Functional near-infrared spectroscopy
Functional near-infrared spectroscopy (fNIRS) is a non-invasive optical neuroimaging technique that measures changes in oxygenated (HbO₂) and deoxygenated (HbR) hemoglobin concentrations in the cerebral cortex by emitting near-infrared light (typically 650–950 nm) through the scalp and detecting the reflected or transmitted signals.[1] This method indirectly assesses neural activity via hemodynamic responses, similar to functional magnetic resonance imaging (fMRI), but offers greater portability and tolerance to participant movement.[2] Developed from the discovery of the near-infrared optical window by Frans F. Jöbsis in 1977, fNIRS has evolved into a widely used tool in cognitive neuroscience over the past 25 years, with applications expanding rapidly since the 1990s.[1][2] The technique operates on the principle of light absorption and scattering in brain tissue, quantified using the modified Beer-Lambert law, which relates changes in light attenuation to hemoglobin concentration variations.[1] Systems vary in design, including continuous-wave (CW), frequency-domain (FD), and time-domain (TD) configurations, with CW being the most common due to its simplicity and cost-effectiveness.[3] fNIRS primarily images superficial cortical regions (penetration depth of 1.5–2 cm, spatial resolution of 2–3 cm), making it suitable for studying prefrontal, motor, and temporal areas but limited for subcortical structures.[1][2] Key advantages of fNIRS include its non-invasive nature, safety for all age groups (including infants and pregnant individuals), low cost (systems often under $50,000), silent operation, and high temporal resolution (up to 10 Hz sampling rates), enabling use in naturalistic settings like walking or social interactions.[1] Unlike fMRI or positron emission tomography (PET), it requires no magnetic fields or radiation, reducing contraindications and allowing integration with other modalities.[2] However, challenges persist, such as susceptibility to motion artifacts, superficial signal contamination from scalp hemodynamics, and the need for standardized data analysis protocols.[1] Applications of fNIRS span cognitive neuroscience, neurodevelopment, and clinical domains, including assessment of working memory, language processing, and social cognition through hyperscanning (simultaneous imaging of multiple brains).[1] In clinical contexts, it aids in diagnosing disorders of consciousness (DoC), such as vegetative and minimally conscious states, with studies from 1993 to 2024 showing up to 92% accuracy in brain-computer interface (BCI) classifications for residual awareness detection.[3] Recent advances, including wearable devices and integration with neuromodulation therapies like deep brain stimulation, have enhanced its utility in rehabilitation for stroke, neurodegenerative diseases (e.g., Alzheimer's), and psychiatric conditions like schizophrenia.[1][3] The Society for functional Near-Infrared Spectroscopy, founded in 2014, supports ongoing research and standardization efforts.[1]Overview
Description
Functional near-infrared spectroscopy (fNIRS) is a noninvasive neuroimaging technique that employs near-infrared light in the wavelength range of 650–950 nm to monitor changes in oxygenated hemoglobin (HbO₂) and deoxygenated hemoglobin (HHb) concentrations in the cerebral cortex, reflecting hemodynamic responses coupled to neural activity.[4][5] This method leverages the relative transparency of biological tissues to near-infrared light, allowing detection of functional brain activation without the need for ionizing radiation or confinement to a scanner.[6] In fNIRS, near-infrared light interacts with brain tissue primarily through absorption by key chromophores—HbO₂, HHb, and to a lesser extent cytochrome c oxidase—and multiple scattering events that cause the light to follow a curved, banana-shaped path beneath the scalp.[1][5] The scattered light emerging as diffuse reflectance is analyzed to infer concentration changes in these chromophores, as HbO₂ and HHb exhibit distinct absorption spectra in this wavelength range.[4] These measurements are typically quantified using the modified Beer-Lambert law, which accounts for both absorption and scattering effects in highly scattering media like tissue.[1] A standard fNIRS setup consists of light sources such as light-emitting diodes (LEDs) or laser diodes, paired with detectors including photodiodes or avalanche photodiodes, integrated into optodes that are positioned on the scalp.[7][8] Source-detector separations of 2–4 cm are commonly used to optimize signal from cortical regions, with most systems operating in continuous-wave mode for intensity measurements.[1][7] fNIRS provides a spatial resolution of approximately 1–2 cm and penetrates 1–3 cm into the cortex, enabling superficial monitoring of brain activity.[1][9] The technique was first demonstrated in 1977 by F. F. Jöbsis for noninvasive monitoring of cerebral oxygenation using near-infrared spectroscopy.[6]Modified Beer-Lambert Law
