Participants
The experiment was performed in accordance with the Declaration of Helsinki and was approved by the local ethics committee. Seven participants (all male, students, age 21–26 years) volunteered in this experiment. One was excluded, because he felt uncomfortable inside the scanner and aborted the experiment. Participants had medical histories free of neurological abnormalities and gave written informed consent for participation before the experiment. They received a monetary compensation of 20 €.
Experiment Description
The aim of the present work was to record EEG, specifically alpha rhythm amplitude differences and evoked brain responses and to compare those from measurements inside and outside the MRI scanner. Hence, each participant performed the experiment twice. First, recordings were performed outside the MRI scanner, in the room where the EEG cap was prepared and then a second time inside the MRI scanner. We used a modified version of the experiment in our last RLAF work (Steyrl et al. 2017). The experiment itself was divided into two parts. During the first part, evoked brain responses were recorded. Participants had their eyes opened and were looking at a computer monitor, where a checkerboard was presented. The checkerboard had 8 × 8 black and white square fields with a small red dot in the center. The black and white fields were inverted every 0.5–0.6 s to trigger visual evoked potentials (VEP). 600 VEPs were collected per experiment. In the second part of the experiment, participants closed their eyes and were instructed to relax, but not to fall asleep, to provoke changes in the alpha rhythm. The experiments outside and inside the scanner differed in three points: (1) Outside the scanner, participants were upright sitting in a chair. Inside the scanner, participants were lying in supine position. (2) The distance between monitor and eyes was about 1 m in the experiments outside the scanner (visual angle 20°), and approximately 2.5 m in the experiments inside the scanner (visible angle 15°). (3) Outside the scanner, the environment was quiet. Inside the scanner, we used earplugs to reduce the scanner noise. One experiment lasted in total about 12 min with approximately 6 min opened eyes and 6 min closed eyes. The overall time per participant was about 2 h with 20 min for instructions and information, 40 min cap preparation and testing, 12 min experiment outside, 20 min preparation inside scanner, 10 min testing inside scanner, 12 min experiment inside scanner, and 5 min for removing the equipment from the participant.
Reference Layer Cap Prototype
In this work, we used the second version of a reference layer cap prototype, developed by GUGER TECHNOLOGIES OG, Austria (patents pending). This prototype cap offers the opportunity of dedicated reference recordings from a separate layer. The new cap version has Ag/AgCl sinter-pellets as electrode contact areas instead of pure Ag. For a description and an evaluation of the first version please see (Steyrl et al. 2015, 2017). A rendering of the cap is depicted in Fig. 1a and see Fig. 1b for a photo of the new cap version. The cap size is optimized for a head circumference of about 58 cm. However, the cap is flexible enough for head circumferences between 56 and 58 cm. To use this cap with larger heads is not recommended, because in that case the cap can cause pain due to high contact pressure. The cap is equipped with 29 double-layer EEG electrode pairs, a common ground/reference electrode, and connectors for two additional self-adhesive MRI compatible electrocardiogram (ECG) electrodes at the participants back. Each double-layer EEG electrode is made of a pair of Ag/AgCl sinter-pellets with a diameter of approximately 2 mm and a thickness of approximately 1 mm. The pellets are glued with conductive epoxy to both sides of an approximately 1 mm thick printed circuit board (PCB). One pellet connects to the scalp via conductive EEG gel and the other to the reference layer. The PCB with sinter-pellets is fixed into an isolating plastic housing. The whole electrode is about 8 mm thick and has a diameter of approximately 14 mm. For a schematic of a double layer electrode see Fig. 1c. The reference layer itself is a grid of silicon tubes filled with physiological saline solution and is electrically isolated from the scalp, except at the common ground/reference electrode. At this electrode, the scalp is connected to the reference layer to pull them at the same potential. Electrodes are connected to the EEG amplifier via thin copper cables. 5kOhm current limiting resistors were placed between the sinter-pellets and the cables, and additional 5kOhm resistors were placed at the end of the cables before a coupling board connects to the EEG amplifiers via a flat ribbon cable. ECG connectors are equipped with 10kOhm current limiting resistors at the electrodes. The cable length is approximately 50 cm. The electrode arrangement is according to the international 10/20 system and depicted in Fig. 1d. We put foam pads between the occipital EEG electrodes to prevent pain from lying on a few small electrodes, see Fig. 1e. Temperature measurements were carried out during SAR intensive sequences to rule out a harmful heating of the electrodes.
