Deep learning-based motion artifact removal in functional near-infrared spectroscopy

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Machine Learning

Background:

• Functional near-infrared spectroscopy (fNIRS) is a valuable neuroimaging tool for monitoring brain activity in unconstrained subjects.
• Motion artifacts are a significant source of noise in fNIRS data, often requiring expert intervention and parameter tuning for removal.
• Existing methods for motion artifact correction in fNIRS are limited by their reliance on prior knowledge and post-hoc adjustments.

Purpose of the Study:

• To develop an assumption-free deep learning method for motion artifact removal in fNIRS data.
• To introduce the application of a denoising autoencoder (DAE) architecture for mitigating motion artifacts in fNIRS.
• To establish a novel approach for enhancing the interpretability of fNIRS recordings.

Main Methods:

• A deep learning approach utilizing a denoising autoencoder (DAE) was developed for motion artifact removal.
• A specialized loss function was designed to facilitate the training of the DAE model.
• Synthetic fNIRS data mimicking real-world signal properties were generated for model training and validation.

Main Results:

• The DAE model demonstrated superior performance compared to conventional methods in reducing residual motion artifacts.
• The proposed method achieved a lower mean squared error, indicating improved signal fidelity.
• The deep learning approach significantly increased computational efficiency for artifact removal.

Conclusions:

• Deep learning models, specifically DAEs, show significant potential for accurate and efficient motion artifact removal in fNIRS data.
• This assumption-free method offers a robust alternative to existing artifact correction techniques.
• The findings pave the way for more reliable and accessible analysis of fNIRS neuroimaging data.