Dear Drew,
what about using New Segment (Segment in SPM12) to estimate and save the
bias field from one of your fMRI images and then apply it to all of your
images (eg with i1.*i2 in ImCalc)?
All the best,
Guillaume.
On 13/03/14 15:35, Chris Watson wrote:
> I don't see how % signal change would "deal with" inhomogeneity.
> What does an example image look like?
>
> On 03/13/2014 11:02 AM, Sevel,Landrew S wrote:
>> I wanted to follow up on this briefly. In terms of the subjects with data
>> removed from the first level mask image, this seems to be due to
>> inhomogeneity in almost every case.
>>
>> I've also noticed that using SPM's brainmask image as an explicit mask and
>> setting the mask threshold to -inf seems to result in hugely intense
>> voxels (F=6000, in one case) in the cerebellum that must be error.
>>
>> It was suggested, and I noticed previous mention of converting images to
>> percent signal change to deal with inhomogeneity. Is this a sensible
>> alternative? If so, when in processing should this be done and what are
>> acceptable calculations to do so?
>>
>> Many thanks,
>>
>> Drew
>>
>> On 3/11/14 2:33 PM, "H. Nebl" <[log in to unmask]> wrote:
>>
>>> The preprocessing sounds alright (but make sure you used correct
>>> parameters during slice timing). You could slightly increase the
>>> smoothing kernel, for example 8 mm, but I don't think this will have much
>>> influence. See below for some ideas:
>>>
>>> Some issues related to motion:
>>> ------------------------------------------
>>> - Slice timing will result in spreading of artefacts from volumes with
>>> large motion into adjacent, originally unaffected volumes to some extent.
>>> Make sure the correction with ART toolbox accounts for that aspect (e.g.
>>> correcting/replacing/deweighting the adjacent volumes as well if
>>> necessary)
>>> - As far as I remember the default threshold in ART toolbox is pretty
>>> liberal. I would look at the data again and check how many volumes show
>>> "fast head motion" defined as say >0.5 mm/TR.
>>> - If a larger portion of the volumes is affected exclude the entire run.
>>> What about the beta/contrast estimates based on the other two runs then?
>>> - Are there any correlations between stimulus presentation and motion?
>>> Some subjects might move every time a stimulus is delivered.
>>>
>>> Some issues related to design matrix:
>>> -------------------------------------------------
>>> - Make sure you use an appropriate high-pass filter
>>> - Make sure the derivates of condition A don't correlate with regressor
>>> for condition B (this might be the case in a fast-event related design)
>>> - You might want to turn to FIR models (as suggested by Chris, canonical
>>> HRF might not be appropriate)
>>>
>>> About the mask:
>>> ----------------------
>>> Find out why some of the voxels were missing in the second-level mask.
>>> Independent of the issue with your coil, these regions might not have
>>> been covered at all in some of the subjects or there might be large
>>> susceptibility aretefacts (amygdala, OFC). If these voxels have a low
>>> intensity they are excluded from individual first-level masks
>>> automatically. Adjusting the masking threshold won't necessarily help,
>>> the voxels become part of the analysis but you might well obtain beta
>>> estimates close to zero, as there is still hardly any signal.
>>>
>>> About the between-session variability:
>>> --------------------------------------------------
>>> - Is it related to motion or signal loss in some of the session (see
>>> above)? One "bad" session might destroy your effects.
>>> - Do the behavioral scores vary as well?
>
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|