Estimation of the haemodynamic response to epileptic ... fileEstimation of the haemodynamic response...
-
Upload
hoangkhanh -
Category
Documents
-
view
221 -
download
2
Transcript of Estimation of the haemodynamic response to epileptic ... fileEstimation of the haemodynamic response...
Estimation of the haemodynamic response to epileptic activity in EEG-fMRI data
Marco Filipe Pinto Leite
Dissertação para obtenção do Grau de Mestre em
Engenharia Biomédica
Júri
Presidente: Prof. Fernando Lopes da Silva
Orientadores: Prof. Patrícia Figueiredo
Dr. Alberto Leal
Vogais: Prof. João Sanches
Prof. Sónia Gonçalves
Outubro 2010
“Eigentlich weiß man nur, wenn man wenig weiß;
mit dem Wissen wächst der Zweifel.”
— Johann Wolfgang von Goethe
“Indeed, one only knows when one knows little;
with knowledge grows the doubt”
ii
iii
ACKNOWLEDGEMENTS
Before the presentation of this Thesis I would like to acknowledge Prof. Patrícia Figueiredo for the
most supportive supervision of this work, for all the time invested and all the clarity granted to this
Thesis. For the motivation and inspiration provided I am also thankful to her.
I would like to thank Dr. Alberto Leal for the insightful morning meetings and for the providence
of the patient’s data, without which the creation of this Thesis would not be possible.
I am grateful to Prof. João Sanches for the conceptual contributions and for the initial momentum
that enabled this Thesis to originate.
I would also like to thank Prof. Sónia Gonçalves for the contributions on the EEG artifact
interpretation, Prof. Lopes da Silva for the contributions on the interpretation of the neurological
bases for EEG and BOLD signal, and Prof. Louis Lemieux for the insights on the methodology
validation.
I would like to express my gratitude for the most enjoyable work environment that my colleagues
at IST provided during the conception of this work. I would like to give a word of recognition to Inês
and all my friends outside IST for keeping me sane and creative.
For all the support, comprehension and kindness, I thank my family: my mother, my father and
my step-mother, my sister and my brother and also my grandparents. They are one of the most
valuable sources of inspiration.
iv
ABSTRACT
Electroencephalography correlated functional Magnetic Resonance Imaging (EEG-fMRI) is a multi-
modal imaging technique with growing application in the clinical evaluation of epilepsy. In this
work, a new approach for simultaneously recorded EEG-fMRI data integration of ictal events in
epilepsy is proposed. To improve the fMRI artefact correction of EEG data, a fully automated
algorithm was developed based on existing tools. Independent component analysis (ICA)
decomposition was performed on the corrected EEG data and multiple model based metrics were
applied to the resulting time courses. These were used to predict the Blood Oxygen Level Dependent
(BOLD) fMRI data using a General Linear Model (GLM) approach. When compared with the
conventional fMRI data analysis based on square waveform descriptions of seizure activity, clinically
valid and more significant activations were found with the method proposed here, for the four
patients studied. In particular, frequency-weighted EEG metrics were found to best describe the
BOLD signal, in support of previous theoretical and experimental work. In general, the results were
consistent with the neurophysiologist’s expectation, but further validation using more direct
measurements of seizure activity is necessary. A detailed study on the hemodynamic response
function (HRF) to the EEG metrics was performed for one patient. The HRFs estimated were broader
than the canonical HRF and the distributions of its delay and dispersion were mapped throughout the
patient’s brain. In summary, this work contributed to a better understanding and improved
integration of EEG-fMRI data collected during epileptic seizures.
v
RESUMO
A Imagem funcional por Ressonância Magnética correlacionada com Electroencefalografia (EEG-
fMRI) é uma técnica multi-modal de imagiologia com crescente aplicação clínica na avaliação da
epilepsia. Neste trabalho é proposta uma nova abordagem à integração de dados de EEG-fMRI,
registados em simultâneo, em eventos ictais de epilepsia. Para melhorar a correcção dos artefactos de
fMRI nos dados de EEG, foi desenvolvido um algoritmo totalmente automático baseado em
ferramentas já existentes. Foi efectuada ICA sobre os dados de EEG e foram aplicadas múltiplas
métricas, baseadas em modelos, aos sinais temporais resultantes. Estes foram utilizados para prever o
sinal BOLD fMRI utilizando uma abordagem de GLM. Quando comparada com a análise
convencional dos dados de fMRI, baseada numa descrição em onda quadrada da actividade da crise,
com o método proposto, foram obtidas activações clinicamente válidas e mais significativas para os
quatro pacientes estudados. . Em particular, as métricas ponderadas na frequência do EEG obtiveram
um melhor desempenho, em concordância com trabalhos teóricos e experimentais anteriores. Em
geral, os resultados foram consistentes com as expectativas do neurofisiologista, mas é necessária
validação extra utilizando medidas mais directas da actividade da crise. Foi efectuado para um dos
pacientes um estudo detalhado da HRF às métricas do EEG. As HRFs estimadas foram mais dispersas
temporalmente do que a HRF canónica e a distribuição do seu atraso e a dispersão foram mapeadas
ao longo do cérebro do paciente. Sumariamente, este trabalho contribuiu para uma melhor
compreensão e integração de dados de EEG-fMRI recolhidos durante crises epilépticas.
vi
TABLE OF CONTENTS
Acknowledgements ......................................................................................................... iii
Abstract .............................................................................................................................. iv
Table of Contents .............................................................................................................. vi
List of Tables ................................................................................................................... viii
List of Figures .................................................................................................................... ix
List of acronyms ............................................................................................................. xiii
Chapter 1 Introduction ................................................................................................ 1
1.1 EEG-fMRI ................................................................................................................... 1
1.1.1 EEG signal ...................................................................................................................................... 1
1.1.2 fMRI and the BOLD signal ........................................................................................................... 3
1.1.3 EEG and fMRI integration ........................................................................................................... 7
1.2 Epilepsy ...................................................................................................................... 9
1.3 Thesis rationale ....................................................................................................... 11
Chapter 2 Methods ..................................................................................................... 12
2.1 Data acquisition ...................................................................................................... 12
2.1.1 Subjects ......................................................................................................................................... 12
2.1.2 EEG acquisition ........................................................................................................................... 14
2.1.3 fMRI acquisition .......................................................................................................................... 14
2.2 EEG analysis ............................................................................................................ 14
2.2.1 Pre-processing ............................................................................................................................. 15
2.2.2 Data decomposition .................................................................................................................... 17
2.3 fMRI analysis ........................................................................................................... 18
2.3.1 General processing ...................................................................................................................... 19
2.3.2 Testing different EEG metrics .................................................................................................... 19
2.4 HRF study ................................................................................................................ 20
2.4.1 ROI based analysis ...................................................................................................................... 20
2.4.2 Pixel-by-pixel analysis ................................................................................................................ 22
Chapter 3 Results ....................................................................................................... 25
vii
3.1 EEG pre-processing analysis ................................................................................. 25
3.1.1 Artefact correction ....................................................................................................................... 25
3.1.2 Data decomposition .................................................................................................................... 27
3.2 EEG-fMRI analysis.................................................................................................. 29
3.2.1 Subject RR ..................................................................................................................................... 29
3.2.2 Subject JQ ..................................................................................................................................... 37
3.2.3 Subject ABC .................................................................................................................................. 42
3.2.4 Subject JB ...................................................................................................................................... 43
3.2.5 Subject GM ................................................................................................................................... 46
3.2.6 Summary ...................................................................................................................................... 48
3.3 HRF study ................................................................................................................ 49
3.3.1 ROI based analysis ...................................................................................................................... 49
3.3.2 Pixel-by-pixel analysis ................................................................................................................ 53
Chapter 4 Discussion ................................................................................................. 60
4.1 Summary .................................................................................................................. 60
4.2 EEG processing and fMRI integration ................................................................. 60
4.3 HRF study ................................................................................................................ 63
4.4 Conclusion ............................................................................................................... 64
References .......................................................................................................................... 66
Annex ................................................................................................................................. 70
viii
LIST OF TABLES
Table 1.1: Typical physiological prior values for the biophysical parameters....................................... 6
Table 2..1: Clinical, anatomical and neurophysiologic characteristics of the subjects. ....................... 13
Table 2.2: fMRI sequence parameters for each patient............................................................................ 14
ix
LIST OF FIGURES
FIGURE 1.1: Examples of common HRF shapes used as impulsive responses. In blue an HRF
presenting a peak at 4 s and a post-stimulus undershoot. In green an HRF modelled as a gamma
distribution function, with a peak at 4.5 s, broader dispersion and no post-stimulus undershoot. ... 5
FIGURE 1.2: Balloon Model scheme: z is the neural input, which induces a vasodilatory activity
dependent signal s that increases the flow f. Altered blood flow causes changes in volume and
deoxyhaemoglobin concentration (v and q). These two haemodynamic states enter the output
nonlinearity λ and produce the observed BOLD. θ = {κ, γ, τ, α, ρ, V0} are the biophysical
parameters to be estimated, k1 k2 and k3 are function of the fMRI sequence parameters. ( may be a
different function if other imaging method would to be employed) ...................................................... 6
FIGURE 1.3: EEG-fMRI integration: (i) integration through prediction, (ii) integration through
constraints and (iii) integration through fusion with forward models. (Figure in (Kilner, et al.
2005)) ................................................................................................................................................................ 8
FIGURE 2.1: Three half-cosines FLOBS HRF parameterization. ................................................................... 23
FIGURE 3.1:ERP images (top) and corresponding average time courses (bottom) of an example channel
(F7) triggered on the fMRI slice events: using fixed time slice triggers (left) and using the slice
timing identification algorithm developed here (right). ......................................................................... 26
FIGURE 3.2: ERP images (top) and corresponding average time courses (bottom) of an example
channel (F7) after fMRI gradient corrections and triggered on the aligned slice events: using the
Neuroscan procedure (left) and the EEGLAB procedure developed here (right). ............................. 26
FIGURE 3.3: ERP images of an example channel (F7) triggered on EKG QRS events. On the left is the
image before balistocardigrphic corrections. On the right is the image after balistocardiographic
artefact correction procedure. Below are the plots of the mean time courses of both images. ......... 27
FIGURE 3.4: IC maps obtained for sequence 3, patient RR. The components are ordered by decreasing
explained variance (reference channel: FCZ). .......................................................................................... 28
FIGURE 3.5: Detail of the 5th IC. The red marker corresponds to the channel with highest absolute
weight on this component, F3. (reference channel: FCZ) ....................................................................... 28
FIGURE 3.6: Spectrograms of EEG channel F3 (top) and ICA’s 5th component (bottom), obtained for
sequence 3 of patient RR. Channel F3 was the channel that contributed most to the 5th IC. The
boxes in red represent the periods marked as ictal events by the neurophysiologist. Spectral
changes associated with the ictal events are visible in both spectrograms, although with a time
shift. These correlations are clearer in the spectrogram obtained for the 5th IC. ................................. 29
x
FIGURE 3.7: Scalp maps of the ICs for sequences 2 and 3 of patient RR (reference channel: FCZ) (top
and centre, respectively). In the bottom image, the intensity corresponds to the absolute value of
the correlation of the normalized IC weights of both sequences........................................................... 30
FIGURE 3.8: Activation map volumes obtained for each metric and each component of sequence 2,
patient RR. For each IC, the volumes are plotted in descending order for the different metrics. .... 31
FIGURE 3.9: Activation map volumes obtained for each metric and each component of sequence 2,
patient RR. For each metric, the volumes are plotted in descending order for the different ICs. .... 31
FIGURE 3.10: Activation map volumes obtained for each metric and each component of sequence 3,
patient RR. For each IC, the volumes are plotted in descending order for the different metrics. .... 32
FIGURE 3.11: Activation map volumes obtained for each metric and each component of sequence 3,
patient RR. For each metric, the volumes are plotted in descending order for the different ICs. .... 32
FIGURE 3.12: Activation Z statistic maps (top) and corresponding average time courses, partial and
full model fits (arbitrary units), plotted as a function of the volume number (bottom), obtained for
sequence 2, patient RR. ................................................................................................................................ 34
FIGURE 3.13: Activation Z statistic maps (top) and corresponding average time courses, partial and
full model fits (arbitrary units), plotted as a function of the volume number (bottom), obtained for
sequence 3, patient RR. ................................................................................................................................ 35
FIGURE 3.14: EEG scalp maps (reference channel: FCZ) (left) and fMRI activation Z statistic maps of
central slices (right), for the EEG ICs of sequences 1, 4, 5 and 6 that exhibited the greatest spatial
correlation with components 16 and 10 of sequences 2 and 3. The EEG metrics presented are the
ones yielding the largest activation maps for the respective component. ........................................... 37
FIGURE 3.15: EEG scalp maps (left) and spectrogram (right) for the 14th IC from the EEG of subject JQ.
). The boxes in red represent the periods marked as ictal events by the neurophysiologist. ............ 38
FIGURE 3.16: EEG time course of channel 25(CP4) (top) and 14th IC (bottom). The abnormally high
amplitudes present in channel 25 time course indicate the artefactual character of the channel, thus
removed from the data. 14th IC amplitude profile accompanies that of the channel 25. .................... 38
FIGURE 3.17: Activation map volumes obtained for each metric and each component of patient JQ. For
each IC, the volumes are plotted in descending order for the different metrics. ................................ 39
FIGURE 3.18: Activation Z statistic maps (top) and corresponding average time courses, partial and
full model fits (arbitrary units), plotted as a function of the volume number (bottom), obtained for
patient JQ with the neurophysiologist’s regressor. ................................................................................. 40
FIGURE 3.19: EEG scalp maps (reference channel: FCZ) and activation Z statistic maps for the ICs with
compatible fMRI activation maps with the neurophysiologist’s regressor. The EEG metric shown is
the one yielding the largest activation map for that component. .......................................................... 41
xi
FIGURE 3.20: 12th IC spectrogram (top) and corresponding regressor and mean BOLD time course over
the activation map yielded in fMRI analysis (bottom). The boxes in red represent the periods
marked as ictal events by the neurophysiologist. .................................................................................... 41
FIGURE 3.21: Spectrograms for the ICs that correlated with the neurophysiologist’s ictal markers,
using Neuroscan procedure pre-processing (top) and EEGLAB procedure (bottom). ............................. 42
FIGURE 3.22: EEG time courses form channel F8 (top) and channel FC4 (bottom) of subject ABC using
as pre-processing method the Neuroscan procedure. ................................................................................. 43
FIGURE 3.23: Relative mean displacement yielded by MCFLIRT (black) and neurophysiologist’s ictal
marker (red) for patient JB ictal sequence. ................................................................................................ 43
FIGURE 3.24: Activation map volumes obtained for each metric and each component of patient JB. For
each IC, the volumes are plotted in descending order for the different metrics. ................................ 44
FIGURE 3.25: EEG scalp maps (reference channel: PZ) (left) and activation Z statistic maps for the IC
EEG metrics with largest fMRI activation maps (right). It is also presented the activation map
obtained with the neurophysiologist’s regressor (top). .......................................................................... 45
FIGURE 3.26: Relative mean displacement yielded by MCFLIRT (black) along with the EEG metric
regressors used in fMRI analysis that yielded the largest activation maps. ........................................ 46
FIGURE 3.27: Scalp maps for the ICs that yielded significant activation maps in fMRI analysis for
patient GM (reference channel: CZ). ......................................................................................................... 46
FIGURE 3.28: With reference to patient GM. For each metric it is plotted, in descending order and as a
function of the IC used, the volume of the activation maps obtained on the fMRI analysis. ............ 47
FIGURE 3.29: Activation maps for patient GM, sequence 1, using the neurophysiologist’s regressor
(top) and the regressor obtained using RMSF metric on the 4th IC (bottom). ...................................... 47
FIGURE 3.30: HRFs estimated by Biophysical Model (blue), IIR filter (green) and FIR filter (cyan), for
sequence 2 (left) and sequence 3 (right) using the whole activation map as ROI. .............................. 49
FIGURE 3.31: BOLD time courses obtained using the HRFs presented in Figure 4.30: Balloon Model
(blue), IIR filter (green) and FIR filter (cyan). The dashed grey is the ROI BOLD mean time course.
