Hi,
We are doing an empirical comparison between different GLM implementations
in three packages: our proprietory NPAIRS, FSL and SPM2 based on a simple
simulating image dataset. We sequencially fed the data into NPAIRS, FSL and
SPM2. We can see obvious activation blobs in tstat maps from NPAIRS and SPM2
as expected, while no obvious activation found in FSL results. The volume
correlation R between tstat maps of SPM2 and NPAIRS is 0.64, while the R
between FSL and NPAIRS is way below 0.1. We are wondering what is causing
such a different result?
The simulation data were generated using a 2-slice sub-volume(64*64*2)
extracted from a brain mask volume(64*64*32) from one subject for a 1.5T
fMRI experiment in which every volunteer was asked to perform two runs of
static force task alternating six rest and five force periods/run
(44s/period, TR=4s). Four artificial Gaussian blob(FWHM=1, 1.5, 2, 4
pixels) activations, each restricted to a 7*7 square, were added to
different locations in the second slice. To form the simulated time sequenc,
the blobs were then multiplied by the on-off reference function for two
parametric static force runs convolved with a Poisson shaped(lamda=7.3)
hrf. After adding white noise to the sequence and normalizing the CONTRAST
and CNR(CONTRAST-to NOISE Ratio) at the blobs' center to be 2, the
simulating data set was obtained.
Here is what we did in FSL:
1. avwmerge 120 slices into a 4d analyze image
2. input 4d data into FEAT, high pass filter cutoff = 100, TR=3.98
3. IF pre-stats was included, we used the default pre-stats setting except
that we changed spatial smoothing FWHM=0mm
4. in Stats part, we used full model setup, original EV = 1.
- Basic Shape: we prepared a text file as custom entry, the file looks like:
00000000111111111100000000001111111111....
- Convolution if used: Gamma, phase=0, stddev=3, mean lag=6
- contrast: mean, EV1=1
5. in Post-stats, Z threshold=2.3, p=0.01
We first ran the data in FEAT with pre-stat and post-stat turned on, we found
- no expected activation blobs found in the color rendered stat image in
report.html.
- from the time series for the voxel with max z, the full model fit doesn't
reflect the contrast = 2.
- there is no expected high intensity blobs in the unthresholded tstat1.img
or zstat1.img
(we used a self-developed idl program to view the t/z map, what the program
does is to map the t/z score, eg. [-3.9, 4.3] to [0..255] and display the
greyscale image. when we did comparison, the t/z maps from NPAIRS and SPM
were viewed through the same program as well.)
Then we tried again with pre-stat, post-stat, and convolution function
turned off, the results are similar as before, no activion blob seen.
Heather Luo
International Neuroimaging Consortium
Minneapolis VA Medical Center
|