The modified Beer-Lambert law (MBLL) serves as the foundational mathematical framework for quantifying changes in hemoglobin concentrations in functional near-infrared spectroscopy (fNIRS), adapting the classical Beer-Lambert law to account for the diffusive scattering of light in turbid biological tissues.[10] The original Beer-Lambert law describes light attenuation in non-scattering media as A = \epsilon C l, where A is the absorbance, \epsilon is the molar extinction coefficient, C is the chromophore concentration, and l is the path length; however, in tissues like the human head, photons undergo multiple scattering events, resulting in a longer, curved trajectory rather than a straight line.[10] To derive the MBLL, the effective path length is scaled by a differential pathlength factor (DPF), and a term for scattering losses is included, yielding the integral form A(\lambda) = -\log_{10}(I/I_0) = \epsilon(\lambda) C \cdot d \cdot \overline{DPF}(\lambda) + G, where I and I_0 are the detected and incident intensities, \lambda is the wavelength, d is the source-detector separation, \overline{DPF} is the mean DPF, and G is a wavelength-dependent scattering term.[10][11] For fNIRS applications, which focus on small hemodynamic changes relative to a baseline, the differential form of the MBLL is used, assuming constant scattering (\Delta G = 0) and negligible higher-order effects: \Delta A(\lambda) = \epsilon_{HbO_2}(\lambda) \Delta [HbO_2] \cdot d \cdot DPF(\lambda) + \epsilon_{HHb}(\lambda) \Delta [HHb] \cdot d \cdot DPF(\lambda) Here, \Delta A(\lambda) = -\log_{10}(I(\lambda)/I_0(\lambda)) represents the change in optical density (attenuation) at wavelength \lambda, \Delta [HbO_2] and \Delta [HHb] are changes in oxygenated and deoxygenated hemoglobin concentrations, and the \epsilon values are the wavelength-specific extinction coefficients for each chromophore.[12] The DPF corrects for the increased path length due to photon diffusion, which follows a characteristic "banana-shaped" trajectory through the tissue, with typical values ranging from 6 to 7.25 for the adult head at near-infrared wavelengths (e.g., 760–850 nm), depending on age, tissue type, and \lambda.[12] The term G accounts for non-absorptive losses from scattering at the tissue boundaries and within the medium, but in the differential form, it cancels out under the assumption of stable scattering properties.[11] The MBLL relies on several key assumptions, including a homogeneous medium with uniform chromophore distributions, negligible fluorescence or other non-absorptive wavelength-dependent effects, and prior knowledge of the DPF, which is often estimated from time- or frequency-domain measurements or empirical models.[12][11] Limitations arise when these assumptions are violated, such as in heterogeneous tissues where focal activation leads to partial pathlength effects (requiring partial pathlength factors instead of DPF), wavelength-dependent variations in DPF not fully accounted for, or large concentration changes that introduce non-linearity.[12] These constraints can introduce errors in concentration estimates, particularly cross-talk between \Delta [HbO_2] and \Delta [HHb] signals if the DPF is inaccurately specified.[12] The MBLL enables the separation of \Delta [HbO_2] and \Delta [HHb] signals by exploiting their distinct absorption spectra across multiple wavelengths (at minimum two, typically 3–4 in fNIRS systems, such as 780 nm and 850 nm).[12] At isosbestic points (e.g., ~800 nm), the \epsilon values are equal, but elsewhere, differences allow solving the system of equations via matrix inversion. For N wavelengths, the vector of concentration changes \Delta \mathbf{C} = [\Delta [HbO_2], \Delta [HHb]]^T is obtained as: \Delta \mathbf{C} = \left( \boldsymbol{\epsilon}^T \boldsymbol{\epsilon} \right)^{-1} \boldsymbol{\epsilon}^T \cdot \frac{\Delta \mathbf{A}}{d \cdot DPF} where \boldsymbol{\epsilon} is the N \times 2 matrix of extinction coefficients, and \Delta \mathbf{A} is the vector of \Delta A(\lambda).[12] This inversion yields independent estimates of each hemoglobin species, providing insights into cerebral oxygenation and blood volume changes.[12]History
Early Developments
