[MINC-users] minctracc with mutual information and spline-based transformation

[Andrew Janke]

> I had some correspondence with a fellow Minc user, and he indicated to me that mutual information-based variant of minctracc does not work well with a spline-based freeform interpolation, due to insufficiency of data for the local histogram that determines the local spline. Has anyone used minctracc in this way (in a manner similar to Daniel Ruickert), rather than the usual Animal correlation method?

Spline-based in minctracc?  minctracc uses a linear-elastic model
(with local smoothing after each iteration) for nonlinear fitting.
But I presume you mean using mutual information for the nonlinear
fitting.

In any case, yes people have used mi for the nonlinear fits and the
best answer I can give is that it is more hit and miss that using the
correlation co-efficient.  Note that this does not mean that it doesnt
work or that it will work in you case.. :)  now that was very specific
wasn't it? :)

Mind you the use of mutual information has not been incorporated in
the released versions of minctracc so you would have to dig in the
code (in minctracc/Main/minctracc.c -- get_nonlinear_objective()) The
methods that are coded in for nonlinear are:

   xcorr, diff, label, chamfer, corrcoeff and sqdiff

You use them as such:

   minctracc -nonlinear xcorr source.mnc target.mnc xfm.mnc

> Secondly, I'm thinking of weighting a registration method for critical tissues, using a mask computed from each patient's scan. Can anyone comment about the feasibility, with either Animal or Mutual information, of using a mask that correlates with these, or perhaps tweaking the algorithm to weight critical tissues more strongly?

This has been discussed at length at the BIC between various people
and everyone agrees that it should be done.. just no-one has done it
(nicely) yet. To do it you would just turn the -weight argument
variable into a volume.

You can of course achieve this in a script as a quick hack by a bit of
judicious masking (with a weighted mask) before fitting and a bit of
blurring of edges.

Good luck!  Let me know if you need more help with the guts of the
code of minctracc.


a