2.1.1.

High-density diffuse optical tomography (HD-DOT)

High-density diffuse optical tomography (HD-DOT) is an emerging approach for functional neuromonitoring that uses denser optode arrays than traditional CW-NIRS systems.^2–4 HD-DOT provides numerous advantages over standard NIRS, including dramatically improved image quality, resolution, localization, and brain specificity. The density of the imaging arrays provides many crisscrossing measurements that enable tomography algorithms to aggregate measurements and partially deblur the diffuse propagation of signals through tissues. Further, different distances offer a limited ability for depth sectioning and flatter sensitivity to deeper depths (up to 2 cm from the scalp surface). When the imaging arrays are coupled with anatomically derived head models, the imaging becomes more compatible with fMRI style data analysis.⁵ When matched within-subjects against functional MRI (fMRI), HD-DOT can obtain localization errors $<5\mathrm{mm}$, and resolution $<15\mathrm{mm}$ full width half maximum (FWHM), sufficient to localize functions to gyri.⁵ Feasibility studies have established HD-DOT for mapping sensory networks (visual and motor) as well as distributed cognitive networks, including the frontal, parietal, and default mode networks.⁴ A diverse set of functional neuroimaging paradigms has also been illustrated with HD-DOT, including traditional tasks,² resting-state functional connectivity,^6,7 naturalistic movie mapping,⁸ and most recently decoding studies.⁹

One of the main limitations of HD-DOT has been the mass of the fibers used in fiber-based HD-DOT systems. While fiber-based systems have thus far set the standard regarding specifications^2,6,10 (including detectivity, dynamic range, crosstalk, frame rate, optode-scalp coupling, and modulation/demodulation strategies for encoding source illuminations⁴), several research groups have been demonstrating fiberless implementations of HD-DOT.^11–15 Although technical issues are still to be worked out, we expect HD-DOT systems to transition to mostly fiberless implementations over the next five years, removing a main barrier to their use instead of sparse NIRS devices.

2.1.2.

Wearable CW-NIRS technology

Wearable CW-NIRS systems use lightweight and fiberless sources and detectors placed directly on the scalp, allowing for data acquisition in more ecologically valid environments.^14,16 LEDs are the most widely used light sources, while APDs and silicon photodiodes are commonly used detectors in wearable CW-NIRS systems. Newer detectors, such as single-photon avalanche diode arrays (SPADs)¹⁷ and silicon photomultipliers, promise higher sensitivity than photodiodes.¹⁸ Recent wearable systems integrate digital conversion within the optodes, thereby eliminating the need for electrical cables to transfer the signal and making the design even more lightweight and compact.^15,19 Moreover, a myriad of modular designs has been proposed that are flexible and adaptable to user needs.^15,19,20

In the coming years, we anticipate that wearable systems will enable whole-head, high-density fNIRS measurements in everyday settings.¹¹ In such settings, protection from ambient light and other environmental factors, robustness to motion artifacts, integration with multimodal measurements (e.g., eye tracking, motion sensors, visual/auditory input), and the synchronization of these additional measurements with the fNIRS signal will become more critical. Moreover, reliable data transmission and online detector sensitivity/signal quality control will be necessary. Early steps have been taken towards this end to extend Lab Streaming Layer (LSL) to support CW-NIRS data.²¹ LSL allows efficient and correct real-time transfer of data between computers and is commonly used in the brain-computer interface community. Its integration may also be used for broader application, such as real-time data quality management.

2.5.1.

High-density single photon avalanche diode (SPAD) array detection

Recent work has incorporated high-density SPAD arrays for parallelized DCS detection of over a thousand individual speckle modes at close to megahertz sampling rates.^64–66 By treating each SPAD pixel within an array of $N$ pixels as an independent decorrelation rate detector and averaging the result across all pixels, SPAD array-based parallelization can improve the SNR of DCS by a factor of the square root of $N$.^65–67 SPAD arrays have recently enabled the detection of mm-scale perturbations buried up to 1 cm deep underneath tissue-like phantoms at up to 33 Hz, as well as monitoring of blood flow from above the prefrontal cortex at up to 30 ms temporal resolution to resolve ventricle contraction and repolarization within the pulse.⁶⁵ Parallelized DCS systems have also been piloted in monitoring CBF in humans during tasks expected to elicit prefrontal cortex activation.⁶⁵ Apart from directly averaging the DCS signal across all SPAD pixels within an array, the signal from individual SPAD pixels may also be digitally combined and processed in alternative configurations, for example, to form images and video of deep-tissue dynamics.⁶⁸

Currently, one main limitation of SPAD array-based DCS detection is its relatively low per-SPAD temporal sampling rate. While a single SPAD can sample at over 100 MHz,⁶⁹ SPAD arrays currently operate at approximately 1 MHz frame rates. This acquisition rate currently limits their application for detecting rapid speckle fluctuations arising from tissue depths necessary to sense the brain (sub-ns to $\mu \mathrm{s}$). A second current limitation of SPAD arrays vs. single SPADs is a lower photon detection efficiency, which we anticipate will increase as integrated solid-state CMOS detector fabrication processes improve. In addition, demonstrations of DCS with SPAD arrays to date have utilized just slightly over one thousand pixels, compared with the millions of pixels currently available in a standard CMOS or CCD pixel array. Luckily, high-density SPAD arrays with up to one million pixels have recently emerged.^70,71 In the future, we can likely expect the speed, sensitivity, and pixel counts of such novel integrated arrays to follow the general trends of improvement that are well-known within the semiconductor industry to benefit upcoming DCS platforms continually.

2.5.2.

DCS at 1064 nm

Traditionally, DCS has employed source illumination in the NIR between 785–852 nm. Recent work has investigated the use of longer source wavelengths for DCS. Between 1050 and 1100 nm, water has a local absorption minimum, offering an additional transmission window for deep tissue measurements, and tissue scattering is progressively reduced. Although this spectral window is less useful for NIRS because hemoglobin absorption is very low, DCS is not impacted as it relies on the dynamic scattering of light by red blood cells, which remains substantial in this wavelength range. The 1064 nm wavelength represents a favorable operating point due to the wide availability of laser sources, fiber amplifiers, and other optical components first developed for the communication industry.⁷²

One of the benefits of 1064 nm operation is that the effective attenuation coefficient (${\mu }_{\mathrm{eff}}(\lambda )\equiv \sqrt{3{\mu }_a(\lambda ){\mu }_s^{\prime }(\lambda )}$) is 10–20% lower than in the 650–850 nm range, offering improved light penetration and brain sensitivity. In addition, ANSI standard limits for safe skin exposure (ANSI Z136.1) allow 2.7–3.7 times more energy to be delivered per unit area than 785–852 nm. Further, because photons carry less energy at longer wavelengths, $\sim 40\%$ more photons are expected per unit of energy at 1064 nm compared to 785 nm. Finally, the intensity autocorrelation decay is slower at 1064 nm due to the decreased scattering, resulting in an increased SNR. Altogether, these factors have been shown to result in 10–20 times more photons being detected at 1064 nm compared to 785-852 nm, corresponding to 10–20 times SNR gain for DCS measurements at 1064 nm. This SNR advantage can be traded for either increased temporal resolution or increased source-detector separation. The 1064 nm operation can offer more than 1 cm extension of the maximum source-detector separation at the same SNR when compared to 785–852 nm DCS measurements, making 3–3.5 cm separations practical and resulting in a 2–3 times increase in the ratio of cerebral to extracerebral blood flow sensitivity.