The foundations of functional near-infrared spectroscopy (fNIRS) emerged from advancements in near-infrared (NIR) spectroscopy applied to oximetry during the 1970s. In Japan, biomedical engineer Takuo Aoyagi developed the core principle of pulse oximetry in 1974 while working at Nihon Kohden Corporation. This technique exploited the pulsatile nature of arterial blood flow to noninvasively measure oxygen saturation using NIR wavelengths (around 660 nm and 940 nm), where oxygenated and deoxygenated hemoglobin exhibit distinct absorption differences, laying the groundwork for tissue oxygenation monitoring without invasive procedures.[13] A pivotal theoretical advancement came in 1977 with the work of physiologist Frans F. Jöbsis at Duke University. In his seminal paper, Jöbsis demonstrated that NIR light (700–1000 nm) could penetrate biological tissues to detect changes in the redox state of cytochrome c oxidase, a key enzyme in mitochondrial oxidative metabolism, in the brains of anesthetized cats. Using reflectance spectrophotometry, he showed real-time monitoring of cerebral oxygen sufficiency and circulatory parameters, such as hemoglobin oxygenation, without surgical intervention, inspiring the extension of NIR to brain function assessment. This discovery highlighted the "NIR window" where tissue scattering and absorption are minimized, enabling deeper penetration than visible light.[6] During the 1980s, early animal studies further validated NIR for cerebral oxygenation measurements, primarily led by David T. Delpy and colleagues at University College London. Their research focused on quantifying NIR signals in vivo, including characterization of absorption spectra for cytochrome c oxidase and hemoglobin in rat and lamb models under varying oxygenation conditions. For example, Delpy's group developed methods to differentiate oxygenated from deoxygenated hemoglobin using multiwavelength spectroscopy, demonstrating reliable detection of hypoxia-induced changes in brain tissue oxygen delivery and metabolism. These experiments, often conducted on newborn animal models to mimic human neonatal conditions, established proof-of-concept for clinical translation and emphasized the need for scattering corrections in turbid media like brain tissue.[14][15] The shift toward functional brain imaging began in the early 1990s with initial human studies exploiting NIR-detectable hemodynamic responses to neural activation. In 1993, multiple research groups independently reported the first observations of brain activity-linked changes in oxygenation. Notably, Yoko Hoshi and Masahide Tamura in Japan used multichannel NIR topography to image prefrontal cortex activation during a word fluency task, detecting increases in oxygenated hemoglobin and decreases in deoxygenated hemoglobin, akin to blood-oxygen-level-dependent (BOLD) signals in functional MRI. Similar findings from groups in Germany (e.g., Hellmuth Obrig and colleagues) and the United States confirmed task-evoked cortical responses, marking the birth of fNIRS as a neuroimaging tool. These studies relied on the modified Beer-Lambert law for pathlength-compensated quantification of chromophore concentrations.[16] Initial implementations of fNIRS grappled with substantial technical hurdles, foremost among them low signal-to-noise ratios stemming from extracerebral contamination, particularly scalp hemodynamics and superficial blood flow. NIR light's diffuse path through layered tissues (scalp, skull, cortex) amplified interference from non-neuronal sources, often overwhelming subtle functional signals from deeper parenchyma. This limitation necessitated a strategic emphasis on superficial cortical regions, such as the prefrontal and motor cortices, where optode placement could maximize brain-specific contributions while minimizing artifacts from hair, skin, and vascular pulsations.[17][18]Key Milestones and Regional Contributions
In the late 1990s and early 2000s, significant advancements in the United States and United Kingdom focused on commercializing fNIRS systems and integrating topographic mapping for broader cortical imaging. NIRx Medical Technologies, founded in the early 2000s based on concepts introduced in 1988, developed one of the first commercial fNIRS platforms emphasizing tomographic imaging with multi-distance measurements for improved depth resolution.[19] In parallel, UK researchers contributed to multichannel systems that enabled 2D topographic mapping of hemodynamic changes across the prefrontal and occipital cortices, facilitating studies on cognitive tasks like mental arithmetic.[20] By the 2020s, US and UK efforts advanced portable and wearable fNIRS devices, such as fiberless systems tolerant to motion, supporting real-world applications in naturalistic settings.[21] Japanese researchers pioneered multi-channel fNIRS systems in the 2000s, with Hitachi Medical Corporation developing the ETG-4000 optical topography system around 2004, which provided 52 channels for high-density prefrontal monitoring and received FDA clearance for cerebral oximetry.