HAL Id: pastel-00583104
https://pastel.archives-ouvertes.fr/pastel-00583104
Submitted on 4 Apr 2011
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of
sci-entific research documents, whether they are
pub-lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
To cite this version:
Sheraz Khan. MEG Source Imaging and Dynamic Characterization.. Signal and Image processing.
Ecole Polytechnique X, 2010. English. �pastel-00583104�
DE L'ÉCOLE POLYTECHNIQUE
Spé ialité : Mathématiques Appliquées
par
Sheraz Khan
MEG Sour e Imaging and
Dynami Chara terization
Soutenue le15 janvier2010 devant la ommission d'examen formée de :
Jury :
Habib Ammari - E ole Polyte hnique
Sylvain Baillet - Medi al College of Wis onsin
Ali Mohammad Djafari - Laboratoire des Signaux & Systèmes (Supéle )
Josselin Garnier - Université de Paris 7
Fran ois Jouve - Université de Paris 7
Roman Novikov - E ole Polyte hnique
AsIlookthroughthe lastthreeyears,therearesomanypeoplewithoutthehelp
of them thisthesis might neverbe possible, rstand foremost my PhD dire tors
SylvainBailletandHabibAmmari,whose ontinuousguidan eandsupportmake
this thesis realizable. Julien Lefevrefor his ex ellent mentorshipduring the
se -ond andthird yearsofPhD,without himHHDpartofthethesiswouldnot have
been possible.
My olleagues at Paris and Milwaukee, Manoj, Rey, Benoit, Guillaume,
So-phie,Bethandmanyothers,there onstanthelpandmotivation, makethisthesis
through,theyarealwayswithmewhenthingsarenotworkingand heermewhen
things work out.
LastlyIa knowledge myparents,mywife and mydaughter, for their
Mythesis hasaddressedtwo omplementary aspe ts of magneti sour e imaging
using Magnetoen ephalography:
1. Imaging ofneural urrent sour esfromMEG surfa ere ordings;
2. Dynami hara terization of neural urrent patterns at thesurfa e of the
ortex.
MEG Sour e Imaging
A urate estimation of the lo al spatial extent of neural urrent a tivity is
very important for the quantitative analysis of neural urrent sour es, as
esti-mated from Magnetoen ephalography (MEG) surfa e re ordings. In asso iation
withthe ex ellent time resolution oered byMEG, this wouldrepresent amajor
advan ement innon invasive, time-resolved fun tional brainimaging.
We addressed this issue through a new method alled Multipole Corti al
Remapping (MCR) to a urately spe ify the spatial extent of neural urrent
sour es.
InMCR,thezeroth-orderTikhonovregularized imageofthe urrent
distribu-tion onthe ortex isrstestimated fromMEGsurfa e datafor whi hwesought
for a realisti model of neural generators. Then the resulting fun tional image
is thresholded usinga simple histogram-based prin iple. This thresholded image
is thende omposed intogroups ofa tivation patternsfollowing anautomati
equivalent multipolar de omposition of ea h urrent pat h isthen obtained. By
default, themultipolar momentsarenot readily relatedto thea tualanatomi al
supportofthea tualneural urrentsdete tedusingMEG.Hen eweintrodu edan
image remapping te hniquesof themultipolar parameters ba konto theoriginal
orti al manifold, ina Bayesian framework in luding physiologi al and
anatomi- al priors.
Chara terization of MEG Sour e Dynami s
For dynami hara terization of neural urrent patternsat thesurfa e ofthe
ortex, we used a modied Helmholtz-Hodge De omposition (HHD), whi h was
appliedonve toreldsdes ribingtheowofneural urrentsour es. Thismotion
eld stems from a generalized approa h to opti al ow estimation, developed
earlier inour team.
Opti alowistheapparentmotion duetovariationsinthepatternof
bright-ness and, under spe i onditions, may mimi the velo ity eld of an obje t.
Normally, the opti al ow is obtained in a two-dimensional domain, whi h may
prevent a ess to some essential features of the obje t's motion with respe t to
the topology or geometry of the domain onto whi h it is evolving. A new
vari-ational method to represent opti al ow on non at surfa es using Riemannian
formulationwaspreviously introdu ed byour group to over omethis issue.
Webroadened this framework andintrodu ed anewformalism to dete t
fea-turesintheresultingopti alowmodelusingamodiedandextendedframework
to the HHD on 2-Riemannian manifolds, whi h we used to hara terize neural
urrent sour es.
HHD is a te hnique used to de ompose a two-dimensional (resp.
three-dimensional) ontinuous ve tor eld into thesum of 3 distin t omponents: (1)
a non-rotational element, deriving from thegradient of a s alar potential
U
;(2) a non-diverging omponent, deriving from the rotational of a s alar potentialA
(resp. ve torial potential); (3) a harmoni ve torial part, i.e., whose Lapla ianvanishes.
WeshowedhowHHDenablesthede ompositionandtra kingoftime-resolved
neural urrent ows asobtained from MEGsour e imagingas sour esand sinks
e.g., by dete ting relative maxima of the non-rotational s alar potential. We
hen eforth suggest to extend the analysis of brain a tivity in terms of tra king
travelling obje ts onto the orti al manifold by dete ting ve tors of largest
am-plitudes inzeroLapla ian harmoni ve tor elds.
We also onsidered HHD through a series of stru tural and fun tional brain
imaging appli ations, withvery en ouragingpreliminaryresults.
