Introduction to ElectroEncephaloGraphy (EEG) and ... · Introduction to ElectroEncephaloGraphy (EEG) and MagnetoEncephaloGraphy (MEG). Outline zOrigin of EEG and MEG signals zEEG

Christophe Grova Ph.DBiomedical Engineering DptNeurology and Neurosurgery DptMontreal Neurological Institute, McGill, UniversityCentre de Recherches Mathématiques

christophe.grova@mcgill.ca

NEUR 570 Human Brain Imaging BIC Seminar 2011

Introduction toElectroEncephaloGraphy (EEG) and MagnetoEncephaloGraphy (MEG).

Outline

Origin of EEG and MEG signals

EEG and MEG data acquisition

Source localization

Outline

Origin of EEG and MEG signals

EEG and MEG data acquisition

Source localization

A bit of history …

1st EEG human recording: Dr. Hans Berger in 1925

1st MEG human recording: Dr. David Cohen in 1972

Several brain imaging techniques

Temporal scale

FDG PET

Spa

tial s

cale

(in

mm

)

8

6

4

2

10

Anatomical MRI

1 ms 1 s 40 s static

SimultaneousEEG/fMRI

Hemodynamic brain response associated to

EEG activity

Glucose metabolism

Intra-cerebralEEG

recordings

Reference

Scalp topographies

SPECT

Brain perfusion

EEG64 electrodes

MEG275 sensors

Epileptic spike

Electric vs magnetic fields

Moving electric charges, an electric current, create a magnetic field B

Electric vs magnetic signals in the brain

Magneto-EncephaloGraphy (MEG):measures changes in magnetic fields on the scalp: sensitive to neuronal currentsElectro-EnphaloGraphy (EEG): measures differences of electric potentials on the scalp: sensitive to conduction (volume) currents

If there was air between the brain and the skull, no EEG could be measured on this subject, but it would be possible to measure MEG

The electro-magnetic dipole model

r’

Jp(r’)

Current dipole

r Measurement point

Electric potential (in free space)

Magnetic field (in free space)

Main generators or electro-magnetic scalp activity: pyramidal cells

Generation of a signal one can detect from scalp measurements

Synchronisation of potentiels post-synaptic potentials along the cortical surface

Organization of pyramidal cells along the cortical surface

Neuronal conduction vs volume conduction

Neuronal conduction: (primary currents)

– Action potentials (1ms), – Post synaptic potentials (10ms)– Active conduction.– Origin of MEG signals

Volume conduction:(secondary currents)

– The brain is a conductive medium– Instantaneous propagation of

electric fields– Passive conduction. – Origin of EEG signalsBaillet et al 2001

Magneto-encephalography (MEG)

Measures magnetic fields generated outside the head from neuronal currentsMagnetic and electric fields are perpendicularThere is no influence of the skull on the propagation of magnetic fieldsRadial sources do not contribute to MEG signals

Differences between EEG and MEG

Although the same neurophysiological processes generate EEG and MEG:

1. Magnetic fields are not distorted by resistance from skull and scalp ⇒ better spatial resolution in MEG?

2. Electrical and magnetic fields are oriented perpendicularly to each other

Hamalainen et al 1993

Differences between EEG and MEG cont

3. Scalp EEG is sensitive to both tangential and radial components of a current source in a sphere, while neuromagnetometers detect only its tangential components

- a radial dipole does not produce a magnetic field outside the sphere

David Cohen NMH/MIT

Differences in EEG and MEG cont

- Thus, MEG selectively measures activity in the sulci.

- EEG measures activity from both the gyri and the sulci (dominated by radial sources)

David Cohen NMH/MIT

Epileptic activity: complementarity btw MEG and EEG

Merlet et al 1997

Outline

Origin of EEG and MEG signals

EEG and MEG data acquisition

Source localization

Electro-encephalography (EEG)

10/20 System (19electrodes)

Measurement of scalp Electric potentials

• Measures potentials generated by volume conduction currents (from 19 to 256 electrodes)• Scalp potentials are attenuated and distorted by the skull (highly resistive)

Electro-encephalography (EEG)

10-20 system 19 electrodes

Background Signal EEG

256 electrodes

EEG during wakefulness

From Dr. Gotman BMDE501 lecture

EEG during sleep

From Dr. Gotman BMDE501 lecture

EEG during an epileptic seizure

From Dr. Gotman BMDE501 lecture

MEG measures magnetic fields related to brain activityfrom femtoTesla (10-15 T) to picoTesla (10-12 T)

Earth’s magnetic field: 4,710-5 T.

small magnetic field measurements lead to artifacts

MEG brain signals = hearing the noise of a pin falling on a sofa in a dance club !