[22] This system facilitated the introduction of hyperscanning techniques, allowing simultaneous recording of brain activity in multiple individuals during social interactions, with early demonstrations in collaborative tasks by the mid-2000s.[20] Ongoing Japanese contributions include extensive pediatric applications, leveraging fNIRS's non-invasive nature for studying infant brain development and neurodevelopmental disorders like autism.[21] Globally, the 2000s saw key regulatory milestones, including FDA approvals for cerebral oximetry devices like the INVOS 3100 in 1993, which laid groundwork for functional applications by validating NIRS for brain oxygenation monitoring during surgery.[23] Post-2010, fNIRS publications surged exponentially, doubling roughly every 3.5 years, driven by its adoption in cognitive neuroscience and clinical studies.[1] In the 2020s, integration of artificial intelligence enhanced data analysis, with machine learning models improving artifact removal and classification of hemodynamic signals for applications like cognitive load estimation.[24] Recent developments from 2024 to 2025 include refined short-channel regression methods to better isolate superficial physiological artifacts, such as scalp hemodynamics, using orthogonalization within generalized linear models to boost signal quality in multimodal paradigms.[25] The global fNIRS market grew to approximately USD 188 million in 2023, reflecting increased demand for portable systems in research and clinical settings.[26] Post-2020 adoption in Latin America accelerated, with the first Latin American NIRS Meeting in 2022 fostering regional collaboration on applications like premature infant brain monitoring in Brazil and Parkinson's disease studies in Mexico.[27]Measurement Techniques
Continuous-Wave fNIRS
Continuous-wave functional near-infrared spectroscopy (CW fNIRS) operates on the principle of illuminating biological tissue with continuous, non-modulated near-infrared light from sources such as light-emitting diodes or laser diodes, typically at wavelengths between 650 and 900 nm, and measuring the attenuation of the detected intensity at scalp-placed photodetectors after the light undergoes diffuse scattering and absorption in the tissue. These attenuation changes over time reflect variations in the concentrations of oxyhemoglobin (HbO₂) and deoxyhemoglobin (HHb), which serve as indirect indicators of neuronal activity through neurovascular coupling. The method relies on the modified Beer-Lambert law (MBLL) for quantitative interpretation, as outlined in foundational work on near-infrared applications.[28] A key advantage of CW fNIRS is its simplicity and low cost, achieved through straightforward instrumentation without the need for complex timing or modulation electronics, making it accessible for widespread use in research and clinical settings. It supports real-time monitoring with typical sampling rates of 10–50 Hz, enabling the capture of hemodynamic responses on timescales relevant to cognitive tasks. Additionally, its portability allows deployment in naturalistic environments, such as with ambulatory subjects, unlike more stationary modalities. In typical setups, optodes are arranged in multi-distance configurations on the scalp, with short source-detector separations of approximately 1 cm to capture superficial signals for correction and longer separations of 3–4 cm to probe deeper cortical layers, often following the international 10–20 EEG system for positioning.[29] Signal extraction begins by converting raw intensity measurements I(t) to optical density (OD) using the formula \text{OD}(t, \lambda) = -\log_{10} \left( \frac{I(t, \lambda)}{I_0(\lambda)} \right), where I_0(\lambda) is the initial intensity and \lambda denotes wavelength; changes in OD are then applied to the MBLL to estimate relative changes \Delta[\text{HbO}_2] and \Delta[\text{HHb}] at dual wavelengths, such as 760 nm and 850 nm. Despite these strengths, CW fNIRS has limitations, including the inability to directly measure absolute hemoglobin concentrations or absolute scattering coefficients, which confounds depth-resolved quantification. It is particularly susceptible to motion artifacts from head movements and extracerebral contamination from scalp blood flow, which can obscure cerebral signals without advanced corrections. CW fNIRS is commonly employed for basic topographic mapping of brain activation in applications like cognitive neuroscience and pediatric monitoring, where intensity-based measurements suffice for relative hemodynamic trends.Frequency-Domain fNIRS
