2.1.

Subjects

30 young adults (15 females, age $22.6\pm 2.46\text{years}$ ) with no history of neurological diseases were recruited from Beijing Normal University. All subjects were right handed as assessed by the Edinburgh Handedness Inventory.¹³ Before the experiment, informed consent was obtained according to the procedure approved by the Review Board at the State Key Laboratory of Cognitive Neuroscience and Learning of Beijing Normal University.

2.2.

Functional Near-Infrared Spectroscopy Measurement

The experiment consisted of an $11\text{-}\min$ resting-state measurement session followed by a $7\text{-}\min$ localizer task measurement session. To avoid the confounding effect of the task on the resting-state measurement, subjects did not know the task prior to the experiment. During the resting-state measurement session, subjects were instructed to remain still with their eyes closed, staying as motionless as much as possible, and to perform no specific cognitive exercise. During the task measurement session, a blocked design task was performed to induce task-related signal changes and help locate language cortices. In brief, subjects were instructed to overtly generate verbs immediately after hearing auditorily presented concrete nouns in eight $17.5\text{-}\mathrm{s}$ task periods alternating with $30\mathrm{s}$ of rest (viewing a fixed “+”). The concrete nouns were matched in aspects of conceptual familiarity, word frequency, and verb-generation difficulty as assessed by another 29 subjects who were not involved in the present study. The stimuli were presented and synchronized with the fNIRS equipment using E-prime software (v.1.2, Psychological Software Tools, Pittsburgh, Pennsylvania).

During fNIRS measurement, subjects were seated in a chair in a silent room with dim lighting. fNIRS measurements were conducted using an ETG-4000 optical topography system (Hitachi Medical Company, Tokyo, Japan). The absorption of near-infrared light at two wavelengths (695 and $830\mathrm{nm}$ ) was measured with a sampling rate of $10\mathrm{Hz}$ . Based on the modified Beer-Lambert law,¹⁴ the concentration changes in oxygenated hemoglobin (oxy-Hb) and deoxygenated hemoglobin (deoxy-Hb) for each channel were obtained. Because the previous studies by Hoshi¹⁰ and our group¹¹ both showed that oxy-Hb was the most sensitive indicator of changes in regional cerebral blood flow in fNIRS measurements, this study focused on the concentration changes of oxy-Hb.

Two $3\times 5$ optode probe sets (eight emitters and seven detector probes, 22 measurement channels, and $30\text{-}\mathrm{mm}$ optode separation) were used in this study. The probe holders were placed over the bilateral frontal and temporal area, with the left channel 21 (left CH 21) above T3 and the right channel 3 (right CH 3) above T4 [Fig. 1 ] in accordance with the international 10–20 system.¹⁵ Moreover, an additional structural MR image was scanned from one subject to assist in anatomical localization of the fNIRS measurements. In Fig. 1, all of the measurement channels marked by vitamin E capsules are visible. After the MR image was normalized into the Montreal Neurological Institute (MNI) standard brain space using SPM5 software (see http://www.fil.ion.ucl.ac.uk/spm), a balloon-inflation model¹⁶ was used to project the measurement channels (i.e., marked by the vitamin E capsules) onto the cortical surface. The estimated locations were then labeled according to the widely adopted automated anatomical labeling (AAL) template¹⁷ and marked in color, as shown in Fig. 1.

Fig. 1

Anatomical positions of the fNIRS measurements and the activation/RSFC map. (a) The structural MRI image with the probe holder. All of the measurement channels are marked by vitamin E capsules. (b) Anatomical position of each measurement channel. Red represents the channels covering the regions of interest (i.e., the IFC and STC). (c) Group-level $t$ maps of the language localizer task. (d) Group-level RSFC map from the language seed channel. (e) and (f) Group-level RSFC maps from both control seed channels (CH6 and CH5). Blue circles indicate the seed channels. (Color online only.)

2.3.

Data Analysis

During data preprocessing, five subjects were excluded from any further analysis because they either showed large head movements, or the data were too noisy due to hair obstruction. The following analysis was performed using Matlab (Mathworks version 7.1, Natick, Massachusetts) and NIRS-SPM software.¹⁸

2.3.2.

Resting-state functional connectivity analysis

For measurement of resting-state data, the first $20\mathrm{s}$ and the last $40\mathrm{s}$ were discarded for steady-state control. The data were then bandpass filtered to extract the low frequency fluctuation (LFF) signals $\left(0.01\text{to}0.08\mathrm{Hz}\right)$ .²³ Individual RSFC estimation was then performed using the GLM approach, in which the time course of the seed channel was used as the independent variable. The autocorrelation of the time course was taken into consideration as mentioned before. The estimated regression parameter (i.e., $\beta$ ) indicated to what extent the LFFs of each channel correlated to that of the seed channel. Finally, a group-level random effect analysis was conducted on the individual $\beta$ images to identify the RSFC map of the language system.