Themethods dis ussedintheHHDse tionofthethesiswereimplementedin
Matlabasplug-intotheBrainstorm(MEG/EEGdatapro essingsoftware)and
an be downloaded from: http://neuroimage.us .edu/brainstorm. A short
Contents 6
List of Symbols and Abbreviations 9
List of Figures 10
Ba kground 15
Te hniquesfor the observation ofthe Human brain . . . 15
Introdu tion to MEGand EEG: . . . 19
Neural bases ofbrain ele tromagneti signature . . . 20
Forward problem . . . 22
Maxwell's equations . . . 23
Modelingprimary urrents . . . 26
Headmodeling . . . 27
Spheri al head model . . . 27
Realisti head model . . . 29
Linear formulation . . . 32 Inverse problem. . . 33 Parametri methods . . . 35 Dipoletting . . . 35 Beamforming approa hes . . . 37 Mat h lter . . . 37
Linearly onstrained minimum-varian e (LCMV) . . . 38
Imaging methods . . . 40
Bayesian formulation . . . 41
Choi e oftheregularization parameter
λ
. . . 42Linear estimators . . . 43
Properties ofthesour e ovarian e matrix . . . 44
Nonlinear estimators ofsour e amplitudes . . . 46
Con lusion . . . 47
Multipolar Corti alRemapping 51 Introdu tion . . . 51
Multipolarexpansionsof as alar potential . . . 52
Spheri al multipolar expansions . . . 52
Multipoleexpansionsof adistributeddipole sour e. . . 54
Multipole momentsof urrent distributions . . . 57
Multipolar orti al remapping. . . 59
Compa t parametri de omposition of orti al urrents . . . 61
Sparse-fo al imagingmodel . . . 62
Results . . . 66
Simulated data . . . 67
A ura y riteria . . . 67
Single sour e ase . . . 67
Robustnessof MCRagainst hangesinthethreshold. . . 70
Two-sour e ase . . . 72
Experimental data . . . 74
Con lusion . . . 76
Helmholtz-Hodge De omposition 79 Introdu tion . . . 79
Regularization . . . 84
Variationalformulation . . . 85
HelmholtzHodgede omposition on2-Riemannian manifold . . . 87
Theory . . . 87
Denitions . . . 87
Theorem . . . 88
Dis retization. . . 89
Feature dete tion as riti al points ofpotentials . . . 92
Simulations andresults . . . 95
Con lusion . . . 99
Appli ations of HHD 101 Introdu tion . . . 101
HDDof MEGexperimental data . . . 102
Chara terizingepilepti a tivity . . . 103
Chara terizingof epilepti a tivityinECoG . . . 105
Chara terizingepilepti a tivitywithMEGsour e imaging . . . . 107
Identi ation of orti aldevelopment inthe neonate brain . . . 109
Chara terizingtumor growth patterns . . . 113
Con lusion . . . 115
Con lusion and Future A tions 117
Grid Generation 121
MEG-ECoG sour e lo alization and dynami s omparison 125
Brainstorm's HHD-Opti alow plug-in Tutorial 133
and Abbreviations
Abbreviation Des ription
EEG Ele troen ephalography
HFO High-Frequen y Os illations
HHD Helmholtz-Hodge De omposition
MCR MultipolarCorti al Remapping
MEEG Magnetoen ephalography and
Ele troen- ephalography
MEG Magnetoen ephalography
1 EEGSetup . . . 16
2 Comparisonof brainsignals . . . 17
3 MEGSetup . . . 17
4 Spatialandtemporal resolution ofdierent brainimagingmethods. . 18
5 MEEGeldpattern . . . 20
6 Cerebral Cortex . . . 21
7 Modeling of head regions . . . 25
8 Current Distribution Model . . . 27
9 Spheri alhead model, where asphere is ttedto thehead geometry. . 28
10 Realisti head Modeling . . . 31
11 DipoleFitting . . . 36
12 MUSICs an . . . 39
13 L-Curve . . . 43
14 GCV urve . . . 44
15 LCMVvs Minimnorm . . . 45
16 Spheri alharmoni multipole omponents . . . 54
17 DistributedDipoleModel . . . 56
18 Lo al urrent distribution . . . 58
19 A ura y ofMCR . . . 68
20 Subspa e orrelation estimationof MCR. . . 69
22 A ura y ofMCR . . . 71
23 Illustrationof the bootstrapestimateof onden eintervals. . . 72
24 Re onstru tionby MCRintwosour e s enario. . . 73
25 MCRinsomatosensoryexperiment . . . 75
26 A tive surfa e areas in
S
1
andS
2
regions. . . 7627 HHDinreal worldappli ations . . . 81
28 Basisve torsof tangent plane. . . 82
29 FEMformalism . . . 90
30 Sour es andsinks onat manifold . . . 92
31 Sour es andsinks onspheri al manifold . . . 93
32 vorti eson at manifold . . . 93
33 Vortexonspheri al manifold . . . 94
34 HHDon rabbitmanifold . . . 96
35 HHDon elephant manifold . . . 97
36 Tra king on rabbitmanifold. . . 98
37 HHDon orti almanifold . . . 99
38 Typi alECoG setup . . . 102
39 De ompositionof experimental data . . . 104
40 Divergen e representation . . . 105
41 HHDon ECoGGrid . . . 106
42 HHDfor seizureonset dete tion . . . 108
43 Two neonate orti al surfa es . . . 109
44 Displa ement eld between two orti al surfa es . . . 110
45 Dete tion ofgrowth seed . . . 111
46 Thereprodu ibilityof the growthseeds . . . 112
47 Tumor Dete tion . . . 114
48 Interpolated grid . . . 122
51 MEGGranger . . . 127 52 ECoGlo alization . . . 128 53 ECoGGranger . . . 129 54 EEGSlow. . . 131 55 BrainstormLaun h . . . 134 56 Plug-in GUI . . . 135 57 Opti al Flow Tab . . . 136 58 HHDTab . . . 137 59 VisualizationTab . . . 138
Te hniques for the observation of the
Hu-man brain
Exploration of the Human brain is of utmost intelle tual interest: de iphering
brain using brain is a hallenging task. Although a great deal has been learnt
aboutbrainanatomyand physiology,thefundamentalquestionshowbrainstore,
retrieve and pro esses information is still largely unknown and full dis overy of
these me hanisms is thefoundational purposeof neuros ien e.
When brain pro esses information, ele trophysiologi al urrents ow within
and outsideneural ells, thus produ ing ele tri and magneti elds thatare
a - essible toexternal measurements. Indeed,signsof this ele tri al neurala tivity
in the brain an be measured with ele trodes at the s alp or with very
sensi-tive magneti dete tors pla ed very near the s alp. The te hnique of ele tri al
measurements from the s alp is alled ele troen ephalography (EEG) [8℄.
His-tori al and re ent EEGsetups areshownin Figure 1. The te hnique measuring
magneti signalsgenerated byneural urrentsis alledMagnetoen ephalography
(MEG) [15 ℄.
Themagneti eldprodu ed byneural urrent sour esarevery weakandare
at least 8 orders of magnitude smaller than the earth stati magneti eld, as
shown Figure 2. Theseelds are urrently pi ked using series of magnetometers
Figure 1: (a) EEG setup in 1970's. (b) Modern EEG setup with qui k-x
ap.
isasensitivedete torofmagneti ux,whi hwasdevelopedbyJamesZimmerman
[114 ℄ inthe late 1960's.
The seminal, original MEG measurements were performed at MIT in May,
1971 byCohen. Alphawaves (ele tromagneti brainos illationsinthefrequen y
rangeof [8,12℄Hz)werere orded asshownFigure3 .a. Atypi al, state-of-the-art
MEGsetup using151 hannels isshown Figure3.b.
Brain imaging te hniques an be divided into two ategories: stru tural and
fun tional. Anatomi al stru tures an be investigated using omputer-aided
to-mography(CT)s ansandbettersousingmorere entmagneti resonan eimaging
approa hes(MRI).For fun tionalimaging beside neural ele tromagneti signals,
brainmetabolism,bloodowand volume(hemodynami s) anbea essedusing
Figure2: Comparisonofbrainsignalswithothersour esofele tromagneti
waves.
Figure 3: (a)First MEGre ording at MIT inside aspa eship like magneti
shielded room using single hannel SQUID. (b) MEG Setup at La
Pitié-Salpêtrière Hospital, Paris inside modern multilayer shielded room using
151 SQUIDs overing whole brain.
inthespatialdistributionoftheseprobes,astheyaretransportedand hemi ally
modied within the brain, an be imaged using positron emission tomography
(PET). These images an rea h a spatial resolutions ashigh as 3mm. However,
temporalresolutionislimitedtominutesbythedynami softhephysiologi al
As neurons be ome a tive,they indu e very lo alized hanges inblood ow and
oxygenationlevels that an be imaged asa orrelateof neural a tivity[65 ℄.
Hemodynami hanges anbedete ted using PET,fun tionalMagneti
Res-onan e Imaging (fMRI), and trans ranial opti al imaging methods. Of these,
fMRI is urrently the most widely used and an be readily performed using a
standard 1.5T lini alMRImagnetalthough anin reasing fra tionofstudiesare
now performed on higher eld (3-7T) ma hines for improved SNR and
resolu-tion. Fun tionalMRIstudiesare apable ofprodu ingspatialresolutionsashigh
as 2-4mm; however, temporal resolution is again limited by the relatively slow
hemodynami response, when ompared to ele tri al neural a tivity,to
approxi-mately one se ond. In addition to limited temporal resolution, interpretation of
fMRI data is hampered by the rather omplex relationship between the blood
oxygenationleveldependent (BOLD) hanges thataredete ted byfMRIandthe
underlying neural a tivity. Regions of BOLD hanges in fMRI images do not
ne essarily orrespond one-to-one withregionsof ele tri alneural a tivity[62 ℄.
