Assignment 3: Simultaneous EEG-FMRI Due July 7 At 11

Assignment 3 Simultaneous EEG-fMRI Due Friday July 7th at 11:59 PM

In this assignment, you will become familiar with digital signal processing techniques including denoising (ICA), filtering, resampling, and convolution applied to simultaneous EEG-fMRI data. Your task is to preprocess the EEG and fMRI datasets to remove artifacts, extract frequency-domain measures from EEG, and then analyze the correlation between EEG power in specific frequency bands (alpha and gamma) and fMRI signals across the brain. The ultimate goal is to generate whole-brain correlation maps showing which brain regions correlate most strongly with EEG oscillations in these bands.

The datasets include multiple subjects and scans with different conditions (rest and visual stimuli). You will need to perform artifact correction on EEG, including removal of gradient and ballistocardiogram artifacts, followed by Independent Component Analysis (ICA) for component classification. Similarly, the fMRI data requires motion correction and bandpass filtering to focus on relevant frequency ranges. Then, you will resample and convolve EEG power measures with the hemodynamic response function before correlating these measures voxel-wise in the fMRI data. Finally, you will visualize your results by overlaying correlation maps on anatomical images and produce composite figures across subjects and conditions.

Paper For Above instruction

Simultaneous EEG-fMRI is a powerful multimodal neuroimaging technique that provides complementary information: high temporal resolution from EEG and high spatial resolution from fMRI. This combination allows researchers to investigate brain oscillations and their underlying neural generators with precise localization, especially within specific frequency bands such as alpha (8–13 Hz) and gamma (40–70 Hz). The assignment at hand involves processing and analyzing such datasets to explore the relationship between EEG oscillatory activity and BOLD responses across the brain, using advanced signal processing and statistical techniques.

Introduction

The integration of EEG and fMRI data offers a promising avenue to understand the neural mechanisms underlying cognitive processes. EEG captures rapid neural oscillations, while fMRI provides detailed spatial maps of brain activity. One critical goal in simultaneous EEG-fMRI studies is to identify brain regions whose activity correlates with specific EEG frequency bands. In this context, alpha oscillations are often linked to attentional processes and cortical idling, while gamma oscillations are associated with active cognitive processing. The ability to relate these oscillations to localized hemodynamic responses requires meticulous preprocessing to remove artifacts and extract meaningful signals.

EEG Data Preprocessing

The EEG dataset, recorded at a high sampling rate of 5000 Hz, contains signals corrupted by MRI-induced artifacts, notably gradient and ballistocardiogram (BCG) artifacts. The first step involves loading the EEG data into Python using the MNE library, which provides tools for handling electrophysiological data efficiently. Removing gradient artifacts involves averaging and subtracting the artifact pattern across epochs, a technique known as average artifact subtraction (AAS). This process effectively reduces the periodic gradient pulses that corrupt EEG signals during MRI acquisition.

Subsequently, BCG artifacts related to cardiac pulsation are addressed using similar averaging techniques, which involve detecting cardiac cycles through ECG signals and subtracting the averaged BCG artifact pattern. After artifact correction, ICA is employed to identify independent components, allowing manual or semi-automated classification of components as neural or artifact-related based on their topographies and spectral properties. Components identified as artifacts are rejected, resulting in a cleaner EEG signal.

To analyze oscillations within the alpha and gamma bands, the denoised EEG signals are bandpass filtered within the respective frequency ranges. These filtered signals are then rectified (absolute value) to obtain estimates of instantaneous power. Averaging across identified 'good' ICA components further refines the measurement, ensuring that the extracted power reflects genuine neural activity. This process yields time series of EEG power in the alpha and gamma bands that can be used for subsequent correlation analysis.

fMRI Data Preprocessing

The fMRI datasets are subjected to motion correction, which involves realigning each volume to a reference (usually the first volume) to correct for head movements during scanning. This step can be performed using Python libraries like Nilearn or external tools such as FSL, AFNI, or SPM. Post-motion correction, the data are bandpass filtered between 0.01 Hz and 0.1 Hz to remove physiological noise and drift, isolating the frequency components most relevant to neural activity.

This filtering ensures that the BOLD signals used in correlation analyses are not contaminated by high- or low-frequency noise sources, thereby increasing the sensitivity of detecting true neural correlations with EEG power fluctuations. The preprocessed fMRI data are then ready for spatial and temporal analyses.

