Real-time motion artifact suppression using convolution neural networks with penalty in fNIRS

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Motion artifacts (MAs) are a significant challenge in functional near-infrared spectroscopy (fNIRS) data acquisition.
• Standardized methods for MA removal in fNIRS are lacking, hindering practical applications.
• Artificial neural networks show promise for signal processing tasks but are underutilized in fNIRS.

Purpose of the Study:

• To develop and validate an innovative neural network-based approach for online processing of fNIRS signals.
• To address the challenge of motion artifact removal in fNIRS data.
• To create a method that is tailored to individual subjects and requires minimal prior experimental data.

Main Methods:

• Introduction of a one-dimensional convolutional neural network with a penalty network (1DCNNwP).
• Incorporation of a moving window and input data augmentation for signal processing.
• Training using simulated data from the balloon model and semi-simulated data for validation.

Main Results:

• 1DCNNwP effectively suppresses motion artifacts in fNIRS signals.
• Achieved over 11.08 dB improvement in signal-to-noise ratio, outperforming existing methods.
• Demonstrated enhanced contrast-to-noise ratios and significant outperformance in subject experiments with minimal processing time.

Conclusions:

• The 1DCNNwP approach offers a novel and effective univariate method for fNIRS signal processing.
• This method requires minimal prior experimental data and adapts to various experimental paradigms.
• The approach shows strong potential for real-time fNIRS data processing applications.