Hi Ben,
FMRI data is quite noisy when you examine the timecourse of a single voxel.
In a clean, low motion subject, my 3T data shows a noise range of
2-3% after realignment, and ~1% after 8 mm smoothing. If the interpolation
method plunked in a value with an error of less than 1%, it would barely be
visible among the native noise.
The ArtRepair package includes a function that will deweight repaired scans
so that they don't enter into the final estimate. This method seems cleaner,
but it is tricky to use ("the Repair and Compare" function) and it is
not clear how much more accuracy it provides.
It is very difficult to quantify the effects of deweighting. Ideally, one would want
a phantom with known values to try out different methods. But the best phantoms
that I'm aware of (POSSUM from FSL, and the Vanderbilt phantom (Morgan 2001))
may not yet fully capture the noise properties of actual ugly data.
Regards,
Paul
----- Original Message -----
From: "Ben Yerys" <[log in to unmask]>
To: "Paul Mazaika" <[log in to unmask]>
Sent: Wednesday, December 17, 2008 12:43:24 PM GMT -08:00 US/Canada Pacific
Subject: Re: Artifact detection and removal: Criteria
Hi Paul--
I hope you don't mind the follow up question, but I was wondering whether
the use of the "interpolation" repair method affected the accuracy of
modeling the hemodynamic response when adjacent (unrepaired) scans were from
a different condition than the bad scan? It seems reasonable that if we
have a block design with a bad scan in the middle of the block that
interpolation would be an ideal method to repair artifacts, but what about
event related designs or on scans that border two block types? I'm sorry if
this question reflects a naivety about fMRI, but my training is in clinical
child psychology and I've only been using fMRI for a few years.
Thanks so much,
Ben
On Wed, Dec 17, 2008 at 3:24 PM, Paul Mazaika <[log in to unmask]> wrote:
> Dear all,
>
> I'd like to comment on repairing/removing slices and volumes, which
> is somewhat controversial. We have chosen to do repairs in our analyses
> (ArtRepair toolbox) for several reasons.
>
> Often the artifacts are huge, perhaps >10% signal change, which is
> far greater than the signals of interest, so it is important to get
> them out of the analysis.
>
> Sometimes there are only a few artifacts, and we would really like
> to use the data because it is expensive to collect, and 98% of the
> data is perfectly fine. Rescanning may be impossible.
>
> One could choose to deweight the bad scans (RobustWLS) or regress
> them out (Lemieux 2007). These methods have the advantage of preserving
> the assumed Gaussian properties of the noise. My opinion is that
> the noise is not exactly Gaussian, as it is a combination of white noise
> and physiological variations (see the Statistical Parameter Mapping book),
> and the user may not have all the effects modeled correctly.
> Of course, as long as the noise is near enough to Gaussian,
> then the statistics would nearly be correct.
>
> Repairs generally make the overall noise properties
> closer to Gaussian then the original data including the artifacts.
>
> For group studies using a mixed level analysis, we want the
> most accurate contrast estimates for each subject. While repairing data
> might affect the single subject t-map by breaking the Gaussian assumption,
> the group analysis does not use those t-maps.
>
> Repairs are easy to automate, and can be applied slice wise and voxel
> wise. (Someday, I hope there will be an AFNI style 3dDespike, for
> SPM in the future.)
>
> Best regards,
> Paul
>
>
>
>
>
>
> ----- Original Message -----
> From: "Dorian P" <[log in to unmask]>
> To: [log in to unmask]
> Sent: Wednesday, December 17, 2008 3:09:23 AM GMT -08:00 US/Canada Pacific
> Subject: Re: [SPM] Artifact detection and removal: Criteria
>
> Dear all,
>
> I've used another method to account for artifacts. It's called
> Robust-Weighted Least Mean squares
> (http://www.bangor.ac.uk/~pss412/imaging/robustWLS.html<http://www.bangor.ac.uk/%7Epss412/imaging/robustWLS.html>
> )
>
> Instead of removing images from analysis it corrects their
> participation to the final analysis by weighting their "importance".
> It doesn't correct big artifacts (I had a subject moving 20mm), and it
> doesn't alter the results compared normal SPM analysis (I compared
> them very well). But it gives slightly stronger results at the 1st
> level. This is translated in similar paterns of activation in the 2nd
> level, but some "spurious" activations are smaller and the global
> maxima is shifted to the most interesting cluster theoretically (at
> least in my case).
>
> Hope this helps, or... If I am going in the wrong direction to have
> some feeback :)
>
> Dorian.
>
>
> 2008/12/17 James Rowe <[log in to unmask]>:
> > Dear Chris and Volkmar,
> >
> >> the simple answer is: you should use ArtifactRepair/TSDiffAna/some other
> >> tool to assess the quality of your data, but you should never attempt to
> >> "repair" your time series by removing/replacing slices/scans.