Frequency-domain functional near-infrared spectroscopy (FD-fNIRS) utilizes intensity-modulated near-infrared light sources operating at radio frequencies, typically ranging from 100 to 1000 MHz, to investigate hemodynamic changes in biological tissue. This modulation introduces a sinusoidal variation in light intensity, and the detected signal at the receiver optode captures both the amplitude and phase alterations caused by tissue absorption and scattering, enabling separation of these effects for more quantitative analysis than simpler techniques.[30][31] The core measurements in FD-fNIRS include the direct current (DC) component, representing average intensity; the alternating current (AC) amplitude, indicating modulation depth; and the phase delay (φ), which reflects the temporal shift of the modulated signal. These parameters relate to the tissue's absorption coefficient (μ_a) and reduced scattering coefficient (μ_s') through models derived from photon diffusion theory, allowing absolute estimation of optical properties without relying solely on relative changes. Seminal work by Chance and colleagues in the late 1980s and early 1990s established the foundation for these measurements, with subsequent developments by Fantini et al. demonstrating quantitative absorption spectroscopy using phase data.[30][31] A key relation for the phase delay arises from the mean time-of-flight of photons, approximated under low-absorption conditions as \phi \approx \left( \frac{\omega}{c} \right) \left( \frac{L}{n} \right), where ω is the angular modulation frequency, c is the speed of light in vacuum, L is the effective optical pathlength, and n is the refractive index of the tissue. This approximation facilitates estimation of pathlength and scattering contributions, enhancing depth sensitivity.[30][31] Compared to continuous-wave fNIRS, FD-fNIRS offers significant advantages, including direct quantification of absolute μ_a and μ_s', which supports accurate chromophore concentration calculations (e.g., oxygenated and deoxygenated hemoglobin), and improved discrimination of superficial versus deeper tissue layers due to phase sensitivity to scattering dynamics. These features enable better rejection of systemic artifacts and higher spatial resolution in functional mapping. However, implementation demands specialized instrumentation, such as radio-frequency modulators for laser diodes and heterodyne detection schemes using avalanche photodiodes or photomultiplier tubes to resolve phase with high precision, leading to increased system cost and complexity relative to continuous-wave setups.[30][31] In practice, FD-fNIRS is particularly valuable in hybrid imaging systems, where its quantitative depth information complements other modalities like diffuse correlation spectroscopy to enhance localization of cerebral activation.[31]Time-Domain fNIRS
Time-domain functional near-infrared spectroscopy (TD-fNIRS) employs ultrashort laser pulses, typically lasting 50-200 ps, to probe tissue optical properties by measuring the time-of-flight of photons through scattering media. The time-resolved signal is captured using detection techniques such as time-correlated single photon counting (TCSPC), which achieves picosecond resolution by recording the arrival times of individual photons relative to the excitation pulse, or streak cameras, which temporally disperse light across a photocathode for high-speed imaging. These methods enable the reconstruction of the temporal point spread function (TPSF), representing the distribution of photon arrival times at the detector.[32][33] From the measured TPSF, the absorption coefficient (μ_a) and reduced scattering coefficient (μ_s') are derived by fitting to forward models based on the radiative transfer equation or its diffusion approximation, allowing absolute quantification without relying on pathlength assumptions inherent in simpler techniques. In the diffusion approximation for a semi-infinite homogeneous medium, the TPSF for reflectance R(ρ, t) at source-detector separation ρ and time t is modeled as: R(\rho, t) = \frac{c}{(4\pi D c t)^{3/2}} \int_0^\infty k \, J_0(k \rho) \exp\left( -\mu_a c t - D k^2 c t \right) \, dk where c is the speed of light in the medium, D = 1/(3 μ_s') is the diffusion coefficient, and J_0 is the zero-order Bessel function of the first kind; this integral form arises from the Hankel transform of the Green's function solution to the time-domain diffusion equation.[32] TD-fNIRS offers the highest accuracy for absolute chromophore concentration measurements, such as oxygenated and deoxygenated hemoglobin, and supports depth-resolved imaging up to approximately 4 cm by analyzing early- or late-arriving photons, which correspond to shallower or deeper tissue layers, respectively.[33][32] However, it faces significant challenges, including high costs due to picometer-scale timing electronics, limited laser repetition rates of 10-80 MHz that constrain signal averaging, and a requirement for dark environments to minimize ambient light interference.[33][32] Recent advances in the 2020s have focused on miniaturization, yielding wearable and wireless TD-fNIRS systems that integrate compact pulsed diodes, SPAD arrays for TCSPC, and low-power electronics, enabling portable functional brain monitoring with reduced form factor while maintaining temporal resolution.Diffuse Correlation Spectroscopy