Figure 4: Spatialand temporalresolution of dierent brain imaging
Introdu tion to MEG and EEG:
EEG and MEG measure the ombined a tivity of multiple areas of the brain
as a mixture of omplex signalpatterns. A primary obje tive is to interpret the
omplexpatternsofthemeasuredele tri potentialsandmagneti elds,interms
of therespe tivelo ationsand time- oursesoftheir underlying sour es. Thekey
to this taskisto design aphysi aland numeri al modelto a ount for theorigin
of the eld patterns aptured by MEG/EEG surfa e re ordings. Estimation of
theele tri andmagneti eldpatternsforagivenmodelofthevolume ondu tor
is a forward problem, following the nomen lature of modeling dataformation as
en ounteredinalargevarietyofappli ations(fromgeophysi stomedi alimaging)
[96℄.
The estimation of neural urrents from measured eld patterns is a typi al
inverseproblem. InEEGorMEGstudies,thesimplestwaytomodelthegeometry
of thehead is to usea single sphere approximationor on entri spheri al shells
ea h withhomogeneous isotropi ondu tivity[76℄.
Themainreasonwhy onsidering spheri algeometry istheavailabilityof
an-alyti al solutions, andtherefore fastimplementations, to solvethe forward
mod-eling problem. Howeveraspheri alapproximationofthehead omplex geometry
is likely to indu elarge sour e lo alization errors[72℄.
Using MRI, itis possible to provide more realisti geometri al models of the
head. Numeri alte hniques su h as theBoundary Element Method (BEM) and
Finite Element Method (FEM) provide the exibility of utilizing a realisti
ge-ometry [51 ℄.
EEGandMEGs alp patternsarequalitativelyorthogonal toea h other(see
gure 5), providing distin tive information about the underlying neural urrent
distributions. They therefore might be viewed as omplementary rather than
as ompeting modalities[24 ℄. Most state-of-the-art MEG fa ilities are equipped
for simultaneous a quisition of EEG and MEG data. Inverse methods for the
Figure 5: Left hand side gure represent the topographi sensitivity maps
of MEG and EEG for radial and tangential dipoles. Figure on right hand
side shows the orthogonalityof MEG and EEG eld patterns. patterns
Neural bases of brain ele tromagneti signature
MEGandEEG(MEEG)aretwote hniquesbasedonwhatGalvani,attheendof
the18th entury, alled"animalele tri ity", todaybetter knownas
ele trophys-iology [85 ℄. Despite the apparent simpli ity in the stru ture of the neural ell,
the biophysi s of neural urrent ow relies on omplex models of ioni urrent
generation and ondu tion [48 ℄. Roughly,when aneuronisex itedbyother
neu-rons via an aerent volley of a tion potentials, postsynapti potentials (PSPs)
are generated at its api al dendriti tree. When the ex itatory PSP's be ome
largerthan inhibitoryPSP's, theapi al dendriti membrane be omestransiently
depolarized and onsequently extra ellularly ele tronegative with respe t to the
ell soma and the basal dendrites. This potential dieren e auses a urrent to
ow throughthe volume ondu tor from the non-ex itedmembrane of thesoma
and basal dendrites to the api al dendriti tree sustaining the PSP's. Some of
the urrent takes the shortest route between the sour e and the sink by
travel-ling within the dendriti trunk (see gure 6). Conservation of ele tri harges
imposes that the urrent loop be losed with extra ellular urrents owing even
se ondary,return,or volume urrents.
Figure 6: The orientation of pyramidal neurons is normal to the ortex
surfa e. MEG signalspreferentially ree tthe urrent owfrompyramidal
ells oriented tangential tothe skull surfa e.
Bothprimaryandse ondary urrents ontributetomagneti eldsoutsidethe
headandtoele tri s alppotentials,butspatiallystru turedarrangementsof ells
are of ru ial importan e to the superposition of neural urrents su h that they
produ e measurable elds. Ma ro- olumnsoftens ofthousands ofsyn hronously
a tivatedlargepyramidal orti al neuronsarethus believedtobethemainMEG
and EEG generatorsbe ause of the oherent distribution of their large dendriti
trunks lo ally oriented in parallel, and pointing perpendi ularly to the orti al
surfa e. The PSPs generated among their dendrites are believed to be at the
sour e of most of thesignals dete ted in MEG and EEG be ause they typi ally
last longer than therapidly ring a tion potentials travelling along theaxons of
ex ited neurons. Indeed, al ulations su h as those shown in [44℄ suggest ea h
synapse along a dendrite may ontribute as little as a 20 fA.m urrent sour e,
probablytoo smalltomeasure inMEEG. Empiri alobservations insteadsuggest
of neuronal density and orti al thi kness suggest that the ortex has a
ma ro- ellular urrent density ofthe orderof 100
nA.mm
−2
[44℄. If we assumethatthe
ortexisabout4
mm
thi k,thenasmallpat hofsize 5mm
x5mm
wouldyield a net urrent of 10nA.m
, onsistent with empiri al observations and invasive studies [44 ℄.In MEEG studies, one is usually on erned with the uppermost layer of the
brain; the erebral ortex, whi h isa 2 to 6
mm
thi ksheet of graytissue where most of the measured neural a tivity takespla e. The se tion of ortex isillus-trated inFigure 6. At least 10 billion neurons residein thewhole ortex tissue.
The total surfa e area of the ortex is about 2500
cm
2
,folded ina ompli ated
way,sothatittswithintheinnerskullvolume. Thetruespatialextent of
realis-ti urrent sour es asso iated withbrain a tivationvariesa ordingto the ause
of thea tivation. Typi ally sensory stimuli a tivate orti al areas starting from
a few
mm
2
up to afew
cm
2
,whereas for spontaneous a tivityand epilepti fo i
an involve ana tivation areaup totens of
cm
2
[95℄.
At a larger s ale, distributed networks of ollaborating and syn hronously
a tivated orti alma ro- olumnsaremajor ontributorstoMEGandEEGsignals
[80℄. This is ompatible with neuro-s ienti theories thatmodel basi ognitive
pro esses intermsof dynami allyintera ting ellassemblies[105 ℄.
Mostregionsofthe ortexaremappedfun tionally. Forexample,theprimary
somatosensory ortex re eives ta tile stimuli from theskin. Areas of thefrontal
lobe are on erned with the integration of mus ular a tivity. Primary motor
ortex isinvolved inthe movement of a spe i partof thebody. Largeareas of
ortexaredevotedtobodyparts,whi haremostsensitiveto tou h(e.g.,lips)or
to theparts where a urate ontrol ofmus les isneeded(e.g.,ngers).
Forward problem
suredele tri /magneti eldsandthe urrentdistributionwhi hprodu eit. This
relationship isknownasforward modelingwhi h translatesasalead-eld matrix
or a gain matrix that binds theamplitude of sour e urrents to the sensor data
as we shall detail below. If the primary sour e and the surrounding
ondu tiv-ity prole of tissues are known, the ele tri potential and magneti eld an be
al ulated fromMaxwell's equations (see[7℄ for a omprehensive reviewofMEG
forward and inversemodeling).
