Dear SPM list

Problem in short:

We have encountered a curious problem regarding our MEG source localisation procedure.
In short, the inverted model returned consists only of NaN's when we run group inversion based on the gradiometers. However, it works fine for the magnetometers. See general information and detailed description of the problem below:

General information about the experiment:

14 subjects
Scanned with Elekta equipment (magnetometers and gradiometers)
ICA applied, eye blink and heart rate components have been removed.
306 channels, 326 samples (250 Hz), and 6 conditions.
SPM-version used: SPM8, revision number 4667 in MATLAB 2011
Preprocessing: bandpass filtered 0.5 Hz - 15 Hz, epoched (-200 ms - 1100 ms), robust averaging on each condition, and finally bandpass filter again to remove artefacts induced by robust averaging. Conditions have been sorted such that D.condlist for each subject is '0', '1', '2', '3', '4', '5'.
Individual MR-scans obtained for each subject and co-registered to the SPM-template.
New fiducials registered and saved for each subject based on co-registered MR-image.


Observations concerning the problem:
Source localisation runs as expected, for all subjects, when only magnetometers are used. This is not the case for gradiometers. According to our preliminary tests, NaN's are returned whenever we include more than 8 subjects in our group inversion. It seems to be independent of whether or not ICA has been applied or not. Furthermore, we have tested whether it could  be a single subject that caused NaN's to be returned, but it doesn't seem like it.

NaN's returned            Subjects included in source localisation after having excluded ICA-dependency as the problem
Yes                                 1-14  subj > 8
Yes                                 1-9    subj > 8
Yes                                 2-10  subj > 8
No                                  1-8    subj = 8
No                                  9-14 subj < 8
No                                  2-9   subj = 8

This shows that each subject can be included in a group inversion without NaN's being returned. It seems, however, that NaN's are returned whenever there are more than 8 subjects in the inversion, which we find a bit curious. Our tests cannot exclude that the problem stems from a combination of (a) certain subject(s) being included in the group inversion and/or the number of subjects being greater than 8.

What we further have investigated:

We have investigated the SPM gain matrices and have observed no obvious correlations between successful and unsuccessful inversions (as measured by whether or not NaN's occur).

We have investigated whether any subject time series contain curious values (e.g. NaN's, inf, NA), and have found none.

We have downloaded the newest version from the SPM-website, revision number 4667, and tried the analysis again, where we removed all NIfTI and GIfTI images from earlier analyses, but with the same result.

We are open to any suggestions you might have.

Thanks in advance

Michael Nygaard Pedersen and Lau Møller Andersen

P.S. Please make sure that you send answers to the emails of both of us!