Thomas,

There is some variability in freesurfer results due to the use of a random number generator in some algorithms. This can be eliminated by adding the -norandomness flag to recon-all, which is what is done for our testing to ensure that recon-all produces the same results from build-to-build.

How that randomness manifests itself in the results produced by freesurfer varies. Cortical thickness, for instance, should not vary by more than 0.5%. But a complete analysis quantifying the variability of results across all possible measures has not been performed. For the measurements you mention, which file contained this measure? wmparc.stats? And which measure was it?

Per your question about runtime differences between computing platforms, can you send me the recon-all.log files from each run? I am interested to see where the time difference occurred. I would expect the faster platform to have finished first.

Nick

On Mon, 2007-10-22 at 14:17 -0300, Thomas Doering wrote:

Hi Pedro, thanks for your reply. I think, the MPRAGE data is ok with 1mm³ isotropic voxelsize aquired on a Trio system. The orientation is probably a problem because i just feed the images the way they come from the scanner direct into freesurfer. Even that freesurfer is not multithreaded, the higher processor 3GHz with higer RAM should run freesurfer faster than the lower 1.6GHz, isn ´t it? Cheers, Thomas ----- Original Message ----- From: Pedro Paulo Oliveira Jr To: 'Thomas Doering' Cc: freesurfer@nmr.mgh.harvard.edu Sent: Monday, October 22, 2007 10:31 AM Subject: RE: [Freesurfer] intrasubject segmentation variance

    Thomas,



    There are some comments on your question.



    Freesurfer is not multithreaded so it won’t benefit from
    multicore unless you start one analysis on each core.



    Your running times are high. My Opteron usually takes 23 hs to
    perform the recon-all in 3.0.5 version of FreeSurfer. One
    issue that can make the running time much higher is a poor
    SPGR/MPRAGE data. Maybe if you send one slice. Other issue
    that can raise the time is wrong volume orientation..



    Best Regards,



    Pedro Paulo Oliveira Jr



    From: freesurfer-bounces@nmr.mgh.harvard.edu
    [mailto:freesurfer-bounces@nmr.mgh.harvard.edu] On Behalf Of
    Thomas Doering
    Sent: segunda-feira, 22 de outubro de 2007 09:05
    To: freesurfer@nmr.mgh.harvard.edu
    Subject: [Freesurfer] intrasubject segmentation variance




    Dear freesurfers,


    i was running Freesurfer on the same subject on two
    independent Computers. 


    The first Computer PC1has 2 physical Intel processors with 2
    logical each - (dual core) - 1.6GHz with 4Gb Ram and the
    second PC2 has 2 physical Intel processors with 4 logical -
    (quad core) - 3 GHz with 8 Gb Ram.


    After running on these two Computers - i start at the exactly
    same time - i got a surprising result: My faster Computer PC2
    took 41h to run the subject and the slower Computer took 31h. 


    I wonder if somebody have some comment on this, because i have
    no idea what is wrong ( When i have the Linux System Monitor
    on, watching the CPU History, it seems that the slower PC hat
    a more effective occupation/usage than the faster - switching
    faster between one CPU and the other one...)


    Another small question: 


    In the final segmentation table, i noticed that there are some
    small diferences between the results of PC1 and PC2: for ex.:


    PC1:ctx - rh-caudalmiddle frontal = 5029 5029


    PC2:ctx -             -"-                     5360 5360


    , so a diference of ~6%. I guess these are variations due to
    the algorithm..


    Are there some established limits for that diferences?


    Thanks for help,





    Thomas 

---

Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer