sLORETA Poor Performance

unes1111 · August 30, 2018, 8:48pm

Hello,
I am analyzing the performance of the sLORETA under different scenarios.
It performed very well with one source even when increasing the Noise, However, With two different sources activate at the same time, (even with low noise) it is performing very poorly. Does anyone know what is the reason for that? Based on the other papers/journals that I have read, sLORETA should perform pretty good in estimating the sources (With only drawback of having the low resolution). Why is it missing a source in this case? is it a computational problem?
Current Density Map Can estimate properly but Sloreta CANT

Here is the specification of my experiment as well as attachments:
- I have used BESA SIMULATOR to simulate the data
- Both Sources are Exactly Same: This Means (SAME STARTING AND ENDING POINT, SAME SHAPE, SAME FREQUENCY, SAME AMPLITUDE, ) ONLY different locations
- Noise is 0.21 MicroVolt (Around 15-20dB SNR im guessing)
- No Pre-Processing , Only setting channel location + Average Referencing

Francois · August 31, 2018, 6:47pm

@John_Mosher @Sylvain?

rey · August 31, 2018, 11:16pm

It would be helpful to remove the threshold so that all activity can be seen.
Perhaps the activity it’s there just below the arbitrary threshold.
-Rey

unes1111 · September 2, 2018, 6:35pm

Hello Rey
Thank you for your suggestion, I was wondering if you know how we can remove the threshold?
thanks

Francois · September 4, 2018, 7:17am

@unes1111 Set the amplitude threshold slider to 0% in the Surface tab.

@rey Wow you’re still reading the posts on this forum, good to know!

rey · September 4, 2018, 10:07pm

@Francois Can’t claim I read every post but I’m still here to help if help is needed. Just ask.
@unes1111 The sLORETA should work pretty well, as it has zero localization bias, but if you want something better, something that gives you an actual current density estimate (i.e., not just some arbitrary unit “activity”) AND has also zero-localization bias, I highly recommend eLORETA for distributed non-sparse solutions. That could be added to BrainStorm but it requires an extra iterative optimization for each leadfield matrix and regularization parameter (for time-dependent SNR, probably just take the mean regularization parameter to avoid too much computation).

unes1111 · September 5, 2018, 3:07am

Thank you @rey and @Francois for your answers.
I will definitely check eLORETA and try to find the computation for it. Considering that for the inverse problem we need the lead field matrix as well as ways to show the result (on the surface) it is a bit difficult for me to go beyond what Brainstorm is offering, I easily get lost on when / where to add lead filed matrix how to demonstrate the result etc…
I am researching on the inverse problem as the topic of my Master Thesis in engineering. In order to be able to develop a new novel approach, I am first trying to understand where current novel solutions like sLORETA or eLoreta fail or even where sLORETA performs better than simple psedu inverse of Leadfield matrix (as Source =L+ M). Sofar based on the adjustment of thresholding that you just mentioned, both current density map and sLORETA are performing quite well. And this includes different High and Low SNRs, one or many sources. Surprisingly even simple psedu inverse problem is performing good and similar to mentioned methods. Through another discussion topic, Dr. @Sylvain told me that the simple inverse is non-unique and non-stable, but on the computational side, I have not seen any failure in of the inverse methods. Any suggestion on under what kind of condition I should see the failure of the current methods?

Note: I am using BESA Simulator to simulate my data since I still have not found the perfect (MATLAB function) that I can add in brainstorm and use that to simulate data (simple sine cosine is very very unrealistic).

As always thank you for the great discussion platform
With warm regards
Younes

Alexandre · September 6, 2018, 10:12am

@unes1111 chapter 3 in https://tel.archives-ouvertes.fr/tel-00880483/document is a good ref.

