Removal of MRI Artifacts from EEG Recordings
Removal of MRI Artifacts from EEG Recordings
Abstract and Keywords
The simultaneous recording of electroencephalogram (EEG) and functional magnetic resonance imaging (fMRI) provides several advantages over multimodal integration based on separate EEG and fMRI recording protocols. However, the recording and analysis of simultaneous EEG-fMRI is not without pitfalls. The potential benefits of simultaneous recordings come at the expense of a massive, inevitable presence of artifacts, which corrupt the EEG signals recorded in the MR environment. This chapter presents different methods of EEG artifact correction. It also discusses the limitations of currently available approaches as well as possible future directions.
The simultaneous recording of electroencephalogram (EEG) and functional magnetic resonance imaging (fMRI) provides several advantages over multimodal integration based on separate EEG and fMRI recording protocols (Debener et al., 2006). However, the recording and analysis of simultaneous EEG-fMRI is not without pitfalls. The potential benefits of simultaneous recordings come at the expense of a massive, inevitable presence of artifacts, which corrupt the EEG signals recorded in the MR environment. It therefore comes as no surprise that many researchers find it difficult to obtain a reasonable EEG data quality in simultaneous EEG-fMRI recordings, or prefer separate protocols (Bledowski et al., 2007). However, the past few years have seen significant progress. With the advent of reliable hardware solutions, and increasing experience in artifact processing, multimodal integration has become feasible. The present chapter surveys the major types of artifacts evident in inside-scanner EEG recordings. With a focus on the most prominent artifacts, it presents different methods of EEG artifact correction; it also discusses the limitations of currently available approaches as well as possible future directions. For a discussion of MRI image quality, see Chapter 2.4.
Prominent MRI Artifacts in EEG Recordings
Depending on the MRI scanner field strength (B0), the MR protocol applied, the subject's behavior in the scanner, and the recording equipment used, EEG data recorded inside the MRI appear massively obscured to totally indiscernible. A typical example of continuous EEG recorded inside the scanner is illustrated in Figure 2.3.1, showing a few channels of an EEG recording taken during an fMRI echo planar imaging (EPI) recording in a 3 Tesla MRI scanner. As can be seen, visual inspection of the raw EEG traces appears impossible and necessitates further processing of the data. In fact, the application of offline artifact removal techniques is mandatory to enable a meaningful interpretation of the EEG signal. The principle assumption of such artifact removal is that the original EEG is linearly mixed with the artifact, the latter dominating the recording and thus rendering the visual inspection of typical EEG features such as occipital alpha oscillations impossible. Indeed, one particular type of artifact, which is related to MR image acquisition, contributes signals that can easily be several orders of magnitude larger than neural contributions to the EEG recording. Moreover, it is not just one artifact, but rather a number of qualitatively different artifact types that obscure the signal and contaminate the data over a broad frequency range overlapping with the frequency range of interest (0–100 Hz) of the physiological EEG. Following Allen et al. (2000), three groups of artifact can be distinguished:
Motion Related Artifacts
The movement of conductive materials, such as EEG electrodes and cables, in the static magnetic field induce an electromotive force. The strength of the induced current varies with a number of factors, such as the area of the conductive loop, the velocity, and the amount of motion and the strength of the static magnetic field (Huang-Hellinger et al., 1995). Accordingly, all kinds of motion listed below should be reduced to a minimum in simultaneous EEG-fMRI. This holds true in particular for transient, gross head motion, which can distort the EEG signal beyond recovery. Since head motion also deteriorates fMRI image quality, both modalities profit from the application of procedures preventing transient head motions (e.g., vacuum pillows, cushions) as well as more slowly occurring changes in head position. It is important to recognize that, compared (p.96) to mere fMRI recordings, the problem of head motion can be significantly amplified in simultaneous EEG-fMRI recordings, simply because of the discomfort caused by the posterior EEG electrodes, which the subject is lying on. In addition, at least two types of vibration-induced motion can contaminate the EEG signal. Vibration-induced motion can be related to the MRI scanner cryogen pump, a problem that could be bypassed by transiently switching off the cooling system. Vibration can also be caused by the gradient switching itself, which, of course, cannot be fully avoided. It is important to note that both types of vibration can induce motion of the subject and/or of EEG recording equipment (i. e., cables and amplifier). Therefore, beyond head fixation, the fixation of EEG cables and amplifiers may be beneficial (this and related recording procedures are covered in Chapter 2.1). And finally, endogenous and inevitable artifacts are induced by subtle head movement related to respiration and the cardiac cycle. The latter in particular forms another group of artifacts that will be discussed in more depth below.
MR Imaging–Related Artifacts
The switching of magnetic gradients is necessary for MRI image acquisition; and the related EEG signal distortions cannot be avoided by any shielding. Interleaved designs, where EEG recordings are analysed during the “silent” interval between MR scanning, have been used, but they do not solve the principle problem and restrict the efficiency of the experiment. The two types of imaging artifact that can be distinguished are the gradient artifact (GA), which causes a massive distortion of the EEG, and radio frequency (RF) artifact. Because the latter has a much higher frequency than the EEG signal, it can be effectively suppressed by analog low pass filtering. Analog low pass filtering is mandatory anyway to avoid EEG amplifier saturation and aliasing problems. Accordingly, the GA represents the major source of MR imaging artifact. The GA is a technical or exogenous artifact that reflects the imaging slice acquisition. Its major contribution is a very steep rising, transient signal with gradients that can be in the order of millivolts per millisecond. Most of the signal distortion visible in Figure 2.3.1 represents the GA. The GA completely dominates the EEG recording during MR image acquisition periods. Unlike the RF, the GA distorts the EEG spectrum over a broad frequency, including the frequency range of interest (< 100 Hz), and therefore cannot be fully accounted for by filtering (Allen et al., 2000). In fact, for many purposes, the GA is one of the two major artifacts that need to be dealt with statistically. Different ways of GA correction, and the challenges involved in it, will be discussed below.
