Elsevier

NeuroImage

Volume 46, Issue 1, 15 May 2009, Pages 133-143
NeuroImage

Real-time imaging of human brain function by near-infrared spectroscopy using an adaptive general linear model

https://doi.org/10.1016/j.neuroimage.2009.01.033Get rights and content

Abstract

Near-infrared spectroscopy is a non-invasive neuroimaging method which uses light to measure changes in cerebral blood oxygenation associated with brain activity. In this work, we demonstrate the ability to record and analyze images of brain activity in real-time using a 16-channel continuous wave optical NIRS system. We propose a novel real-time analysis framework using an adaptive Kalman filter and a state–space model based on a canonical general linear model of brain activity. We show that our adaptive model has the ability to estimate single-trial brain activity events as we apply this method to track and classify experimental data acquired during an alternating bilateral self-paced finger tapping task.

Introduction

Over the last few decades, functional near-infrared spectroscopy (NIRS) has been attracting interest from the psychology and medical imaging communities for its ability to non-invasively measure the cerebral hemodynamic changes associated with functional brain activity (Boas et al., 2004, Cannestra et al., 1998, Delpy et al., 1987, Villringer and Chance, 1997). NIRS is an optical spectroscopy technique which uses time-resolved, multi-wavelength measurements to infer changes in the optical absorption of tissue and thereby to report changes in oxy- and deoxy-hemoglobin, the two primary absorbing chromophores in biological tissue that vary dynamically with a functional task. Similar to functional magnetic resonance imaging (fMRI), NIRS-based brain studies can be used to record changes in cerebral blood oxygenation during the performance of a specific activation task. This technique has been applied to the study of numerous brain systems and several clinical conditions (see Obrig and Villringer, 2003 for review).

Traditionally, brain imaging studies have generally been performed in two-stages; collection of the experimental data followed by an off-line analysis of the signals and interpretation of brain activity. In recent years, however, the combination of more powerful and faster computers, improved data analysis schemes, and a desire to develop experimental designs using biofeedback given to the test subject, has motivated attempts by several groups to develop real-time data analysis schemes to assess brain activation at the same time as data collection. To date, real-time brain imaging has been performed with NIRS (Coyle et al., 2007, Sitaram et al., 2007), EEG (Mason et al., 2004, Funase et al., 2006, Muller et al., 2008), and fMRI modalities (LaConte et al., 2007, Weiskopf et al., 2007, Christopher Decharms, 2008). The ability to provide real-time estimates of brain activity has become increasingly attractive to researchers — offering the potential to apply these methods towards applications in the development of brain–computer interfaces (Weiskopf et al., 2004), the study of complex subject–interaction tasks such as the field of neuroeconomics (Sanfey et al., 2003, Singer, 2007, Walter et al., 2005), and potential clinical assessment tools capable of providing immediate feedback to guide medical and behavioral therapies (Chute, 2002, Hintz et al., 2001, Shingu et al., 2003). It is believed that these new neuroimaging tools may one day have the possibility to offer cerebral-based diagnostics in a variety of clinical examinations to enhance existing qualitative, observational, or behavioral assessments (Chute, 2002, Christopher Decharms, 2008, Weiskopf et al., 2007).

Over the last several years, real-time functional MRI, in particular, has been used for a growing number of applications (reviewed in Weiskopf et al., 2007, Christopher Decharms, 2008). However, some of the drawbacks of fMRI-based approaches are the cost and restrictions of the fMRI scanner, which limit its usability to lab facilities. In comparison to fMRI, generally, NIRS instrumentation is smaller, more portable, and less expensive to purchase and maintain. For these reasons, NIRS techniques are ideally suited for the development of portable and potentially clinical real-time systems.

