Integrating MEG, EEG and fMRI data

MEG/EEG Brain Mapping Course 

HBM2006 

Florence, June 11, 2006 

Integrating MEG, EEG and fMRI data 

Gian Luca Romani 

Institute of Advanced Biomedical Technologies – ITAB 

“G. D’Annunzio University” Foundation 

and 

Department of Clinical Sciences and Bioimaging 

“G. D’Annunzio” University, 

Chieti, ITALY

Outline 

• MEG - EEG – fMRI integration: respective 

advantages and limitations 

• Basic methods 

• Examples of integration in the study of primary 

areas, in cognitive neuroscience, and in the 

clinical field 

– Somatosensory system: pain vs. no-pain activations and 

somatotopic properties 

– Passive listening to sounds from different locations 

– Functional reorganization in cases of malformation of 

cortical development 

• Some critical remarks 

• Conclusions and perspectives

MEG and fMRI equipment 

at ITAB – University of Chieti 

MEG 

fMRI 

165-channel wholehead 

system 

(by end 2006, a new 

500-channel system) 

1.5T Siemens Vision 

(by end 2006, a new 

3T Philips Achieva)

MEG - EEG – fMRI integration: 

respective advantages and 

limitations

fMRI – MEG (EEG) relationship 


Generated by oxygenated 

hemoglobin in the blood 

Functional images are 

directly related to 

structural images 

No source model required 

High spatial resolution (but 

depends on vascularization) 

Low time resolution (limited 

by hemodynamic response) 

Resolution independent of 

source depth 

MEG (EEG) 

Generated by intracellular 

currents 

Functional images are not 

directly related to 

structural images 

Imaging depends on source 

and head models (and 

conductivity, for EEG) 

Low spatial resolution 

(related to source type) 

High time resolution (better 

than 1 ms) 

Poor resolution for deep 

sources

MEG, EEG, fMRI spatio-temporal 

properties 

Spatial resolution (mm) 

10 

8 

6 

4 

2 

EEG 

MEG 


0 

10 -3 10 -2 10 -1 10 0 10 1 10 2 

Temporal resolution (s) 

10 3

Basic methods

How can we integrate MEG, EEG, and fMRI ? 

a. Simple superposition of the equivalent current 

dipoles (ECDs) and the fMRI active areas 

b. Dipole seeding, i.e. utilize fMRI activation 

maps to constrain MEG inverse solutions 

(Ahlfors et al., J Neurophysiol 1999) 

c. Use of distributed sources, and regularization 

of Linear inverse Estimation (LE) for EEG 

and/or MEG data, with inclusion of fMRI 

constraints (Liu et al., PNAS 1998; Babiloni et al., 

Int J Bioelectromagnetism 1999; etc.)


a. Simple superposition of the equivalent current 

dipoles (ECDs) and the fMRI active areas 

• Data from MEG and fMRI are independently 

analyzed 

• The active regions identified by the two 

techniques are merged in a common reference 

system (MRI) using fiducial markers 

• MEG sources (ECDs) are additionally 

transformed into the Talairach space


b. Dipole seeding, i.e. utilize fMRI activation maps 

to constrain MEG inverse solution 

• The brain areas identified as active by fMRI 

during a specific task are used to guide the MEG 

inverse solution 

• It is not possible to simply position the MEG 

source in a fixed location, since fMRI and MEG 

detect different physical phenomena 

• The ECDs are let free to move inside a small 

volume, their orientation is let free as well, and 

a figure of merit is optimized, in order to obtain 

the time course of each source


c. Use of distributed sources and regularization of 

Linear inverse Estimation (LE) 

• The solution of the MEG (EEG) inverse problem 

with a distributed source model requires the 

solution of a linear equation such as: 

Ax – b = 0 

where A is the lead field matrix: 

# rows = # field (or potential) measurements (knowns) 

# columns = # dipole components (unknowns) 

x is a column vector of dipole components 

b is a column vector of measured field components 

(for a review, see: Del Gratta et al., Reports on Progress in Physics, 2001)


d. Use of distributed sources and regularization of 

Linear inverse Estimation (LE) 

• since the number of dipole components (5,000-10,000) is 

much larger than the number of measurements (150-500), 

we must minimize the following expression (cost 

function): 

⎜⎜Ax - b⎜⎜ + λ 2 ⎜⎜Cx ⎜⎜ 

where λ is the so-called regularization parameter and C is 

simply a weight matrix that depends on the head model. 