The main challenge for DCS at 1064 nm is the lack of suitable semiconductor detectors. Current commercially available silicon SPADs have very low efficiency, while indium gallium arsenide (InGaAs) SPADs have excessive after-pulsing. Thus, the full potential of DCS at 1064 nm has so far been demonstrated only using cryo-cooled superconducting nanowire single-photon counting detector (SNSPD) devices that have limitations due to cost, size, and noise.⁷³ Novel gated InGaAs detector arrays are currently under development to be used with a pulsed laser source. These are expected to offer sufficient channels to implement cross-correlation detection to remove after-pulsing effects and will be able to power large functional DCS imaging probes.

2.5.3.

Time-domain DCS (TD-DCS)

The development of time-domain DCS (TD-DCS)⁷⁴ has enabled the selection of photons with long path lengths by their time-of-flight to minimize the influence of extracerebral layers on the measured blood flow index. This approach can be applied even at short separations where more detected photons are available—the same principle behind the “null-separation” TD-NIRS approach.⁷⁵ TD-DCS systems comprise a picosecond pulsed laser ($\sim 300\mathrm{ps}$ FWHM pulses appear to offer the best balance between coherence and the ability to resolve deep traveling photons) and time-correlated or gated single-photon counting detectors that allow the selection of photons arriving in a specific time window (gate) after the laser pulse. The intensity autocorrelation is computed using only photons arriving “in the gate” over the same time intervals typically used for DCS (sub-ns to ms). TD-DCS has been piloted in measuring the response to physiological manipulations and preliminary assessment of functional brain responses.^69,76,77 A key advantage of TD-DCS is the MHz laser repetition rates used, which allows for the implementation of dense optode arrays without crosstalk, as multiple illumination states can be employed without any loss of temporal resolution or SNR. Currently, the main limitation of TD-DCS is the lack of efficient detectors without a “diffusion tail” affecting their timing response, as well as the higher cost of components versus traditional CW-DCS systems.

In the near future, we expect TD-DCS to benefit from long-wavelength operation at 1064 nm and the development of new detector technologies. On the one hand, SNSPDs have recently emerged as the ultimate detector technology with very high photon detection efficiencies across broad wavelength ranges, nearly negligible dark counts, and no after-pulsing.⁷³ However, as mentioned in the previous section, SNSPDs are expensive, large, and loud, though miniaturization efforts are ongoing. At the same time, SPAD cameras with thousand element and larger detector arrays are reaching higher performance levels in terms of detection efficiency and timing performance.⁷⁸ These improvements will permit the practical implementation of TD-DCS imaging arrays for functional brain measurements across the whole head.

2.5.4.

Interferometric diffusing wave spectroscopy (iDWS)

Recently, interferometric optical methods have also been investigated for measuring diffuse light fluctuations arising from CBF dynamics.⁷⁹ The basic idea is to split a portion of the source light into a reference path that recombines (or interferes) with the light that has reached the detector after diffusing through the tissue. In contrast to traditional DCS, which measures intensity fluctuations to determine the temporal intensity autocorrelation, interferometric DWS (iDWS) measures electric field fluctuations of the sample light that are in-phase with the reference light⁷⁹ to determine the temporal field autocorrelation. In iDWS, the strong reference light field multiplies and amplifies the weak sample light field from the tissue, enabling users to replace the photon-counting detectors of DCS with noisier, inexpensive detectors. Because the sample signal is amplified, iDWS is relatively insensitive to ambient light. Array detectors in conjunction with multimode fiber collection can parallelize iDWS and improve light throughput and SNR.⁷⁹ Interferometric techniques have been demonstrated with an APD,⁸⁰ a fast-sampling linear array sensor (e.g., non-scientific line-scan CMOS camera⁷⁹), and in the Fourier domain, with a two-dimensional sensor.⁸¹ Path length discrimination, akin to TD-NIRS and TD-DCS, can also be achieved in an interferometric setup by reducing the coherence length of the light source.^82,83

The SNR of a single interferometric channel can exceed that of a standard DCS photon-counting channel⁸⁰ if the reference light level is high enough. Therefore, with parallel detection, iDWS greatly improves SNR and reduces cost, accessing higher temporal resolution and allowing for measurements at larger source-detector separations.⁷⁶ For instance, pulsatile iDWS BFI measurements at 3.5-4.0 cm source-detector separation were recently demonstrated with a 333 kHz line-scan CMOS camera, where $\sim 192$ interferometric channels were achieved.⁷⁶ Further improvements of the SNR-to-cost ratio will be possible with iDWS methods that utilize low frame rate two-dimensional sensors^84,85 with many megapixels.^80,85 Keeping in mind that the field autocorrelation decays at half the rate of the intensity autocorrelation, detectors that provide access to $\mu \mathrm{s}$ time scales are needed to maximize brain specificity. While the original iDWS approach⁸⁶ required rapid sampling, the multi-exposure approach⁸⁶ allows access to these time scales with short exposure times. Thus, we anticipate that the next five years will see increasing adoption of interferometric optical CBF sensing and mapping in human subjects.

2.5.5.

Speckle visibility spectroscopy (SVS)

Speckle visibility spectroscopy (SVS), also known as speckle contrast optical spectroscopy (SCOS)⁸⁷ or diffuse speckle contrast analysis (DSCA),⁸⁸ is based on the idea that a rapidly fluctuating speckle field imaged on a camera with a fixed integration time will appear washed-out and more uniform than a more slowly fluctuating speckle field.⁸⁹ A sensitive determination of the speckle decorrelation can be made by characterizing the overall speckle statistics. SVS is categorically a DCS method, but it differs from most traditional DCS methods in that it measures speckles in space (ensemble) rather than in time (temporal) to arrive at an index of blood flow in the underlying tissue. The sensitivity of both strategies is unified by the number of independent observables (NIO) measured.⁹⁰ In a traditional temporal DCS measurement architecture, a single speckle grain is measured over time with a sensitive single-pixel detector, and the NIO is the number of full decorrelation events measured.⁶¹ In speckle ensemble DCS measurements or SVS, speckle grains are measured in parallel on an area detector such as a camera, and the NIO is the number of speckle grains imaged on the sensor.⁸⁹ The SVS approach allows higher NIO to be attained by leveraging high pixel count cameras.

SVS can be further improved by adding a reference beam to enable interferometric measurements—a method termed interferometric speckle visibility spectroscopy (iSVS).⁸⁴ The iSVS scheme is illustrated in Fig. 1. A standard SVS setup is modified by interfering the captured speckle field from the sample with a tilted reference beam, forming a hologram on the sensor. The hologram can be analyzed using standard off-axis processing techniques to retrieve the sample field.⁹¹ The amplitude of the speckle pattern captured in the hologram is related to the medium’s decorrelation time, which is ultimately associated with the flow within the medium.

