Functional Near-Infrared Imaging for Biomedical Applications

Abstract

Functional near-infrared spectroscopy (fNIRS) is utilized as an optical approach for biomedical applications, especially for the brain-computer-interfaces (BCIs) applications due to their absorption contrast between oxygenated hemoglobin (oxy-Hb) and deoxygenated hemoglobin (deoxy-Hb). In this chapter, we first make a brief introduction about the research background of fNIRS; then, the basic work principle of the fNIRS instrument was also reviewed, the performance of which was greatly affected by the light source (LEDs and lasers) and detectors (pin photodetector, avalanche photodiodes, and photomultiplier tube); afterward, we thoroughly introduce the fNIRS and hybrid fNIRS-EEG BCIs with a focus on the data classification methods, for instance, machine-learning (ML) algorithms and deep-learning (DL) algorithms, thereby forming better classification accuracies; lastly, challenges of fNIRS were pointed out, and an outlook was also made to foster the rapid research and development of this technology toward neuroscience and clinical applications.

1. Introduction

Neuroscience is a multidisciplinary field dedicated to understanding the nervous system, comprising the brain, spinal cord, and peripheral nerves, which spans a variety of scientific fields. Consequently, research in related aspects is crucial for advancing our understanding of the nervous system, improving human health, and enhancing our capabilities to interact with and understand the world [1, 2]. Moreover, it also bridges disciplines and continues to reveal the intricate mechanisms that govern our thoughts, behaviors, and experiences. With the development of neuroscience, numerous technologies were developed to help researchers understand the anatomy and operation of nervous system, such as magnetic resonance imaging (MRI) [3, 4], functional MRI (fMRI) [5], positron emission tomography (PET) [6], electroencephalography (EEG) [7], and fNIRS [8, 9, 10]. Compared with other technologies, fNIRS technology offers the advantages of non-invasive, safe, high spatial and temporal resolution, portable and flexible, real-time monitoring, cost-effectiveness, etc. Thus, many scientists are working on the fNIRS technology.

Typically, fNIRS instrument consists of a near-infrared (NIR) light source, optical detector, optical fibers, control system, and data acquisition system, which are employed to measure the variations of oxy-Hb and deoxy-Hb amounts. Since oxy-Hb and deoxy-Hb absorb the NIR light at dissimilar rate, relative concentration variations of oxy-Hb and deoxy-Hb were determined by quantifying the absorption amount and changes of NIR light in brain tissue, thereby reflecting the metabolism of cerebral cortex situation. Specifically, blood supply will rise when a certain brain area is active, thus increasing the oxy-Hb concentration and decreasing the deoxy-Hb concentration. fNIRS technology illuminates the cerebral cortex with NIR light and measures the absorption and changes of NIR light by the brain tissue. It can obtain the variations of oxy-Hb and deoxy-Hb while the brain is active, as well as analyze the metabolic activity and neural function of cerebral cortex; exhibiting fNIRS technology is a practical method that can be used in the brain-computer-interfaces (BCIs) applications [11, 12, 13].

In this chapter, fNIR technology with two or three light sources and detectors was always used for the BCI applications. Both the fNIRS instrument and fNIRS-based BCIs were developed to push the fNIRS to be used for non-invasive BCI applications, especially for drowsiness detection. Additionally, hybrid fNIRS-EEG-based BCIs present a promising strategy for assessing cerebral activity in BCIs.

2. fNIRS instrument

Since 1980s, the fNIRS technique has been found to be a feasible method to detect adult cortical oxygenation. Afterward, considerable work was done to develop the fNIRS instrumental prototypes and even products. Basically, there are several reasons why the fNIRS technique can be used for the neuroscience field: (I) human tissues exhibit a considerable level of transparency to the light ranging from 650 to 1000 nm; (II) NIR light is also susceptible to the pigmented compounds or scattering; (III) light ranging from 650 to 1000 nm possesses the capability to penetrate into the human tissues; (IV) comparatively elevated optical attenuation in the brain tissues. The optical principle of fNIRS is mainly based on the scattering of NIR light. When light penetrates the cerebral cortex, it is scattered and absorbed by the oxy-Hb and deoxy-Hb in the brain tissue. Since the two Hbs have different absorption rates of NIR light, relative concentration changes of oxy-Hb and deoxy-Hb were determined by evaluating the absorption amount and changes, thereby reflecting the metabolism status of the cerebral cortex.

