A deep convolutional neural network for estimating hemodynamic response function with reduction of motion artifacts...

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Signal Processing

Background:

• Functional near-infrared spectroscopy (fNIRS) non-invasively monitors hemoglobin changes.
• Subject motion introduces significant artifacts, complicating data analysis.
• Current artifact suppression methods have limitations in parameter selection.

Purpose of the Study:

• To develop a deep convolutional neural network (CNN) method for motion artifact reduction in fNIRS.
• To improve the accuracy of hemodynamic response function (HRF) estimation.
• To address limitations of conventional artifact suppression techniques.

Main Methods:

• Utilized a U-net architecture for the CNN.
• Generated large-scale training data by combining hemodynamic response function (HRF) variants with motion noise.
• Trained the CNN to reconstruct neuronal activity-related hemodynamic responses while reducing motion artifacts.

Main Results:

• The proposed CNN method demonstrated superior accuracy in estimating task-related HRF compared to wavelet decomposition and autoregressive models.
• CNN-based HRF estimates exhibited the lowest mean squared error and variance.
• Performance improvements were particularly notable with semi-simulated data featuring diverse HRF shapes and amplitudes.

Conclusions:

• The CNN method accurately estimates HRF amplitude and shape with significant motion artifact reduction.
• This technique shows promise for real-world fNIRS applications with substantial motion.
• The approach enhances the reliability of fNIRS for monitoring brain activity in dynamic environments.