Dear Freesurfer experts,
we are testing a grid computer infrastructure for running freesurfer analysis and are evaluating consistency and reliability of freesurfer results across platforms. Before we start our tests across platforms we conducted some freesurfer analysis on the same computer and had some unexpected results.
Subsequent executions of the freesurfer pipeline ("New invocation of recon-all") on a certain subject - without any manual editing in between - give different results. We have invoked recon-all up to four times and all results are different. We tested in particular the aseg.stats results, where we found different volume estimations (up to 4%) depending on the number of processing loops. Intermediate results have also been found to be different The effect is reproducible, meaning the nth invocation of recon-all on the subject gives always the nth result.
We suspect that the cause is (following the pipeline documentation: http://surfer.nmr.mgh.harvard.edu/fswiki/ReconAllDevTable) the twofold generation of "talairach_with_skull.lta", once in the autorecon-1 and a second time in the autorecon-2 phase:
autorecon1: mri_em_register -skull nu.mgz $FREESURFER_HOME/average/RB_all_withskull_2008-03-26.gca transforms/talairach_with_skull.lta
autorecon2: mri_em_register -skull -t transforms/talairach.lta nu_noneck.mgz $FREESURFER_HOME/average/RB_all_withskull_2008-03-26.gca transforms/talairach_with_skull.lta
and it seems that the step in autorecon1 is skipped if the file "talairach_with_skull.lta" is already present.
We were wondering if this behaviour is intended, so are the FreeSurfer results expected to improve with multiple invocations, even without manual intervention? Shall they converge after sufficient re-runs?
Any help or comments are highly appreciated.
Best, Michael