Specifically, when a certain area of the brain is active, the blood supply to that area increases, causing the oxy-Hb concentration to surge while the deoxy-Hb concentration recedes. By measuring these Hb concentration changes, fNIRS can indirectly reflect the metabolic activity and neural function of the cerebral cortex. To gain a fully understanding of fNIRS workflow, a schematic diagram illustrating the path of NIR light as it travels through the human head has been presented in Figure 1 [14]. It is clearly observed that the emitter and detector are the main components of the fNIRS instrument.

Figure 1.

Schematic diagram of NIR light traveling through the head, including, scalp, skull, gray matter, detector, emitter, and path of detected light [14].

Generally, the fNIRS instrument uses two or more NIR light sources with specific wavelengths (typically at least one in the red region and the other in the infrared region), which was employed to differentiate the properties of oxy-Hb absorption and deoxy-Hb absorption (Table 1). The light sources used in the fNIR instrument primarily include LEDs and lasers. Among them, LEDs are widely employed in fNIR systems due to their advantages of low power consumption, long lifespan, and high stability. However, specific type of light source varies depending on instrument model, configuration, and application requirements. Compared with LEDs, lasers are less common in fNIR systems, which is maybe a better choice for higher sensitivity and higher resolution applications due to their ability to produce high-intensity, monochromatic, and highly directional beams. The selection of light source wavelengths is crucial in fNIR systems, which allows for good penetration through the scalp and skull while differentiating the characteristics of oxy-Hb absorption and deoxy-Hb absorption.

Light source type	LEDs	Lasers
Penetration depth	2 cm	3 cm
Portability	Yes	No (needs optical fibers)
Wavelength range	Broad	Narrow
Peak intensity	Low	High
Accuracy	Moderate	High
Price	Cheap	Expensive
Safety	Yes	Yes (Class I or Class II)

Table 1.

Comparison of LEDs and Lasers for fNIRS instrument.

The detector is used to receive NIR light that is reflected or transmitted after passing through the scalp and brain tissue. These NIR detectors are usually highly sensitive photodiodes, such as pin photodetector (PD) [15, 16, 17, 18, 19], avalanche photodiode (APD) [20, 21, 22], and photomultiplier tube (PMT) [23, 24, 25]. Compare with PD, APD possesses the internal gain mechanism, which can generate the secondary carriers and amplify the photocurrent through the avalanche multiplication effect. The internal gain mechanism enables the Si APDs to detect weaker optical signals with higher sensitivity, faster response speed, and exceptional signal-to-noise ratio (SNR). PMT is a specialized vacuum electronic device that converts weak optical signals into electrical signals and significantly enhances the intensity through multiple stages of multiplication amplification, which can detect extremely weak optical signals, shorter response time, broader spectral range (ultraviolet, visible light, and near-infrared), and exceedingly low noise. Despite the detector sensitivity, there are also several other possible factors that will affect the SNR during fNIRS imaging: (i) optical coupling between the light source and tissue; (ii) light scattering in the different brain tissues and subject; (iii) brain tissues and subject movement will disrupt the light path and also lead to the detected signal change; (iv) data processing method; (v) physiological noise, such as changes in blood flow or heart rate can also introduce the noise, which will also cause the SNRs.

Even there are also GaAs and InGaAs NIR detectors, SiGe(Sn) semiconductor is the most favorable choice [26, 27, 28, 29, 30]. Meanwhile, NIR detector choice also determines the sensitivity of fNIR instrument (Table 2).

