Dear Spmers
Thanks for the responses Daniel and Marko. I have looked over the brett
website
as I have in the past and Daniel's data and I am still wondering about
which is
the best way to proceed. Marko suggested using the structurals /
coregistration
and Daneil suggested using the epi to epi template using masking with my low
resoultion epi data (7x4x4 mm on 1.5 T Philips scanner) with a structural of
same dimensions and no field map.
Marko, f I do a check reg on my unsmoothed normalized using my prior
suboptimal
technique of epi normalization to epi template (no masking other than defaults
to account for warped regions), each subjects appears pretty much aligned for
epi image to epi template (eyeball method and pixel surfing) except for some
pulling in distorted regions which end up outside of the group mask. Still,
there are peaks of activation in the ventrilces on the smooth statistical
images.
I guess the bottom line initial question that I need help with is
understanding
what coregistration mutual information is doing when you have epi distortion /
signal dropout. Is there a way to mask out or not give weight to these regions
in determining epi to structural coregistration. What is the program actually
doing to account for such issues- I use spm2 at the moment with mutual image
registration.
Thanks a lot
Sincerely,
Jeff L.
Jeffrey Lorberbaum, MD
Assitant Professor of Psychiatry
Penn State Univeristy Hershey Medical Center
Quoting Daniel Simmonds <[log in to unmask]>:
> Hi Jeff,
> There are some tutorials by Matthew Brett here on how to deal with this:
>
> http://imaging.mrc-cbu.cam.ac.uk/imaging/ProcessingNormalization
>
> We have tried both the EPI cost function masking method and the
> intermediate grey matter method (the two talked about on the page) on some
> of our older Philips 1.5T data (two separate samples of children, one with
> n=13, the other n=25). Both seemed to work reasonably well; however, we a
> different disadvantage present in each method.
>
> Problem with the EPI masking method:
>
> The biggest issue is that you have to do this by hand. In order to do
> it in a reasonable amount of time, you'll probably overestimate when
> masking the dropout regions, which will then not be taken into account when
> normalizing. This leads to (what we saw in our data) less coverage in the
> dropout regions on the group level than with the structural intermediate
> level.
>
> Problem with the structural intermediate method:
>
> On the group level, the structural intermediate method slightly
> underestimated the size of the template when normalizing, and the EPI
> template slightly overestimated it. While this didn't seem too significant
> on the unstatistical "eyeballing" level, it made a big difference when it
> came to group modeling, because 1) on a single subject level, modeling
> discards voxels that don't exceed a certain intensity threshold, and 2) on
> a group level, the group model only uses voxels that are present in every
> individual model. This led to a "shrinking brain effect" (I attached a
> picture to show it). As the task was a finger-tapping task, we had robust
> activation in the SMA and contralateral primary motor cortex. We ran
> paired t-tests between the two normalization methods within each subject,
> which showed to a significant degree on the group level that cluster in the
> SMA was partially dragged into the cingulate (the primary motor was also
> significantly dragged down).
>
> My conclusion from this is that it depends what areas you are primarily
> interested in. I don't know how this would extrapolate to 3T data or
> higher resolution EPI data in general (my gut feeling is that the
> structural method does a poorer job because of the coregistration with low
> resolution EPI data). Part of my reason for writing such a long response
> is that I wonder if anyone has thought about any of these things or has any
> bits of wisdom to share.
>
> Hope this helps.
>
> Dani
>
> Daniel Simmonds
> Developmental Cognitive Neurology
> Kennedy Krieger Institute
> [log in to unmask]
>
> On Mon, 8 Jan 2007 13:22:02 -0500, [log in to unmask] wrote:
>
>> Dear SPM group
>>
>> We have been using spm2 to analyze fMRI data (n=40) obtained from Philips
> 1.5 T
>> scanner (saggital slices; voxels 7 x 4 mm x 4mm). We have a T1 structural
> scan
>> obtained in the same planes for all but no field map was obtained. There
> is epi
>> distortion mostly in OFC regions.
>>
>> Using an approach with spatially normalizing the epi images to the epi
> template
>> (4x4x4 voxel output in template bounding box [-90to 91, -126 to 91, -72 to
>> 109]) and spatial smoothing 8x8x8, I end up with peak activations in the
>> ventricles that appear to be due to a shift / misregistration.
>>
>> This ventricular activity occurs within some subjects and at a group level
>> (n=40). Within-subjects, the spatial normalization comparing the mean epi
>> normalized image to the epi template looks generally good to my eye using
> check
>> reg and surfing at different points in the brain although though the
> images are
>> obviously blurry.
>>
>> My questions mainly is what is the best way to spatially normalize my data
> given
>> that I don't have a field map.
>>
>> (Ideas:
>> (1) Is there a way to get rid of epi distortion post-processing without a
> field
>> map
>>
>> (2) Am I stuck with spatially normalizing the epi images to the epi
> template
>> using some masking of distorted regions?
>>
>> (2) Should I co-register structural and epi images to each other even
> though
>> there is epi distortion in the OFC and spatially normalize:
>> (a) the structural to the structural template
>> or
>> (b) gray matter template to gray matter template
>>
>> (3) Is there a way to mask out distorted areas in coregistration (if I
> were to
>> coregister structural and functional images) and then do #2))
>>
>> Thanks a lot,
>> Jeff
>>
>>
>> Jeff Lorberbaum, MD
>> Assistant Professor of Psychiatry
>> Penn State University- Herhsey College of Medicine
>
>
|