Voxel based nonrigid image registration using local and partial volume similarity measures

Abstract

Recently, different approaches have emerged for voxel based nonrigid image registration using local instead of global similarity measures. Benefits are more accurate registration or the ability to subdivide the global similarity in local contributions. Within this article, we provide a general method to localise similarity measures using overlapping regions. Moreover, we extend the concept of partial volume estimation, introduced for mutual information (MI), to other similarity measures. We compare local and global sum of squared differences (SSD), cross correlation (CC) and MI for different sizes of the local regions. In general, local MI gives the highest accuracy, even for image pairs of the same modality. Partial volume estimation slightly improves the accuracy for local measures; the improvement is more pronounced for label images.