1. Simultaneous EEG-fMRI: from acquisition to application.
Karen Mullinger
Sir Peter Mansfield Magnetic Resonance Centre,
School of Physics and Astronomy
University of Nottingham

2. Overview Introduction Aspects of getting good quality data
Optimising experimental set-up
General pointers
Facilitating good:
gradient artefact correction
pulse artefact correction
Summary
Application
Neurovascular coupling.
Latest results (food for thought)

3. Why Simultaneous EEG –fMRI?
Very powerful spatiotemporal tool
Same experimental environment
Same attention and awareness
Same brain activity
Necessary when brain activity can’t be predicted
fMRI
EEG

4. EEG Artefact Sources
Gradient Artefact (GA): Switching of the gradient fields, causes large changes in magnetic flux inducing electrical signals within the EEG.
Average slice Artefact

5. 1) Pulsatile blood flow effects (Hall effect).
EEG Artefact Sources
Pulse Artefact (PA): Precise source unclear but linked to the cardiac cycle.
The ions in a conducting fluid, such as blood, experience a force when the fluid flows in a magnetic field.
This results in a charge distribution building up at the vessel wall, which generates an electric field that causes current flow in conducting tissue surrounding the vessel.
1) Pulsatile blood flow effects (Hall effect).
2) Small head nod
3) Scalp expansion

6. The Result!
200µV

7. Good quality EEG data Two aspects to EEG-fMRI:
Experimental set-up and data collection
Best post-processing methods

8. Good quality EEG data
Experimental set-up and data collection

9. General advice Low impedances of EEG channels
Less noisy EEG signals
Subject comfort and padding
Minimise movement → reduced artefacts

10. General advice: Motion
Aim:
To investigate effect of motion artefacts on EEG-BOLD correlates
Method:
4 subjects
Standard 32 channel EEG recording.
EEG data were recorded during Dual Echo EPI:
40 slices, 84×84 matrix, 3×3×4 mm3 voxels
TR=3s TE1/TE2 =20/48ms
Episodic memory task: required to move a cursor with a roller-ball to respond.
Jansen, M. et al, NeuroImage 59, (2012)

11. General advice: Motion
Analysis:
EEG
Gradient (AAS) and Pulse (OBS) artefact correction
ICA to remove residual artefacts
Noisy channels removed
Filtered 4-8Hz (Theta band)
fMRI
Motion and physiological correction
Echoes combined
Regressors:
Continuous theta regressor
Head motion (from motion parameters)
Artefacts remaining after correction (from visual inspection)
Jansen, M. et al, NeuroImage 59, (2012)

12. General advice: Motion
Not convolved with HRF
Convolved with HRF
Jansen, M. et al, NeuroImage 59, (2012)

13. General advice: Motion
Task: Foot motion
Not convolved with HRF
Convolved with HRF
CAREFUL how you interpret results!
Jansen, M. et al, NeuroImage 59, (2012)

14. General advice Low impedances of EEG channels
Less noisy EEG signals
Subject comfort and padding
Minimise movement → reduced artefacts
Isolate amplifiers/cables from scanner bed
Minimise vibration of equipment

15. General advice 7T, no scanning Amplifier on the scanner bore
Amplifier suspended.
Mullinger, K.J. et al, MRI 26(7), (2008)

16. General advice Low impedances of EEG channels
Less noisy EEG signals
Subject comfort and padding
Minimise movement → reduced artefacts
Isolate amplifiers/cables from scanner bed
Minimise vibration of equipment
Turn cyrocooler compression pumps off
Minimise noise sources

17. General advice 7T, no scanning Everything on Cryopumps off..
...and room lights, gradient and patient airflow
Mullinger, K.J. et al, MRI 26(7), (2008)

18. Gradient artefact

19. Average Artefact Subtraction (AAS)
Allen, P.J. et al. NeuroImage 12, (2000)

20. Artefact Correction requirements
AAS Requires:
Artefact to be highly repeatable across cycles
Precisely recording the artefact waveform and the beginning of each volume.
These requirements must be closely adhered to as the unfiltered GA is at least 10,000 times larger than an evoked response
Residual artefacts are problematic

21. Precise sampling Acquire EEG data at 5kHz
Ensure your slice TR is a multiple of the scanner clock period (i.e. 200μs)
WARNING:
TR entered into console is not always the TR outputted due to rounding issues!!
Philips System for equidistant EPI: TR Calculator*
*Need clinical science agreement for this

