Motion artifacts in functional near-infrared spectroscopy: a comparison of motion correction techniques applied to real cognitive data

Abstract

Motion artifacts are a significant source of noise in many functional near-infrared spectroscopy (fNIRS) experiments. Despite this, there is no well-established method for their removal. Instead, functional trials of fNIRS data containing a motion artifact are often rejected completely. However, in most experimental circumstances the number of trials is limited, and multiple motion artifacts are common, particularly in challenging populations. Many methods have been proposed recently to correct for motion artifacts, including principle component analysis, spline interpolation, Kalman filtering, wavelet filtering and correlation-based signal improvement. The performance of different techniques has been often compared in simulations, but only rarely has it been assessed on real functional data. Here, we compare the performance of these motion correction techniques on real functional data acquired during a cognitive task, which required the participant to speak aloud, leading to a low-frequency, low-amplitude motion artifact that is correlated with the hemodynamic response. To compare the efficacy of these methods, objective metrics related to the physiology of the hemodynamic response have been derived. Our results show that it is always better to correct for motion artifacts than reject trials, and that wavelet filtering is the most effective approach to correcting this type of artifact, reducing the area under the curve where the artifact is present in 93% of the cases. Our results therefore support previous studies that have shown wavelet filtering to be the most promising and powerful technique for the correction of motion artifacts in fNIRS data. The analyses performed here can serve as a guide for others to objectively test the impact of different motion correction algorithms and therefore select the most appropriate for the analysis of their own fNIRS experiment.

Keywords: Functional near-infrared spectroscopy; Hemodynamic response; Motion artifact; Motion correction; fNIRS.

Fig.1

a) Probe placement: detectors in blue and sources in yellow. Numbers represent the channels. b) Example of motion artifacts present in the time-series of one participant. The blue line shows the 830 nm wavelength time-series, the red line the 690 nm wavelength time-series. The 690 nm time-series has been shifted by -0.3 for visualization purposes. Vertical lines indicate when the stimulus is presented to the participant. Note how the motion artifact is correlated to the task.

Fig.2

Signal processing steps for all techniques. The processing streams for every technique are represented by colored arrows: black for rejection, blue for no motion correction, red for PCA_80, cyan for Spline, green for Wavelet, magenta for CBSI, orange for Kalman filter and grey for PCA_97.

Fig.3

Box plots of the AUC_0-2, AUC ratio and within-subject SD computed for all techniques and for both HbO (upper row) and HbR (bottom row). The red line in the box plot indicates the median, while the two extremities of the box plot represent the first and third quartile. Red crosses indicate outliers. The lines above linking the different techniques represent the significant statistical difference (p<.05 if the line is blue, p<.01 if the line is red).

Fig.4

Bar plots with the mean number of trials averaged for each technique normalized to the mean number of trials averaged with the no motion correction technique; the error bars represents the standard deviation. The lines above indicate whether the techniques that they link together differ significantly from each other (p<.05 if blue, p<.01 if red).

Fig.5

Examples of recovered mean HRFs for four selected subjects, channels and tasks for every technique for both HbO (solid line) and HbR (dashed line). HbR HRFs have been shifted in baseline towards negative values for visualization purposes only. In a) Wavelet, and CBSI provide some minimization of the motion artifact, while PCA_80 increases it. In b) all techniques but PCA_97 are able to recover physiological HRFs, no motion correction included; PCA_97 highly underestimates the HRF. c) is an example of PCA_80 and PCA_97 adding a motion artifact in a motion-free channel and d) is an example of a channel in one subject where the Kalman filter is unstable. Gray line represents the actual task duration, 850 ms, which is the grand average of the reaction times, i.e. the time needed by participants between the appearance of the word and the color being pronounced.

Fig.6

a) and b) Scatter plots of the AUC_0-2 metric computed with the rejection technique (y axis) vs. that computed with no motion correction technique (x axis) for both HbO (a) and HbR (b). Trial rejection decreases AUC_0-2 36% of the time for HbO and 37% for HbR, but increases it in almost the same percentage of cases. 28% of the times the AUC_0-2 value is identical for both techniques. c) and d) Scatter plots of the within-subject SD metric computed with the rejection technique (y axis) vs. that computed with no motion correction technique (x axis) for both HbO (c) and HbR (d). Trial rejection decreases the standard deviation 54% of the time for HbO and 51% for HbR compared to no motion correction.

Fig.7

a,b: Scatter-plots of the AUC_0-2 metric for both HbO (a) and HbR (b): no correction (x axis) vs. Wavelet, CBSI, Kalman, Spline, PCA_80 and PCA_97 (y axis).

Fig.7

a,b: Scatter-plots of the AUC_0-2 metric for both HbO (a) and HbR (b): no correction (x axis) vs. Wavelet, CBSI, Kalman, Spline, PCA_80 and PCA_97 (y axis).

Fig.8

a,b: Scatter-plots of the within-subject SD metric for both HbO (a) and HbR (b): no correction (x axis) vs. Wavelet, CBSI, Kalman, Spline, PCA_80 and PCA_97 (y axis).

Fig.8

a,b: Scatter-plots of the within-subject SD metric for both HbO (a) and HbR (b): no correction (x axis) vs. Wavelet, CBSI, Kalman, Spline, PCA_80 and PCA_97 (y axis).

Fig.9

a,b: Scatter-plots of the between-subject SD metric for both HbO (a) and HbR (b): no correction (x axis) vs. Wavelet, CBSI, Kalman, Spline, PCA_80 and PCA_97 (y axis).

Fig.9

a,b: Scatter-plots of the between-subject SD metric for both HbO (a) and HbR (b): no correction (x axis) vs. Wavelet, CBSI, Kalman, Spline, PCA_80 and PCA_97 (y axis).