Also you have eLORETA implemented in Fieldtrip

github.com

fieldtrip/fieldtrip/blob/master/inverse/ft_eloreta.m

function [dipout] = ft_eloreta(dip, grad, headmodel, dat, Cf, varargin)
%
% Use as
%   [dipout] = ft_eloreta(dipin, grad, headmodel, dat, cov, varargin)
% where
%   dipin       is the input dipole model
%   grad        is the gradiometer definition
%   headmodel   is the volume conductor definition
%   dat         is the data matrix with the ERP or ERF
%   cov         is the data covariance or cross-spectral density matrix
% and
%   dipout      is the resulting dipole model with all details
%
% The input dipole model consists of
%   dipin.pos   positions for dipole, e.g. regular grid, Npositions x 3
%   dipin.mom   dipole orientation (optional), 3 x Npositions
%   dipin.filter precomputed filter cell-array of 1 x Npositions, each cell
%               containing nchannels x 3 matrix
%
% Additional options should be specified in key-value pairs and can be

This file has been truncated. show original

and MNE-Python

github.com

mne-tools/mne-python/blob/master/mne/minimum_norm/_eloreta.py

# Authors: Eric Larson <larson.eric.d@gmail.com>
#
# License: BSD (3-clause)

import numpy as np

from scipy import linalg

from ..defaults import _handle_default
from ..fixes import _safe_svd
from ..utils import warn, logger


# For the reference implementation of eLORETA (force_equal=False),
# 0 < loose <= 1 all produce solutions that are (more or less)
# the same as free orientation (loose=1) and quite different from
# loose=0 (fixed). If we do force_equal=True, we get a visibly smooth
# transition from 0->1. This is probably because this mode behaves more like
# sLORETA and dSPM in that it weights each orientation for a given source
# uniformly (which is not the case for the reference eLORETA implementation).

This file has been truncated. show original

you also have lots of methods which are non-linear such as (TF-)MxNE [Gramfort et al. 2012, Gramfort et al. 2013 and Strohmeier et al. 2016] or Champagne by the group of Sri Nagarajan.

Hope this helps
Alex

unes1111 · September 7, 2018, 4:55am

Hello Alex
Thanks for the links
Do you know if there is an easy way to use the constructed head and source model from Brainstorm in field trips or python?

My problem with the field trip is that once it gets to tutorials about source modeling its all for Linux or Mac so I am not able to use it in windows. I basically get stuck at Source model: Volumetric processing in FreeSurfer.

My issue is on how to use let the source /head model built on Brainstorm in field trips or etc.
Let say I want to try the python code you sent. Where are the inputs? is the anyway I can just put the Data (EEG) and liedfiled matrix (observed from brainstorm code) and plug it in the python code to perform inverse problem?

Francois · September 7, 2018, 9:53am

Do you know if there is an easy way to use the constructed head and source model from Brainstorm in field trips or python?

There are existing processes that call FieldTrip functions, like ft_sourceanalysis, which use Brainstorm functions that convert Brainstorm structures into FieldTrip structures. Look at the corresponding code for information on how to do this:

github.com

brainstorm-tools/brainstorm3/blob/master/toolbox/process/functions/process_ft_sourceanalysis.m#L131


iChannelsData = setdiff(iChannels, iBadChan);
% Error: All channels tagged as bad
if isempty(iChannelsData)
    bst_report('Error', sProcess, sInput, 'All the selected channels are tagged as bad.');
    return;
end
% Convert data file to FieldTrip format
ftData = out_fieldtrip_data(DataMat, ChannelMat, iChannelsData, 1);
% Add data covariance
ftData.cov = NoiseCovMat.NoiseCov(iChannelsData,iChannelsData);
% Convert head model to FieldTrip format
[ftHeadmodel, ftLeadfield] = out_fieldtrip_headmodel(HeadModelMat, ChannelMat, iChannelsData, 1);


% === CALLING FIELDTRIP FUNCTION ===
bst_progress('text', 'Calling FieldTrip function: ft_sourceanalysis...');
% Prepare FieldTrip cfg structure
cfg           = [];
cfg.method    = Method;
cfg.grid      = ftLeadfield;
cfg.headmodel = ftHeadmodel;
% Additional options for the method

We don’t provide anything similar for MNE-Python, but this is a goal for the next few months.

John_Mosher · October 29, 2019, 8:48pm

@unes1111 I would suggest two things (1) lower the Data Options Amplitude slider (under the Surface Tab) to zero, so that we can see the entire signal distribution, and (2) in addition to sLORETA, please also run dSPM, which should give a similar looking result.