(M. Hamalainen)

Noise sources

Challenges in MEG data acquisition

Brain magnetic fields are tiny (pT,10-12T):– MRI: 1,000,000,000,000,000 (=1T)– Earth magnetic field: 100,000,000,000– Magneto-CardioGram (MCG): 100,000– MEG signals: 1,000– Sensitivity of magnetometer: 10

Highly sensitive sensors: SQUID (Superconducting QUantumInterference Device) in liquid helium (-269ºC)Noise reduction:

– Magnetically shielded room– First order gradiometer to eliminate remote interference and record

only local field– Reference sensors

MEG is sensitive to head motion: head localization is requiredMEG setup costs approximately $3M

MEG Systems

CTF – 275 sensors 4-D – 148 or 248 sensors,

– www.4dneuroimaging.comNeuroMag – 306 sensorsKIT - KanazawaOther (Los Alamos)

An highly sensitive sensor of magnetic field (fT)

SQUID: Superconducting Quantum Interference Devices

invented in 1965 by James Edward Zimmerman and Arnold Silver at Ford Research Lab.

All the system requires superconducting state: Liquide Helium + Cryogenic Dewar

Environmental Noise Reduction

Shielded Room Gradiometers Reference sensors used to pick up noise(3rd order gradient of env. noise)

Opening a MEG device

Reference sensors

Passive noise reduction:

s

Power lines orother current lines

Movingmagnetic dipoles

Reference system

Magnetically shielded room(mu metal)

Magnetic sensors are subjected not only to the measured MEG signal S, but also to unwanted signals (environmental noise, signals from parts of brain not being measured and other body parts).

Active noise cancellation: coils + external magnetometer to compensate external noise

MEGsystem

(a)

shieldedroom

(b)

Noise cancellation coils

Active noise cancellation. (a) coil system in an unshielded environment; (b) coil system combined with a shielded room.

Synthetic noise cancellation: flux transformers estimating local gradient of magnetic field

Unshielded, B

Powerline

Shielded, B

5 fT rms/√Hz

Vibrationalnoise

Frequency (Hz)1 fT

10 fT

100 fT

1 pT

10 pT

100 pT

1 nT

10 nT

100 nT

1 µT

10 µT

0.01 0.1 1 10 100

Magnetometers

3rd-order gradiometer

(b)

Noi

se (

B o

r G

(3) ·d

1·d 2

·d3)

(rm

s/√H

z)

Unshielded, G

Powerline

Vibrationalnoise

Frequency (Hz)1 fT

10 fT

100 fT

1 pT

10 pT

100 pT

1 nT

10 nT

100 nT

1 µT

10 µT

0.01 0.1 1 10 100

1st-order gradiometers

Shielded, G

5 fT rms/√Hz

(a)

3rd-order grad.Noi

se (

G(1

) ·d o

r G

(3) ·d

1·d 2

·d3)

(rm

s/√H

z)

MEG is sensitive to head motions during data acquisition

Continuous head localization system

Three emitting coils usedfor continuous head localization:Nasion and Peri-auricular points

3D localization of the three localization coils, the head shape and EEG electrodes on the subject’s head

Use of a magnetic device (Polhemus) for 3D localization

Co-registration with the subject’sanatomy (skin surface segmented from anatomical MRI)

MEG sensors (red) +EEG electrodes (blue) Co-registered on skinsurface

EEG electrodes (blue) + Digitized Head Shape (red)Co-registered on skinsurface

Few EEG/MEG examples

Epileptic activity : interictal spikes

Interictal spikes are spontaneous activity generated by the brain without any clinical sign

Multimodal exploration is feasible

Intra-cerebral EEG recordings showed that interictal spike generators are rarely focal (Merlet I. et al. Clin. Neurophys. 1999)

A minimum brain activated area of 6 cm2 is needed to generate a spike on the scalp (Ebersole J. Clin. Neurophys. 1997), spike generators may also be quite more extended than 6 cm2

A minimum brain activated area of 3 cm2 is needed to generate a spike on MEG

EEG interictal spike

Epileptic spikes in EEG

Epileptic spikes in MEG

Generation of evoked activity: averaging will increase the Signal-to-Noise ratio

Average of the 20 trials

Trial 1

Trial 2

Trial 3

….