Maxwell's equations
In 1873,Maxwell showed thatele tromagneti elds an bedes ribedusing only
4 ve tor dierential equations [70℄:
∇ × E +
∂B
∂t
= 0,
(1)∇ · B = 0,
(2)∇ · E =
ρ
ǫ
0
,
(3)∇ × B = µ
0
(J + ǫ
0
∂E
∂t
),
(4)where
E
istheele tri eld,B
themagneti eld,ρ
the hargedensity,andǫ
0
andµ
0
representthepermittivityandthepermeabilityoftheemptyspa ewithvalues8.85
10
−12
Fm−1
and4π10
−7
Hm−1
, respe tively (the magneti permeability
µ
0
of braintissues is onsidered identi al tothat ofthefree spa e).Negle tingtheee tsof thetime-dependent termsisthequasi-stati
approx-imation ofMaxwell's equations. Thisdepends on thetypi al frequen y range of
the signals of interest and the properties of the medium. The frequen y of the
signals obtained from bio-ele tromagneti measurements in MEG and EEG are
typi ally below 1 KHz. It has therefore been veried that the physi s of MEG
and EEG are well des ribed using the quasi-stati approximation of Maxwell's
equations [44 ℄. Quasi-stati Maxwell'sequations an bewritten as:
∇ · B = 0,
(6)∇ · E =
ǫ
ρ
0
,
(7)∇ × B = µ
0
J.
(8)Equation (5) an further be satised by representing the ele tri eld
E
as thegradient of as alarfun tionV
:E
=
−∇V.
(9)From (8 ), we obtain the relation between the urrent distribution
J(r
′
)
at
point
r
′
and the magneti eld
B(r)
measuredatr
whi h reads:B(r) =
µ
0
4π
Z
J(r
′
)
×
r
− r
′
||r − r
′
||
3
dv
′
,
(10)where
||.||
representsthe Eu lidean norm.Thisrelationship (10 ) ispopularly known asBiot-SavartLaw.
The urrent distribution
J(r)
an be dividedinto two parts:1. Primary urrent
J
p
(r)
produ ed bythe neural urrent a tivity;
2. Volume urrent
J
v
(r)
produ edinallthevolumetoprevent hargebuildup.
Primaryandse ondary urrentsareshowninFigure5.b. The urrentdistribution
J(r)
now an be representedasJ(r
′
) = J
p
(r
′
) + J
v
(r
′
) = J
p
(r
′
) + σ(r
′
)E(r
′
) = J
p
(r
′
)
− σ(r
′
)
∇V (r
′
),
(11)where
σ(r
′
)
is the ele tri al ondu tivity of the tissue at lo ation
r
′
, whi h we
will onsidertobeisotropi throughout thisthesis. SeeFigure7) wherethehead
onsistsof regionsof onstant ondu tivities
σ
i
, i = 1, 2, . . . , N + 1
.Now we an rewrite the Biot-Savart equation (10) and use (11 ) to divide it
into twoparts: therstpart onsistsof
B
0
(r)
,themagneti elddueto primary urrentsonlywhilethese ondtermisduetothe ontributionofvolume urrents,formed asasumofsurfa eintegralsoverthe brain-skull,skull-s alpands alp-air
boundaries. In fa t,wehave
B(r) = B
0
(r) +
µ
0
4π
X
ij
(σ
i
− σ
j
)
Z
S
ij
V (r
′
)
r
− r
′
||r − r
′
||
× dS
′
ij
.
(12)Figure7: Shell modelof the head.
This general equation states that the magneti eld an be al ulated if we
knowthe primary urrent distributionandthepotential
V (r
′
)
on allthesurfa es
S
ij
. We an reate asimilar equationfor thepotential itself,yielding(σ
i
+ σ
j
)V (r) = 2σ
0
V
0
(r)
−
1
2π
X
ij
(σ
i
− σ
j
)
Z
S
ij
V (r
′
)
r
− r
′
||r − r
′
||
× dS
′
ij
,
(13)where
V
0
(r)
isthe potential atr
due to the primary urrent distribution. Ifwespe ifyaprimary urrent distributionJ
p
(r
′
)
,we an al ulateaprimary
potential and aprimary magneti eldasfollows
V
0
(r) =
1
4πσ
0
Z
J
p
(r
′
)
·
r
− r
′
||r − r
′
||
× dS
′
ij
,
(14)B
0
(r) =
µ
0
4π
Z
J
p
(r
′
)
·
r
− r
′
||r − r
′
||
× dS
′
ij
.
(15)Theprimary potential is thenused to solve (13)for the potentials on all the
surfa es, and therefore ompletes the resolution of the forward problem. These
solve (12) for theexternal magneti elds. Unfortunately,the solution to (13 )is
analyti onlyinaspe ialshapesandellipti volume ondu torandmustotherwise
besolved numeri ally. Thisthesiswill onsiderusingspheri al headmodels only.
Inthe next two se tions,modelsfor neural urrent distribution will be
intro-du ed and subsequently modelsfor volume ondu tor will be dis ussed.
Modeling primary urrents
Consider a small pat h of a tive ortex
S(r
′
)
entered at
r
′
and an observation
point
r
at some distan e from this pat h. The primary urrent distribution in this ase an be well representedbythe multipolar representationΩ
n
S((r
′
))
given byΩ
n
S((r
′
))
=
1
n!
Z
r
′
⊂S((r
′
))
(r
′
− l)
n
J
p
(r
′
)dr
′
,
(16)where
l
isthe point ofexpansion for multipoles.It is important to note that the brain a tivity does not a tually onsist of
dis rete sets of physi al urrent dipoles, but rather that the dipole is a
onve-nient representation for oherent a tivation ofa largenumberof pyramidal ells,
possiblyextendingovera fewsquare entimeters ofgraymatter.
If the primary urrent distribution is very fo al then it an be well
approxi-mated by anequivalent urrent dipole(ECD) dened as:
Ω
0
= q
≡
Z
J
p
(r
′
)dr
′
.
(17)TheECD an be representedasa point sour e
J
p
(r
′
) = qδ(r
′
− l),
(18)where
δ(r)
is the Dira delta distribution. Note that an ECD is a multipolar expansion oforder0
.Ifthe urrent distributionis not fo al,thenmultipolar expansionsarebetter
suited for themodeling ofneural sour es. The ontributions reported[74, 54,53℄
Figure 8: Current Distribution
S(r
′
)
entered atr
′
and measured atr
. Adapted from [54℄. Head modelingSpheri al head model
Headmodeling usingasspheri al approximationofits geometry hasbeenwidely
used in the MEG ommunity, the reason for its popular use is the simpli ity it
oerswithrespe tto omputationrequirements. Computings alppotentialsand
indu edmagneti eldsrequiresolvingtheforwardequations(13)and(12)
respe -tively for a parti ular sour e model. We have seenabove that when the surfa e
integralsare omputed overrealisti headshapes, theseequationsmustbesolved
numeri ally. However, analyti solutions exist for simplied geometries, su h as
when the head is assumed to onsist of a set of nested on entri homogeneous
spheri al shells representing brain, skull, and s alp respe tively. These models
are routinely used in most lini al and resear h appli ations to E/MEG sour e
lo alization. Figure 9des ribes aspheri al head model approximation. Consider
spheri al head,and a MEG systeminwhi h we only measurethe radial
ompo-nent of the external magneti eld, i.e., the oil surfa e of the magnetometer is
oriented orthogonally to a radial line from the enter of the sphere through the
enter of the oil. It is relatively straightforward to showthat the ontributions
of thevolume urrents vanish inthis ase, andwe areleft withonly theprimary
term. Taking the radial omponent ofthis eld forthe urrent dipole redu es to
theremarkablysimple form:
B
r
(r) =
r
r
· B(r) =
r
r
· B
0
(r) +
µ
0
4π
·
X
ij
(σ
i
− σ
j
)
Z
S
ij
V (r
′
)
r
r
r
− r
′
||r − r
′
||
× dS
′
ij
.