Unlike the GA, which occurs only during slice acquisition periods, the ballistocardiogram (BCG) is always present in the scanner's magnetic field. The BCG contributes to the low (p.97) frequency portion of the EEG signal (< 15 Hz). A concurrent recording of the EEG with the electrocardiogram (ECG) reveals that the periodic distortion present in most EEG channels is related to the cardiac cycle. While the exact origin of the BCG is not known yet, it is likely that the pulsatile flow of blood associated with the cardiac cycle induces a rocking, nodding head motion (Anami et al., 2002). Figure 2.3.2 provides a schematic. Another source of influence could be that EEG electrodes (or cables) over, or adjacent to, pulsatile blood vessels are in steady motion. And finally, according to the Hall effect, the acceleration of blood, which is electrically conductive, could be a source of current induction that is registered in the EEG.
Consequently, it is important to recognize that the BCG is of mesogenous rather than purely exogenous or endogenous origin. It is a major problem for EEG recordings acquired in the MRI environment and approximately scales with the static magnetic (B0) field (Debener et al., 2008). However, its temporal variability reflects changes in heart rate and blood-flow parameters, which are under autonomous nervous system control. Accordingly, the BCG is subject to substantial temporal fluctuations, making its removal challenging.
Removal of the GA
The GA is generated by the magnetic field inside the scanner changing due to rapidly switching varying gradient fields and by the RF pulse, both of which are necessary to produce MR images. They represent a technical type of artifact that shows little fluctuation over time (see Figure 2.3.1). Accordingly, consecutive occurrences of MR volume (or slice) acquisitions can be expected to lead to similar EEG signal distortions. Another important characteristic of the GA is that it generates fast transients with large amplitudes on the order of 25000 μV/ms. Fortunately, MR-compatible EEG hardware solutions provide effective analog low pass filters that prevent the amplifier from drifting into saturation. Thus, the recorded signal, though carrying very (p.98) large artifact amplitudes, contains the normal EEG signal “riding on top” of the GA. This linear mixing renders the statistical removal of the GA possible.
Under the assumption of similarity of repetitive GAs, a GA template can be created by using the onset of each GA (per slice or per volume, respectively) as the time-locking event and then averaging over repeated GAs. Given enough trials, this effectively averages out the non–volume locked EEG signal and leaves a GA template that can be used subsequently to subtract out the GA from every single trial. This general procedure is known as average artifact subtraction (AAS) and was originally proposed by Allen et al. (2000). Previous attempts to deal with the GA included, among others, adaptive filtering (Sijbers et al., 1999) and frequency domain methods (Hoffmann et al., 2000), but proved less practical. Note that the morphologies and amplitudes of the GA (template) systematically differ between EEG channels, due to the different positions and orientations of the electrodes and cables with respect to the gradients, and other factors (see Hoffmann et al., 2000, for details). Hence, the GA template calculation and the corresponding GA correction is done for each channel separately.
A second important aspect concerns the temporal precision assumption, that is, the accuracy by which the GA is recorded and digitized. Given the enormous slope of the GA, an EEG sampling rate of, e.g., 250 Hz would mean that the GA amplitude could easily jump in steps of 100 μV (or more) from sample to sample. Accordingly, sampling rates are generally high in simultaneous EEG-fMRI experiments, in order to capture the steep GA slopes. Any temporal jitter that could for instance be introduced by asynchronous MR scanner and EEG recording computer clocks would then result in systematic errors, that is, residual artifact amplitudes that could still be larger than the EEG signal of interest.
In fact, both assumptions mentioned above are in reality not perfectly met: the GA amplitude and morphology fluctuates a little over time, and the precision by which the GA is recorded can suffer from temporal jitter, causing even more problems. Taking these two problems into account is the challenge for all GA correction procedures, and this has inspired a number of researchers to develop more sophisticated procedures, which generally include a number of processing steps, not just template subtraction. Allen et al. (2000) already combined the template subtraction approach with an upsampling procedure and adaptive noise cancellation, and these two procedures are still included in several currently used GA reduction implementations, albeit in different ways.
Challenges in GA Removal
As already noted, one complicating factor is that the GA can change over time, for instance due to different types of head movement (slow, involuntary movement; fast voluntary, temporally restricted movements). Because these types of head movement result in a different position and orientation of the EEG electrodes in the gradient field, the GA can be substantially different. Thus, a mean template based on the average across all recorded volumes may not be representative for single GA events. To compensate for the potential problem that single artifact instances could compromise the quality of the GA template (because they could be overlayed by other artifacts, say, gross head movement), Allen et al. proposed to include only those GA events that correlate with an initial GA template (based on the mean of the first five volumes) above r = 0.975, thus effectively removing outlier events.
