Computer Vision-Driven Movement Annotations to Advance fNIRS Pre-Processing Algorithms

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Vision

Background:

• Functional near-infrared spectroscopy (fNIRS) enables brain activity studies in naturalistic settings due to movement tolerance.
• Motion artifacts in fNIRS data can compromise results, necessitating effective correction algorithms.
• Evaluating fNIRS motion artifact correction is hindered by the lack of reliable ground truth data for head movements.

Purpose of the Study:

• To investigate the feasibility and reliability of a deep learning computer vision (CV) approach for automated head movement detection and annotation from video recordings.
• To establish a reliable ground truth for head movements to aid in the development and evaluation of fNIRS motion artifact correction methods.

Main Methods:

• Fifteen participants performed controlled head movements across rotational axes at varying speeds and types.
• Video recordings captured movements; head orientation signals were extracted using SynergyNet.
• A 1-dimensional UNet (1D-UNet) model processed orientation signals for movement detection, with manual annotations serving as ground truth.

Main Results:

• The CV model demonstrated strong performance in detecting head movements, with Jaccard indices of 0.954 (training) and 0.865 (test).
• Consistent performance was observed across different movement axes and speeds.
• Performance varied by movement type: repeated movements (J=0.941) > complete movements (J=0.872) > half movements (J=0.826).

Conclusions:

• The proposed CV approach provides accurate and reliable ground truth information for head movements.
• This automated method can significantly aid researchers in evaluating and improving fNIRS motion artifact correction algorithms.
• Future research can leverage this CV technique to enhance the quality and validity of fNIRS studies.