In comparison to the more standardized and traditional off-line analysis approaches for characterizing evoked hemodynamic signals (for example, see Friston, 2007), real-time brain imaging requires several modifications which have been discussed in (Coyle et al., 2004a, Coyle et al., 2004b, Matthews et al., 2008, Weiskopf et al., 2007). In particular, not only do these computations need to occur quickly and efficiently, which in general is now fairly straightforward with moderate-grade computers, but more importantly many of the quality assurance tests performed on data need to be automated, including detection, rejection, and correction of motion artifacts, removal of systemic physiological signals, and automated thresholding for visualization and secondary analysis of activation images. In this regard, real-time NIRS brain imaging methods pose several additional and unique challenges in comparison to fMRI, as was recently discussed in Matthews et al. (2008). One of the challenges in the analysis of optical signals is its sensitivity to superficial and global physiological changes. Linear time-series methods, such as bandpass filtering, can be used to remove high-frequency cardiac signals and low-frequency drifts in the optical signal (e.g. Boas et al., 2004) and can be implemented as analog or digital real-time filters. However, bandpass filters are not effective at removing several specific physiological noise signals such as respiratory and blood pressure (e.g. Mayer wave) fluctuations (Franceschini et al., 2006, Toronov et al., 2000) since these fluctuations are difficult to distinguish from typical brain activations by frequency signatures alone. Additionally, bandpass filters cannot remove motion artifacts (e.g. sudden, large, non-physiological changes in the signal due to probe movement relative to the head). As reviewed in Boas et al. (2004), many of the existing sophisticated analysis methods that have been proposed for removing this systemic physiological noise, are based on retrospective correction of signal using partial linear regression of independently measured physiology either from an external monitor or from local recordings of superficial responses by additional short-distance NIRS measurements (e.g. (Diamond et al., 2005, Diamond et al., 2006, Saager and Berger, 2005); also see RETROICOR methods developed for fMRI (Deckers et al., 2006, Glover et al., 2000)). Alternatively, methods to filter systemic effects by reducing spatial covariance due to the physiology using principal or independent component analysis have been proposed (Franceschini et al., 2006). While principal component analysis has been used to remove these additional sources of physiological noise (Zhang et al., 2005a, Franceschini et al., 2006), as well as for removal of motion artifacts (Wilcox et al., 2005), these algorithms are based on the projection of the spatial covariance structures from resting-state signals and, thus, require training data, for example using a separate baseline data file to determine the physiological components. Principal component methods also have the limitation of not being able to attribute each component to a specific physiological signal prior to the start of the study since the relative contributions of the major physiological signals (blood pressure, respiration, and cardiac) may vary such that it is difficult to know how many and which principal components need to be removed. Thus, these methods are difficult to implement in a real-time analysis scheme. More recently, Zhang et al., 2007a, Zhang et al., 2007b) described the potential to use adaptive filters to remove physiology from NIRS signals. Although in their paper adaptive filtering was used as a post-processing step in analysis of their functional signals, they expressly point out that an advantage of this approach is the potential of its application to real-time analysis.

In our current paper, we present an online data processing scheme using a similar adaptive filter for the purpose of estimating the cerebral brain activity in real time from NIRS. Our approach is based on a time-variant (or adaptive) version of the canonical general linear model (GLM) of brain activity. The canonical general linear model was introduced by Friston et al. (1994) and has extensively been used in analysis of functional MRI (e.g. in the software SPM; statistical parametric mapping (Friston, 2007); and AFNI, analysis of functional neuroimages (Cox, 1996) among others). Recently, several optical studies have also begun to adapt this approach (Cohen-Adad et al., 2007, Diamond et al., 2005, Zhang et al., 2005a, Schroeter et al., 2004). Unlike the conventional linear time-invariant regression analysis used in these earlier works, in our model we propose an adaptive algorithm based on the Kalman filter (Welch and Bishop, 1995) to estimate the coefficients of this general linear model for each individual time-point. In this way, our adaptive modeling approach provides a real-time estimate of brain activation on a single-trial basis and allows us to take advantage of several features of the Kalman filter that are beneficial to analysis of optical signals. In particular, a Kalman filter weighs the current estimation of brain response between a prediction of the model based on the history of evoked responses from all previous trials and new information from the current data point. After each new data point is made available, this weight it reevaluated. This provides an optimal estimate of brain activity and a means of tracking inter-trial changes in the evoked response. In comparison to the fixed window averaging methods, which evaluate real-time signals by averaging data from a sliding window covering the last several trials, approaches which have been used in the majority of real-time imaging applications to date (e.g. Nakai et al., 2006), we believe that our adaptive filtering method provides a more flexible means to perform real-time imaging by incorporating prior knowledge about the expected reproducibility of the evoked response within the prior estimates of the processes noise of the Kalman filter. In comparison to the initial adaptive model proposed in Zhang et al., 2007a, Zhang et al., 2007b, which used an auto-regressive model as the basis of the state prediction, we use a framework similar to the canonical general linear model which allows us to introduce both knowledge of the timing of stimulus events and a prior expectation of the shape of the evoked response. Note also that our adaptive version of the standard general linear model allows for more flexibility in the temporal dynamics of the hemodynamic response of oxy- and deoxy-hemoglobin in comparison to the static canonical temporal model, since the coefficients multiplying this canonical function (typically denoted β) can also vary in time. In this way, the canonical hemodynamic response is used as a prior on the shape of the evoked response, rather than a strict canonical assumption.

