Limitations in the signal quality of optical signals have hampered the acceptance of fNIRS in clinical practice and cognitive neuroscience research. To date, a number of groups have pursued varying techniques to regress or discriminate physiological noise from optically acquired functional brain signals. One potential method is to use direct peripheral measurements of the hypothesized noise sources (e.g., via a pulse-oximeter or respiration belt), which can then be regressed from the acquired data [as is commonly performed in fMRI (Glover et al., 2000
)]. This approach, however, is limited in the number of noise sources it can measure, and additionally it assumes a similarity between blood flow in the periphery and in the head that may not hold in practice. Our lab, as well as others, have used statistical techniques that assume orthogonality or independence between functional brain signals and physiological noise to identify these components (Morren et al., 2004
; Zhang et al., 2005
; Markham et al., 2009
). An additional method for regression is the use of adaptive filtering techniques (Zhang et al., 2007
), which have been used in single-source, multi-detector geometries.
In this paper, we have implemented and evaluated a theoretically and computationally simple regression procedure within the context of a high-density DOT imaging array. This method is amenable to real-time imaging and has been shown to be capable of CNR improvement in each hemoglobin species for ≥
80% of subjects, both in SD-pair data and in image reconstructions. The use of tomography individually is still helpful, but less impressive. This result might be due to the fact that we are only using first- and second-nearest neighbors in the present analysis. As there is substantial overlap in the sensitivities of these two measurement distances, we might predict that their ability to completely distinguish two depths to be limited. We expect that future DOT systems with sufficiently higher signal-to-noise to include further source–detector distances would have better depth profiling (Dehghani et al., 2009
). The combined effect of SSR and image reconstruction improves the CNR in all 10 datasets for HbO2
and HbT, and nine of 10 datasets for HbR. SSR and DOT have a synergistic effect to improve the depth-sectioning capabilities of optical imaging.
Additionally, we see that it is the subjects with the worst initial SNR that have the largest improvement after the use of SSR and DOT. This result is possibly because there are two contributions to noise: physiological noise and instrument noise. Since we expect instrument noise to be relatively constant between studies, the variance between subjects results predominantly from the amount of physiological noise present. Subjects with initially poor CNR (and thus high physiological noise) will benefit most from the noise removal techniques described here. Subjects with initially low physiological noise have CNRs limited by instrument noise. We would then expect the present techniques (which are designed to solely remove physiological noise) to have little effect on the CNR of these subjects.
It is important to note that what one refers to as “physiological noise” varies with the context of the experiment. When performing a functional activation study (as here), many normal physiological processes (including pulse, respiration, blood pressure oscillations, and spontaneous neural activity) all qualify as noise as they are undesired variance not related to your experimental paradigm. However, in other circumstances it could be the “noise” that you wish to measure. Resting-state functional connectivity (Biswal et al., 1995
; Fox et al., 2005
; Fox and Raichle, 2007
; White et al., 2009
) was originally dismissed as noise in fMRI signals. Additionally, one can extract important information through the examination of optical measurement of pulse and respiratory fluctuations (Wolf et al., 1997
; Franceschini et al., 2002
). The SSR method in this paper was designed for examination of task-evoked neural activity; care should be taken in choosing the appropriate noise removal method for each experiment.
While a linear regression should be less powerful than an adaptive filter, our results are more consistent and comprehensive than those previously presented for adaptive filters. Zhang et al. (2009
) found that with their filter “71% of the [HbO2
] measurements revealed CNR improvements after adaptive filtering, with a mean improvement of 60%. No CNR improvement was observed for [HbR].” In contrast, our improvement of about 200% in CNR is seen across all hemoglobin species. These results show that effective filtering can be obtained with simple, easily implemented algorithms. However, it also demonstrates that more research needs to be done on the nature of physiological interference in optical signals. We expect that greater knowledge of the sensitivities of different contrasts to physiological processes will yield even higher performing filter algorithms.
While one of the advantages of the proposed method is its ease of use, this simplicity does come at the cost of making assumptions that may not hold in all cases. One limitation is that there are likely several sources that contribute to the measured noise and that the linear combination present in the first-nearest neighbor pairs may not be the same as the linear combination in the second-nearest neighbors. In such a case, simple regression cannot remove all noise, but still has been shown to provide noise-reduction benefits (Saager and Berger, 2008
). An extension of the current technique would be a multiple linear regression method combined with direct measurements of other systemic signals (e.g., heart rate, breathing rate, and arterial blood pressure), which would, in principle, provide still better performance. A second limitation is that the averaged first-nearest neighbor signal might include some component sensitive to the brain. In the present adult study this effect is minimal since the brain sensitivity of first-nearest neighbors is low (<5%) and the activations were localized. However, in other applications, such as imaging infants, where the scalp and skull are much thinner and activations cannot be as easily localized, one would have to take care that the regression did not remove too much of the desired brain response. A third concern is that the superficial noise signal might vary over the surface of head. In this case, one could determine a local measure of the noise using nearby first-nearest neighbors. However, although future work will develop better algorithms, the present paper shows that current data can still be dramatically improved with simply implemented methods.
The average CNR improvement (with the use of both SSR and DOT) of approximately 2.5-fold in hemoglobin absorption data corresponds to a greater than six-fold reduction in the necessary acquisition time required for an equivalent signal quality (assuming that CNR scales as the square root of the total scan time). Decreased acquisition time will have two complementary effects that aid the use of fNIRS in novel environments, such as with young children and hospitalized patients. With increased signal quality, one needs to rely less on the long-term cooperation of the subject and can obtain useful data even if only short scanning windows are available. Conversely, for a fixed session length, decreased time per stimulus allows the use of multiple stimulus paradigms and attempts to decode more complex brain functionality.
We have shown that high-density DOT's overlapping, depth-dependent measurements of the head and brain can be leveraged to perform SSR in addition to three-dimensional image reconstructions. Both of these methods reduce the effect of physiological noise in the acquired neuroimaging data. The extension of SSR to DOT data provides an additional benefit beyond DOT's inherent depth-sectioning capabilities. Improvements in repeatability and signal quality advance DOT towards real-time imaging and greater utility in basic neuroscience and clinical care.