A. Abu-Tarif and G. Nagy (USA)
Multi-Modal Registration, Medical Image Processing.
We observed that in 3D images of the brain, Mutual
Information (MI), the average Euclidean distance, and
the ground-truth distance for 3D registration are highly
correlated, but neither the average Euclidean distance
nor the MI is perfectly correlated with the ground truth.
This suggests that combining feature-based and MI
based methods may be better than registration based on
either separately. We therefore compared 21 variations
on models, criterion functions, and optimization
algorithms on images from a single patient, and then
applied the best of the resulting methods to CT-MR
image pairs from 16 patients for which we did not have
the ground truth. The results were evaluated
independently. Our median ground-truth distance of
1.42mm over all patients ranked best among those
submitted by eight other groups.