source localizing frequency bands

Hello,

I have a generic question about source localization. From what I've
seen so far, when the goal is to localize the activity in certain
frequency bands, it's common to use methods such as DICS or the
variant of MCE for frequency bands. I was wondering if it is also
correct to do source localization using a different method (e.g. MNE)
with the whole signal (i.e. prior to band-pass filtering to specific
bands) and then convert the signal in the localized sources to
frequency domain. Another alternative would be to band-pass the signal
to a certain frequency band in source space, and then run a different
source localization method on it (e.g. MNE).

Could anyone elaborate on the advantages / disadvantages ( / validity)
of these two methods?

Thank you,

Gus

Just a few thoughts...

I'm not familiar enough to comment on the DICS method, but the MNE method returns a distribution based on the provided sensor waveforms. It would seem intuitive to me that performing a restrictive (say, only alpha activity) bandpass on the sensor waveform prior to performing MNE would significantly change the topology of the waveform, and thus significantly affect how the sources are localized.

The real question being asked is whether the steps are commutative; i.e., do the following two analysis streams produce identical results:
1) bandpass, MNE
2) MNE, bandpass
My gut intuition is that they are NOT, since the first method performs the bandpass on the raw MEG data, so to speak, while the second one performs the bandpass on reconstructed data. However, I can't prove this, and it bears further investigation. Anyone else have any thoughts?

Elli Kanal

Hi Gus and Elli,

Thanks for the discussion. I have been doing something similar.

If your bandpass filter is a FIR filter then I think bandpass/MNE are
commutative processes. This is because FIR filtering is a convolution
which can be implemented by the right-multiplication of a 'time x time'
BPF matrix constructed from the filter coefficients, with a 'channels x
time' raw data matrix [step 1]. The linear MNE inverse operator on the
other hand (literally) is implemented by the left-multiplication of a
'source pts x channels' matrix [step 2].

It is easy to see that steps 1 and 2 are interchangeable. However, I am
not completely sure about whether step 1 holds for FIR filtering, since
presumably Matlab does it row by row for each channel...

Pavan

Just a few thoughts...

I'm not familiar enough to comment on the DICS method, but the MNE

method

returns a distribution based on the provided sensor waveforms. It would

seem intuitive to me that performing a restrictive (say, only alpha
activity) bandpass on the sensor waveform prior to performing MNE would
significantly change the topology of the waveform, and thus
significantly

affect how the sources are localized.

The real question being asked is whether the steps are commutative;

i.e.,

do the following two analysis streams produce identical results: 1)

bandpass, MNE

2) MNE, bandpass
My gut intuition is that they are NOT, since the first method performs

the

bandpass on the raw MEG data, so to speak, while the second one performs

the bandpass on reconstructed data. However, I can't prove this, and it
bears further investigation. Anyone else have any thoughts?

Elli Kanal

Hello,
I have a generic question about source localization. From what I've

seen so far, when the goal is to localize the activity in certain
frequency bands, it's common to use methods such as DICS or the variant
of MCE for frequency bands. I was wondering if it is also correct to do
source localization using a different method (e.g. MNE) with the whole
signal (i.e. prior to band-pass filtering to specific bands) and then
convert the signal in the localized sources to

frequency domain. Another alternative would be to band-pass the signal

to a certain frequency band in source space, and then run a different
source localization method on it (e.g. MNE).

Could anyone elaborate on the advantages / disadvantages ( / validity)

of these two methods?

Hi Kanal,
   You are right, they are are not the same. The reason they are not the
same practically is that the noise covariance when generated after
bandpass filtering the signal first can be very very different (in
scale and possibly also in the spatial structure) and hence the inverse
operator is different.

I would vouch for computing one inverse operator with the entire band of
interest (0-100Hz for example) and then looking at narrower bands once the
data is in source space. My reason for that would be to not impose too
much temporal correlation in the noise by band-pass filtering in sensor
space. MNE as such is equivalent to a maximum aposteriori probability
(MAP) estimate of the source space activity only under the assumption that
the noise is uncorrelated in time. It is robust to violations of the
assumption to a large extent (correct me if I'm wrong here) but filtering
around 7Hz, for instance to look at alpha makes the noise heavily
different from its assumed no temporal correlation behaviour.

Having said that, if its an event related paradigm and you have enough
trials and excellent SNR, doing it either way should not make a
difference.

Regards,
Hari

Thank you for all the responses. Elli did get to the heart of the
question, and my gut feeling said the same (i.e. that they would not
result in the same thing). It's good to see that the following answers
seem to provide explanations for it.

However, if we abstract ourselves from the details of the math behind
it, shouldn't these processes be the same? For example, let's say we
recorded raw MEG signal with a strong 10Hz component to it.

1) If I run MCE/DICS in this 10Hz activity, I should get some source
activity that caused it.
2) If I band-pass the signal to 10Hz and run MNE on it, I will get
some source activity that caused it.
3) If I run MNE on the signal prior to any filtering, band-pass the
activity in all sources estimates to 10Hz, and focus on the sources
with highest power in 10Hz.

In all 3 cases, I'm looking for the sources that caused the 10Hz
activity I see in the MEG signal. Then, shouldn't the results of all 3
methods above agree to a certain degree?

Before I end this e-mail, I realize that "abstracting from the math"
is not a luxury we can have, since they're the heart of every source
localization technique. But maybe one technique or the other would be
more in accordance with the intuition above?

Thanks,

Gus