Supervised autoencoder for gradient and BCG artifact removal in EEG during simultaneous EEG-fMRI.

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Artificial Intelligence

Background:

• Simultaneous electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) offer complementary insights into brain function.
• Magnetic resonance imaging (MRI) introduces gradient (GA) and ballistocardiogram (BCG) artifacts that contaminate EEG signals, hindering analysis.
• Existing artifact removal methods often struggle to balance artifact suppression with physiological signal preservation.

Purpose of the Study:

• To introduce Deep Artifact Removal (DAR), a novel deep learning approach for suppressing MRI-induced artifacts in simultaneous EEG-fMRI.
• To evaluate DAR's effectiveness in artifact reduction and its ability to preserve physiological neural activity.
• To compare DAR's performance against established artifact removal techniques.

Main Methods:

• Developed DAR, a supervised 1D convolutional autoencoder.
• Trained DAR on the public Carbon Wire Loop EEG-fMRI dataset using paired artifact-contaminated and MR-corrected data segments.
• Validated DAR's performance using metrics such as Root Mean Square Error (RMSE) and Structural Similarity Index Measure (SSIM), and assessed signal-to-noise ratio (SNR) gain.
• Employed leave-one-subject-out cross-validation to test generalization to unseen subjects.
• Utilized saliency analysis to identify influential patterns for artifact reduction.

Main Results:

• DAR achieved excellent artifact suppression, with low RMSE (0.022 ± 0.015) and high SSIM (0.888 ± 0.091) across subjects.
• Demonstrated significant SNR gain (14.63 dB, p < 0.001) with DAR.
• Showcased strong generalization capabilities, maintaining good performance on unseen subjects (RMSE 0.063 ± 0.011, SSIM 0.666 ± 0.088).
• Preserved physiological fidelity by retaining over 75% of occipital alpha power (8-12 Hz).
• Outperformed PCA, ICA, average artifact subtraction, and wavelet denoising in key performance metrics.

Conclusions:

• Deep Artifact Removal (DAR) is a highly effective deep learning method for mitigating MRI artifacts in simultaneous EEG-fMRI.
• DAR significantly enhances EEG signal quality without compromising neurophysiologically relevant brain activity.
• DAR represents a superior alternative to conventional methods for artifact removal in combined EEG-fMRI studies.