A Flexible Framework for Non-Invasive Source Localization in Pediatric Focal Epilepsy Tiferet Levine-Gazit Medical Vision Group CSAIL, MIT

1
A Flexible Framework for Non-InvasiveSource
Localization in Pediatric Focal EpilepsyTiferet
Levine-GazitMedical Vision GroupCSAIL, MIT
2
Background

• Focal Epilepsy
• Epilepsy affects 1 of the population under the
  age of 20
• Seizures prevent healthy development and may
  cause brain damage
• Focal epilepsy is triggered by pathological
  electrical activity in a small clump of neurons
• Current Treatments
• 35 of focal epilepsy patients do not respond to
  medication, and must undergo surgical resection
  of the epileptic focal points
• Surgery requires accurate localization of the
  foci
• Subdural EEG is currently needed to reconstruct
  focal sources from voltage measurements, and even
  with it only about 65 of patients are
  seizure-free after surgery

3
Project Overview

• Project Goals
• More accurate source localization to improve
  post-surgical prognosis
• Source localization based on non-invasive tests
  and scans instead of subdural EEG
• Contributions
• Flexible, modular framework for non-invasive
  source localization based on scalp EEG
• Incorporates prior information from MRI and other
  sources
• Utilizes state-of-the-art patient specific head
  modeling
• Allows easy switching of prior map, field-solver,
  or inverse method
• Tools for processing raw patient data for use
  with sophisticated field-solving packages

4
Source Localization

• The Source Localization Problem
• EEG gives us voltage readings at electrodes on
  the scalp
• From the quasi-static Maxwell equations we get
  the state equation relating source currents in
  the head to voltages on the scalp
• ? ( ? ?V ) ? JP
• Forward problem Find voltages at electrodes
  given source configuration
• Inverse problem Find source configuration given
  voltages at electrodes
• The inverse problem is highly ill-posed and must
  in practice be solved through iterative solution
  of the forward problem to search for a current
  configuration that best explains the voltages
  measured

5
Modeling the Head

• Head Models
• In order to solve the forward problem we need a
  model of the head as a volume conductor
• Current clinical practice uses simple multishell
  spherical head models
• BEM allows realistic modeling of scalp and skull,
  but not different brain tissues
• FEM allows realistic modeling of scalp, skull,
  CSF, GM, and WM

• Creating a Head Model
• We use NeuroFEM as the field-solving package to
  solve the forward problem on a high-resolution
  FEM head model
• First we segment whole-head MRI using modified
  watershed segmentation
• Then we mesh the head volume, assigning a
  conductivity to each mesh node based on its
  tissue classification

6
Modeling the Electrodes

• Clinical Electrode Placement
• There are several ways electrodes may be placed
  in an EEG acquisition
• The most common method is the 10-20 system,
  requiring manual placement of 19-32 electrodes
  based on anatomical landmarks and relative
  distances
• Better methods include dense electrode nets with
  standard locations, electrodes with MR-visible
  markers, and electrode locations recorded with a
  3D tracker

7
Modeling the Electrodes

• Aligning Electrodes to Our Head Models
• In order to use NeuroFEM, we need to give it the
  location of each electrode in the reference frame
  of the head model
• To do this, we first build a very simple model of
  the patient head surface in Slicer
• If the 10-20 system or electrode markers were
  used, we then place fiducials in Slicer at each
  electrode location, transform the coordinates
  from Slicer coordinates to NeuroFEM coordinates,
  and output the electrode file
• If a tracker or standard net locations were used,
  we place fiducials at the four reference points
  on the Slicer model, and from these four points
  we find the matrix to transform the given
  acquisition coordinates to the Slicer reference
  frame. We use this matrix to transform all
  electrode coordinates. We then transform the
  electrodes from Slicer into NeuroFEM and output
  the electrode file.

8
Solving the Inverse Problem

• Flexible Framework for Source Localization
• In order to solve the EEG inverse problem, one
  must optimize the dipole configuration through
  repeated forward simulations
• We have implemented and tested several
  optimization techniques
• Exhaustive search over the entire brain or a
  specified ROI (good for clinical source
  localization)
• Simultaneous Perturbation Stochastic
  Approximation (SPSA) for very efficient
  stochastic optimization (good for research when
  many localizations must be carried out quickly)
• NeuroFEMs built-in Simplex optimization (useful
  only if no prior map is to be used)
• Other inverse methods may very easily be
  implemented in this framework, and the framework
  may be used to compare different methods in a
  research setting