Visual evoked field / potential

Visual evoked field / potential (left stimulation)

MEG topography EEG topography Source localization

Somatosensory vs motor evoked field (potential)

90ms after left thumb pneumatic stimulation

20ms after left thumb

tapping

Somatosensory vs motor evoked field (potential)

90ms after left thumb pneumatic stimulation

20ms after left thumb

tapping

Outline

Origin of EEG and MEG signals

EEG and MEG data acquisition

Source localization

Is it possible to localize source of brain activity from scalp measurements ?

Source localization

?

• Forward problem = modelling : knowing where are the sources, computation of the EEG/MEG signals generated by these sources

?

• Inverse problem: estimation of sources of brain activity from EEG/MEG scalp recordings

Forward problem

• Knowing brain sources and a model of the head, one can compute corresponding electric potentials or magnetic fields generated on the scalp

?

Spherical model: analytical solution

Realistic models of the head

a : spherical model (analytical solution)b : realistic surface model (BEM) → BrainStormc : realistic volume model (FEM)

The Direct Problem is Relatively Simple

“We know the geometry and the electrical and magnetic properties of the brain, CSF, skull and scalp”The geometry is complex: we need a mathematically manageable representation of the head. The simplest is a sphere; the more complex is a realistic head model.The electrical properties are complex: electrical conductivities cannot be measured in vivo. Bone conductivity is quite variable and is the most important factor.Magnetic properties are more homogeneous across tissues (less influence of the bone)

The Inverse Problem

Given a distribution of electric potentials or magnetic fields at the surface of a volume conductor, where are the sources inside the volume giving rise to this distribution, and what are their characteristics?

Inverse problem

The Inverse Problem is Hopeless

There is an infinite number of distribution of sources inside a volume conductor that can give rise to the samepotential distribution at the surface of the conductor (Helmholtz, 1853).There is an infinite number of possible arrangement of sources inside the brain giving rise to a particular EEG or MEG signalThe inverse problem is hopeless in theory. It can only be solved with simplifying assumptions = constraints

Selection of the more appropriate model

One need to add assumptions in the model to be able to find a unique solution

Are they realistic ?

Models of the sources of brain activity

Equivalent current dipolenon-linearnb of sources ?what is an ECD ?

Distributed sourcesanatomical constraint linearp = 103 sources n= 102 electrodes ill-conditioned pbneeds regularization

Summary: any source localization method relies on some a priori assumptions

Number of generators well-knownECD approaches

Few decorrelatedsources, number unknownDipole scanningapproaches

Distributed network and/or extended sourcesDistributed sources approaches

Model of signal generation

M

JLead field (forward pb)

signal

Noise

Model of signal generation

M

J sources

Lead field (forward pb)

signal

Noise

Model of signal generation

M

J sources

Lead field (forward pb)

signal

Noise

Model of signal generation

M

J sources

Lead field (forward pb)

signal

Noise

Model of signal generation

M

J sources

Lead field (forward pb)

signal

Noise

Estimate of J using and Equivalent Current Dipole

Model of signal generation

M

J sources

Lead field (forward pb)

signal

Noise

Estimate of J using and Equivalent Current Dipole

Estimate of J constrained on The cortical surface

The Equivalent Current Dipole

If one assumes that the EEG or MEG signals are generated by one or a small number of dipole sources, then it is possible to solve the inverse problem.A dipole is a point source, defined by its location (3 parameters), its orientation (2 parameters) and its moment (1 parameter).Is the dipole model reasonable?

Estimation of the Equivalent Current Dipole

Several types of dipole models:– Moving dipole: Position ?, Orientation ?, Amplitude ?– Rotating dipole: fixed position, Orientation ?, Amplitude ?– Fixed dipole: fixed position, fixed orientation, Amplitude ?