Results displayed for sequence 2 (top) and sequence 3(bottom). .......................................................... 50
FIGURE 3.32: HRFs estimated using IIR model (top), FIR model (centre) and Biophysical Model
(bottom), for sequences 2 (left) and 3 (right). The ROIs were the entire activation map, hamartoma,
left hippocampus, left occipital lobe, left frontal lobe and left parietal lobe. ....................................... 52
FIGURE 3.33: HRF basis set, derived from IIR filter model estimation for the whole activation map of
sequence 3. ‚Centre‛ HRF (blue), time derivative (green) and dispersion derivative (red). ............ 53
FIGURE 3.34: Distribution of the normalized β2 and β3 across the brain (top). Example HRFs,
translating the colormap of the image (bottom left). Scatter plot on which each dot represents a
xii
voxel with the coordinates being its normalized β2 and β3 (bottom right). Dots in black correspond
to frontal lobe voxels, in red to occipital lobe, cyan to parietal lobe, green to hippocampus and blue
to hamartoma. ............................................................................................................................................... 54
FIGURE 3.35: Distribution of the normalized β2 and β3 across the brain (top). Example HRFs,
translating the colormap of the image (bottom left). Scatter plot on which each dot represents a
voxel with the coordinates being its adjusted β2 and β3 (bottom right). Dots in black correspond to
frontal lobe voxels, in red to occipital lobe, cyan to parietal lobe, green to hippocampus and blue to
hamartoma. ................................................................................................................................................... 55
FIGURE 3.36: FLOBS sample HRFs (left) and resulting set of basis functions (right) for parameters
m1∈[0-0s], m2∈[4-12s], m3∈[20-30s], m4∈[0-0s] and c∈[0-0]. ................................................................ 56
FIGURE 3.37: FLOBS basis functions, with symmetric ‘dispersion’ base function, used for results
display. .......................................................................................................................................................... 56
FIGURE 3.38: Distribution of the normalized β2 and β3 across the brain (top). Example HRFs,
translating the colormap of the image (bottom left). Scatter plot on which each dot represents a
voxel with the coordinates being its adjusted β2 and β3 (bottom right). Dots in black correspond to
frontal lobe voxels, in red to occipital lobe, cyan to parietal lobe, green to hippocampus and blue to
hamartoma. ................................................................................................................................................... 57
FIGURE 3.39 Distribution of the normalized β2 and β3 across the brain (top). Example HRFs,
translating the colormap of the image (bottom left). Scatter plot on which each dot represents a
voxel with the coordinates being its adjusted β2 and β3 (bottom right). Dots in black correspond to
frontal lobe voxels, in red to occipital lobe, cyan to parietal lobe, green to hippocampus and blue to
hamartoma. ................................................................................................................................................... 58
FIGURE 3.40: Volume of the activation maps (number of voxels) as a function of Tup and Tmax for
sequence 2 (left) and sequence 3 (right). The largest activation for sequence 2 (5956 voxels) is
obtained with Tup 11 s and Tmax 6 s and for sequence 3 2 (6638 voxels) with Tup 17 s and Tmax 2
s. ...................................................................................................................................................................... 59
FIGURE A.1: Electrode scalp map identifying the channel names with their locations. ........................... 70
xiii
LIST OF ACRONYMS
ADC Analog-to-Digital Converter
BOLD Blood Oxygen Level Dependent (signal)
DC Direct Current
dHb deoxyhaemoglobin
EEG Electroencephalogram
EKG Electrocardiogram
ERP Event Related Potential
EPI Echo Planar Imaging
FIR Finite Impulse Response
fMRI Functional Magnetic Resonance Imaging
FWHM Full Width Half Maximum
GLM General Linear Model
Hb Haemoglobin
HRF Haemodynamic Response Function
IC Independent Component
ICA Independent Component Analysis
IIR Infinite Impulse Response
MF Mean Frequency
MRI Magnetic Resonance Imaging
MR Magnetic Resonance
RMS Root Mean Square
RMSF Root Mean Square Frequency
ROI Region Of Interest
SNR Signal to Noise Ratio
TP Total Power
TR Repetition Time
uRMSF un-normalized Root Mean Square Frequency
uMF un-normalized Mean Frequency
1
Chapter 1
INTRODUCTION
This work is focused on the integration of two techniques central to the measurement of brain
activity, the electro-encephalography (EEG) and the functional Magnetic Resonance Imaging (fMRI),
in the context of the characterization of ictal epileptic events. A special attention was given to the
haemodynamic response to specific EEG features. This chapter will be dedicated to a brief description
of these techniques and their physiological origin and correlates, and also their clinical application in
the field of epilepsy, including an overview of the literature that inspired and influenced this work
the most.
1.1 EEG-FMRI
The EEG is a long established technique for measuring brain activity. The electrical manifestations
of the brain were discovered more than a century ago, but only in the 1930s, with the demonstration
that the brain electrical activity could be recorded at scalp level, progressive use was given to the EEG
in the analysis of brain function (reviewed in Lopes da Silva 2010). As to the fMRI, its introduction in
the early 1990s proclaimed a new era in imaging neuroscience, allowing for the identification of
haemodynamic correlates anywhere in the brain (Friston 2010). In 1993 John Ives and colleagues
pioneered the simultaneous recording of EEG and fMRI impelled by the desire of mapping the
epileptic activity in the brain (Ives, et al. 1993).
1.1.1 EEG signal
To help the interpretability of the main results of this Thesis, a brief explanation of the EEG signal
origin, along with an introduction to Independent Component Analysis (ICA) and its application to
EEG data, will follow.
EEG origin
The EEG is the scalp recording of electrical rhythms produced by the brain. These potentials are
originated by transmembrane currents at neuronal level. Two main forms of neuronal activation may
be distinguished: the fast depolarization of neuronal membranes (~2 ms) and the slower changes in
membrane potential due to synaptic activations. The latter may be further distinguished into
excitatory and inhibitory post-synaptic potentials. In general terms, at the level of an excitatory
synaptic potential, the excited neuron intakes positive ions (e.g. Na+), which correspond to a current
sink. On the contrary, the inhibitory synaptic potentials involve the intake of negative ions (e.g. Cl -) or
2
outtake of positive ions (e.g. K+), which correspond to a current source. In both cases, a passive
counter current is generated re-establishing the medium mean (constant) charge, i.e., there is another
associated current source or sink distributed by the neuron soma. (reviewed in Lopes da Silva 2010)
Consequently there is a dipolar, source-sink, configuration in the extra cellular medium caused by
synaptic activity. Considering the palisade arrangement of the cortex pyramidal neurons, with their
apical dendrites aligned perpendicularly to the cortical surface, when activated with a certain level of
synchrony, coherent electric dipolar configurations are generated and susceptible of being recorded
by electrodes placed at a relatively close distance. This way, pyramidal cortex neurons, near the scalp,
with synchronous activity, are the major contributors for the EEG signal. (reviewed in Lopes da Silva
2010)
ICA and EEG
One of the major problems in EEG interpretation is the fact that, in general, this signal is a complex
mixture of several electrical potentials originated in distinct parts of the brain, making it difficult to
characterize the contribution from each particular brain region. The relatively recent introduction of
powerful techniques of blind source separation, such as ICA (Comon 1994), opened new possibilities
for significant improvements in the approach to this problem.
ICA is a method for blind source separation in which the data is modelled as a mixture (linear
transformation) of statistically independent signals:
(1.1)
where x is the observed data, A the mixing matrix and s the sources signal, being the latter two
unknown. The goal is to find the inverse of matrix A, W; this is done by maximizing the statistical
independence of the resulting s estimates:
(1.2)
From the central limit theorem, a classical result in probability theory, comes that, the distribution
of a sum of independent random variables tends towards a Gaussian distribution. This is the base
towards the minimization of the ‘non-gaussianity’ performed in ICA, which, in popular algorithms
(e.g. fastica and infomax), is achieved by the minimization of the mutual information of the s vectors
(Comon 1994, Hyvärinen and Oja 2000).
As W is constant along time, when applying ICA, a static character for the sources of interest is
being assumed. In EEG applications, ICA is a popular technique for the removal of muscular activity
or eye movement (Vigário 1997) and, because of the large amplitude of epileptic interictal activity and
the fact that its sources can generally be considered static, ICA has also proved to be useful in the
3
separation and identification of such activity (Marques, et al. 2009, Kobayashi, et al. 1999, Urrestarazu,
et al. 2006).
1.1.2 fMRI and the BOLD signal
In this section the basic principles of the BOLD fMRI signal will be described, starting with a
general introduction to the magnetic resonance (MR) signal and then detail into the BOLD signal. The
basics of fMRI signal modelling will also be introduced.
The MRI s ignal
In magnetic resonance imaging (MRI), the signal of interest being measured arises from the nuclei
of the tissue’s hydrogen atoms (i.e. single protons). When subjected to an external magnetic field, the
proton’s spin tends to align with it, acquiring a lower energy state. This is the case of the tissue inside
the MR scanner bore: spins will be distributed into two energy levels, parallel or anti-parallel with the
external field B0, with a small majority in the former state, given the usual field strengths and the
tissue’s temperature. At a macroscopic level, this gives rise to a net magnetization M that is parallel to
B0. Measurements of M will be the source of information for the imaging techniques.
The measurement of M is possible due to the excitability of the nuclei if exposed to an
electromagnetic wave with a specific frequency, entitled Larmor frequency (hence the term
‘resonance’). The mean magnetization M is therefore tilted and rotates, i.e. precesses, around the
magnetic field vector B0 with frequency equal to the Larmor’s frequency, which for hydrogen nuclei is
about 43.6 MHz/T, corresponding for example in a 1.5 T scanner to approximately 64 MHz (in the
range of radio frequencies). The tilted magnetization gradually returns to its equilibrium, a process
called relaxation, emitting electromagnetic waves with the Larmor frequency in the process.
Relaxation consists in two simultaneous processes, resulting in separate relaxation profiles for the
longitudinal (i.e. parallel to the magnetic field B0) and the transversal (i.e. perpendicular to B0)
components of M. The longitudinal relaxation, also called spin-lattice relaxation, involves the transfer
of energy from the excited spins to the surrounding environment (the lattice) and results in an
exponential decay to the equilibrium magnetization, with its time constant defined as T1. The
transversal relaxation, also called spin-spin relaxation, results from the interaction of the different
hydrogen nuclei and exhibits also an exponential decay to the equilibrium, but with a different time
constant defined as T2. Yet another mechanism of loss of transversal magnetization is caused by the
local inhomogenities of the magnetic field, which causes the Larmor frequency to slightly vary from
nucleus to nucleus, precessing at slightly different frequencies, thus resulting in a loss of coherence
between the different nuclei spins. This process is central to functional imaging as will be further
explored in BOLD signal explanation, it also presents an exponential decay profile with time constant
defined as T2*, which is smaller than the T2 previously referred. In MRI, employing different
4
acquisition sequences, it is possible to obtain images pondered in the referred parameters, T1, T2 and
T2*, across the brain. (Deichmann, Nöth and Weiskopf 2010, Logothetis and Wandell 2004)
The BOLD signal
The acronym BOLD stands for ‘blood-oxygen-level dependent’ (signal), heralding the principle
underneath the physiologic origin of this signal. The BOLD contrast mechanism relies on the
measurement of the variations of T2* along time and across the brain. As mentioned, T2* is associated
with the local inhomogenities of the magnetic field and these will be indirectly affected by neural
activity. Deoxyhaemoglobin (dHB) is paramagnetic and therefore influences the magnetic
susceptibility of the surrounding tissue reducing T2*, unlike haemoglobin (Hb) which is diamagnetic
(Pauling and Coryell 1936). Therefore images pondered in T2* will vary along time with blood
oxygenation demand and blood flow of the local tissues (Ogawa, et al. 1190). BOLD signal is, this
way, an indirect measure of neural activity; the causal mechanisms that translate neural activity into
BOLD signal are a central question, yet to be fully closed, in the interpretation of fMRI images
(Deichmann, Nöth and Weiskopf 2010). Some studies addressed the question of which aspect or
aspects of the neural activity BOLD best reflects, identifying the neuronal input to a relevant area and
its processing there, rather than the long-range signals transmitted by action potentials to other
regions of the brain, as primary sources for BOLD. (Logothetis, et al. 2001, Shmmuel, et al. 2006)
The coupling between the neural activity and the associated haemodynamic changes is a relevant
issue that was addressed in the present Thesis. The modelling of this process has been the object of
various studies and, though nonlinearities in the process have been extensively described (Glover
1999, Logotetis 2002), the most common approach is to model the haemodynamic response as a linear
time invariant system, with the impulsive response including the most relevant features of the
experimentally observed haemodynamic responses: the time delay between the stimulus and the
haemodynamic response function (HRF), the temporal smoothing of the stimulus and a possible post-
stimulus undershoot. Examples are shown in Fig. 1.1.
5
FIGURE 1.1: Examples of common HRF shapes used as impulsive responses . In blue an HRF
presenting a peak at 4 s and a post-stimulus undershoot. In green an HRF modelled as a gamma
distribution function, with a peak at 4.5 s, broader dispersion and no post-stimulus undershoot.
This impulsive response is then convolved with the paradigm describing the time series of
neuronal activity (elicited by a stimulus or task, or related with some kind of spontaneous brain
activity) and introduced as explanatory variable (regressor) for the fMRI data analysis.
More sophisticated modelling has been done on the transduction of neural activity into BOLD
signal. One prominent example is the Balloon Model firstly introduced by Buxton and colleagues
(Buxton, Wong and Frank 1998), complemented with the flow dynamics and implemented by Friston
and colleagues (Friston, Mechelli, et al. 2000) in order to achieve a better description of BOLD
dynamics. A schematic representation of the Balloon Model, including its mathematical formulation,
is presented in Figure 1.2; a brief description is included on its legend. The estimation of such
biophysical model is usually performed resorting to a Bayesian framework, for physiological priors
are easily obtained and, if not introduced, the estimation problem would be badly conditioned for the
levels of SNR typically available in fMRI. A set of typical physiological values of the system
parameters, used as prior information in model estimation, is presented in Table 1.1. (Friston, Penny,
et al. 2002)
5 10 15 20
0
0.5
1
Time (s)A
mplit
ude (
std
units)
6
FIGURE 1.2: Balloon Model scheme: z is the neural input, which induces a vasodilatory activity
dependent signal s that increases the flow f. Altered blood flow causes changes in volume and
deoxyhaemoglobin concentration (v and q). These two haemodynamic states enter the output
nonlinearity λ and produce the observed BOLD. θ = {κ, γ, τ, α, ρ, V0} are the biophysical parameters
to be estimated, k1 k2 and k3 are function of the fMRI sequence parameters. ( may be a different
function if other imaging method would to be employed)
TABLE 1.1: TYPICAL PHYSIOLOGICAL PRIOR VALUES FOR THE BIOPHYSICAL PARAMETERS
Parameter Description Prior Mean Prior Covariance
κ Rate of signal decay 0.65 s-1 0.015
γ Rate of flow dependent elimination 0.41 s-1 0.002
τ Haemodynamic transit time 0.98 s 0.0568
α Grubb’s exponent 0.32 0.0015
ρ Resting oxygen extraction fraction 0.34 0.0024
fMRI data model l ing
Functional studies are usually analyzed resorting to classical inference on a statistical parametric
map. Statistical parametric maps are spatial random processes that, under the null hypothesis, follow
a known probability density function, usually student’s T or F distributions. A general linear model
(GLM) is used to estimate some parameters that could explain the spatially continuous data and
Gaussian random field (GRF) theory provides a method for correcting p values for false positive rates
7
derived from the multiple comparisons made throughout the brain. (Friston, Holmes, et al. 1995,
Worsley, Evans, et al. 1996)
The GLM may be expressed as:
(1.3)
with y being the observed response, function of a linear combination of explanatory variables
(regressors) of the matrix X plus an error term (assumed white Gaussian noise). X is often designated
‚design matrix‛ and its columns contain the designed effects of the experiment (e.g. stimulus
convolved with the HRF) and the confounds (e.g. movement parameters). The relative contribution of
each column is incorporated in the parameters β and, for each voxel, a β vector is estimated. Statistical
inference about the parameters consists on determining whether a specific linear combination (T-test)
or any of them (F-test), explain the data or not, i.e. one rejects the null hypothesis or not. (Friston,
Models of Brain Function in Neuroimaging 2005)
1.1.3 EEG and fMRI integration
The simultaneous acquisition of EEG and fMRI data poses a number of specific technical and
conceptual problems.
Firstly regarding the technical challenges of EEG-fMRI, a number of artefactual components
appear in EEG data specific to the MR environment acquisitions. The most relevant ones are
originated by the echo planar imaging (EPI) sequence slice gradients, which induce currents in the
EEG inter-electrode loops and completely preclude the true EEG activity. Therefore, these artefacts
must be corrected through adequate data processing methods. Also, small movements of the
electrodes, under the static magnetic field B0, induce changes in the magnetic flux on the inter-
electrode loops, consequently, small currents appear associated with balistocardiographic effects and
may also be corrected through adequate processing. Subjects’ movement artefacts also appear
intensified under the magnetic field B0, however, as they do not possess a reproducible shape, a
posteriori correction is not achievable, and these may only be diminished instrumentally (minimizing
the inter-electrode loops). (Gonçalves, et al. 2007)
Conceptually, the integration of EEG and fMRI data is a challenge far from being disclosed.
Approaches to the integration of both modalities may be classified into three categories: integration
through prediction, integration through constraints and integration through fusion. These are
represented in Figure 1.3.
8
FIGURE 1.3: EEG-fMRI integration: (i) integration through prediction, (ii) integration through
constraints and (iii) integration through fusion with forward models . (Figure in (Kilner, et al. 2005))
Integration through prediction uses the temporally resolved EEG signal as a predictor of
haemodynamic changes in the spatially resolved fMRI data. This is the case of the majority of studies
involving EEG-fMRI, approaching matters such as alpha rhythm BOLD modulations (Lafus, et al.
2003) or epileptic correlated BOLD changes (Salek-Haddadi, et al. 2006).
Integration through constraints makes use of the spatial resolution of focal activations in fMRI to
constrain the distribution of equivalent dipole location in EEG source estimation (Phillips, Rugg and
Friston 2002).
Forward models explaining the relation between the neural activity of interest and the
haemodynamics and electrical manifestations measured are crucial to a symmetrical approach to the
interpretation of EEG-fMRI data. Perhaps the simplest approach to this was made by Kilner and
colleagues (Kilner, et al. 2005). They proposed a heuristic explanation for EEG and BOLD prediction
based on neural activation. This heuristic specifies activations as accelerations on neural dynamics
and, this way, BOLD activations are accompanied by an increase in the ‘average’ frequency of EEG
neuronal activity, where average is defined in a root mean square sense. Conversely, decreases in the
mean frequency of EEG activity would correspond to decreases in BOLD. This results from the
following equation:
(1.4)
where is the activation parameter, and b correspond to the mean BOLD signal for activated and
base neuronal activity, and are the power spectrum of the EEG for activated and base neuronal
activity.
An experimental work, partially motivated by the heuristic model, was performed by Rosa et al.
(Rosa, et al. 2010) and will be central to the present Thesis. In her work, a group of EEG spectral
measures were tested as transfer functions between the EEG and the BOLD signal in a visual
stimulation experience. Some level of detail will be given to these spectral measures, as they will be
further employed on this study.
9
The first model applied was inspired on the result of Wan et al. (Wan, et al. 2006) that assumes the
neurovascular coupling as a power transducer; this way, the corresponding feature to be extracted
from the EEG time series spectral profile is the total power:
(1.5)
where P is the power associated to frequency f and time t, and is the resulting metric time course.
The second model followed Goense and Logothetis work (Goense and Logothetis 2008), assuming
that BOLD is best explained by a linear combination of activity in different frequency bands, for each
band resulted an explanatory time course for BOLD:
(1.6)
where fbmin and fbmax correspond to the band frequency limits.