(19)Figure 9: Spheri al head model,where a sphereis ttedtothe head
geom-etry.
In this same ase, it is very simple to show that the ontribution of volume
urrents will also redu eto zero. Hen e these ond term in 19 vanishes and this
equation writethe following simpler form:
B
r
(r) =
r
r
· B
0
(r) =
µ
0
4π
r
× r
′
r
||r − r
′
||
3
· q.
(20)Noti e here that the magneti eld
B
r
(r)
is linear with respe t to the dipole momentq
buthighly nonlinearwithrespe tto dipolelo ation:r
q
.In nutshell,
B
r
(r)
is zero everywhere outside the head ifq
points towards theradialdire tionr
q
. Amore general resultisthatradially-orienteddipoles do not produ e anyexternal magneti eld outside a spheri ally symmetri volumeondu tor, regardlessofthe sensororientation [89℄.
Importantly,this isnot the asefor EEGwhi h issensitive to radialsour es,
whi hdemonstratesoneofthe omplementarydieren esbetweenMEGandEEG
prin iples.
Realisti head model
In reality, the head has anisotropi tissue properties, is inhomogeneous and not
spheri albutsurprisingly,thespheri alapproximationworksreasonablywell,
par-ti ularly for MEG, whi h is less sensitive than EEG to volume urrents. These
latter are more ae ted than primary urrents by deviations from the idealized
model. By usingthe individual MRIdata fromthesubje t, itispossibleto
on-stru t amore detailed head model byisolating dierent regions ofinterest using
fully-automati segmentationte hniques[16 ℄. Figure10showstypi alsurfa eand
volume tessellations for usewith BEM and FEM(see [33℄ for a omplete review
of thehead geometriesusedinMEG).
Two typesofapproa hesareavailable for realisti head modeling:
1. BoundaryElementMethod(BEM)BEMisanumeri alte hniqueofsolving
linearpartialdierentialequationswhi hhavebeenformulatedina
bound-ary integral form. Normally in MEG, single-shell and three-shell BEM
methods are used. BEM methods still assume homogeneity and isotropy
within ea h region of the head. It therefore ignores, for example, the
on-du tivity anisotropy indu ed by white matter tra ts, where ondu tion is
higher along axonal bers ompared to a transverse dire tion. Similarly,
2. Finite Element Method (FEM)) FEM is a numeri al te hnique for nding
approximatesolutions ofpartialdierential equations(PDE).InFEM,
dis- retizationof thePDE isperformedintheentire head volume. Anisotropy
and heterogeneity in dierent tissue types an therefore be modeled and
therefore represents a very omprehensive approa h to solving theMEEG
forward problem.
Typi ally,BEM andFEM al ulationsareverytime onsumingandtheiruse
may be onsidered as impra ti al when in orporated as part of an iterative
in-verse solverfor urrent sour es. In fa t, through use of fastnumeri al methods,
pre- al ulation,andlook-uptablesandinterpolationofpre- al ulatedelds,both
FEM and BEM an be made quite pra ti al for appli ations in MEG and EEG
[31℄. One problemremains: thesemethods reauirethe ondu tivity propertiesof
head tissues be known. Most of head models used in the bio-ele tromagnetism
ommunity onsider typi al values for the ondu tivity of the brain, skull and
skin. Skull is typi ally assumed to be 40 to 90 times more resistive than brain
and s alp, whi h are assumed to have similar ondu tive properties. These
val-ues were measured in vitro from postmortem tissue samples, with ondu tivity
values that may be signi antly altered from those in in vivo tissues however.
Consequently,somere ent resear heortshavefo usedoninvivomeasurements
of tissue ondu tivity. Ele tri al Impedan e Tomography (EIT) pro eeds by
in-je ting a small urrent (1-10 mi roA) between pairs of EEG ele trodes and by
measuring the resulting potentials at all ele trodes. Given a model for the head
geometry, EIT solves an inverse problem by minimizing the error between the
measuredpotentialsontherestoftheEEGleadsandthemodel-based omputed
potentials, interms of parameters of the ondu tivity prole. Simulation results
with three or four-shell spheri al head models have demonstrated the feasibility
of this approa h thoughtheasso iated inverseproblem isalso fundamentally
ill-posed [32 ℄. These methods are readily extendible to realisti surfa e models as
of spatiallyvaryinganisotropi ondu tivity. A se ondapproa hto imaging
on-du tivity is to usemagneti resonan e. One te hnique uses the shielding ee ts
of indu ed eddy urrents on spin pre ession and ould in prin iple help
deter-mine the ondu tivity prole at anyfrequen y [113℄. The se ondte hnique uses
diusion-tensor imaging withMRI(DT-MRI) thatprobes themi ros opi
diu-sion properties of water mole ules within thetissues of the brain. The diusion
values an then be tentatively related to the ondu tivity of these tissues [100℄.
None ofthese MR-based te hniques have rea hed ommon pra tise by far . F
ur-ther, giventhepoorsignal-to-noiseratio(SNR)oftheMRinboneregions,whi h
is of riti alimportan e forthe forward EEGproblem, thepotential for fully3D
impedan etomography withMRremains spe ulative.
Figure10: (a) FEM modelingof the forward model; (b) BEM modelingof
Linear formulation
Theforwardproblemnow anbeexplainedusingthemodelsforsour esandhead
geometry dis ussed above. Themagneti eldand s alppotential measurements
are linear with respe t to the dipole moment
q
and nonlinear with respe t to its lo ationr
′
. For larity, it is onvenient to separate the dipole magnitude
q =
||q||
from itsorientationu
= q/
||q||
,whi hwewrite inspheri al oordinates byΘ = [φ, ρ]
. Letb(r)
denote the magneti eld generated by a dipole having xed orientationΘ
:b(r) = g(r, r
q
, Θ)q,
(21)where
g(r, r
q
, Θ)
isa leadeldsolutionof the magneti eldfor adipolehaving unit amplitude andorientationΘ
.ForN dipoleslo atedat
r
q
i
,their ombinedmagneti elds anbeexpressedusing linearsuperposition ofMaxwell's equations as
b(r) =
N
X
i=1
g(r, r
q
i
, Θ
i
)q
i
.
(22)The simultaneous MEGmeasurements made at m sensors for N dipoles, an be
expressedas B
=
B(r
1
)
. . .B(r
m
)
=
G(r
1
, r
q1
, Θ
1
)
. . .
G(r
1
, r
qN
, Θ
N
)
. . . . . . . . .G(r
m
, r
q1
, Θ
1
) . . . G(r
m
, r
qN
, Θ
N
)
q
1
. . .q
p
.
(23)It an be written inamatrix form as
B
=
G(
{r
qi
, Θ
i
})
J,
(24)where
G(
{r
qi
, Θ
i
})
isthem
× N
gainmatrixrelatingN dipolestothem sensors. Ea h olumn ontains the ontribution of one dipoleto ea hsensorin thearray.The matrix
J
ontains thesetof instantaneous amplitudes ofall thedipoles. Inthismodel,theorientationofthedipoleisnotafun tionoftime. Thistypeof modelis often referredto asa "xed" dipole model. Alternative models that
Inverse problem
To produ e estimates of the neural urrent sour es that generated the observed
MEGsignals,wemustsolve theasso iated quasi-stati ele tromagnetism inverse
problem. The inherent ill-posedness of this problem, oupled with the limited
number of spatialmeasurements available with urrent MEGand EEG systems,
(150-300 measurements) and signal-to-noise ratio (SNR) make this estimation
very hallenging [44℄.
