| We present a general framework of registration, review several variations on the Iterative Closest Point algorithm, discuss ground truthing, review data interpolation techniques, and compare geometry/feature-based and intensity-based registration.; We show that for a pair of multi-modal brain images, the average Euclidean distance and the normalized Mutual Information (MI) are highly correlated with the ground truth distance, but are not perfectly correlated. We introduce general model construction and object feature extraction procedures based on volumetric triangulation, and validate four new methods of combining intensity-based and geometry-based registration. A comparative study of 21 algorithms on a pair of multi-modal brain images shows that registration through the optimization of a convex function combining the Mutual Information and the average Euclidean distance is the most accurate. We also show that the accuracy of registration can be improved while reducing the number of voxels if isotropic resealing and sub-sampling is performed as a preprocessing step. The registration of 106 pairs of multi-modal brain images from 18 patients confirms that the convex function optimization is more accurate than the methods it was compared with.; The study of initial and final registration metrics for multi-modal brain images, performed through registration experiments with many starting points in the rotational and translational solution spaces, shows that the convex function combining MI and the average Euclidean distance has characteristics closer to those of the ground truth distance than either the average Euclidean distance, the MI, or the normalized MI. The accuracy of the ICP algorithm with an M-Estimator improves with isotropic resealing and sub-sampling. The plots of local convergence points indicate that the convex function combining MI and the average Euclidean distance has fewer local convergence points than the average Euclidean distance, the MI, and the normalized MI. The configuration of local convergence points also shows that the convex function has desirable characteristics in both the translational and the rotational solution spaces.; We further illustrate the generality, robustness, and adaptability of our approach through the registration of cell-nuclei images labeled with a DNA stain and a pair of images of dye-injected pyramidal neurons.; We conclude that combining intensity and geometry features in registration can improve the accuracy and robustness of the process without any loss of generality. The selection of the weights in the convex function combining the average Euclidean distance and the Mutual Information proved crucial in favoring the most reliable optimization function during registration, and in reducing the number of local extrema in the solution space of registration. |