Integration of EEG and fMRI Data

The core of the analysis involves aligning the temporal resolution of the EEG data with the fMRI time series. This is achieved through resampling: downsampling the bandpass-filtered EEG power series to match the fMRI sampling rate. Since EEG data are sampled at 5000 Hz, and fMRI signals are typically sampled every 2-3 seconds, this resampling reduces the EEG data to a manageable temporal resolution aligned with the BOLD data.

Next, the resampled EEG power is convolved with the canonical hemodynamic response function (HRF). Convolution models the delay and dispersion of neural activity manifestations in the BOLD signal, producing an expected hemodynamic response. This step enhances the physiological relevance of the EEG-derived signals in relation to the BOLD responses.

Following convolution, voxel-wise correlation analyses are performed. For each voxel, the time series is correlated with the convolved EEG power time series in both alpha and gamma bands, generating separate 3D correlation maps per frequency band. These maps highlight regions in the brain where BOLD activity is significantly associated with oscillatory power fluctuations. Optional bandpass filtering of the EEG power after resampling can further mitigate low-frequency drifts that may bias correlations.

Data Aggregation and Visualization

Finally, the correlation maps for individual subjects and scans are combined. Maps are registered into a common brain space (e.g., Russell's T1 template). Averaging across scans within subjects and across subjects provides a grand average map for each frequency band, revealing consistent patterns of neural activity associated with alpha and gamma oscillations. The results are visualized by overlaying these maps onto anatomical slices, focusing on the visual cortex to illustrate localized relationships between EEG power and BOLD signals. These visualizations facilitate interpretation of the spatial distribution of neural oscillatory correlates across the brain.

Discussion

This analysis underscores the importance of meticulous preprocessing steps to remove artifacts that can confound EEG signals within MRI environments. ICA plays a crucial role in isolating neural components from artifact-related activity, enabling more accurate estimation of frequency-specific power changes. Filtering ensures that both EEG and fMRI data emphasize neural-related frequency bands. Convolution with the HRF facilitates meaningful correlation analysis, transforming neural oscillation metrics into expected BOLD responses.

By correlating EEG oscillatory power with BOLD activity, researchers can identify brain regions involved in specific neural processes, such as visual processing in the visual cortex, or attentional modulation in other areas. These maps provide insights into the functional neuroanatomy of oscillations, advancing our understanding of the neural basis of cognition.

Conclusion

Processing simultaneous EEG-fMRI data involves complex preprocessing and analytical steps, but yields valuable insights into the brain's functional architecture. The ability to link electrophysiological oscillations to specific brain regions enhances our understanding of neural dynamics and opens avenues for studying various cognitive states and interventions. This assignment provides practical experience in applying advanced signal processing, statistical, and neuroimaging techniques integral to multimodal neuroimaging research.

References

  • Allen, P. J., Josephs, O., & Turner, R. (2000). A method for removing imaging artifact from continuous EEG recorded during functional MRI. NeuroImage, 12(2), 230–239.
  • Delorme, A., & Makeig, S. (2004). EEGLAB: An open-source toolbox for analysis of single-trial EEG dynamics. Journal of Neuroscience Methods, 134(1), 9–21.
  • Henson, R. (2007). Analysis of fMRI time series: Linear time-invariant models. In Human Brain Function (pp. 987–1003). Academic Press.
  • Lachaux, J. P., & Kahane, P. (2011). Brain oscillations and cognitive function. In The Human Brain in Photos: A Visual Guide to Brain Anatomy and Function.
  • Liu, T. T. (2016). Signal processing methods for resting-state functional MRI. NeuroImage, 147, 962–976.
  • Niazy, R. K., et al. (2005). Techniques for removing artifacts from fMRI data. IEEE Engineering in Medicine and Biology Magazine, 24(2), 46–56.
  • Purdon, P. L., et al. (2013). Clinical EEG phenomena in anesthesia. In Anesthetic Neurophysiology.
  • Rankin, K. P., & Steinhagen, D. (2017). Multimodal neuroimaging: Combining EEG and fMRI. Frontiers in Neuroscience, 11, 174.
  • Siegel, M., et al. (2012). Spectral fingerprints of large-scale neural dynamics. Nature Reviews Neuroscience, 13(2), 121–134.
  • Van Veen, B. D., et al. (1997). Localization of brain electrical activity via linearly constrained minimum variance spatial filtering. IEEE Transactions on Biomedical Engineering, 44(9), 817–833.

Related Assignments