Hi everyone,
I just wanted to give you a quick feedback about source reconstruciton
methods in SPM12.
I've used the same subject and the same session (168*5 secs epoch) as a
test. I've run MSP (no random patches) -in SPM- and DICS + rPCA (Minka
truncation as regularization) -in DAiSS-. I'm using 63 EEG channels and
EO condition. Artifact rejected by ICA.
Overall the performances appear comparable: the wave morphology is very
similar. Also the drawbacks are the same: more or less the same epochs
get corrupted: e.g. some epochs show huge power at some sources and flat
EEG at others. It means that the corrupted epoch has nothing to do with
its previous and next one. Instead using spatial regularization (13
sources centered at a MNI coordinates) DICS -but also eLORETA, Minimum
Norm, IID, COH- don't show any unrealistic flat EEG.
What worries me at most are the flat EEG source: these could screw
completely the connectivity analysis since spurious connections are
expected.
Computational time: 1 Hr X 1 epoch by MSP; 8 Hrs X 1 epoch by DICS +
rPCA. No problem for me since I'm using clusters.
In conclusion, it seems that the bayesian approaches work suboptimally,
at least to my data.
All the best,
Paolo Ranzi
P.S. for Barnes: I'm trying to use MSP + 'Source inversion, iterative'
in order to see if I can solve the problem of the "flat EEG" sources.
Unfortunately, no clear guideline is present in the SPM manual about how
to set the parameters.
Combining info from different articles (e.g. Belardinelli et al. 2012,
PLOS ONE; Lopez et al. 2014, Neuroimage;) I've figured out empirically
the following parameters: Random patches= 100 (1Hr computation per 5
secs epoch); Number of iteration= 20; Patch smoothness= 0.6; Spatial
modes= 13 (I've 13 DMN sources); Temporal modes= 16. No priors file
provided. To me they look reasonable.
Further, within the 'Inversion function' the 'Classic' option doesn't
work, but only 'Custom' does.
Lastly, about the 'Inversion priors' module: not sure its utility, since
it should be better to compute the iterative source inversion for EACH
epoch independently. Am I right here? Obviously the computational time
is huge...
P.S. for Litvak: LCMV performs exactly like DICS. This is already known
from Belardinelli's papers.
|