9
Prior Information

• Defining Prior Probabilities for Focal-Point
  Locations
• There are many methods in the literature for
  locating focal-point hot-spots and assigning
  prior probabilities on dipole locations
• Any such method may be used to define a prior
  probability map for use within our framework
• Automatic methods of assigning prior
  probabilities may look at factors such as
• Anatomical considerations for where focal points
  are likely to lie
• Machine-vision techniques for locating specific
  anomalies such as cysts or lesions
• Asymmetry measures between the two hemispheres,
  measures of GM/WM blurring, volumetric measures
  of various structures, measures of cortical
  thickening, etc.
• Our framework is very useful in testing and
  comparing different prior formulations

10
Incorporating Prior Information

• The Error Function
• Any inverse method optimizes an error function
• We incorporate the prior information we wish to
  use through a prior probability map that is used
  as a weighted additive Bayesian prior term in the
  error function

Slice from a patient prior probability map
11
Experiments and Results

• Sanity Checks
• To verify that a given head model and inverse
  method works well, the first step is always a
  sanity check localizing a known simulated dipole
• We build a head and electrode model, place a
  dipole with known configuration, and solve the
  forward problem to obtain simulated voltages for
  this dipole
• We then solve the inverse problem with these
  voltages on the same head model, to make sure we
  get back the dipole we started with
• We ran such sanity checks on three different head
  models spherical, isotropic FEM, and
  anisotropic FEM using various inverse methods
  such as NeuroFEM Simplex, SPSA, and exhaustive
  search, in some cases using both deep and surface
  dipoles.
• All our sanity checks returned the dipoles we had
  initially place to very high degree of accuracy

12
Experiments and Results Robustness

• Tests of Robustness to Noise
• Real data always contains noise. It is important
  to test how various models and methods are
  affected by this noise.
• To obtain noisy data with known parameters we
  again use simulated EEG, but this time
  artificially add noise in the various inputs used
  by the source localization software
• Noise in voltages Add zero-mean Gaussian noise
  to the voltage at each electrode, simulating
  various levels of noise by adjusting the STD of
  the Gaussian
• Errors in electrode locations Add zero-mean
  Gaussian noise to the three location components
  of each electrode, simulating various levels of
  location uncertainty by adjusting the STD of the
  Gaussian

13
Experiments and Results Robustness

• Robustness to Noise in Voltages
• Two head models spherical and anisotropic FEM.
  SPSA optimization.
• At least ten different localizations for each
  noise level on each head model

14
Experiments and Results Robustness

• Robustness to Errors in Electrode Locations
• Two head models spherical and anisotropic FEM.
  SPSA optimization.
• At least fifteen different localizations for each
  noise level on each head model

• These results indicate the benefits of using
  accurate electrode placement methods such as
  electrode nets or tracking devices in clinical
  EEG acquisitions

15
Experiments and Results Prior Probabilities

• Experiments Incorporating a Prior Probability Map
• Isotropic FEM head model from real patient data
  with focal cortical dysplasia (FCD) (figure 1)
• Electrode net aligned to head model from standard
  locations (figure 2)
• Prior probability map based on the patients
  anatomy (figure 3)
• Hot-spots found through asymmetry analysis of MRI
  data
• Probabilities assigned according to anatomical
  considerations based on voxel tissue
  classifications
• Simulated dipole within the anomalous region
• Noise added to voltages, electrode locations, and
  tissue conductivities

16
Experiments and Results Prior Probabilities
Summary of Prior Probability Experiments
Noise level Low Medium High
Voltage noise STD (microVolts) 5x10-4 (SNR 22dB) 1x10-3 (SNR 16dB) 5x10-3 (SNR 2dB)
Electrode errors STD (mm) 3 4 5
Conductivity errors 5 7.5 10
NeuroFEM Simplex localization Worked well enough Significant errors Very large errors
Localization error (mm) 2 4.9 14.6
Other localization N/A Exhaustive search with ROI determined by NeuroFEM localization results Exhaustive search with broad ROI determined by MR-visible anomaly
Localization error (mm) N/A 0.5 0.5
Total time for source localization (minutes) 10 Under 40 Under 50
17
Conclusion

• Future Work
• Spatio-temporal preprocessing of EEG to separate
  out signal from multiple dipoles
• Anisotropic conductivity tensors in our head
  models
• Inclusion of more inverse methods, prior
  formulations, and field solvers within our
  source-localization framework
• Development of a more automated pipeline with a
  nice user interface
• More rigorous testing of different models and
  methods, together with clinical trials and
  surgical validation

18
Thank you!