• By appropriately shaping C we can take into account the 

possible constraints used in various LE analysis methods 

(Minimum Norm, SLORETA, etc.), but also the fMRI 

constraints


• The functional information is used to force the linear inverse 

estimation weighting the minimization procedure with respect to 

the bold activation. 

• Anatomical and functional information can be used to delimit the 

source space to a given ROI. The activity of this specific ROI 

during time can be estimated using a linear inverse estimation. 

• λ (regularization parameter): by choosing the correct value for λ 

an equilibrium between goodness of fit and closeness to the model 

term ⎜⎜Cx ⎜⎜ is achieved. 

A possible method to calculate λ is the L- 

λ ≈ 0.0001 

curve criterion: if the model term ⎜⎜Cx ⎜⎜ 

10 

is plotted vs. the data term ⎜⎜Ax - b⎜⎜ as 

3 

λ ≈ 0.001 

a curve parametrized by λ,an L-shaped 

10 2 

λ ≈ 0.01 

curve is obtained. The corner of L can be 

λ ≈ 0.1 

10 

seen as an equilibrium point between the 

1 

data term and the model term. (see 

10 -3 10 -2 10 

schematic and the reference: Hansen, P.C., SIAM 

-1 

⎜⎜Ax - b⎜⎜ 

Rev. 34, 1992) 

⎜⎜Cx ⎜⎜

Sensorimotor EEG-MEG data can be 

“fused” in a finger movement 

paradigm, using the LE approach 

Regularization of LE permits to infer the time 

course of the source current density and to 

better discriminate the contribution from 

different neural districts in the pre- and postmotor 

period (Babiloni et al., Human Brain 

Mapping 2001)

Integration of fMRI-EEG (MRP) with regularization of LE 

allows better identification of the active areas, if we add the 

fMRI constraints 

Right index finger movement

Somatosensory system: 

pain vs. no-pain activations 

and 

somatotopic properties

Exp.1 - Effect of stimulus intensity from motor 

to weak painful, MEG study I 

(Torquati et al., NeuroReport 2002) 

• Median nerve stimulation in 8 healthy subjects 

• Ten different levels of current intensity. 

Sensory Motor Weak painful 

I 0 I 1 I 2 I 3 I 4 I 5 I 6 

4.2 7.5 11 14 17 21 24 

I 7 I 8 I 9 

31 45 59 mA

Results 

motor strong motor weak painful 

I 1 I 5 

I 9 

SI - increase in amplitude from I1 to I5, then saturation: increasing 

synchronization mechanism of post-synaptic potentials, then “ceiling” effect on the 

involved myelinated Aβ fibers (nociceptive innervation in SI mainly represented by 

much slower Aδ and C unmyelinated fibres) 

SII - decrease for strong motor stimulation: “gating” effect due to the poor 

nociceptive specificity of the electrical stimulation involving possible interference 

in the evoked activation of the “non painful” and “painful” populations; increase for 

weak painful stim.: stronger attention to the stimulus, and amount of activated “non 

painful” neurons decreasing with the “gating” effect 

•Possible existence of two adjacent neuronal pools indistinguishable by MEG

Exp.2 – identification of anterior and posterior 

SII: spatial features, fMRI study 

Activation for non-painful stimuli 

cSI 

cSIIa 

iSIIa 

Ferretti et al., NeuroImage, 2003

Activation for painful stimuli 

cSI 

cSIIa 

iSIIa 

cSIIp 

iSIIp 

Ferretti et al., NeuroImage, 2003

Results 

• Evidence for a spatial segregation 

occurring in SII for neural population 

responding to painful stimulation with 

respect to that activated by non-painful 

stimulation 

• The “painful” population located more 

posterior than the “non-painful” area 

• Small distance between the two areas 

• No information on timing

Exp.3 - Identification of anterior and posterior 

SII: temporal features, MEG study II 

[example of dipole seeding] 

Torquati et al., Neuroimage 2005

Identification of anterior and posterior SII: 

temporal features 

[example of LE (MNE) with fMRI constraints] 

SI 

SIIa 

SIIp

Motor threshold 

Painful threshold 

SI 

iSII 

cSII

Results 

Mean Latencies 

ms 

90 

88 

86 

84 

82 

80 

( * p

Passive listening to sounds from different 

locations (fMRI and MEG) 

(Brunetti et al., Human Brain Mapping 2005) 

This study aimed at highlighting how the auditory system 

identifies the direction a sound is coming from. Activations of 

cortical areas during stimulation with sounds coming from a 

fixed source or, randomly, from five different sources 

located in a horizontal half-space were studied with both 

MEG and fMRI


locations (fMRI results) 