fMRI Scanner and EEG Recording System
Functional MRI data were acquired at a Siemens Skyra 3.0T (Siemens, Erlangen, Germany) at the MRI-Lab Graz (Austria) using a 20 channel head coil. The helium pump was active and the ventilation was set to the lowest level possible. A standard EPI sequence was implemented (TR = 2250 ms, TE = 28 ms, base resolution = 64, 3.5 × 3.5 × 3.5mm3 voxel size, 0.4 mm gap, 36 slices, field of view = 224 × 224). EEG and ECG was recorded with a 64 channel MRI compatible EEG system (BrainAmp MR plus, Brain Products GmbH, Gilching, Germany). The EEG system was positioned inside the borehole at the head end of the MRI scanner on a wooden panel. Cables and amplifiers were fixed with sand bags. All amplifier settings were chosen according to the manufacturer’s recommendations. Hence, the sampling rate was set to 5 kHz, the cut-off frequency of the hardware high pass filter to 0.016 Hz and the cut-off frequency of the hardware low pass filter to 250 Hz. The voltage range was set to ± 16.384 mV, resulting in a resolution of 0.5 μV/bit. The EEG system clock was synchronized with the gradient clock of the MRI scanner via the Brain Products SyncBox device to ensure a highly accurate GA sampling. Sync status has been monitored. BrainVision Recorder (Brain Products GmbH, Gilching, Germany) software version 1.21.0102 was used for EEG data recording. The two ECG channels were treated like EEG channels, hence, EEG settings also apply to ECG recordings. We carefully prepared the electrode skin contact with abrasive electrode gel, but we were not able to control the skin impedances. It would appear that separate ground and reference electrodes must be mandatory to measure impedances with that EEG system.
Pre-processing Procedure of Outside-MRI-Scanner EEG
After the experiments, outside-MRI-scanner EEG recordings were down-sampled from 5000–250 Hz, using the “Change sampling rate” transformation in the BrainVision Analyzer software (Brain Products GmbH, Gilching, Germany, version 2.1.1.327). That included a 112.5 Hz low-pass anti-aliasing filter with 24 dB/oct damping before the down-sampling. The down-sampling itself is based on spline interpolation. See also Fig. 2a for a summary of the pre-processing procedure. We refer to the EEG after this procedure of outside-MRI-scanner EEG recording and offline EEG pre-processing, when we write of “outside EEG” in upcoming chapters.
