thresholding Z-stats images is quite a complex topic, I suggest reading
through some previous discussions on the list re different practices, a
good starting point probably is an email Joe Devlin sent earlier this
On 22 Dec 2004, at 00:04, Fornito, Alexander wrote:
> Thanks. The references help.
> Although I can (for the most part) understand the principles, I have a
> little difficulty understanding the practical implications and
> implementation in FEAT.
> Is there any recommended Z-threshold and p value, or p value for a
> given z threshold?
Not really, it depends on the expected size, the expected amplitude
modulation etc. The Feat default values are sensible across a wide
range of studies.
> Also, I've been playing around with different thresholds on my images
> and have noticed that there is little difference between the
> activation maps produced when I use an uncorrected p < .01 and a
> cluster threshold of z = 2.3 p < .05. The only difference is the
> appearance of some very small additional 'blobs' in the uncorrected
> analysis. Do you know why this would be the case?
I guess cluster thresholding is more conservative. In GRF the Z-images
are thresholded and for each post-threshold 'cluster' a probability is
calculated which depends on the cluster size but also the smoothness of
the image. If the 'blobs' are very small it probably means that they
have quite a low probability given the smoothness.
> I also get some contrasts that show no activation whatsoever. This is
> strange since these are my main contrasts and the task is based on one
> that has already been reported in the literature. This only happens
> for some individuals (I'm just doing single-subject analyses atm), and
> I was wandering whether this was just due to normal variation, or the
> symptom of something more sinister that I'm not aware of?
Maybe, have you tried running melodic to see if there are any nasty
'structured noise' effects in the data? Also, loading the data into
fslview and playing around with the time series display is not a bad
> On contrasts (and in subjects) where there is activation, the output
> only seems to report results for very large (>100 voxels) clusters,
> when I can see that there are also smaller clusters. Is there anyway
> to get the z values and coordinates of the smaller clusters?
> I've noticed that the z-map images (thresh_zstat) have the image
> properties of the 4D, rather than the template. Is this because the
> analysis is done in native space before being registered to the
Yes, stats is done in the native data space - the processing order
corresponds to what you see in the GUI from left to right.
> Despite this, although my 4D images are 16 bit, the thresh_zstat
> images are 32 bit. Why does this happen?
All analysis is done in floats/doubles - it does not make sense to use
the native image type for this (e.g. in your case the data probably is
16bit integers but I guess you would not be very happy if all Z-scores
were integer, too)
> Finally, do you forsee any problems if I convert my images, which are
> acquired at 16 bit, to 8 bit before running my analyses? It would save
> disk space and promote efficiency.
I recommend you don't - even if Feat will not crash (which I should
not, but then again I have not tried this) I don't think this is a good
idea. FMRI data is expensive (in many ways) and I cannot see any reason
why one would want to mess with the dynamic data range. This is very
likely to change the results of the analysis.
> Sorry about all the questions - I'm new to fMRI!
> Many many many thanks,
> -----Original Message-----
> From: Mark Jenkinson [mailto:[log in to unmask]]
> Sent: Tuesday, December 21, 2004 8:06 PM
> To: [log in to unmask]
> Subject: Re: [FSL] AW: [FSL] Timeseries in FSL view
> You might find our brief technical report and the references
> in that (particularly Kiebel et al 1999) to be useful in
> explaining exactly what GRF does. The report is at:
> Matthew Brett also has an excellent web page explaining GRF
> and cluster-based inference at:
> Hope these help.
> All the best,
> On 21 Dec 2004, at 00:07, Fornito, Alexander wrote:
>> That worked -thanks!
>> You're answer raised another issue for me. I'm still trying to
>> understand how thresholding is achieved in FEAT. Let me see if I
>> For voxel-based thresholding, is each voxel treated independently?
>> Is the corrected p-value set by the user corrected according to GRF
>> theory (eg., if I set the value to .05, the threshold is corresponds
>> to a GRF corrected .05. In this case, is there any way of determining
>> what the actual threshold is?
>> For cluster-based thresholding, are the voxels smoothed according to
>> the FWHM set in the pre-stats tab of the FEAT GUI? Thus, for eg., if
>> the FWHM is 5mm then the stats are run on clusters of 5 or more
>> contiguous voxels?
>> What do the Z values and p values represent in cluster based
>> thresholding, and are there any rules of thumb for determining what
>> the appropriate values for these are?
>> Once again, many thanks for your assistance,
>> -----Original Message-----
>> From: Andreas Bartsch
>> [mailto:[log in to unmask]]
>> Sent: Monday, December 20, 2004 8:01 PM
>> To: [log in to unmask]
>> Subject: [FSL] AW: [FSL] Timeseries in FSL view
>> Hi Alex,
>> ok - this one is easy:
>> Load the 4D-file (e.g., filtered_func_data) into FSLView. The selected
>> file appears in the lower box on the left. Once you have loaded the
>> stats image (e.g., thresh_zstat1) you need to select the 4D-file in
>> the box again to view the time series. I assume you had the stats
>> image selected (which is 3D), so you could not access the time series
>> of the 4D file.
>> And yes - all clusters surviving the threshold are listed. But note
>> that are always some stipulations involved in defining what is
>> regarded as a cluster (which depend on the precise approach chosen).
>> To define contiguous clusters, FEAT applies an initial Z-threshold you
>> Best regards-
>> -----Ursprüngliche Nachricht-----
>> Von: Alex Fornito [mailto:[log in to unmask]]
>> Gesendet: Mo 20.12.2004 03:39
>> An: [log in to unmask]
>> Betreff: [FSL] Timeseries in FSL view
>> I'm having trouble accessing timeseries data in FSL view.
>> I load the 4D image, and add the thresh_zstat image for the
>> contrast of
>> interest. The stat image gets overlaid on the 4D image, but I
>> am uable o
>> select "Timeseries" from the "View" menu. Am I missing
>> Also, as an aside, in the report.html page, do the cluster
>> lists obtaine
>> by clicking on a contrast image represent an exhaustive list
>> of all
>> clusters that syurvived thresholding?
>> Many thanks for your help,
Christian F. Beckmann
Oxford University Centre for Functional
Magnetic Resonance Imaging of the Brain,
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK
Email: [log in to unmask] - http://www.fmrib.ox.ac.uk/~beckmann/
Phone: +44(0)1865 222782 Fax: +44(0)1865 222717