The solutions to the neuromagneti inverse problem will depend on whi h
forwardmodelisused. Infa t,agiveninversealgorithmwillyieldslightlydierent
results if dierent forward models are used; hen e, the importan e of using an
a urate realisti forward model. However, these two problems are relatively
independent of one another. In the forward problem, we attempt to model the
lassi al physi s ofMEGandEEGasrealisti allyaspossible. In ontrast, inthe
inverse problem, we often deal with purely mathemati al on epts and a priori
assumptions that are in orporated in a sour e model. The independen e of the
inverse problem from the model's physi s allows one to use the same inverse
algorithm for MEG or EEG. On the other hand, many dierent estimates of
a tivity anbeobtainedforaparti ulardatasetusingdierentinversealgorithms
but sharing the same forward model. This brings us to the main issue with
neuromagneti inverseestimation: nonuniqueness. Thereisnounique solutionto
the physi ally and mathemati ally ill-posed neuromagneti inverse problem. In
fa t,aninnitenumberof urrentsour edistributions anintheorygenerateany
parti ular magneti eld measurement ve tor due to the existen e of magneti
silentsour es [47 ,44, 89℄.
In both MEG and EEG, silent sour es an be added to any given inverse
solution without hanging the forward eld and/or potential that the ombined
sour e generates. Thus, there are indeed an innite number of solutions that
explain anygiven MEG/EEGdataset equallywell. Therefore, a priori
though mathemati ally unique solutions an be obtained by postulating spe ial
sour e properties, physi al non-uniqueness is intrinsi to the neuromagneti
in-verse problem.
The two major approa hes to the estimation of neural urrent sour es are
"imaging" and "parametri /lo alization"methods.
Imaging methods typi ally onstrain sour es to a tessellated surfa e
represen-tation of the ortex, assume an elemental urrent sour e in ea h area element
(vertex) normal to the ortex surfa e, and solve the linear inverse problem that
relates these urrent sour es to the measured data. A urate tessellation of the
ortexrequiresontheorderof
10
5
elements. Sin ethemaximumnumberofMEG
sensors is about 300, theproblem is highlyunder-determined. By using
regular-ized linear methods based on minimizing a weighted
l
2
-norm on the image, we
an produ e unique stablesolutions.
Parametri /lo alization methods assume a spe i parametri form for the
sour es. Byfarthemost widelyusedmodels inMEGaremultiple- urrent-dipole
approa hes[112 ,90℄. Theseassumethatthenumberofneuralsour esisrelatively
smalland ea hsu iently fo althatthey anberepresentedbyafewequivalent
urrent dipoles with unknown lo ations and orientations. In both imaging and
parametri methods, the MEG/EEGforward problem an be written as
B
= G(θ)J + ǫ,
(25)where
B
istheM
×time
ve torrepresentingMEGmeasurements,J
istheN
×time
ve tor representing thedistribution urrents. For imaging methods, itistheam-plitude ofelementary urrentsat ea h orti al vertex. In parametri methods,it
isthevaluesofamplitudeparametersforea h urrentmodelelement.
G(θ)
istheM
×N
leadeldmatrixrelatingadditionalparametersofthe urrentdistribution to the magneti eld measured byM
sensors.θ
gathers the parameters whi h the lead elds depend uppon, i.e., urrent sour es, lo ationsr
q
i
, orientationsΘ
i
andtheiramplitudesq
i
. TheM
× time
noiseve torǫ
representsa ombination ofa tivity ofheart andeyes, et ) onsensors.
Parametri methods
Parametri methods an be broadly lassied into "Dipole tting" and
"Beam-forming".
Dipole tting
Therstinversemethodforequation(25 )isbasedontheassumptionthatneural
a tivity an be modeled by a few sparse, elementary sour es
α
. The problem redu es to theestimationfromthedataof theparametersθ
forα
sour es, whi h are des ribed as their positionsr
q
i
, their orientationsΘ
i
and their amplitudesq
i
(withi
∈ [1, α]
). This may be written as an optimization problem of a ostfun tion to be minimized.
Theestimateinthe least-squares (LS)sense writes:
J(θ)
LS
= arg min
J
||B − G(θ)J||
2
F
(26)where
||.||
F
denotes the Frobenius norm. LetG
+
(θ)
be the pseudo-inverse ofG(θ)
:G
+
(θ) = US
+
V
t
,
(27) whereUSV
t
is the singular value de omposition (SVD) of
G(θ)
andS
+
is the
diagonal matrix ontaininginverse ofsingularvalues of
G(θ)
[39℄. Equation(26) an bewritten intheform:J(θ)
LS
=
||B − G(θ)[G
+
(θ)B]
||
2
F
=
||(I − G(θ)G
+
(θ))B
||
2
F
,
(28)where
I
istheidentitymatrix of rankα
. Thus, theLSproblem an be optimally solvedinthelimitedsetofnonlinearparametersr
q
i
,Θ
i
withaniterative minimiza-tion pro edure. The linear parameters inq
i
are then optimally estimated from 26;see[75 ℄. MinimizationmethodsrangefromMarquardt-LevenbergandFigure 11: (a) Dipole Fitting in axial view; (b) Dipole Fitting in oronal
view ( ) Dipoletting insagittal view.
Thisleast-squaresmodel aneitherbeestimatedfromdatafromasingletime
snapshot or atimewindow. When applied sequentiallyto a setof timesamples,
thisresultsina"movingdipole"model,sin ethelo ationisnot onstrained[112 ℄.
Alternatively, by using a ontiguous time blo k of data in the least-squares t,
the dipole lo ations an optionally be xed over the entire interval. The xed
and moving dipole models have both proven useful inboth EEG and MEG and
remain the most widely usedapproa hes to pro essing experimentaland lini al
data. A key problem with the LS method is that the number of sour es to be
used must be de ided a priori. Estimates an be obtained by looking at the
ee tive rank of thedatausing a SVDor through information-theoreti riteria,
su iently largenumberofsour es an be madetot anydataset, regardlessof
its quality. Furthermore, as the number of sour es in reases, the non- onvexity
of the ost fun tion results inin reased han es of trapping in undesirable lo al
minima. This latter problem an be approa hed using sto hasti or multistart
sear h strategies [50℄. The alternatives to LS des ribed below avoid the
non- onvexity issue by s anning a region of interest that an range from a single
lo ation to the whole brain volume for possible sour es. An estimator of the
ontributionofea hputativesour elo ationtothedata anbederivedeithervia
spatial ltering te hniques or signal lassi ation indi es. An attra tive feature
of these methods is that they do not require a prior estimate of the number of
underlying sour es.
Beamforming approa hes
A beamformer performsspatial ltering on data from asensor array to
dis rim-inate between signals arriving from a lo ation of interest and those originating
elsewhere. Beamformingoriginated inradar and sonar signalpro essing but has
sin e found appli ations in diverse elds ranging from astronomy to biomedi al
signalpro essing [103℄.
Mat h lter
Thesimplestspatiallter,amat hedlter,isobtainedbynormalizingthe olumns
of the lead eld matrix and transposing this normalized di tionary. The spatial
lter for lo ation
r
i
is given byW
i
(T )
=
G
T
: i
kG
: i
k
F
.
(29)Thisapproa hessentiallyproje tsthe dataontothe olumnve torsofthe
di tio-nary. Although thisguarantees thatwhen onlyone sour e isa tive,theabsolute
the spatial resolution of the lter is so low given the high orrelation between
di tionary olumns. Thisapproa h an beextendedto fast re ursive algorithms,
su h as mat hing pursuit and its variants, whi h sequentially proje t the data
or residual to the non-used di tionary olumns to obtain fastsuboptimal sparse
estimates.