Now consider the scenario that an individual's head position shifts a little, but continuously, over the EEG-fMRI recording duration. In this case, one would expect nearby GA events to be highly similar, whereas those being more distant in time would correlate to a much lesser amount. Following this line of reasoning, several AAS removal implementations now offer to calculate the GA template based on a moving average of adjacent volumes. Combined with an outlier identification procedure such as the one used by Allen, this would then account for involuntary head movements over time. To account for possible drifts in the GA over time, Beckeret al. (2005) introduced a weighted average based on an exponential decay function, thereby ensuring that adjacent artifact volumes have a stronger influence on the local template than artifact volumes that are more distant and thus might have a different morphology.
In addition to slow head movements, participants also move their head abruptly during data acquisition. It was shown that abrupt head movements larger than 1 mm head deflection do happen in about 30% of fMRI studies (Moosmann et al., 2009). This ratio is likely to be higher in simultaneous EEG-fMRI studies, in which discomfort may be caused by the EEG electrodes on which the subjects are lying. Abrupt head movements alter the geometry of electrodes and cables in the magnetic field and consequently the induced GA properties change. This leads to an increased heterogeneity of the GA and thus impairs EEG signal quality after GA correction (Laufs et al., 2008). The correction techniques mentioned above are adequate for homogeneous data or slow drifts of the artifact properties but cannot optimally represent transitions, i.e., when abrupt changes of the artifact properties occur. Whereas head movements in fMRI data are commonly corrected by the realignment preprocessing procedure (Friston et al., 1996), this issue has not been addressed for MR gradient–contaminated EEG data. Therefore modifications of the moving average algorithm have been proposed recently (Moosmann et al., 2009; Sun and Hinrichs, 2009) that take the head movement parameters from the fMRI preprocessing into account to estimate improved artifact templates that serve as local filters to correct the distorted EEG. More precisely, the subject's movement information is used to calculate a correction matrix that codes the position of the window of GA volumes being part of a specific template. Movements above a certain threshold act as a barrier in order to avoid averaging over discontinuities of artifact properties. Templates for GA volumes before/after a head (p.99) movement are generated from GA volumes before/after the movement only (See Figure 2.3.3 for an illustration of the method). The application of this method (Moosmann et al., 2009) will result in a better signal-to-noise ratio and a smaller residual variance around events of head movements compared to standard template correction methods. The realignment parameter–informed algorithm has now been realized as a Matlab plug-in for the open-source EEGLAB environment (Delorme and Makeig, 2004).
As mentioned above, the GA amplitude and morphology fluctuates over time due to a lack of temporal precision, the so-called jitter-problem. Because of an inaccuracy of the internal clocks of the EEG and MR system, the temporal properties of the gradient switching process (GHz) and limits in the sampling rate of the EEG (kHz), the GA is not always digitized at the same location in time, introducing a variation in the shape of the GA from volume to volume. An upsampling up to 100 kHz was proposed to adequately sample the slope of the gradient followed by a phase correction via a temporal alignment using a cross-correlation function (Allen et al., 2000; Becker et al., 2005; Niazy et al., 2005). Goncalves et al. (2007) proposed a selective average subtraction on slice and on volume level to meet jitter-related GA variations. Niazy et al. (2005) further introduced the fMRI artifact template removal (FASTR) procedure, which is implemented as part of the FMRIB toolbox, a plug-in for EEGLAB. FASTR reduces jitter-related temporal variations of the GA by generating a unique artifact template for each volume. It combines a local moving average as a template with a linear combination of basis functions, which are derived from a temporal principal component analysis (PCA) on the residual artifacts.
However, jitter-related issues can be avoided by synchronizing scanner clock and the EEG recording computer clock (Anami et al., 2003; Mandelkow et al., 2006; Mullinger et al., 2008). The synchronization of the systems results in more homogeneous GA volumes, which better meet the assumptions of the template-based correction methods and thus result in cleaner EEG signals after GA correction.
Anami and colleagues went one step further and modified the fMRI imaging sequence timing parameters. They programmed MR gradients to be switched during the times when no EEG data point was sampled, theoretically resulting in minimal GA (Anami et al., 2003).
As discussed above, template-based GA correction methods have been proven to be very successful to correct GA. However, since the assumptions for applying any of the proposed algorithms are not always met, the corrected EEG data might still be contaminated by residual artifacts. To correct residual artifacts a second processing step was proposed. Niazy and colleages (2005) employed optimal basis sets of orthogonal components to filter residual imaging artifacts. This method is discussed in more detail in the paragraph on the removal of ballistocardiographic artifacts below. Brookes et al. (2008) suggested a spatial beamformer technique, which uses spatial filters that extract the components of a signal with (p.100) a specific spatial characteristic. It is able to localize electrical activity as well as to remove residual artifacts that may not be eliminated by the AAS technique. Mathematically, the filter is based on a weighted sum of measurements made at each of the EEG electrodes. This weighted sum gives an estimate of local electrical source strength at some predetermined location in the brain. Sequential application of the spatial filter to a number of locations (voxels) in the brain will then yield a volumetric image of source power. To correct residual artifacts related to head movements Masterton et al. proposed a method using wire loops that were attached to the electrode cap to measure subject movements (Masterton et al., 2007). Linear adaptive filtering based on recursive least squares was used to reduce the artifact power online while preserving the physiological EEG signal.
