[Insight-users] MI registration update

Dill, John john-dill at uiowa . edu
Mon, 1 Dec 2003 14:19:40 -0600


I have finished looking at the results of an experiment I did for observing
the effect of the sample size on the mi search algorithm with sample size of
50, 100, 250, 500, 750, 1000.  It was done over nine patients and here are
some things I have questions about.

Out of the 18 registrations (ap and lat), there were 7 that were not good.
About half of these I think I can account for by claiming data faults, where
a significant portion of the DRR was missing and hence was biased by
mis-information in the drr.  The others are not particularly clear why the
registrations failed, perhaps artifacts from the treatment table are the
cause.  All of my lateral registrations were good (except for the one which
had half the drr missing).  I think I can correct those bad registrations
with a proper window.

In the registrations that worked, the sample sizes of 250, 500, 750, and
1000 all converged to the same x and y translation within about +/-0.2
pixels, but what differed was the rotation component, which varied about
+/-3 degrees in the extreme cases.  I found that out of those 4 sample
sizes, the registration I preferred (not always though) typically had a
rotation component closer to 0.  I was wondering if anyone had suggestions
for trying to tighten the uncertainty in the rotation, perhaps increasing
the rotation scale parameter?  I saw a tightening of the rotation
variability when I increases the sample size, but not enough to be within
say +/- 0.5 degree if possible, but could live with 1.  I can not have a +/-
3 degree uncertainty for sure.

Thanks,
John