1. Automated Subject-Specific Peak Identification
and Ballistocardiographic Artifact Correction
in EEG-fMRI
Sara Assecondi
Dep. Electronics and Information Systems
MEDISIP, Ghent University
Promotors: Prof. Dr. Ir. S. Staelens, Prof. Dr. P. Boon
2. The secret dream of a neuroscientist
is to map the human brain: WHERE and WHEN
personality
sensory
emotions
problem solving
reasoning
hearing
language
speech
vision
bla…. bla….
bla….
Introduction: general background 2
3. Brain mapping has useful applications
Pathologies
• Dyslexia • Epilepsy
(5% school aged children) (0.5-1 % of the population)
EL
DYS XIA
Cognition
• mind, reasoning, …
• perception, …
• intelligence, learning…
Introduction: general background 3
4. Nowadays we have reliable means
to localize brain functions
EEG exam fMRI exam
Take an image of the brain-
fMRI
Introduction: general background 4
5. What an ElectroEncephaloGram (EEG) is
100 billion
(100000000000)
Neurons communicate by means of electrical impulses: the brain is
characterized by an electrical activity that can be measured in μV
Signal amplitude
(μV)
Time
(ms)
Introduction: general background 5
6. EEG exam
EEG electrode
Signal amplitude
(μV)
Time (ms)
Introduction: general background 6
7. EEG exam:
Event Related Potentials (ERP)
Signal amplitude
μV ERP
(μV)
STIMULUS
- electric,
- auditory
- visual Time ms
(ms)
- cognitive...
Introduction: general background 7
8. Information in EEG and ERP
The latency of peaks , i.e. the time interval between the
stimulus and the occurrence of a peak, gives temporal
information about the sequence of events in the brain
μV
perception decoding comprehension
μV
.
.
.
ms stimulus ms
stimulus latency
Each peak is related to a specific stage Scalp
of the information process, through its map
latency
Introduction: general background 8
9. Basics of Magnetic Resonance Imaging (MRI)
Up to 70 % of body weight is made up by water
70%
O N
H
H H
1H nucleus
≈ S
Water has a spin
molecule and a magnetic moment
By making use of the magnetic
properties in brain tissues, an image
of the brain can be reconstructed
Introduction: general background 9
10. MRI exam
MR-scanner
strong
magnet Sagittal view
Radio-frequency
waves and
time-varying
magnetic fields
Axial view
Introduction: general background 10
11. MRI exam: MR-scanner
functional MRI
Activity:
on on on
off off off off
time
Activity:
• Endogenous (epileptic spikes)
• Exogenous (external stimuli)
Introduction: general background 11
12. Do we have everything we need?
personality
We have the means:
sensory
emotions • EEG μV
problem solving
reasoning hearing
language ms
speech vision
• fMRI
Two new important concepts:
• Temporal resolution
• Spatial resolution
Introduction: general background 12
13. Temporal resolution
The smaller the interval the higher the temporal resolution
μV μV μV
ms ms ms
EEG-ERP (ms) fMRI (s)
μV
ms
Introduction: general background 13
14. Spatial resolution
The finer the grid the higher the spatial resolution
EEG-ERP (cm) fMRI (mm)
Scalp map
Introduction: general background 14
15. EEG-fMRI achieves high temporal and spatial resolution
Invasive
0
Brain
PET
-1 EEG+ERP
Map -2 fMRI
NIRS
Log size (m)
-3
Column
Layer -4
Adapted from Cohen and Bookheimer, 1994
Neuron -5 Non-invasive
-3 -2 -1 0 1 2 3 4 5
Log time (s)
