The cluster in question consists of 20 voxels that seem to be just above
the cluster forming threshold (why p<0.005?) and the threshold on
cluster extent is set to 20 voxels so you don't need much variation in
the numerical computations to see this cluster appearing or disappearing
of the results table. Slightly different data due to different
preprocessing steps or different computing platforms might be enough.
One way to automatically reorient images is to use this script:
On 06/06/18 00:13, Heard, Matt wrote:
> Hello SPM Community,
> I wanted to share something I found when working with my data today, and
> to ask the community for advice. I've been replicating an analysis I ran
> earlier this year using a Linux server (same versions of SPM and Matlab
> 2017a). One particular subject caused me trouble. When set at
> uncorrected P < 0.005 and cluster size threshold k = 20, my old analysis
> featured an entire cluster that was absent in the newer analysis, as
> well as some small differences in peak/cluster statistics and difference
> in number of voxels included.
> Tracing back my steps, I determined that I did not set the origin of the
> anatomical image to the anterior commisure/pitch the brain to align the
> axial slice with the posterior commisure. I ran the analysis again using
> the origin-set anatomical image and the cluster disappeared. I also
> noticed that when run on my local (Windows) computer, the analysis
> included the same number of voxels. The Linux server, however, used one
> fewer voxel. I've attached a screenshot of the three results windows,
> with added notes and emphasis.
> I've been advised (and seen the effects of doing so) to set the AC-PC
> line before coregistering and normalizing when some subjects are not
> processing correctly
> The same stands for using skull-stripped anatomical images, which is
> standard practice by some functional connectivity toolboxes
> My questions are as follows:
> 1) Is it commonly accepted to set the AC-PC in the anatomical images
> before coregistration and normalization?
> 2) How about using skull stripped images?
> 3) Are there any toolboxes that set the AC-PC line automatically?
> 4) Do these differences phase out at a more stringent threshold, such as
> P < 0.001, and across subjects?
> 5) What about the difference in number of voxels? Is this reason for
> 6) As long as all of the subjects undergo the same pipeline, do these
> steps even matter?
> Matthew Heard
> Laboratory Manager
> *The Ohio State University*
> Arts and Sciences Speech and Hearing Science
> 004 Pressey Hall, 1070 Carmack Road, Columbus, OH 43201
> 614-247-8320 Office / 214-458-7255 Mobile
> [log in to unmask] <mailto:[log in to unmask]> https://osu-slam.github.io/
> /Buckeyes consider the environment before printing./
> To unsubscribe from the SPM list, click the following link:
Guillaume Flandin, PhD
Wellcome Centre for Human Neuroimaging
University College London
12 Queen Square
London WC1N 3BG