You can use 3dAllineate's '-weight' option. You have to make a mask dataset that is set to 1 for the voxels that you like, and to 0 for the voxels that you don't like. There are many ways to create mask datasets in the AFNI package. 3dcalc is probably the simplest one for your application.

Note that the histogram-based similarity metrics are dependent on the number of grid intervals used to build the joint histogram, so the values computed in this program probably won't agree with the values from other software. And also note that some of the cost functionals are inverted in 3dAllineate, since the program tries to minimize these values -- for example, the MI cost functional reported is actually minus the mutual information, since mutual information is supposed to be largest at the the bestest alignment. See the output of the latest version of 3dAllineate -HELP for all the details.