Four other metrics, based on the heuristic model, were applied:
(1.7)
(1.8)
(1.9)
(1.10)
where (f,t) is the frequency normalized power for frequency f and time t. The differences between
metrics account for the non-linearities and for the power normalization effects on the quality of the
BOLD prediction. The physiological principle underlying the frequency-based metrics is the energy
dissipative character of the neuronal transmembrane currents, from which faster dynamics represent
higher energy dissipation and subsequent increases in BOLD signal.
The application of these metrics to the EEG-fMRI data of visual activation collected from 3 subjects
indicated that the normalized heuristic based models superiorly predicted the BOLD time courses.
1.2 EPILEPSY
As defined by the International League Against Epilepsy and the International Bureau for
Epilepsy ‚an epileptic seizure is a transient occurrence of signs and/or symptoms due to abnormal
excessive or synchronous neuronal activity in the brain‛ and ‚Epilepsy is a disorder of the brain
characterized by an enduring predisposition to generate epileptic seizures and by the neurobiologic,
cognitive, psychological, and social consequences of this condition‛, still stating that ‚the definition of
epilepsy requires the occurrence of at least one epileptic seizure‛ (Fisher, et al. 2005).
Epileptic activity can be categorised into interictal or ictal activity. The ictal activity, or seizure,
corresponds to a symptomatic transient state, with a clear start and ending. The termination may not
10
be as clear as the onset, for often the post ictal symptoms preclude the ending of the crisis. These
timings may be determined either behaviourally or on EEG grounds, which may not always coincide.
Interictal activity corresponds to sporadic and asymptomatic abnormal neural activity or
synchronisation. (Fisher, et al. 2005)
The EEG has been, over the years, the tool of excellence for diagnose and characterization of
epilepsy. With the advent of fMRI, and specifically of the simultaneous acquisition of EEG-fMRI,
studies of epileptic correlated BOLD activity proliferated (Salek-Haddadi, et al. 2006, Hoffman, et al.
2000, LeVan and Gotman 2009). Because of its clinical potential in the localization of the sources of
epileptiform activity, synchronous EEG-fMRI is becoming a more common tool in the study of
epilepsy.
Waites and colleagues used a non parametric approach to assert on the reliability of interictal EEG-
fMRI studies, concluding that significantly different BOLD activations are obtained for interictal
activity when compared to ‚random event‛ analyses (Waites, et al. 2005). Yet, the meaning of the
activation maps obtained from interictal EEG-fMRI is still a matter of debate, specifically, whether or
not the epileptogenic areas of interictal and ictal activity are related
Also studies with interictal activity investigated the relevance of using different HRF models on
the event related design. Some have proposed patient-specific HRFs achieving higher sensitivity rates
(Kang, et al. 2003); others have explored the lag of the HRF, improving the detection of negative
BOLD responses (Bagshaw, et al. 2004). When exploring a parameterised HRF space accounting for
dispersion and delay, Grouiller and colleagues found optimal HRF parameters for epileptic activity
that were significantly different from the canonical HRF found in visual, motor, facial and scene
encoding stimuli. These results indicated a poor sensitivity associated with the canonical HRF for
interictal epilepsy studies. (Grouiller, et al. 2010).
Ictal studies appear less often in literature due to the ictal event’s rare occurrence inside the MRI
scanner. However, they have proven to be insightful in terms of describing, to some extent, onset and
propagation patterns on ictal activity. An example of these is Thornton and colleagues’ work
presenting a study including 83 patients, from whom 9 had seizure occurrences during the scan, and
modelling of the ictal events was made including three different phases: onset, early propagation and
late propagation; thus allowing for haemodynamic flexibility and variability across the brain
(Thornton, et al. 2010). A method also employed in this work was spatial ICA analysis on fMRI data,
which is a data driven approach that excludes the need of an a priori HRF modelling, the spatially
independent components’ (IC) associated BOLD time courses are then compared in order to infer on
the event propagation. The GLM analysis on EEG-fMRI data revealed localized BOLD changes
concordant with the ictal onset zone when the scalp EEG reflected the seizure onset. ICA was
11
independent from the observed EEG and avoided this issue. . However, as a data driven approach,
ICA lacks an explanatory forward modelling. Other studies have also successfully employed ICA for
the distinction between the onset and the propagation of BOLD patterns of ictal events (Iriarte, et al.
2006) and cross-validated ICA methods with GLM analysis (Moeller, Levan and J. 2010, LeVan and
Gotman 2009).
1.3 THESIS RATIONALE
The initial goal of this thesis was to investigate the HRF of ictal epileptic events. As ictal events are
rare and with variable time durations, the usual event-related averaging procedure employed in
interictal studies is inadequate. Therefore, a direct analysis of the EEG data was explored in order to
extract information on the seizure dynamics to be used in the HRF estimations.
An important issue when obtaining direct measures on EEG time courses collected simultaneously
with fMRI is their contamination with artefacts. These artefacts will likely have components in the
frequency band of interest of the EEG, and their mitigation is crucial if one wants to rely on measures
upon spectral characteristics of the EEG time courses. Hence, a special attention was also given to the
correction for MR environment derived EEG artefacts.
The following chapters will describe the methods used and implemented in this Thesis and the
main results and conclusions that were achieved by their means.
12
Chapter 2
METHODS
This chapter will be dedicated to the description of the subjects involved in this study, a brief
description of the data collected from these subjects and a more detailed report on the methods
employed and developed for the EEG-fMRI data analysis.
2.1 DATA ACQUISITION
The data used in this thesis were acquired as part of the pre-surgical evaluation of epilepsy
patients, under the Program for Epilepsy Surgery of the Hospital Centre of West Lisbon. (Leal and
Secca 2010)
2.1.1 Subjects
Five patients were selected from a pool of 30 patients undergoing EEG-fMRI studies supporting
evaluation for surgery. The choice of these patients relied on the occurrence of ictal events on at least
one of the EEG-fMRI sequences acquired. Furthermore, patients with excess movement, ruining the
quality of the data, more than 2 mm absolute movement throughout the sequence, were also
discarded. A clinical, anatomical and neurophysiologic characterization of the selected patients is
presented in Table 2.1 (page 13) (In the resume paper, patient ABC was also excluded, for there was a
ubiquitous artefact presence for this patient’s EEG data which suppression was not achieved, as will
be described in section 3)
During the scanning sessions, patients ABC, JB, JQ and RR were administered light anaesthesia
with Sevoflurane at 1% (Abbott Laboratories, Abbot Park, IL, USA), through mask, as established by
the MRI protocol for small children and uncooperative patients. Patient GM had no anaesthesia
administered.
13
T
AB
LE
2..1
: CL
INIC
AL, A
NA
TO
MIC
AL
AN
D N
EU
RO
PH
YS
IOL
OG
IC C
HA
RA
CT
ER
IST
ICS
OF
TH
E S
UB
JEC
TS.
Su
rger
y o
f
epil
epsy
Infe
rio
r
occ
ipit
al l
ob
e
cort
ical
rese
ctio
n.
En
gel
lb
aft
er 2
yea
r
Tw
o s
tag
e
ham
arto
ma
dis
con
nec
tio
n.
Sei
zure
s
red
uce
d t
o 1
-3,
dai
ly, a
fter
1yea
r.
Inte
rict
al E
EG
Slo
w b
ack
gro
un
d,
wit
h a
bu
nd
ant
mu
ltif
oca
l sp
ike
acti
vit
y.
No
rmal
bac
kg
rou
nd
.
Ab
un
dan
t
rhy
thm
ic s
pik
es
and
slo
w w
aves
ov
er P
7-O
1.
No
rmal
bac
kg
rou
nd
. Sh
arp
wa
ves
ov
er f
ron
tal
lob
es.
No
rmal
bac
kg
rou
nd
. Sp
ikes
at r
igh
t fr
on
tal
area
, F4.
No
rmal
bac
kg
rou
nd
, wit
h
foca
l sl
ow
act
ivit
y
asso
ciat
ed w
ith
abu
nd
ant
spik
es
ov
er l
eft
occ
ipit
al
lob
e.
EE
G p
atte
rn o
f u
sual
seiz
ure
s
Bac
kg
rou
nd
des
yn
chro
niz
atio
n, w
ith
rhy
thm
ic b
urs
ts o
f fa
st
acti
vit
y o
ver
rig
ht
fro
nta
l
lob
e.
Bac
kg
rou
nd
des
yn
chro
niz
atio
n,
pro
gre
ssiv
e b
uil
d u
p o
f
rhy
thm
ic s
pik
es o
ver
P7
-O1.
Bac
kg
rou
nd
des
yn
chro
niz
atio
n w
ith
pro
gre
ssiv
e sh
arp
wav
es
ov
er b
ilat
eral
fro
nta
l lo
bes
.
Pro
gre
ssiv
e b
uil
d u
p o
f
rhy
thm
ic s
pik
es a
t v
erte
x
(Cz)
wit
h l
ater
in
vo
lvem
ent
of
left
cen
tral
der
ivat
ion
s.
Bac
kg
rou
nd
des
yn
chro
niz
atio
n, r
hy
thm
ic
spik
es i
n l
eft
occ
ipit
al l
ob
e
and
lat
er b
uil
d u
p o
f
rhy
thm
ic s
pik
es i
n l
eft
fro
nta
l lo
be.
Sei
zure
freq
uen
c
y
>20
dai
ly
>20
dai
ly
20-2
8
dai
ly
10-2
0
dai
ly
>50
dai
ly
Sei
zure
sem
iolo
gy
Rh
yth
mic
sp
ams,
wit
h
exte
nsi
on
of
bo
th u
pp
er l
imb
s
and
up
war
d t
on
ic d
evia
tio
n.
Du
rati
on
1-3
min
ute
s.
Sp
eech
arr
est,
slo
win
g o
f m
oto
r
acti
vit
y, e
ye
dev
iati
on
to
th
e
left
, in
terr
up
tio
n o
f
con
scio
usn
ess.
Du
rati
on
20
-30
seco
nd
s.
Par
ox
ysm
al a
nxi
ety
, fea
r,
mo
tor
hy
per
acti
vit
y. D
ura
tio
n
1-3
min
ute
s.
Infa
nti
le s
pam
s in
fir
st y
ear
of
life. Seizures with ‚odd
sensation in right arm‛
foll
ow
ed b
y c
ho
reic
/bal
list
ic
mo
vem
ents
. Pro
gre
ssiv
e ri
gh
t
hem
ipar
esis
. Du
rati
on
1-2
min
ute
s.
Mo
tor
arre
st, r
hy
thm
ic e
yel
id
mo
vem
ents
, fea
rfu
l fa
ce
occ
asio
nal
lau
gh
ter.
Du
rati
on
<20
seco
nd
s.
An
ato
mic
al M
RI
Ag
enes
is o
f C
orp
us
Cal
losu
m. M
ult
iple
cort
ical
dy
spla
sias
,
mo
re s
ever
e at
rig
ht
occ
ipit
al l
ob
e.
No
rmal
Lef
t A
mig
dal
a
dy
spla
sia.
No
rmal
Gia
nt
hy
po
thal
amic
ham
arto
ma.
Ag
e o
f
seiz
ure
on
set
(yea
rs)
0.25
7 3 5 0
Ep
ilep
sy
Ty
pe
Ep
ilep
tic
spam
s
Occ
ipit
al
lob
e
epil
epsy
Tem
po
ral
lob
e
epil
epsy
Par
ieta
l
lob
e
epil
epsy
Gel
asti
c
epil
epsy
Ag
e
(Yea
rs)
2 9 4 7 2
Su
bje
ct
AB
C
GM
JB
JQ
RR
14
2.1.2 EEG acquisition
The EEG was recorded with a 37-channel system (electrode map presented in annex) plus two
electrocardiogram (EKG) channels, all using sintered AgCl electrodes connected by carbon fiber wire
to an amplifier located outside the scanner room (Maglink, Neuroscan, Charlotte, USA). The DC
amplifier used (NuAmps, Neuroscan, Charlotte, USA) had a low-pass filter set at 70 Hz and a
sampling rate of 1000 Hz. To obtain skin-electrode impedances bellow 10 KΩ, patients’ skin was
prepared with an abrasive paste and a conductive gel was applied.
The neurophysiologist analysed all EEG sequences for identification of inter-ictal and ictal activity
and provided indication of the time periods where these occurred in each case.
2.1.3 fMRI acquisition
All imaging was performed on a 1.5 T MRI scanner (GE Cvi/NVi). Four to ten fMRI echoplanar
imaging (EPI) sequences were obtained for each patient. The details of the parameters used for each
patient are present in Table 2.2. An anatomical image was also acquired for each subject performing a
T1-weighted spoiled gradient recovery (SPGR) three dimensional (3D) sequence, with an in-plane
resolution of 0.940.94 mm2 and slice thickness of 0.6 mm.
TABLE 2.2: FMRI SEQUENCE PARAMETERS FOR EACH PATIENT
Subject
Number of
volumes per
sequence
Number of
initial volumes
rejected
TR (s) Spatial Resolution Field of view
ABC 170 2 2.100 646424 pixel
4.3754.3755.000 mm3 280280120 mm3
GM 100 2 3.120 646416 pixel
3.7503.7507.000 mm3 240240112 mm3
JB 150 0 2.330 646424 pixel
3.7503.7505.000 mm3 240240120 mm3
JQ 170 2 2.100 646424 pixel
4.3754.3755.000 mm3 280280120 mm3
RR 154 4 2.275 646424 pixel
3.7503.7505.000 mm3 240240120 mm3
2.2 EEG ANALYSIS
The analysis of the EEG data was performed in two main stages:
1. Following visual inspection for rejection of bad channels, a number of basic pre-processing
procedures were first executed and algorithms were then employed for the correction of the
gradient and balistocardiogram artefacts;
15
2. After pre-processed, the EEG data were decomposed using ICA and the resulting ICs were
time-frequency analyzed in order to extract a number of different metrics to be fed into the
regressors in the fMRI analysis.
These stages will now be described in detail.
2.2.1 Pre-processing
The EEG pre-processing was executed through two different methods:
1. EEGLAB procedure: based on tools implemented on the EEGLAB toolbox (Delorme and Makeig
2004) for Matlab;
2. Scan 4.3 procedure: based on the software Scan 4.3(Neuroscan, Charlotte, USA).
In both cases, EEG data was pre-evaluated for the presence and consequent rejection of channels
exhibiting too large amplitude artefacts or channels not acquired during the imaging process. As
result, after this stage, the number of channels remaining for each patient was variable. A 2 Hz high-
pass filter was applied so as to remove the baseline drifts of the signal.
EEGLAB procedure
A series of simple algorithms was developed in order to achieve a fully automated method for the
EEG pre-processing using the EEGLAB tools in Matlab.
Firstly it was necessary to identify the period of the EEG time course corresponding to the fMRI
acquisition sequence. Because the EEG and fMRI acquisition clocks were not synchronized, the
identification of the fMRI acquisition periods was accomplished through the identification of the
corresponding gradient artefacts on the EEG time course. One of the characteristics of the slice
gradient artefacts is the increased power at high frequencies in the EEG recorded during the
application of the imaging gradients, when compared to the EEG recorded in between gradients
(Niazy, et al. 2005). This characteristic was therefore emphasized in order to help identify the exact
times of occurrence of the gradients. A second finite difference was computed over the EEG signal in
order to amplify the higher frequencies. Then, a simple threshold, with value of the mean plus a
standard deviation of the signal data, and an undersized gap elimination procedure were applied to
identify the temporal location of the MR gradient activity. The EEG was then trimmed to retain only
the MR gradient contiguous activity compatible with the functional scanning time (TR Nvolumes).
Gradient artefact correc t ion
The fMRI gradient artefact correction algorithm, FMRIB’s FASTR (Niazy, et al. 2005), as
implemented in EEGLAB, requires the indication of a timing event identifying the sampling of each
fMRI slice. This indication should be accurate enough so as to generate an optimal set of basis
functions describing the temporal variations of the artefact. This is then used as a template, which is
16
fitted and removed from the data. The slice timings provided are further refined to a sub-sampling
resolution by the FASTR algorithm, so the initial estimates of these timings need not have a sub-
sampling rate precision. Using the initial point of the acquisition sequence, T0, identified by the
method described on the previous paragraph, as a reference point, an initial estimate of the slice
timings can then be obtained by:
(2.1)
where is the timing of the ith slice, is the fMRI repetition time and is the number of slices
per volume. Yet, as it will be shown in Chapter 3 Results, these timings were not an accurate
estimation for the sequences presented in this work and a correction was therefore needed. A
recursive algorithm was implemented for this purpose: starting on , a local maximum on the first
finite difference of the EEG signal was located within a window of radius of each slice time. The
occurrence of this well defined local maximum on the first finite difference of the EEG is persistent for
each slice (Niazy, et al. 2005). The lag between this maximum and the slice time is recorded and the
next maximum is located recursively:
∈
(2.2)
where DEEG is the first finite difference of the EEG channel time course under study and is the
time between the timing of the first slice and the corresponding maximum on DEEG. These times
were computed for every EEG channel and the one exhibiting the lowest standard deviation for the
interval duration was considered to provide the best estimate of the slice timings.
After this correction, it was possible to apply the FASTR algorithm over the data. The EEG data
was upsampled to 10 times its original sampling rate and a 30 slice averaging window was used to
compute the artefact template. The number of residual components to be removed was left to be
determined automatically.
Bal istocard iographic artefact correct ion
For the balistocardiographic artefact removal, FMRIB’s QRS complex identification algorithm was
applied to both ECG channels. To correct for missing heart beats, the events detected on both
channels were merged and sorted by time of occurrence and the events lying within a third of the
median time interval between events were considered to belong to the same heart beat. Still, some
QRS events remained undetected and triggers identified on different channels were not aligned in
time. The correction algorithm implemented within FMRIB tools for EEGLAB was applied on this
processed data, yielding the final event list. Finally, the data were low-pass filtered at 45 Hz,
downsampled to 100 Hz to improve its manageability, and an optimal basis set of 3 principal
components was then employed for pulse artefact removal.
