Abstract
The non-invasive delivery of laser light to the brain, termed transcranial photobiomodulation (tPBM), has been shown to stimulate vascular and metabolic function in the rodent brain. Given the large number of neurological disorders marked by impaired blood flow and metabolism, it is important to objectively evaluate the reported effects in humans. To date, however, neurophysiological evidence of tPBM effects in the human brain is scarce. Here we combined tPBM with functional magnetic resonance imaging (fMRI) by recording the blood-oxygenation-level dependent (BOLD) signal before, during, and after application of tPBM to the right forehead of 20 healthy participants. We found both acute and outlasting BOLD increases as high as 31% in regions of the prefrontal cortex within 30 mm of the site of light incidence. The observed effect was consistent with rapid inflow of blood to the stimulated brain region, coupled with an increase in oxidative metabolism. Our findings underscore the therapeutic potential of tPBM in the human brain.
Introduction
Ever since Endre Mester serendipitously discovered that low-levels of coherent light stimulate hair growth in mice [41], the delivery of light to tissue for therapeutic purposes - now known as photobiomodulation (PBM) - has been investigated for an array of conditions, including wound healing and joint pain reduction [14]. Intriguingly, PBM has recently shown promise as a treatment for disorders of the central nervous system, including both neurological (e.g. stroke, traumatic brain injury) and psychiatric (e.g. depression) disorders [24]. The purported mechanism of action in transcranial PBM (tPBM) is the absorption of light by cytochrome c oxidase (CCO), the terminal enzyme in the mitochondrial electron transport chain, leading to increased energy metabolism [12]. Absorption of near-infrared light has been hypothesized to upregulate CCO activity and increase the rate of ATP production [33, 62], which could then fuel a number of neuronal processes [4].
Several findings from animal and cell culture models support the idea that near-infrared light boosts brain metabolism. tPBM has been shown to increase cerebral blood flow (CBF) [60] and cortical ATP [42], while reducing reducing inflammatory markers [43, 66] and apoptosis [53, 62, 65]. Moreover, tPBM has been reported to ameliorate amyloid beta induced neurodegeneration [37] and oxidative damage from sleep deprivation [54]. Translation of these findings to human neurology and psychiatry has the potential to yield new treatments for disorders in which metabolic or vascular dysfunction have been implicated (e.g. Alzheimer’s Disease [22], depression [6]).
However, there are significant cross-species differences in skull thickness and head size, and it is unknown whether tPBM similarly modulates biological processes when applied through the thicker human cranium and when stimulating only a portion of the brain. Moreover, the measurement of CBF and cerebral blood oxygenation in humans is presently only possible with neuroimaging, whose sensitivity is lower than the invasive techniques employed in animal studies. In order to assess the potential therapeutic benefits of tPBM, objective neurophysiological evidence of its effects on the human brain are necessary.
To date, however, human investigations of tPBM have largely employed cognitive and behavioral outcome measures. Improvements in sustained attention and working memory after a single session of tPBM have been reported [5, 30]. Recent investigations have found that tPBM improves rule-based learning [8] and performance on tests of executive function [7]. Neurophysiological changes in the human brain have only been probed with near infrared resonance spectroscopy (NIRS), reporting an increase in oxygenated hemoglobin over both cerebral hemispheres following application to the right forehead [59]. Due to absorption and scattering of light through the human head, NIRS is limited in its spatial resolution and the ability to measure deep structures [23]. Functional magnetic resonance imaging (fMRI), on the other hand, confers excellent spatial resolution and the ability to image proxies of brain metabolism. In particular, the blood-oxygenation-level-dependent (BOLD) signal is sensitive to both CBF and the cerebral metabolic rate of oxygen (CMRO2) [10], processes that are expected to vary along with a local change in energy metabolism.
fMRI measures the level of deoxygenated hemoglobin in a small volume of tissue. Its vast use in studies of brain function stems from the basic finding that during neural activation, there is an increase in CBF that outweighs the activation-driven increase in oxidative metabolism [45]. Consequently, the relative level of deoxyhemoglobin is actually decreased during activation, leading to an increased BOLD signal due to a more homogeneous local magnetic field [10]. Conventional BOLD imaging samples the T2* decay only once (“single-echo”), whereas multi-echo BOLD acquires multiple (usually 3) samples, with the BOLD contrast from neural activation increasing with echo time [34, 49]. Importantly, the initial (zero-echo) signal intensity is sensitive only to perfusion and independent of deoxygenated hemoglobin content [35].
