[Insight-developers] Where mutual information works, and doesn't

[Paul Hughett]

I've had the time to think a little bit more about registration using
the mutual information metric since the telcon on Friday, and realized
that my experience and intuition about the way it works is consistent
with the following generalization:  MI works well for cross-modality
registration on a single subject, but not for cross-subject
registration; this is because it requires a reasonably close match to
begin with and inter-subject differences in geometry overwhelm it.
Note also that MI would regard intensities 100 and 101 as completely
unrelated, so smoothing and measurement noise will affect MI
adversely.  Conversely, the L2 metric (aka sum of squared differences)
often works well for cross-subject registration but not for
cross-modality registration, because it cannot cope with the
differences in the tissue-to-intensity mapping.

Does anyone else's experience support or deny this?

Paul Hughett