Multiple signal lassi ation (MUSIC)
The MUSIC algorithm was adopted from spe tral analysis, Dire tion of
Ar-rival(DOA)estimation te hniquesand modiedfor spatiallteringof MEGdata
[75, 73℄. TheMUSIC ost fun tion isgivenby
W
i
(T )
=
I
− U
s
U
T
s
G
: i
2
2
kG
: i
k
2
2
=
P
⊥
U
s
G
: i
2
2
kG
: i
k
2
2
,
(30) whereB
= USV
T
isthesingularvaluede ompositionofthedata,
U
s
isamatrix with the rstd
s
right singular ve torsthat formthe signalsubspa e, andG
: i
is thegainve torforthedipolelo atedatr
i
andwithorientationθ
i
(obtainedfrom anatomy or using thegeneralized eigenvalue de omposition). The operatorP
⊥
U
s
is an orthogonal proje tion operator onto the datanoise subspa e. TheMUSIC
mapis there ipro alofthe ost fun tionat all lo ationss anned. Thismap an
be usedto guidea re ursive parametri dipoletting algorithm. The number
d
s
is usuallyset byan expert user.For more omplete explanation of subspa e methods likeMUSICsee [55℄.
Linearly onstrained minimum-varian e (LCMV)
Beamformers,asusedintheeldofbrainimaging,arespatiallteringalgorithms
that s anea h sour e-point independently to passsour e signalsat a lo ation of
interestwhilesuppressinginterferen efromotherregionsusingonlythelo algain
ve tors and the measured ovarian e matrix. One of the most basi and often
Figure12: A typi alMUSIC s an for epilepti spikes.
to a unitygain onstraint:
min
W
i :
tr W
i
:
Σ
B
W
T
i
:
subje ttoW
i
:
G
: i
= I,
(31)where
Σ
B
isthe data ovarian e matrix,G
: i
isthed
b
by3gainmatrix ofthei
th
sour e point, and
W
i
:
is the 3 byd
b
spatial ltering matrix [104 ℄. The solution to this problemis given byW
i
(T )
= G
T
: i
Σ
B
−1
G
: i
−1
G
T
: i
Σ
B
−1
.
(32)The parametri sour e a tivity at the
i
th
a tivity. Thisbeamformingapproa h anbeextendedtoamoregeneral Bayesian
graphi al model that uses event timing information to model evoked responses,
whilesuppressinginterferen eandnoise sour es[115℄. Thisapproa husesa
vari-ationalBayesian EMalgorithm to ompute thelikelihoodofadipoleatea hgrid
lo ation.
Imaging methods
Imagingapproa hestotheMEGinverseproblem onsistofmethodsforestimating
the amplitudesof a dense setof dipoles distributed at xed lo ationsand
orien-tation within the head volume. In this ase, sin e thelo ations and orientation
are xed, only the linear parameters need to be estimated and theinverse
prob-lem redu estoalinearone withstrongsimilaritiesto thoseen ounteredinimage
restoration and re onstru tion. By putting lo ations and orientation onstraint
theequation (25)be omes
B
= GJ + ǫ.
(33)Herethegainmatrix
G
isxedandonlydipoleamplitudesJ
havetobeestimated. The most basi approa h onsists of distributing dipoles over a predenedvolumetri grid similar to theones usedin s anning approa hes. However, sin e
primary sour es are essentially restri ted to ortex, the image an be plausibly
onstrainedto sour eslyingonthe orti alsurfa e,asextra tedfroman
anatom-i al MR images of the subje t [22℄. Following segmentation of the MR volume,
dipolar sour esarepla ed at ea h node ofa triangular tessellationof thesurfa e
ofthe orti almantle. Sin ethepyramidal ellsthatprodu e themeasuredelds
areorientednormalto thesurfa e,we anfurther onstrainea hofthese
elemen-tal dipolar sour esto benormal to the surfa e. The highly onvoluted nature of
the human ortex requires that a high-resolution representation ontains of the
order of ten to one hundred thousand dipole "pixels". The inverse problem is
this hasbeen a omplished through the use of regularization or Bayesian image
estimation methods.
Bayesian formulation
Bayesian approa h to neuronmagneti inverse problem was rst introdu ed by
Clarkein1989[14℄. IntheBayesianformalism,theneuromagneti inverseproblem
isdened astheproblemofestimating thematrix
J
of dipoleamplitudesat ea h tessellation element from the spatio-temporal data matrixB
, whi h arerelated inthenoiseless asebyB
= GJ
. Thei
-throwofJ
ontains theamplitude image a rossthe ortexattimei
. FromBayestheorem,theposteriorprobabilityp(J
|B)
for theamplitude matrixJ
onditioned onthedataB
is given byp(J
|B) =
p(B
|J)p(J)
p(B)
,
(34)where
p(B
|J)
gives the forward probability density of getting magneti eldB
onditioned onJ
.p(J)
is a prior distribution ree ting our knowledge of the statisti al properties of the unknown image. While Bayesian inferen eoers thepotentialforafullstatisti al hara terizationofthesour esthroughtheposterior
probability, images are typi ally estimated in pra ti e by maximization of the
posterior or log-posterior probability.
Theestimation of
J
inthe maximum a posteriori (MAP)sense isgiven byˆ
J
M AP
= arg max
J
p(B
|J)p(J).
(35)
The log-likelihood of (35)isgiven by
ˆ
J
M AP
= arg max
J
(log[p(B
|J)] + log[p(J)]).
(36)
Typi ally, MEG and EEG data are assumed to be orrupted with additive
Gaussian noise that we assume here to be spatially identi ally distributed over
all sensors (generalization is straightforward). The log-likelihood is thensimply
given,within a onstant, by
ln[p(B
|J)] = −
√
1
2σ
2
||B − GJ||
2
The priorisa probabilisti modelthatdes ribesour expe tations on erningthe
statisti alpropertiesofthesour eforwhi hwewillassumeanexponentialdensity
p(J) =
1
z
exp[
−βf(J)],
(38)where
z
andβ
andf (J)
depends on the imageJ
. This form en ompasses both multivariate Gaussian models and the lass of Gibbs distributions or Markovrandom eld models [13℄. Combining the log-likelihood and log-prior gives the
generalformofthenegativelog-posteriorwhoseminimizationyieldsthemaximum
a posteriori estimate:
ˆ
J
M AP
= arg min
J
||B − GJ||
2
F
+ λf (J),
(39) whereλ = 2βσ
2
.
λ
is the regularization parameter. The parameterλ
should be onsidered as a regularization parameter tuning between the priorf (J)
and t to the data. Ifλ = 0
estimation of the urrent distribution be omes simply least squares. This type of solution to the inverse problems was introdu ed byTikhonovin[97℄.
Choi e of the regularization parameter
λ
There are many approa hes to estimate the valueof
λ
. We summarize a few as explained below:1.L-Curve: Whenplotted onalog-logs ale,theparametri urveofoptimal
valuesof
||W||
anddatat||B − GJ||
oftentakesonanLshape. Forthisreason, the urve is alledanL- urve[45 ℄. The value ofλ
intheL- urve riterion isthe value ofλ
that gives the solution losest to the orner of the L- urve, asshown inFigure 13 .2. Generalized ross validation (GCV) is an alternative method for
estimating the regularization parameter
λ
[107 ℄, that has a number of desirable statisti al properties. Considerf (λ) =
||B − GJ||
T race(I
− GG
†
)
=
V (λ)
Figure13: Typi alL- urve for lassi shaw inverse problem.