Offline AAS Artifact Reduction Procedure of Inside-MRI-Scanner EEG
BrainVision Analyzer was used to perform artifact reduction offline and included the following steps: (1) Removing signal offsets with a high-pass filter (Butterworth zero phase, cut-off at 1 Hz, 4th order). (2) The next step was GA reduction with AAS as implemented in BrainVision Analyzer. The MRI scanner was sending TTL level triggers during the data recording, to mark new volumes. These markers were used to divide the EEG recordings into GA epochs. GA templates have been calculated separately for each epoch by averaging over 100 adjacent artifact epochs, 50 before and 50 after the current epoch. Subsequently, GA templates were subtracted from EEG recordings and all recordings were down sampled to 250 Hz (low-pass anti-aliasing filter, 112.5 Hz cutoff frequency, 24dB/oct damping). (3) AAS was carried out a second time for PA reduction. To divide the EEG recordings into PA epochs, the semiautomatic R-peak detection mode of the BrainVision Analyzer software was used. In that mode, R-peaks are detected automatically in separate ECG recordings, manually readjusted and subsequently used as markers. As in the GA reduction step, a separate template for subtraction was computed for each PA epoch. 50 adjacent PA epochs, 25 epochs before and 25 epochs after each PA, have been averaged to obtain the PA templates. The procedure is summarized in Fig. 2b. The number of epochs for averaging is a crucial parameter in AAS. It determines the adaptiveness of AAS templates as well as the EEG residuals in the AAS templates. Unfortunately, no gold standard has emerged yet for determining the number of epochs. We base our choice on the following argument: In one of the original papers on AAS (Allen et al. 2000), the aim was to obtain a clean artifact template, in which small events in the EEG are not covered by EEG residuals. They authors assumed that small EEG events have an amplitude of 10 µV and large EEG events have an amplitude of 250 µV, which leads to the use of 25 epochs (Allen et al. 2000). Beside the events argument, using 25 epochs implies that the RMS amplitude of the residual EEG in the template is reduced to 20% of the original RMS amplitude of the EEG, since the RMS amplitude is reduced by a factor of \({\sqrt {number\;of\;epochs} }\). Our goal was to at least maintain that level of residual EEG in two subsequent applications of AAS. Therefore, a reduction to 14% of the original RMS amplitude is necessary in each single step to maintain an overall reduction to 20%. 50 epochs for averaging are necessary to achieve that reduction to 14% and was therefore our choice for the minimum number of epochs. We name the EEG after this procedure of inside-MRI-scanner EEG recording and subsequent offline AAS, as “offline AAS EEG” throughout this work.
Online AAS Artifact Reduction Procedure of Inside-MRI-Scanner EEG
Inside-MRI-scanner EEG recordings were stored with BrainVision Recorder and were simultaneously sent to BrainVision RecView with the remote data access option of the BrainVision Recorder. Online artifact reduction in RecView included the following steps: (1) High-pass filtering to remove offsets (Butterworth filter, 1 Hz cut-off frequency, 24 dB/oct damping). (2) Online GA reduction with AAS. The TR was used to divide the past EEG into artifact epochs. The first 10 epochs per channel were averaged to compute initial individual GA templates. New epochs were added to the templates if the correlation of the new epoch with the current template exceeded a predefined threshold of 0.975. Subsequently, the current templates were subtracted online from the artifact afflicted EEG. (3) Subsequently, the EEG was down-sampled to 250 Hz (Butterworth low-pass anti-aliasing filter, 112.5 Hz cutoff, 24 dB/oct damping). (4) PAs were tackled with online AAS too. The past EEG was divided into epochs of PAs via online R-peak detection. Online R-peak detection in RecView is based on a template correlation approach. The method searches for a prototypical ECG epoch and subsequently compares it with the ongoing EEG. If certain thresholds are exceeded an epoch is found (settings: minimal pulse period 650 ms, minimal correlation 0.6, minimal amplitude 0.6, maximal amplitude 1.2). Separate PA templates were computed per channel by averaging over the last 50 PA epochs of the respective channel. The current templates were subtracted online from the artifact afflicted EEG. For an overview of this procedure see Fig. 2c. It can be assumed that this online artifact reduction procedure has a maximum delay of 150 ms. It takes 80–100 ms until the EEG data are available in RecView, including the hardware delay of the EEG system, transport of the EEG data via USB and the delay of the BrainVision Recorder software. The actual online artifact reduction in RecView is carried out sample-by-sample and hence, only a small additional delay is added. We assume that this delay is below 50 ms. We abbreviate the EEG after this artifact reduction procedure of inside-MRI-scanner recording and online AAS, with “online AAS EEG” in the following chapters.