17
Neuroscan procedure
Using the tools provided in Scan 4.3 (Neuroscan, Charlotte, USA), the EEG data was high-pass
filtered at 3 Hz, corrected for gradient artefacts and then low-pass filtered at 45 Hz. The EEG data was
downsampled to 100 Hz in order to improve its manageability. No balistocardiographic artefact
correction was performed for this procedure.
2.2.2 Data decomposition
After pre-processing, the EEG data were decomposed by ICA, performed using the infomax
algorithm as implemented on the EEGLAB toolbox (Delorme and Makeig 2004). Although the
referencing method for the EEG channels does not affect the final IC time courses, as it is linearly
separable from the data, it will affect the component scalp maps. As some of the patients had to few
good electrodes or an unbalanced electrode distribution, an average electrode referencing would not
be an adequate procedure if one would want to keep a standard processing for every patient. Given
this, the reference channel was arbitrarily kept as the one chosen by the electrophysiologist during the
acquisition, still allowing intra subject quantitative comparisons between scalp maps.
All resulting ICs were time-frequency analyzed by convolving the signal with Morlet wavelets, G,
which, as a function of time, t, and frequency, f, are given by:
(2.3)
where A=( t )-1 2
, t= ( f), f= and =7 is the ‘wavelet factor’. The power of the signal over
time, t, and around frequency, f, is then (Tallon-Baudry and Bertrand 1999):
(2.4)
where y(t) is the time course of the IC under analysis.
The resulting power spectra were visually inspected for correlations with the time series of ictal
events indicated by the neurophysiologist. Components with visually detectable spectral changes
around the time intervals indicated as ictal events were used for further analysis. On the sequences
where no ictal events were detected, all components were further analysed.
This analysis consisted on the extraction of five different metrics based on the power spectra of the
IC. The metrics used were inspired by those presented in Rosa et al. (2010) and they are:
Tota l power
(2.5)
Un-normalized root mean square frequency
(2.6)
18
Un-normalized mean Frequency
(2.7)
Root mean square f requency
(2.8)
Mean frequency
(2.9)
For the computational implementation of the spectral decomposition and metric application, a
discretization of the time-frequency domain was required. As referred in section 2.1.1, the final EEG
data had a time sampling frequency of 100 Hz and the spectral content of interest was between 2 Hz
and 45 Hz. Therefore, a signal frequency interval between 1 and 50 Hz was chosen.
Furthermore, the frequency resolution of the Morlet wavelets is linearly dependent of the
frequency being analyzed, f= . Hence, the frequency domain sampling interval, , should be kept
bellow this value:
(2.10)
Consequently, the frequencies sampled in the frequency domain, fn, are given by:
(2.11)
) (2.12)
(2.13)
with . This represents a logarithmic sampling in the frequency domain. The integral
expressions for the metrics were then approximated by:
(2.14)
where A is the function to be integrated, = to account for the change of variable from f to fn,
and fn is given by Eq. (2.13).
In this study, the parameters used were: R=7 for the Morlet wavelet decomposition,
K=exp(log(50)/50), f0=1 Hz and N=50. In this way, K verifies Eq. (2.10).
2.3 FMRI ANALYSIS
Analysis of fMRI data was performed with FMRIB’s software library, FSL
(www.fmrib.ox.ac.uk/fsl) (Woolrich, Jbabdi, et al. 2009, Smith, et al. 2004), specifically using the
methods implemented within FEAT (Woolrich, et al. 2001). Following standard pre-processing
procedures, the data were analysed for the identification of BOLD signal changes associated with the
different metrics of interest extracted from the EEG, using a GLM approach. These metrics were then
19
compared in terms of the resulting activation maps. The general processing procedure and the EEG
metrics comparison will now be described in detail.
2.3.1 General processing
Pre-processing of the fMRI data consisted on the following steps. Motion correction was carried
out using MCFLIRT (Jenkinson, et al. 2002). Slice timing correction was performed by using
(Hanning-windowed) sinc interpolation to shift each time-series by an appropriate fraction of the
repetition time (TR) relative to the middle of the TR period. This step will be important for the
estimation accuracy of the HRF’s delays. The brain region extraction was performed using BET (S. M.
Smith 2002). All fMRI sequences were temporally high-pass filtered, rejecting periods above 100 s. A
spatial Gaussian filter with FWHM of 8 mm was also applied.
For statistical analysis, a GLM was generated using the regressors of interest extracted from the
EEG, as detailed in the next sub-section. Six motion parameters were also included in the GLM matrix
as confounds of no interest, in order to account for residual motion jitter not removed by the motion
correction procedure. The GLM was then fitted to the data using the FILM algorithm, which
incorporates a robust and accurate nonparametric estimation of time series autocorrelation to
prewhiten each voxel's time series (Woolrich, et al. 2001). After the model fit, a t test was then applied,
resulting in a Z (Gaussianised T/F) statistic map for each contrast. The Z maps for each subject were
thresholded using a clustering procedure, whereby each cluster is determined by a voxel Z > 2.3 and a
(corrected) cluster significance threshold P = 0.05 (Worsley 2001).
2.3.2 Testing different EEG metrics
For each IC of interest and each metric, the corresponding time course was convolved with a
canonical HRF. This HRF is the impulse response function associated with the biophysical model
implemented in SPM8 (www.fil.ion.ucl.ac.uk/spm/), which results from the Balloon Model originally
proposed by Buxton and colleagues (Buxton, Wong and Frank 1998) and further complemented with
the flow dynamics by Friston and colleagues (Friston, Mechelli, et al. 2000). The prior expected values
presented in Table 1.1 were used as parameters.
The final regressors were obtained by re-sampling the data to match the middle of the acquisition
time period of each volume of the fMRI sequences. The time derivative of the regressor was also
included in the GLM matrix in order to allow for uncertainties in the delay of the haemodynamic
response across the brain.
For quantitative comparisons between the different EEG metrics, the number of voxels of the
activation cluster was used.
20
2.4 HRF STUDY
An investigation of the HRF was performed for patient RR. The goal of the HRF studies performed
in this Thesis is the identification of the transfer function taking the EEG metric under analysis as
input and the subject’s BOLD response as output. Two main approaches were taken in these HRF
studies: pixel-by-pixel methods and ‚region of interest‛ (ROI) based methods. Each of these methods
will now be presented.
2.4.1 ROI based analysis
A number of ROI’s are first defined based on the activation clusters combined with functional
anatomical criteria.
ROI def init ion
Maps of relevant anatomical regions were identified based on MNI’s anatomical probability maps
(Lancaster, et al. 2007): left frontal lobe, left occipital lobe, left parietal lobe and left hippocampus.
These maps were transformed into the functional space of each sequence of interest and then
thresholded at 0.5 probability value. The hamartoma region was delineated manually based on the
identification of the lesion on the patient’s high-resolution T1-weighted image, and was subsequently
transformed into the functional space. The ROIs were obtained by intersecting each of these maps
with the thresholded activation maps obtained in each sequence.
All transformations were achieved using FLIRT (Jenkinson, et al. 2002) toolbox. To obtain the
transformation matrices between the MNI’s standard space, the anatomical space and the functional
space, firstly the functional reference image of every sequence was registered into the anatomical
space using a 12 parameter affine model. Afterwards, the anatomical image was registered into the
standard space using the same model. The functional to standard space registration corresponds to
the sequential application of the transformations described.
Analysis
In the ROI based analysis, a BOLD time course, representative of a certain region, is used. This
representative time course is obtained by averaging the voxels time courses across the ROI and then
projecting this average onto the null space of the confounds (the six movement parameters).
The EEG metric time courses were low-pass (anti-aliasing) filtered and resampled to the fMRI
mean volume acquisition times in order be in concordance with the ROI BOLD time course sampling.
The models estimated were a finite impulse response linear filter (FIR deconvolution model), an
infinite impulse response (IIR) linear filter and the biophysical model. The estimation procedures will
subsequently be described.
21
Biophysical model
The biophysical model used is the one implemented in SPM8 (www.fil.ion.ucl.ac.uk/spm/), which
results from the Balloon Model already introduced in Chapter 1. This model describes the changes in
blood oxygenation, cerebral blood flow and cerebral blood volume, as a consequence of the regional
increase in brain metabolism associated with neuronal activity. The estimation of the parameters of
the Balloon Model was achieved using the expectation-maximization algorithm implemented on
SPM8 (Friston, Harrison and Penny 2003).
Tests involving the contraction or relaxation of the priors were done by multiplying the original
prior parameter covariance matrix by a constant.
I IR model
The IIR model implemented was defined in terms of a difference equation from which an estimate
of BOLD response is given by:
(2.15)
where bi and ak are the IIR filter parameters to be estimated for some given nb and na; and r(n) is the
regressor time course.
The estimation of the IIR model was done by an error minimization, on a least squared error sense,
between the predicted BOLD response and the observed one. The energy function to optimize is then:
(2.16)
where is the BOLD observed time course and is the BOLD estimate given by Eq. (2.15).
A gradient descent algorithm was applied for the minimization of E, introduced in Eq. (2.16), as a
function of the IIR parameters bi and ak. Differentiating Eq. (2.16) in order to ak yields:
(2.17)
A similar result is obtained when differentiating E in order to bi. Finally, differentiating Eq. (2.15)
in order to the IIR filter parameters yields:
(2.18)
and
(2.19)
The total gradient is computed sequentially, imposing all terms and derivatives for equal to
zero for any n < 0. An adaptive steps acceleration method (Almeida, et al. 1998) was included in the
implementation for a faster convergence. The initial estimate for the parameters was obtained by
considering the observed BOLD signal, y, equal to the predicted one, , so that:
22
(2.20)
where and are the vectors of the IIR parameters, and are matrices of the corresponding lags of
the observed BOLD and regressor time courses, respectively, and is the column vector of the
observed BOLD time course.
In the present study, the values na=3 and nb=3 were used, following the physiologically inspired
modelling made in (Afonso, Sanches and Lauterbach 2007).
FIR model
The FIR model is the most unconstrained approach explored for the HRF estimation. The only
assumptions made are that the stimulus has a limited temporal effect on the haemodynamic response
observed and that this response is linear and time invariant. Also, on the present work, only causal
HRFs were allowed.
The estimation of the FIR model was performed by fixing a response length, nb, and, once again,
minimizing the squared residue of the predicted BOLD response and the observed one. The
estimation procedure is identical to that of the IIR model for na equal to zero. For its estimation, only
Eq. (2.20) needs to be computed as no recursive dependencies from previous times of BOLD estimates
are present ( and are empty matrixes). This procedure corresponds to a deconvolution process
(Gray 2006).
2.4.2 Pixel-by-pixel analysis
Two major HRF modelling approaches were adopted in this section: the use of basis functions and
the exhaustive search on a parameterized space of HRFs. In both cases, the goal was to gain insight
into the distribution of the time delay and the time dispersion of the HRFs across each subject’s brain.
Basis functions
The investigation of haemodynamic features through the use basis functions is based on the
linearity of the convolution operation. This means that, when using a set of haemodynamic
explanatory regressors obtained by convolution of a basis set of functions with the same EEG metric,
one single haemodynamic function associated to a given voxel is possible to be derived:
(2.21)
where βi are the coefficients resulting from the GLM fit, Hi corresponds to the ith element of the set of
n basis functions used and R is the regressor under analysis. For inter voxel comparisons of HRFs, the
β coefficients are convenient to be normalized in order to be directly comparable with each other:
(2.22)
This way the normalized HRF for a given voxel is:
23
(2.23)
In order for a limited, linearly combined, set of basis functions to better represent, in a Taylor
series expansion sense, a non-linearly parameterized HRF, it is convenient to ‘center’ it around the
mean HRF object of study. Meaning that, when using a standard HRF, its time derivative and its
dispersion derivative, it will be convenient for the coefficients of the derivative terms, resulting from
the GLM fit to the data, to be as close to zero as possible across the ROI under analysis. With this
rationale, a ROI based analysis over all activation maps was performed before engaging on studies
based in basis functions. This way it was possible to build a more data informed basis set. Once the
HRF was estimated using any of the methods in 2.4.1, its time and dispersion derivatives were
computed. All three time courses were then convolved with the EEG metric under study and
resampled to the middle of the volume fMRI acquisition times, resulting on the final basis set to be
included in the GLM analysis.
A related approach using basis functions was also explored by using FMRIB’s FLOBS toolbox and
filmbabe script (Woolrich, Behrens and Smith 2004). FLOBS uses a three half-cosine parameterization
of the HRF, resulting in five free parameters, m1, m2, m3, m4 and c, Figure 2.1. It generates 1000
random uniformly sampled HRFs from a prescribed set of parameter value intervals. After this, the
first N, in this case 3, principal components of the samples generated are kept as the basis functions.
Also in this stage, using the random samples as learning set, a prior mean and covariance matrixes for
the basis set components are computed reprojecting the basis functions onto the data.
FIGURE 2.1: Three half-cosines FLOBS HRF parameterization.
The core difference between filmbabe script and regular GLM analysis using FEAT is the inclusion
of a priori knowledge of the HRF shape when fitting the basis functions of the GLM on the data. The
basis coefficients probability function is modelled as a multivariate normal distribution with mean
and covariance learned as mentioned in the previous paragraph. This represents a constraint on the
admissible HRF shapes, reducing the occurrence of implausible HRFs.
24
The EEG metric time courses were low-pass (anti-alaiasing) filtered and resampled to the fMRI
volume acquisition times in order to be usable by filmbabe script.
Exhaust ive search
This approach is inspired on the one presented in (Grouiller, et al. 2010). The HRF model is based
on a gamma function, which described the main behaviour of the BOLD signal without undershoot:
(2.24)
with a=6s and b=0.9. The bivariate modelling of the HRF comes as:
(2.25)
where tup is abτ. In the present work, a re-parameterization of the HRF turning was considered in
order to make its peak time, tmax, and its raising time, tup, explicit:
(2.26)
(2.27)
Another motivation for this re-parameterization is of the reduction of the co-dependency between
the explicit variables and the quality of model fit to the data. Specifically, the dependency of the
model fit on t0 will be correlated with its dependency on τ. This is because the ‚peaking time‛ feature
explains more signal variance than the ‚raise beginning time‛ feature alone and tmax= t0 + abτ. This is
observable in the results obtained in (Grouiller, et al. 2010). The choice of re-parametrizing τ into tup
was made in order to obtain a more intelligible measure of dispersion, as tup directly corresponds to
the interval of time from the beginning of the HRF raise to the moment of its peak.
A discretization of the parameter space was performed and, for every pair of parameter values, a
GLM analysis was executed. The one regressor included in the GLM (besides the confounds) was
obtained by convolving the corresponding HRF with the time course of the EEG metric of interest.
The total number of voxels in the activation map obtained and the maximum Z value for that map
were plotted for every pair of values of the parameterized HRF space. For every voxel, the pair of
parameter values yielding the greatest Z value was also mapped.
25
Chapter 3
RESULTS
In this chapter the main results obtained through the methodology previously described will be
presented. It should be noted that the fMRI images are displayed following the radiologic convention:
the image left side corresponds to the patient’s right, while the EEG IC scalp map images are
presented following the neurological convention: the image left side corresponds to the patient’s left.
3.1 EEG PRE-PROCESSING ANALYSIS
The present section will be dedicated to the depiction of each of the processing steps leading from
the raw EEG data to the final regressors to be included in GLM analysis of the fMRI data. Here, only
the processing of sequence 3, form subject RR, will be described, as an example illustrating the effects
and importance of each of the pre-processing steps.
3.1.1 Artefact correction
The performance of the artefact correction procedures employed in this work is illustrated here
using event-related potential (ERP) images, obtained by triggering the EEG signal on the fMRI slice
timing events or the EKG QRS. The corresponding average time courses are also shown, for an
example channel. The presence of signal structure correlated with the slice gradients or the QRS
events is reflected in high vertical coherence in the ERP images and results in high average signal
amplitude at specific moments in time.
The performance of the slice timing identification algorithm developed here is illustrated in Figure
3.1. Observing the ERP image on the left, it is clear that the fMRI sequence does not maintain a
constant time between slices, using as reference the EEG ADC clock. Some sporadic shorter slice
durations are patent, for example near the 3000th slice, as well as a more frequent small jitter on slice
timing all along the ERP image. After the application of the routine developed in this work, one may
see a clear improvement in the alignment of the MRI slice artefacts, resulting in a mean time course of
more than four times the amplitude of the one not corrected, which reflects the higher vertical image
coherence.
The Neuroscan procedure and the EEGLAB procedure are compared in Figure 3.2: one can observe a
larger residual artefact left by the first approach. On the left ERP image, the higher vertical coherence
results in a mean time course of more than 100 times the amplitude of the one on the right.
26
FIGURE 3.1:ERP images (top) and corresponding average time courses (bottom) of an example
channel (F7) triggered on the fMRI slice events : using fixed time slice triggers (left) and using the
slice timing identification algorithm developed here (right).
FIGURE 3.2: ERP images (top) and corresponding average time courses (bottom) of an example
channel (F7) after fMRI gradient corrections and triggered on the aligned slice events: using the
Neuroscan procedure (left) and the EEGLAB procedure developed here (right).
-44.5
-22.2
0
22.2
44.5
Trials
F7
100
200
300
400
500
-200 -100 0 100 200 300-3-2-1012
Time (ms)
V
-243
-122
0
122
243
Trials
F7
1000
2000
3000
0 10 20 30 40 50 60 70 80 90
-850
85170255
Time (ms)
V
-242
-121
0
121
242
Trials
F7
1000
2000
3000
0 10 20 30 40 50 60 70 80 90
-850
85170255
Time (ms)
V
-243
-122
0
122
243
Trials
F7
1000
2000
3000
0 10 20 30 40 50 60 70 80 90
-44-22
02244
Time (ms)
V
-44.5
-22.2
0
22.2
44.5
Trials
F7
100
200
300
400
500
-200 -100 0 100 200 300-3-2-1012
Time (ms)
V
-45.7
-22.8
0
22.8
45.7
Trials
F7
1000
2000
3000
0 10 20 30 40 50 60 70 80 90
-20-10
01020
Time (ms)
V
-35.8
-17.9
0
17.9
35.8
Trials
F7
1000
2000
3000
0 10 20 30 40 50 60 70 80 90
-20-10
01020
Time (ms)
V
-45.7
-22.8
0
22.8
45.7
Trials
F7
1000
2000
3000
0 10 20 30 40 50 60 70 80 90
-20-10
01020
Time (ms)
V
μV
μV
27
The effect of the balistocardiographic artefact correction is shown in Figure 3.3. On the left, a
vertical structure around 100 ms after the ERP triggers reflects the QRS-related artefact. This structure
is once again translated into high amplitudes around that time in the mean time course of the ERP
image. On the right, after the correction procedure, the vertical coherence is attenuated and the mean
time course of the image becomes close to zero after the QRS trigger.