Properties and Removal of the BCG
In most recording conditions the BCG is clearly visible after gradient artifact removal. In the absence of scanning it contributes signals to the frequency range that are close to the range of the EEG. Although it exists in EEG recordings outside the scanner, the large amplitude BCG in the scanner arises due to the interaction between the active cardiovascular system (endogenous contribution) and the main static (B0) field inside the MRI scanner (exogenous contribution), rendering it a mixed, or mesogenous artifact. This characterization already outlines some potential problems associated with the variability inherent in the BCG, i.e., the MRI environment (e.g., the scanner field strength) and the cardiovascular characteristics (e.g., heart rate variability) influence the resulting BCG.
The BCG is usually present in all recorded EEG channels, albeit to a different extent; as a rule of thumb, EEG electrodes far from the EEG reference electrode express larger amplitudes, and a high field MRI scanner causes larger amplitudes than a scanner with lower field strength (Debener et al., 2008), as the BCG amplitude is proportional to the MRI scanner B0 field strength (Tenforde et al., 1983). Note the general temporal pattern inherent in the BCG is synchronized to the cardiac cycle. This becomes apparent when comparing the EEG traces with the simultaneously recorded electrocardiogram (ECG) signal. Importantly, this comparison reveals a delay of approximately 200 ms between the ECG R peak and the peak amplitude of the BCG in the EEG traces (Allen et al., 1998). Due to its close relationship to the cardiac cycle, however, fluctuations in the subject's heart rate (and likely in other cardiovascular parameters) result in fluctuations in the BCG. Also, BCG peak latencies and BCG morphologies can be different across channels, suggesting a complex spatiotemporal activity pattern that the BCG contributes to the EEG.
The BCG represents a rather complicated, dynamic contribution to the EEG and therefore requires attention when BCG removal procedures are developed, applied, and compared. The spatial complexity of the BCG has only recently (Nakamura, et al., 2006) been investigated in more detail. A spatial analysis may help to characterize important properties of the BCG with possible implications for its reduction or removal (Debener et al., 2008). Figure 2.3.4 illustrates the main spatial features of the BCG, showing the time-domain averaged signal. The upper traces in the figure show the evoked BCG activity at all EEG electrodes together with the mean global field power (GFP). It can be seen that the BCG starts approximately 150 ms after the ECG Q peak and is characterized by two main power maxima at approximately 230 and 330 ms. Also, the BCG lasts at least until approximately 500 ms after the ECG Q peak. Corresponding voltage maps at selected GFP peak latencies reveal several interesting features: First, in most cases the BCG topography can be characterized by a low spatial frequency, meaning that most electrodes contribute to generally smooth topographies. Therefore, the BCG can best be studied based on adequate spatial coverage of the head sphere. Second, as noted above, these topographies change substantially over time. At first glance, it appears that several topographies recur at later latencies (for example, compare maps at 176 and 464 ms in Figure 2.3.2). However, closer inspection reveals that these maps appear slightly rotated to each other. This rotational aspect demonstrates that the BCG does not simply contribute a single topography. Rather, it appears that the BCG represents very dynamic (moving, rotating, and polarity-inverting) activity.
The lower part of Figure 2.3.4 shows all single ECG trials, along with the averaged ECG. Note that the ECG morphology is also compromised by the magnetic field influence. Bear in mind that this illustrates a single subject, based on 30-channel EEG recordings obtained in a 1.5 Tesla MRI scanner. The illustrated BCG features appear to be similar across subjects and MRI scanners. However, some important differences can be expected as well. First of all, because the BCG scales approximately in proportion to the MRI B0 field (Tenforde et al., 1983), the BCG is much smaller in amplitude at 1.5 Tesla compared to 3 or 7 Tesla recordings, which has consequences for the choice of the BCG removal technique (Debener et al., 2008). Second, individuals differ in fluctuations of cardiac activity, such as heart rate changes. Therefore, for those individuals with a higher heart rate, the BCG activities between adjacent cardiac cycles may overlap to some extent, which could further complicate BCG removal issues (Vincent et al., 2007). The peak latencies (and the exact morphologies) shown in Figure 2.3.2 will differ between subjects (Allen et al., 1998). The features discussed above seem to be fairly consistent across subjects and scanner sites. Among them is the ∼200 ms delay as well as the general topography, with its dynamic, rotational, and polarity-inverting aspects. In fact, it was consistently found that the moving topographical BCG pattern is irrespective of the number of EEG channels (30 to 62), the MRI scanner manufacturer (Siemens, Philips), the MRI scanner type (head scanner, whole body scanner), or the MRI B0 field strength (1.5, 3, and 7 Tesla).
Allen and colleagues introduced the average artifact subtraction (AAS) approach (Allen et al., 1998), which has been among the most influential and frequently used methods for BCG removal. While different implementations and developments of the AAS exist (Laufs et al., 2008), the basic principle of the AAS is common to all variants.
Figure 2.3.5 illustrates and highlights several features of the AAS. First, the AAS approach requires estimates of the onset of each cardiac-cycle from the concurrently recorded ECG. The next step is then to define an artifact template; this is done with a moving average procedure for each EEG channel separately through averaging the EEG time-locked to each cardiac cycle onset. The resulting average represents the evoked BCG, and this template can then be subtracted from each EEG epoch, thus removing the major fraction of the BCG. This procedure often yields satisfactory EEG data quality (e.g., Sammer et al., 2005; Hamandi et al., 2008). Another strength of the AAS is the straightforward possibility for a real-time implementation that then affords an online evaluation of the EEG.
