Query over medial wall activity

✉️ Mailing List Archive (read-only)
1

Hi MNE-ers

I am working with auditory data, running my analysis on source estimations
reconstructed from MEG and EEG sensor recordings.

My analysis takes the form of pattern matching over the estimated activity
of each of the vertices in a source space, and as such, is reliant on the
reconstruction being of good quality. I am very pleased with the quality
of the results using MNE - my pattern matching technique should locate
those vertices along Heshl's Gyrus, and indeed it does - an indication,
presumably, of the high quality of the reconstruction. (so a big thank you
to everybody involved with constructing and maintaining MNE!)

However, I did want to ask this mailing list about one concern: my pattern
matching technique also picks up vertices directly 'under' HG - on the
medial wall in the 'unknown' label of the Destrieux Atlas
(aparc.a2009s.annot) (see figure 1 attached). It seems pretty clear why:
the inverse solutions given by MNE give both these regions similar evoked
responses (figure 2 of the attached), which is why my pattern matching
technique flags both areas up. While it is possible that these results may
be correct (the auditory thalamus is in this area, and so might plausibly
causing this medial activity) I wanted to poll this mailing list to get a
feel for how likely you think this activity is being correctly estimated
here, or if you feel it is a simple case of mislocalisation from the
auditory cortex (and if so, whether it can be fixed). I'm not really sure
what my grounds for suspicion are, except that the affected vertices on
the medial wall are directly under HG - implying the HG source activity
might be 'seeping' through to these more medial sources during
reconstruction.

I have observed this phenomenon in two independent experiments. And
although I can't do my pattern matching on the MNE example 'audvis' data,
this too seems to show the same phenomenon (figure 3).

I have tried pretty much every flag and option MNE offers - depth on/off,
sLORETA vs. MNE vs. DSPM, different SNRs, pick_normal on/off, different
looseness's - all end up with pretty much identical results (which is
good, I guess, as it means the reconstruction is pretty robust).

I appreciate that for many people this isn't an issue if they are doing
analysis only in predetermined regions of interest (I can't imagine that
many people are looking for results in a label called 'unknown'). But as
my analysis works by searching vertex-by-vertex, I want to say truthfully
that I looked through all vertices the reconstruction gave back, or at
least give a reason why I excluded vertices in the `unknown' label from my
analysis.

Anyway, I don't know if it is a common occurrence, or is something I have
done wrong (although the fact that we see the 'audvis' data behave in the
same way is evidence against this). Or maybe you think it is correct - a
number of my co-authors have suggested we take it as correct, and say it
is evidence of a cortico-Thalamic loop.

I attach some figures that demonstrate the phenomenon.

Thanks in advance for any thoughts.

Andy

-------------- next part --------------
A non-text attachment was scrubbed...
Name: For-MNE-mailinglist.jpg
Type: image/jpeg
Size: 187373 bytes
Desc: For-MNE-mailinglist.jpg
Url : http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20130206/6fe60832/attachment.jpg

2

Generally, I am dubious about the likelihood of there being sufficient
pyramidal dendrites in these regions to provide the necessary
architecture for generating signal. I have considered removing these
locations from my forward solution. I am also curious about others'
thoughts.

D

3

Although it's reasonable that the majority of detectable MEG sources are due to population post synaptic currents, it also seems reasonable that transient longitudinal currents in axon bundles due to synchronized volleys of action potentials could also produce detectable fields. Somatosensory evoked responses from the hand typically produce a sharp response in the MEG recordings at about 25 msec post-stimulus (n20) and an opposite going response a few msec later. This latter response usually localizes well using equivalent current dipole localization to the primary somatosensory cortex with the current's direction pointed posterior consistent with the orientation of the cortex. Additional waves over the next 20 msec or so typically also localize close by and with the current direction nearly exactly parallel to that of the first p wave, suggesting that they too are cortical.

But the n20 wave, while it also localizes nearby, shows a non-parallel orientation. Now this wave is known to derive from thalamo-cortical projections. It is preserved in humans in electrical recordings in the face of profound cortical ischemia due to carotid cross-clamp. So I speculate that at least in that case, the n20 is due to passage of volleys of action potentials through axons perhaps terminating in the somatosensory cortex. And given the timing, nearby localization, and the orientation of the equivalent current solution, it seems to originate from the bundles where they curve to enter the cortex.

Perhaps you are seeing something comparable in the auditory system.

Don

Don Krieger, Ph.D.
Department of Neurological Surgery
University of Pittsburgh

4

Hi MNEers,

Thanks for the interesting question and discussion.

About 5 years ago, I too encountered this problem of additional sources
(typically on the medial wall) being reconstructed by MNE, from a dataset
consisting of a simulated source in the primary auditory cortex. We since
resorted to calling them ghost sources. Since then, I have also
encountered sources in other lateral cortical regions 'ghosting' to the
medial wall, especially in the context of resting-state analysis.

I haven't seen a general treatment of this problem, except (very elegant)
case-by-case analysis based on neuroanatomical constraints (such as Donald
Krieger's discussion for S1).