NIR detector type	PDs	APDs	PMTs
Internal gain	No	10 to a few 100	reach up to 10^7
Sensitivity	Low	Higher than PDs	Meet gold standard
Speed	Fast	Faster than PDs	Faster than PDs
Portability	Yes (on head)	No	No
Voltage supply	Low	High	High
Safety	Yes	No	No
Cooling system	No	Yes	Yes
Dynamic range	100 dB	60 dB	60 dB

Table 2.

Comparison of NIR PDs, APDs, and PMTs for fNIRS instrument.

The typical interaction between NIR light and brain tissue usually consists of three possible paths: absorption, scattering, and transmission. The energy variation quantified through the absorption is specified as:

$$\mathit{\Delta E}=\mathit{hv}=h\frac{c}{\lambda }$$E1

In this equation, c represents the speed of the NIR light, λ denotes the NIR wavelength, and h stands for Planck’s constant. The correlation involving the absorption of oxy-Hb and deoxy-Hb at specific wavelengths can be derived from this equation. Figure 2 shows the absorption spectrum of oxy-Hb and deoxy-Hb that are important for the fNIRS; 690 nm and 830 nm NIR light were highlighted as the light source due to their absorption contrast for oxy-Hb and deoxy-Hb [31].

Figure 2.

Absorption characteristics of oxy-Hb and deoxy-Hb range from 650 to 900 nm [31].

The absorber property will affect the NIR optical routine as well as behaviors, which also have the Beer-Lambert law equation:

$$I=I_O{e}^{-\epsilon \left[C\right]L}$$E2

This means that light (I) attenuation is proportional to the product of absorber ([C]) and optical routine distance (L). The light attenuation (A) can be obtained by applying the logarithm and calculating the reciprocal of the output divided by the input.

$$A=log\left(\frac{I_O}{I}\right)=\epsilon \left[C\right]L$$E3

To quantify the absorption variations of oxy-Hb and deoxy-Hb concentration, Beer-Lambert law was modified as the following equation:

$$\mathit{\Delta OD}=\mathit{\text{εΔCLB}}$$E4

In this equation, ε represents the extinction coefficient, ΔC indicates the variation of oxy-Hb and deoxy-Hb concentration, L denotes the light source-detector distance, and B refs to the differential path-length factor (DPF).

3. fNIRS-based BCIs

3.1 fNIRS-BCI for drowsiness detection

In the realm of neurotechnology, BCIs offer unprecedented opportunities for enhancing human tissue cognition, treating neurological disorders, and facilitating communication beyond natural boundaries. BCIs also establish the direct communication link between human brain and external device simultaneously, which enables the transfer of information in both directions. Based on whether craniotomy surgery is performed, BCIs are classified into invasive and non-invasive. Invasive BCIs involve placing chips directly on the brain’s cortex, offering high signal precision but requiring a craniotomy. While modern minimally invasive procedures can achieve small incisions of just one or two centimeters, introducing foreign objects into the body can trigger immune responses. Over time, electrodes may become encapsulated, leading to signal loss, and there is also infection risk. Non-invasive BCIs place electrodes on the wearable caps, with signal strength generally lower than invasive methods, but surgery is not involved. In clinical settings, particularly for treating epilepsy, invasive BCIs are currently predominant and well-developed, demonstrating significant efficacy. Meanwhile, commercially available non-invasive BCI products are designed for improving sleep or monitoring fatigue during driving, indicating non-invasive BCI technology can be widely used in our daily life without craniotomy surgery risk and complexities, especially for special populations, such as senior citizens, children, and pregnant woman. To date, there are three categories of non-invasive BCIs: EEG, fNIRS, and fMRI. Compared with EEG and fMRI, fNIRS offers unique advantages, particularly in terms of low noise, real-time imaging, security, portability, and economic viability, making it valuable for the drowsiness state application.

To reduce the safety risks associated with fatigue driving, fNIRS-based BCIs were proposed to monitor the drowsiness state of the driver. Studies have shown that drowsiness often manifests in the prefrontal cortex (PFC) while operating a vehicle. The main body of studies is primarily focused on identifying the neural associations crucial for drowsiness detection. Meanwhile, it is crucial to study both the hemodynamic response characteristics and neuronal responses associated with drowsiness to prevent false alarms. For this reason, Khan and Hong feasibility of drowsiness detection was studied by leveraging hemodynamic brain activity in an fNIR-based BCI setup [32]. The drowsiness indicators were captured with continuous-wave (CW) fNIRS technology, the operation wavelengths are 760 nm and 830 nm, respectively (Figure 3).

