[MINC-users] hires vs. lowres registration

[Andrew Janke]

On 5/5/05, elo@neurorx.com <elo@neurorx.com> wrote:
> 
> I've encountered a curious situation with image registration and would
> like to know if anyone has any insight or comments:

I'm sure Louis is the one who can provide the most insight here, but
nevertheless here's a few things to think about.

> I'm aligning two images that represent MRI scans of the same object (i.e.
> the human brain) but in two different modalities (i.e. T2 and T1).  The
> registration process is essentially mritoself which calculates the
> transform function needed to register one image to the other, followed by
> mincresample, which applies the transform and places one image in the same
> space as the other.
> 
> Initially we decided to use the hi-res T1 images and to align them to the
> lower resolution T2 images, with the rationale that providing more
> information about the image should lead to a better fit.  But we have
> since found that using the lo-res T1 images actually produces better
> registration (i.e. alignment) results.  Does this make sense and what
> could be happening here within the mritoself and mincresample algorithms
> to cause this?

First, this difference has (well should have!) nothing to do with
mincresample so let's concentrate on minctracc/mritoself.

>From what I can gather you are registering the T1 to the T2 as such:

    minctracc/mritoself T1.mnc T2.mnc T1-T2.xfm

Also given that you are using differing modalities, I presume you are
using mutual information:

     minctracc -mi T1.mnc T2.mnc T1-T2.xfm

Given this information a few things could be happening:

   * mritoself as part of its own internal fitting process does do
some blurring.
 however the size of this blurring is somewhat dependant on the input
resolution.  As such it may just happen that the higher FWHM blurring
caused by initially downsampling the data just happens to be getting
over a local minima in the high res version.

   * Given that you are using mutual information and that the input T1
may be somewhat noisy, the mutual information "image" that is
calculated internally as part of the fit may be somewhat spurious. 
Down sampling the data first is effectively blurring it.  Thus your
input T1 is more smooth in the low-res version.

    * Lots of other things....

You can easily test the second by blurring the input high-res data
before fitting at the same level as the downsampling would achieve.
-- 
Andrew Janke      (a.janke@gmail.com || www.cmr.uq.edu.au/~rotor)
Australia->Brisbane     H: +61 7 3390 6332  || M: +61 4 2138 8581