Hi Stephen, hi Wolf,
On Jun 21, 2010, at 12:33 AM, Stephen Smith wrote:
> Hi - to answer Matthew's question first - as far as I know, it is not common practice to bias-correct 4D FMRI data. This is probably because "typical" bias field, while being a problem for structural segmentation, is not a large enough effect to worry about in FMRI data.
>
> However, Wolf's points are good ones, and I suspect that as more people start using large-array coils with stronger bias fields than has commonly previously been seen, we should quite possibly make some correction. For example, our new 32-channel head coil (is excellent but) does produce stronger bias than we've previously been used to.
I have seen data from Lawrence Wald (at MGH, also see http://www.ncbi.nlm.nih.gov/pubmed/19904727) comparing coils of ascending channel counts (up to 96 I think), showing that this bias is mainly coming from higher signal close to the coil, not so much from less signal in the center. The upshot of this is that it is not the center that would need correction but rather areas close to the coil. But that seems not like a good idea to scale down the signal there :). (It should be relatively simple and statistically "conservative" to raise the required significance thresholds in voxels close to the coil, but at the same time pretty counter productive; in comparison relaxing thresholds in areas where the measurements are more uncertain feels rather dangerous).
> The thermal noise is still the same everywhere, while the BOLD contrast and physiological noise will fall off in the areas where the mean signal is lower (e.g. in the centre of the image). Hence you have variable zstats across the image, which will be hard to remove.
Now, here is where I am going to embarrass myself, often we go and compare signal (say a beta value describing the amount of BOLD modulation in some way) with residual noise (say using a t-test). Simple intensity correction of the Bias field will affect both values pretty much the same way. I would expect that the t-test comparing signal and noise should be pretty much independent from common scaling (but hey I did not go and actually confirm this so I could be "out to lunch").
>
> If you bias correct the 4D data, the BOLD contrast (and hence PE and COPE values) will then be more comparable across space and across subjects; arguably maybe the best way of doing this is simply to convert PEs and COPEs to % signal change before doing group-stats (etc), though that is poorly conditioned in some areas e.g. edge voxels which are dark because of partial voluming, and interactions of these effects with head motion cold be nasty. Also, things may get even more complicated when one considers parallel imaging.
Doesn't % sinal change also depend on measurement parameters like SNR? I have a hunch that thermal noise would create a lower floor for any signal and distance from receivers will limit the ceiling of the signal, creating limits on achievable % signal change? Again I am unsure about the validity of my thoughts and would love to learn how to correct for this...
Best
Sebastian
>
> Cheers.
>
>
>
> On 20 Jun 2010, at 12:04, wolf zinke wrote:
>
>> Hi Matthew,
>>
>> Thanks for this question, this gives me an opportunity to ask something related I was often wondering about.
>>
>> First a remark to your question. Why do you want to bias correct your 4D data? This would correct for (static) differences of the signal intensity over the brain. But when collecting functional data you are interested in the BOLD signal fluctuations, not the static bias, hence I doubt that you gain anything with a bias correction. I would even assume, that this is likely to introduce some problems. In my opinion, the susan smoothing copes somehow with the bias problem on a local scale. However, if you have got non-BOLD signal fluctuations over time (e.g. B0 fluctuations, physiological artefacts) you might be interested to correct for these. Anyway, I am not an FSL expert, so I hope that the expert will come up with some more competent comments on this issue.
>>
>> My follow-up question here is related to the consequences of a signal bias on statistical maps. If there is a strong gradient in the data (macaque data, surface coil), the SNR changes within the brain volume, hence a common statistical threshold would result in biased maps, making it more difficult to detect activity down the signal gradient. Should this be accounted for when applying significance thresholds to the data? In my case, detecting 'strong activations' in parietal and frontal regions, but 'weak activations' in lower temporal regions might simply be attributed to the signal gradient and not to the underlying physiological processes. What is a common practice o cope with such a problem?
Erm you could place your surface coil over each area of interest in separate experiments and report the differences due to coil position? (I would also push for more channels as this makes a huge difference for NHP imaging in my experience).
>>
>> Thanks for any comments,
>> wolf
>>
>>
>> On 06/20/2010 02:41 AM, R. Matthew Hutchison wrote:
>>> Hi FSL users,
>>>
>>> Is it common practice to bias correct the 4D EPI data?
>>> There doesn't seem to be an option in the FAST GUI to do this, so perhaps there are clear reasons not to.
>>>
>>> Any insight would be much appreciated.
>>>
>>> R. Matthew Hutchison
>>>
>>> --
>>> R. Matthew Hutchison, PhD. Candidate
>>> Centre for Functional and Metabolic Mapping
>>> Robarts Research Institute
>>> Cuddy Wing - 9.4T Suite
>>> P.O. Box 5015, 100 Perth Drive
>>> London, Ontario, Canada N6A 5K8
>>
>
>
> ---------------------------------------------------------------------------
> 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
> ---------------------------------------------------------------------------
>
>
>
--
Sebastian Moeller
telephone: 626-395-6713
fax: 626-395-8826
German GSM: 0 15 77 - 1 90 31 41
[log in to unmask]
Division of Biology
MC 114-96
California Institute of Technology
1200 East California Boulevard
CA 91125, Pasadena
USA
|