However, experience with the AAS has led to the identification of several potential pitfalls, and has motivated the further development of the original approach. For example, the ECG channel itself is also contaminated with gradient and BCG artifact contributions, making it sometimes hard to identify the onset of each heartbeat cycle in the ECG. As a result, automatic R-peak detection algorithms that work on (p.102) artifact-free ECG recordings sometimes fail on inside scanner recordings and result in an inaccurate positioning of the event markers (Debener et al., 2008). Some software packages such as BrainVision Analyzer (Brain Products GmbH, Munich, Germany) already take jitter information into account, and automatically align markers statistically such that the overall jitter is minimized before the AAS correction is performed. A second main problem of this approach is the assumption of similarity between adjacent BCG occurrences. That is, within the chosen moving average window size, it is assumed that the BCG contribution to the EEG channel artifact is very similar at adjacent cardiac cycles and changes only slowly over time. This assumption may not always be correct. Shortening the moving average window size can not fully address this problem (as a smaller moving average window would leave more residual EEG activity in the template). Several groups have suggested alternative template constructions that are based on weighted averages (Goldman et al., 2000) or median instead of mean values (Sijbers et al., 2000). Also, some implementations allow for the selection of trials contributing to the template generation to depend on whether they correlate sufficiently with other trials or not. This option helps to ensure that trials containing other EEG artifacts are excluded, thus improving the quality of the final template being used. While all these features may improve the BCG correction quality, they do not fully address the basic problem, as they still rest on the assumption of local BCG similarity, and assume that one template is sufficient for each BCG epoch.
These latter problems have been addressed by Niazy and colleagues (Niazy et al., 2005), and similarly by Negishi et al. (2004), who proposed a new way of constructing a BCG template. These authors suggested generating BCG templates based on a channel-wise temporal principal components analysis (PCA), thereby not assuming any local BCG similarity. Niazy and colleagues named this approach the optimal basis set (OBS) method, which refers to the view that the first few principal components represent several distinct BCG templates and at the same time explain most of the BCG variance in any given EEG channel. These templates are jointly used to regress out the BCG from the EEG data. The OBS approach does not assume that adjacent BCGs are more similar than more distant ones, and accounts for the possibility of different artifact shapes. It has received considerable attention, and indeed several groups have used it successfully. This tool is a freely available Matlab plug-in that interacts with the open-source EEGLAB environment (Delorme and Makeig, 2004). However, the number of principal components used as BCG templates in the OBS approach has to be chosen by the user, such that bias and over- and underfitting become relevant issues.
A further problem that is common to all channel-by-channel template subtraction approaches is the choice of template length. With a constantly changing interbeat (R-R) interval, the length of the template may also need some adjustment over time. A template covering an inappropriately short or long interval may either leave residual activity in, or add spurious activity to, the EEG traces. Indeed, because the interbeat interval changes over time (e.g., due to respiratory arrhythmia), an AAS template that may be of appropriate length for some cardiac cycles may not be adequate for other cycles. This has led to the introduction of alternative template (p.103) generation schemes that either scale the BCG template with a percentage of the mean R-R period in the moving average window (Ellingson et al., 2004) or that build a template that incorporates the BCG data for all R-R period lengths present in the current moving average window (BrainVision Analyzer software). In the latter case, this template is then adaptively applied to each heartbeat based on its R-R period, thus ensuring that no portion of the BCG artifact remains uncorrected due to a suboptimal template length.
Other channel-by-channel correction approaches exist that account for the template duration problem. Bonmassar and colleagues for instance utilized an adaptive Kalman filter approach (Bonmassar et al., 2002), but this requires an extra motion sensor signal to be recorded as a reference signal and, while producing good corrections, appears to be computationally demanding. Other methods such as the wavelet-based non-linear reduction of the BCG (Wan et al., 2006) are also computationally demanding and therefore unlikely to replace the AAS in the near future. Recently, Vincent and colleagues (2007) proposed a moving general linear model approach (mGLM). These authors point to a potentially important problem of the AAS and related BCG temporal template correction approaches: they fail to account for BCG artifacts that last longer than a cardiac cycle. It remains to be determined whether this is indeed a significant problem, and whether the mGLM approach provides a substantial improvement over the AAS.
Removing the BCG Using Spatial Pattern Removal Approaches
The BCG can also be characterized by a number of prototypical topographies. The potential virtue of spatial approaches, aiming at removing those topographies, is that exact knowledge about the onset of each cardiac cycle is not necessarily required. In principle, a spatial approach can avoid problems that are inherent in AAS and OBS as discussed above.
Motivated by the success of spatial approaches for the removal of ocular artifacts, two spatial BCG correction approaches have been proposed (Bénar et al., 2003): principal components analysis (PCA) and independent component analysis (ICA). The assumption behind these is that the BCG contribution is statistically independent of, or in the case of PCA uncorrelated to, ongoing EEG activity. Therefore, BCG activity can be expected to be identified by few components, whereas all other EEG activity should be represented by other components. In the original work by Bénar and colleagues, BCG components were visually identified by exploring the similarity of all component time courses to the simultaneously recorded ECG signal (Bénar et al., 2003). Back-projection of all but the identified components then reduced the BCG in the EEG recordings. Bénar and colleagues found both ICA and PCA well suited for this task, as they eliminated BCG activity while preserving the relative amplitude of epileptic spikes.
