Hi Maurice
I am not sure what is going on.
To summarize:
So you take random source locations, simulate and reconstruct and all is OK.
But when you take new random source locations it doesn’t work.
Could there be an error here (like not overwriting the original SPM file you use for the stats) ?
A couple of checks-
Firstly I would check what is happening at a single subject level (maybe just using one source to make things simpler).
Remember there may be lots of deeper sources which are difficult to resolve even in simulated data.
See if you can pinpoint whether your problem lies within subjects or at the group level.
Best
Gareth
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of Maurice Goldi
Sent: 12 May 2015 09:36
To: [log in to unmask]
Subject: Re: [SPM] EEG source localization unrealistic activation found
Hi Gareth
Thanks for your suggestion.
I had indeed not looked at the contrast, expecting there should be no activation if I just look at the trials where there is no activity at that frequency.
I have done some more testing and unfortunately this has not solved all my source localisation problems.
I have generated 20 subjects with 112 trials ranging from -300 to 2000ms each. These trials contain 2 conditions.
One is a 6Hz condition: 6Hz sinusoid from 0-1000ms plus random noise across the whole 2.3 seconds.
The 15Hz condition: 15 Hz sinusoid from 0-1000ms plus random noise across the whole 2.3 seconds.
Activations are localized in the same place
I do group inversion with the standard inversion setting.
As a window of interest i choose 0-1000ms and 15Hz.
when i contrast the 15Hz vs 6Hz condition i localize the correct source.
All other contrasts give me no source.
This is what I would expect.
If I however test further and do exactly the same setup, but with the 6Hz and 15Hz source not at the same place i find no activation at all for any contrast.
Not expected.
Expanding on this I tried with even more sources and different frequencies.
What I tend to find is a correct localization if the contrast has an activation at the same location PLUS 'ghost' localizations where ever there is activity at other frequencies.
Not expected.
Any suggestions what is going on?
How to test further?
Cheers
Maurice
---
University of Zürich
Maurice Göldi
Department of Psychology
Biopsychology
Binzmühlestr. 14 / Box 5
CH - 8050 Zürich
Tel. +41 (0)44 635 74 55
www.psychologie.uzh.ch
[log in to unmask]
-----"SPM (Statistical Parametric Mapping)" <[log in to unmask]> schrieb: -----
An: [log in to unmask]
Von: "Barnes, Gareth"
Gesendet von: "SPM (Statistical Parametric Mapping)"
Datum: 28.04.2015 11:02
Betreff: Re: [SPM] EEG source localization unrealistic activation found
Hi Maurice
Sounds interesting.
So everything is fine when you look within 300 ms and 20Hz.
This means I guess that you base the inversion computation (i.e. the weights mapping sensors to sources) on just the first 300ms and maybe 0-40Hz ?
The key thing is that you should try the statistical contrast (can you specify what you mean by activation) based on this same window.
Hopefully if you take your 300ms period when there is no signal and do both computations within the same window there shouldn't be much there.
Can you confirm if this is the case ?.
Best
Gareth
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of Maurice Göldi
Sent: 27 April 2015 18:36
To: [log in to unmask]
Subject: [SPM] EEG source localization unrealistic activation found
Hi SPM experts
I am trying to do source localization of my EEG data. Because I am finding unexpected activation I have created synthetic data (using the code provided in the suplementary material in Lopez et al; Algorithmic procedures for Bayesian MEG/EEG source reconstruction in SPM; 2014).
What I expect:
1.) Source reconstruction for the synthetic data at the generated frequency and in the time window where it was generated.
2.) No activation (or minimal leakage due to filtering) in time windows where there was no activity generated.
3.) No activation (or minimal leakage due to filtering) in frequency windows where there was no activity generated.
What I get:
1.) Nice source reconstruction.
2.) Activation in time windows where there should be no activation at all. This activation looks just like the activation that I generated.
3.) Activation for frequency windows where I did not generate any data. Again the sources found are similar to what I generated.
I have generated many different combinations of sources, frequencies and time windows.
E.g. I have one 20Hz source (no additional noise added) from 0 to 300 ms.
1.) I find the source when I window 0-300ms and at 20 Hz.
2.) I find nearly the same pattern when I window from 700 to 800ms and 20Hz. (the time window should be far enough away to have minimal leakage)
3.) I again find nearly the same pattern when I have a window at 0-300ms and 6Hz.
The magnitude of activity in the 3 different result is always about the same.
I have tried different sources, time and frequency windows but always with the same problem.
The coregistration seems fine and also the lead field seems ok as far as I can tell.
I am using SPM12.
I have also generated datasets with multiple trials and multiple conditions. unfortunately these problems do not average out.
My questions:
Does anybody else have this problem?
Are my expectations wrong?
What could be the problem?
My guess at the moment is that it might be the temporal projector having an unwanted effect.
In spm_eeg_inv_result.m on line 98, TTW is computed.
In contrast to W it shows activity across the whole epoch (when I window in 700 -800ms for example). I dont understand the theory well enough to definately say if this a problem or not. Perhaps some of you may have an idea on this?
Thanks in advance for any input.
Cheers
Maurice
|