Diffuse correlation spectroscopy (DCS) is an optical technique that complements functional near-infrared spectroscopy (fNIRS) by measuring cerebral blood flow through the analysis of temporal intensity fluctuations in diffusely scattered coherent light. It employs long-coherence-length near-infrared lasers, typically at wavelengths between 650 and 1064 nm, to illuminate tissue, where moving scatterers such as red blood cells induce speckle patterns in the backscattered light. These fluctuations are quantified using the normalized intensity autocorrelation function, g_2(\tau), computed from high-speed detection of the diffuse intensity over time delays \tau, providing a direct probe of microvascular perfusion dynamics.[34][35][36] The core analysis in DCS relies on the Siegert relation, which connects the intensity autocorrelation g_2(\tau) to the electric field autocorrelation g_1(\tau) via g_2(\tau) = 1 + \beta |g_1(\tau)|^2, where \beta is the optical coherence factor. For Brownian motion of scatterers in a diffusive medium, g_1(\tau) decays exponentially as: g_1(\tau) = \exp(-2 k^2 D_B \tau) Here, k is the wavevector magnitude, and D_B is the motion diffusion coefficient, which is proportional to blood flow velocity; the derived blood flow index (BFI) scales with D_B and serves as the primary metric for relative perfusion changes. This framework, originally developed for multiply scattered light, was first applied to biological tissues in seminal work establishing the correlation diffusion equation. Instrumentation for DCS includes single- or few-mode optical fibers for source-detector separations up to 4 cm, enabling depths of approximately 1-2 cm, along with photon-counting detectors such as avalanche photodiodes (APDs), photomultiplier tubes (PMTs), or single-photon avalanche diode (SPAD) arrays to achieve the necessary temporal resolution (e.g., microseconds) for real-time autocorrelation computation via hardware correlators or software algorithms.[34][35][37] Hybrid DCS-fNIRS systems integrate blood flow measurements with oxygenation assessments from fNIRS to estimate the cerebral metabolic rate of oxygen (CMRO₂) using models that relate perfusion, oxygen extraction, and hemoglobin concentrations, as demonstrated in early animal studies and subsequent human applications. DCS offers key advantages as a non-invasive method for real-time perfusion imaging, with high sensitivity to microvascular alterations that are often undetectable by other modalities, and it requires no exogenous contrast agents or ionizing radiation. However, its penetration is limited to superficial cortical regions due to strong tissue scattering, and it remains highly sensitive to motion artifacts from bulk tissue movement, necessitating advanced correction techniques such as short-time averaging or multi-distance probing.[34][35][38]Instrumentation
System Components
Functional near-infrared spectroscopy (fNIRS) systems rely on specific hardware components to emit, transmit, detect, and process near-infrared light for measuring brain hemodynamics. The core elements include light sources, detectors, supporting electronics, and optical interfaces, each tailored to ensure safe, reliable operation across various measurement techniques. Light sources in fNIRS systems primarily consist of light-emitting diodes (LEDs) or laser diodes, selected based on the required spectral properties and power output. LEDs provide broadband emission with lower power levels, making them suitable for portable, continuous-wave setups due to their cost-effectiveness and ease of integration. In contrast, laser diodes offer narrowband, coherent light, which is essential for techniques like diffuse correlation spectroscopy (DCS) that demand high temporal resolution for blood flow measurements. Wavelength selection typically involves pairs such as 780 nm and 850 nm, chosen to optimally differentiate absorption changes in oxygenated (HbO) and deoxygenated (HbR) hemoglobin based on their distinct spectral signatures in the near-infrared range (700–900 nm).