Fig. 1

Measurement of scattering dynamics with interferometric speckle visibility spectroscopy (iSVS). (a) Coherent light illuminates a dynamic scattering media, generating a time-varying speckle field. The speckle field is combined with a tilted reference beam to form a hologram on the sensor. (b) The movement of scatterers (e.g., red blood cells) within the sample create captured holograms with normalized intensities proportional to the speed of the scatterer movement. (c) A depiction of how the recorded holograms are formed. If the medium is dynamic, the speckle pattern changes during the camera exposure time, leading to a reduced overall speckle contrast and maximum intensity in the recorded hologram. In contrast, if the scattering medium is static, the captured speckle pattern integrated over the exposure time is static, leading to a higher contrast captured hologram. (Figure modified from Xu et al.⁹⁰)

iSVS can perform shot-noise limited detection even in photon-starved situations, enabling the use of commercially available, large pixel count, high-speed cameras with higher noise levels than the sensors needed for SVS. The higher NIO can be leveraged to either improve the measurement’s SNR or to enable parallel measurements. One limitation of techniques based on spatial ensemble measurements is that if the underlying decorrelation process is associated with multiple decorrelation time constants, SVS and iSVS will not be able to separately quantify these constants. On the other hand, SVS and iSVS can reliably output an aggregate decorrelation time measurement, which is well suited for characterizing dynamic changes in the brain due to brain activity. iSVS also shares the limitation of other DCS type measurements in that the depth sensitivity relies on the distribution of the photon paths and thus degrades with penetration depth. However, the interferometric nature of the detection provides the ability to modify the coherence of the light source to improve the depth selectivity.

In the next five years, both practical and theoretical concerns of iSVS should be addressed to improve the performance and to enable it to be more easily used in laboratory and clinical settings. Practically, designing a stable and robust light collection system is critical to isolating the scattering dynamics from other environmental factors such as vibration. Other theoretical avenues for exploration include optimizing the source parameters (e.g., wavelength, coherence length) to maximize the sensitivity to the dynamics of interest and exploring the use of multi-exposure methods.

3.1.1.

Cardiac-based assessment of signal quality

Typically, fNIRS investigators consider the presence of cardiac pulsation in raw amplitude signals as a reliable and readily ascertainable indicator of successfully measured tissue hemodynamics. This feature has been translated into quantitative measures such as the Scalp Coupling Index (SCI)¹¹² and the Peak Spectral Power (PSP).¹¹³ These complementary metrics quantify the strength of cardiac pulsation of raw fNIRS signals in temporal and spectral domains, respectively. Critically, they can be computed on each short-timed segment (3–5 seconds) of an entire fNIRS recording to capture its inherently time-variant quality adequately. Notably, the combination of SCI and PSP allows for robust discrimination of clean signals from movement artifacts and/or noisy signals.¹¹³

Of practical note, such quality assessment can be performed and displayed for all fNIRS optical source-detector separations (i.e., channels) both in real-time during the headgear fitting to optimize the setup (e.g., PHOEBE)¹¹³ as well as post hoc to assess the data quality prior to other processing steps (e.g., QT-NIRS).¹¹⁴ In the latter case, data quality assessment and pruning of channels with little to no cardiac pulsation can be carried out automatically alongside data analysis by directly interfacing QT-NIRS with analytical tools such as Homer 3 or Brain AnalyzIR as part of the same graphical or programmatical pipeline. In addition, QT-NIRS can evaluate multiple recordings as a batch to generate a study-wide data quality report. At this time, extensive validation of these measures and technologies is being performed using both ground-truth datasets and real-world datasets encompassing a wide variety of experimental protocols and populations. It is expected that additional quality measures from existing and new literature will be incorporated into fNIRS data analysis in the near future.

3.1.2.

Motion artifact correction

Motion artifacts are one of the primary sources of noise in optical data, particularly in experiments involving infants or challenging clinical populations, as well as during tasks requiring participant movement (e.g., physical exercise or speech).^115,116 Moreover, the development of wearable CW-NIRS devices has allowed monitoring participants while performing real-world activities,¹³ further increasing the chance that motion artifacts will affect the recorded signals. Several strategies and motion correction techniques have been proposed in the literature to reduce the occurrence and/or correct motion artifacts in fNIRS data.^117–122 Recent comparison papers have demonstrated that a hybrid approach for motion artifact correction that combines spline and wavelet filtering seems to work best across several populations and tasks, especially when motion artifacts are highly contaminating the signal.^123,124 This approach aims to first reduce baseline shifts, which are hard to correct with wavelet filtering, with the spline interpolation and then apply wavelet filtering to minimize all other types of motion artifacts. The hybrid method was applied to different infant¹²³ and adult¹²⁴ datasets acquired with different NIRS devices, tasks, and laboratories worldwide with improved performance compared to other motion correction techniques.

One of the greatest limitations of most motion correction techniques is the requirement for users to set some input parameters, which might be dataset dependent, subjective, and quite challenging for non-expert users. In the future, we can expect the development of novel automatic motion correction techniques that will not require user input. For example, machine learning approaches might be applied to large fNIRS datasets to try to train a network able to correct any type of motion artifact,¹²⁵ or one might demonstrate that the input parameter of the wavelet technique might be automatically computed based on some known features of the experimental design (e.g., length of stimulus or SNR). Furthermore, integrating motion sensors in recently developed modular wearable fibreless NIRS devices will be straightforward, and novel motion correction approaches could be extended to exploit this additional information.¹²⁶

3.1.3.

Spatial registration

Previous test-retest fNIRS studies have shown that different sessions conducted on the same subject within a short interval can yield highly variable results.^127–132 This lack of reproducibility limits group-level analysis as well as within-subject comparisons across time in longitudinal studies. Although the high fNIRS variability is often attributed to low SNR, motion artifacts, and physiological noise,^127–132 recent works have shown that the lack of spatial information plays an important role in the observed fNIRS variability, particularly in low-density probes.^132,133 Variability in probe placement across subjects or sessions increases the fNIRS variance and reduces statistical power in data analysis. In some cases, marginal changes that are deemed statistically significant may merely reflect heterogeneity from underlying anatomy rather than brain functional changes.

Methods employing probe registration have been previously proposed in the fNIRS literature to overcome this limitation.^134–137 Typically, a digitizer and/or camera are used to acquire the position of the optodes, which can later be registered onto an anatomical MRI/CT image or a 3D model of the brain. Despite the effort to record this spatial information, there has been little development of data analysis methods to incorporate the probe location into the fNIRS analysis. One readily available solution recently presented is to use a forward model of light propagation to perform a region of interest (ROI) or voxel analysis of the fNIRS data rather than standard examination in the channel space.¹³⁸ Another powerful alternative approach employs real-time neuronavigation software to guide the optode positioning.^132,133,139 With real-time feedback, one can precisely position the optodes on the subject’s head and guarantee the NIRS optodes are on the target ROI before any measurement is taken, thus reducing variability due to spatial imprecisions. When applied to a longitudinal study, this approach doubled the intra-subject reproducibility and increased the fNIRS sensitivity to detect hemodynamic changes during a motor task (Fig. 3).¹³² As fNIRS advances towards routine functional brain imaging, stable longitudinal results at the subject level will become indispensable. Therefore, we envision that assisted probe positioning will be standard in longitudinal fNIRS protocols. Moreover, we envision this strategy could be extended to clinical settings to collect accurate optical data above a focal brain injury with a few sources and detectors, which will speed up experiment setup and data collection.