However, an obvious problem of this approach is selection of components based of subjective criteria. Indeed, it can be rather difficult to visually identify and select the components representing BCG activity beyond the first few strongest and most obvious. Therefore, other methods of BCG component identification have been proposed. Srivastava and colleagues (2005), for instance, proposed the identification-by-correlation approach, by which all ICA time courses are correlated with the simultaneously recorded ECG channel and those components expressing the highest correlations with the ECG are identified as BCG-components. Alternatively, BCG-related components can be identified by the amount of variance they contributed to the evoked BCG. When compared to the identification-by-correlation criterion, this latter approach gives better results (Debener et al., 2008).
Several groups have reported success in using ICA for BCG removal (Bénar et al., 2003; Eichele et al., 2005; Briselli et al., 2006; Nakamura et al., 2006; Mantini et al., 2007) in 1.5 T MRI scanners, while the experiences in a 3 Tesla MRI scanner were much less positive (Debener et al., 2007). Note that a model assumption of temporal ICA is that the sources contributing to the linearly mixed signals are spatially stationary. In a recent study, the properties of the BCG artifact and the performance of ICA at 1.5, 3 and 7 Tesla were compared. Compared to ICA decompositions based on outside MRI scanner data, it was found that the ICA results obtained from 1.5 Tesla recordings were only moderately affected in their quality. At 3 and 7 Tesla, however, the typical topographies could hardly be recovered anymore. While it appears that ICA is robust to violation of the stationarity assumption at lower field strength, it is more sensitive and requires more components to model the convolutive (moving, rotating) properties of the BCG artifact at higher field strengths.
Currently, the available evidence suggests that the spatial filtering approaches such as ICA and PCA are not as efficient as template methods and in particular at fields of 3 Tesla and higher. Nevertheless, as will be pointed out in the next section, the application of ICA in combination with other approaches such as AAS or OBS seems very attractive, as it provides several advantages that may otherwise be impossible to achieve, and does so even at higher field strengths (Debener et al., 2006).
Combining and Comparing Different BCG Removal Approaches
It appears that no single method exists that performs optimally under all circumstances, and many authors consider the combination of different BCG removal algorithms a very attractive option. Kim and colleagues, for instance, combined a wavelet-based denoising approach with recursive adaptive filtering as post-processing only in the case that AAS gave not satisfactory results (Kim et al., 2004). Likewise, the FMRIB (p.104) plug-in developed by Niazy et al. (2005) offers application of adaptive noise cancellation following OBS to increase performance. Also, BrainVision Analyzer provides flexibility in combining different methods, such as optimized AAS and subsequent ICA with automatic determination of the BCG components to remove.
In the study mentioned above (Debener et al., 2007), three different BCG removal approaches were compared, namely ICA with the identification-by-correlation approach (Srivastava et al., 2005), OBS (Niazy et al., 2005), and a combination of both, where ICA was applied after BCG removal with OBS (Debener et al., 2005b). It was found that ICA, when used on its own, while reducing the BCG substantially, also reduced the SNR of ERPs, suggesting that artifact and signal could not be well disentangled by ICA. However, when used after OBS, which, as expected, reduced the BCG and improved the ERP SNR, infomax ICA could further improve the ERP SNR and the ERP topography. This pattern of finding suggests that ICA, after most of the BCG is removed with a channel-wise procedure such as AAS of OBS, is able to deal with residual artifact, which is well in line with the earlier argument. Moreover, in this combination, the other advantages of ICA can be utilized, that is, the removal of further EEG artifacts such as eye blinks, and the separation of brain-related signals from each other (Debener et al., 2005a; Debener et al., 2005b; Debener et al., 2006). It is important to note that not only the amount of BCG artifact removed should be considered. Analyzing whether the BCG is reduced or not does not make it possible to determine whether the separation from ongoing (or event-related) EEG activity was successful. It is therefore a better strategy to always analyze the data with regard to the amount of BCG correction and with regard to one or more measures of the quality of the EEG variable of interest. When the focus is on event-related potentials (ERPs) consider the signal-to-noise ratio (SNR) of the ERP component of interest. The topographical quality of the resulting ERPs could be quantified if a reference topography is available, or if information about the source configuration of the ERP component of interest exists (Debener et al., 2007). Also, measures of the recovery of the EEG spectrum can be included, in particular if the EEG variable of interest is in the frequency or time-frequency domain (Allen et al., 1998).
This chapter describes the different types of artifacts caused by the static and changing magnetic fields of the MR and the possibilities of their removal. The GA, due to its large amplitudes, is quite dominant on first sight but it has been shown to be more easily controllable than the the cardiac-related artifacts. This is mainly due to the fact that the GA is caused by an external source that is reliable in its temporal properties. Template-based correction methods have been shown to be most successful and convenient as a first step of GA removal. Modifications for dealing with motion-induced alteration of the geometry of the electrodes and cables were discussed. We believe that future work should focus on how to avoid or attenuate MR-related imaging artifacts altogether, or at least reduce the heterogeneity of the GA in order to further increase signal quality. Dedicated fMRI sequences should be designed that are tuned for EEG-fMRI recordings as it was pioneered by Anami and colleagues (2003). Recently developed “silent” fMRI sequences (Schmitter et al., 2008) could be advantageous because acoustic resonance peaks of the MR system during the gradient switching process are avoided. Consequently this would result in less vibration-induced motion of the EEG electrodes and cables.