Take the advantage of both EEG and fMRI
to have, at the same time,
high temporal and high spatial resolution
Introduction: general background 15
16. What do we need for EEG-fMRI?
• MR-compatible EEG equipment
Amplifier
syncbox Power pack
Galvanic isolation
• MR scanner
Introduction: general background 16
17. EEG-fMRI exam
Scanner room Control room
EEG electrode
EEG wires
PC
amplifier
Activity:
on on on
off off off off
Adapted from Bénar, PhD thesis
Introduction: general background 17
19. How we approach the problem of
cerebral localization of brain functions
Integration of
EEG and fMRI
Time domain Time and spatial domain
Improving the temporal Data quality of simultaneous
localization of brain functions EEG-fMRI recordings
Quantitative and automated
identification of ERP peaks Data quality of simultaneous
ERP-fMRI recordings
Data interpretation
Effect of different recording
environments on the process
under investigation
Overview 19
20. How we approach the problem of
cerebral localization of brain functions
Integration of
EEG and fMRI
Time domain Time and spatial domain
Improving the temporal Data quality of simultaneous
localization of brain functions EEG-fMRI recordings
Quantitative and automated
identification of ERP peaks Data quality of simultaneous
ERP-fMRI recordings
Data interpretation
Effect of different recording
environments on the process
under investigation
Overview 20
21. The quantification of amplitudes and latencies
in ERPs may be challenging
PRE-LEXICAL LEXICAL POST-LEXICAL
N3 N4
N0 N1 N2
P0 P1 P4 P600
P2b
P2a
0ms 160ms 420ms 800ms
• Challenges • Information to retrieve
– Number of peaks – Latency of each peak
– Labelling – Amplitude of each peak
– Inter-subject variability – Label of each peak
Time domain: Automatic identification of ERP peaks 21
22. Features of peak quantification approaches
Reliability
Reproducibility, over subjects and over time
Objectivity
Gold standard: expert clinician Automated methods: Peak-picking,
• Time consuming Dynamic Time Warping
• subjective • Fast
• Objective
N1 N1
μV
μV
P2
P2
stimulus ms stimulus ms
Time domain: Automatic identification of ERP peaks 22
23. Method 1: Peak-picking is a search for extrema
in predefined time intervals
N3 N4
N0 N1 N2
P0 P1 P4 P600
P2b
P2a
Predifined
time interval
Example: Looking for the N3 peak Drawbacks
• Define the time interval
• Highly dependent on the
• N3 is a negative peak: defined intervals
search for a minimum
• Does not take into account
• Assign the N3 label the inter-subject variability
to the minimum found
Time domain: Automatic identification of ERP peaks 23
24. Method 2: Dynamic time warping (DTW) is a
nonlinear mapping of two signals
The ERP is compared with a reference signal containing the
peaks of interest
Linear alignment Dynamic Time Warping
reference reference
subject subject
Time (ms) Time (ms)
Drawback A peak is always found: it may not be a peak
Time domain: Automatic identification of ERP peaks 24
25. We proposed a method (ppDTW) that integrates
peack-picking and Dynamic Time Warping
Compared by means of DTW
REFERENCE SUBJECT
Sequence of
candidate peaks Peak-picking in a
time interval
centered at the
candidate peak
Measured latencies
and amplitudes
Time domain: Automatic identification of ERP peaks 25
26. The reference signal must contain
all the peaks of interest
We computed the reference by interpolating mean
amplitudes and latencies derived from a normal population
Peak variability derived
Reference
from a normal population
Time domain: Automatic identification of ERP peaks 26
27. Case study
(Assecondi et al., Clinical Neurophysiology 2009)
• Normal children and children diagnosed with
developmental dyslexia
• Reading related ERP
• Latencies automatically determined:
– Peak picking
– ppDTW
• Validation
– Comparison with the visual scoring of an expert clinician
Time domain: Automatic identification of ERP peaks 27
28. Performance of the methods
on normal and dyslexic subjects
normal dyslexic
N0 N2 N3 N4
N1N2 N3 N4
P1 P2b P4
P2b P4 P600
ppdtw
ppdtw
visual scoring
visual scoring
peak picking
peak picking
Time domain: Automatic identification of ERP peaks 28
29. We evaluate the methods by comparing
the automated scoring with the visual scoring
C = correct identifications
S = substitutions C S D I
D = deletions expert
I = insertions method
N1
precision recall F-score
Latency N1
1
C C 1 1