To account for the physiological noise specific to the optical data, we have introduced additional regressors into our adaptive model based on the time-varying physiological model developed by Prince et al. (2003). In addition to modeling the canonical functional response, we include a trigonometric series with adaptive amplitude and phase components in order to model the specific physiological noise contributions from the respiratory, cardiac, and blood pressure (Mayer) wave signals. This model allows a tunable bandwidth to remove these systemic physiological signals.

In this paper, we described the real-time implementation of our adaptive model using an extended Kalman filtering algorithm. In order to demonstrate the application of our model, we show data from a self-paced motor activation in which a NIRS probe was placed over the left and right motor areas. During this study, the subject tapped the fingers of one of their two hands and our real-time NIRS system was used to estimate and display the brain activation signal, as well as to classify the handedness of the tapping per trial and to feedback the results to the individual.

Section snippets

Near-infrared brain imaging

The NIRS technique is a non-invasive method used to measure functional brain signals (Jobsis, 1977, Wray et al., 1988, Delpy et al., 1988). NIRS instruments use low levels of light (typically between 5–10 mW) within the wavelength region of 650–850 nm light to record changes in the optical absorption of tissues over time. During functional brain activation, increases in blood flow, volume and oxygenation result in changes in absorption due to underlying changes in the concentration of oxy- and

Experimental design and subject characteristics

In order to demonstrate the effectiveness of our adaptive linear model and real-time system, optical studies were performed on three healthy individuals. All subjects gave voluntary consent and this study was approved by a local institutional review board at the University of Pittsburgh Medical Center. The subjects' ages ranged from 26 to 43 years old and all three were male. Two subjects were right-handed, while one subject is left-handed (self-reported, see Table 1). During the study,

Results

Using the Cw6 optical imaging system with our customized control and analysis software, we were able to detect and visualize brain activation in real-time using NIRS at data rates of 4 Hz for the full compliment of up to 256 source-detector combinations. A total of 56 combinations were used in the probe for this study. This approach allows spatially resolved imaging of the brain signals. In order to demonstrate this system, we recorded NIRS signals from the bilateral motor cortex regions during

Discussion

We have described an adaptive version of the general linear model, which can be used for real-time assessment of brain function. To demonstrate this model, we have performed a motor-task activation study, consisting of an alternately left and right handed self-paced finger tapping tasks. This study allows us to highlight several important features of our model. First, the adaptive model provides a means to perform single-trial estimation of the evoked brain response. We have chosen this

Limitations of adaptive model

While our adaptive GLM approach offers these several advantages, adaptive filtering, in general, is not without some limitations. In the Kalman filtering model, prior distributions must be assumed for both the process and observation noise (Q and R respectively). These two distributions determine the behavior of the Kalman tracker and, in particular, how closely the model will follow data and how sensitive the model is to noise. This tuning is done prior to the experiment and requires some

Future extensions

The focus of this paper was the development of the adaptive general linear model capable of providing real-time estimates of brain activity through tracking of the states beta, which multiplied the canonical functional response profiles. In this current work, the states were input into a simple linear classifier to determine left versus right-handed self-paced finger tapping. While this classifier scheme performed well for this simple task and for demonstrating the adaptive linear more, more

Conclusions

NIRS is a neuroimaging modality with appealing properties as a relatively cheap and compact fashion of obtaining functional brain images at a high temporal resolution. In this work, we have introduced and demonstrated a real-time algorithm for tracking functional brain signals using a multi-channel NIRS system. Our algorithm is based on an adaptive version of the general linear model. This model allows both single-trial and real-time tracking of brain signals and used extended Kalman filter. As

Acknowledgments

This work was supported by department funds from the University of Pittsburgh's department of radiology.