Removal of cardiac-related artifacts remains challenging since their origins are partly physiologically caused. As the fMRI community moves on to larger field strengths to increase sensitivity, and the pulse artifacts scale with it, the issues of removing cardiac-related artifacts become even more challenging. It was shown that different methods and especially their combination have to be adapted to the given situation.
Improved artifact correction techniques may result from a better understanding of the mechanisms that give rise to the MR induced artifacts (Yan et al., 2009). In other words, a better knowledge about the different mechanisms causing GA and BCG should help to optimize recording conditions as well as offline data correction approaches, and thus contribute to a better EEG signal quality. Given the promises of simultaneous EEG-fMRI (e.g., Debener et al., 2006), this seems a worthwhile goal.
Allen PJ, Polizzi G, Krakow K, Fish DR, Lemieux L (1998) Identification of EEG events in the MR scanner: the problem of pulse artifact and a method for its subtraction. Neuroimage 8:229–239.
Allen PJ, Josephs O, Turner R (2000) A method for removing imaging artifact from continuous EEG recorded during functional MRI. Neuroimage 12:230–239.
Anami K, Saitoh O, Yumoto M (2002) Reduction of ballistocardiogramwith a vacuum head-fixating system during simultaneous fMRI and multi-channel monopolar EEG recording. Recent Adv Hum Brain Mapp 1232:427–431.
Anami K, Mori T, Tanaka F, Kawagoe Y, Okamoto J, Yarita M, Ohnishi T, Yumoto M, Matsuda H, Saitoh O (2003) Stepping stone sampling for retrieving artifact-free electroencephalogram during functional magnetic resonance imaging. Neuroimage 19:281–295.
Becker R, Ritter P, Moosmann M, Villringer A (2005) Visual evoked potentials recovered from fMRI scan periods. Hum Brain Mapp 26:221–230.
Bénar C, Aghakhani Y, Wang Y, Izenberg A, Al-Asmi A, Dubeau F, Gotman J (2003) Quality of EEG in simultaneous EEG-fMRI for epilepsy. Clin Neurophysiol 114:569–580.
Bledowski C, Linden DE, Wibral M (2007) Combining electrophysiology and functional imaging: different methods for different questions. Trends Cogn Sci 11:500–502.
(p.105) Bonmassar G, Purdon PL, Jaaskelainen IP, Chiappa K, Solo V, Brown EN, Belliveau JW (2002) Motion and ballistocardiogram artifact removal for interleaved recording of EEG and EPs during MRI. Neuroimage 16:1127–1141.
Briselli E, Garreffa G, Bianchi L, Bianciardi M, Macaluso E, Abbafati M, Grazia Marciani M, Maraviglia B (2006) An independent component analysis-based approach on ballistocardiogram artifact removing. Magn Reson Imaging 24:393–400.
Brookes MJ, Mullinger KJ, Stevenson CM, Morris PG, Bowtell R (2008) Simultaneous EEG source localisation and artifact rejection during concurrent fMRI by means of spatial filtering. Neuroimage 40:1090–1104.
Debener S, Makeig S, Delorme A, Engel AK (2005a) What is novel in the novelty oddball paradigm? Functional significance of the novelty P3 event-related potential as revealed by independent component analysis. Brain Res Cogn Brain Res 22:309–321.
Debener S, Ullsperger M, Siegel M, Fiehler K, von Cramon DY, Engel AK (2005b) Trial-by-trial coupling of concurrent electroencephalogram and functional magnetic resonance imaging identifies the dynamics of performance monitoring. J Neurosci 25:11730–11737.
Debener S, Ullsperger M, Siegel M, Engel AK (2006) Single-trial EEG-fMRI reveals the dynamics of cognitive function. Trends Cogn Sci 10:558–563.
Debener S, Strobel A, Sorger B, Peters J, Kranczioch C, Engel AK, Goebel R (2007) Improved quality of auditory event-related potentials recorded simultaneously with 3-T fMRI: removal of the ballistocardiogram artefact. Neuroimage 34:587–597.
Debener S, Mullinger KJ, Niazy RK, Bowtell RW (2008) Properties of the ballistocardiogram artefact as revealed by EEG recordings at 1.5, 3 and 7 T static magnetic field strength. Int J Psychophysiol 67:189–199.
Delorme A, Makeig S (2004) EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. J Neurosci Methods 134:9–21.
Eichele T, Specht K, Moosmann M, Jongsma ML, Quiroga RQ, Nordby H, Hugdahl K (2005) Assessing the spatiotemporal evolution of neuronal activation with single-trial event-related potentials and functional MRI. Proc Natl Acad Sci U S A 102:17798–17803.
Ellingson ML, Liebenthal E, Spanaki MV, Prieto TE, Binder JR, Ropella KM (2004) Ballistocardiogram artifact reduction in the simultaneous acquisition of auditory ERPS and fMRI. Neuroimage 22:1534–1542.
Friston KJ, Williams S, Howard R, Frackowiak RS, Turner R (1996) Movement-related effects in fMRI time-series. Magn Reson Med 35:346–355.
Goldman RI, Stern JM, Engel J, Jr., Cohen MS (2000) Acquiring simultaneous EEG and functional MRI. Clin Neurophysiol 111:1974–1980.