Fig. 3

fNIRS reproducibility with a neuronavigation system. Frequency of activated brain regions during a motor task for the standard and guided approaches for probe positioning across five participants. The standard procedure used a tape to record head-size and find the optode location relative to the 10-20 system, while the guided approach used real-time neuronavigation software to place the optodes on the target region of interest (motor cortex, shown in green in the reference brain located at the bottom of the figure). Data were collected at three sessions on three different days for every subject. The frequency of activation represents reproducibility.

3.2.1.

Systemic physiology augmented fNIRS

The primary source of the non-neuronal influence on fNIRS signals arises from systemic physiology, which can induce hemodynamic changes in both the extracerebral and cerebral tissue compartments. Thus, not all fNIRS-measured signal changes result from neurovascular coupling nor reflect brain activity. Changes in respiration, blood pressure, or autonomic nervous system activity can all induce changes in fNIRS signals that may mimic hemodynamic responses induced by neurovascular coupling.^140,141 One way to optimally interpret fNIRS signals is by adding independent measurements of key systemic physiology parameters, including arterial oxygenation (${\mathrm{SpO}}_2$), respiration rate (RR), heart rate (HR), end-tidal carbon dioxide (${{\mathrm{P}}_{\mathrm{ET}}\mathrm{CO}}_2$), continuous blood pressure (mean arterial pressure, MAP, and pulse pressure, PP), and skin conductance (SC). Measuring systemic physiology in combination with NIRS has been termed “systemic physiology augmented fNIRS” (SPA-fNIRS).¹⁴³ SPA-fNIRS has been applied in several studies to investigate the physiological reaction of subjects during fNIRS acquisitions, and significant changes in systemic physiology were found even during easy tasks that are not strenuous, such as exposure to colored light.^144–146 SPA-fNIRS can also be used in a hyperscanning setting; the first SPA-fNIRS hyperscanning study was recently published.¹⁴⁷

In general, SPA-fNIRS is an excellent approach to enable a deeper understanding of the fNIRS signals and prevent misinterpretation. It allows one to gain insight into what causes the changes in fNIRS signals. In addition, SPA-fNIRS can generate a novel framework to investigate the complex inter-relationship between brain activity-related cerebral hemodynamics and systemic physiological activity. Current development of SPA-fNIRS concerns the optimization of the multimodal measurement setup as well as signal-processing and data analysis frameworks tailored to exploit the full potential of SPA-fNIRS. When possible, SPA-fNIRS should be increasingly used in future fNIRS studies since there is growing evidence that multi-distance approaches contain most but not all systemic information.¹⁴⁸

3.2.2.

Multi-distance fNIRS

In fNIRS, various blind signal separation methods, such as independent component analysis (ICA), principal component analysis (PCA), or empirical mode decomposition, have been used to decompose the measured signal at a long separation into its components and to isolate the component that arises from the brain. (Note: Blind signal separation methods refer to approaches that attempt to isolate a signal when the source of the signal comes from a set mixed signals in which the mixing methodology is unknown.) However, these methods have the risk of overcorrecting the signal by removing the frequency band of interest.¹⁴⁹ Using the average of all long source-detector separation measurements as a regressor representing the global systemic physiology also runs the risk of eliminating the evoked brain response, especially for sensor designs that cover only local regions of interest.

Short source-detector separation channels that measure only scalp hemodynamics are a robust solution for temporally filtering the confounding components in the fNIRS signals.^150–152 With this approach, it is advisable to place short-separation detectors homogenously across the regions covered by the probe as the pial vasculature across the brain surface and the scalp vasculature are spatially inhomogeneous.¹⁵³ A typical method of using the short-separation channel signal for the removal of confounding components in the fNIRS signal is to add the signal as a regressor in a General Linear Model (GLM) framework.¹⁵⁴ This approach allows for a robust estimation of the brain response while regressing out the confounding signals. When enhanced with temporally embedded canonical correlation analysis, the method has been shown to further improve the estimation of the brain response by taking care of the time delays between different physiological signals.¹⁵⁵

High-density multi-distance measurements are another approach that naturally allow for spatially filtering the hemodynamic changes in superficial layers through image reconstruction.¹⁵⁶ Methods that incorporate short-separation regression directly into the image reconstruction scheme can further improve the estimation of the brain response.

While all these approaches are suitable for offline data analysis, regressing out the physiological confounds and extracting the brain response from fNIRS data in real-time requires more advanced methods. In this context, the Kalman filter allows a real-time estimation of hemodynamic changes for such data with dynamic statistical properties.^157–159 As mobile NIRS systems will facilitate studies in the real world and in real-time, the fNIRS signal will become more prone to motion-induced artifacts and systemic interference. Novel data analysis approaches can benefit from multimodal regressors for systemic confounds and motion artifacts dynamically adapted in a Kalman filtering scheme.¹⁶⁰

3.2.3.

Dual-slope method for enhanced depth sensitivity

Another approach to enhance depth sensitivity with multi-distance measurements utilizes two long channels (e.g., 2.5 and 3.5 cm) to generate data that have comparable contributions from superficial tissue and different contributions from cerebral tissue so that their combination can cancel out (or strongly suppress) extracerebral tissue contributions.¹⁶¹ This approach yields a slope of the optical signal versus source-detector distance, which is commonly accomplished in the field using a single source and multiple detectors (or a single detector and multiple sources).^162,163 Because this approach uses a single element (source or detector), it is referred to as a single-slope method. An extension of the single-slope method was proposed in the late 1990s to achieve more robust measurements of the tissue optical properties with FD-NIRS and a special arrangement of two sources and two detectors (self-calibrating approach).¹⁶⁴ Its main advantage is the insensitivity to probe-tissue coupling and instrumental drifts, thus allowing measurements without the need for any preliminary calibration. The special arrangement of two sources and two detectors for this self-calibrating approach has been used for cerebral oximetry with FD-NIRS¹⁶⁵ and CW-NIRS¹⁶⁶ and has been implemented in commercial cerebral oximeters based on CW-NIRS.^167,168

Further characterization and development of the self-calibrating approach for individual intensity and phase slopes measured with FD-NIRS led to a so-called dual-slope method.¹⁶⁹ The basic idea is to measure two paired single slopes (using two sources and two detectors), one where a given optode (say, a detector) collects data at the shorter distance and another where the same optode collects data at the longer distance. The dual-slope is the average of these paired slopes. This average cancels out or strongly suppresses contributions from source power, detector sensitivity, probe-tissue coupling, attenuation, delays in optical fibers, or any other instrument feature. In terms of probed tissue volume, dual-slope data were found to feature a greater relative sensitivity of deep vs. superficial tissue than single-slope and single-distance data, and the dual-slope phase showed a deeper region of sensitivity compared to dual-slope intensity.^170,171 The value of a technique that is largely insensitive to instrumental drifts, probe-tissue coupling, and motion artifacts, and that is preferentially sensitive to deeper tissue in a spatially confined volume is highly attractive and particularly important in non-invasive optical measurements of the brain. The dual-slope method was used with CW-NIRS and FD-NIRS to measure cerebral hemodynamics,^170,172 perform tissue imaging,^173,174 and generate absolute broadband absorption spectra of turbid media and biological tissue.^175,176

The current limitations of the method are related to signal-to-noise, especially in the case of phase measurements in FD-NIRS, and to the potential impact of lateral heterogeneity of tissue on the accuracy of the measured cerebral absorption changes. Nevertheless, the robustness and the preferential sensitivity to deeper tissue featured by the dual-slope method, especially with phase measurements in FD-NIRS, render it a valuable diffuse optical technique for functional brain imaging and the assessment of cerebral hemodynamics. Future developments include the characterization of the spatial region of sensitivity of dual-slope data in the presence of anatomical heterogeneity, the refinement of source-detector arrangements for imaging applications and optimal signal-to-noise conditions, and the selection of most effective modulation frequencies for specific applications.

