First we wrote:
We have unfortunately run into some more problems with our analysis in FSL.
Our study gives us 3 different runs from each subject. Each run lasts for
about 18 minutes, and contains 6 different modes (3 active, 2 baseline, and
pause). From this, we have made 16 different contrasts that should be more
or less interesting.
We have run all the first-level analysis without any problems. Higher lever
analysis runs OK with fixed-effects, but when trying any kind of ME, the
output is always corrupted; we (almost always) get this error:
“ndtri domain error”
Which appears in the log for the higher-level stats, almost in the
beginning. We have searched the online FSL-forum for help regarding this
error, but none of the suggestions there helped us. They suggested that the
first-level analysis could be corrupted, but everything there seems ok. It
also seems strange that the FE should run if data was corrupted.
To further investigate the problem, we have also tried to analyze the
lowest possible amount of data, which is one run from 2 different subjects,
and only one condition (Vanlig-Strek). Still, we get the same error (ndtri
domain error), and no output. So, we think we can exclude hardware as the
source of our problems. We have tried different computers, different
versions of Linux, and different versions of FSL (3.3 and 4.0). Nothing
seems to help.
Then Steve answered:
Most likely what's going on is that the ME maths is unhappy because you
have strong outliers in the data, but is doing it's best. Does the final
output look at all reasonable? Probably the best solution (without removing
the outliers that is) is to use just FLAME stage 1 - does that run ok?
However:
1. "Does the final output look at all reasonable?" - Yes, all of the
first-level outputs appear ok.
2. "Probably the best solution (without removing the outliers that is)
is to use just FLAME stage 1 - does that run ok?" - No, we are unable
to run any ME analysis on higher level. We have tried FLAME 1, FLAME 2
and even OLS. All of these failed, and produce the “ndtri domain error”.
3. "ME maths is probably unhappy because you have strong outliers in
the data" - We have even tried to analyse only 2 runs from two
different subjects, that both have almost the excact same first-level
activation. This we have tried numerous times on different subjects and
with all the different ME. Even this small amount of data makes FSL
crash. Based on this, we are pretty sure that the problem is not caused
by outliers.
4. And from a colleauge we got a suggestion that the problem might lie in
the DICOM to NIFTI conversion. We think we have ruled out that
possibility, since we have tried numerous different converters and
different scripts. All of these produce the same result, and in all
cases the first-level analysis looks resonable.
Anyone got any suggestion?
---------------------------------------------------------------------------
Hallvard Røe Evensmoen, masterstudent
Center for circulation and medical imaging
Norwegian university of science and technology
---------------------------------------------------------------------------
|