Here we conducted the first human brain imaging study of tPBM by recording the multi-echo BOLD signal from N = 20 healthy participants receiving tPBM to the right forehead. We found a significant increase in BOLD at brain regions predicted to have high absorption by a simulation of light propagation through the human head. The effect persisted beyond the stimulation period, appearing at all three echos and decreasing in magnitude with echo time.
Results
Simulation of light propagation predicts substantial absorption in grey matter
In order to estimate the spatial extent of significant light absorption in the human brain, we simulated photon propagation through a heterogeneous medium whose optical properties matched those of the human head. The model consisted of four infinitely-wide slabs corresponding to the scalp, skull, cerebrospinal fluid, and brain (Fig 1A; see Simulating light propagation). Although the model did not account for the real geometry of the head, we expected its predictions to be reasonable due to the small size of the beam (i.e., 1 cm diameter) in relation to the curvature of the human scalp. The total simulated dose (150 J) applied to the model matched that of the experiments. The simulated absorption exhibited a maximum in the scalp and decreased exponentially with depth (Fig 1B, logarithmic colormap). Interpolating the absorption pattern onto the anatomical MRIs of the study participants led to maxima in the grey and white matter of 0.83 ± 0.18 J/cm3 (standard deviation across N = 20 subjects) and 0.49 ± 0.13 J/cm3, respectively (Fig 1C; see “Interpolating light absorption onto human head”). The total absorption within the grey and white matter totaled 0.88 ± 0.19 J and 0.36 ± 0.10 J, respectively (Fig 1D). Thus, despite the thickness of the human cranium, substantial light absorption is expected in the cerebral cortex.
Modulation of BOLD at multiple echos
To determine whether tPBM modulated the BOLD signal, and if so, over what extent of the brain, we conducted a whole-brain analysis, probing for both acute and outlasting effects. Statistical Chow tests [13] were conducted separately at each echo of the group-averaged BOLD in order to detect significant changes during and after tPBM, relative to the pre-stimulation baseline period. The Chow test detects structural changes between two time series (here implemented by comparing two sets of autoregressive (AR) coefficients; see Chow Test for detecting an effect on BOLD), thus representing a more flexible test of significant differences between temporal data records than, for example, comparing the levels of the signals.
During tPBM, we found a significant modulation of early-echo (13 ms) BOLD in a cluster of 214 voxels in the right medial frontal gyrus (MFG), 24 mm from the site of light incidence (Fig 2A). A cluster of 154 voxels in the right MFG, 22.4 mm from incidence, was also found at echo 2 (34 ms). Four smaller clusters were detected at echo 3 (56 ms), including one in the right superior frontal gyrus (SFG), 22.5 mm from incidence. The other 3 clusters were located in the paracentral lobule (includes the motor cortex) and cerebellum (>90 mm from the laser), likely reflecting increased movement and movement control during tPBM (study participants were asked to remain still for a 30 minute duration).
Interestingly, we found that tPBM produced an effect on the early-echo BOLD that outlasted the stimulation. A cluster of 218 significant voxels centered in the MFG was found at a distance of 28.7 mm from incidence (Fig 2B). Outlasting effects were also detected in smaller clusters of the paracentral lobule (echo 1) and cerebellum (echo 3). The locations of all significant clusters, for both acute and outlasting effects, are listed in Table 1.
tPBM increases low-frequency BOLD
The Chow test detects structural changes in a time course, but is agnostic to what those changes are. In order to determine the nature of the BOLD effect induced by tPBM (i.e., increase versus decrease, low versus high-frequency), we examined the time courses and AR spectra of the BOLD at the voxels marked significant by the Chow test. Changes to the AR spectrum of the BOLD shed light into which frequencies were modulated (see “Computation of AR spectrum”).
The average time course of significant voxels did not exhibit sustained level differences either during or after tPBM relative to baseline for any of the echos (Fig 3A). Instead, the tPBM effect was evident when examining the spectral content of the BOLD. During stimulation, the AR spectra showed a marked increase in low frequency power for all echos, with the magnitude of the effect decreasing with echo (Fig 3B). The largest increase was observed near DC (0 Hz), indicating a slow modulation from tPBM. The low-frequency increases persisted after stimulation, albeit to a lesser extent than that observed during tPBM (i.e., spectra fell closer to baseline values). We also examined the AR spectra of the voxels not marked significant by the Chow test. These spectra showed remarkable consistency across the pre-tPBM, peri-tPBM, and post-tPBM segments (Fig 3C).
