Simulations

**Figure 2:** Simulations: ANIMAL-only vs ANIMAL+sulci
$\begin{figure} \begin{center} \centerline{\epsfxsize=11cm \epsfbox{./paper-92-fi... ...e standard \begin{sc}animal\end{sc}~ registrations.} } \end{center}\end{figure}$

Standard ANIMAL: In the first experiment (Fig. 2-b) we can see that the global brain shape is properly corrected and that the main lobes are mostly aligned since the central and Sylvian sulci are better aligned than in the linear registration. Even though previous experiments in [2] have shown that basal ganglia structures (e.g., thalamus, caudate, putamen, globus pallidus) and ventricular structures are well registered (with overlaps on the order of 85% to 90%) for both simulated and real data, the simulations presented here indicate that the standard $\begin{sc}animal\end{sc}$ technique cannot always register cortical structures by using only blurred image intensities and gradient magnitude features.

ANIMAL+selected sulci: In Fig. 2-c, 10 extracted sulci (superior and middle frontal, Sylvian, olfactory and central, on both hemispheres) were used as additional features in $\begin{sc}animal\end{sc}$ to address problems in establishing correspondence. Visual inspection of Fig. 2-c shows that the previous misalignments in the motor-sensory area have been corrected, and alignment for neighbouring structures has improved greatly. Indeed, when evaluated on the central sulcus, the ${D_{\mathsf{surf}}}$ measure improves dramatically: 3.2mm for linear; 2.0mm for standard non-linear; 0.7mm for $\begin{sc}animal\end{sc}$ with labelled sulci. The standard deviation of ${D_{\mathsf{surf}}}$ decreases as well, indicating tighter grouping around the target sulci. The 0.54mm average ${D_{\mathsf{surf}}}$ reduction is highly significant (p<0.0001, T=157, d.o.f.=320, paired T-test over all sulci and all subjects).

ANIMAL+selected sulci+all sulci: In this experiment, the $\begin{sc}animal\end{sc}$ +selected sulci transformation is used as input where all sulci extracted by $\begin{sc}seal\end{sc}$ are used as features in an attempt to further improve cortical alignment in regions that are not close to the previously selected sulci. The improvement is evident qualitatively in Fig. 2-d and quantitatively in Table 1 where the average ${D_{\mathsf{surf}}}$ , evaluated over all sulci and all subjects drops from 1.6mm for the standard $\begin{sc}animal\end{sc}$ to 1.5mm when using all sulci. While this improvement is small, it is statistically significant (p<0.0001, T=88.5, d.o.f.=320, paired T-test).