Online AAS + RLAF Artifact Reduction Procedure of Inside-MRI-Scanner EEG
In accordance with previous works, we implemented the adaptive filtering as an additional processing step after GA and PA reduction with AAS (Chowdhury et al. 2014; Steyrl et al. 2017). Online AAS artifact reduction was carried out in BrainVision RecView (see description above). Subsequently, EEG data were transmitted to MATLAB (Mathworks Inc., Natick, MA, USA, Version 2012b) via the BrainVision RecView BCI2000 bridge. This bridge opens a TCP/IP server and the data can be received with any TCP/IP client. Brain Products recommends the pnet TCP/IP client from the TCP/UDP/IP Toolbox for receiving the data in MATLAB. Brain Products provide sample code on their homepage on how to use pnet. In MATLAB, the EEG data were adaptively filtered. The adaptive filtering was directly implemented in MATLAB with the following equations,
$${{\text{Subtraction step}}\quad eeg{{\left( n \right)}_{adaptive}}=eeg\left( n \right) - weight\left( n \right) \cdot ref\left( n \right)}$$
(1)
$${{\text{Weight update step}}\quad weight\left( {n+{\text{1}}} \right)=weight\left( n \right)+step \cdot eeg{{\left( n \right)}_{adaptive}} \cdot ref\left( n \right)}$$
(2)
where “n” is the current time sample, “eeg” is the signal of a scalp electrode, “ref” is the signal of the respective reference electrode, “weight” is the respective scaling factor, which we initialized with 1, and “eegadaptive” is the adaptively filtered EEG. “weight” can change its value over time, whereas “step” defines the speed of change. Finding a suitable value for “step” is a trade-off between speed of adaptation (large value) and preventing over-fitting (small value). Based on our experience, we choose a rather small value for “step” of 8 × 10e−7. Our implementation establishes first order models, hence the reference signals are scaled, but no bandwidth limiting filters are learned. The procedure is depicted in Fig. 2d. From here on we term the EEG after this procedure of inside-MRI-scanner recording and online AAS combined with online RLAF as “online AAS + RLAF EEG”.
Analysis and Performance Metrics
After a visual inspection of an EEG example, we analyze two very common EEG phenomena that were already used as performance criteria for artifact reduction methods in other publications (Chowdhury et al. 2014; Vanderperren et al. 2010). Namely, alpha rhythm amplitude changes and evoked potentials (EPs).
Alpha Rhythm Amplitude Changes
Oscillatory EEG components often show a brain activity related relative difference in their amplitude compared to a baseline. A prominent example is the occipital alpha rhythm. The amplitude at occipital EEG electrodes rises when one closes his/her eyes. The typical frequency range of that rise is 8–13 Hz. To visualize the amplitude changes, we computed spectra for the opened eyes period and the closed eyes period of the experiment respectively (Welch approach, window length 5 s, overlap 50%). We report the average spectra over the occipital channels (P3, Pz, P4, POz, O1, O2) separate for each participant.
To obtain a performance metric that describes the amplitude change of the alpha rhythm, we calculated the ratios of alpha amplitude between closed and opened eyes with the following equation
$${rati{o_\alpha }=\frac{{{A_{close{\text{8}} - {\text{13}}Hz}}}}{{{A_{open{\text{8}} - {\text{13}}Hz}}}}}$$
(3)
in which Aclose8−13Hz is the amplitude during the closed eyes period and Aopen8−13Hz is the amplitude during the opened eyes period. We report the average of the alpha amplitude ratio over occipital EEG channels (P3, Pz, P4, POz, O1, O2) separate for each participant.
Artifacts or noise in the EEG can cover the amplitude change. Hence, one expects that clean EEG shows a higher alpha amplitude ratio than artifact afflicted EEG. This is generally the case, however, the ratio metric can be distorted by artifacts that (1) have the same frequency range and (2) change with closed and opened eyes. This may apply to PAs. Their frequency range is overlapping with the alpha rhythm and if the PA detection rate is different between opened eyes and closed eyes, then omitted PA artifacts distort the alpha ratio metric. One can avoid this problem in offline PA reduction with AAS, since it is possible to manually search for omitted PAs and to mark them for PA reduction. However, it becomes a problem in online AAS, where a manual intervention is not possible. Therefore, we analyzed the PA detection rate in the online EEG data, and computed the percentage of detected PAs during opened eyes and closed eyes separately for each participant.
