Multi-variate Mutual Information for Registration

An extension of the mutual information metric to a three-variate cost function for driving the registration of a volume to pair of co-registered volumes is presented. While mutual information has typically been applied to pairs of variables, it is possible to compute multi-variate mutual information. The implementation of multi-variate mutual information is described. This metric is demonstrated using the problem of registering a deformed t2 slice of the visible male magnetic resonance data set to either a single t1 slice or a pair of co-registered t1 and proton density slices. Two-variable and three-variable metric registration results are compared. Adding the extra proton density information to the registration cost metric leads to faster optimization convergence and better final accuracy. Multi-variate mutual information has potential application in problems where the addition of more information can lead to solution convergence or improve accuracy.