C S I C S D 2P 2R
Peak-picking 93% 80% 85%
PP-DTW 93% 86% 89%
Time domain: Automatic identification of ERP peaks 29
30. Conclusion
We developed an automated peak detection method that
takes into account the inter-subject variability and a-priori
knowledge about the ERP (in the reference)
A-priori knowledge may be derived from literature,
hypothesis about the ERP experiment, expertise of the
clinician
The method is valuable with different ERPs, when huge
databases have to be measured or when the same subject
must be examined during a therapy or rehabilitation
Time domain: Automatic identification of ERP peaks 30
31. How we approach the problem of
cerebral localization of brain functions
Integration of
EEG and fMRI
Time domain Time and spatial domain
Improving the temporal Data quality of simultaneous
localization of brain functions EEG-fMRI recordings
Quantitative and automated
identification of ERP peaks Data quality of simultaneous
ERP-fMRI recordings
Data interpretation
Effect of different recording
environments on the process
under investigation
Overview 31
32. EEG-fMRI achieves high temporal and spatial resolution
Invasive
0
Brain
PET
-1 EEG+ERP
Map -2 fMRI
NIRS
Log size (m)
-3
Column
Layer -4
Adapted from Cohen and Bookheimer, 1994
Neuron -5 Non-invasive
-3 -2 -1 0 1 2 3 4 5
Log time (s)
EEG - fMRI implies the interaction of
different physical and physiological systems ARTIFACTS
(human being, EEG system, MR scanner)
Time and spatial domain: Data quality of simultaneous EEG-fMRI 32
33. An artifact is a contamination
of the data of interest
Suppose you want to listen to the radio while a bell is ringing
din
don
Source of noise o Data of interest
l i
d
e ? i
a
i s
s
n i
n
d
s
i i
Time and spatial domain: Data quality of simultaneous EEG-fMRI 33
34. Two main artifacts affect EEG-fMRI recordings
• Image acquisition artifact
caused by the RF-waves
and time-varying
gradients
~0.5 s
Very deterministic
BCG-peak
• Ballistocardiographic
artifact is a blood related
effect due to the high
static magnetic field R-peak
Only QUASI-deterministic
~1 s
Time and spatial domain: Data quality of simultaneous EEG-fMRI 34
35. The BallistoCardioGraphic artifact (BCGa)
R R R R R R R R
100 µV
10 seconds
• It depends on the proximity of the EEG electrodes to blood vessels
• It is synchronous to the heart beat and spread throughout the
heart beat and all over the scalp
Time and spatial domain: Data quality of simultaneous EEG-fMRI 35
36. Distribution of scalp arteries
Adapted from Gray, 1918
Time and spatial domain: Data quality of simultaneous EEG-fMRI 36
37. The BCG artifact can be removed
if an additive model is assumed
Noise and
Recorded other non-MR
signal = EEG + BCGa + related
artifacts
Estimate of the BCGa
that depends on the
algorithm used
subtraction BCGa
model
Time and spatial domain: Data quality of simultaneous EEG-fMRI 37
38. Intra-subject variability of the BCGa
Heart beat
-0.2 0 0.2 0.4 0.6
seconds
Time and spatial domain: Data quality of simultaneous EEG-fMRI 38
39. Inter-subject variability of the BCGa
S1 S2 S3 S4 S5 S6
channels
seconds
Time and spatial domain: Data quality of simultaneous EEG-fMRI 39
40. The BCG artifact removal algorithms must take the
inter- and intra-subject variability into account
Group 1: methods based on Average Artifact Subtraction
Moving Average
Weighted Average
Selective Averaging based on clustering
Group 2: methods based on Blind Source Separation
Optimal Basis Set (OBS)
Independent Component Analysis (ICA)
Canonical Correlation Analysis (CCA)
Time and spatial domain: Data quality of simultaneous EEG-fMRI 40
41. Blind Source Separation identifies the sources
generating the recorded signal
s1
Blind Cerebral
Source s2 sources
Separation s3
(BSS) Non-cerebral
x2 x3 X
s4 sources
x1
x4
constraints
s1
s2 s3 s4
BSS as an artifact removal technique is twofold:
• Step 1: Identification of sources
• Step 2: Selection of sources
(brain signal or artifact)
Time and spatial domain: Data quality of simultaneous EEG-fMRI 41
42. We proposed a BSS method that uses Canonical
Correlation Analysis to identify the sources
Step 1: Identification of sources Step 2: Selection of artifact sources
EEG epoch 1 EEG epoch 2 Criteria have to be met simultaneously:
- correlation of sources common to two
consecutive EEG epochs
- Sources must not be sinusoidal EEG
rhythms
Sources or
Canonical BSS-CCA
variates
time
.
