Hi John et al.,
First I'll clarify our terminology:
- example_func = mid-timepoint image taken from the FMRI (normally
EPI) data
- initial_highres (not normally used) = normally a full brain EPI,
when example_func is restricted FOV
- highres = normally a brain-extracted whole-brain T1 structural
- standard = normally the brain-extracted MNI152 T1 average
The normal usage in FEAT (FSL's FMRI GLM tool) is a two-stage
registration:
1 example_func -> highres (normally 6-9 DOF linear)
2 highres -> standard (normally 12 DOF linear)
The more advanced usage in question here is the three-stage:
1 example_func -> initial_highres (typically 3 or 6 DOF linear)
2 initial_highres -> highres (typically 6-9 DOF linear)
3 highres -> standard (typically 12 DOF linear)
We find that in both cases, using the subject-specific structural to
help the registration is a GOOD THING (e.g., see brief mention and
investigation of this in our NI variability paper). In all cases the
transforms are indeed concatenated mathematically, avoiding repeated
interpolation of the data.
We don't generally find that using a low-res/session-specific T1 as
the initial_highres is particularly useful. The main use is when the
FMRI data is just a few slices, and initial_highres is a single EPI
of the same specification, but covering the whole head.
The latest version of FEAT also includes the ability to unwarp the
EPI data using PRELUDE+FUGUE fieldmap unwarping: this is applied to
the EPI data before any registration begins, and can be really useful.
FEAT doesn't currently use nonlinear registration (apart from the
unwarping mentioned above). We are still evaluating the stability of
this....watch this space. What we are fairly sure about at this point
is that applying nonlinear registration to EPI data (as opposed to
the structural) is REALLY badly conditioned.
For other details on registration inside FEAT, see the FEAT manual:
http://www.fmrib.ox.ac.uk/fsl/feat5/detail.html#reg
and also the relevant lecture from the FSL course:
http://www.fmrib.ox.ac.uk/fslcourse/lectures/flirt_fugue/
flirt_fugue_slides.pdf
Cheers, Steve.
------------------------------------------------------------------------
---
Stephen M. Smith, Professor of Biomedical Engineering
Associate Director, Oxford University FMRIB Centre
FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
+44 (0) 1865 222726 (fax 222717)
[log in to unmask] http://www.fmrib.ox.ac.uk/~steve
------------------------------------------------------------------------
---
On 21 Sep 2006, at 10:25, Ashburner John (PSYCHOLOGY) wrote:
> Steve, if you're reading this could you comment? Feel free to do some
> FSL sales pitch.
>
> I'm guessing that the low quality T1 has the same (phase encode
> direction) distortions as the EPI data, and that the additional
> step is
> in order to correct some of these distortions. It is easier to
> obtain a
> more accurate registration if the relationship between the intensities
> in one image are simply related to the intensities of the other - so
> nonlinear matching a T1 weighted to another T1 weighted image is
> easier
> than matching a T2* weighted to a T1.
>
> If the transformation is indeed a nonlinear one, then there is no easy
> way of doing this within SPM. If it is a rigid-body, then the
> Coregister button should do a reasonable job. Coregistration of
> the low
> quality T1 and the EPI may not be necessary if the images were
> acquired
> in the same session and there was no subject motion.
>
> Best regards,
> -John
>
>
> I have a question. I have seen papers and spoken with people using
> FSL,
> who do the normalization process in several steps. First, they
> coregister
> the low-quality T1 image (that is taken at the same time and in the
> same
> space as the functional data, and thus will coregister much better) to
> the
> functional image. Next, they determine the paramaters of the non-
> linear
> transformation from the low-quality T1 to the high-quality T1 (eg
> MPRAGE)
> and from the MPRAGE to the standard-space template. They then
> multiply
> those two transformation matrices (low quality T1 -> MPRAGE ->
> template)
> and
> apply the parameters to the functional images in order to normalize
> them
> to
> standard space. Hence, my question is: how would I do this in SPM? I
> can
> do each transformation and apply the parameters separately, but
> warping
> the
> functional data twice will create more opportunities for the data
> to be
> mishandled; I would prefer a single warp with the combined parameters.
> Thanks for the help!
>
> Daniel Simmonds
> Developmental Cognitive Neurology
> Kennedy Krieger Institute
> [log in to unmask]
|