I cannot do this in medx for two reasons: 1) the FSL utilities work only
with 3D registration; 2) even when I do the registration with air medx
cannot save a 2D transform, it can only save a 3D (that is really odd!!!,
and disappointing).

I tried to to 2D transforms in the flirt GUI, without changing cost
functions. I attempted first with my original images (t2 as reference, t1 as
input). I got the message "child killed: SIGABRT", the output file was not
written, but the .mat file was. Since the reference image has very different
contrast (and pretty bad contrast for that matter as compared to the t1
weighted), I also used a t1 weighted image as reference, which I had
obtained from a successful 2D registration in medx. I got the same error
message. I enclose here the transformation (.mat) files I got with the 2
procedures. Should I try to apply the transforms to a 2D image and to a
reference volume from the command line?
Thanks for your help. Stefano



-----Original Message-----
From: Mark Jenkinson [mailto:[log in to unmask]]
Sent: Monday, January 21, 2002 3:32 AM
To: [log in to unmask]
Subject: Re: [FSL]


Dear Stefano,

> Here is my starting problem, it is still quite unclear to me how to do
> things based on your previous email.
>
> My reference image is a single slice. I have another 3 slices in the same
> acquisition, but these slices are 7.5 mm thick, with a 11 mm gap.
>
> Then I have a volume that I want to register to the single slice. I assume
> that the head has moved only in-plane between the two scans (I think that
is
> a reasonable approximation for what I want to do).

Yes - this all sounds fine.
For a 2D registration you do want a single slice for input and a single
slice for reference.

> I thought I would register the best estimate slice of the input with the
> reference slice, save the transform and apply it to the full input volume,
> the same transform for each slice.

Again, this is right.  The first step is to get the 2D registration
using only single slices.  Once you've done this you need to apply
the transformation and only then do you need 3D volumes.

> When I tried this, I used different cost functions and a 2D to 2D rigid
body
> 3 parameter model. In both cases I get an error message:
> child killed SIGABRT
> It saves a .mat file but not the output file. What should I do?

OK, if you used a single slice as the input AND a single slice as the
reference then using the 2D->2D parameter model was correct.  There
should be no need to change the cost functions.  If this isn't
working then we need to debug your images.  However, for the moment
I will assume that this will work if you try it again with single
slice inputs and references.

> I tried to save the original reference volume as a 4 slice 18.5 mm thick
> slice volume. Then I registered the equivalent t1 weighted volume to that.
> The registration was not satisfactory.

The 4 slices volume is unnecessary for you.

How many slices do you have in the input volume?
Lets assume for the moment that you have 10 slices in a volume that you
want to apply the 2D transformation to.

At this stage you have already done the registration - all we want to
do is apply the transformation.   Now in MEDx the transformation file
(saved with save shadow transform) contains information about the original
image size, so you can't easily apply this to a volume.  In fact,
in MEDx the easiest thing is to apply the 2D transformation (the MEDx
shadow transform) to each of the 10 input slices separately and
then group the results together as a 3D volume.

> The other option (which I think you describe in your email) is to create a
> reference volume with the same dimensions as the desired output of the
> shadow transformed t1 weighted file. I tried doing that starting from my 4
> original slices, but with little success. I can resample the volume to
make
> the slices of the right thickness, but should I add extra slices on top
and
> on the bottom to fit the final output volume?

The *alternative* to the above is to do this outside of MEDx.  In that case
you can make a reference volume that has 10 slices (each slice being the
same size as the *reference* slice).  Then you apply the flirt transform
(the .mat file) to the input volume (10 slices), using this new
reference volume (10 slices) to get an output volume (10 slices).
But, I repeat, this is only for doing it *outside* of MEDx.

Hope this makes it clear,
        Mark


> -----Original Message-----
> From: Mark Jenkinson [mailto:[log in to unmask]]
> Sent: Thursday, January 17, 2002 9:58 PM
> To: [log in to unmask]
> Subject: Re: [FSL]
>
> Dear Stefano,
>
> In answer to some of your questions...
>
> > When I use sinc interpolation, which filter should I use and what size?
> > I used Haninng 7 voxels for registering the images prior to
segmentation,
> > how does that sound (there must be a reason why it's default) ?
>
> There is no general concensus on what interpolation gives the best
results.
> We left 7 voxel Hanning as the default as it gave good results in a
> reasonable time.  But the only way to really know if other kernels might
> be better is to try them.  However, my suspicion is that it won't be
> that significant - there isn't anywhere near as much difference as there
> would be between sinc and trilinear.
>
> > The other question is where are the transform files saved after
> > registration? In the GUI I see no option to save the transform files,
yet
> I
> > imagine that the .mat files that are created may be the xform. Can I
take
>
> Yes, that's right.  The .mat files are the transformation files that
> flirt saves.  They are in ascii format (not MATLAB) and can be viewed
> as text files (a 4x4 affine matrix).  However, they are not MEDx
transforms
> and so if you want a MEDx transform then you should use the save shadow
> transform option immediately after getting the transformed image.
> An alternative is to use the command line program "convert_xfm" to
> convert the .mat file into a MEDx .xfm file.
>
> > these files and shadow-transform other volumes? How do I do that?
>
> Once you have the .xfm MEDx version you can happily shadow-transform away
> as normal.
>
> Note: If you want to use it outside of MEDx then you need to
> use the -init and -applyxfm options to the flirt command line program.
>
> > What are MRF's?
>
> Markov Random Fields.
>
> They are a mathematical form for imposing neighbourhood relations between
> the probabilities that helps to do spatial regularisation (like smoothing)
> to the internal segmentation-finding parts of FAST.
>
> Cheers,
>         Mark
>
> > -----Original Message-----
> > From: Stephen Smith [mailto:[log in to unmask]]
> > Sent: Wednesday, January 16, 2002 5:20 PM
> > To: [log in to unmask]
> > Subject: Re: [FSL]
> >
> > Hi
> >
> > On Tue, 15 Jan 2002, Marenco, Stefano (NIMH) wrote:
> >
> > > Is there any particular order I should apply susan and brain
extraction?
> > > I seem to get different results when I invert the order of these
> > operations.
> > > Is this to be expected?
> > > And what does susan actually do?
> >
> > They are both nonlinear processes so yes, the order will affect things.
> >
> > You should not normally need to apply susan before BET. In general, we
do
> > not use SUSAN before our other programs as they are all designed to be
> > robust. For example, the FAST segmentation effectively includes its own
> > "spatial regularisation" (MRFs on the class labels), so SUSAN is
> > definitely not recommended in that case.
> >
> > SUSAN is a nonlinear noise reduction algorithm - see
> > http://www.fmrib.ox.ac.uk/~steve/susan
> >
> > ttfn  :)