In order to quantify the magnitude of the observed BOLD effect, we computed a measure of percent signal change from the AR spectra. To that end, we computed a frequency-weighted average of percent signal change using the baseline AR power for the frequency weighting. For the acute tPBM effects, the resulting percent signal changes were 31%, 26%, and 11% for echos 1, 2, and 3, respectively (Fig 3D). tPBM produced outlasting percent signal changes of 25%, 16%, and 3% for echos 1, 2, and 3, respectively.
BOLD effect decreases linearly with echo
To gain insight into the physiological mechanisms bringing about the observed changes in BOLD, we analyzed the magnitude of the effect as a function of echo time: the early echo signal is only sensitive to CBF, whereas the slope of the percent change across echo time indicates whether the level of deoxygenated venous blood increased or decreased [34, 49]. We thus regressed the percent signal change (see Fig 3D) onto echo time, considering only the voxels exhibiting a significant effect at any of the three echos (Fig 4A; N = 258). Averaging across all such voxels, we found a decreasing linear trend in the percent signal change across echo time (Fig 4B; R2 = 0.99, p = 0.084, F = 69.7, n = 3).
Estimating energy required to modulate BOLD
Having both the predicted absorption levels (Fig 1) and the spatial pattern of modulated voxels throughout the brain (Fig 2), we sought to estimate the amount of light needed to produce a significant effect on BOLD. For this we separated the voxels into those that responded to tPBM and those that did not. For all 3 echos, the absorption distributions of the two classes showed a high amount of separability (Fig 5A), with the absorption of responsive voxels (means of 0.15 J/cm3, 0.17 J/cm3, and 0.19 J/cm3 for echos 1,2, and 3 respectively) greatly surpassing those of the non-responsive voxels (means of 0.008 J/cm3, 0.008 J/cm3, and 0.01 J/cm3). Employing the logistic regression framework (Nedler and McCullers, 1986), we were able to achieve nearly perfect linear classification of responsive versus non-responsive voxels based on the absorption (Fig 5B, Area Under ROC curve = 0.97, 0.98, and 0.99 for echoes 1, 2, and 3, respectively). The decision boundaries were 0.01, 0.03, and 0.04 J/cm3 at echos 1, 2, and 3. Importantly, the absorptions at which the probability of a significant effect exceeded 0.95 were found to be 0.10, 0.10, and 0.11 J/cm3 for echos 1, 2, and 3, respectively. These values may be interpreted as the amount of absorbed light that is required to produce a reliable effect on BOLD.
Discussion
By applying tPBM to human participants concurrently to scanning their brain activity with multi-echo fMRI, we found that non-invasive laser stimulation of the brain produced acute and outlasting changes to the BOLD signal. A low-frequency increase in the AR spectrum was observed in the medial and superior frontal gyri (MFG/SFG) within 30 mm of the site of light incidence. The increase in BOLD was largest at echo 1 (31%) and decreased linearly with echo time (26% and 11% in echos 2 and 3, respectively).
Our findings are consistent with a superposition of two distinct physiological effects. The first is an increase in CBF, as evidenced by the large BOLD increase at an early echo (i.e., 13 ms) at which the effect of oxygenation is nearly zero [35, 49]. The increased perfusion manifests as an apparent T1 effect where fresh blood enters the imaging voxel and increases the initial signal intensity of the T2* signal [21]. Importantly, the contribution of the blood inflow to BOLD signal changes is constant across echo [34]. The second effect is a simultaneous increase in oxidative metabolism, as evidenced by the fact that the magnitude of the net BOLD effect decreased linearly with echo time (see Fig 4). This indicates that the increase in CBF was coupled with a relatively larger increase in CMRO2, thus yielding an overall increase in the oxygen extraction fraction (OEF), defined as the ratio of CMRO2 to CBF weighted by the arterial oxygen concentration [10]. Thus, our findings indicate that tPBM achieved a focal increase in blood perfusion coupled with a larger increase in oxidative metabolism. To illustrate the superposition of these two inferred effects, we have performed a simulation of the BOLD where simultaneous increases of CBF and OEF were modeled (Fig S2). Note that the fact that after-effects were only observed at echo 1 actually suggests that outlasting changes occurred to both perfusion and oxygenation. At echos 2 and 3, the increase in deoxygenated venous blood reduced the overall BOLD to baseline values, leading to a non-significant net effect.