3.2.4.

Depth selectivity with time-domain methods

The influence of extra-cortical contributions can be accounted for more easily using time-resolved methods,¹⁷⁷ where the time-of-flight of photons can be related to the light penetration depth. In classical TD-NIRS, this can be done by recording the DTOF of photons using the time-correlated single-photon counting technique.¹⁷⁸ Several data analysis approaches were proposed to estimate changes in the absorption coefficient of the medium at different depths in the tissue. These techniques are based on the parametrization of the measured DTOF: analysis of time windows, statistical moments, or Mellin-Laplace transform parameters.^43,179,180 Combining time-domain with distance-resolved NIRS or multi-wavelength measurements improves the method’s depth-selectivity,^181,182 including approaches based on two sources and two detectors that are arranged as described in the previous section for dual-slope measurements.¹⁸³ An approach based on the null-distance technique was proposed in which depth discrimination is based on the direct measurement of time spent by photons in the tissue when the inter-optode distance is very short.¹⁸⁴ Similar to TD-NIRS, one can also design an approach to estimate tissue perfusion at different depths using TD-DCS by analyzing the autocorrelation function with photons that arrived at a selected time-of-flight.⁷⁴

These approaches have been validated in multiple in vivo experiments using neurophysiological tests wherein the cortical responses were selectively obtained and differentiated from extracerebral contamination.^54,185–187 Potential utilization of the time-domain technique was also tested in several clinical applications.¹⁸⁸ Application of modular, highly integrated multichannel systems providing measurements at multiple source-detector pairs will improve depth resolution and reduce the influence of lateral heterogeneity of extracerebral tissue.

3.2.5.

Multi-distance DCS

Compared to NIRS, the higher blood flow in the brain compared to the scalp gives DCS an inherent advantage in terms of depth sensitivity. Collecting DCS data at different source-detector separations provides the most straightforward method of monitoring flow changes in the scalp and brain. An example of this multi-distance (MD) DCS approach is seen in Fig. 4(a), which shows blood flow responses recorded at source-detector separations of 1 and 2.7 cm during a hypercapnia challenge. Notice how the response recorded at 2.7 cm was larger and had a more rapid decline after hypercapnia, indicative of the expected fast response of the cerebral vasculature to changes in arterial carbon dioxide tension.¹⁸⁹

Fig. 4

(a) Average increase in the blood flow index ($\mathrm{\Delta }\mathrm{BFi}$) during a 2-min period of hypercapnia as indicated by the shaded region (N = 9). Time courses are presented for source-detector separations of 1 (top) and 2.7 cm (bottom). Shading surrounding each line represented the standard error of the mean. (b) Left: relative changes in BFI (rBFI) in response to increases in tourniquet pressure recorded at source-detector separations (${\mathrm{r}}_{\mathrm{SD}}$) of 1 and 3 cm. Shading around each line represents the standard error (N = 5). Right: rBFI for brain and scalp derived from the three-layer model. Error bars represent the standard error of the mean. For reference, the shading represents the individual time courses recorded at ${\mathrm{r}}_{\mathrm{SD}}=1$ and 3 cm.

One approach to separate scalp and cerebral blood flow is to analyze MD DCS data using a three-layered solution to the diffusion approximation representing scalp, skull, and brain.¹⁹⁰ The model requires independent estimates of the optical properties for the different layers and the thicknesses of the scalp and skull. The fitting parameters are blood flow indices for scalp and brain, assuming negligible flow in the skull layer. Despite the added complexity of the multi-layer model relative to the standard homogenous solution, Fig. 4(b) shows the benefits of separating the contributions from scalp and brain. In this example, DCS data were acquired at separations of 1 and 3 cm while inflating a cuff wrapped around the head to impede scalp blood flow. The model was able to predict negligible flow changes in the brain, as expected, despite large changes in the scalp, particularly at the end of the experiment when the cuff was released.¹⁹¹ A limitation with this modeling approach is the necessity to have independent measurements of layer thicknesses.¹⁹² In clinical applications, this information is often available from clinical MR or CT scans. Alternatively, it may be possible to use ultrasound to obtain these measurements,¹⁷² or to fit for this information by incorporating a simple pressure modulation challenge (designed to solely influence scalp perfusion) and constraining the fitting parameters accordingly.¹⁹³

Other approaches to separate scalp and cerebral blood flow using MD-DCS include a modified Beer-Lambert law for flow or regressing out the signal from a short source detector separation using a general linear model (as in Sec. 3.2.2).^194–196 While both of these approaches avoid several of the limitations of the three-layer analytical model and show promise as a means to isolate cerebral blood from extracerebral signals, work is needed to validate these approaches against other “gold standard” perfusion modalities before they become more widely utilized.

3.5.1.

Systemic low-frequency oscillations

Although functional connectivity maps derived from low-frequency oscillations (LFOs) are often attributed to neural activity, at least part of these oscillations appears to have a systemic contribution. The systemic low-frequency oscillations (sLFOs: $0.01\sim 0.1\mathrm{Hz}$) can be a confounding factor for interpreting functional connectivity maps. On the other hand, sLFOs measured with NIRS can carry useful information on their own and may be biomarkers of disease severity in various pathological conditions that alter circulation (e.g., stroke).²³⁹

The origin of sLFOs, which can be seen in both fNIRS and fMRI data, is still unclear. Several studies have explored the underlying mechanisms.^240–243 In one of the early concurrent fNIRS/fMRI studies of resting state, LFOs in oxyhemoglobin concentration measured by NIRS in the prefrontal region were cross-correlated with the fMRI signal from each voxel. High correlations between fNIRS bandpass filtered in the LFO range and the fMRI signal were found in many brain voxels with various time delays (up to few seconds), and the dynamic patterns in the time delays of LFOs mimicked the patterns of blood flow distribution throughout the brain.²⁴⁴ The magnitude of the time delays of these correlations along with their spatio-temporal distribution indicates that the LFOs are likely associated with the global (i.e., systemic) blood flow and can propagate through the vasculature. For further validation, in another concurrent fNIRS/fMRI study, the NIRS sensors were positioned to monitor peripheral circulation (e.g., fingertips and toes). Again, high correlations were found between the NIRS sLFOs from the periphery and many voxels within the brain measured with fMRI. Moreover, the spatiotemporal patterns observed were similar to those found when using NIRS LFOs from the prefrontal cortex as a regressor (Fig. 5). Interestingly, the sLFOs located at the left and right fingertips were similar, while the sLFOs found at the toe were delayed approximately 3 s compared to that from the finger. These studies supported that sLFOs are global physiological oscillations associated with blood circulation. As a result, sLFOs can be used as biomarkers to assess circulatory patterns in the brain and even the whole body.^245,246 Currently, the main limitation of this approach is the accurate identification of sLFOs because many physiological parameters (e.g., blood pressure) can contribute to sLFOs. In the future, we can expect that sLFOs will be used extensively in (1) identifying deficits in systemic circulation, and (2) denoising the fNIRS signal to expose the neuronal components.