. .
.
. .
Time and spatial domain: Data quality of simultaneous EEG-fMRI 42
43. Case study
(Assecondi et al., Physics in Medicine and Biology 2009)
• Patients affected by epilepsy
• Recording of 20 minutes of simultaneous EEG-fMRI
• Artifact removal
– AAS
– CCA
• Validation
– Frequency content at the harmonics of the ECG
– Global Field Power
– Signal Envelope
Time and spatial domain: Data quality of simultaneous EEG-fMRI 43
44. Comparison of cleaned and raw EEG
BSS-CCA
AAS
RAW
Time and spatial domain: Data quality of simultaneous EEG-fMRI 44
45. Comparison of cleaned and raw EEG
BSS-CCA
AAS
RAW
Time and spatial domain: Data quality of simultaneous EEG-fMRI 45
46. Conclusion
The BCG artifact has an intrinsic inter- and intra-subject
variability that make the performance of artifact removal
algorithms subject-dependent
We proposed a method to remove the BCG artifact that
deals with the intra- and inter-subject variability
The proposed method is an added value especially in those
cases where the AAS fails, because of the excessive intra-
subject variability of the artifact
Time and spatial domain: Data quality of simultaneous EEG-fMRI 46
47. How we approach the problem of
cerebral localization of brain functions
Integration of
EEG and fMRI
Time domain Time and spatial domain
Improving the temporal Data quality of simultaneous
localization of brain functions EEG-fMRI recordings
Quantitative and automated
identification of ERP peaks Data quality of simultaneous
ERP-fMRI recordings
Data interpretation
Effect of different recording
environments on the process
under investigation
Overview 47
48. The methodologies of analysis of ERPs and EEG
are intrinsically different
s s s s s
• Smaller amplitude
• Involve additional averaging
• Different signal-to-noise ratio
The method to remove the BCG artifact needs
modifications to select the non-cerebral sources
Time and spatial domain: Data quality of simultaneous ERP-fMRI 48
49. The proposed method is adapted to ERPs
Step 1: Identification of sources Step 2: Selection of artifact sources
Average BCG EEG epoch Criteria have to be met simultaneously:
- correlation of sources extracted from
the average BCG and the EEG epoch
- Sources that maximally contribute at the
same time to the artifact and to the EEG
Sources or
Canonical BSS-CCA Sources important for the EEG but not
variates
for the artifact are not removed
.
. .
.
. .
Time and spatial domain: Data quality of simultaneous ERP-fMRI 49
50. Case study
(Assecondi et al., submitted to Clinical Neurophysiology)
• Healthy volunteers
• Three tasks:
– Visual: Visual detection task (5 subjects)
– Cognitive: Go-nogo task (5 subjects)
– Motor: Motor task (7 subjects)
• In two situations:
– Outside the MR-scanner room (0T)
– Inside the MR-scanner, without scanning (3T)
• BCG artifact removal
– AAS (Average Artifact Subtraction)
– CCA (Canonical Correlation Analysis)
Time and spatial domain: Data quality of simultaneous ERP-fMRI 50
51. Three different types of stimuli were used
Detection GoNogo Motor
Left visual field No Go Left keypress
Central visual field Go Withhold
Right visual field Right keypress
Time and spatial domain: Data quality of simultaneous ERP-fMRI 51
52. In most cases both methods are able to recover
ERP time series
Detection task
0T
3T
Time and spatial domain: Data quality of simultaneous ERP-fMRI 52
53. BSS-CCA seems more effective to recover small
amplitude low frequency components
Motor task
AAS BSS-CCA
0T
3T
Time and spatial domain: Data quality of simultaneous ERP-fMRI 53
54. How we approach the problem of
cerebral localization of brain functions
Integration of
EEG and fMRI
Time domain Time and spatial domain
Improving the temporal Data quality of simultaneous
localization of brain functions EEG-fMRI recordings
Quantitative and automated
identification of ERP peaks Data quality of simultaneous
ERP-fMRI recordings
Data interpretation
Effect of different recording
environments on the process
under investigation
Overview 54
55. Do the very different environments in which the data are
recorded have an effect on the physiological process under
investigation?