In studies of brain activation, the BOLD effect arises from a mismatch between CBF and CMRO2: neural activation produces an increase in oxidative metabolism along with an even larger increase in blood flow [10]. As the BOLD signal is sensitive to the homogeneity of the local magnetic field, the relative decrease of paramagnetic deoxygenated venous blood increases the signal. On the other hand, the effect observed here was marked by an increase in deoxygenated hemoglobin that went uncompensated by CBF. Moreover, the effect of perfusion that was found here to increase initial signal intensity is apparently not observed in studies of neural activation, which employ echo times in the order of 30 ms (closest to our second echo time). Thus, we observed two phenomena distinct from those found during neural activation studies: rapid perfusion that increased initial signal intensity, and an increase in CMRO2 that outweighed the effects of CBF, leading to decreasing BOLD across echo time.
The difficulties of interpreting the single-echo BOLD signal are evident when considering our findings from the middle echo in isolation (i.e., 34 ms). In this hypothetical case, one would interpret the finding as a BOLD increase and hence incorrectly infer a relative increase in CBF that outweighed CMRO2 (as in neural activation). However, the availability of the two additional echos allowed us to determine that in fact the CMRO2 increased more than CBF, and that the net increase was due to the apparent T1 decrease from an inflow of blood into the imaging slice. In this manner, our study illustrates the importance of multi-echo BOLD in developing a quantitative understanding of the underlying physiological changes. Moreover, it suggests that caution should be exercised when interpreting single-echo BOLD effects.
Our findings are consistent with the absorption of near infrared light by mitochondrial CCO, leading to increased ATP production. The increased metabolism would manifest as an increase in the CMRO2, which would in turn yield increased CBF into the stimulated region. A rapid increase of CBF that did not fully compensate for the CMRO2 increase could then produce the BOLD changes observed here. However, it is important to point out that the observed effects on BOLD may also have come about from an effect of heating. The rates of chemical reactions, including oxidative metabolism in the brain, are temperature dependent. The heating of brain tissue that is expected to occur during tPBM may increase local CMRO2. The increased metabolic rate will then decrease the oxygen saturation in the venous blood and hence decrease the BOLD signal. Under this scenario, if the CBF recruited to cool the tissue with an inflow of cooler arterial blood was not sufficient to counteract the evoked increase in CMRO2, then the BOLD effects observed here would be predicted. During prolonged visual stimulation that increased the BOLD by approximately 5%, tissue temperature decreases of 0.2° have been estimated in the brain [63]. Incidentally, under the assumption that all of the absorbed light from tPBM ends up as heat, we estimate that the peak temperature increase in the grey matter was also 0.2° (see “Estimating heating induced by tPBM”). The processes regulating temperature in the brain are not well understood, and it is difficult to predict how an exogenous increase in brain temperature would alter the dynamics of CBF and CMRO2. We note, however, that our findings cannot be explained by a simple cooling effect of increased CBF to remove the exogenous heat from tPBM. This would be marked by increased CBF (CMRO2 unchanged) which would increase BOLD but in a manner that increases linearly with echo time (here we found a decrease with echo).
Regardless of the mechanism of action, the ability to non-invasively increase perfusion and metabolism in the human brain encourages the investigation of light-based approaches to the treatment of neurological and psychiatric disorders associated with impaired CBF or metabolism. Most notably, it has been argued that Alzheimer’s Disease (AD) is a cerebrovascular pathology, with chronic brain hypoperfusion linked to AD risk factors and preclinical detection [16]. Intriguingly, there is also evidence that mitochondrial dysfunction has an early role in the progression of AD [44]. tPBM has already been shown to reduce Amyloid Beta (Aβ) plaque levels in a transgenic mouse model of AD [17]. Moreover, laser stimulation of the mouse tibia bone marrow reduced Aβ deposition in progressive AD stages [19]. Taken together with these promising findings from animal studies, our results compel the investigation of tPBM in the treatment and prevention of human AD. The design of such trials will necessitate carefully selecting several parameters, including the site of stimulation on the scalp. Note that in rodent models, the diameter of the laser beam is on the same order as the entire dorsal cortical surface [17]. The results of the mouse bone marrow study [19] suggest that circulation may transfer PBM effects to the brain, although we did not find any “network” effects in this transcranial stimulation study.