References (56)

  • VillringerA. et al.

    Non-invasive optical spectroscopy and imaging of human brain function

    Trends Neurosci.

    (1997)
  • WalterH. et al.

    Motivating forces of human actions. neuroimaging reward and social interaction

    Brain Res. Bull.

    (2005)
  • WangL. et al.

    MCML—Monte Carlo modeling of light transport in multi-layered tissues

    Comput. Methods Programs Biomed.

    (1995)
  • WeiskopfN. et al.

    Real-time functional magnetic resonance imaging: methods and applications

    Magn. Reson. Imaging

    (2007)
  • WrayS. et al.

    Characterization of the near infrared absorption spectra of cytochrome aa3 and haemoglobin for the non-invasive monitoring of cerebral oxygenation

    Biochim. Biophys. Acta

    (1988)
  • BoasD.A. et al.

    Improving the diffuse optical imaging spatial resolution of the cerebral hemodynamic response to brain activation in humans

    Opt. Lett.

    (2004)
  • CannestraA.F. et al.

    Refractory periods observed by intrinsic signal and fluorescent dye imaging

    J. Neurophysiol.

    (1998)
  • Christopher DecharmsR.

    Applications of real-time fMRI

    Nat. Rev. Neurosci.

    (2008)
  • ChuteD.L.

    Neuropsychological technologies in rehabilitation

    J. Head Trauma. Rehabil.

    (2002)
  • CopeM. et al.

    Methods of quantitating cerebral near infrared spectroscopy data

    Adv. Exp. Med. Biol.

    (1988)
  • CoyleS. et al.

    Physiological noise in near-infrared spectroscopy: implications for optical brain computer interfacing

    Conf. Proc. IEEE Eng. Med. Biol. Soc.

    (2004)
  • CoyleS. et al.

    On the suitability of near-infrared (NIR) systems for next-generation brain–computer interfaces

    Physiol. Meas.

    (2004)
  • CoyleS.M. et al.

    Brain–computer interface using a simplified functional near-infrared spectroscopy system

    J. Neural. Eng.

    (2007)
  • DelpyD.T. et al.

    Cerebral monitoring in newborn infants by magnetic resonance and near infrared spectroscopy

    Scand. J. Clin. Lab. Invest. Suppl.

    (1987)
  • DelpyD.T. et al.

    Estimation of optical pathlength through tissue from direct time of flight measurement

    Phys. Med. Biol.

    (1988)
  • DiamondS.G. et al.

    Physiological system identification with the Kalman filter in diffuse optical tomography

    Med. Image Comput. Comput. Assist. Interv. Int. Conf. Med. Image Comput. Comput. Assist. Interv.

    (2005)
  • FranceschiniM.A. et al.

    Diffuse optical imaging of the whole head

    J. Biomed. Opt.

    (2006)
  • FristonK.J.

    Statistical Parametric Mapping : the Analysis of Functional Brain Images

    (2007)
  • Cited by (172)

    • Tracking differential activation of primary and supplementary motor cortex across timing tasks: An fNIRS validation study

      2020, Journal of Neuroscience Methods
      Citation Excerpt :

      fNIRS provides an alternative to fMRI for assessing hemodynamics specific to the cortex. The NIRS method is based on changes in near-infrared light absorption reflecting corresponding concentration changes in oxygenated (oxy-Hb) and deoxygenated (deoxy-Hb) hemoglobin in the cortex (Abdelnour and Huppert, 2009; Jahani et al., 2015; Pollonini et al., 2014; Rahimpour, Dadashi, Soltanian-Zadeh, et al., 2017; Rahimpour, Noubari, et al., 2018). We used a two-wavelength fNIRS system (NIRScout, NIRx, Glen Head, NY) with 16 light-incident emitters; 20 light-detecting fibers, and a sampling rate of 6.25 Hz.

    View all citing articles on Scopus
    View full text