One approach which I meant to try a long time ago, but never did, was to
generate an atlas of "ghost sources" for every simulated source on a
surface vertex, for the MNE parameters in question. Essentially this atlas
would be a point-spread function for each source location. Such an atlas
could either be used to compensate surface current estimates directly (by
incorporating them into some kind of Bayesian priors), or then at least
use alongside our MNE estimate to 'interpret with caution'.

With more groups getting interested in the precise spatiotemporal
localization of scene- and face-processing regions, I would imagine that
ghosting would be an important signal analysis issue to resolve. So thanks
again for the question. Look forward to others' views.

Kind regards,
Pavan

5

Hi Andy,

Best solution to find out, your activity on medial wall is real or not,
put a simulated source on Heshl's Gyrus of apropriate size, multiply it
with the forward operator and then add some empty room noise to it. Solve
again the inverse solution to find the spread.

This can be done easily in mne-matlab or mne-python.

Sheraz

6

Hi everyone,

Many thanks for all your thoughts so far. Sheraz - will test this and post
my results to this list, hopefully later today. As I understand it from his
email, Pavan's ghosting appeared when simulating a source in auditory cortex
- so, as a first step, let's see if I can replicate that. I'll try and look
at a small cross-section of lateral vertices.

Perhaps also worth mentioning, as I didn't put it in the original email - I
also did some MEG-only and EEG-only reconstructions to see which information
is contributing greatest to this (real or unreal) 'ghosting effect', and it
seems to be EEG. This doesn't seem to me to invalidate one view or another -
as I understand it, EEG is better at picking up deep sources, with some
groups using EEG to record brainstem responses (although this effect
requires thousands of trials to become reliable); on the other hand, 'EEG is
better at localising deep sources' is another way of saying 'MEG is worse at
localising deep sources', so if this effect is a mislocalisation to the
medial wall then perhaps it is not surprising that MEG is not contributing
as much to it.

Anyway, will report back later with results.

Andy

7

Hi everyone,

I have done the reconstruction for a simulation of Heshls Gyrus activity and
I end up with Figure 1 (attached), which confirms the findings of Pavan,
namely that sources simulated in Heshls Gyrus can end up (very strongly) on
characteristic parts of the medial wall [1]. I can confirm that the
positioning of this activity is identical to where I find my medial wall
matches. From the comments so far, this seems to be a wide spread
phenomenon.

As Don suggests, this doesn't necessarily mean that all medial wall activity
found is erroneous, but I feel it more than likely that, in my particular
case, it seems misleading to portray my medial wall matches as correct,
given that I have strong prior reasons from the literature to believe that
HG activity is correct.

As I see it, there are a couple of options open to those that find strong
spread from HG (or other areas) onto the medial wall. Please tell me if you
don't agree with these, or if you think there are other options :slight_smile:

1. Exclude 'unknown' and other medial regions from the final analysis,
saying that the reconstruction mislocalises 'real' activity to this area.
But this is a bit difficult - who is to say which area is prone to
mislocalisation and which isn't? The medial wall isn't the only place that
simulated activity mislocalises too - one could use the same logic to cut
out any regions one didn't want to see activity in. As I mention above, I
might be OK claiming this, but only because I have strong evidence from the
literature that tells me HG is correct (and thus that HG is mislocalising to
'unknown', and not the other way around).

2. Do what Dan suggests and exclude this area from the forward
solution. This would have the added advantage that it would presumably make
the resulting reconstruction more accurate. But one would need to be pretty
sure that the sensors aren't (ever) picking up medial wall activity (or at
least, only very weakly), because once it's gone, true medial activity is
then localised to the lateral surface. As Dan points out, perhaps we might
justify this by assuming there aren't sufficient pyramidal dendrites in this
region, but Don has cautioned that sensors might pick up medial transient
longitudinal currents. I would be very interested to hear any further views
on this.

3. Further priors of some sort, perhaps related to physiology of the
medial wall. But this would require some sort of agreement of number 2.

Hope his is helpful, thanks for all your thoughts so far,

Andy

[1] the 'no depth' option was present in this reconstruction.

8

hi everyone,

just a quick note. The SNR used in dSPM extends the spread. If you have
high SNR data you should reduce these "ghost sources". It would worth
performing a simulation to see what's a good SNR to avoid this. Also
now that mne-python has LCMV beamformer it'd be curious to see
how much this will happen too (it should at a certain level).

Cheers,
Alex

9

Hi Alex,

Thanks for this. Regarding SNR - I am working with single trials and I get
the best results when I specify an SNR of 1, so this could be exacerbating
the ghosting problem for me. However, the example plots I gave as mock ups
were of grand averages, where I specified an SNR of 3. I might do a test
where I make it higher value to see if this improves things.

With regards to the LCMV beamformer, I will perform this shortly, both on my
data and the example mne audvis_data, and post it to the list.

I will also try excluding the 'unknown' region from my forward model to see
what this result looks like (independent of whether any of us think it is a
good idea), and post the result to the list.