[7][39][40] Detectors capture the attenuated light after its diffusion through tissue and must match the system's sensitivity needs. Photomultiplier tubes (PMTs) are employed in time-domain fNIRS for their superior temporal resolution and low noise, enabling precise photon time-of-flight measurements. For continuous-wave and frequency-domain systems, photodiodes are more common due to their simplicity and low-voltage operation, while avalanche photodiodes (APDs) enhance performance in low-light conditions through internal gain mechanisms, improving signal-to-noise ratios without excessive power consumption. Component choices, such as using PMTs for time-domain setups, are driven by the need for fast timing to resolve picosecond-scale photon paths.[7][41][42] Electronics form the backbone for signal generation, conditioning, and digitization. In frequency-domain fNIRS, modulation circuits produce radiofrequency signals (e.g., 100–200 MHz) to drive light sources and enable phase/amplitude detection. Time-domain systems incorporate precise timing electronics, often using time-correlated single-photon counting, to synchronize picosecond pulses. Across all configurations, low-noise amplifiers boost weak detected signals, and high-resolution analog-to-digital converters (ADCs, typically 16–24 bits) digitize them for subsequent processing, ensuring fidelity in capturing subtle hemodynamic changes.[43][44][45] Optical fibers and optodes facilitate light delivery and collection from the scalp. Multimode fibers, with core diameters of 200–400 μm, are standard for transmitting light efficiently over short distances while minimizing losses in diffuse scattering media like tissue. Optodes serve as the non-invasive interfaces, housing emitters and receivers, with typical integration times of 50–100 ms per channel to achieve adequate signal averaging without sacrificing temporal resolution for dynamic brain activity.[46][47] System scalability supports diverse applications, from compact portable devices with 8–16 channels for ambulatory studies to high-density arrays exceeding 100 channels for whole-head coverage and improved spatial resolution. Safety is paramount, with designs adhering to ANSI Z136.1 standards that limit skin irradiance to under 20 mW/cm² for prolonged exposures, preventing thermal effects while maintaining measurement efficacy.[48][49][50]Optode Configurations
Optode configurations in functional near-infrared spectroscopy (fNIRS) typically involve pairs of sources (light emitters) and detectors placed on the scalp, with standard source-detector separations of 2 to 4 cm to achieve adequate penetration into cortical tissue while minimizing signal attenuation.[51] Basic montages often use bilateral or unilateral arrangements, such as 8-16 channels over the prefrontal cortex for cognitive tasks, ensuring coverage of targeted regions without excessive complexity.[52] Topographic layouts align optodes with the international 10-20 electroencephalography (EEG) system, using anatomical landmarks like the nasion and inion to position probes systematically at 10% or 20% intervals along the scalp.[52] This alignment facilitates reproducible placement and integration with other neuroimaging modalities. High-density arrays extend this approach, incorporating multi-distance pairs to yield up to 204 channels for whole-head coverage, enhancing spatial resolution to approximately 1 cm.[53] Short-separation channels, with source-detector distances of 0.5 to 1 cm (commonly 8 mm), are integrated alongside long-separation pairs to capture superficial signals from the scalp and extracerebral vasculature, enabling regression of physiological noise like blood pressure fluctuations.[54] These channels improve signal quality by isolating cerebral hemodynamics, particularly in motion-prone setups.[55] Wireless and wearable designs, such as flexible caps and headbands, support mobility in naturalistic environments, with configurations like 27 channels in modular arrays for prefrontal or sensorimotor monitoring.[56] These systems, prominent in the 2020s, use lightweight optodes embedded in neoprene or elastic materials to maintain contact during movement.[57] Key considerations in optode placement include mitigating hair interference in dense or curly hair; techniques like brush optodes or comb attachments part hair to ensure consistent light delivery.[58] Adequate probe pressure is essential to avoid motion artifacts, while region-specific coverage, such as over the prefrontal cortex for executive function studies, prioritizes sensitivity to underlying gyri.[59] Recent advancements feature flexible printed optodes on circuit boards, enabling conformal fitting to the scalp and integration with virtual reality headsets for immersive neuroimaging, as demonstrated in systems with real-time 3D shape estimation for up to 64 channels.[60] [61]Data Processing and Analysis
Preprocessing Methods
