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

Abstract

Significance: Functional near-infrared spectroscopy (fNIRS), a well-established neuroimaging technique, enables monitoring cortical activation while subjects are unconstrained. However, motion artifact is a common type of noise that can hamper the interpretation of fNIRS data. Current methods that have been proposed to mitigate motion artifacts in fNIRS data are still dependent on expert-based knowledge and the post hoc tuning of parameters. Aim: Here, we report a deep learning method that aims at motion artifact removal from fNIRS data while being assumption free. To the best of our knowledge, this is the first investigation to report on the use of a denoising autoencoder (DAE) architecture for motion artifact removal. Approach: To facilitate the training of this deep learning architecture, we (i) designed a specific loss function and (ii) generated data to mimic the properties of recorded fNIRS sequences. Results: The DAE model outperformed conventional methods in lowering residual motion artifacts, decreasing mean squared error, and increasing computational efficiency. Conclusion: Overall, this work demonstrates the potential of deep learning models for accurate and fast motion artifact removal in fNIRS data.

Keywords: deep learning; denoising autoencoder; functional near-infrared spectroscopy; motion artifact.

Fig. 1

Illustration of the fNIRS data simulation process and the designed DAE model. (a) The green lines are the experimental fNIRS data, including noisy HRF and resting fNIRS data, while the blue and red lines are simulated ones. The AR models are fitted to the experimental resting-state fNIRS time series data, based on whose parameters the simulated resting fNIRS data are generated. The HRFs are simulated from gamma functions. The shift and spike noise are simulated based on the same distribution of the parameters from the experimental HRF. The simulated noisy HRF data (black line) is the sum of the simulated HRF, the shift noise, the spike noise, and the resting-state fNIRS. (b) DAE model: the input data of the DAE model are the simulated noisy HRF, and the output is the corresponding clean HRF without noise. The DAE model incorporates nine convolutional layers, followed by max-pooling layers in the first four layers and upsampling layers in the next four layers, with one convolutional layer before the output. The parameters are labeled in parentheses for each convolutional layer, in the order of kernel size, stride, input channel size, and kernel number.

Fig. 2

An example of the fNIRS data simulation process and the designed DAE model. (a) An example of experimental data. (b) The model artifact extracted from the experimental data in (a). (c) The resting state period extracted from the experimental data in (a). (d) Simulated evoked responses. (e) Motion artifact data simulated based on the parameters extracted from the motion artifact in (b). (f) Resting state data simulated based on the data in (c). (g) Synthetic noised HRFs, which is the sum of data in (d)–(f). (h) The expected output of DAE model, which is the same with the data in (d).

Fig. 3

The denoising results. (a) The number of residual motion artifacts for the simulated testing dataset. (b) The number of residual motion artifacts for experimental data. (c), (f) An example of processed data by different models in the simulated dataset, (d), (g) in experimental data under “No act.” condition, and (e), (h) under “Act.” condition. “No correction” indicates that no motion artifact correction model was used. An enlarged two-dimensional (2D) view of (c)–(h) is in Fig. S4 in the Supplemental Material.

Fig. 4

The denoising results in the new dataset. (a) The number of residual motion artifacts for experimental data. (b), (c) An example of processed data by different models. (d), (e) An enlarged 2D view of (b) and (c).