Fig. 5

Concurrent peripheral NIRS measurements and resting-state fMRI. (a) NIRS sensors on the finger and toes. (b) sLFOs of changes in oxyhemoglobin concentration ($\mathrm{\Delta }[\mathrm{HbO}]$) measured from the toe (blue) and finger (red). These sLFOs are highly correlated with a delay about 3 s. c) Map of voxels wherein the BOLD fMRI signal was highly correlated with sLFOs of $\mathrm{\Delta }[\mathrm{HbO}]$ measured in the finger at any time lag. Figure modified from Tong et al.²⁴⁵

4.2.1.

Autism spectrum disorder (ASD)

Autism spectrum disorder (ASD) is a serious psychiatric disorder defined by behavioral deficits in social functioning, communication, restricted interests, and repetitive behaviors.^293–295 It is estimated that ASD affects 1/100 children worldwide.²⁹⁶ Early behavioral and educational interventions, starting at 18-24 months of age, improve outcomes in subsets of patients.^297,298 Neuroimaging methods, including both task-based fMRI and task-free functional connectivity MRI, have demonstrated sensitivity to neural signatures of ASD that may inform diagnosis and track responses to interventions.^299,300 However, the loud and constraining nature of MRI-based neuroimaging severely limits studies on direct within-room social communication, auditory processing, language generation, and overt motion. In addition, it presents an excessively challenging setting for sensitive participants, such as school-aged children and particularly young infants, toddlers, and those severely affected with ASD.³⁰¹ Functional neuroimaging with diffuse optics is a potential solution to providing a more naturalistic functional neuroimaging environment that will be amenable to studies on children with ASD.

Multiple studies have shown progress in applying optical neuroimaging methods to patients with ASD,^302,303 and recent developments in HD-DOT techniques have dramatically improved image quality, enabling fMRI-comparable image quality in an open, upright seated environment, even in low-resource settings.^4,304,305 With optical methods, important and novel questions about ASD throughout development, including the effects of clinical interventions, can be addressed in a comfortable, free-moving, and naturalistic setting. The wearability and the open scanning environment of fNIRS and HD-DOT will also allow for motor tasks that are more challenging in the confining environment of an fMRI to be administered in the ASD population, e.g., studies of infants or toddlers in dyadic interaction and paradigms exploring joint attention-related phenomena. These studies may also provide a foundation for future investigations pairing genotyping with functional neuroimaging. Future wearable, fiber-less HD-DOT imaging¹¹ may facilitate the identification of common mechanisms by which disparate genetic pathways to ASD result in the ASD phenotype.

4.2.2.

Attention deficient hyperactivity disorder (ADHD)

fNIRS also has a great potential for assessing children with attention deficit hyperactivity disorder (ADHD), for many of the same reasons outlined in Sec. 4.3.1. Regular fNIRS assessment with drop-out rates of less than 5%, even for ADHD children, are commonly observed owing to the modality’s ability to enable monitoring in naturalistic environments. In a series of recent studies, fNIRS has been used to determine how commonly-used treatments for ADHD – including atomoxetine,^306,307 osmotic release oral system methylphenidate,^308,309 and guanfacine extended release³¹⁰—affect the neurocognitive functions of ADHD children during an inhibitory and attentional task. Specially, it has been shown that, unlike age- and gender-matched typically developing controls, pre-medicated ADHD children do not activate the right inferior and middle prefrontal gyri (IFG/MFG) during the inhibitory task. The lack of activation of right IFG/MFG was acutely normalized after administration of atomoxetine and methylphenidate, but not after placebo administration. Interestingly, the administration of guanfacine-extended release activated the right inferior parietal cortex (IPC), but not the right IFG/MFG. On the other hand, the attentional task recruited the right IFG/MFG and IPC in typically developing children. The right prefrontal activation was normalized after atomoxetine and methylphenidate administration in ADHD children, but the right IFG normalization was specific to atomoxetine. These results demonstrate that fNIRS is sensitive to the differential neuropharmacological effects of atomoxetine, methylphenidate, and guanfacine-extended release as they up-regulated the noradrenaline and dopamine systems in the inhibitory and attentional networks in the brains of ADHD children.

Interestingly, such differences in cortical activation patterns, which may reflect underlying neuropharmacological effects, seem to differ between subtypes of ADHD children, namely, medication-naïve ADHD children with or without ASD.³¹¹ During the inhibitory task, medication- naïve ADHD children without ASD exhibited reduced right prefrontal activation, but methylphenidate medication significantly elicited right prefrontal activation. However, medication-naïve ADHD children with ASD showed marginal activation before the first day’s medication and reduced activation after methylphenidate medication in the right prefrontal area. Thus, ADHD with and without ASD can be characterized by a different underlying neurofunctional pathology of inhibition control using fNIRS. In a typical clinical diagnosis, distinction between ADHD with or without ASD takes careful behavioral observation over six months, but the fNIRS-based monitoring would provide a quick but useful hint on the presence of ASD-comorbidity and help clinicians to formulate drug-treatment plans for the ADHD children.

4.2.3.

Developmental coordination disorder (DCD)

Developmental coordination disorder (DCD) is a neurodevelopmental condition characterized by deficits in acquiring and executing motor skills within the Diagnostic and Statistical Manual 5^th edition. DCD is often manifested in school age children as clumsiness and slowness, impacting personal, social, academic, and occupational functioning.³¹² Although there is currently no cure, early steps can be taken to reduce the emotional, physical, and social consequences associated with this disorder. However, identifying children with DCD is difficult as the etiology of DCD remains unclear. Previous studies using traditional neuroimaging under rigid, confined, and unrealistic conditions, resulting in limited understanding of the neural correlates of the motor-cognitive dysfunction.³¹³

As the deficits in DCD are motor-related, mobile fNIRS is an ideal candidate tool, as it allows monitoring brain activity during gross motor movements. One recent study collected fNIRS during active ambulatory behavioral task execution from both typically developing and DCD children in one of the largest neuroimaging studies of DCD.³¹³ Children performed both cognitive and physical tasks in both single and dual conditions. The study confirmed that DCD is a motor-cognitive disability, as gross motor complex tasks revealed dysfunction within the right middle and superior frontal gyri of the prefrontal cortex. By incorporating behavioral performance, decreased neural efficiency in these regions were revealed in children with DCD, specifically during motor tasks. Evaluating the disorder in such ecologically valid contexts could help optimize the interventions and open the door for precision therapies.

4.3.1.

