In medical imaging processing research, correct automatic labelling of each
voxel in a 3-D brain image remains an unsolved problem. Towards this goal,
we have developed a program called 
(Automatic Nonlinear Image
Matching and Anatomical Labelling) [2] which, like many other
non-linear registration procedures, is based on the assumption that
different brains are topologically equivalent and that non-linear
deformations can be estimated and applied to one data set in order to bring
it into correspondence with another. These non-linear deformations are
then used (i) to estimate non-linear morphometric variability in a given
population [3], (ii) to automatically segment MRI data
[2], or (iii) to remove residual alignment errors when
spatially averaging results among individuals.
Our previous validation of
showed that it worked well for deep
structures [2], but sometimes had difficulty aligning sulci
and gyri. In a follow-up paper presented at VBC96 [1], we
described an extension of the basic non-linear registration method that
used additional image-based features to help align the cortex. Simulations
showed that Lvv-based features and blurred sulcal traces significantly
improved cortical registration, however these results were not confirmed
with real MRI data. In this paper, we demonstrate the use of automatically
extracted and labelled sulci as extra features in conjunction with a
chamfer distance objective function [4,5].
Experimental results presented here (for both simulated and real data) show
significant improvement over our previous work.
In contrast to existing methods for cortical structure alignment that depend on manual intervention to identify corresponding points [6,7] or curves [8,9,10,11], the procedure presented here is completely automatic. While other 3D voxel-based non-linear registration procedures exist (e.g., [12,13,14,15,16,17,18,19,20]), none of these have looked specifically at the question of cortical features for cortical registration. The method presented here explicitly uses sulcal information in a manner similar to surface-matching algorithms described in [10,21,22] to improve cortical registration. However, our method is entirely voxel-based, completely automatic, and does not require any manual intervention for sulcal identification or labelling.
The paper is organized as follows: Section 2 describes the brain-phantom used to create simulated data, the acquisition parameters for real MRI data, the non-linear registration method, and the features used in the matching process; Section 3 presents experiments on simulated MRI data with known deformations in order to compare our standard linear and non-linear registration procedures with the new technique and to determine a lower-bound for the registration error. Section 3 with experiments on 11 real MRI data sets, comparing the old and new registration methods. The paper concludes with a discussion and presents directions for further research.