Registering EPIs to structurals is really hard when there is a lot of
It is unlikely that your problems are to do with initial alignment and
space so much as the amount of distortion present and the typically poor
resolution and anatomical contrast of EPI.
Non-linear registration is not a cure in these circumstances either because
of the amount of signal loss that usually occurs, since non-linear
can only cope with geometric distortion.
Therefore, the best method is to use field maps.
However, if you do not have fieldmaps then I would still not advise
non-linear (or even 12 dof) as it will tend to incorrectly stretch the
brain to fill in the "gaps" in the EPI that occur due to signal loss.
So in this case the best thing to do is to use cost function weighting
and draw (or derive by some other means) a volume which has zero in
the areas where there is bad signal loss or distortion (a large blob
around the inferior frontal and temporal areas normally does the
trick) and has one in all other areas (including the background).
Then if you register using this volume as a cost function weight and
stick with 6 dof (or 7 dof if you mistrust the precise volumetric
calibration on the scanner - which is sensible if the structural was
not acquired in the same session) the registration should be improved.
If this is the case then you will need to do this registration
manually and generate the appropriate matrices to be put in the "reg"
subdirectory of the ".feat" directory in order for feat to work
correctly for higher level analyses.
All the best,
Fornito, Alexander wrote:
>That's a big help - thanks!
> Another question re: flirt, this time in FEAT.
>I am trying to run FEAT on our 3T EPIs and have been running into some
>problems with the initial step of registering the epi to the structural
>(using 7 dof). In the registration results window, the epi-to-structural
>results are way off; the epi appears with the anterior end facing down
>(dorsal surface to the left), and only slight portions of the red lines
>representing the structural can be seen.
>The structural to standard is fine, and the final epi to template isn't
>totally off, but could possibly be better.
>Unfortunately, we do not have field maps for this data. Is the gross
>abnormality in the initial registration due to the biases in the
>uncorrected epi, or because the images might not be in the same
>orientation to begin with (I've used full search for the registration,
>but is this sufficient)?
>Would I be better off skipping the epi-to-structural step and
>registering my epis to an epi template using nonlinear registration (as
>implemented in SPM), creating a 4D file out of these images, and then
>Many thanks again,
>From: Mark Jenkinson [mailto:[log in to unmask]]
>Sent: Monday, July 19, 2004 7:13 PM
>To: [log in to unmask]
>Subject: Re: [FSL] fasting after flirting
>We always recommend running FAST on untransformed images.
>That is, on the originally acquired images, after running BET to remove
>non-brain structure, but without applying any spatial transformations.
>The reason that intensities change with transformation is that
>is required to look-up intensities between the original grid points.
>If you want to see your images in MNI space or to do the tracing in this
>space then that is fine - just register the images, transform them to
>MNI space, do any tracing in this space and then transform your
>masks back into the original space. To do this last step you need to
>invert the transformation (use InvertXFM or convert_xfm) and then
>make sure that you make a floating point output image (for the mask)
>and rethreshold it to the desired value (near 0.5 keeps the size about
>the same, near 1.0 will "shrink" the mask to leave only voxels which
>overlapped greatly with the mask in the MNI space, and near 0.0
>will "enlarge" the mask, including any voxels with even small overlaps
>with the MNI-space mask). Then, once you've rethresholded, you have
>a mask in the original space, and can use this in conjunction with the
>FAST output from native space.
>Also, FAST works best when it has access to the entire brain (excluding
>non-brain structures) so don't use it just on an ROI in case that was
>something you were thinking of. To get the amount of gray matter in an
>ROI just multiply a binarised mask (that only contains 1's and 0's) of
>ROI with the gray matter PVE output from fast and sum the output.
>Hope this all makes sense.
>All the best,
>On Tuesday, July 13, 2004, at 02:36 am, Alex Fornito wrote:
>>I am doing manual tracing on some T1 images. I've used flirt to align
>>by registering the stripped T1 to the MNI template, and then applying
>>transform back onto the unstripped image.
>>I've noticed that doing this changes the pixel intensities. Since I
>>want to segment these brains using FAST, I was wondering if you could
>>a - exactly why do the pixel intensities change?
>>b - does this introduce some error when I last segment with FAST?
>>Basically, if I trace regions on the unsegmented image, and then
>>to segment the grey matter within the ROI using fast will this be an
>>adequate representation of the region's grey matter volume?
>>Thanks for your help,