Figure 3.

The positioning of source, detector across the regions of the prefrontal and dorsolateral prefrontal cortex [32].

Signals were captured using a 28-channel NIRS setup; the numbers of source and detectors in each channel were 7 and 16, respectively. Specifically, channels 1–8 recorded the statistics from the right side of the dorsolateral prefrontal cortex (DPFC), while channels 21–28 recorded data from the left side of the DPFC. All the channels were categorized into three regions: channels 1–8 constituted region A, channels 9–20 were designated as region B, and region C encompassed channels 21–28. The modified Beer-Lambert law [33] was employed to transform the unrefined data into concentration changes of the oxy-Hb and deoxy-Hb (∆HbO and ∆HbR), which was expressed as:

$$A\left(t;\lambda \right)=ln\frac{I_{\mathit{\text{in}}}\left(\lambda \right)}{I_{\mathit{\text{out}}}\left(t;\lambda \right)}=\alpha \left(\lambda \right)\times c\left(\lambda \right)\times l\times d\left(\lambda \right)+\eta$$E5

$$\left[\begin{array}{l}\Delta c_{\mathit{HbO}}\left(t\right)\\ \Delta c_{\mathit{HbR}}\left(t\right)\end{array}\right]=\frac{\mathit{\Delta A}\left(t;{\lambda }_1\right)}{\left[\begin{array}{cc}{\alpha }_{\mathit{HbO}}\left({\lambda }_1\right)& {\alpha }_{\mathit{HbR}}\left({\lambda }_1\right)\\ {\alpha }_{\mathit{HbO}}\left({\lambda }_2\right)& {\alpha }_{\mathit{HbR}}\left({\lambda }_2\right)\end{array}\right]\cdot l\times d\left(\lambda \right)\cdot \mathit{\Delta A}\left(t;{\lambda }_2\right)}$$E6

where A is the light absorbance, I_in corresponds to the density of incident light, I_out denotes the measured light flux, α denotes extinction coefficient at a certain wavelength, c refers to the absorption concentration, l indicates the source-detector distance, d represents differential path-length factor (DPF), and η signifies light attenuation because of light scattering. An essential consideration in feature extraction is determining the optimal temporal window size for effectively capturing drowsiness and alert states’ distinctive features. Based on all these equation and calculation methods, three distinct time windows (0–5, 0–10, and 0–15 seconds) were extensively explored. For each segment, they analyzed eight different characteristics by averaging the signal within different regions.

Figure 4a illustrates the average variations in classification accuracy across different time windows within region A, region B, and region C, indicating region A consistently exhibits higher average accuracy compared to the other regions (regions B and C). Although there is a slightly decrease in the mean accuracy in the 0–5 sec time interval when the sec time window was changed to 0–15, it remains above 70%, making the 0–5 sec window the most suitable choice among the three for fNIRS-based BCIs applications. Figure 4b presents the mean and standard deviation across subjects within the three regions. These results are averaged across all the time intervals for each subject in different regions. Broadly, drowsiness activity detection is more effective in region A. Figure 4c illustrates the statistics variability across the channels for all subjects, thereby highlighting higher signal variations in region A. This study demonstrated that the right DPFC yielded higher drowsiness and alert state accuracy compared to the PFC and DPFC regions on the left side. In the PFC region, mean accuracies across the second time windows of 0–15 to 0–5 range from 84.9% to 64.4%. The t-tests were employed to assess the significance of accuracies. Compared with the precisions in other regions, region A yields the p-values of 0.0001 for both 0–5 seconds and 0–10 second time windows, indicating notable activities are clearly observed from channels 1 to 8 in the drowsy state detection. Furthermore, no substantial discrepancies in mean precisions in the midst of the different distinct time intervals were observed in Figure 4a, suggesting that the time intervals with 0–5 seconds range are more suitable for identifying the drowsiness via fNIRS-based BCIs. It should be noted that support vector machines (SVM) method was employed to improve the accuracy of classification. Generally, ML algorithms and DL algorithms were used to compute the classification accuracy. This part gives a detailed introduction to different types of ML and DL algorithms.