22. Precise sampling
Synchronise the MR Scanner and EEG clocks using the output from the MR scanner.
Philips system: use the 10MHz output from the MR scanner clock to drive the EEG clock
Mandelkow, H. et al, NeuroImage 32(3) (2006)
Mullinger, K.J. et al, JMRI 27(3): p (2008)

23. Standard Deviation associated with average slice artifact
Experimental Results
TR = 2s, synchronised
TR = 2s, not synchronised
TR = s, synchronised
Standard Deviation associated with average slice artifact
Average slice artifact
180 dynamics, 20 slices, 3 subjects
Results from electrode F7 for a single subject
Mullinger, K.J. et al, JMRI 27(3): (2008)

24. Minimising GA amplitude
Why?
Prevent channel saturation
Allow higher EEG recording bandwidth
Improve artefact correction
How?
Position subjects 4cm in foot direction (naision at isocentre = 0cm). Approximately at Fp1&2.
Yan, W.X., et al. NeuroImage 46(2): (2009)
Mullinger, K.J. et al, NeuroImage, 54(3): (2011)

25. Optimal Position: standard fMRI
Aim:
Compare GA produced by a multi-slice EPI sequence at standard and optimal subject positions.
Method:
6 subjects
Experiments were carried out with the nasion at:
iso-centre
optimal (+4 cm) z-offset
Standard 32 channel EEG recording, 250 Hz low pass filter.
EEG data were recorded during standard EPI:
32 slices, 84×84 matrix, 3×3×4 mm3 voxels
TR=2.5s TE =40ms; slice repetition frequency = 12.8 Hz
Cued foot movement: 5s every 30s (total: 8 minutes): cumulative head movements of <1 mm.

26. Optimal position: Results
RMS of average artefact before correction
40% average reduction in RMS over all channels
Optimal position
Isocentre
STD across slices after correction
36% reduction in RMS at slice harmonics after correction

27. Pulse artefact

28. Pulse Artefact Correction
Many methods of PA correction
Average artefact subtraction (AAS)1
Optimal basis sets (OBS)2
Independent component analysis (ICA)3
Varying levels of success reported
Most require correctly identifying the QRS complex
within the ECG trace.
ECG
[1] Allen, P.J. et al, NeuroImage 8(3), (1998)
[2] Srivastava, G. et al, NeuroImage 24, (2005)
[3] Niazy, R.K. et al, NeuroImage 28, (2005)

29. Pulse Artifact Problems: ECG is affected by gradients as well.
Sometimes hard to get a good ECG trace.
Trace is sometimes saturated.

30. Solution on a Philips system*:
Use vector cardiogram (VCG) from MR Scanner which is unaffected by gradients1.
R peak markers are also placed automatically in the physlog file2 which can be used for pulse artefact correction directly.
[1] Chia et al. JMRI, 12: (2000)
[2] Fischer et al. MRM, 42: (1999)
*Need research login to access physlog file

31. Results Data gradient-corrected and low-pass filtered at 70 Hz
EEG trace from Tp10 averaged over all cardiac cycles in 2 minute period.
0 time=R peak marker from VCG
No correction
Using ECG markers
Using VCG markers
Mean Standard Deviation

32. Pulse Artefact Precise source unclear but linked to the cardiac cycle.
Variation between cardiac cycles makes correction of difficult
Problems increase with field strength
Need a greater understanding of pulse artefact
Average pulse artefact
1) Pulsatile blood flow effects
R-peak T7
2) Small head nod
3) Scalp expansion
Debener, S. et al, Int. J. Psychophys, 2008, 67(3), p

33. Measuring the PA constituents
6 subjects
Recorded EEG data in 3T MR scanner
4 conditions:
Relaxed
Bite Bar and vacuum cushion (stop head nod)
Swimming cap (stop Hall effect)
2&3 (left with scalp expansion).
Yan, W.X., et al., HBM, (4): p
Mullinger, K.J. et al, #667 WTh HBM Quebec.

34. PA Experimental Results
Relax B C D
Restrained
Insulated
Restrained & insulated
Amplitude (µV) A
EEG
Average RMS
Subject RMS

35. Summary
SNR of EEG data inside the MR scanner still lower than outside.
Higher MR fields → increasing EEG artefact problems.
Experimental set-up is important.