Cheers,

Andy

10

Hi all,

One factor I believe has not not been mentioned is that the correction of depth bias induced by converting MNE into dSPM is somewhat too aggressive and, therefore,
you are likely to get these false activations in the medial wall. Note also that the medial surface of the temporal lobe is not so terribly far from the auditory areas in the Sylvian fissure.

Best,
- Matti

11

Hi Everyone,

I use the low pass filtering build into mne_process_raw().
By default, it appears to use a window of 2048 points and a taper of 5 Hz.

I would like to make sure I understand what the filter is doing.
I assume that these mean the following:

2048 points at a time are run through an FFT.
A cosine^2 window is applied to the fourier coefficients beginning at a point where, close to the selected ? amplitude point, for the entire set of coefficients?
Or perhaps the fourier coefficients are tapered down to zero over a 5 hz interval?
I'm definitely hazy on this part.

The resulting "filtered" fourier coefficients are then run through an inverse fourier transform to replace the original time domain signal with a low pass filtered one.

What is the correct understanding of this?
And is there information about the transfer function of the filter?

For instance, what is the fall off per octave depending on the selected ? amplitude point and width of the taper?

Thanks for any help you can provide.

Regards,

Don

Don Krieger, Ph.D.
Department of Neurological Surgery
University of Pittsburgh

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20130312/8ba6558d/attachment.html

12

hi Don,

MNE uses what is called an overlap-add approach. Indeed fft is applied to
buffers of 2048 samples by default. The filter is applied in the
frequency domain
using a transition band of 5Hz. The benefit is that it is much faster
and you only
need to store a fraction of the data in memory. FYI the MNE-Python code
uses it too but also supports IIR filters.

HTH
Alex

13

Thanks for responding, Alex.

Please bear with a further question:
If the low-pass is set at 40 Hz, then the cos^2 drop off is applied beginning at the 40 hz components and falling to zero at 45 ? Or is it applied beginning at something like 38 hz so that the fourier coefficients are attenuated by a fact of 2 at 40 ?

And pardon this stupid question: Does this produce a gain of 0.0 for all higher frequencies in the filtered signal above 45? Or is there some kind of ringing which occurs?

Thanks again.

Don

Don Krieger, Ph.D.
Department of Neurological Surgery
University of Pittsburgh

14

hi Don,

Please bear with a further question:
If the low-pass is set at 40 Hz, then the cos^2 drop off is applied beginning at the 40 hz components and falling to zero at 45 ? Or is it applied beginning at something like 38 hz so that the fourier coefficients are attenuated by a fact of 2 at 40 ?

I am not 100% about the C code but I bet for 40 to 45 with 0 at 45 Hz

And pardon this stupid question: Does this produce a gain of 0.0 for all higher frequencies in the filtered signal above 45? Or is there some kind of ringing which occurs?

it does remove all frequencies about 45Hz up to some numerical errors
and if there is some ringing it will be in time. The less the stop
band (here 5) the more ringing in time domain.

Alex

15

Hi Don,

In the C code the default width of the lowpass transition is 5 Hz. For 40 Hz lowpass this means that the falloff starts at 37.5 Hz and ends at 42.5 Hz. You can adjust this value in the mne_browse_raw or with the --lowpassw option in mne_process_raw and mne_browse_raw.

The manual tells this in sections 4.5.1 and 4.2.1.

- Matti

16

Thanks very much.

Don

Don Krieger, Ph.D.
Department of Neurological Surgery
University of Pittsburgh
(412)648-9654 Office
(412)521-4431 Cell/Text

17

Hi Don, Alex and Matti,
    Not to introduce any more confusion but I have a clarification question:
Which of the following is true about the C-code?
(1) The filter is realized in the frequency domain fully (i.e) the FFT
coefficients for blocks of 2048 time samples are tapered to have a
transition band of 5 Hz around the cutoff and then going back to time. In
this case, the filter is a non-causal IIR filter with zero-group delay.

(2) An FIR filter is designed first and then then implemented in the
frequency domain using the overlap-add method with FFT blocks of 2048
points each. In this case, the filter is FIR and non-causal with
zero-group delay but there are (small) sidebands extending upto the
Nyquist rate. This would be like MATLAB's fftfilt().

Thanks,
Hari

18

Hari,

Please excuse the long silence. In the C code, method (1) is used, i.e.,
the filter is designed in the freq. domain with smooth transitions at
the corner frequencies (using a cosine). In MNE-Python on the other hand
method (2) is used. It can use both IIR or FIR filters, the FIR filters
are designed using

http://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.firwin2.html

the actual implementation uses an overlap-add FFT applied in forward and
backward direction to get zero phase (like filtfilt, as you mentioned).

I hope this helps,

Martin

19

Thanks! I should apologize for being too lazy to look at the C code.

Regards,
Hari

20

Thank you for continuing to try to address this question.

Please pardon my asking for clarification.

My question boils down to this: If I use the standalone mne_process_raw and use the standard filter settings of 40 Hz with 5 Hz dropoff, is everything above 42.5 Hz cut to zero?

Thanks,

Don

Don Krieger, Ph.D.
Department of Neurological Surgery
University of Pittsburgh
(412)648-9654 Office
(412)521-4431 Cell/Text