Neuroergonomics

Understanding brain function in real-world applications is the next frontier for diffuse optics in human neuroscience. Existing studies with traditional neuroimaging approaches have accumulated overwhelming knowledge but are limited in scope, i.e., only in artificial lab settings and with simplified parametric tasks.³¹⁴ Diffuse optical techniques can fill the current gap by providing a reliable tool to the emerging and new interdisciplinary field of neuroergonomics, which aims to investigate human brain function in the wild, during unrestricted real-world tasks in everyday life contexts, and its relationship to action, behavior, body, and environment.^315,316

This paradigm shift opens exciting experimental opportunities in relevant fields not yet fully explored by other traditional neuroimaging techniques. The ultra-portable wearable and wireless NIRS devices already mentioned overcome limitations and constraints imposed by traditional neuroimaging tools on experimental protocols, data collection settings, and task conditions at the expense of ecological validity.^16,317–319 Recent fNIRS studies demonstrate examples of experimental scenarios that were not feasible until recently, such as speaker-listener coupling,^320–322 mental workload decoding in real-time during a flight over the clouds,³²³ or while walking outdoors.^324,325 In addition, fNIRS can provide continuous and repeated measurement for longitudinal evaluation, and this can be used not only to assess but also to enhance the training/skill acquisition for complex tasks³²⁶ or to provide online real-time feedback as clinical intervention.³²⁷ There is a vast potential offered by continuous neural measures to reshape our understanding of the brain mechanisms and how we engage with the world. Neuroergonomic applications with NIRS technology may eventually be an integral part of the way we learn new skills, do our work, and entertain ourselves.

Despite the progress, there are still significant shortcomings that plague the full utilization of NIRS for neuroergonomic applications, including low-to-no sensitivity for deeper brain areas, low spatiotemporal resolution, and low SNR. Non-portable, room-based NIRS instruments could address some of the challenges with denser sensor arrays and tighter coupling of sensors, but mobile and newer ultra-mobile versions are more affected by these limitations.³²⁸ In the meantime, ongoing sensor and signal processing improvements make systems more portable, reliable, and affordable. Newer generation ultra-portable wearable NIRS systems are already positioned to go outside the lab for continuous measurements over longer periods and less constrained setups. These newer systems can allow more natural multimodal integration with EEG for faster time-resolution or neurostimulation for guided modulation of brain activity and enable adaptive computer/machine interaction. Improvements in artificial intelligence/machine learning approaches to recognize and label behavioral patterns will provide task context to brain signals for expanded and more automated analysis. Together with the capability for continuous interpersonal neural measures from fNIRS, these improvements may soon yield new means of cooperation, training, gaming, and work.

4.3.2.

Social interactions

Humans are profoundly social. Understanding the neurobiology that underlies social behaviors is a frontier in social neuroscience.^329,330 Conventional investigations of the neural underpinnings for social behaviors are based on single brain investigations. These conventions have led to a long-standing experimental paucity of two-person interactive experimental paradigms to investigate social behaviors. However, recent developments in fNIRS for hyperscanning (i.e., simultaneous brain scanning of two or more individuals during live interactions) pave the way for long-awaited studies of live interactions between individuals.³³¹ Technical advances in fNIRS (see Secs. 2 and 3) and the growing importance of understanding the biological components of live and interactive human social behaviors have supported the emergence of dyadic neuroscience.^332–334

From a Dyadic Neuroscience perspective, a single brain is only one half of a dynamic social unit.^335–337 In the case of a dyad, pairs of individuals are “linked” together by meaningful and reciprocal exchanges of information.^338,339 Quantification of dyadic properties includes measures of neural coupling, sometimes referred to as inter-brain synchrony or cross-brain coherence,^340–342 which quantifies the extent to which the temporal patterns of neural signals across brains are correlated.^{321,343–345} The main questions include: What are these dyadic mechanisms? How can they be isolated? How can they be modulated in health and disease?

Multimodal acquisitions of many simultaneous measures, including EEG, face recognition, auditory recordings, eye-tracking, and physiological variables, enhance interpretations of fNIRS findings.^346–348 Wearable systems with full-head and high-density spatial resolution promise continued advances. A vast array of development opportunities exists in potential fusions of collaborating disciplines. However, the absence of sensitivity in subcortical layers of the brain is a significant limitation, but it is balanced with the advantages of probing neural processes during live and natural social interactions. These advantages bring social neuroscience closer to applications that benefit ordinary lives and patient care.³⁴⁹

4.3.3.

Learning and education

Since NIRS is considerably more flexible regarding head motion tolerance and environmental conditions, it has become an attractive neuroscientific tool for investigating learning in real educational contexts. Moreover, some portable devices are also available, making it possible to collect the data at school, i.e., a natural learning environment for the child. In addition, experiments with children usually demand the preparation be short so the participants are more collaborative during the acquisition. Thus, the ease of cap/band/net preparation in fNIRS is also helpful. 

The main progress in fNIRS applications in education is the exploration of interpersonal interactions during learning using hyperscanning.^322,350 These situations connect with pedagogical constructs such as Vygotsky’s proximal development zone. Moreover, combining fNIRS signals and machine learning for mental states decoding made it possible to monitor and quantify attention, engagement, and learning.³⁵¹

Current limitations in this field may also be seen as opportunities for technological development. In comparison to biomedical research, budgets in education are very limited. Thus, the availability of low-cost portable quality systems³⁵² is crucial for impacting elementary schools. Furthermore, considering the specificities of educational contexts in which block or event-related designs are not feasible, novel experimental designs and the respective analytical methods are necessary. In the next five years, we can expect that fNIRS in education will go beyond the frontiers of academic research. Neurotechnology startups focused on education will provide hardware and analytical products to quantify engagement and learning. This innovation will allow experimental sciences to be more relevant in educational strategies and instructional resources.

4.3.4.

Perceptual-cognitive development

The advances in hardware and data analysis mentioned in the previous sections have enabled novel perceptual and cognitive development studies. Here, we highlight a number of exciting avenues of research that have recently emerged and that we think will be highly influential in fNIRS studies of perceptual-cognitive development in the next five years.

As caregivers and the environment are important in infants’ perceptual-cognitive development and its disorders, the use of fNIRS in social contexts and in interaction with caregivers is promising for revealing mechanisms of development and for understanding risk and resilience. To this end, hyperscanning allows researchers to assess how infants interact with and learn from their environment, such as within the mother–child dyad.³⁵³ These studies highlight that during real-life communication, infants’ brains and behaviors both shape and reflect those of adults.³⁵⁴

Another way in which fNIRS is supporting more ecologically-valid research in perceptual-cognitive development is through its application in a range of global health research projects.^292,355,356 These projects are opening new horizons for understanding environment-specific impacts on brain development as well as increasing the diversity of research participants and the communities who are represented in studies of perceptual-cognitive development.^357,358

Finally, multi-modality recordings (e.g., EEG with NIRS) are on the rise^359–361 and in the next five years will provide a critical cross-modal comparison of the physiological processes underlying infants’ perception and cognition. Indeed, NIRS is uniquely poised to allow for these kinds of multimodal recordings as it does not have the same constraints as the MR environment.

5.6.1.

Cerebral autoregulation

Cerebral autoregulation refers to the constraint of blood flow in the brain in the presence of changes in blood pressure. This process is mediated by vessel reactivity, and it acts to prevent potentially harmful fluctuations in cerebral blood flow. When arterial blood pressure decreases from normal levels, arteries and arterioles within the brain dilate to decrease vascular resistance and maintain CBF at normal levels. However, when arterial blood pressure is reduced to a degree whereby the vasodilatory response is exhausted, CBF decreases, and the brain is at risk for ischemia. This transition point is called the lower limit of autoregulation, and it serves as a clinically significant reference point for arterial blood pressure management. Unfortunately, because there is no approved method to delineate this lower limit in clinical practice, and because it can vary between subjects and across disease states, clinical care of conditions that impair cerebral perfusion pressure and are associated with ischemic injuries—such as head trauma, subarachnoid hemorrhage, sepsis, and cardiac surgery—involves management of arterial blood pressure in ignorance of this lower limit.

