Feature-based vs. Intensity-based Neuroimage Registration: Comprehensive Comparison Using Mutual Information

Abstract

We propose a mutual information-based method for quantitative evaluation of the deformable registration algorithms at three levels: global, voxel-wise and anatomical structure. We compare two fully deformable registration algorithms: feature-based HAMMER and a set of intensity-based algorithms (FEM-Demons) in the ITK package. Evaluation is carried out using the AAE template image with 116 labeled anatomical structures and a set of 59 MR brain images: 20 normal controls (CTE), 20 Alzheimer's disease patients (AD) and 19 mild cognitive impairment patients (MCI). We show that both HAMMER and FEM-Demons perform significantly better than an affine registration algorithm, FLIRT, at all three levels. At the global level, FEM-Demons outperforms HAMMER on the images of AD and MCI patients. At the local and anatomical levels, FEM-Demons and HAMMER dominate each other on different brain regions.