> >
> > I broadly agree, but there is another approach that may help. If you have
> an
> > objective criterion (predefined), applicable to the tsdiffana output, for
> an
> > artefact (eg spike) then this can be used to create a nuisance regressor
> > identifying an affected volume, effectively removing that volume from
> > estimation of the parameters for experimental effects. This is not
> removing
> > or repairing data, but accommodating the artefact within the GLM.
> >
> > This is similar to what happens when including movement parameters and
> their
> > temporal derivatives: a sudden large displacement gives a 1 or 2 volume
> > 'peak' in the rp' regressor, against a low value for other volumes.
> However,
> > with an objective threshold, this can be binarised, to 'ok scans = 0' and
> > 'this bad volume =1'. If the threshold is too low, you will have too many
> > nuisance regressors (one for each bad volume) burning up degrees of
> freedom
> > and adversely affecting the normal distribution of residuals. However, a
> > handful of such regressors for a long study would seem a reasonable way
> to
> > remove the effects artefacts like occasional spikes.
> >
> > Best wishes,
> >
> > James
> >
> >
> >> If there
> >> are technical reasons for bad image quality, you should improve your
> >> setup and scan again. If there is head movement in some subject that
> >> severely distorts your data, you should scan another subject. If head
> >> movement is due to your experimental design, you should check whether
> >> the design, experiment setup and subject instructions can be improved.
> >>
> >> The only acceptable reason for "repairing" bad data would be that your
> >> individual subject has very special abilities that can not be found in
> >> any other subject. There is a tutorial for ArtifactRepair at
> >>
> >> http://cibsr.stanford.edu/tools/ArtRepair/ArtRepair.htm
> >>
> >> which discusses the available options and their pros and cons.
> >>
> >> Volkmar
> >>
> >>
> >> Am Mittwoch, den 17.12.2008, 01:22 +0000 schrieb Christopher Benjamin:
> >>
> >>>
> >>> Hi,
> >>>
> >>> I didn't receive a response to this post; apologies for re-posting if
> >>> this is quite basic but I'm having trouble finding a practical
> discussion of
> >>> these issues. Any help would be invaluable,
> >>>
> >>> With thanks,
> >>>
> >>> Chris
> >>>
> >>> ---------------------------
> >>>
> >>> Hi Spmmers,
> >>>
> >>> I'm removing artefact from event-related fMRI data using Sue Whitfield
> >>> Gabrieli's automated artifact detection toolbox
> >>> (http://web.mit.edu/swg/software.htm).
> >>>
> >>> I'm wondering if people are using any standard criteria for removal of
> >>> images in terms of -
> >>>
> >>> 1. Image signal variation (i.e., deviation of signal from the series
> >>> mean);
> >>> 2. Participant movement, and
> >>> 3. Participant rotation.
> >>>
> >>> In terms of criteria I've seen, 1. often seems to be used to exclude
> >>> images that deviate from the mean by more than two or three standard
> >>> deviations; (which I think makes sense);
> >>> 2. to exclude images in which movement beyond .5 or 1mm occurred, and
> for
> >>> 3. exclusion of iamges where rotation of more than .05 radians
> occurred.
> >>>
> >>> I realise there are unlikely to be rigid criteria, especially given the
> >>> limitations inherent in scanning some populations, but I'd greatly
> >>> appreciate anyone point me to a stand-out reference/discussion...
> >>>
> >>> With thanks
> >>>
> >>> Christopher
> >>>
> >>>
> >
> >
> > --
> >
> > --------------------------------------
> > Dr James Rowe
> > Senior Clinical Research Associate and
> > Consultant Neurologist, Cambridge University Department of Clinical
> > Neurosciences,
> > Box 83, R3 Neurosciences,
> > Addenbrooke's Hospital, Cambridge, CB2 2QQ
> > UK
> >
> > Tel: +44 (0)1223 273630
> >
>
> --
> Paul K. Mazaika, PhD.
> Center for Interdisciplinary Brain Sciences Research
> Stanford University School of Medicine
> Office: (650)724-6646 Cell: (650)799-8319
>
> CONFIDENTIALITY NOTICE: Information contained in this message and any
> attachments is intended only for the addressee(s). If you believe
> that you have received this message in error, please notify the
> sender immediately by return electronic mail, and please delete it
> without further review, disclosure, or copying.
>
--
Benjamin Yerys, PhD
Instructor
Center for Autism Spectrum Disorders
Children's Research Institute - Neuroscience
Children's National Medical Center
111 Michigan Ave, NW
Washington DC, 20010
202-476-5358 (office)
301-765-5425 (lab)
--
Paul K. Mazaika, PhD.
Center for Interdisciplinary Brain Sciences Research
Stanford University School of Medicine
Office: (650)724-6646 Cell: (650)799-8319
CONFIDENTIALITY NOTICE: Information contained in this message and any
attachments is intended only for the addressee(s). If you believe
that you have received this message in error, please notify the
sender immediately by return electronic mail, and please delete it
without further review, disclosure, or copying.
|