Information theory in MEG: Physics Foundations for Source Space Encoding

•

Zachary Bednarke presented his work.

UW Neuroinformatics SIG 10/30/18

Zachary Bednarke Information theory in MEG: Physics Foundations for Source Space Encoding

What is MEG?

  • Non-invasive ST imaging of brain imaging (discretized mag field distribution, reconstruct underlying current)
  • Pre-surgical mapping, epilepsy
  • Neuromagnetic field=Post-synaptic potentials from neurons (primarily cortex), perpendicular to EEG
  • Used for resting state studies, connectivity analysis, neurofeedback, infant studies Getting MEG Ready for Traditional Neuro IT Goals
  • Combining physics with info theory, establishing set of basis functions ideal for processing multichannel magnetic field data.
  • Portable MEG: wearable source-space BCI, UW/Sandia Deep Brain Array Roots of Zach’s Work
  • Channel Capacity: specific ROI, max # of messages transmitted via sensor array?
  • Optimal Encoding: Achieve same DOF as from PCA but w/physical interpretation of importance hierarchy→ get to principled approach to encoding brain function Physics Foundations for Source Space Encoding
  • B(mag field vector space), J (current VS) field microstates: info capacity of Shannon
  • Lossless compression of magnetic field signal: Vector Spherical Harmonics basis in lieu of PCA Vector Spherical Harmonics
  • 3D vector of mag field coords fed into VSH fxns. Hierarchical in angular, radial complexity–lost if you decomposed w/fourier space.
  • Lowest number of model parameters that explains data = low-variance estimators for J
  • Transform to VSH space from sensor space is stable
  • Natural basis of special functions for solutions to Maxwell equations Total Information Extracted from MEG
  • Minimize bit error in transmission of message (inverse problem)
  • Calculate I(input, output) for multichannel MEG
  • Measure noise covariance (Probability Density Fxn), calc confidence interval of delta x
  • Divide input space volume by uncertainty volume = # mutually distinguishable messages Proof of Lossless Compression w/ 70 DOF Long term goals
  • Estimate bandwidth of brain ROIs for SS BCI
  • Sampling theory for VSH: Device design at Sandia (optimal placement)
  • We may not need MRIs for babies (only get L of 4 or 5)