•Two main areas were found to be 

activated during auditory stimulation: the 

bilateral Heschl’s gyrus and the right 

superior temporal gyrus. Larger activation 

was found in the MIXED condition than in 

the LEFT and RIGHT conditions 

•Additionally, the right supramarginal 

gyrus was activated during the MIXED 

condition only 

•A right hemisphere specialization for 

auditory spatial processing seemed 

confirmed, as opposed to the language 

processing in the left hemisphere 

(Gazzaniga et al., 1998) 

•No information about timing was obtained 

from fMRI 

MIXED 

RIGHT 

LEFT

Passive listening to sounds from different locations 

(example of simple superposition and of dipole seeding) 

MEG allowed identification of a bilateral activation in 

the Heschl’s gyrus, but no reliable localizations for the 

superior temporal gyrus and the supramarginal gyrus 

were obtained (sources too close) 

The last two areas were identified by constraining the 

sources in a cube with 4 mm side centered in the 

corresponding fMRI activation.


locations (time course results) 

MEG permitted identification of 

the time sequence of activation for 

the three sources: 

1.Heschl’s gyrus (m. latency 138 ms) 

2.superior temporal gyrus (156 ms) 

3.supramarginal gyrus (162 ms) 

thus suggesting a hierarchical 

structure for the processing of 

sound localization features

Functional reorganization in cases of malformation of 

cortical development (MCD) - a TMS-fMRI study 

• Structural MRI, TMS, and fMRI 

findings (paretic hand movement) 

obtained in 3 cases 

• A–C: Axial reconstructions from 

the T1-weighted 3D data sets 

• D–F: Results of TMS for 

stimulation of the affected and 

contralesional hemispheres, with 

MEPs recorded simultaneously 

from target muscles of both the 

paretic hand (yellow MEPs) and 

the non paretic hand (white 

MEPs). 

• H, J and L: fMRI activations for 

movement of the paretic hand. 

• M–O: illustration of TMS and 

fMRI findings; 

(Staudt et al., J. Neurosurg., 2004)

A MEG coherence study on the same grop 

of patients (Belardinelli et al., 2006, in preparation) 

Coherent brain areas with paretic hand EMG in the β frequency range 

during forearm contraction. The coherence 3D mapping within the 

brain is obtained by means of different spatial filters (Linear 

Constrained Minimum Variance Beamformer and Sloreta)

A further possibility: identification of coupling 

direction 

Different direction evaluators were used: a Granger causality index (Kaminski et al, 

Biological Cybernetics, 2001), as well as two different indexes for the detection of 

coupling direction in chaotic oscillators (Rosenblum et al, Phys. Rev. Lett. 2004) 

have shown values indicating M1 as generator of the activity which comes to the 

cerebellum in case of paretic hand use. The cerebellum coherent activity is not 

present in case of contralateral (healthy) hand use

Pyramidal neurons: only 

1% of these neurons (low 

metabolic demand), if 

synchronously active 

produce 90% of scalp 

EEG/MEG 

Some critical remarks 

Stellate neurons (15% of neocortical 

neurons): strong metabolic/rCBF demand 

but no scalp EEG or MEG signals (closed 

electromagnetic fields) 

Moreover, MEG/EEG and fMRI may be sensitive to different 

phenomena (i.e. rhythm modulations vs. regional neural 

activation), or be affected by different “boundary” conditions

Cognitive fMRI-EEG-MEG data could not be “fused” in a 

working memory (WM) paradigm, since ERD/ERS of rhythmic 

activity can be studied with MEG/EEG much better 

Alpha ERD is evaluated by LE after the 1st (T1) and 2nd (T2) 

second of the delay phase 


Babiloni et al., Clin. Neurophysiol. 2004

Neural plasticity after stroke 

MEG ECDs were always 

identified in both 

hemispheres of all 

patients featuring a 

good clinical function 

recovery; 

BOLD-fMRI activation 

was seldom observed 

(20% of same patients 

only): this could be 

related to the 

significantly impaired 

vasomotor reactivity 

those patients were 

still featuring 

Rossini et al., Brain 2004

Conclusions and perspectives 

• A combined use of MEG, EEG and fMRI may 

provide a unique tool for studying brain 

activity with both high spatial resolution and 

high temporal resolution 

• Integration requires skill and care, but can be 

handled routinely 

• Integration makes sense only if the same kind 

of activity is being studied with the different 

techniques 

• Even in this case, and particularly before 

applying the method to the clinical field, one 

must be aware of some respective limitations 

of the three approaches in this area 

• Integration with TMS!

Integrating MEG, EEG and fMRI data

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?