36. Data Acquisition Summary
To improve gradient artefact correction:
Chose TR and number of slices wisely
Synchronise scanner clocks
Optimally position the subject
To improve pulse artefact correction:
Use VCG to monitor cardiac trace

37. Application

38. Investigating origin of Negative BOLD
Negative BOLD Response (NBR): Regions where there is a stimulus related decrease in BOLD signal.
Reported in visual1, motor2 and somatosensory3 cortices.
From: [2] Stefanovic et al. Neuroimage 22;2004.
[1] Shmuel et al. Neuron 36(6);2002. [3] Kastrup et al. Neuroimage 41(4);2008.

39. Negative BOLD NBR origin unclear:
Neuronal basis
Haemodynamic artefact (blood steal)
Invasive recordings in monkeys show a decrease in local field potentials (LFP) and spiking activity in regions of NBR, and suggest at least 60% of NBR is neuronal in origin1.
Clarification in humans is needed.
NBR-decrease in BOLD signal amplitude with stimulation.
Neuronal basis: reduction in CBF due to neuronal activity with a lesser reduction in CMRO2, increased OEF-decreased BOLD
Invasive recordings in animals in the visual cortex.
[1] Shmuel et al. Nat Neurosci. 9(4);2006.

40. Aim
To use simultaneous measurements of BOLD, ASL and EEG to investigate the relationship between natural fluctuations in the NBR and somatosensory evoked potentials (SEPs) during median nerve stimulation (MNS)1
[1] Mullinger et al Proc. ISMRM #109; 2011

41. Method Simultaneous EEG-fMRI:
Philips Achieva 3T MR scanner; 8 channel SENSE head coil.
64 channel Brain Products EEG system.
Localiser: GE-EPI BOLD sequence used for planning.
Experiment:
FAIR Double Acquisition Background Suppression1 sequence used for simultaneous BOLD and background suppressed ASL data acquisition (TR=2.6s, TE=13/33ms (ASL/BOLD), label delay=1400ms, 3x3x5mm3 voxels, 212mm FOV, SENSE factor 2; background suppression TI1/TI2=340ms/560ms).
Cardiac and respiration monitored.
MR and EEG scanner clocks synchronised.
EEG electrode positions digitised (Polhemus system, Isotrack).
Localiser used to place 10 slices over sensorimoter cortex.
FAIR-flow sensitive alternating inversion recovery
[1] Wesolowski et al. Proc. ISMRM, #6132;2009.

42. Paradigm 13 right handed subjects (8 males, 26±3 yrs)
Stimulate median nerve of right wrist
Amplitude: just above motor threshold to cause thumb distension
2 Hz stimulation, 0.5ms pulses (Digitimer DS7A)
10s
20s
40 blocks
20 pulses per block
2Hz chosen to allow entire evoked response to be recorded.
Jitter-to prevent stimulation occurring in same position in TR period in successive blocks

43. Analysis EEG pre-processing
Gradient and pulse artefact correction using average artefact subtraction (Brain Vision Analyzer2)
Data inspection:
3 subjects excluded due to gross (>3mm) or stimulus-locked movement.
Noisy channels and/or blocks rejected
Down-sampled: 600Hz
Re-referenced: Average of non-noisy channels
Filtered: 2-40 Hz
2Hz high pass used to avoid any artefact from the stimulation device.

44. Analysis EEG Beamformer1
Fitted2 basis set to SEP for each block to find peak-to-peak P100-N140 amplitude
T-stat map:
active window: s passive window: s
VE timecourse for single block
Auto linear regression method used:
Black is the SEP averaged over a block
Blue is the fit to the P100 peak
Green is the fit to the N140 peak
Averaged over 20 responses in a block
[1] Brookes et al. NeuroImage 40(3);2008
[2] Mayhew et al. Clin. Neurophysiol. 117(6);2006

45. Analysis fMRI pre-processing Motion corrected (FLIRT, FSL)
BOLD data physiologically corrected (RETROICOR)
Interpolated to effective TR=2.6s
ASL: perfusion weighted image: Tag-Control
BOLD image pairs averaged
Normalised to MNI template
Smoothed: 5mm FWHM kernel

