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

[Asad A. Abu-Tarif]

>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.

This is partly correct and partly not very accurate.
As for the reasonably close match to begin with, it depends on what
optimizer
you use. For example, Powell's algorithm shows better initial registration
requirements than Gradient Descent.
One of my PhD objectives is to develop registration evaluation procedures.
One
of them is determining what we call "Initial Registration Kernel". This is a
plot
of the registration accuracy in the registration parameter space.

As for inter-subject registration, MI in its standard format can be
onverwhelmed.
However, some researchers used local registration using MI to achieve
elastic
registration (which fits intersubject registration).
Others defined modifications on the standard MI registration and called it
higher order MI. They showed that it applies for intersubject registration.


>Note also that MI would regard intensities 100 and 101 as completely
>unrelated, so smoothing and measurement noise will affect MI
>adversely.

Again, this applies to MI in its standard format. However, if you register
(using MI)
two segmented (or quasi-segmented) images you can overcome this problem.
This requires no change on the implementation of the registration algorithm,
it
only requires that the two images be segmented (not necessary very accurate
segmentation) before being passed to the registration algorithm.
Another work I've seen in the literature is to modify the MI to register 3
data sets,
the two raw images and a third segmented image (from the two). What they've
done
is to run some Gray-Scale Connected Component Analysis on the two images
to generate a third one and then register the three using the modified MI.
Part of my PhD work is to oversome this. What I'm doing is to perform
multi-scale
registration both spatially and in density (gray-scale values). I'm still
working on
this part and don't have results yet!!

Regards,
Asad

================================================
Asad Abu-Tarif
Research Assistant, NYS Center for Automation  Technologies
PhD Candidate, Computer Engineering, Rensselaer Polytechnic Institute
tarifa@rpi.edu
================================================