FIGURE 3.3: ERP images of an example channel (F7) triggered on EKG QRS events. On the left is the
image before balistocardigrphic corrections. On the right is the image after balistocardiographic
artefact correction procedure. Below are the plots of the mean time courses of both images.
3.1.2 Data decomposition
Here, we show the effects of ICA decomposition in terms of the separation of the ictal activity into
a limited number of components and the spectral profiles obtained. For this purpose, the results of the
ICA decomposition of the example sequence are illustrated by the maps presented in Figure 3.4. A
more detailed analysis of these maps will be made in section 3.2, integrated with the corresponding
fMRI results. In this figure, the components are ordered with decreasing EEG explained variance.
Some components revealed spectral correlations with the ictal markers indicated by the
neurophysiologist. One of them was component 5, whose map is detailed in Figure 3.5.
-44.5
-22.2
0
22.2
44.5
Trials
F7
100
200
300
400
500
-200 -100 0 100 200 300-3-2-1012
Time (ms)
V
-43.7
-21.8
0
21.8
43.7
Trials
F7
100
200
300
400
500
-200 -100 0 100 200 300
-12-606
12
Time (ms)
V
-44.5
-22.2
0
22.2
44.5
Trials
F7
100
200
300
400
500
-200 -100 0 100 200 300
-12-606
12
Time (ms)
V
-44.5
-22.2
0
22.2
44.5
Trials
F7
100
200
300
400
500
-200 -100 0 100 200 300-3-2-1012
Time (ms)
V
μV
-44.5
-22.2
0
22.2
44.5
Trials
F7
100
200
300
400
500
-200 -100 0 100 200 300
-12-606
12
Time (ms)
V
28
FIGURE 3.4: IC maps obtained for sequence 3, patient RR. The components are ordered by decreasing
explained variance (reference channel: FCZ).
FIGURE 3.5: Detail of the 5 th IC. The red marker corresponds to the channel with highest absolute
weight on this component, F3. (reference channel: FCZ)
The data separation provided by the ICA decomposition is patent in Figure 3.6. When comparing
the spectral profile of component 5 with the spectral profile of the single channel with highest
absolute weight in this component, F3, the better co-localization of the IC’s power changes relative to
1 2 3 4 5
6 7 8 9 10
11 12 13 14 15
16 17 18 19 20
21 22 23 24 25
-
+
CNT file resampled
5
1 2 3 4 5
6 7 8 9 10
11 12 13 14 15
16 17 18 19 20
21 22 23 24 25
-
+
CNT file resampled
29
the neurophysiologist’s markers becomes evident. This gives rise to a clearer correspondence between
the spectral alterations and the neurophysiologist’s markers, not achieved by any other single
channel.
FIGURE 3.6: Spectrograms of EEG channel F3 (top) and ICA’s 5 th component (bottom), obtained for
sequence 3 of patient RR. Channel F3 was the channel that contributed most to the 5 th IC. The boxes
in red represent the periods marked as ictal events by the neurophysiologist. Spectral changes
associated with the ictal events are visible in both spectrograms, although with a time shift. These
correlations are clearer in the spectrogram obtained for the 5 th IC.
3.2 EEG-FMRI ANALYSIS
This section will be structured patient-by-patient, as each one of them has its own specificities. For
each patient, the results obtained by the GLM analysis of the fMRI data using the EEG-derived
regressors will be presented. Specifically, comparisons will be performed regarding the volume of the
activation maps, as well as their associated time courses.
The volume of the activation map, defined as the total number of activated voxels identified as
activated, will be used as a measure of how well a certain regressor explains the BOLD signal across a
network of brain areas. We chose to use this measure, instead of other extensively used measures of
model fit such as the maximum Z value, because our aim is to explain the greatest amount of brain
areas involved in the epileptic events using the EEG regressors.
3.2.1 Subject RR
Subject RR had six EEG-fMRI sequences, from which two had ictal events identified by the
neurophysiologist: sequences 2 and 3. By visual inspection of their spectra, the ICs exhibiting a
correlation with the ictal events were identified for each sequence (as illustrated in the previous
section). For sequence 2, some of these components did not yield significant activation maps, so they
Time (s)
Fre
quency (
Hz)
50 100 150 200 250 300 350
2
4
8
16
32
Time (s)
Fre
quency (
Hz)
50 100 150 200 250 300 350
2
4
8
16
32
30
were discarded a posteriori. At the end, 10 ICs from sequence 2, and 9 from sequence 3, were kept.
Some equivalence between the components of sequences 2 and 3 was observed: in Figure 3.7, the
scalp maps of the ICs of interest are shown and an absolute correlation image between the normalized
ICA weights found for both sequences.
FIGURE 3.7: Scalp maps of the ICs for sequences 2 and 3 of patient RR (reference channel: FCZ) (top
and centre, respectively). In the bottom image, the intensity corresponds to the absolute value of the
correlation of the normalized IC weights of both sequences.
1 4 6 9 10
12 16 17 21 23
-
+
Subject RR, Sequence 2
1 2 5 6 10
11 14 18 24
-
+
Subject RR, Sequence 3
Independent components (sequence 3)
Independent
com
ponents
(sequence 2
)
1 2 5 6 10 11 14 18 24
1
4
6
9
10
12
16
17
21
230.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
31
Activat ion volumes
The next set of figures synthesizes the activation map volumes obtained for each metric and each
component of sequences 2 and 3. As a reference values, the analysis with the neurophysiologist’s
regressor yielded activation maps with 1194 voxels for sequence 2 and 1751 voxels for sequence 3.
SEQUENC E 2, PATIENT RR
FIGURE 3.8: Activation map volumes obtained for each metric and each component of sequence 2,
patient RR. For each IC, the volumes are plotted in descending order for the different metrics.
FIGURE 3.9: Activation map volumes obtained for each metric and each component of sequence 2,
patient RR. For each metric, the volumes are plotted in descending order for the different ICs.
0
500
1000
1500
2000
2500
FP
uR
MS
FT
PM
FR
MS
F
Vo
lum
e (
vo
xe
ls)
Component 1
0500
100015002000250030003500
RM
SF
MF
uR
MS
F
FP
TP
Component 4
0
500
1000
1500
2000
2500
3000
RM
SF
MF
uR
MS
F
FP
TP
Component 6
0
500
1000
1500
2000
2500
RM
SF
MF
FP
uR
MS
F
TP
Component 9
0500
100015002000250030003500
uR
MS
F
FP
MF
RM
SF
TP
Component 10
0
500
1000
1500
2000
2500
3000
uR
MS
FF
PM
FR
MS
FT
P
Vo
lum
e (
vo
xe
ls)
Component 12
0
1000
2000
3000
4000
5000
RM
SF
MF
uR
MS
F
FP
TP
Component 16
0500
100015002000250030003500
uR
MS
F
FP
TP
MF
RM
SF
Component 17
0
500
1000
1500
uR
MS
F
FP
MF
RM
SF
TP
Component 21
0
500
1000
1500
2000
uR
MS
F
FP
RM
SF
TP
MF
Component 23
010002000300040005000
16 4 6 10 9 12 21 1 17 23Vo
lum
e (
vo
xe
ls)
Component Inex
MF
010002000300040005000
16 4 6 9 10 12 21 23 1 17
Component Inex
RMSF
0
1000
2000
3000
4000
16 10 17 12 1 4 23 9 21 6Vo
lum
e (
vo
xe
ls)
Component Inex
uMF
0
1000
2000
3000
4000
16 10 17 12 1 4 23 21 9 6
Component Inex
uRMSF
0
1000
2000
3000
17 16 10 12 1 23 4 9 21 6
Component Inex
TP
32
SEQUENC E 3, PATIENT RR
FIGURE 3.10: Activation map volumes obtained for each metric and each component of sequence 3,
patient RR. For each IC, the volumes are plotted in descending order for the different metrics.
FIGURE 3.11: Activation map volumes obtained for each metric and each component of sequence 3,
patient RR. For each metric, the volumes are plotted in descending order for the different ICs.
When carefully comparing the stats and the scalp IC maps some consistencies appear. Firstly, the
1st and 10th components from sequence 3, which yielded the largest activation volumes with MF and
RMSF metrics, are highly correlated, in a scalp map senesce, with the 6th and 16th components,
respectively, which also were amongst those with the largest activation volumes for MF and RMSF
0
500
1000
1500
2000
2500
3000
3500M
FR
MS
Fu
RM
SF
FP
TP
Vo
lum
e (
vo
xe
ls)
Component 1
0
500
1000
1500
2000
2500
uR
MS
F
FP
MF
RM
SF
TP
Component 2
0
1000
2000
3000
4000
uR
MS
F
FP
TP
MF
RM
SF
Component 5
0
500
1000
1500
2000
TP
FP
uR
MS
F
MF
RM
SF
Component 6
0
1000
2000
3000
4000
MF
RM
SF
uR
MS
F
FP
TP
Component 10
0
500
1000
1500
2000
2500
MF
uR
MS
FF
PR
MS
FT
P
Vo
lum
e (
vo
xe
ls)
Component 11
0
1000
2000
3000
4000
uR
MS
F
FP
TP
MF
RM
SF
Component14
0
200
400
600
800
1000
TP
FP
uR
MS
F
MF
RM
SF
Component 18
0
200
400
600
800
uR
MS
F
FP
MF
RM
SF
TP
Component 24
0
1000
2000
3000
4000
10 1 11 2 5 14 24 6 18Vo
lum
e (
vo
xe
ls)
Component Inex
MF
0
1000
2000
3000
4000
10 1 11 2 5 24 14 6 18
Component Inex
RMSF
0
1000
2000
3000
4000
5 14 10 1 11 2 6 24 18Vo
lum
e (
vo
xe
ls)
Component Inex
uMF
0
1000
2000
3000
4000
5 14 10 2 1 11 6 24 18
Component Inex
uRMSF
0
1000
2000
3000
4000
5 14 6 10 11 18 1 24 2
Component Inex
TP
33
metrics for sequence 2. Considering now the unormalized regressors, uMF and uRMSF, an equivalent
observation can be made: the 5th and 14th components from sequence 3 correlate with the 17th and 10th
components from sequence 2, respectively, and they are all amongst the components with largest
activation volumes for these regressors (excluding the 16th component from sequence 2, already
discussed for MF and RMSF metrics, these components were effectively the ones yielding the largest
activation volumes).
Making general considerations upon the EEG metrics, no single one clearly outstood positively
from the others, yet the TP metric produced poorer results when compared against the remaining
metrics. The normalization was an important metric feature as the components which yield the
largest activations clearly depended on this. The frequency averaging method, simple or root mean
square, seemed to have little effect on the activation volumes produced.
On the components’ topographies, it is observable that those that yielded larger activation maps
have higher absolute weights over the left scalp hemisphere, concurring with the EEG interpretation
made by the neurophysiologist, which values features mainly in the time courses from those
channels.
Activat ion maps and t ime courses
The first comparison worth to be made is the difference on the activation maps and BOLD time
courses obtained for the standard regressors defined by the neurophysiologist and the most relevant
regressors according to the criteria presented in the previous section. These maps and time courses
are shown in Figure 3.12 and Figure 3.13 for sequences 2 and 3.
34
SEQUENC E 2, PATIENT RR
16TH IC RMSF REGRESSOR: 2.3 6.8 NEUROPHYSIOLOGIST’S REGRESSOR : 2.3 5.0
FIGURE 3.12: Activation Z statistic maps (top) and corresponding average time courses, partial and
full model fits (arbitrary units), plotted as a function of the volume number (bottom), obtained for
sequence 2, patient RR.
Neurophysiologist’s regressor
16th IC RMSF regressor
R L
35
SEQUENC E 3, PATIENT RR
10TH IC MF REGRESSOR: 2.3 6.7 NEUROPHYSIOLOGIST’S REGRESSOR : 2.3 4.9
FIGURE 3.13: Activation Z statistic maps (top) and corresponding average time courses, partial and
full model fits (arbitrary units), plotted as a function of the volume number (bottom), obtained for
sequence 3, patient RR.
10th IC MF regressor
Neurophysiologist’s regressor
R L
36
One may see that, for both sequences, the activation maps obtained with the EEG metric regressors
are broader and more significant when compared to the obtained with the neurophysiologist’s
marked regressors. Although they do not overlap, in general, the activation maps are concordant in
terms of the anatomical regions which appear with activated areas: left occipital and parietal lobes,
left frontal lobe and left hippocampus. The presumed epileptogenic focus, the hamartoma, appeared
activated, in sequence 2, only with the EEG metric regressors and, in sequence 3, with both the
neurophysiologist marked regressor and the EEG metric regressors.
For both sequences, taking into account the mean BOLD time courses of the activation masks, the
fit of the model with the EEG metric regressors produced smaller residues than the fit with the
neurophysiologist’s regressor model. This is particularly patent for sequence 2, in which the mean
time course of the activation mask for the neurophysiologist’s regressor has two unexplained major
‘bumps’ which appear to be predicted by the EEG metric regressor.
The activation maps obtained for the other components and metrics yielding large activation maps
(>1000 voxels) are generally concordant with those presented here (not shown).
When comparing the activation maps presented with the respective EEG ICs’ scalp maps, for both,
occipital/parietal configurations are present. Also, other EEG ICs yielding large activation maps
emerged with more frontal configurations, in concordance with the activation maps presented. The
scalp maps configurations of the ICs must be interpreted with caution, since the reference channel
used was FCZ and not the average between all the electrodes.
Non- icta l sequences
When the same methodology was applied to the sequences on which no ictal events were
identified by the neurophysiologist (1, 4, 5 and 6), some consistencies emerged. For sequences 1, 4 and
6, the components yielding the largest activation maps were also the components with higher spatial
correlation with the 16th and 10th components of sequences 2 and 3, respectively. For sequence 5, the
component most correlated with these yielded the second largest activation maps. Again, for this
component topography the most significant metrics were frequency normalized, being the only
exception sequence 1. The scalp topographies and activation masks for these components are
presented in Figure 3.14.
The topography of the fMRI activation masks obtained for these sequences were in concordance
with the ones of sequences 2 and 3 in terms of the anatomical regions activated: mostly left occipital,
parietal and frontal lobes and left hippocampus. The activation topography of sequence 5 was less
defined towards the left hemisphere and appeared more widespread, including various clusters in
the right hemisphere. The epileptogenic focus, the hamartoma, appeared with activated voxels, in
sequences 4 and 6.
37
FIGURE 3.14: EEG scalp maps (reference channel: FCZ) (left) and fMRI activation Z statistic maps of
central slices (right), for the EEG ICs of sequences 1, 4, 5 and 6 that exhibited the greatest spatial
correlation with components 16 and 10 of sequences 2 and 3 . The EEG metrics presented are the ones
yielding the largest activation maps for the respective component.
3.2.2 Subject JQ
Only one EEG-fMRI sequence from patient JQ was available. This sequence had three periods
marked as ictal events by the neurophysiologist. From the ICs’ spectral analysis, one component
outstood in terms of the correlation of the spectrogram with the neurophysiologist’s ictal markers. Its
spectrogram and scalp map are presented in Figure 3.15. Yet, this component’s time course was
highly correlated with the only rejected channel from this sequence, as is shown in Figure 3.16. This,
along with the multi-pole configuration of the scalp map, suggests an artefactual nature for this
component. As no other IC was spectrally correlated with the neurophysiologist’s markers as clearly
as component 14 and this correlation could represent artefact residuals not separated by the ICA
procedure, the pre-selection of ICs was ignored in this case and the EEG metrics were applied on all
of them, each proceeding to fMRI analysis.
IC 14 from Subject RR, Sequence 1
-
+
IC 12 from Subject RR, Sequence 1
-
+
IC 17 from Subject RR, Sequence 1
-
+
IC 19 from Subject RR, Sequence 1
-
+
7.5
2.3
Seq
. 6
, 19th
IC
, MF
6.0
2.3 S
eq. 5
, 17th
IC
, RM
SF
7.0
2.3
Seq
. 4, 1
2th I
C, M
F
6.0
2.3
Seq
. 1, 1
4th I
C, u
MF
R L L R
38
FIGURE 3.15: EEG scalp maps (left) and spectrogram (right) for the 14 th IC from the EEG of subject
JQ. ). The boxes in red represent the periods marked as ictal events by the neurophysiologist .
FIGURE 3.16: EEG time course of channel 25(CP4) (top) and 14 th IC (bottom). The abnormally high
amplitudes present in channel 25 time course indicate the artefactual character of the channel , thus
removed from the data. 14 th IC amplitude profile accompanies that of the channel 25.
Activat ion volumes
In Figure 3.17 a set of plots that synthesizes the activation mask volumes obtained for each metric
and each component of subject JQ is presented. Using the neurophysiologist’s regressor, the
activation map obtained had 53 active voxels and several components yielded larger activation maps
than the neurophysiologist’s regressor. Yet, when examining their activation map topographies, only
3 components yielded maps consistent with the one obtained with the neurophysiologist’s regressor
and the patient’s clinical history, presenting activated voxels on the left parietal lobe.
IC 14 from Subject JQ, Sequence 1
-
+
Time (s)
Fre
quency (
Hz)
50 100 150 200 250 300 350
2
4
8
16
32
50 100 150 200 250 300 350
-4
-2
0
2
4
x 104
Time (s)
Am
plit
unde (
V)
EEG channel 25
50 100 150 200 250 300 350
0
Time (s)
Am
plit
unde (
std
units)
Independent component 14
39
FIGURE 3.17: Activation map volumes obtained for each metric and each component of patient JQ.