ERP lab MR scanner
position sitting laying
Light dimmed dimmed
Noise no yes
Screen orientation In front mirrors
Magnetic fields no yes
Time and spatial domain: Effect of different environments 55
56. To disentangle effects of factors affecting EEG-fMRI
data, different situations must be compared
Effect of the static
field +position
Static
0T field
No MR-related artifacts BCG artifact
Time and spatial domain: Effect of different environments 56
57. Case study
(Assecondi et al., submitted to Clinical Neurophysiology)
• Healthy volunteers
• Three tasks:
– Visual: Visual detection task (5 subjects)
– Cognitive: Go-nogo task (5 subjects)
– Motor: Motor task (7 subjects)
• In two situations:
– Outside the MR-scanner room (0T)
– Inside the MR-scanner, without scanning (3T)
• BCG artifact removal
– AAS (Average Artifact Subtraction)
– CCA (Canonical Correlation Analysis)
Time and spatial domain: Effect of different environments 57
58. We found differences between
0T, 3T and cleaned data
• Amplitude decrease in 3T
• Latency increase in 3T
• Strong contamination of the data due to the artifact
• Difference in the scalp distribution
0T 3T After BCG removal
goP3
GoNogo task
Ipsi N1
Detection task
Time and spatial domain: Effect of different environments 58
59. Reaction time: rt@3T-rt@0T
***
***
***
50 ***
***
*** *** ** * * **
*
0 - -
ms - subject
-
**
***
-50
- p>0.05
* p<0.05
*** *** ** p<0.01
*** p<0.001
-100
Leuven
-150 ***
Detection Go-Nogo Motor
Time and spatial domain: Effect of different environments 59
60. Conclusion
The quality of average ERP is less sensitive to the method
used to remove the BCG. However, BSS-CCA seems to be
more effective with low frequency low amplitude
components and when less trials are available
Differences are found between different environment.
More subjects are needed to obtain robust results
Different situations must also be taken into account
(dummy scanner, actual fMRI acquisition)
Time and spatial domain: Effect of different environments 60
61. How we approach the problem of
cerebral localization of brain functions
Integration of
EEG and fMRI
Time domain Time and spatial domain
Improving the temporal Data quality of simultaneous
localization of brain functions EEG-fMRI recordings
Quantitative and automated
identification of ERP peaks Data quality of simultaneous
ERP-fMRI recordings
Data interpretation
Effect of different recording
environments on the process
under investigation
Overview 61
62. Overall considerations and future prospects
• The localization of brain functions can be improved by taking into
account a-priori knowledge of the process and the inter- and intra-
subject variability
• New multimodal approaches, e.g. EEG-fMRI, offer a deeper insight
into how the brain works (great help by the availability of
commercial MR-compatible EEG systems)
Fundamental research Development of methods
• to understand the relation between • to improve the integration of
EEG and fMRI the two modalities and to
• to comprehend the effect of the actually benefit by the different
very different environments on EEG information extracted by
data different modalities
Overall considerations and future prospects 62
63. Automated Subject-Specific Peak Identification
and Ballistocardiographic Artifact Correction
in EEG-fMRI
Sara Assecondi
Dep. Electronics and Information Systems
MEDISIP, Ghent University
Promotors: Prof. Dr. Ir. S. Staelens, Prof. Dr. P. Boon
Thank you!
64. Related publications
• S. Assecondi, et al. Effect of the static magnetic field of the MR-
scanner on ERPs: evaluation of visual, cognitive and motor
potentials Submitted.
• S. Assecondi, et al. Automated identification of ERP peaks through
Dynamic Time Warping: an application to developmental dyslexia
Clinical Neurophysiology, DOI: 10.1016/j.clinph.2009.06.023.
• S. Assecondi, et al. Removal of the ballistocardiographic artifact
from EEG-fMRI data: a canonical correlation approach Physics in
Medicine and Biology. Vol. 54 (2). 2009. pp. 1673-1689
64