Figure 4.

(a) Mean classification accuracies across 13 subjects for each time window (0–5, 0–10, and 0–15 seconds); (b) mean classification accuracies and variability across the subjects in different regions; (c) occurrence count of drowsy state across 13 subjects (e.g., Chapter 1 exhibited the drowsy state in four subjects) [32].

There several types of ML algorithms were developed, including SVM, k-Nearest Neighbor (k-NN), and Linear Discriminant Analysis (LDA) [34, 35, 36, 37].

SVM is widely recognized for processing the data from the fNIRS-BCI system. Hyperplanes, created by SVM classifier, were employed to maximize the separation distance from the nearest training points. The separating hyperplane is expressed as:

$$f\left(x\right)=r\cdot x+b$$E7

where b functions as the factor for scaling, while r, x are the elements of R², b belongs to the R¹. The equation for r* (optimal solution) is formulated as:

$$\frac{1}{2}{‖w‖}^2+C\sum _{i=1}^n{\xi }_i$$E8

$$y_i\left(w^T{x}_i+b\right)\ge 1-{\xi }_i,\xi \ge 0$$E9

y_i stands for the class label of the sample, T signifies the transpose operation, and n is the aggregate samples quantity, ||w||2 = w^Tw, w^T and x_i ∈ R², b ∈ R¹, C serves as a parameter for trade-off, and ξ_i denotes the error for the training.

In the k-NN approach, the advantages of minimal computational demands and straightforward implementation make it popular in the fNIRS-based BCIs. The following equation given the Euclidean distance:

$$D_E\left(p,q\right)={\left[\sum _{i=1}^n{\left(p_i-q_i\right)}^2\right]}^{1/2}$$E10

where n denotes n-dimensional space, p and q represent the two points in n-dimensional space, and the corresponding two pairs of vectors are denoted as p_i and q_i, respectively.

Normally, discriminant hyperplanes were utilized by LDA to effectively distinguish the different classes, the advantages of simplicity and rapid execution make LDA well-suited for different types of BCI systems. Fisher’s criterion is also optimized to minimize the intra-class variance and maximize the inter-class separation, the equation for which is given as follows:

$$J\left(v\right)=\raisebox{1ex}{$v^T{S}_{b}v$}\!\left/ \!\raisebox{-1ex}{$v^T{S}_{w}v$}\right.$$E11

S_b and S_w refer to the scatter matrices that account for between-class and within-class variations, respectively, which were given by:

$$S_b={\left(m_1-m_2\right)}^{T+1}$$E12

$$S_w=\sum _{x_n\in C_1}\left(x_n-m_1\right){\left(x-m_2\right)}^T+\sum _{x_n\in C_2}\left(x_n-m_1\right){\left(x-m_2\right)}^T$$E13

Samples are represented as x_n, m₁ denotes the class mean for groups C₁, while m₂ denotes the class means for C₂. In the ML algorithms, peak accuracies of SVM, k-NN, and linear LDA are 78.90, 77.01, and 66.70%, respectively [38], suggesting that SVM provides the best peak accuracy (Figure 5).

Figure 5.

ML classification accuracies of SVM, k-NN, and LDA algorithms in nine subjects [38].

3.1.1 DL algorithms

The data biases and data overfitting are the drawbacks encountered in the fNIRS-based BCIs classification with ML algorithms, which also consumes a lot of times. To avoid the above-mentioned problems, DL algorithms are one of the most promising methods to extract the appropriate features of complex fNIRS-BCI signals. Typically, DL algorithms are classified as convolutional neural networks (CNNs), long short-term memory (LSTM), and bidirectional LSTM (Bi-LSTM) [38, 39, 40]. This part introduces the three types of DL algorithms.

