Resolution: Won't Fix
Affects Version/s: None
Fix Version/s: None
Component/s: API (Simulator - Framework)
The projection matrices provide a mapping from activity on nodes (sources) to a
set of sensors. The OpenMEEG package used for this can, in principal, generate
mappings to EEG, MEG, and "internal" sensors.
- A single, externally computed, ProjectionMarix for mapping from our current
default surface (reg13) to a single set of 62 EEG channels exists.
- A rough draft of python code facilitating access to OpenMEEG exists – This
will most likely need to be refactored for easier UI integration, possibly
in the form of one or more "ProjectionMatrix" datatypes.
- Right now we have importers for Projection Matrices EEG (Surface and region);
- At least one, preferably multiple, complete datasets to serve as a default
dataset available to users who can't or aren't interested in providing their
own. Of specific importance here are the surfaces defining head geometry.
- Mechanisms for UI access will need to be created, currently the solution
being considered will be similar with the DTI Pipeline – That is, to have
all required software installed on the cluster, and from every TVB instance
access that machine(s), schedule a computation task there, then check
periodically if the results are ready.
- Runtime on a highend workstation, circa 2010 (comparable to a cluster node):
custom ATLAS: ~3 hours (45min head + 1h15 invert + 1h source matrices)
stock ATLAS: ~8 hours
- Most computation time is in the Source+Head geometry and conductances, once
these parts are computed ProjectionMatrices for multiple Sensors can be
quickly computed for the same source/head.
- Storage requirement if intermediate data is saved, based on our current
default data, is of order 10GB.
- On interaction with us and our requirements, bug fixes were made to the
OpenMEEG project to enable it to work with our data. TVB's externals
directory contains a snapshot from their repo including these fixes, however,
once the fixes are pushed to their release version we should move to that, as
their releases are available as debian packages this should hopefully help
facilitate cluster installation.
- Although OpenMEEG can generate forward solutions for independent dipole
sources (ie, directly from regions), it will generally be better to map
region timeseries to a surface and then perform the projection to MEEG, etc.