To quantitatively compare the RSFC strength inside the language areas with that outside, a region of interest (ROI) analysis was further conducted. Four ROIs were selected based on their anatomical localization [see Figs. 1 and Fig. 2 ], with two inside and two outside the language system. The two language ROIs (L-ROIs) were the left IFC (L-ROI-IFC, left CH 15, 16, and 20) and STC (L-ROI-STC, left CH 13, 17, and 21), respectively. The two ROIs outside of the language system (OL-ROIs) were located in the left middle frontal cortex (MFC). One OL-ROI (OL-ROI-adj, left CH 6 and 10) was adjacent $(⩽3\mathrm{cm})$ to the dominant language areas, while the other OL-ROI (OL-ROI-rem, left CH 1 and 5) was spatially remote from the dominant language areas [ $>4\mathrm{cm}$ , see Fig. 2]. For each subject, the connectivity strength of each ROI to the seed channel was defined as the maximum $\beta$ value in corresponding ROI.²⁴ Finally, one-way ANOVA analysis and post hoc tests were performed to detect differences in connectivity strength between these ROIs (SPSS v15, Incorporated, Chicago, Illinois).

Fig. 2

ROI analysis results. (a) Schematic position of each ROI. (b) The RSFC strengths of each ROI across all subjects. The RSFC strengths of these four ROIs were significantly different [ $F(3,72)=22.263$ , $p=2.64\times {10}^{-10}$ ]. Error bar indicates standard error. (c) Post hoc tests between each pair of ROIs. For (c), gray color, $p<0.05$ , and dark gray, $p<0.001$ (Bonferroni corrected).

3.1.

Localizer Task Results

The activation map of the language localizer task is shown in Fig. 1. From the figure, the strongest foci of activation were found in the IFC and STC of the left hemisphere. The right homologous areas also showed activations. Because the left CH 11 within the left IFC showed the peak activation, it was defined as the seed channel for the RSFC analysis in the language system.

3.2.

Resting-State Functional Connectivity Results

Figure 1 shows the resulting RSFC map of the language system. A significant correlation was found between the RSFC map and the language task-induced activation map, as shown in Fig. 1 [ $r=0.495$ , $p=7.32\times {10}^{-4}$ , see Fig. 3 ], suggesting the similar functional anatomy of both the spontaneous and evoked brain activity related to the language system. ROI analysis results are shown in Figs. 2 and 2. The RSFC strengths of the four ROIs to seed channel were significantly different [ $F(3,72)=22.263$ , $p=2.64\times {10}^{-10}$ ]. Specifically, the RSFC strengths of the two L-ROIs were significantly stronger than that of the two OL-ROIs ( $t=4.123$ , $p=0.0023$ between the L-ROI-IFC and OL-ROI-adj; $t=5.711$ , $p=4.18\times {10}^{-5}$ between the L-ROI-IFC and OL-ROI-rem; $t=3.864$ , $p=0.0045$ between the L-ROI-STC and OL-ROI-adj; and $t=5.218$ , $p=1.44\times {10}^{-4}$ for the L-ROI-STC and OL-ROI-rem, Bonferroni corrected).

Fig. 3

Similarity to the language task activation map. (a) The language system RSFC map ( $r=0.495$ and $p=7.32\times {10}^{-4}$ ). (b) The control RSFC maps ( $r=0.0072$ and $p=0.963$ for left CH 6, and $r=-0.160$ and $p=0.307$ for left CH 5). A dot in the scatter plot indicates a measurement channel. The solid lines are the linear regression lines.

3.3.

Lateralization Index Results

As shown in Fig. 4 (left), the average LI of the language RSFC across subjects was $+0.381$ , indicating a relatively strong left side lateralization. Because the left seed channel can bias the resulting language RSFC map toward the left hemisphere due to spatial correlation, the LI was also computed for the RSFC map derived using left CH 5 (outside the language system) as the control seed channel. Although the average control LI $(+0.138)$ showed a weak left side lateralization, as shown in Fig. 4 (right), the language RSFC LI was significantly stronger than that of the control RSFC map (paired $t$ -test, $t=2.418$ , $p=0.024$ ).

Fig. 4

Lateralization indexes of the language RSFC (left bar) and the control RSFC (right bar) across subjects. Error bar indicates standard error. The difference between the two LIs was statistically significant (paired $t$ -test, two-tailed, $t=2.418$ , $p=0.024$ ).

3.4.

Validity and Reliability of the Language Resting-State Functional Connectivity

As shown in Figs. 1 and 1, the RSFC maps from both control seed channels (i.e., left CH 6 and left CH 5) showed different patterns from the language RSFC map ( $r=-0.030$ and $p=0.851$ for left CH 6, and $r=-0.325$ and $p=0.036$ for left CH 5). These control RSFC maps were also significantly different from the language task activation map ( $r=0.0072$ and $p=0.963$ for left CH 6, and $r=-0.160$ and $p=0.307$ for left CH 5) [see Fig. 3]. These results indicated that the revealed fNIRS-based RSFC is specific to the language system rather than a general phenomenon.

