Hi Sebastien,

A bug causing this seems much less likely that just some
registration step not working. The method used here does
not seem so robust.

On thing that strikes me is that you are using different
volreg bases across runs, which is generally not advised.
While that might be a little thing, I could also imagine
it having a compounding effect that takes you off a cliff.

In any case, evaluate the alignment at the sub-steps, not
just using the final result. The alignment is surely
failing somewhere, and you just need to see where.


Look at the alignment between the 2 mean datasets,
trun01_mc_mean.nii.gz and trun10_mc_mean.nii.gz, to be
sure that the volreg steps seems to be succeeding.

Also, compare how blurry those images are. I suspect the
01 dataset might be much more blurry since you used a
blurry base image (the average from run 10). That may
lead to a compound blur in the 01 average due to weak
registration (meaning inconsistent registration across
volumes).

And if the 01 mean is very blurry, it might then fail
to register well to the deformation dataset.

Also, there is a chance that the distance between runs
1 and 10 is too great, and 3dvolreg might need options
like -twopass and/or -zpad 4 to succeed. You should be
able to tell that by comparing the blur averages, at
least probably.

You might also just scroll through time when viewing
the registered run 1 data. If it does not look smooth,
that is another flag.

Also, I see no reason to use both 3dvolreg to align run
01 to run 10 and 3dQwarp to align both means to the
deformation dataset. If you have already successfully
aligned the EPI data across runs, just apply the same
trun10_WARP, rather than computing a new one for run 01.
Doing it 10 times gives you 10 chances to fail.

- rick