One potential limitation of tPBM pertains to the penetration depth. Here we found effects on brain regions within 3 cm of the site of light incidence. This empirical finding matched the prediction of our light propagation model, and we estimated that approximately 0.1 J/cm3 was required to achieve a reliable effect on BOLD. An absorbed energy density of 0.1 J/cm3 corresponds to a fluence of 1.2 J/cm2 in the brain. Biphasic dose responses have been reported for PBM [29], with maxima near 3 J/cm2 [25]. Based on our model, the maximum deposited energy density in the grey matter was 0.83 J/cm3, corresponding to a fluence of 9.9 J/cm2. An observed BOLD effect at these fluences is, therefore, in line with previously reported efficacious doses. Brain regions such as the dorsolateral prefrontal cortex (DLPFC), a frequent target for transcranial magnetic stimulation (TMS) [47], appear to be well within the reach of tPBM. To reach subcortical regions such as the subthalamic nucleus [36] and hippocampus [51], targets for other forms of brain stimulation, it may be required to employ multiple applicators such that the net absorption in these deep targets reaches the threshold found here.
Finally, our findings extend the recent body of evidence showing that light stimulation modulates physiological processes in diverse ways. Blue light has been shown to increase CBF by dilating arterial vessels [52], introducing potential confounds in studies combining optogenetics and fMRI. In a similar light, our findings suggest that caution should be exercised when interpreting NIRS measurements, especially those made over long periods of time where significant absorption of light may have increased perfusion and oxidative metabolism per se. Pulsed infrared stimulation has been shown to alter the electrical capacitance of oocytes via absorption of photons by water [57]. Here we applied light transcranially to the intact human brain, thus hinting that such physiological effects may be achieved non-invasively and in humans.
Materials and Methods
Subjects
All study procedures were approved by the Institutional Review Board of the City University of New York. We recruited N = 20 participants (10 females) from the local New York City population. In an attempt to achieve uniform baseline measures of cerebral oxygenation in our sample, only subjects aged 18-40 were considered: the cerebral metabolic rate of oxygen (CMRO2) has been found to decrease monotonically with age [39, 64]. The mean age of the participants was 24.8 ± 4.6 years. During recruitment, we employed those exclusionary criteria common to MRI (e.g., patients with cardiac pacemakers, neurostimulation systems, or claustrophobia were excluded). All subjects completed the experiments and there were no major adverse effects. One subject complained of a headache following the scan, which may have been caused by the headgear that was worn throughout the (~1 h) experiment.
Experimental Design
We employed a within-subjects design where the BOLD signal was continuously acquired for 30 minutes. The laser was turned on 10 minutes after the onset of the scan and remained active for a 10-minute duration. The equal durations of the pre-tPBM, peri-tPBM and post-tPBM segments were designed to allow a consistent and reliable measure of BOLD during the three experimental periods. Subjects were not made aware of when the laser was turned on or for how long. From verbal post-experimental surveys, subjects did not perceive any sensations (thermal or otherwise) during tPBM.
Transcranial Laser Stimulation
Participants received tPBM at a wavelength of 808 nm, selected based on a previous study that demonstrated an improvement in cognitive task performance after tPBM in healthy volunteers [5]. This wavelength falls within the so-called “optical window” in which absorption by common chromophores (water, melanin, and hemoglobin) is low [26], allowing for deeper penetration through the human scalp and skull and into the brain as compared to other candidate wavelengths [48, 58].
A class IV 10W diode laser (Ultralasers MDL-N-808-10000) powered by a laser driver (Ultralasers PSU-H-LED) provided the monochromatic light. A calibrated photodetector was employed to set the laser power to 250 mW prior to each experiment; the output power of the laser varied by less than 1%. Given a 1 cm diameter aperture in the headgear worn by participants, this resulted in an intensity of 318 mW/cm2, which is within the ANSI safety limit for human tissue (330 mW/cm2).
The laser output was coupled directly into a custom made, multimode optical fiber (Thorlabs FT400EMT) whose core diameter was 400 μm. In order to ensure MR compatibility, the distal end of the optical fiber was fitted with a ceramic ferrule, which was affixed to a custom 3D-printed headgear worn by participants. The headgear, measuring 5.5 cm × 3.2 cm × 2.7 cm, contained a clamp which secured the ferrule. The headgear was secured against the subject’s head such that its 1 cm diameter aperture was flush against the forehead. Given that the duration of tPBM was set to 10 minutes, the total incident light energy was 150 J. It should be noted that the coherent nature of the laser is not believed to contribute to any of the observed effects. A given photon undergoes multiple scattering events on its path to the brain, thereby losing its original phase information. Light emitting diodes (LEDs) have elicited effects in previous PBM experiments [14]. A laser was employed in this case to facilitate light delivery at the desired power.
