Dear Jon,
Registering warped EPI images to structurals can be quite difficult
when there is a lot of distortion as you have. In addition, the fact
that all images are 90 degrees or more out of initial alignment makes
the registration quite hard. Combining large search spaces with
deweighting masks that ignore large image areas, unfortunately,
makes the registration less robust. One quick way to help is to
reslice the images into the same slicing (say axial, like the standard)
by using avwswapdim. Once you've done this you don't need to
use the large search spaces anymore.
I've had a quick look at your data and have a registration which
might be good enough. However, I was a bit lost in amongst the
many registrations you'd previously done. Which was your best
EPI to structural registration? The structural to standard, and
hence the EPI to standard should be easy to get working. It is
always the EPI to structural that is hardest, especially with
distortion present.
So, could you please let me know which EPI to structural registration
you thought was the best, and then I can compare the one I have
made to see if that is a sufficient improvement, or whether it needs
more work.
I'm fairly confident that we can get quite a reasonable registration,
although it must be kept in mind that there will be some deformations
even in the more superior EPI slices, and so we will never get a really
good match. You can only get rid of these distortions by using
fieldmaps or heavily constrained non-linear registration (which,
unfortunately, is not currently available).
Anyway, let me know about the best registration, and you might like
to play with avwswapdim yourself.
All the best,
Mark
Jon Zadra wrote:
> Hello,
>
> I have some fMRI data from an older Varian scanner at UCSD which has
> some warping and signal dropout on the dorsal brain areas.
> Unfortunately, I do not have field maps.
>
> All the information I found about dealing with this problem in FSL
> seems to suggest that best option is to use de-weighting masks for the
> FLIRT runs. None of this information, however, indicates any
> advantages or disadvantages of using masks for the input versus
> reference images. As such, I have attempted to register the data using
> every possible combination of masks for the EPI, structural, and
> template images.
>
> Because my data appeared to be incorrectly oriented, I also attempted
> several of the de-weight mask variations with both -90/90 and -180/180
> x, y, and z search parameters.
>
> I also found some references regarding difficult image registration to
> changing DOF transformations for 2-stage registration. The last
> variations I attempted were, for 2 stage registrations, 1) 12 DOF
> transformations for both stages; or, 2) a 7 DOF transformation for EPI
> to structural and 12 DOF transformation for structural to standard.
>
> I had little or no luck with any of this variations. At best, some
> variations yielded results similar to those obtained without any
> de-weight masks, while others were off by over 90 degree rotations on
> some axes.
>
> Any help or suggestions would be greatly appreciated; I’ve run out of
> ideas!
>
> I’ve posted all relevant information on a public ftp server:
>
> - The original data
>
> - The data resulting from each registration variation
>
> - Slices output (.gif’s) for each registration variation
>
> - The scripts I created and used to create this mess
>
> - A description of the variations I attempted, as well as the results
> from my visual inspections and the results from rmsdiff. (The major
> numbers indicate the main registration steps (varying by input image
> and reference images or images used for 2-stage transformations). The
> decimals indicate variations under their main registration step number
> for each possible combination of de-weight masks.)
>
> The anonymous login server is: ftp://roswell.sdsu.edu
> <ftp://roswell.sdsu.edu/>
>
> Login: ftp
>
> Pass: (your email address)
>
> All files are located in /incoming/zadra
>
> Thanks in advance,
>
> Jon Zadra
>
|