Goncalves SI, Pouwels PJ, Kuijer JP, Heethaar RM, de Munck JC (2007) Artifact removal in co-registered EEG/fMRI by selective average subtraction. Clin Neurophysiol 118:2437–2450.
Hamandi K, Laufs H, Noth U, Carmichael DW, Duncan JS, Lemieux L (2008) BOLD and perfusion changes during epileptic generalised spike wave activity. Neuroimage 39:608–618.
Hoffmann A, Jager L, Werhahn KJ, Jaschke M, Noachtar S, Reiser M (2000) Electroencephalography during functional echo-planar imaging: detection of epileptic spikes using post-processing methods. Magn Reson Med 44:791–798.
Huang-Hellinger FR, Breiter HC, McCormack GM, Cohen MS, Kwong KK, Savoy RL, Weisskoff RM, Davis TL, Baker JR, Belliveau JW, Rosen BR (1995) Simultaneous functional magnetic resonance imaging and electrophysiological recording. Hum Brain Mapp 3:13–23.Kim KH, Yoon HW, Park HW (2004) Improved ballistocardiac artifact removal from the electroencephalogram recorded in fMRI. J Neurosci Methods 135:193–203.
Laufs H, Daunizeau J, Carmichael DW, Kleinschmidt A (2008) Recent advances in recording electrophysiological data simultaneously with magnetic resonance imaging. Neuroimage 40:515–528.
Mandelkow H, Halder P, Boesiger P, Brandeis D (2006) Synchronization facilitates removal of MRI artefacts from concurrent EEG recordings and increases usable bandwidth. Neuroimage 32:1120–1126.
Mantini D, Perrucci MG, Cugini S, Ferretti A, Romani GL, Del Gratta C (2007) Complete artifact removal for EEG recorded during continuous fMRI using independent component analysis. Neuroimage 34:598–607.
Masterton RA, Abbott DF, Fleming SW, Jackson GD (2007) Measurement and reduction of motion and ballistocardiogram artefacts from simultaneous EEG and fMRI recordings. Neuroimage 37:202–211.
Moosmann M, Schonfelder VH, Specht K, Scheeringa R, Nordby H, Hugdahl K (2009) Realignment parameter-informed artefact correction for simultaneous EEG-fMRI recordings. Neuroimage 45:1144–1150.
Mullinger KJ, Morgan PS, Bowtell RW (2008) Improved artifact correction for combined electroencephalography/functional MRI by means of synchronization and use of vectorcardiogram recordings. J Magn Reson Imaging 27:607–616.
Nakamura W, Anami K, Mori T, Saitoh O, Cichocki A, Amari S (2006) Removal of ballistocardiogram artifacts from simultaneously recorded EEG and fMRI data using independent component analysis. IEEE Trans Biomed Eng 53:1294–1308.
Negishi M, Abildgaard M, Nixon T, Constable RT (2004) Removal of time-varying gradient artifacts from EEG data acquired during continuous fMRI. Clin Neurophysiol 115:2181–2192.
Niazy RK, Beckmann CF, Iannetti GD, Brady JM, Smith SM (2005) Removal of FMRI environment artifacts from EEG data using optimal basis sets. Neuroimage 28:720–737.
Sammer G, Blecker C, Gebhardt H, Kirsch P, Stark R, Vaitl D (2005) Acquisition of typical EEG waveforms during fMRI: SSVEP, LRP, and frontal theta. Neuroimage 24:1012–1024.
Schmitter S, Diesch E, Amann M, Kroll A, Moayer M, Schad LR (2008) Silent echo-planar imaging for auditory FMRI. Magma 21:317–325.
Sijbers J, Michiels I, Verhoye M, Van Audekerke J, Van der LA, Van Dyck D (1999) Restoration of MR-induced artifacts in simultaneously recorded MR/EEG data. Magn Reson Imaging 17:1383–1391.
Sijbers J, Van Audekerke J, Verhoye M, Van der Linden A, Van Dyck D (2000) Reduction of ECG and gradient related artifacts in simultaneously recorded human EEG/MRI data. Magn Reson Imaging 18:881–886.
Srivastava G, Crottaz-Herbette S, Lau KM, Glover GH, Menon V (2005) ICA-based procedures for removing ballistocardiogram artifacts from EEG data acquired in the MRI scanner. Neuroimage 24:50–60.
(p.106) Sun L, Hinrichs H (2009) Simultaneously recorded EEG-fMRI: Removal of gradient artifacts by subtraction of head movement related average artifact waveforms. Hum Brain Mapp Apr 13 [Epub ahead of print].
Tenforde TS, Gaffey CT, Moyer BR, Budinger TF (1983) Cardiovascular alterations in Macaca monkeys exposed to stationary magnetic fields: experimental observations and theoretical analysis. Bioelectromagnetics 4:1–9.
Vincent JL, Larson-Prior LJ, Zempel JM, Snyder AZ (2007) Moving GLM ballistocardiogram artifact reduction for EEG acquired simultaneously with fMRI. Clin Neurophysiol 118:981–998.
Wan X, Iwata K, Riera J, Ozaki T, Kitamura M, Kawashima R (2006) Artifact reduction for EEG/fMRI recording: nonlinear reduction of ballistocardiogram artifacts. Clin Neurophysiol 117:668–680.
Yan WX, Mullinger KJ, Brookes MJ, Bowtell R (2009) Understanding gradient artefacts in simultaneous EEG/fMRI. Neuroimage 46:459–471.