[Eeglablist] ICLabel: paper online

Brian Roach brian.roach at ncire.org
Wed Oct 30 13:08:43 PDT 2019


Hi All,

I have an ICLabel paper question that might be of interest to the eeglab list.  There is a nice comparison of existing methods in Table 1, which includes two versions of the IC_MARC algorithm: SF for spatial features and EF for established features.  In the IC_MARC toolbox (v1.6) documentation for the ICMARC interface function takes an optional argument, feature_set, listed in bold below:


function [EEG,com] = pop_ICMARC_interface(EEG, varargin)
% pop_ICMARC_interface() offers a graphical user interface to receive
% inputs for ICMARC_interface() if all the inputs required by
% ICMARC_interface() are not given.
%
% Input:
%   EEG: Current dataset structure or structure array.
%
% Optional input, if both are not given, a window will pop up to ask for
% their values.
%   feature_set: (Optional) string containing the name of the the feature set to use
%   to classify ICs. Possible values are 'established_spatial_features',
%   'spatial2', and 'established_features'. Default:
%   'established_spatial_features'.
%   Possible values:
%   'established_features': Calculate the features in the established
%       feature set consisting of 14 features as described in the
%       accompanying paper.
%   'established_spatial_features': Calculate the features in the
%       established spatial feature set (described in the
%       accompanying paper.)
%   'spatial2': Calculate a set of spatial features without the
%   computationally demanding dipole features. This feature set was not
%   optimized systematically, but by some fiddling by hand.

Did you use the 'established_spatial_features' or 'spatial2' in your evaluation of the ICMARC_sf method in the ICLabel paper?  If you used established_spatial_features, then I am curious to know if there is a trick to reducing its runtime, as Figure 6 implies it takes 1 to 10s per component, but I'm finding dipole fitting takes much longer on what I think is a similar computer.

thanks,
Brian



On 6/20/19 1:34 PM, Scott Makeig wrote:

Luca Pion-Tonachini's successful defense of his dissertation recently
included presentation of his successful project to build a machine-learning
based EEG independent component (IC) classifier with several novel and
useful properties. A paper on ICLabel has now been published in NeuroImage
- follow this authors' link <https://authors.elsevier.com/c/1Z6AA3lc~r73Nv><https://authors.elsevier.com/c/1Z6AA3lc~r73Nv>
to download free copies.

The ICLabel website <https://labeling.ucsd.edu/tutorial/><https://labeling.ucsd.edu/tutorial/>
(*labeling.ucsd.edu/tutorial/
<http://labeling.ucsd.edu/tutorial/><http://labeling.ucsd.edu/tutorial/>*) continues to be a valuable resource
for learning to appreciate differences in spatial and temporal properties
of ICs accounting for brain, eye, muscle, line noise, and other effective
EEG sources. (Note: the site is quite usable for EEG course assignments).

The ICLabel website also continues to accept ('crowd-sourced') help from
users who wish to contribute labels for the still more than 100,000 ICs on
the SCCN servers that yet have no suggested labels... As new label
suggestions accumulate, the ICLabel algorithm will be regularly retrained
on the growing set of labeled ICs -- thus new label suggestions will
contribute to further improving the robustness of the ICLabel plug-in,
which is now already in use in many EEG labs and analysis pipelines around
the world.

The website also has a leading contributor 'leaderboard,' and your label
suggestions can receive a (private) 'Compared to Experts' estimate for your
information ...

Scott Makeig
for the EEGLAB team and Luca Pion-Tonachini

Scott





More information about the eeglablist mailing list