If I want to play with amount of regularization (try to look at the registration in ventricle as mentioned in the early emails), should I try different lambdas for each resolution, or only plan with the last lambda (the highest resolution)?
it is a little hard to say without knowing what your images looked like prior to registration. I think that for "normal" subjects I would expect the ventricles to line up a little better than that.
As for the lambdas it is trial and error really. You could e.g. try to split the last step (the one with ten iterations) into two steps of five iterations each, and reduce lambda by a factor ~2 for the last step.
Good Luck Jesper