With regard to the alpha amplitude ratio metric, it is important to asses its topological distribution. We show the spatial distribution of the metric in separate topo-plots for each participant.
Visual Evoked Potentials
Evoked potentials are often investigated with respect to their amplitude. We computed the average visual evoked potential (VEPs) of each participant for all different artifact reduction procedures. The depicted channels were selected by the highest outside EEG VEP amplitude of the respective participant.
The VEP signal-to-noise-ratio (SNR) and the similarity of single VEPs to the respective mean VEP are important metrics to quantify VEP quality. We calculated both metrics. The SNR was calculated for each EEG channel separately using
$${VEPSN{R_{db}}={\text{20}} \cdot {\text{lo}}{{\text{g}}_{{\text{10}}}}\left( {\frac{{{A_{signal}}}}{{{A_{noise}}}}} \right)}$$
(4)
where VEP SNRdb is the signal-to-noise-ratio in dB, Asignal is the amplitude of the signal, and Anoise is the amplitude of the noise. We defined the signal amplitude (Asignal) as the peak-to-peak amplitude of the first and the second peak in the average VEP. Average VEPs were calculated by averaging band-limited (1–15 Hz) EEG over the VEP trials of the respective EEG channel. We defined the noise amplitude (Anoise) as the root-mean-square (RMS) amplitude of the band-limited (1–15 Hz) plus-minus (±) reference of the EEG signal of the respective EEG channel (Schimmel 1967). For the (±) reference, odd and even VEPs were averaged separately and subsequently, the average odd VEP was subtracted from the average even VEP. This difference is an estimator of the residual noise in the EEG (Schimmel 1967). The RMS amplitudes of Asignal and Anoise and therefore the SNR too, are dependent on the bandwidth of the EEG. A smaller bandwidth implies a smaller RMS amplitude and hence a higher SNR, as long as the EP amplitude stays constant. However, the choice of the bandwidth is not crucial as long as it is the same for all calculations, since our intention is to unveil relative differences between the methods. We report the average SNR over occipital EEG channels (POz, O1, O2) separately for each participant.
The root-mean-square (RMS) distance of single VEPs to the average VEP measures the similarity of single VEPs to the respective average VEP. This similarity to the average VEP is equivalent to the variability of single VEPs. The variability has two causes: noise in EEG and the inherent variability of VEPs. One cannot separate these two. However, offline AAS EEG, online AAS EEG and online AAS + RLAF EEG used the same raw EEG data, hence, the underlying inherent VEP variability was the same. Which means that a variability reduction was caused by the artifact reduction method that either reduces the noise in EEG or the inherent VEP variability, or both. It is important to keep in mind, that comparing the RMS distances of inside MRI scanner recordings with outside EEG is problematic since changes in distance could be caused by differences in the inherent VEP variability. RMS distances were normalized to the amplitude of the respective average VEP, since RMS distances are dependent on the absolute signal amplitudes. The distances were calculated per participant and per EEG channel with
$${RMS\;distanc{e_j}=\sqrt {\frac{{\text{1}}}{N}\sum\limits_{{n={\text{1}}}}^{N} {{{\left( {avgVEP\left( n \right) - VE{P_j}\left( n \right)} \right)}^{\text{2}}}} } }$$
(5)
$${NRMS\;distance=\frac{{avgRMS\;distance}}{{VE{P_{amplitude}}}}}$$
(6)
where NRMS distance is the average RMS distance divided by the amplitude of the respective average VEP. The “RMS distance” of the jth VEP to the average VEP was calculated using Eq. (5), where “n” is the nth time sample and “N” is the total number of time samples of the EEG data epochs. EEG data epochs had a length of half a second. We report the average NRMS distance over occipital EEG channels (POz, O1, O2) separate for each participant.