Features:

The objective function maximized in the optimization procedure measures the similarity between features extracted from the source and target volumes. We use blurred image intensity and blurred image gradient magnitude, calculated by convolution of the original data with zeroth and first order 3D isotropic Gaussian derivatives in order to maintain linearity, shift-invariance and rotational-invariance in the detection of features. Since both linear and non-linear registration procedures are computed in a multi-scale fashion, the original MR data was blurred at two different scales: =8, and 4mm, with the $=(2.35\sigma)$ of the Gaussian kernel acting as a measure of spatial scale.

The enhanced version of   uses geometrical landmarks such as sulci to improve cortical alignment. We have previously detailed a method called   (Sulcal Extraction and Automatic Labeling) to automatically extract cortical sulci (defined by the 2D surface extending from the sulcal trace at the cortical surface to the depth of the sulcal fundus) from 3D MRI [33,31]. Each sulcus is extracted in a two step process using an active model similar to a 2D Snake that we call an active ribbon. First, the superficial cortical trace of the sulcus is modeled by a 1D snake-spline [34], that is initialized to loci of negative curvature of the cortical surface. Second, this snake is then submitted to a set of forces derived from (1) tissue classification results, (2) differential characteristics of the image intensities near the cortex and (3) a distance-from-cortex term to force the snake to converge to the sulcal fundus. The set of successive iterative positions of the 1D snake define the sulcal axis. These loci are then used to drive a 2D active ribbon which forms the final representation of the sulcus.

When applied to an MRI brain volume, the output of this automatic process consists of a set of Elementary Sulcal Surfaces (ESS) that represent all cortical sulcal folds. A number of these ESSs may be needed to represent a sulcus as defined by an anatomist. Simply blurring the voxel masks that represent the sulci at scales of =16 and 8mm  so that they could be directly incorporated into the standard   procedure for correlative matching was found not to improve cortical registration for real MRI data [1]. Here, in order to incorporate the geometric sulcal information represented by the ESS, the sulcal ribbons are voxelated. Like Sandor et. al.[10], we use a chamfer distance function to drive sulci together. This methods brings sulci that are initially far apart into registration.