For each IC, the volumes are plotted in descending order for the different metrics.
Activat ion masps and t ime courses
In Figure 3.18 the activation map and the mean BOLD time course obtained for the
neurophysiologist’s regressor are shown. A small cluster on the left parietal lobe is observable.
Taking into account the activation maps obtained for the EEG metrics applied in the different ICs,
as referred before, only 3 had clusters in the left parietal lobe and the activation map obtained for the
largest activation yielding metric is displayed in Figure 3.19. The spectral correlation of component 14
with the neurophysiologist’s markers produced, for the un-normalized EEG metrics, regressors very
close to the one obtained from the neurophysiologist’s appointment. This way, the activation map
obtained is also very similar to the one of Figure 3.18, yet with an additional cluster on the right
parietal lobe. But as previously adverted, this component is likely to be contaminated with artefacts.
Taking into account the 19th IC, some clusters are obtained with locations beside the left parietal
lobe, namely, activations in right frontal and right and left occipital lobes. When observing the EEG
scalp map associated with this component, the major weights are for electrodes located in right
temporal areas, with some degree of concordance with the right frontal cluster observed in the fMRI
activation map, yet again these observations must be made cautiously as the reference was not made
to the mean electrode time course..
The 12th IC yielded a single activation cluster located in left apical regions, mostly in the parietal
lobe but also extending to the frontal lobe. When observing the IC scalp map, a frontal oriented
0
100
200
300
400
500
23 15 25 19 12 6 16 17 22 24 14 5 9Vo
lum
e (
vo
xe
ls)
Component Inex
MF
0
200
400
600
19 25 15 23 6 16 12 28 24 14 22 17 5 9
Component Inex
RMSF
0
50
100
150
200
11 22 6 14 23 30 3 21Vo
lum
e (
vo
xe
ls)
Component Inex
uMF
0
50
100
150
22 11 14 6 30 23 3
Component Inex
uRMSF
0
200
400
600
11 23 21 18 20 9 14 30 3 16 17
Component Inex
TP
40
structure appears slightly lateralized to the left hemisphere, yet again these observations should be
interpreted in the sense of confirming the non-artefactual nature of the component. This component’s
spectrogram, the EEG metric regressor MF and the mean BOLD time course over the activation mask
are presented in Figure 3.20. This is a good example to illustrate the effect of the normalization in the
EEG metric: observing the spectrogram in the period around 200 s marked as ictal event by the
neurophysiologist, a desynchronization of the EEG on the frequencies around 2~4 Hz is patent, which
translates into a higher mean frequency and an increase in BOLD signal estimation, which is then
identified in the fMRI analysis around TR 100 (corresponding to ~200s).
FIGURE 3.18: Activation Z statistic maps (top) and corresponding average time courses, partial and
full model fits (arbitrary units), plotted as a function of the volume number (bottom), obtained for
patient JQ with the neurophysiologist’s regressor .
3.8
2.3
R L
41
FIGURE 3.19: EEG scalp maps (reference channel: FCZ) and activation Z statistic maps for the ICs
with compatible fMRI activation maps with the neurophysiologist’s regressor . The EEG metric
shown is the one yielding the largest activation map for that component.
FIGURE 3.20: 12 th IC spectrogram (top) and corresponding regressor and mean BOLD time course
over the activation map yielded in fMRI analysis (bottom). The boxes in red represent the periods
marked as ictal events by the neurophysiologist .
IC 12 from Subject JQ, Sequence 1
-
+
IC 19 from Subject JQ, Sequence 1
-
+
IC 14 from Subject JQ, Sequence 1
-
+
Time (s)
Fre
quency (
Hz)
50 100 150 200 250 300 350
2
4
8
16
32
3.9
2.3
14th
IC
, u
MF
4.3
2.3
19th
IC
, R
MS
F
4.1
2.3
12th
IC
, MF
L R
R
R L
42
3.2.3 Subject ABC
Subject ABC had one sequence with marked ictal events. The EEG from this sequence was pre-
processed with both EEGLAB base procedure and Neuroscan procedure, being the later the one used by
the neurophysiologist to examine the data. Significant differences between the outputs of the
procedures were observable, specifically in the spectral profile of the one IC that correlated with the
neurophysiologist’s markers, which is presented in Figure 3.21. In both ICs, channel F8 was the one
that contributed most significantly. For this channel, an example time course detail of the second
marked ictal episode is presented in Figure 3.22, using the Neuroscan pre-processing. The time course
from a neighbour channel, which was rejected, is also presented for comparison. It is clear that the
spectral alterations around 10 Hz on the IC are highly correlated with the slice gradient artefacts (with
period of ~0.09 s), thus probably themselves also of artefactual nature. The correlation with the slice
gradient artefacts in the EEG pre-processed with the EEGLAB procedure was not as clear, for the
method adapts faster to artefact shape alterations, still it would not be appropriate to interpret this as
true EEG features. Therefore, the fMRI analysis performed on this dataset may not be meaningfully
interpreted, even though some of the EEG metric regressors based on these components yielded
similar results to the neurophysiologist’s regressor (not shown).
FIGURE 3.21: Spectrograms for the ICs that correlated with the neurophysiologist’s ictal markers,
using Neuroscan procedure pre-processing (top) and EEGLAB procedure (bottom).
Time (s)
Fre
quency (
Hz)
50 100 150 200 250 300 350
2
4
8
16
32
Time (s)
Fre
quency (
Hz)
50 100 150 200 250 300 350
2
4
8
16
32
43
FIGURE 3.22: EEG time courses form channel F8 (top) and channel FC4 (bottom) of subject ABC using
as pre-processing method the Neuroscan procedure.
3.2.4 Subject JB
Subject JB had one sequence with only one marked ictal event. Comparing the motion parameters
obtained in the fMRI pre-processing with the ictal marker indicated by the neurophysiologist shown
in Figure 3.23 , a pronounced head movement (about 0.7 mm) correlated with the beginning of the
ictal event is observable.
FIGURE 3.23: Relative mean displacement yielded by MCFLIRT (black) and neurophysiologist’s ictal
marker (red) for patient JB ictal sequence.
When analysing the spectrograms of the IC analysis, the majority of the components showed
spectral correlations with the onset of the ictal event. These correlations were high amplitude waves
with disperse spectral content which affected all EEG channels and was not linearly separable from
the rest of the signal using ICA, all of which suggest the association with motion artefacts. When
compared with the mean relative displacement parameter, the ICs displayed power correlations with
both peaks presented in Figure 3.23, at about 100 s and 210 s.
From the group of EEG metrics applied to the ICs, the MF and RMSF, which are power
normalized, were less sensitive to the presence of the motion artefacts and supported proceeding to
the fMRI analysis. All components and all metrics were included in the fMRI studies.
Activat ion volumes
For the components that yielded significant activation maps, the volume of the activation maps
obtained for each component and each metric is plotted in Figure 3.24. A larger number of
0 50 100 150 200 250 300 350
0
0.5
1
Time (s)
Mean d
ispla
cem
ent
(mm
)
1 mV
0.2 s
44
components appeared with significant activation maps for the metrics not power normalized (uMF,
uRMSF and TP), still two of the three largest activation maps were obtained by components 1 and 7
with power normalized metrics (MF and RMSF).
FIGURE 3.24: Activation map volumes obtained for each metric and each component of patient JB. For
each IC, the volumes are plotted in descending order for the different metrics.
Activat ion maps and t ime courses
The fMRI activation maps and IC scalp maps for the three regressors yielding the largest
activation volumes are presented in Figure 3.25, along with the activation map obtained for the
neurophysiologist’s regressor.
Patient JB had temporal lobe epilepsy with presumed epileptic focus in the left temporal pole. This
area only appeared activated with regressors from the 1st and 7th components, and more vastly with
the later. Regarding the scalp maps from these components, some consistency with the activation
maps is obtained. Specifically, component 1 scalp map presents a wide dipolar topography inclined
towards the posterior left hemisphere (reference channel: PZ), which comes in agreement with the
fMRI activation clusters obtained mainly in the left hemisphere. The scalp map of the 7th IC is less
defined and more difficult to interpret, not being referenced to a mean channel. Nevertheless a
dipolar configuration in the left temporal lobe is apparent, which would be consistent with the
activation cluster obtained in the left temporal pole.
0
500
1000
1500
2000
2500
1 7 21Vo
lum
e (
vo
xe
ls)
Component Inex
MF
0
500
1000
1500
2000
7 1 21
Component Inex
RMSF
0
500
1000
1500
2000
22 20 6 19 11 16 1Vo
lum
e (
vo
xe
ls)
Component Inex
uMF
0
500
1000
1500
2000
22 20 6 11 16 19
Component Inex
uRMSF
0
1000
2000
3000
4000
22 20 6 18 10 19 16 1 2 12 4
Component Inex
TP
45
FIGURE 3.25: EEG scalp maps (reference channel: PZ) (left) and activation Z statistic maps for the IC
EEG metrics with largest fMRI activation maps (right). It is also presented the activation map
obtained with the neurophysiologist’s regressor (top).
When observing the scalp map from the 22nd IC, a multipolar configuration is patent. This along
with the ‚ring‛ configuration of the fMRI activation map obtained points towards a movement
artefactual nature for this component. When plotting, in Figure 3.26, the regressors from Figure 3.25
along with the mean displacement estimated by MCFLIRT, the correlation between the 22nd IC TP
metric and the movement parameters is clear. The other regressors appear less correlated with the
IC 22 from Subject JQ, Sequence 1
-
+
IC 1 from Subject JQ, Sequence 1
-
+
IC 7 from Subject JQ, Sequence 1
-
+
3.9
2.3
7th I
C,
RM
SF
6.1
2.3
1st I
C, M
F
4.1
2.3
22n
d I
C, T
P
3.9
2.3
Neurophysiologist’s
reg
ress
or
L R
R L
46
movement parameters (the regressors presented are convolved with the HRF, thus delayed
correspondingly).
FIGURE 3.26: Relative mean displacement yielded by MCFLIRT (black) along with the EEG metric
regressors used in fMRI analysis that yielded the largest activation maps.
3.2.5 Subject GM
Subject GM had one sequence with one ictal event indicated by the neurophysiologist. From the
ICA decomposition and spectral analysis, only one IC had a spectral profile not correlated with the
marked ictal event. From the 16 ICs that proceeded to fMRI analysis, only 10 yielded significant
activation maps and their scalp maps are presented in Figure 3.27.
FIGURE 3.27: Scalp maps for the ICs that yielded significant activation maps in fMRI analysis for
patient GM (reference channel: CZ).
Activat ion volumes
A set of plots that summarizes the volume of the activation maps obtained for each metric and
each component of patient GM is presented in Figure 3.28. It is observable that the normalized
metrics yielded larger activation maps, and for more ICs.
0 50 100 150 200 250 300 350-0.5
0
0.5
1
Time (s)
Mean d
ispla
cem
ent
(mm
)
Mean displacement
IC 1 MF
IC 22 TP
IC 7 RMSF
1 3 4 5 6
11 12 14 15 16
-
+
Subject GM, Sequence 1
47
FIGURE 3.28: With reference to patient GM. For each metric it is plotted, in descending order and as a
function of the IC used, the volume of the activation maps obtained on the fMRI analysis.
Activat ion maps
When examining the activation maps obtained, component 4 outstood from the others by yielding
an activation map remarkably concordant with the activation map obtained with the
neurophysiologist’s regressor and the clinical profile of the patient. This activation map, along with
the one yielded by the neurophysiologist’s regressor, are presented in Figure 3.29.
FIGURE 3.29: Activation maps for patient GM, sequence 1, using the neurophysiologist’s regressor
(top) and the regressor obtained using RMSF metric on the 4 th IC (bottom).
0
200
400
600
800
1000
12 16 4 6 14 1 11 3Vo
lum
e (
vo
xe
ls)
Component Index
MF
0
500
1000
1500
4 12 16 3 14 6 11 1 15 5
Component Index
RMSF
0
20
40
60
80
100
16 6Vo
lum
e (
vo
xe
ls)
Component Index
uMF
0
50
100
150
16
Component Index
uRMSF
0
50
100
150
200
4 16 6
Component Index
TP
4.5
2.3
4.0
2.3
R L
48
Comparing the activation maps, it is observable that both regressors yielded activated clusters
involving mostly the left occipital lobe, which, as referred in Table 2.1, was the target of epilepsy
surgery. The volume of the activation map and the maximum Z score were higher for the EEG metric
regressor. This regressor also yielded an active cluster in the cerebellum that appears activated for the
neurophysiologist’s regressor if the clustering threshold is lowered. Additional activation clusters
appear in the frontal cortex associated with the 4th IC RMSF regressor, but these are not as significant
as the previously referred clusters, although they are relatively large.
Comparing the activation map of the 4th IC with its scalp map, a correlation between the main
activation clusters and the most weighted channels becomes apparent, with both showing a trend
towards the left occipital lobe. This also comes in agreement with the neurophysiologist’s
interpretation of the EEG, which values events occurring mostly on channels O1 and P7.
3.2.6 Summary
When examining the results presented in this section, patient RR clearly outstands from the other
patients in terms of the data SNR, reflected in the activation volumes and maximum Z values
obtained both for the neurophysiologist’s regressor and for the EEG metric regressors.
For this patient, in sequence 2, 10 out of 10 of the pre-selected ICs yielded, for some EEG metric,
larger activation maps than the neurophysiologist’s regressor. For sequence 3 this was achieved for 6
out of 9 of the ICs. All of the activation maps referred were generally concordant with each other and
with the neurophysiologist’s regressor’s activation map. The increased activation volumes and
maximum Z values obtained reflect the superior prediction of the BOLD signal provided by the
frequency weighted EEG metrics. To note is the fact that the only ICs for which the total power metric
yielded the largest activation maps, were the ones with smaller activations than the original
neurophysiologist regressor.
For the sequences of patient RR on which no ictal events were identified, the methodology
presented was able to identify the same brain network that was found on the ictal sequences. This
was achieved by using regressors with the same IC scalp topography as the most relevant ICs for the
ictal sequences. These findings support the idea of an underlying network of functionally connected
brain regions, which occasionally manifests itself with an ictal character.
Subjects JQ, JB and GM, also presented at least one EEG metric, for one IC, which predicted the
BOLD signal better than the neurophysiologist’s regressor. The activation maps obtained were in
good agreement with the patient’s clinical history and, for these patients, all the metrics considered
concordant were power normalized. The scalp maps of the corresponding ICs and their general
spatial correlation with the activation maps obtained from the fMRI analysis, suggested a non-
artefactual nature of the results.
49
Subject ABC EEG data contained residual slice gradient artefacts which precluded further fMRI
analysis with spectral metrics based upon them.
3.3 HRF STUDY
The HRF studies were focused only on patient RR as this was the only one whose data had
sufficiently high temporal SNR. Furthermore, only sequences 2 and 3 were used, since they were the
only ones where ictal events were identified by the neurophysiologist. The regressors used for the
HRF estimations were those that yielded the largest activation maps, i.e., 16th IC with RMSF metric for
sequence 2 and 10th IC with MF metric for sequence 3. Firstly the ROI based analyses will be
addressed as they were also the starting point for the pixel-by-pixel approaches.
3.3.1 ROI based analysis
Using the entire activation maps as ROI, the HRFs obtained by the three methods implemented in
this study are presented in Figure 3.30. The BOLD time courses resulting from the HRFs estimated are
displayed in Figure 3.31. The FIR estimations were carried out using a convolution kernel with length
17 TR and the Biophysical Model was estimated using the priors in Table 1.1.
FIGURE 3.30: HRFs estimated by Biophysical Model (blue), IIR filter (green) and FIR filter (cyan), for
sequence 2 (left) and sequence 3 (right) using the whole activation map as ROI.
0 20 40 60 80 100-0.2
-0.1
0
0.1
0.2
0.3
Time (s)
Am
plit
ude (
std
units)
0 20 40 60 80 100-0.2
-0.1
0
0.1
0.2
0.3
Time (s)
Am
plit
ude (
std
units)
50
FIGURE 3.31: BOLD time courses obtained using the HRFs presented in Figure 3.30: Balloon Model
(blue), IIR filter (green) and FIR filter (cyan). The dashed grey is the ROI BOLD mean time course.
Results displayed for sequence 2 (top) and sequence 3(bottom).
Comparing the HRFs obtained by each of the methods, the wider character of the ones estimated
by the less constrained methods, IIR and FIR, contrasts with the narrower (close to the prior) estimate
of the Biophysical Model. Regarding the BOLD time course fits, the IIR and FIR models appear again
more flexible, yielding better fits than the Biophysical Model. Although the Biophysical Model fit
yielded HRF’s not significantly different from the canonical, tests were performed by relaxing the
priors and disperse HRFs similar to those yielded by the IIR model approach were obtained, with
minimal residues for the BOLD estimates (results not presented). It should also be noted that the
lower frequencies of the HRFs were filtered from the BOLD time course and hence their estimation is
badly conditioned in the more free methods. This way, the mean value of the FIR HRF, for example, is
poorly estimated.
The activation maps obtained from patient RR involved five main anatomical areas: left frontal
lobe, left occipital lobe, left parietal lobe, left hippocampus and the hypothalamic hamartoma. These
ROIs were the target of further study and their HRFs are presented in Figure 3.32.
Fixing each model used, the HRF estimations are consistent across sequences for each ROI.
Specifically, using the IIR filter model, the HRFs estimated for the hippocampus and for the
hamartoma appear peaking later than the remainder, and, between these, the HRF for hippocampus
appears sharper than the one for the hamartoma. Moreover, the similarity between the HRFs for the
frontal, occipital and parietal lobes is consistent between sequences. When using the FIR model, the
similarities between the estimates for the different sequences are not as clear. In regard to the
Biophysical Model estimates, again the hippocampus and the hamartoma appear with HRFs more
0 50 100 150
-0.2
-0.1
0
0.1
0.2
Time (TR)
Am
plit
ude (
std
units)
0 50 100 150
-0.2
-0.1
0
0.1
0.2
Time (TR)
Am
plit
ude (
std
units)
51
closely related, yet here the reason for this is the lower level of evidence provided by the model fit to
their mean BOLD time course. They have smaller volumes than the frontal, occipital and parietal
ROIs, consequently they also present lower levels of SNR and their HRF estimates become tighter to
the prior.