CNNs are convolutional neural networks that were designed for autonomously extracting the meaningful characteristics from the fNIRS-based BCIs statistics, which comprise the four types of layers. fNIRS-based BCI statistics, representing the variations for the oxy-Hb concentrations across the channels, are processed using the CNNs method. Within the convolutional layer, convolution kernels extract the features. CNNs were able to enhance the classification accuracy by iteratively improving filter weight across the whole propagation. Mathematically, the convolutional process is represented as:

$$C_t=f\left(w\ast X_{t-1/2:t+l/2}+b\right)$$E14

In this equation, C represents convolving output, T denotes filter, X denotes length, ω denotes length, l and b represent the parameters with bias, f denotes non-linear activation function.

LSTM and Bi-LSTM: LSTM is another type of DL algorithm, which can classify the fNIRS data with high accuracy, processing, and forecasting. The internal mechanism includes forget, input, and output gates. Here are the equations for all the gates:

$$f_t=\sigma \cdot \left(W_f\cdot \left[h_{t-1,x_t}\right]+b_f\right)$$E15

$$i_t=\sigma \cdot \left(W_i\cdot \left[h_{t-1},x_t\right]+b_i\right)$$E16

$${\mathrm{o}}_t=\sigma \cdot \left(W_o\cdot \left[h_{t-1},x_t\right]+b_{\mathrm{o}}\right)$$E17

Here, W_f, W_i, and W_o denote weight matrices associated with three gates, h_t − 1 represents the concealed state. All these three gates are utilized to regulate the values flow throughout the network [41]. The function of the sigmoid represents as follows:

$$f\left(x\right)={\left[1+e^{-k\left(x-x_O\right)}\right]}^{‐1}$$E18

Here, x_o represents the midpoint of the x-value for sigmoid function, e denotes the natural logarithm base, and k denotes the rate of growth. The bidirectional LSTM (Bi-LSTM) integrates forward and backward LSTM networks together [42], which makes Bi-LSTM operate better than LSTM networks. For ML algorithms, the peak accuracies of CNN, LSTM, and Bi-LSTM are 95.47, 95.35, and 95.54%, respectively, suggesting that DL algorithms provide higher peak accuracy than ML algorithms (Figure 6).

Figure 6.

DL classification accuracies of CNN, LSTM, and Bi-LSTM algorithms in nine subjects [38].

4. Hybrid fNIRS-EEG-based BCIs

Both fNIRS and EEG serve as beneficial tools to monitor brain activities in BCI applications. fNIRS detects the changes in NIR light intensity following its moves across the scalp and cerebral tissue, providing insights into the oxy-Hb and deoxy-Hb activity. Different from fNIRS, EEG captures the electrical signals from groups of neurons over short periods by placing electrodes over the head. Currently, fNIRS as well as EEG are widely utilized in BCIs on account of the non-invasive feature, cost-effectiveness, portability, and suitability for prolonged monitoring. Nevertheless, fNIRS as well as EEG have their own benefits and drawbacks in terms of temporal resolution, spatial resolution, spatial specificity, motion artifacts, sensitivity, etc. Therefore, hybrid fNIRS-EEG-based BCIs present a promising strategy to provide a comprehensive assessment of cerebral activity in BCIs [43, 44, 45, 46].

With the rapid expansion of consumer electronics, there has been a corresponding surge in research exploring the integration of fNIRS and EEG technologies. This convergence has led to various innovative methodologies aimed at enhancing BCI systems. One prominent approach involves the development of full-head caps that seamlessly combine optodes for fNIRS and electrodes for EEG into a single integrated setup (Figure 7A). This design not only facilitates simultaneous data acquisition but also ensures spatial alignment of measurements, which is crucial for accurate fusion and interpretation of fNIRS-EEG data [47].

Figure 7.

Hybrid fNIRS-EEG-based BCIs. (A) Schematic diagram of optodes and electrodes; (B) Optodes and electrodes positioning of fNIR and EEG; (C) Hybrid fNIRS-EEG-based BCLs robot system [47].

