Hi All,
I just wanted to make something clarified if you could help please. I wonder what is the difference between extracting your sources using constrained source map directly or using unconstrained source map first and then convert [Nchannels, 3 * Nsources] gain matrix to a constrained map using bst_gain_orient function?
Is this a way to combine 3 differently oriented and positioned dipoles in a one single position and orientation for only the sake of decreasing the computational costs or would it still provide better and smoother estimation as it would be obtained using unconstrained model directly?
I would really appreciate any comments on those.
Thank you
Lisa
Hi Lisa,
If you compute the MNE source activity with unconstrained dipoles orientations and then project the 3 values (x,y,z) at each point on the normal to the cortex at this point, it would produce the same results as using the option “constrained orientation: normal to cortex”.
You can use the computational costs of your analysis by using constrained dipoles orientations, but you are are not using a very accurate anatomical model (subject MRI + precise sensor positions) it very difficult to justify this this approach.
You can compare this approach with a downsampled cortex surface (5000 vertices) and unconstrained sources (15000 dipoles).
Cheers,
Francois
Hi Francois,
thank you for the answer, now it is clear. May I ask one last very simple question please? In this entry http://neuroimage.usc.edu/forums/showthread.php?2316-ImagingKernel-Size-Question&highlight=bst_gain_orient you mentioned doing the computation first and then averaging the result.
In the case of my both source data A and S is three dimensional, then the coherence will be calculated between the matching coordinates isn’t it. I mean,
Cx = coherence (Ax, Sx), Cy = coherence (Ay, Sy), Cz = coherence (Az, Sz), and C = mean (Cx, Cy, Cz).
Sorry for asking that much easy detail, but I started to became very suspicious about every step after doing a lot of paper, forum and tutorial reading so I really appreciate your confirmation.
Thank you very much in advance,
Lisa
No, it’s not happening the way you describe it.
If you have two grid points with three orientations each, A(Ax,Ay,Az) and B(Bx,By,Bz), the function bst_connectivity.m is going to compute first all the pairs of connectivity measures [B][3x3] matrix[/B], then take the [B]maximum[/B].
If A=B, then the result is always 1.
It’s not necessarily the best approach, maybe it will change in the future.
It’s a good thing to be critical with these tools, they are still under development and poorly documented.
Let us know if you think this should be done in a different way.
Cheers,
Francois
Hi Francois,
thank you for the explanation, it was a bit different solution than I expected. May I ask what is the reasoning behind having the maximum out of this adjacency matrix to represent connectivity? Is that mean each dipole stands as it own during whole estimations of connectivity?
Edit: May I also ask, isn’t each source represent a vector in the frequency domain after Hilbert transformation? Why we cannot apply vector operations on those three dipoles in the frequency domain to obtain a single representation of a source?
Thank you Francois,
Lisa
Hi Lisa,
Yes, the estimation of the connectivity measure is done for each dipole (= each orientation) separately.
There is no clear justification behind this approach, this is partly why these tools are not documented yet.
I haven’t had any input yet from our collaborators specialized in the estimation of connectivity measures.
If you can find any publication indicating that this should be done in a different way, you’re welcome to share it with us.
In the time-frequency decomposition processes (Hilbert transform and Morlet wavelets), we sum the power of the signals along the three directions.
Two values of signal power in the same frequency bands can be summed, but coherence or PLV values cannot…
Cheers,
Francois
Hi Francois,
thank you for the explanations. Actually I am reviewing many source localisation papers, but I haven’t came across the one explaining or using unconstrained method with connectivity analysis yet.
May I ask which function in the toolbox is used for unconstrained source localisation, is it bst_gain_orient.m? I mean I am trying to understand how does brainstorm extract three orthogonal dipoles really and what are do three unconstrained dipoles represent really. I mean I assume they are not simply projection of a single dipole to the Cartesian axes isn’t it?
Thank you,
Lisa
Hi Lisa,
The forward models are computed with bst_headmodeler.m and many subfunctions, always for 3 orthogonal dipoles, following the three axis (x,y,z).
At each position of our source grid, we place three dipoles, with respective orientations (1,0,0), (0,1,0) and (0,0,1).
The function bst_gain_orient projects each set of three leadfield columns (x,y,z) onto a single column corresponding to the normal orientation. It restricts the source space in the head model only to dipoles oriented normally to the cortical surface.
Most of this is described in the tutorials:
http://neuroimage.usc.edu/brainstorm/Tutorials/HeadModel
http://neuroimage.usc.edu/brainstorm/Tutorials/SourceEstimation
The orientation constraint is applied during the source estimation process.
In the old version (Compute sources): Call to bst_gain_orient() in file bst_wmne.m at line 351
In the new version (Compute sources 2016): Lines 534-536 in file bst_inverse_linear_2016.m
Francois
[QUOTE=Francois;10236]
…
If you can find any publication indicating that this should be done in a different way, you’re welcome to share it with us.
…
Cheers,
Francois[/QUOTE]
Hi Francois,
have had a chance to look at the paper of Mehrkanoon et al(2014), in which they use PCA to extract the dominant orientation among the 3 orientations. The reference paper of Hipp et al (2012) gives more detailed explanation to it really. I am curious what would be your recommendations on that.
Thank you
Lisa
Hi Lisa,
We’ll look at these papers, thanks for the reference.
Francois