Number of dipoles:– Estimation of the signal subspace (PCA)– Iterative approaches

Solving the inverse problem:– Non-linear optimisation to find dipole locations– Minimisation of residual variance:

M = G(Φ) .J + E G: Lead field matrixΦ: dipoles location and orientationJ: amplutide of the the dipoleE: Noise

Minisation of residual variance (RV)

RV = || M - G(Φ) .J ||2 / || M || 2

Total varianceVariance non explained by the model

RV is not a validation metric !!!A misleading solution can provide a very low RV

Is the Dipole Model Reasonable?

In primary sensory evoked experiments, particularly somatosensory and auditory, the dipole model appears justified (it has been validated by comparing the results of modeling to what we know about sensory physiology).In other situations, particularly in epilepsy, the dipole model requires validation.

Can the Dipole Model be Misleading?

YESIn most instances, it is possible to find a dipolar source that can model well a scalp distribution. This only indicates that the dipole is a possible source of the distribution. It does not prove that it is the source of the distribution.There are systematic errors caused by the fact that the source is likely to have a significant spatial extent.

Confidence intervals

The Dipole Model: Conclusions

Very powerful method to find intracerebral sources from a scalp recording.Only valid if its underlying assumptions are correct (that sources are dipolar).Appears to localize well sources of primary sensory activity.Localizes reasonably well the maximum of extended sources of epileptic activity, although secondary (small amplitude) sources are probably less reliable.No information on extent of actual source.Main limitation: number of dipoles

Solving the inverse problem using distributed sources

BJGM +⋅=Inverse pb: LINEAR, but ill-posed.• Under-determined equation system:

102 measures 104 dipole sources (=unknown)

• Lead field G = ill-conditioned

Regularization is needed to find a solution(requires a priori assumptions)

Classical assumptions

Solution of minimum energy: Minimum NormHamalainen et al. Med Biol. Eng. Comput. 94

Solution maximum spatial smoothness: LORETAPascual-Marqui et al. Int. J. Psychophys. 94

Minimum Norm Estimate

LORETA: maximum of spatial smoothing

Anatomical constraints: sources distributed on the cortical surface

Dale A., Sereno M., 1993. J. Cogn. Neurosci. 5, 162– 176.

Extraction of a distributed sources model from an anatomical MRI

3D T1-weighted MRI acquisition: Matrix = 170x256x256, voxel = 1mmTR = 22 ms, TE = 9.2 ms

Automatic segmentation of the cortical surface = White Matter/Grey Matter interfaceBrainvisa: http://www.brainvisa.info

Regularization

Inverse problem= ill posed problem (10,000 sources vs 100 sensors)

No unique solutionThe problem needs regularization (assumptions)

1. Minimum norm (MN): minimum of energy (may be weighted W). Minimise ||M-GJ||2 + α ||WJ||2

2. LORETA : maximum of smoothness.

Minimise ||∆J||2 under the constraint M = GJwhere ∆ is a discrete spatial Laplacian operator

3. MEM : maximum entropy of the mean

Minimum Norm Estimate within the Bayesian Framework (1/5)

Linear distributed model for source localization

M: n x t signal on the n scalp sensors G: n x p forward model of Gain matrixJ: p x t current density distribution on the sources along the cortical surfacep >> n

Bayes Law:

Minimum Norm Estimate within the Bayesian Framework (2/5)

Bayes Law:

A priori distribution of the sensor noise E:Gaussian distribution with null mean andAssumption of uncorrelated noise

Data likelihood:

A priori distribution of the source distribution J:Gaussian distribution with null mean and

Minimum Norm Estimate within the Bayesian Framework (3/5)

Bayes Law:

Maximum likelihood solution

Data fit term

Regularization hyperparameter

Weighted Min. Energy constraint

Minimum Norm Estimate within the Bayesian Framework (4/5)

Bayes Law:

Maximum likelihood solution

This solution depends on the regularization hyparameter α

α can be estimated using the L-curve technique

Minimum Norm Estimate within the Bayesian Framework (5/5)

Maximum likelihood solution

Estimation of the regularization hyparameter α using the L-curve technique

Examples of source localisation of an epileptic spike using anatomical constraints

fMRI activation for a similar spike

Regularizing the ill-posed linear model: M = GJ +E

2. Maximum Entropy on the Mean (MEM)

Data fit: set of all distributions pJ(J)explaining the data on average

pJ(J)

