Simultaneous acquisition of electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) EEG fMRI is extensively applied for brain mapping, but gradient artifacts (GA) and ballistocardiographic artifacts (BCG) remain a serious problem.
The most widely used method to correct both artifacts is averaged artifact subtraction (AAS)(EEG Electrodes), which averages epochs of data in the time domain to form a subtraction template. However, since the template cannot adapt to all variations across epochs, optimal basis sets (OBS)(Niazy et al., 2005) were proposed using principal component analysis (PCA), assuming that artifact-related components account for most of the signal variance. However, the number of components to form the basis of artifacts (Mandelkow et al., 2010) or to be separated from the signal-related components(Liu et al., 2012) is difficult to control, and may even lead to attenuation of true neuronal signals (EEG Data Analysis).
The method proposed here performs simple modifications of the OBS method to obtain a drastically better differentiation between the noise-related components and the non-noise-related components, thus greatly improving EEG-fMRI artifact removal.
Methods:
Data Acquisition
EEG data were acquired in 10 subjects (5 with simultaneous fMRI and 5 without scanning) in a Siemens Prisma 3T scanner and were sampled at a rate of 5000 Hz and filtered between 0.016-250 Hz using a 64-channel MR-compatible system (BrainProducts, Germany). The EEG clock was synchronized to the 10 MHz scanner clock.
Data Processing
For EEG-fMRI datasets, GAs were first removed by conventional OBS using PCA across TR epochs on each channel. The correction procedure was then repeated using two modifications that should result in a better separation of artifacts from true EEG signals: 1) PCA was instead applied on moving-average AAS templates, and 2) the data were pre-whitened before applying PCA. For both conventional OBS and the new method, components were then regressed out of the original signal by generalized least-squares.
For BCG correction, the same approach was employed across heart cycles.
Result Quantification
To assess GA removal, power spectral densities (PSD) were compared at harmonics of 1/TR. The corrected EEGs were also divided into non-overlapping segments of 10-TR epochs, within which the root-mean-square (RMS) power across TR was calculated. The standard deviation of RMS power was then compared to a null distribution generated by shuffling the signal points within epochs, thus determining whether the epochs still contained significant GAs (LeVan et al., 2016).
A similar quantification was performed across 10-heart-cycle epochs to assess BCG removal.
Results:
Fig 1 shows that the new method removed GAs better than conventional OBS. From Fig. 1b, it can be seen that conventional OBS attenuates signal power in the lower frequency band, even between 1/TR harmonics, indicative of distortions of real signals, while the high-frequency band still shows GA harmonics. In contrast, the new method resolves both of these issues. The average reduction of spectra power at 1/TR harmonics over conventional OBS across all subjects is 36.88%, while the proportion of epochs still containing artifacts was reduced by 14.68% compared to conventional OBS.
Fig 2 shows the BCG removal results. The new method outperformed conventional OBS in almost all channels. The proportion of artifactual epochs was reduced by 72.1% and 45.84% for EEG-fMRI datasets and datasets without scanning, respectively, compared to conventional OBS.
Conclusions:
The researchers (1) demonstrate simple modifications of the OBS method that result in large improvements for the removal of both GAs and BCG artifacts. Moreover, unlike conventional OBS, the new method does not appear to distort underlying neuronal signals, at least when performing GA removal. Future studies will be required to assess the preservation of these signals.
References
Shuoyue Zhang, Jürgen Hennig, Pierre LeVan. (2019).A “more” optimal basis set for EEG artifact denoising of EEG-fMRI
Allen, P.J, (2000). A Method for Removing Imaging Artifact from Continuous EEG Recorded during Functional MRI. NeuroImage 12, 230–239.Allen, P.J. (1998). Identification of EEG Events in the MR Scanner: The Problem of Pulse Artifact and a Method for Its Subtraction. NeuroImage 8, 229–239.Assländer, J. (2013). Single shot whole brain imaging using spherical stack of spirals trajectories. NeuroImage 73, 59–70. https://doi.org/10.1016/j.neuroimage.2013.01.065LeVan, P. (2016). EEG-fMRI Gradient Artifact Correction by Multiple Motion-Related Templates. IEEE Trans. Biomed. Eng. 63, 2647–2653.Liu, Z. (2012). Statistical Feature Extraction for Artifact Removal from Concurrent fMRI-EEG Recordings. Neuroimage 59, 2073–2087.Mandelkow, H. (2010). Good practices in EEG-MRI: The utility of retrospective synchronization and PCA for the removal of MRI gradient artefacts. NeuroImage 49, 2287–2303.Niazy, R.K. (2005). Removal of FMRI environment artifacts from EEG data using optimal basis sets. NeuroImage 28, 720–737.
The content published here is the exclusive responsibility of the authors.