Over the past 15 years, Near-Infrared Spectroscopy (NIRS) (Villringer et al., 1993
; Obrig and Villringer, 2003
; Gibson et al., 2005
; Hillman, 2007
; Hoshi, 2007
) has emerged as a complement to functional Magnetic Resonance Imaging (fMRI) for mapping the hemodynamic response associated with cerebral activity. NIRS non-invasively measures the temporal variations of the two dominant chromophores in the near-infrared window: oxygenated hemoglobin (HbO) and deoxygenated or reduced hemoglobin (HbR).
The advantages of NIRS for the investigation of brain activity include the measurement of both HbO and HbR concentrations, its low cost, and its portability. The portability of NIRS enables long-term monitoring of the hemodynamic reponse associated with, for instance, epilectic activity at the bedside (Roche-Labarbe et al., 2008
). Disadvantages of NIRS include modest spatial resolution of the order of one to three centimeters and limited penetration depth (Boas et al., 2004
A common problem with NIRS recordings is the presence of strong physiology-based systemic interference in the signal which reduces the accuracy of NIRS for detecting brain activation. This interference arises from cardiac activity, respiration and other homeostatic processes (Obrig et al., 2000
; Tonorov et al., 2000
; Diamond et al., 2009
; Payne et al., 2009
). The contribution of this interference in the NIRS signal is amplified because the light is both introduced and collected at the surface of the scalp. This back-reflection geometry makes NIRS very sensitive to the superficial layers of the head which contain no brain signal but exhibit strong systemic fluctuations. As such, the NIRS signal is often dominated by systemic interference occurring in the superficial layers of the head including the scalp and the skull.
Several methods have been described in the literature to remove the systemic interference from NIRS measurements. Some post-processing algorithms include low pass filtering (Franceschini et al., 2003
; Jasdzewski et al., 2003
), principal component analysis (Zhang et al., 2005
; Franceschini et al., 2006
) and wavelet filtering (Lina et al., 2008
; Matteau-Pelletier et al., 2009
; Jang et al., 2009
; Lina et al., 2010
). Multi-distance NIRS measurements with layered models and path length weighted methods have also being investigated (Umeyama and Yamada, 2009a
; Yamada et al., 2009
). Other methods include the subtraction of a NIRS channel acquired in a non-activated region of the brain from the signal of interest to reduce the systemic interference (Franceschini et al., 2003
A more refined version of this method is to simultaneously collect additional NIRS measurements using short source-detector (SD) separation channels (generally shorter than 1 cm), which are sensitive to superficial layers only (Saager and Berger, 2005
). Assuming that the signal collected with these additional short separation measurements is dominated by the same systemic interference present in the longer SD channels, the small separation signals can be used as regressors to filter the systemic interference from the longer SD measurements. Several algorithms have been developed to perform the regression of the small separation measurements. These include linear minimum mean square estimation (LMMSE) (Saager and Berger, 2005
; Gregg et al., 2010
; Saager et al., 2011
), adaptive filtering (Zhang et al., 2007a
) and state-space modeling with Kalman filter estimation (Gagnon et al., 2011
An important question which was not addressed in these previous papers is the impact that the relative location of the short and long SD channels has on the performance of the short separation method. If good performance is obtained using a short separation channel located far away from the standard long SD channel, then a single short separation channel can be used as a regressor for all longer SD channels on the probe. On the other hand, the performance of the short separation method potentially worsen as the relative distance between the short and the long SD channel increases. In this case several short separation channels would be required and only those closest to the long SD channels would be suitable for regression.
The main contribution of this paper is to quantify the performance of the short separation method as a function of the relative distance between long SD NIRS channels (3 cm) containing the brain signal and short separation (1 cm) channels used as regressors. We investigated this relationship with both simulations and real functional data. NIRS measurements including several short separation channels spread across the probe were acquired on 6 human subjects. The simulations were performed by adding a synthetic hemodynamic response to the resting-state NIRS data. NIRS signals were also collected during a series of finger tapping blocks for each of the 6 subjects. In both cases, the performance of the short separation regression was characterized for different short SD regressors located at different distances from the standard 3 cm channel.