Many published methods exist to measure autoregulation.⁴⁶⁰ All methods are composed of three essential elements: 1) a measure of perfusion pressure (such as arterial blood pressure or cerebral perfusion pressure), 2) a measure of a cerebral vascular property (such as blood flow, blood volume, or blood oxygenation), and 3) a measure of the relatedness of the cerebral vasculature to the perfusion pressure (such as correlation, coherence, phase shift or gain of transfer). The modalities employed and the permutations of combinations of these three elements are numerous. Given the numerous advantageous attributes of NIRS and DCS discussed herein, including their low-cost and bedside monitoring capabilities, these tools hold promise to measure cerebral vascular properties (element 2). Indeed, literature reports of autoregulation assessment with commercially available CW-NIRS systems abound, and more recently, assessment with DCS measurements of blood flow has become more prevalent.^{388,432,461–463} While DCS provides direct measurements of CBF, NIRS measurements of blood volume, cerebral oxygenation, or the difference between the concentrations of oxy- and deoxy-hemoglobin have also been used as surrogates of CBF. Dynamic relationships between the optical measurements of CBF (or its surrogate) and mean arterial pressure (MAP), are then examined to yield measures of cerebral autoregulation. Most commonly, these dynamic relationships are determined by transfer function analysis or correlation analysis. In contrast to TCD assessments of cerebral autoregulation that rely on measurements of CBF velocity in the middle cerebral artery,^464,465 a significant feature NIRS/DCS is that the tools provide complementary local measurements of the autoregulatory capacity at the microcirculation level. Moreover, the local measurements made with NIRS and DCS lend themselves well to spatially resolved or imaging implementations.

The main limitation of autoregulation monitoring is the need for oscillations in perfusion pressure at relatively low frequencies. Pulse and respiratory frequency oscillations are too fast to reliably engage the autoregulation mechanism, which has an upper frequency limit of 0.10 – 0.15 Hz.⁴⁶⁶ Transients and oscillations at lower frequencies (i.e., 0.008 – 0.03 Hz) are ideal for monitoring autoregulation, but these phenomena are often irregular. The lack of reliable low-frequency blood pressure oscillations creates random noise in the autoregulation analysis, so time windows used for assessment must be averaged together or excluded. Thus, autoregulation monitoring can be slow, especially when the arterial blood pressure is static. Furthermore, there is inadequate clinical experience with autoregulation monitoring outside of the research arena to know if widespread application will be feasible or safe. Removal of statistical noise using appropriate signal processing will improve the safety of this modality,⁴⁶⁷ but proper phase II studies of safety are needed to demonstrate a lack of harmful side effects from corrective interventions based on autoregulation monitoring. The application of autoregulation monitoring will likely lead to clinical management that utilizes higher arterial blood pressure targets for most conditions and possibly that tolerates lower arterial blood pressure management for other conditions. The safety of this practice change must be determined before widespread use occurs.

Anticipated advancements in autoregulation monitoring include: 1) improved SNR to render precise and rapid delineation of the lower limit of autoregulation; 2) improvement in the clinical user interface such that automated interpretation renders easy-to-read guardrails for arterial blood pressure management without needing to be an expert in the mechanics of autoregulation monitoring; 3) studies demonstrating feasibility (reduction of the burden of hypotension below the lower limit), safety (does optimizing blood pressure to the brain harm other organs?), and efficacy (a reduction in patient morbidity with the application of autoregulation monitoring). Diffuse optical methods such as DCS and NIRS offer a good option for autoregulation assessment and monitoring in that they are non-invasive, portable, safe, and offer a strong sensitivity to local microvascular flow. The first general challenge mentioned above, namely the achievement of a high SNR also implies a suitable suppression of confounding contributions from extracerebral tissue hemodynamics and is currently being actively explored in light of the many research advances that have been described in this article. The second and third challenges mentioned above require clinical validation studies and technological refinements in the user interface that are also being explored in the field.

5.6.2.

Cerebrovascular reactivity

Cerebrovascular reactivity (CVR) describes the process why which the cerebral vasculature dilates or constricts in response to a vasoactive stimulus like dissolved oxygen or carbon dioxide. CVR is an integral mechanism in brain homeostasis. Unfortunately, this mechanism can become damaged in a wide range of conditions, including cerebrovascular disease,^468–471 stroke,^472–475 cardiac arrest,⁴⁷⁶ and traumatic brain injury.^477–480 Moreover, several studies have suggested that impaired CVR may serve as a prognostic biomarker of functional outcome.^481–484 For example, impaired CVR has been shown to be a more robust predictor of poor neurologic outcome than standard imaging-based scoring systems in patients with carotid artery stenosis.⁴⁷³ Thus, bedside quantification of CVR could serve a valuable role in guiding patient care.

Assessment of CVR requires both a vasoactive stimulus (either by spontaneous breath-holding or through administration of acetazolamide or carbon dioxide) as well as the quantification of the cerebral blood flow response to the stimulus. The latter requirement is well suited for either DCS measurements of CBF, or NIRS measures of blood flow surrogates like cerebral blood volume or oxygen saturation. Indeed, both NIRS and DCS have been used to quantify CVR in a range of pilot patient cohorts, including children with congenital heart defects^485–487 and obstructive sleep apnea,⁴⁸⁸ adults with cerebrovascular disease,^{368,369,489,490} as well as healthy controls.¹⁸⁹ However, the main limitations of assessing CVR with diffuse optics are the poor spatial sensitivity, depth sensitivity limited to superficial cortex, and contributions of extracerebral hemodynamics. Spatial and depth sensitivities are most relevant when regionally heterogenous CVR is thought to be present (e.g., in the case of ischemic stroke); however, this limitation is less relevant in patient cohorts wherein global impairments in CVR are commonly observed (e.g., traumatic brain injury, sickle cell disease). Extracerebral contributions to the diffuse optical signal are of concern when assessing CVR, as the stimuli utilized to elicit the CVR response often also elicit systemic changes (i.e., increase in blood pressure and/or heart rate) that influence both scalp and brain perfusion. These systemic changes can confound the estimation of CVR in a manner that is not present with other perfusion modalities.

In the future, we anticipate a growing focus on assessment of CVR with diffuse optics as the recognition of the importance of this biomarker of vascular health becomes more widespread. Towards this end, researchers will need to capitalize on both hardware and software advancements discussed in previous sections to enhance depth penetration and spatial coverage of the measurements, as well as to provide real-time assessments of CVR at the bedside. Moreover, we will need a better understanding of the reproducibility of these measurements, including not just measurement repeatability, but also the sensitivity of CVR to factors such as time of day, caffeine intake, hormone status, age, and sex. Finally, given the relative ease of CVR assessment with diffuse optics (compared to MRI, PET, or even TCD), the field could benefit from the continued development of more tolerable and practical means of applying a vasoactive stimulus that do not require inhalation of ${\mathrm{CO}}_2$ gas or intravenous administration of acetazolamide. Recent work that quantifies the CBF response to fluctuations in end-tidal ${\mathrm{CO}}_2$ that occur naturally during free-breathing while at rest or due to subject-guided breath modulation^491,492 would provide a complementary, minimally burdensome means to assess CVR with diffuse optics.