46. Analysis  fMRI General Linear Models Boxcar: SEP amplitude modulator:
2nd level fixed effects analysis on BOLD and ASL data
Areas of significant positive/negative correlation (P<0.05, FWE corrected) between BOLD and the simple boxcar model were found in contralateral/ipsilateral sensorimotor cortex (S1/M1).
Areas showing significant positive/negative correlations between CBF and the boxcar model (P<0.001, uncorrected) were identified from the ASL data. These data were masked by the BOLD boxcar model (P<0.001, uncorrected) data, and mean timecourses extracted from these CBF ROIs.
Areas showing significant modulation with SEP amplitude in BOLD (P<0.05, FWE) and ASL (P<0.001, uncorrected) data were also identified, masked by BOLD boxcar model data (P<0.001, uncorrected), and timecourses extracted.
Group ROI defined for positive and negative correlation.
BOLD: P<0.05 FWE
ASL: P<0.001 uncorr
Timecourse for each region & subject obtained; averaged over subjects & blocks

47. Results BOLD ASL MNI peak co-ordinates (-42,-20,50) Positive
Positively correlated with Boxcar
Negatively correlated with Boxcar
Negatively correlated with SEP amplitude
MNI peak co-ordinates
(-42,-20,50) Positive
(36,-18,50) SEP
(34,-16,46) Negative
Negative and Positive BOLD and ASL clearly within primary sensorimotor cortex
The negative correlation with the SEP model in the BOLD data has slightly better agreement with the peak location of the positive BOLD signal.
No positive correlation amplitude of SEP and fMRI in S1.

48. Results
Solid line = BOLD, Dashed line= ASL

49. Results Constants: M = 7.2%1, α = 0.38, β = 1.2
Isocontours of CMRO2 (Davis Model2) 2
R = , P<0.1*10-4 Gradient = 0.42
Coupling ratio agrees with Stefanovic3
Error bars = STD over trials for CBF, CMRO2 were calculated from the error in CBF and error in BOLD, other constants were assumed to have no error (although I know this isn’t true it would have been constant for all of the data points).
Coupling ratio shown by gradient of the line is in agreement with previous findings for a motor task
NB Although value of M was taken at 1.5T it seems to represent a value somewhere in the middle of all the literature values
[1] Kastrup et al., Neuroimage 41(4);2008.
[2] Davis et al., PNAS, 95;1998
[3] Stefanovic et al., NeuroImage 22;2004
[1] Kastrup et al., Neuroimage 41(4);2008.
[2] Davis et al., PNAS, 95;1998

50. Discussion
No positive correlation of fMRI and evoked potentials in S11.
Ipsilateral NBR cannot be explained by blood steal2 as bilateral S1 regions are fed from different vascular territories.
CMRO2 shown in NBR region - suggests a neuronal origin of the response.
unlike Klingner et al1 who found correlation of stimulus intensity at PBR
Supra-threshold stimulus compared with sub-threshold stimuli.
Supra-threshold stimulus providing robust response well explained by boxcar, compared with sub-threshold stimuli.
[1] Klingner et al. Neuroimage 53(1); 2010
[2] Wade et al. Neuron 36(6);2002

51. Discussion
Show for first time correlation between ipsilateral S1 NBR and amplitude of concurrent EEG evoked response from contralateral S1/M1.
Agrees with area identified by Klingner where NBR is modulated by intensity of MNS1.
Suggest that NBR-SEP relationship arises because NBR results from inhibition of task irrelevant processing in ipsilateral S1, with corresponding increase in excitability of contralateral S1, as indexed by increasing SEP amplitude.
[1] Klingner et al. Neuroimage 53(1); 2010

52. Why simultaneous recordings....
Trial by trial natural fluctuations in the evoked response → simultaneous recordings are essential.
Can also study changes in oscillatory activity and correlations with BOLD1 and also CBF
Positively correlated with Boxcar
Negatively correlated with Boxcar
Negatively correlated with Mu amplitude
p<0.05, FWE
[1] Mayhew et al, Proc ISMRM #1560, 2011

53. Why Simultaneous recordings....food for thought
Differences in oscillatory activity: providing evidence of a neuronal origin of the post-stimulus undershoot...

54. Acknowledgments MRC Professor Richard Bowtell Dr Susan Francis
Colleagues
Professor Richard Bowtell
Dr Susan Francis
Winston Yan
Jade Havenhand
Dr Thomas White
Dr Marjie Jansen
Dr Elizabeth Liddle
Prof Peter Liddle
Birmingham University
Dr Stephen Mayhew
Dr Andrew Bagshaw
Industry
Robert Stormer (Brain Products)
Dr Matthew Clemence (Philips)
Funding
MRC
EPSRC
Mansfield Fellowships

55. Thank you