Fig. 3.

Overview. This diagram provides an overview of the study for a single nonlinear registration algorithm, placing example preprocessed data from Fig. 1 into the equations of Fig. 2. The three stages include linear registration, nonlinear registration, and evaluation (left to right). The four different datasets (LBPA40, IBSR18, CUMC12, and MGH10) are aligned along the left in four different versions: images, surfaces derived from the images, labels, and borders derived from the labels. A source and target are drawn from each version (image volumes are shown as coronal slices for clarity). A source image I_s is linearly then nonlinearly registered to a target image I_t. The linear and nonlinear transforms (X_s→t and X_[s→t]⇝t) are applied to the corresponding source labels L_s. The resulting nonlinearly transformed labels L_[s→t]⇝t are compared against the target labels L_t. This comparison is used to calculate volume overlap and volume similarity per region. The target surface S_t is intersected with the target labels L_t and warped source labels L_[s→t]⇝t to calculate surface overlap. Borders between each labeled region and all adjacent labeled regions are constructed from L_t and L_[s→t]⇝t, and average distances between the resulting borders B_t and B_[s→t]⇝t are calculated per region.