52
FIGURE 3.32: HRFs estimated using IIR model (top), FIR model (centre) and Biophysical Model
(bottom), for sequences 2 (left) and 3 (right). The ROIs were the entire activation map, hamartoma,
left hippocampus, left occipital lobe, left frontal lobe and left parietal lobe.
0 20 40 60 80 100
-0.1
-0.05
0
0.05
0.1
0.15
0.2
Time (s)
Am
plit
ude (
std
units)
0 20 40 60 80 100
-0.1
-0.05
0
0.05
0.1
0.15
0.2
Time (s)
Am
plit
ude (
std
units)
0 20 40 60 80 100
-0.2
-0.1
0
0.1
0.2
Time (s)
Am
plit
ude (
std
units)
0 20 40 60 80 100
-0.2
-0.1
0
0.1
0.2
Time (s)
Am
plit
ude (
std
units)
5 10 15 20
0
0.05
0.1
Time (s)
Am
plit
ude (
std
units)
5 10 15 20
0
0.05
0.1
Time (s)
Am
plit
ude (
std
units)
53
3.3.2 Pixel-by-pixel analysis
Basis functions
A set of basis functions, presented in Figure 3.33, was firstly derived from sequence 3, whole
activation map ROI, IIR model HRF estimation, and is used in this section. This choice was based on
the higher level of model fit and HRF time course continuity obtained by this estimation in relation to
the other methods. The IIR estimation for the same ROI of sequence 2, although with the same degree
of data fit, presented a significant discontinuity in the initial rise of the HRF, thus yielding time
derivatives too sharp, which severely limit the degree of time delay and dispersion the function basis
of may account for.
FIGURE 3.33: HRF basis set, derived from IIR filter model estimation for the whole activation map of
sequence 3. ‚Centre‛ HRF (blue), time derivative (green) and dispersion derivative (red).
Figure 3.34 and Figure 3.35 synthesize the results from the HRF study using the ROI based basis
functions presented, for sequence 2 and 3. For the HRF basis presented, higher relative values for β2
(time derivative coefficient) represent earlier HRFs and higher relative values for β3 (dispersion
derivative) represent sharper HRFs. This way, for example, orange regions represent late and wide
HRFs, while blue regions display earlier and sharper HRFs. This relation is graphically depicted on
the bottom left image of the figures. It should be noted that different scales were used for the axes of
the colormaps, being the sequence 2 coefficients biased towards sharper and earlier HRFs, and
sequence 3 more centred, as its estimated mean IIR HRF was the one used for the construction of the
function basis.
When examining the scatter plots for both sequences, it can be seen that no specific profile can be
found for the HRFs of a given anatomical region. Though the centre of mass of the dots is different
from ROI to ROI, the disperse character of the dot pattern does not encourage the attribution of a
single representative HRF to a given ROI.
Taking into account the HRF mapping, some regions display a higher homogeneity. An example is
the superior activated region of the occipital cortex which presents mainly a blueish white colour; this
represents a sharper and slightly earlier HRF than the ‘centre’ HRF. This is visible for both sequences.
0 20 40 60 80 100
-1
-0.5
0
0.5
1
Time (s)
Am
plit
ude (
std
units)
54
A feature also present in both sequences is the wider and later HRF in the inferior left hippocampus
region, which is translated into a red/orange colour in the HRF maps (more visible in sequence 3).
This comes in agreement with the ROI HRF estimates for the hippocampus, which also appear
peaking slightly later than the remainder ROIs. The superior left frontal lobe also presents a
consistent trend towards sharper and earlier HRFs, which is also reflected in the ROI estimates.
SEQUENC E 2 - ROI DERI VED BASIS FUNC TION
FIGURE 3.34: Distribution of the normalized β2 and β3 across the brain (top). Example HRFs,
translating the colormap of the image (bottom left). Scatter plot on which each dot represents a voxel
with the coordinates being its normalized β2 and β3 (bottom right). Dots in black correspond to
frontal lobe voxels, in red to occipital lobe, cyan to parietal lobe, green to hippocampus and blue to
hamartoma.
0 10 20 30 40
-1
0
1
2
Time (s)
Am
plit
ude (
std
units)
2 (normalized)
3 (
norm
aliz
ed)
-0.5 0 0.5
0
0.5
1
55
SEQUENC E 3 - ROI DERI VED BASIS FUNC TION
FIGURE 3.35: Distribution of the normalized β2 and β3 across the brain (top). Example HRFs,
translating the colormap of the image (bottom left). Scatter plot on which each dot represents a voxel
with the coordinates being its adjusted β 2 and β3 (bottom right). Dots in black correspond to frontal
lobe voxels, in red to occipital lobe, cyan to parietal lobe, green to hippocampus and blue to
hamartoma.
For the HRF studies using filmbabe script, a FLOBS set of basis functions was created using the ROI
estimates as indication for the parameter choice. The set of 3 basis functions was created with m1∈[0-
0s], m2∈[4-12s], m3∈[20-30s], m4∈[0-0s] and c∈[0-0] (see Figure 2.1). The set of 1000 samples and the
resulting function basis are presented in Figure 3.36. The three functions constituting the basis are
roughly equivalent to a ‘mean’ HRF (blue), its time derivative (green) and its dispersion derivative
0 10 20 30 40
-1
0
1
2
Time (s)
Am
plit
ude (
std
units)
2 (normalized)
3 (
norm
aliz
ed)
-0.5 0 0.5
-0.5
0
0.5
56
(red). One feature to notice is the symmetry between the dispersion derivatives from the ROI derived
basis function and the ones from the FLOBS. From this results that, on the fit of the FLOBS basis,
higher values for β3 correspond to wider HRFs (contrary to the fit of the ROI derived basis). This way,
to improve the interpretability of the data, the results were adapted to represent the basis displayed
in Figure 3.37, thus, in the results displayed, higher values for β3 correspond to sharper HRFs in
concordance with the ROI derived basis results.
FIGURE 3.36: FLOBS sample HRFs (left) and resulting set of basis functions (right) for parameters
m1∈[0-0s], m2∈[4-12s], m3∈[20-30s], m4∈[0-0s] and c∈[0-0].
FIGURE 3.37: FLOBS basis functions, with symmetric ‘dispersion’ base function, used for results
display.
Figure 3.38 and Figure 3.39 synthesize the results from the HRF study with filmbabe script. The
most notorious difference from the regular FEAT is the constraints on the admissible HRFs, which are
translated into more limited intervals on the axes of the colormaps and more plausible HRFs
represented in the bottom left images. The constraint is particularly strong on the third basis function
(dispersion derivative) and this is observable by the narrow range of widths displayed in the bottom
left plots when fixing the value of β2 (corresponding to the delay). This way, changes of colour in a
vertical direction on the colormap have only a small effect on the HRF.
Using this methodology, superior frontal activated areas still show evidence towards an earlier
peaking HRF, displaying purple/blue colours. Regarding the sharpness of the HRF, there was no clear
tendency for this region, as the prior variance for β3 parameter was small and hence the possibility for
fitting different dispersions was limited. The occipital areas showed ‘central’ delays (red white cyan)
but, again, no clear tendency was observed on the dispersion. Inferior hippocampus areas continued
0 10 20 30 40-0.1
-0.05
0
0.05
0.1
Time (s)
Am
plit
ude (
std
units)
57
to show a bias towards later peaking HRFs. These observations come in agreement with the results of
the ROI derived basis functions, but fitting more plausible HRFs than those fitted on the regular
FEAT analysis.
SEQUENC E 2 - FLOBS BASIS FUNC TION (FILMBABE )
FIGURE 3.38: Distribution of the normalized β2 and β3 across the brain (top). Example HRFs,
translating the colormap of the image (bottom left). Scatter plot on which each dot represents a voxel
with the coordinates being its adjusted β 2 and β3 (bottom right). Dots in black correspond to frontal
lobe voxels, in red to occipital lobe, cyan to parietal lobe, green to hippocampus and blue to
hamartoma.
0 10 20 30 40-0.05
0
0.05
0.1
Time (s)
Am
plit
ude (
std
units)
2 (normalized)
3 (
norm
aliz
ed)
-0.2 0 0.2 0.4
-0.02
0
0.02
0.04
58
SEQUENC E 3 - FLOBS BASIS FUNC TION (FILMBABE )
FIGURE 3.39 Distribution of the normalized β2 and β3 across the brain (top). Example HRFs,
translating the colormap of the image (bottom left). Scatter plot on which each dot represents a voxel
with the coordinates being its adjusted β 2 and β3 (bottom right). Dots in black correspond to frontal
lobe voxels, in red to occipital lobe, cyan to parietal lobe, green to hippocampus and blue to
hamartoma.
Exhaust ive search
The exhaustive search approach was also guided by the previous ROI based analysis, centring the
search around the values for the HRFs that better fitted the total activation map, ROI IIR model
estimate for sequence 3. The discretization of the HRF parameter space was Tup ∈ { 5 8 4 7 0}
0 10 20 30 40-0.05
0
0.05
0.1
Time (s)
Am
plit
ude (
std
units)
2 (normalized)
3 (
norm
aliz
ed)
-0.2 0 0.2
-0.04
-0.02
0
0.02
59
and Tmax ∈ {0 6 8 0 }. h vo u s su ti o A ysis usi th s sso s are
presented in Figure 3.40.
FIGURE 3.40: Volume of the activation maps (number of voxels) as a function of Tup and Tmax for
sequence 2 (left) and sequence 3 (right). The largest activation for sequence 2 (5956 voxels) is
obtained with Tup 11 s and Tmax 6 s and for sequence 3 2 (6638 voxels) with Tup 17 s and Tmax 2 s.
The exhaustive search approach was the only method applied which allowed non causal HRFs
(HRF(t)≠0 for t<0). This proved to be important for the parameterization in practice, as the optimal
values for Tup and Tmax, in the sense of yielding larger activation maps, corresponded to non-causal
HRFs (Tup > Tmax). The number of voxels of the activation map showed a co-dependency between
the parameters Tup and Tmax, despite the effort of the reparameterization.
The mapping of the HRF parameters yielding the highest Z value for each voxel gave rise to
results compatible with those of the basis function approaches. The rough discretization of the
parameter space produced maps with rougher resolution on the parameters, thus redundant to be
displayed.
Tmax (s)
Tup (
s)
0 2 4 6 8 10 12
2
5
8
11
14
17
20 1000
2000
3000
4000
5000
Tmax (s)
Tup (
s)
0 2 4 6 8 10 12
2
5
8
11
14
17
20 1000
2000
3000
4000
5000
6000
60
Chapter 4
DISCUSSION
In this chapter the main results obtained in this Thesis will first be briefly summarized and then
discussed in detail. Firstly the EEG pre-processing procedures will be evaluated, followed by the
results and implications of the EEG metric derived BOLD regressors. The HRF studies will then be
addressed and some final remarks will be presented.
4.1 SUMMARY
A number of methodologies were developed for the analysis of EEG-fMRI data acquired
simultaneously during epileptic seizures and applied to five cases of patients undergoing pre-surgical
evaluation. Firstly, the correction of the fMRI gradient and balistocardiographic artefacts present on
the EEG data was addressed, by overcoming specific difficulties and implementing a fully automatic
algorithm. Secondly, a processing pipeline for the EEG data was developed in order to obtain seizure-
related regressors for the analysis of the fMRI BOLD signal. This pipeline included ICA
decomposition, Wavelet spectral analysis and application of different model-based metrics aimed at
testing existing hypothesis concerning the transfer function between EEG and fMRI data. The results
obtained support the heuristic model whereby increases in hemodynamic signals should be
associated with a power increase in higher EEG frequencies, relative to lower frequencies. Finally, a
comprehensive study of the hemodynamic response to epileptic seizures was performed for one
patient, using a range of different HRF estimation approaches. In general, significantly wider
responses were found compared with the standard HRF in the healthy brain. Moreover, the HRF
delay and dispersion characteristics revealed heterogeneity between and within brain regions.
4.2 EEG PROCESSING AND FMRI INTEGRATION
Pre-process ing
The effort taken in terms of improving the slice gradient correction obtained with the preliminary
Neuroscan approach turned out to be fruitful and the main source of improvement was the slice
timing correction performed. The origin of the irregularities observed in the slice timings was not
completely identified; although it could be related to errors in the timing of the MR sequence itself, it
is more likely to be due to some imprecision in the digitalization and pre-processing procedures of
the EEG data performed by the hardware. Independently from this, the algorithm developed
61
successfully identified the slice timings referenced to the nominal sampling frequency of the EEG
data. This allowed the improved artefact removal presented in Section 3.1.1. It is noteworthy that only
with the EEGLAB procedure implementation were plausible results obtained for patients GM, JB and
JQ.
The effects of the correction procedures employed in this work, both slice gradient and
balistocardiographic artefact, may be further examined with spectral-temporal analysis, before and
after data processing, which will be addressed in future work. Nevertheless, the ERP images
displayed reflect the effectiveness of the artefact removal procedures, being only questionable the
degree of information loss or the incorporation of artificial components induced by the methods
employed.
The choice of the frequency band to analyse was a somewhat subjective choice. Although the
frequency limits presented in this Thesis are commonplace (Rosa, et al. 2010), they do discard higher
gamma band activity that presents considerable power above 45 Hz. Also, the low frequencies were
fairly cut at 2 Hz ignoring lower delta band activity and very low frequency activity, which have been
previously correlated with BOLD fluctuations (Leopold, Murayama and Logothetis 2003).
Nevertheless, the frequency band limits used are believed to include most of the ictal activity power
while still retaining good levels of SNR, which are difficult to achieve for higher frequencies due to
the residual presence of slice gradient artefacts and, for lower frequencies, due to instrumental
baseline drifts.
ICA
The use of ICA decomposition of the EEG during seizures may be questionable, because the
assumption of the spatial stationarity of the sources may not be verified due to the propagation of
ictal activity. However, no evidence could be found for such non-stationarity, which was previously
reported as very unlikely (Debener, et al. 2010). Nevertheless, our results indicate that ICA may be a
useful tool for the separation of epileptic activity from background neuronal activity as well as
residual artefacts not fully corrected by the EEG pre-processing.
An additional remark on the EEG ICA incorporation in fMRI analysis is its data-driven character
on the EEG analysis plane and its model-based inclusion in the fMRI analysis. This way, the
hemodynamic changes observed in fMRI are representative of EEG time courses whose topographies
are known and may be interpreted along with the time signal by the neurophysiologist. Also, the
presented methodology avoids the subjective and sometimes difficult task of identifying the onset
and offset of the ictal events. A limitation of this approach is related with the bias of the EEG towards
superficial cortical activity, which possibly precludes the identification of hemodynamic changes not
correlated with some superficial cortical activity.
62
Spectral Analys is Method
The adoption of a wavelet decomposition method for the spectral analysis, as opposed to a regular
Fourier spectrogram analysis which fixes the temporal resolution, was focused on achieving a better
temporal description for higher frequencies present in transient features of the EEG signal. The use of
Morlet wavelets led to a logarithmic sampling of the frequency domain in order to, firstly, not have
‘frequency domain aliasing’ for the lower frequencies and, secondly, to save memory not
oversampling the higher ones. The sampling on the time domain could also be adjusted if a more
memory efficient procedure was to be achieved: the lower frequencies of the Morlet wavelet
decomposition need not have a time sampling frequency as high as the one for the higher frequencies.
EEG metrics and fMRI result s
In this work, the volume of the activation map was used as an indirect measure of how well a
certain regressor explains the BOLD signal across a network of brain areas. Other measures could
have been used, such as the Z score or the residues of the GLM fit, or the variance explained by the
regressors. Although these measures could more directly assess how well each EEG regressor tested
explains the BOLD data, the relative results should remain the same as those obtained using the
activation volume.
In general, for the four patients studied, it was possible to identify at least one IC and one metric,
in each patient, that produced regressors yielding activation maps in agreement with the clinical
expectation for the seizure propagation network and which predicted the observed BOLD signal
better than the neurophysiologist’s regressor (hence producing larger and more significant activation
maps). The scalp maps of the corresponding ICs showed a good spatial correlation with the activation
maps obtained from the fMRI analysis, which suggests a non-artefactual nature.
Our results consistently show superior performance of frequency-weighted EEG metrics, in
support of the heuristic model proposed by Kilner (Kilner, et al. 2005). However, we could not find a
superior BOLD prediction power of the normalized metrics when compared with the un-normalized
ones, as was reported by Rosa and colleagues (Rosa, et al. 2010) in a visual stimulation experiment.
When interpreting the results in light of the heuristic model, the power normalization factor comes
un-accounted, as increases in neuronal activity are modelled as accelerations in neural dynamics,
resulting in an increase of the root mean square frequency of the EEG spectrum, maintaining the total
power constant. No account is made in terms of alterations of synchrony or number of active units,
which would be reflected in the EEG total power. This way, the results presented do not contradict
the model proposed; however unveil one of its limitations.
The methodology presented approaches the EEG-fMRI data integration in a directional fashion:
the EEG high temporal resolution is used to predict the spatially resolved fMRI data and identify the
63
brain regions correlated with the former. The referred directionality may be overcome by including
fMRI information in the optimization of the EEG ICA procedure, resorting to one of the EEG metrics
proposed. Though some non-linearities appear, the un-normalized root mean square frequency
metric may be approximated by the power of the first temporal derivative of the EEG time course,
simplifying the optimization procedure, and resulting in:
(4.1)
where is the first time derivative EEG matrix including all channels, is the weight vector to be
optimized (initialised by the most relevant IC weight vector, for example), and are the
haemodynamic response convolution matrix and the resampling matrix, respectively, is the
observed BOLD time course, possibly actualised iteratively, and is the BOLD measurement noise.
