Hi Mark,
I think I'll go with FNIRT (and maybe avoid spiral in the future!). The mismatch between the MNI template and DLPFC isn't as bad in some subjects, but FNIRT should certainly be more consistent than FLIRT on average. I think this will be particularly true around the basal ganglia.
Thanks again for your feedback and comments.
Cheers,
David
On Jul 13, 2011, at 3:40 AM, Mark Jenkinson wrote:
> Dear David,
>
> I think you are right - as you've used the whole head images to
> do the FNIRT registration, then it is simply unfortunate that it
> amplifies the small errors in the FLIRT registration of the
> example_func to highres. Of your two options I think I would opt
> for number 2 and do FNIRT but simply realise that for DLPFC your
> results will not be quite in the right place. I think this is worth it
> as it will be better for other areas and may well give more consistency
> even in the DLPFC, just consistently misplaced, and it is the
> consistency which will affect your stats.
>
> Sorry I don't have better news.
> All the best,
> Mark
>
>
> On 12 Jul 2011, at 23:28, David V. Smith wrote:
>
>> Hi Mark,
>>
>> Thanks for your input! Unfortunately, I did not get lucky with the non-smoothed image that had a better brain extraction...
>> http://www.duke.edu/~dvs3/example_func2standard_FNIRT_new.png
>> http://www.duke.edu/~dvs3/example_func2standard_FLIRT_new.png
>>
>> I assume I only have two options now (though both still suffer from the drawbacks of spiral data that you pointed out):
>> 1) Abandon FNIRT (and just go with FLIRT) and potentially see better alignment of DLPFC and other superior structures since those areas seem to be the most distorted/compressed by the interaction of FNIRT and an imperfect example_func2highres transform. However, this (FLIRT-only) approach will seemingly increase the cross-participant variability in subcortical structures since an affine transformation won't be able to perfectly align participants whose ventricles vary in size.
>> 2) Stay with FNIRT and accept the imperfect registrations around DLPFC (though these imperfections seem rather bad). With this approach, however, at least the subcortical structures will (I think) be generally better-aligned across participants.
>>
>> Thanks again,
>> David
>>
>>
>> On Jul 12, 2011, at 2:08 PM, Mark Jenkinson wrote:
>>
>>> Hi David,
>>>
>>> Ah, spiral data - this adds another level of complexity!
>>> In spiral data you don't have distortions like you do in rectilinear EPI.
>>> So even fieldmaps wouldn't have helped.
>>> What you are probably seeing it spatially-variable blurring and
>>> signal loss. This is likely to be what is causing the geometric mismatch.
>>> There isn't even a registration solution to this, as the spiral sequence
>>> doesn't mislocate signal, it blurs it. So that limits your options.
>>>
>>> I would definitely try again with a non-smoothed example_func.
>>> It is never a good idea to try registration using a smoothed image,
>>> so you might be lucky. If you can do a better brain extraction that
>>> *might* help, but I'm not confident.
>>>
>>> I think you might have to accept the fact that the registration will
>>> not be perfect and you will have to factor this into the interpretation
>>> of where your final activations are actually located.
>>>
>>> Best of luck!
>>> Mark
>>>
>>>
>>> On 12 Jul 2011, at 18:58, David V. Smith wrote:
>>>
>>>> Hi Mark,
>>>>
>>>> Thanks for your reply. I do not have fieldmaps for these scans. They're from a spiral-in SENSE sequence, which seems to help with the ventral frontal distortions, but it doesn't fix all distortions.
>>>>
>>>> I took a closer look at highres.nii.gz and example_func2highres.nii.gz, and they are generally aligned using 6 DOFs. However, the alignment probably isn't perfect, and I can see that the edges (particularly around DLPFC) aren't matched up: the functional data is compressed/distorted relative to the highres.
>>>>
>>>> Since I don't have fieldmaps for this particular dataset, what are my options (assuming I have some)? Will masking the example_func better help? It seems to have some excess non-brain material, even after BET. Will an unsmoothed example_func help? These were originally smoothed with a 6mm kernel since I made the example func from my preprocessed data.
>>>>
>>>> Thanks!
>>>> David
>>>>
>>>>
>>>>
>>>> On Jul 12, 2011, at 12:32 PM, Mark Jenkinson wrote:
>>>>
>>>>> Hi David,
>>>>>
>>>>> Did you have fieldmaps for these EPI scans?
>>>>>
>>>>> I suspect that the problem is actually in your example_func2highres
>>>>> scan and made slightly more prominent in the standard space. The
>>>>> areas which don't match well look like the result of EPI distortion to
>>>>> me. You highres2standard looks great, so that means that it really
>>>>> isn't FNIRT that is the problem. There's a limit to how well I can really
>>>>> tell from the images for the example_func2highres as the edges are
>>>>> always rather poorly estimated, but have a close look in FSLView and
>>>>> see if in fact the DLPFC is well aligned in these or not.
>>>>>
>>>>> All the best,
>>>>> Mark
>>>>>
>>>>>
>>>>>
>>>>> On 12 Jul 2011, at 17:19, David V. Smith wrote:
>>>>>
>>>>>> Hello,
>>>>>>
>>>>>> I'm trying to improve my FNIRT registration (example_func2standard); however, I'm not sure where to begin because (I think) my highres2standard looks fairly reasonable (both for FNIRT and FLIRT):
>>>>>> http://www.duke.edu/~dvs3/highres2standard_FLIRT.png
>>>>>> http://www.duke.edu/~dvs3/highres2standard_FNIRT.png
>>>>>>
>>>>>> Nevertheless, for the FNIRT version of example_func2standard, you can see that FNIRT is shrinking DLPFC a little too much in the example_func2standard, especially when you compare this to the FLIRT version.
>>>>>> http://www.duke.edu/~dvs3/example_func2standard_FLIRT.png
>>>>>> http://www.duke.edu/~dvs3/example_func2standard_FNIRT.png #note misaligned DLPFC here
>>>>>>
>>>>>> (The example_func2highres also looks fine.)
>>>>>> http://www.duke.edu/~dvs3/example_func2highres.png
>>>>>>
>>>>>> You can see my commands here: http://www.duke.edu/~dvs3/example_reg.sh (sorry, attachment was rejected). I think I really only have three (minor) deviations from the defaults:
>>>>>> 1) using sinc interpolation: -interp sinc -sincwindow hanning
>>>>>> 2) using the header: -usesqform
>>>>>> 3) using our own study-specific template from ANTs and FLIRTed to MNI. I don't think this is a problem, but you can inspect the images, if you like:
>>>>>> http://www.duke.edu/~dvs3/MNI_diffeo.nii.gz
>>>>>> http://www.duke.edu/~dvs3/MNI_diffeo_brain.nii.gz
>>>>>>
>>>>>> Anyway, I've already tried a couple of things on the FNIRT documentation (e.g., adjusting the standard_mask size, specifically making it smaller; and applying the mask only in the final iterations: --applyrefmask=0,0,0,0,1,1), but nothing really preserves the shape of DLPFC (or really even makes a noticeable difference in the output). Is there anything else I could try? Or is this something I shouldn't even worry about?
>>>>>>
>>>>>> Thanks!
>>>>>> David
>>>>>>
>>>>>> --
>>>>>> David V. Smith
>>>>>> Graduate Student, Huettel Lab
>>>>>> Department of Psychology and Neuroscience
>>>>>> Duke University
>>>>>> Durham, NC 27708
>>>>>> http://www.mind.duke.edu/huettellab/
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>
>>
|