Preprocessing in functional near-infrared spectroscopy (fNIRS) involves a series of steps to transform raw light intensity measurements into clean, interpretable signals representing cortical hemodynamics, while mitigating artifacts from motion, physiology, and instrumentation. These methods are essential due to the susceptibility of fNIRS to superficial and systemic noise, ensuring reliable downstream analysis of oxygenated (HbO) and deoxygenated (HbR) hemoglobin concentrations.[62][63] The initial step converts raw intensity data to optical density (OD), calculated as \Delta OD = -\log_{10}(I/I_0), where I is the detected intensity and I_0 the baseline, providing a logarithmic scale insensitive to absolute power variations. Subsequently, OD is transformed into relative hemoglobin concentrations using the modified Beer-Lambert law (MBLL), which accounts for light scattering in tissue via a differential pathlength factor (DPF):\Delta[\mathrm{HbO}] = \frac{ \epsilon_{\mathrm{HbR}}(\lambda_1) \Delta \mathrm{OD}(\lambda_2) - \epsilon_{\mathrm{HbR}}(\lambda_2) \Delta \mathrm{OD}(\lambda_1) }{ \mathrm{DPF} \cdot d \cdot [ \epsilon_{\mathrm{HbO}}(\lambda_1) \epsilon_{\mathrm{HbR}}(\lambda_2) - \epsilon_{\mathrm{HbO}}(\lambda_2) \epsilon_{\mathrm{HbR}}(\lambda_1) ] }
\Delta[\mathrm{HbR}] = \frac{ \epsilon_{\mathrm{HbO}}(\lambda_1) \Delta \mathrm{OD}(\lambda_2) - \epsilon_{\mathrm{HbO}}(\lambda_2) \Delta \mathrm{OD}(\lambda_1) }{ \mathrm{DPF} \cdot d \cdot [ \epsilon_{\mathrm{HbO}}(\lambda_1) \epsilon_{\mathrm{HbR}}(\lambda_2) - \epsilon_{\mathrm{HbO}}(\lambda_2) \epsilon_{\mathrm{HbR}}(\lambda_1) ] }
where λ₁ = 780 nm, λ₂ = 850 nm, ε denotes the molar extinction coefficients for oxygenated (HbO) and deoxygenated (HbR) hemoglobin, and d is the source-detector distance. This conversion, rooted in seminal work adapting the Beer-Lambert law for turbid media, enables quantification of hemodynamic changes but assumes constant DPF, often estimated age- or wavelength-dependently.[64][65][66] Motion artifacts, arising from head movements disrupting optode coupling, are corrected using techniques like principal component analysis (PCA), which decomposes signals into components and removes those correlating with motion epochs; wavelet-based filtering, which thresholds high-frequency noise in wavelet domains; or spline interpolation, which fits smooth curves to corrupted segments identified by abrupt intensity spikes. Comparative studies show spline and wavelet combinations often outperform PCA alone, recovering up to 90% of artifactual segments in infant data while preserving task-related signals. Targeted PCA variants further enhance specificity by focusing on motion-correlated variance.[67][68][69] Physiological noise from scalp hemodynamics, heartbeat, and respiration is addressed via short-separation regression, employing channels with <15 mm source-detector distances to isolate superficial signals subtracted from long-separation (>25 mm) cortical measures in a general linear model framework, reducing systemic variance by 20-50%. Additional bandpass filtering targets cardiac pulsations (0.6-2 Hz), respiration (0.15-0.5 Hz), and Mayer waves (0.01-0.1 Hz), with short-separation regressors improving activation detection in noisy recordings. Multimodal extensions, incorporating auxiliary signals like electrocardiography, further refine this by embedding canonical correlation analysis.[54][70][71] Detrending removes low-frequency drifts from baseline shifts or instrumental instability using linear or polynomial fits (typically 0-3rd order) subtracted from the time series, or temporal derivatives to emphasize changes while suppressing trends. These methods, applied post-artifact correction, stabilize signals for hemodynamic response estimation without over-smoothing task-evoked responses.[72][73] Quality control assesses signal integrity through signal-to-noise ratio (SNR) thresholds (e.g., >2 for acceptable channels) or spectral metrics like cardiac power, leading to rejection of noisy channels (often 10-20% in scalp recordings) based on coefficient of variation or correlation criteria to prevent bias in group analyses. Automated tools, such as those using machine learning on raw photodiode data, streamline rejection by classifying poor optode coupling.[74][75][76] Emerging 2020s approaches leverage deep learning for artifact removal, including autoencoder-based denoisers that learn motion patterns from paired clean-noisy data, achieving 15-30% better SNR than traditional methods without assuming artifact shapes, and LSTM architectures for sequential denoising of physiological confounds. Recent advancements as of 2025 include Transformer-based models, which leverage attention mechanisms for improved signal denoising and task classification, outperforming traditional deep learning in handling variable-length fNIRS sequences. These AI techniques, validated on diverse datasets, promise robust preprocessing for real-world applications like ambulatory monitoring.[77][78][79][80]