The square operates point by point. The total power metric would correspond to the substitution of
the first time derivative by the original EEG time courses. The inclusion of normalized metrics would
require an extra non-linearity. Also, noise still present in the EEG measures would continue to come
unaccounted and is one other possible source of improvement of the methodology proposed in this
Thesis.
4.3 HRF STUDY
The HRF study for the ictal events relied on the superior description of the seizure dynamics
achieved by the EEG metrics explored in the first part of this Thesis, since the statistical power of the
ictal events was not adequate to follow regular HRF estimation procedures as those usually
performed for interictal activity, for example (Lu, et al. 2007). For patient RR, the quality of the BOLD
prediction yielded by the EEG metrics allowed model least squares fit to the experimental BOLD
signal, discarding the usually mandatory priors on the model parameters. However, the use of a least
squares method may be questionable, since the BOLD noise (un-modelled signal) did not display a
white Gaussian noise behaviour. This may have led, to some extent, to an over fit of the models to the
data, particularly when using the IIR model due to its long range oscillatory character.
The most striking feature regarding the HRF estimates obtained was the dispersion of the
temporal responses yielded by the IIR, FIR and pixel-by-pixel based approaches, which was
significantly deviated from the dispersion commonly found in the healthy brain. Although the
Biophysical Model fit yielded HRF’s not significantly different from the canonical, when the priors
where relaxed, disperse HRFs similar to those yielded by the IIR model approach were observed with
minimal residues for the BOLD estimates (results not presented). These findings indicate that the
general behaviour of the haemodynamic response to the EEG metrics explored is still captured by the
model structure.
64
When interpreting the larger dispersion observed for the estimated HRFs, one should also take
into account the administration of anaesthesia during the EEG-fMRI acquisition. Vasomotor
alterations during anaesthesia have been described in the literature. Specifically, anaesthesia with
sevoflurane inducing alterations in default mode network coherence was described in (Deshpande, et
al. 2010). Also, it has been hypothesised that anaesthesia (thiopental in children) may lead to altered
feedback in neurovascular coupling, explaining the observed reduction of the frequency of slow
fluctuations on cortical blood flow (Kiviniemi, et al. 2000). This would be concordant with the
increased dispersion of the estimated HRF, corresponding to slower feedback dynamics in
neurovascular coupling.
The pixel-by-pixel approaches led to some insight on the distribution of the HRF features across
the patient’s brain. A first result was the variability of the haemodynamic response estimates, not
only between, but also within brain areas, which could not be appreciated in the ROI based analysis
due to the averaging of BOLD time courses in the anatomical driven regions. In fact, these exhibited
some degree of HRF variability within themselves. The emergence of zones of higher HRF
homogeneity appeared with some significance, as subgroups of the previously described anatomical
ROIs.
The inclusion of priors on the covariance of the basis functions coefficients yielded more
physiologically plausible results for the estimated HRFs, yet the covariance terms for the third base
function, corresponding to dispersion features, were sub-estimated from the learning set, yielding
limited freedom for the estimation of this parameter across the subject’s brain. A more diverse
learning set in terms of dispersion of the HRFs would be interesting to investigate. Also, the
implementation of the VOI derived HRFs into a FLOBS structure, with priors learnt from an artificial
HRF set would be an approach worth pursuing.
4.4 CONCLUSION
This work presented a new approach to EEG-fMRI data integration in the field of epilepsy. The
methodology allowed better descriptions of the haemodynamic changes associated with ictal events
yielding plausible, broader and more significant activation maps than the usual boxcar event
description approach. An HRF study was performed in one patient, providing insight into the
haemodynamic variability distribution across the brain during epileptic seizures. Evidence towards a
wider HRF to the epileptic activity was found for that specific patient and EEG metrics applied.
Finally, a number of new perspectives for EEG-fMRI integration worth to be pursued are suggested
by the present work.
65
The major limitation of the work presented in the Thesis is the small size and heterogeneity of the
patient population studied, as well as the generally poor SNR of the data. However, these are
precisely the major difficulties in the analysis of seizures using EEG-fMRI and our results provide
clear directions for improving the integration of the two types of data in a potentially clinically useful
way. Of course the proposed methodologies need to be validated. In a first approach, the validity of
the ICs selected for fMRI analysis could be evaluated by comparing their scalp maps with the ones
obtainable from the ictal EEG performed outside the MRI scanner (and hence without the
contamination by artefacts). This approach is intended to be pursued in future work. Also an
interesting validation approach would be source analysis upon the EEG data, confirming the
apparent spatial correlation between both modalities. However, this falls out of the scope of this
work. Naturally, the only conclusive validation of the networks identified by our approach would
have to be obtained from more direct measures of seizure activity, such as the ones obtained by intra-
cranial recordings (Vulliemoz, et al. 2011).
Despite its limitations, which may be overcome in future work, this Thesis provided the following
original contributions to the field of EEG-fMRI studies in epilepsy:
- Improved fMRI-derived artefact correction of EEG data;
- Novel EEG processing protocol for improved EEG-fMRI integration in epileptic seizures;
- Novel evidence supporting the heuristic proposed in the literature for EEG-fMRI integration;
- Study of the HRF shape across the brain in a case of epileptic seizures.
66
REFERENCES Afonso, D, J Sanches, and M Lauterbach. ‚Neural physiological modeling towards a
hemodynamic response function for fMRI.‛ Conf Proc IEEE Eng Med Biol Soc, 2007: 1615-8.
Almeida, LB, L Thibault, J D Amaral, and A Plahkov. ‚Parameter adaptation in stochastic
optimization.‛ In On-line learning in neural networks, by David Saad. 1998.
Bagshaw, AP, et al. ‚Analysis with multiple haemodynamic functions and comparisons with
gadolinium-enhanced MR anciograms.‛ Hum Brain Map, 2004: 179-192.
Buxton, RB, EC Wong, and LR Frank. ‚Dynamics of Blood Flow and Oxygenation Changes During
Brain Activation: The Balloon Model.‛ Magn. Reson. Med., 1998: 855-64.
Comon, P. ‚Independent component analysis, A new concept?‛ Signal Process, 1994: 287-314.
Debener, S, J Thorne, TR Schneider, and FC Viola. ‚Using ICA for the Analysis of Multi-Channel
EEG Data.‛ In Simultaneous EEG and fMRI - recording, analysis and application, by U Ullsperger and S
Debener, 121-33. New York: Oxford University Press, 2010.
Deichmann, R, U Nöth, and N Weiskopf. ‚The Basics of Functional Magnetic Resonance Imaging.‛
In EEG-fMRI, by L Lemieux, 39-62. Heidelberg: Springer, 2010.
Delorme, A, and S Makeig. ‚EEGLAB: an open source toolbox for analysis of single-trial EEG
dynamics.‛ Journal of Neuroscience Methods, 2004: 9-21.
Deshpande, G, C Kerssens, PS Sebel, and X Hu. ‚Altered local coherence in the default mode
network due to sevoflurane anesthesia.‛ Brain Res., 2010: 110-21.
Fisher, R S, et al. ‚Epileptic Seizures and Epilepsy: Definitions Proposed by the International
League Against Epilepsy (ILAE) and the International Bureau for Epilepsy (IBE).‛ Epilepsia, 2005: 470-
72.
Friston, KJ. ‚Foreword.‛ In EEG-fMRI, by L Lemieux, v-vi. Heidelberg: Springer, 2010.
Friston, KJ. ‚Models of Brain Function in Neuroimaging.‛ Annu. Rev. Psychol, 2005: 57-87.
Friston, KJ, A Mechelli, R Turner, and C J Price. ‚Nonlinear Response in fMRI - The Balloon
Model, Volterra Kernels.‛ NeuroImage, 2000: 466–477.
Friston, KJ, AP Holmes, KJ Worsley, JB Pline, CD Frith, and RSJ Frackowiak. ‚Statistical
parametric maps in functional imaging: a general linear approach.‛ Hum. Brain Mapp., 1995: 189-210.
Friston, KJ, L Harrison, and W Penny. ‚Dynamic causal modelling.‛ NeuroImage, 2003: 000-000.
Friston, KJ, W Penny, C Phillips, S Kiebel, G Hinton, and J Ashburner. ‚Classical and Bayesian
inference in neuroimaging: theory.‛ NeuroImage, 2002: 465–483.
Glover, GH. ‚Deconvolution of impusle response in event-related BOLD fMRI.‛ NeuroImage, 1999:
416-29.
67
Goense, JB, and NK Logothetis. ‚Neurophysiology of the BOLD fMRI Signal in Awake Monkeys.‛
Current Biology, 2008: 631-40.
Gonçalves, SI, PJ W Pouwelsa, JPA Kuijera, RM Heethaara, and JC de Muncka. ‚Artifact removal
in co-registered EEG fMRI by selective average subtraction.‛ Clinical Neurophysiology, 2007: 2437-50.
Gray, RM. Toeplitz and Circulant Matrices: A review. Now Publishers, 2006.
Grouiller, F, L Vercueil, A Krainik, C Segebarth, P Kahane, and O David. ‚Characterization of the
Hemodynamic Modes Associated With Interictal Epileptic Activity Using a Deformable Model-Based
Analysis of Combined EEG and Functional MRI Recordings.‛ Hum Brain Map, 2010.
Hoffman, A, L Jager, K J Werhahn, M Jaschke, S Noachtar, and M Reiser. ‚Electroencephalography
during functional echo-planar imaging: Detection of epileptic spikes using post-processing methods.‛
Magn. Reson. Med., 2000: 791-798.
Hyvärinen, A, and E Oja. ‚Independent Component Analysis: Algorithms and Applications.‛
Neural Networks, 2000: 411-30.
Iriarte, J, et al. "Independent component analysis in the study of focal seizures." J Clin Neurophysiol,
2006: 551-8.
Ives, JR, S Warach, F Schmit, R R Edleman, and D L Schomer. ‚Monitoring the patient's EEG
during echo planar MRI.‛ Electroencephalogr Clin Neurophysiol, 1993: 417-20.
Jenkinson, M, PR Bannister, JM Brady, and SM Smith. ‚Improved optimisation for the robust and
accurate linear registration and motion correction of brain images.‛ NeuroImage, 2002: 825-41.
Kang, JK, et al. ‚Using patient-specific hemodynamic response functions in combined EEG-fMRI
studies in epilepsy.‛ NeuroImage, 2003: 1162-1170.
Kilner, JM, J Mattout, R Henson, and KJ Friston. ‚Hemodynamic correlates of EEG: A heuristic.‛
NeuroImage, 2005: 280-286.
Kiviniemi, V, et al. ‚Slow Vasomotor Fluctuation in fMRI of Anesthetized Child Brain.‛ Magnetic
Resonance in Medicine, 2000: 373–378.
Kobayashi, K, CJ James, T Nakahori, T Akiyama, and J Gotman. ‚Isolation of epileptiform
discharges from unaveraged EEG by independent component analysis.‛ Clin Neurophysiol, 1999: 1755-
63.
Lafus, H, et al. ‚EEG-correlated fMRI of human alpha activity.‛ NeuroImage, 2003: 1463-1476.
Lancaster, JL, et al. ‚Bias between MNI and Talairach coordinates analyzed using the ICBM-152
brain template.‛ Hum Brain Mapp, 2007: 1194–205.
Leal, A, and M Secca. ‚BOLD mapping of ictal and interictal epileptic activity in the evaluation for
surgery of epilepsy.‛ (submited) ., no. . (2010).
68
Leal, A, JP Monteiro, MFJ Secca, and C Jordão. ‚Functional brain mapping of ictal activity in
gelastic epilepsy associated with hypotalamic hamartoma: a case report.‛ Epilepsia, 2009: 1624-1631.
Leopold, DA, Y Murayama, and NK Logothetis. ‚Very slow activity fluctuations in monkey visual
cortex: implications for functional brain imaging.‛ Cereb. Cortex, 2003: 422–33.
LeVan, P, and J Gotman. ‚Independent Component Analysis As a Model-Free Approach for the
Detection of BOLD Changes Related to Epileptic Spikes: A Simulation Study.‛ Human Brain Mapping,
2009: 2021–31.
Logotetis, NK. "The neural basis of the blood-oxygen-level.dependent functional imaging signal."
Philos. Trans. R. Soc. London Ser. B, 2002: 1003-37.
Logothetis, N, and B Wandell. ‚Interpreting the BOLD signal.‛ Annu. Rev. Physiol., 2004: 66:735-69.
Logothetis, NK, J Pauls, M Augath, T Trinath, and A Oeltermann. ‚Neurophysiological
investigation of the basis of the fMRI signal.‛ Nature, 2001.
Lopes da Silva, F. ‚EEG: Origin and Measurment.‛ In EEG-fMRI, by L Lemieux, 19-38. Heidelberg:
Springer, 2010.
Lu, Y, C Grova, E Kobayashi, F Dubeau, and Gotman J. ‚Using voxel-specific hemodynamic
response function in EEG-fMRI data analysis: An estimation and detection model.‛ NeuroImage, 2007:
195–203.
Marques, JP, J Rebola, P Figueiredo, A Pinto, F Sales, and M Castelo-Branco. ‚ICA Decomposition
of EEG signal for fMRI Processing in Epilepsy.‛ Human Brain Mapping, 2009: 000-000.
Moeller, F, P Levan, and Gotman J. ‚Independent component analysis (ICA) of generalized spike
wave discharges in fMRI: Comparison with general linear model-based EEG-fMRI.‛ Hum Brain Mapp,
2010: 000-000.
Niazy, RK, CF Beckmann, GD Iannetti, JM Brady, and SM Smith. ‚Removal of FMRI environment
artifacts from EEG data using optimal basis sets.‛ NeuroImage, 2005: 720-37.
Ogawa, S, TM Lee, AS Nayak, and P Glynn. ‚Oxygenation-sensitive contrast in magnetic
resonance image of rodent brain at high magnetic fields.‛ Magn. Reson. Med., 1190: 14:68-78.
Pauling, L, and C Coryell. ‚The Magnetic properties and structure of hemoglobin.‛ Proc. Natl.
Acad. Sci. USA, 1936: 210-16.
Phillips, C, MD Rugg, and KJ Friston. ‚Anatomically informed basis functions for EEG source
localization: combininc functional and anatomical constrains.‛ NeuroImage, 2002: 678-95.
Rosa, MJ, J Kilner, F Blankenburg, O Josephs, and W Penny. ‚Estimating the transfer function from
neuronal activity to BOLD using simultaneous EEG-fMRI.‛ NeuroImage, 2010: 1496-509.
Salek-Haddadi, A, et al. ‚Hemodynamic correlates of epileptiform discharges: An EEG-fMRI
study of 63 patients with focal epilepsy.‛ BRAINRESEARCH, 2006: 148–66.
69
Shmmuel, A, M Augath, A Oeltermann, and N K Logothetis. ‚Negative functional MRI response
correlates with decreases in neural activity in monkey visual area V1.‛ Nat Neurosci, 2006: 000-000.
Smith, S M. ‚Fast robust automated brain extraction.‛ Hum Brain Mapp, 2002: 143-55.
Smith, SM, et al. ‚Advances in functional and structural MR image analysis and implementation
as FSL.‛ NeuroImage, 2004: 208-19.
Tallon-Baudry, C, and O Bertrand. ‚Oscilatory gamma activity in humans and its role in object
representation.‛ Trends Cogn Sci, 1999: 000-000.
Thornton, RC, et al. ‚Imaging haemodynamic changes related to seizures: comparison of EEG-
based general linear model, independent component analysis of fMRI and intracranial EEG.‛
NeuroImage, 2010: 000-000.
Urrestarazu, E, J Iriarte, J Artidea, M Alegre, M Valencia, and C Viteri. ‚Independent component
analysis separates spikes of different origin in the EEG.‛ J Clin Neurophysiol, 2006: 72-78.
Vig{rio. ‚Extraction of ocular artifacts from EEG using independent component analysis.‛
Electroencephalogr Clin Neurophysiol, 1997: 395-404.
Vulliemoz, S, et al. ‚Simultaneous intracranial EEG and fMRI of interictal epileptic discharges in
humans.‛ Neuroimage, 2011: 182-90.
Waites, AB, ME Shaw, RS Briellmann, A Labate, DF Abbott, and GD Jackson. NeuroImage, 2005:
192-99.
Wan, Xiaohong, Jorge Riera, Kazuki Iwata, Makoto Takahashi, Toshio Wakabayashi, and Ryuta
Kawashima. ‚The neural basis of the hemodynamic response nonlinearity in human primary visual
cortex: Implications for neurovascular coupling mechanism.‛ NeuroImage, 2006: 616-25.
Woolrich, M W, B D Ripley, J M Brady, and Smith S M. ‚Temporal autocorrelation in univariate
linear modelling of FMRI data.‛ NeuroImage, no. 14 (2001): 1370-86.
Woolrich, MW, et al. ‚Bayesian analysis of neuroimaging data in FSL.‛ NeuroImage, 2009: 173-86.
Woolrich, MW, TEJ Behrens, and SM Smith. ‚Constrained linear basis sets for HRF modelling
using Variational Bayes.‛ NeuroImage, 2004: 1748-61.
Worsley, KJ. ‚Statistical analysis of activation images.‛ Chap. 14 in Functional MRI: An Introduction
to Methods, by P Jezzard, P M Matthews and S M Smith. 2001.
Worsley, KJ, AC Evans, S Marrett, and P Neelin. ‚A three-dimensional statistical analysis for CBF
activation studies in human brain.‛ J Cereb Blood Flow Metab, 1992: 900-18.
Worsley, KJ, AC Evans, S Marrett, and P Neelin. ‚A unified statistical approach for determining
significant signals in images of cerebral activation.‛ Hum. Brain Mapp., 1996: 58-73.
70
ANNEX
FIGURE A.1: Electrode scalp map identifying the channel names with their locations.
F3 FZ F4 F8
FC3 FCZ FC4FT8
C3 CZ C4
CP3 CPZ CP4TP8
A1
P3 PZ P4 T6
A2
Channel locations
FP1 FP2
F7
FT7
T3 T4
TP7
T5
O1 OZ O2