Alternatively, researchers have explored using distinct modalities in different scalp regions (Figure 7B), exploiting the complementary strengths of fNIRS and EEG. This spatial separation allows for specialized measurements optimized for each modality’s strengths: fNIRS excels in providing spatially resolved hemodynamic responses, while EEG offers high temporal resolution ideal for capturing rapid neural dynamics. Beyond these integrated and segregated approaches, some BCIs utilize the EEG signals to decode brain activities while simultaneously employing an fNIRS device to monitor the cortical activations (Figure 7C). This hybrid method capitalizes on the strengths of both technologies, aiming for enhanced accuracy and robustness in BCI applications. For instance, fNIRS can provide supplementary information about cortical oxygenation levels, which can complement EEG’s electrical signals in decoding cognitive states or motor intentions.

In the realm of BCIs employing fNIRS-EEG hybrids, ML and DL algorithms are crucial for the data processing and classification. These algorithms are applied to preprocess raw data, extract relevant features, and classify neural patterns associated with specific tasks or mental states. Figure 8 illustrates a detailed procedural overview of how these algorithms are implemented in practice, showcasing the step-by-step data flow and decision-making processes involved in fNIRS-EEG BCI systems. Moreover, the integration of ML and DL in fNIRS-EEG BCIs underscores a broader trend toward leveraging advanced computational techniques for neuroscientific research and clinical applications. These algorithms improve both the accuracy of neural signal classification and real-time feedback, enabling adaptive interfaces that can adjust to user-specific neural responses [47].

Figure 8.

Data processing procedure in hybrid fNIRS-EEG-based BCIs [47].

Hybrid fNIRS-EEG technologies in BCIs represent a promising frontier in neuroscience and neuroengineering. By combining the spatial specificity of fNIRS with the temporal dynamics of EEG, researchers are advancing toward more sophisticated and effective brain-computer interface systems. These developments hold potential implications not only for assistive technologies but also for enhancing our understanding of brain function and cognition across diverse populations and applications. As research continues to evolve, further innovations in sensor design, signal processing techniques, and algorithmic approaches are expected to drive the future growth and adoption of fNIRS-EEG hybrid systems in both research and clinical settings.

5. Conclusions

In conclusion, fNIRS technology are facing the problems of limited SNR, limited penetration depth, depth sensitivity, lower SNRs, data processing algorithms dependent, and limited spatial resolution, etc. To solve the above-mentioned problems, future fNIRS research could be improved from the following aspects: Firstly, improve the light source and detector design, especially for the light source and detector operating at the short-wave infrared (SWIR) and long-wave infrared (LWIR) range, which can increase the penetration depth in the brain tissues [48, 49, 50]. Secondly, develop more advanced optics technologies and improve the depth sensitivity. Thirdly, the optical imaging method and spatial resolution should be improved. Lastly, fNIRS was combined with other imaging techniques (such as fMRI and EEG) to achieve a more comprehensive assessment of brain function.

fNIRS technology can be classified as continuous-wave (CW) mode and pulsed mode, but they exhibit significant differences in terms of light source emission methods, data processing, analysis, and application scenarios. Researchers should select the appropriate fNIRS mode based on their specific research needs and experimental conditions. CW fNIRS is widely applied in laboratory research, educational psychology, health psychology, engineering psychology, and clinical medicine due to its portability, low cost, and ease of use. It is particularly suitable for scenarios requiring long-term and repeated measurements, such as cognitive neuroscience studies in infants and special populations. While more complex and costly, pulsed fNIRS excels in applications that require high precision and spatial resolution, which is more suitable for studies that demand detailed optical parameter information, such as fine brain region differentiation and neural activity localization in advanced cognitive neuroscience research.

Acknowledgments

The authors would like to acknowledge the support from Guangdong Province Key Fields Research and Development Plan Project (Grant No. 2024B0101130001), and “Pearl River Talent Plan” Innovation and Entrepreneurship Team Project of Guangdong Province (Grant No. 2021ZT09X479).

Conflict of interest

The authors declare no conflict of interest.