Hi Freesurfers,
we have FreeSurfer running on 10 computers:
7 computers with suse 10 32bit, freesurfer 4.5.0 32bit 3 computers with debian 5 64bit, freesurfer 4.5.0 64bit
recon-all -all -i finished on all of them without any errors, so i assumed the best. out of curiousity i processed one of the subjects on both systems, and checking the results in tkmedit and tksurfer i noticed that they are visibly different. (thus in this case one of the surfaces does not look quite ok, other subjects processed on that system look fine however, i haven't looked through all of them yet). screenshots attached. also, i just started recon-all -all -s bert on both systems, so i can compare them to the original bert.
what could be the reason for this failure at reproducibility?
Hi Andreas,
what version are you running? The 32 and 64 bit versions will definitely have differences - it's unavoidable. In the newest version we did eliminate some of the randomness that reduced reproducibility though.
Bruce
On Sat, 13 Feb 2010, Andreas Berger wrote:
Hi Freesurfers,
we have FreeSurfer running on 10 computers:
7 computers with suse 10 32bit, freesurfer 4.5.0 32bit 3 computers with debian 5 64bit, freesurfer 4.5.0 64bit
recon-all -all -i finished on all of them without any errors, so i assumed the best. out of curiousity i processed one of the subjects on both systems, and checking the results in tkmedit and tksurfer i noticed that they are visibly different. (thus in this case one of the surfaces does not look quite ok, other subjects processed on that system look fine however, i haven't looked through all of them yet). screenshots attached. also, i just started recon-all -all -s bert on both systems, so i can compare them to the original bert.
what could be the reason for this failure at reproducibility?
Hi, thank you for the reply,
what version are you running? The 32 and 64 bit versions will definitely have differences - it's unavoidable.
On Debian: CentOS 4 x86_64 stable v4.5.0 On Suse: CentOS 4 stable v4.5.0
are the screenshots within the expected range of random difference or do i have to assume that one of them must be wrong? and, would you recommend using either only the 32 bit version or only the 64 bit version for all of our processing?
On Sat, 13 Feb 2010 10:26:36 -0500 (EST) Bruce Fischl fischl@nmr.mgh.harvard.edu wrote:
Hi Andreas,
what version are you running? The 32 and 64 bit versions will definitely have differences - it's unavoidable. In the newest version we did eliminate some of the randomness that reduced reproducibility though.
Bruce
On Sat, 13 Feb 2010, Andreas Berger wrote:
Hi Freesurfers,
we have FreeSurfer running on 10 computers:
7 computers with suse 10 32bit, freesurfer 4.5.0 32bit 3 computers with debian 5 64bit, freesurfer 4.5.0 64bit
recon-all -all -i finished on all of them without any errors, so i assumed the best. out of curiousity i processed one of the subjects on both systems, and checking the results in tkmedit and tksurfer i noticed that they are visibly different. (thus in this case one of the surfaces does not look quite ok, other subjects processed on that system look fine however, i haven't looked through all of them yet). screenshots attached. also, i just started recon-all -all -s bert on both systems, so i can compare them to the original bert.
what could be the reason for this failure at reproducibility?
oh sorry, I didn't see the versions in there. It's hard to tell from a single screenshow, but yes, you are always better off using the same hardware for your whole study
cheers, Bruce On Sat, 13 Feb 2010, Andreas Berger wrote:
Hi, thank you for the reply,
what version are you running? The 32 and 64 bit versions will definitely have differences - it's unavoidable.
On Debian: CentOS 4 x86_64 stable v4.5.0 On Suse: CentOS 4 stable v4.5.0
are the screenshots within the expected range of random difference or do i have to assume that one of them must be wrong? and, would you recommend using either only the 32 bit version or only the 64 bit version for all of our processing?
On Sat, 13 Feb 2010 10:26:36 -0500 (EST) Bruce Fischl fischl@nmr.mgh.harvard.edu wrote:
Hi Andreas,
what version are you running? The 32 and 64 bit versions will definitely have differences - it's unavoidable. In the newest version we did eliminate some of the randomness that reduced reproducibility though.
Bruce
On Sat, 13 Feb 2010, Andreas Berger wrote:
Hi Freesurfers,
we have FreeSurfer running on 10 computers:
7 computers with suse 10 32bit, freesurfer 4.5.0 32bit 3 computers with debian 5 64bit, freesurfer 4.5.0 64bit
recon-all -all -i finished on all of them without any errors, so i assumed the best. out of curiousity i processed one of the subjects on both systems, and checking the results in tkmedit and tksurfer i noticed that they are visibly different. (thus in this case one of the surfaces does not look quite ok, other subjects processed on that system look fine however, i haven't looked through all of them yet). screenshots attached. also, i just started recon-all -all -s bert on both systems, so i can compare them to the original bert.
what could be the reason for this failure at reproducibility?
freesurfer@nmr.mgh.harvard.edu
-
Andreas Berger -
Bruce Fischl