µJ(J) Reference distribution = a priori information

Relative entropy

p*J(J)MEM

solution

MEM solution is the one with maximum µ−entropy, i.e., the one that makes the least assumption regarding missing informationPrior information on J:

Parcelling of the cortical surface in K parcels:

Validation Results: 4th order spatial extent (14 cm2)

Gold Standard:

Temporo-Radial Source

MN : AUC = 0.78 MEM : AUC = 0.93 LORETA : AUC = 0.99Gold Standard:

Temporo-Tangential Source

MN : AUC = 0.79 MEM : AUC = 0.93 LORETA : AUC = 0.96

Grova C, Daunizeau J, Lina JM, Benar CG, Benali H, Gotman J. Neuroimage. 2006 Feb 1;29(3):734-53.

Validation Results: MEG source localization (1/2)

Chowdhury R. et al, Proc. of HBM 2010 conference

Application of model evidence for model comparison: Henson et al. Neuroimage 2009

Validation: comparison between MEM sourcelocalization and intracranial EEG recordings

Lina et al, IEEE TBME Sumitted

Summary: any source localization method relies on some a priori assumptions

Number of generators well-knownECD approaches

Few decorrelatedsources, number unknownDipole scanningapproaches

Distributed network and/or extended sourcesDistributed sources approaches

• Requires a good knowledge of the signal to be localized• Requires statistics (SPM-like, non parametric)• Requires the comparaison of several methods

Take home messages

Complementarity between EEG and MEG: – MEG signals are not distorted by the skull (better spatial accuracy)– MEG can see only tangential sources whereas EEG can see tangential and radial

sources– MEG data acquisition is challenging: tiny magnetic fields in a noisy environment

EEG/MEG source localization:– Any source localization requires an a priori model of the underlying sources– Main models:

Equivalent current dipoleDipole scanning approachesDistributed sources (along the cortical surface)

– Model comparison or model selection is required (hypothesis test)Multimodal data fusion:

– There is no one to one correspondance between EEG source, MEG source and BOLD response: COMPLEMENTARITY

– EEG/MEG sources can be associated either with BOLD activation or deactivation– EEG/MEG signals require synchronization of neuronal activity,

the BOLD signal does not

Suggested Bibliography

Niedermeyer’s ElectroEncephalography: Basic Principles, Clinical Applications and Related Fields. 6th Edition, Ed. D.L. Schomer and F.H. Lopes da Silva. Wolters Kuwer, Lippincott Williams and Wilkins 2011.MEG: an introduction to methods. Eds. P.C. Hansen, M.L. Kringelbach and R.Salmelin Oxford University Press 2010.

Review PapersBaillet S., Mosher J., Leahy R., 2001. Electromagnetic brain mapping. IEEE Signal Process. Mag., 14–30.Hamalainen, M, Hari, R, Ilmoniemi, R, Knuutila, J, Lounasmaa, O (1993). Magnetoencephalography--theory, instrumentation, and applications to noninvasive studies of the working human brain. Rev Mod Phys, 65: 1-93.Michel C., Murray M., Lantz G., Gonzales S., Spinelli L., Grave de Peralta R., 2004. EEG source imaging. Clin. Neurophysiol. 115, 2195– 2222.

• 275 magnetometers (MEG sensors)

• 64 simultaneous EEG

• all channels sampled @ up to 12kHz

• multimodal stimulus presentation (video, audio, somesthetic, ...)

• audio, video subject monitoring

• operates in upright or supine positions

• located in the Neuro’s new extension

MEG system @ The Neuro

MEG @ The Neuro:more information

NEW COURSE offered in January 2012: BMDE 610 Functional Neuroimaging fusion

Registration is now open on Minerva for Winter 2012 Space is limited, register ASAP !!!

Acknowledgements

Montreal Neurological Inst.Eliane KobayashiM.L. Jones, M. AiguabellaM. Sangani, D. Rosenberg,M. Porras-Betancourt

The Multimodal Functional Imaging Lab.(Multi-FunkIm)Christophe Grova R. Chowdhury, Y. Potiez,A. Machado, T. Hedrich, A. Blanc

Ecole de Technologie Sup.Jean Marc LinaA.S. DubarryE. Lemay S. Deslauriers

Collaborators: -MNI – EEG/fMRI studies: J. Gotman, F. Dubeau, an their team !