[FieldTrip] mne vs lcmv methods

Wed Sep 28 17:38:35 CEST 2011

Dear Beth,

I think we'd need to know more about your analysis input to answer questions about the dimensionality of the source outputs, but in general, the lcmv beamformer suffers from the same problems as other beamformers - that is, when you have the expectation of highly correlated sources, one or both of the sources will be suppressed, particularly when the time range for the data input to the beamformer is restricted to short intervals. Many papers have been written on this issue, and there are some attempts at solutions (e.g., see Brookes et al. 2006 Neuroimage). I do not believe that the lcmv or dics beamformers in FieldTrip have any such implementation for highly correlated sources, but I may be wrong. Since you want to reconstruct the sources of an auditory evoked field, in general a beamfomer will be a poor choice of methods for you. For simple, passive auditory stimulation, two-dipole models work very nicely. If you have the expectation of more complicated source arrangements because of some cognitive manipulation, then the cortically constrained MNE approach seems reasonable to me.

All that said, the lcmv beamformer in Fieldtrip should be able to give you output at the same times as you input for analysis. The dics beamformer, as it is done in the frequency domain, not the time domain, will generally give you a single output. Also, to clarify something that might be misinterpreted from Lilla's answer to you, at some level (the forward solution) all of these models (single dipole, multiple dipole, cortically constrained mne, beamformer) are all reliant on the dipole model. Distributed models simply have many dipoles, which are fixed in location to either the vertices of the cortical surface tesselation, oriented normal to the surface, or in the case of a beamformer, typically done on a regularly spaced grid with either fixed or optimized source orientation (e.g., lcmv vs. sam).

Best,

Don

On Sep 28, 2011, at 5:08 AM, Magyari, Lilla wrote:

> Hi Beth,
> 
> I do not know how lcmv beamforming is implemented in FieldTrip, and I 
> haven't used it before. But I did use the mne method for 
> sourcelocalization in FieldTrip, so may I can answer some of your 
> questions partially. But if someone is more expert in this area, 
> comments are appreciated.
> 
> Belluscio, Beth (NIH/NINDS) [E] wrote:
>> Hi fellow Fieldtripers-
>> 
>>  I am learning about the pros & cons of different methods for source 
>> localization.  For some of my data sets, I want to localize evoked 
>> activity (ERFs).  In one of them, I anticipated a single unilateral 
>> primary site of activation and MNE worked very nicely for this.  In the 
>> second, I am using auditory stimuli and therefore anticipated 
>> bilaterally synchronous activity.  As was to be expected, (I think) MNE 
>> gave a single site of activity near the midline.  I tried a parallel 
>> approach, but specifying lcmv as the method for ft_sourceanalysis as 
>> follows:
>> 
>> 
>> 
>> cfg = [];
>> 
>> cfg.method = 'lcmv';
>> 
>> cfg.grid = leadfield_02_stim1;
>> 
>> cfg.vol = vol;
>> 
>> cfg.lambda = 1e8;
>> 
>> cfg.checksize = inf;   %I’ve been inserting this and the next line 
>> because I’ve gotten error messages about the size.  This didn’t happen 
>> before, but these lines avoid the problem
>> 
>> cfg.checkconfig ='loose';
>> 
>> source3CB24_02_stim1 = ft_sourceanalysis(cfg, tlkCB24_02_stim1)
>> 
>> 
>> 
>> The output has a time dimension 1x601, and a pos dimension of 8196x3 
>> dimension, but the avg.pow is only 1x8196.  I interpret this to mean 
>> that the analysis averaged the power over some time interval instead of 
>> computing the power in each leadfield position for each time point.  
>> When I use the mne method, the output is clearly power for each time 
>> point, and produces a beautiful movie, as indicated in the tutorial. 
>> 
>> 
>> 
>> My questions about the lcmv technique are:
>> 
>> 1.       Is there a fundamental difference between the lcmv approach and 
>> the mne approach whereby it cannot compute the power at each time point, 
>> but only gives a single average value for a time period?
> 
> As far as I know, the mne approach (compared to other methods, like 
> beamforming) is favored for calculating source estimates of evoked 
> activity over time. Therefore, I am not surprised that the lcmv is not 
> doing the computation over time.
> I found this a useful reading about the differences between beamforming 
> and mne:
> 
> Jensen, O., Hesse, C., 2010, Estimating distributed representation of 
> evoked responses and oscillatory brain activity, In: MEG: An 
> Introduction to Methods, ed. by Hansen, P., Kringelbach, M., Salmelin, 
> R., doi:10.1093/acprof:oso/9780195307238.001.0001
> 
>> 
>> 2.       If so, how do I specify the time interval of interest (ie that 
>> containing the main peak of the ERF) in ft_sourceanalysis?
> 
> I would use ft_redefinetrial. This function can select only a segment of 
>  your trials.
> 
>> 3.       I thought one advantage of either MNE or LCMV was that I could 
>> indicate my a priori hypothesis of the anticipated localization of the 
>> source activity.  Do I do this with cfg.grid?  Is there a way to 
>> indicate the number of anticipated dipoles?
> 
> A dipole model presupposes indeed that the activation is coming from a 
> small area or from multiple separate small areas, while the mne and the 
> beamforming approach are providing a distributed representation of the 
> neural activity. The minimum "a priori" assumption is that the 
> underlying activation is coming from the brain. The brain volume is 
> discretized onto a 3-dimensional grid (this is the cfg.grid). And the 
> current is estimated at these grid locations. So, in this sense, yes, 
> these grid points are your dipoles. Typically, you specify  more 
> thousands of these points.
> 
> I hope this helps.
> 
> Lilla Magyari
> 
> 
>> 
>> 
>> Thanks in advance for your help-
>> 
>> Beth.
>> 
>> 
>> 
>> Beth Belluscio MD-PhD
>> 
>> Clinical Fellow
>> 
>> Human Motor Control Section
>> 
>> NINDS, NIH
>> 
>> 301-402-3495
>> 
>> 
>> 
>> 
>> ------------------------------------------------------------------------
>> 
>> _______________________________________________
>> fieldtrip mailing list
>> fieldtrip at donders.ru.nl
>> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
> 
> -- 
> PhD student
> Language and Cognition Group
> research assistant
> Neurobiology of Language Group
> 
> Max Planck Institute for Psycholinguistics
> Nijmegen, P.O. Box 310, 6500AH, the Netherlands
> Phone: 0031 24 3521561
> 
> _______________________________________________
> fieldtrip mailing list
> fieldtrip at donders.ru.nl
> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip

-----------------------
Don Rojas, Ph.D.
Associate Professor of Psychiatry
U. of Colorado Denver Anschutz Medical Campus
Director, UCD Magnetoencephalography Lab
13001 E. 17th Pl F546
Aurora, CO 80045
303-724-4994