically comprise generation of multiple candidate labels that are then combined at a final label fusion stage. Label fusion strategies usually leverage information contained in these training labels but ignore local neuroanatomical information. Here, we address this limitation by explicitly incorporating local information at the label fusion stage. The proposed method - Autocorrecting Walks over Localized Markov Random Fields (AWoL-MRF) - is initialized using a set of candidate labels from the atlas library to partition a specific structure into high and low confidence regions. The labels of the low confidence regions are updated based on a localized Markov random field model and a novel sequential inference process (walks), which mimics manual segmentation protocols. The approach combines a priori information from the atlas library with the local spatial constraints improving the accuracy and robustness of the existing segmentation methods.