The numerator in (14) is the data mist in the least squares sense and the
dominator measures the loseness of the data resolution matrix to the identity
matrix. In the GCV method, we pi k the value of
λ
that minimizes (14), as shown inFigure14 .Linear estimators
Thesimplest approa h to(39) isto onsider priordistributionof sour e
ammpli-tudes
J
to be Gaussian withzeromean. Introdu ef (J) = tr[JC
−1
J
J
t
],
(41)where
C
−1
J
is the inverse ovarian e matrix of sour es. If we break this inverse matrix as,C
−1
J
= WW
t
,then(39 ) an be written inthefollowing manner:ˆ
J
M AP
= arg min
J
||B − GJ||
2
Figure14: Typi al GCV- urve for lassi shawinverse problem.
The MAPestimator now takesthefollowing simple linearform:
ˆ
J
t
M AP
= WW
t
G
t
(GWW
t
G
t
+ λI)
−1
B.
(43)Inthis ase,
J
ˆ
M AP
alsofollowsaGaussiandistribution. (39 )isnormallyknownas zeroth orderTikhonovregularized solutionofJ
[97 , 26 ℄,where theregularization parameterλ
anbeestimatedfromanyofthete hniquesexplainedintheprevious se tion.Properties of the sour e ovarian e matrix
Sour e ovarian e is the last parameter of the model whi h will ondition the
1. The identity matrix,whi h yields lassi alminimum-normestimators [97℄.
Themajorassumptioninusingtheidentitymatrixisthatsour eamplitudes
J
areindependent andidenti ally distributed.In Figure 15 ,a omparison is shown between LCMV beamformer and the
minimum-norm solution to the inverse problem, showing that though the
minimum-norm solution is widespread, the peak of maximum intensity is
in the right pla e in this median nerve stimulation experiment, where we
expe t a tivitywithin primarysomatosensory areas.
Figure 15: Comparison of LCMVand minimum norm.
2. A diagonal matrix whose elements are given by the norm of the elements
of the orresponding olumn in the lead-eld matrix (i.e.,
W
ii
=
||g
i
||
2
with
g
i
thei
th
olumn of
G
). This solution is a forward-eld normalized solution.3.
W
whi h isbasedontherelationship between sour e neighbors[108 ℄. The matrixW
is given byW
ij
=
1
ifi = j,
−
n
1
ifj
∈ N (i),
0
otherwise,
where
N (i)
denestherstorderneighborofi
th
sour eand
n =
Card[N (i)]
.4.
W
is diagonal with elements equal to some estimate of the sour e power at thatlo ation,whi hmaybe omputed fromtheoutputofabeamformeror MUSICs anevaluated for ea hdipolepixel [69 ℄or weighted fromother
fun tional imagingmodalitiessu hasfMRI, PET, orSPECT [64 , 21 ℄.
Thesemethodshavetheadvantagetobefastandoverallrobusttowardsnoise
[106 ℄. They provide estimates where the enter of gravity of thea tivity is very
lose to the true sour e. However, results are often very smooth spatially and
do not allowfor estimation of thespatial extent ofthe a tivity. This problemof
spatial extent andits solutionwill beaddressed indetailsinChapter 2.
Nonlinear estimators of sour e amplitudes
Itispossibletoobtainsparserimageestimatesofthe urrentdistributionbyusing
alternative (non-quadrati ) ost fun tions
f (J)
in (39). Norms and semi-norms on sour e amplitude priors with valuesp
≤ 2
in (42 ) have been investigated. Solutions will be ome in reasingly sparse asp
is redu ed. For the spe ial ase ofp = 1
, the problem an be slightly modied to be re ast asa linear program. Thisisa hievedbyrepla ingthequadrati log-likelihoodtermwithasetofunder-determined linear inequality onstraints, where the inequalities ree t expe ted
mismat hesinthettothedataduetonoise. The
l
1
- ost anthenbeminimized
overthese onstraintsusing a linearsimplexalgorithm. Properties of linear
pro-grammingproblems guaranteethatthereexistsanoptimalsolutionfor whi hthe
number of non-zero pixels does not ex eed the number of onstraints, or
equiv-alently the number of measurements. Sin e the number of pixels far outweighs
thenumberofmeasurements, thesolutionsarethereforeguaranteedtobesparse.
This idea an betaken even further byusing thequasi-norm for valuesof
p < 1
. In this ase, it is possible to show that there existsa value0 < p < 1
for whi hAnotherapproa hdened liquishrelationshipsbetweenneighborhoodsour es.
The whole network of sour esmay be des ribed asdistributed within a Markov
Random Field(MRF), this relationship wasexploited in [5,84 ℄. Akeyproperty
ofMRFsisthattheirjointstatisti aldistribution anbe onstru tedfromasetof
potentialfun tionsdenedonalo alneighborhoodsystem[83℄. Thus,theenergy
fun tion
f (J)
for the prior an beexpressedasf (J) = L
N
X
i=1
[α
i
J(i) + γ
i
[
X
j∈N
(i)
(J(i)
− J(j))
2
]
Q
]
(44)where
L
isthenumberoftimesamples,α
i
andγ
i
determinestheweightingfa tors between neighborhood sour es.Q
is theindex of theamplitude of the neighbor-hoodgroup.N (i)
neighborhoodofthesour ei
isdenedasthe9
losestneighbors to thesour e. Therstterminequation(44 )expressessparsitywhilethese ondone favorsfo alsour esdistributions.
TheMRF-basedimagepriorsleadtonon- onvex[5℄andinteger[83℄
program-mingproblemsin omputingtheMAPestimate. Computational osts anbevery
highfor thesemethodssin e althoughthe priorshave omputationallyattra tive
neighborhood stru tures, the posteriors be ome fully oupled through the
likeli-hood term. Furthermore, to deal with non- onvexity and integer programming
issues,someformofdeterministi orsto hasti annealingalgorithmsmustbeused
[35℄.
Con lusion
Theex ellenttimeresolutionofMEGprovidesusauniquewindowonthe
dynam-i s of humanbrain fun tions. Though the limitedspatial resolution remains the
problemforthismodality,adequatemodelingandmodernsignalpro essing
meth-odsproveMEGasadependablefun tionalimagingmodality. Potentialadvan es
inforward modeling in lude better hara terization of theskull,s alp and brain
ods for ombining MEG with other fun tional modalities and exploiting signal
Introdu tion
Theequivalent urrent dipolemodelisdire tlyinterpretable asa urrentelement
restri tedto the orti alsurfa erepresenting apointsour e. However, oneof the
per eived key limitations of this model is that, distributed sour es may not be
adequately represented. This problem was one of the prime motivations to the
development of imaging approa hes. An alternative solutionis to remain within
the model-based framework but to broaden the model to allow parametri
rep-resentations ofdistributed sour es. Themultipolar expansionprovides a natural
framework for generating these models [79, 36 ℄. Multipolar expansions are
de-rived fromspheri al harmoni softhemagneti s alarpotential. Iftheexpansion
point is hosen near the enter of a distributed sour e, then the ontribution of
higher-ordertermswilldroporapidlyasthedistan efromsour estothesensors
in reases. Using this framework we expand the setof sour esto in lude urrent
dipoles and rst-order urrent multipoles. These sour es are able to represent
theeldfromadistributedsour e morea uratelythan by urrent dipolemodel,
thoughstillbenetingfroma ompa t,low-dimensionalform[78 ℄. Multipolar
ex-pansionsofmagneti s alarpotentialsoriginate fromgeneralspheri al harmoni s
solution ofthe Poissonequation.
In this thesis, we proposed an approa h for estimating the spatial extent of
orti al urrent sour es using a hybrid methodology alled Multipole Corti al