The RSFC results based on the two randomly generated subgroups of subjects are shown in Fig. 5 . The resulting group-level RSFC maps were fairly similar to each other (Pearson correlation: $r=0.709$ and $p=1.03\times {10}^{-7}$ ). Similarly, the RSFC results generated from the first and second half of the fNIRS time courses are shown in Fig. 6 . They were also similar to each other (Pearson correlation: $r=0.856$ and $p=2.71\times {10}^{-13}$ ), and both are similar to the map computed from the full length time course (Pearson correlation: $r=0.892$ and $p=1.03\times {10}^{-15}$ for the first half; $r=0.982$ and $p=3.52\times {10}^{-31}$ for the second half). Moreover, these consistencies also existed at the individual subject level, as shown Fig. 6. Thus, all of these results indicated a high level of reliability of the fNIRS-based RSFC results.

Fig. 5

Consistency between the RSFC results derived from two randomly generated subgroups of subjects, evaluated by (a) group-level RSFC maps and (b) linear regression analysis. Blue circle indicates the seed channel. A dot in the scatter plot indicates a measurement channel, and the solid line is the linear regression line. (Color online only.)

Fig. 6

Temporal reliability of the language system RSFC map. The RSFC maps are computed from (a) the two split-half datasets, and (b) the group- and (c) subject-level linear regressions between the two results. Blue circles indicate the seed channel. A dot in the scatter plot indicates a measurement channel, and the solid lines are the linear regression lines. (Color online only.)

4.1.

Feasibility of Investigating Language Resting-State Functional Connectivity by the Functional Near-Infrared Spectroscopy Technique

Although fNIRS has been shown to be able to detect RSFC in the primary sensory and motor systems,^{11, 12} whether this technique can be used to detect RSFC in high-level neural systems remains unclear. The present study, for the first time, demonstrated the feasibility of investigating RSFC by fNIRS in a high-level neural system (i.e., the language system).

4.2.

Strengths and Limitations of the Functional Near-Infrared Spectroscopy Technique for Investigating Resting-State Functional Connectivity

The use of fNIRS has several advantages over fMRI for RSFC studies, especially for developmental and clinical studies. First, fNIRS is safe, tolerant of movement, and can accommodate a realistic environment. These attributes make fNIRS an ideal choice for repeated, short-interval measurement and lengthy continuous recording on special populations such as neonates, infants, and intensive care patients who are not amenable to fMRI. Second, fNIRS can be conducted in silence, which makes fNIRS more suitable for resting-state measurements during RSFC studies. Third, fNIRS has a high temporal sampling rate $\left(10\text{to}100\mathrm{ms}\right)$ . It further prevents aliasing of higher frequency physiological interferences into LFFs and makes a more reliable RSFC estimation. Moreover, it provides information about more specific physiological variables (e.g., oxy-Hb, deoxy-Hb, and total-Hb). This attribute allows fNIRS to provide a more comprehensive understanding of RSFC than fMRI. Finally, the optical signal does not interfere with electrical or magnetic fields. Thus, fNIRS is ideally compatible with other technologies, such as EEG and TMS. This attribute makes it possible to provide multiple indices of RSFC in neurological and physiological aspects.

fNIRS also has several potential limitations. First, the measurement area of fNIRS is limited to the outer cortex, because the typical depth sensitivity of most fNIRS systems is about $25\text{to}30\mathrm{mm}$ below the scalp due to the diffusion of photons. Thus, fNIRS is unsuited for investigating brain activation in deeper regions, such as the basal ganglia and amygdale. This constrains functional integration studies related to these deeper structures using the fNIRS-based RSFC approach. Second, most currently available fNIRS systems can only cover part of the brain, even with a multichannel fNIRS system. This makes it difficult to explore the complex connectivity pattern among all of the cortical neural systems. Third, fNIRS has a spatial resolution on the order of $10\text{to}20\mathrm{mm}$ , which is larger than fMRI $\left(1\text{to}3\mathrm{mm}\right)$ . The limited resolution may hinder the specificity of fNIRS-based RSFC results. Finally, the detected fNIRS signals contain significant noise that is related to systemic physiological hemodynamic fluctuations, such as cardiac pulsation, respiratory signals, and blood pressure changes. Eliminating the adverse effect of instrumental noise and physiological interferences on fNIRS-based RSFC is still a challenge.

4.3.

Future Works

As mentioned before, the resting-state fNIRS measurements reflect not only the cerebral hemodynamic fluctuations related to spontaneous brain activity but also hemodynamic changes of systemic origin. Thus, reducing the influence of physiology interference and improving the specificity of fNIRS-based RSFC is an important direction for future fNIRS studies. Moreover, only the language system RSFC was investigated in the present study. Additional experiments are needed to understand other high-level functional systems and further extend the application fields of fNIRS-based RSFC.