Spatial-dependent similarity metric supporting multi-atlas MRI segmentation

Abstract

A locally weighted normalized mutual information (wNMI) metric is proposed to highlight spatial correspondences in an image pair. Aiming to account for local similarities, our proposal computes the normalized mutual information on local regions and linearly weights them depending on their information. Additionally, we introduce a criterion for tuning the number of regions based on the variability maximization of the metric values in a given dataset. To assess our proposal, we compare wNMI to means squares (MS), cross correlation (CC) and global normalized mutual information (NMI) to select the closest atlases on a multi-atlas segmentation scheme for labeling ganglia basal structures. Obtained results show that our proposed measure provides a more robust atlas selection.