Greetings FSLers,
I'm having trouble recreating a step of our analysis pipeline across different versions of FSL. I'm hoping you might have some insight.
The goal is to denoise single run data by running FEAT using only nuisance regressors, and then taking the res4d file for subsequent analyses. In a previous dataset (and an older version of FSL: 5.0.1) we set up a GLM in the Stats tab that included Standard Motion Parameters, the path to a voxelwise confound list containing our PNM regressors, and a single empty EV of all zeros under the Full model setup window. We'd get the warning about the design matrix being rank deficient (not surprising with only a single empty EV specified), but the job would run without any additional trouble.
We have since upgraded to 5.0.8, and I'm getting a FATAL ERROR ENCOUNTERED each time I attempt this job. Additionally, there is an error in the log under Preprocessing:Stage 2 that reads:
/bin/rm -rf prefiltered_func_data*
Completed
while executing
"if { [ catch {
for { set argindex 1 } { $argindex < $argc } { incr argindex 1 } {
switch -- [ lindex $argv $argindex ] {
-I {
incr arginde..."
(file "/share/apps/fsl/bin/feat" line 310)
-stats
Completed
To troubleshoot, I tried adding some task EVs to the design matrix and was able to successfully complete the job. Since the error seems to relate to the empty design matrix, I tried selecting "Don't Add Motion Parameters" in the GUI and instead adding the 6 motion parameters as EVs in the Full Model Setup. This worked in that the job completed without any errors, however the res4d file that is created differs from the res4d file that was created using 5.0.1
from FSL 5.0.1:
non-zero voxel mean: 0.850
non-zero voxel stdev: 69.02
min/max: -2095/3244
from FSL 5.0.8:
non-zero voxel mean: 0.035
non-zero voxel stdev: 73.41
min/max: -2560/3510
I've confirmed that the design.fsf files between the two versions use the exact same inputs, and are as identical as possible (some options are new or different with 5.0.8, but they don't seem relevent to this problem). I also directly compared the design.mat files that appear in each .feat output dir (sorted them so the regressors would be in the same order). They are not identical, but differ by no more than 1E-06
So, two questions really:
1) What is the best way to run a GLM using only motion and physio regressors, now that FEAT doesn't seem to like a single empty EV in the Full Model Setup?
2) What could account for the different res4d files that result from what seem like identical models? Has something changed in how GLM fit is calculated between versions?
Any help or ideas is greatly appreciated.
Many thanks,
Jeff
|