The headgear was centered at location “Fp2” (right frontal pole) of the 10/20 standard system for electroencephalography [32], where each participant’s Fp2 location was measured and marked on the scalp prior to the experiment. Stimulating the forehead yields better penetration through the scalp due to the absence of scattering by the hair. Our targeted scalp location also matched that employed previously [5]. For all but 2 subjects, Vitamin E markers were placed on the headgear so that the location of light incidence could be registered with the antomical MRI.
Simulating light propagation
We have used the freely available Monte Carlo Multi Layered (MCML) program [61] to simulate light propagation through the human head. The software models an infinitely narrow photon beam normally-incident on multiple layers (slabs) of turbid material. Each layer is specified by five properties: thickness, refractive index, absorption coefficient (μa), scattering coefficient (μs), and the anisotropy factor (g). The resulting “impulse response” is then convolved with a user-defined, finite diameter beam (Gaussian or flat-top), yielding the light absorption (in units of J/cm3) across a finite cylindrical coordinate space where cylindrical symmetry is assumed.
The human head was modelled with 4 distinct layers: scalp, skull, cerebrospinal fluid (CSF), and the brain parenchyma. As described above, MCML requires the specification of five optical properties for each layer. The values used in our simulations are summarized in Table S1. References to the literature from which these values have been taken are provided in the table. Whenever possible, we have used in vivo optical properties. In particular, while MCML requires separate specification of μs and g, in vivo values for the reduced scattering coefficient μ’s ≡ μs(1 − g) and g are more readily available. As such, we have inferred values of μs from reported values of μs and g. The references cited for μs report in vivo values of reduced scattering coefficients. All values of μa and μs are in units of cm−1.
The incoming light was modelled with a beam of 1 cm diameter and a flat intensity profile. In the experiments, we employed a multimode fiber with a core diameter of 400 pm to deliver the laser light to the subject’s head. The transverse intensity profile of our beam was neither Gaussian nor flat-top. However, simulating a Gaussian beam is expected to underestimate the lateral spread of the light inside the head. We preferred to obtain an upper bound on the lateral distribution and thus modelled the beam with a flat transverse intensity profile. We simulated an input energy of 150 J, corresponding to a 10 minute application of tPBM at an intensity of 250 mW.
A grid spacing of 20μm was used in both radial and axial directions, and 500, 000 photons were simulated in each of five runs. We report the average absorption across the 5 runs.
Estimation of heating induced by tPBM
From the simulation of light absorption, the maximum energy absorption in the brain was estimated as E = 0.83 J/cm3. To relate this to a change in tissue temperature, we employed average values of specific heat capacity [1] and density [2] of grey matter as compiled from multiple sources. The specific heat capacity of grey matter was taken to be C = 3696 J/kg/°C, and the density of grey matter was assumed to be ρ = 1045 kg/m3. The change in temperature was then estimated as ΔT = E/ρC = 0.21°C. In arriving at this value, we have assumed that all of the absorbed light energy is eventually converted to heat.
MRI and fMRI
Imaging was performed with a Siemens Magnetom Skyra 3 Tesla scanner. All subjects wore protective eye goggles in addition to the laser headgear during scanning. A 16-channel transmit/receive head coil was used for data acquisition. Structural images were acquired with a T1-weighted MPRAGE sequence (FOV 230 mm, in plane resolution 256 × 256, 224 slices with a thickness of 0.9 mm, TI = 1000 ms). Functional BOLD scans were acquired with a multi-echo EPI sequence (FOV 228 mm, in plane resolution 90-by-90, 60 slices with a thickness of 2.5 mm, TR=2800 ms, Flip Angle = 82 degrees). The three echo times were 12.8 ms, 34.3 ms, and 55.6 ms, which allowed for the characterization of the T2* decay of the BOLD signal [49]. The duration of the BOLD scans was 30 minutes (645 volumes). Subjects were instructed to rest but stay awake and to not think about anything in particular.
Interpolating light absorption onto human head
In order to estimate the amount of light absorbed by the brain during tPBM, we interpolated the result from the simulation of light propagation (see Simulating light propagation) from the native simulation grid (two-dimensional radius r and depth z, see Figure 1) onto the three-dimensional anatomy of each participant’s head. This procedure was not aimed at circumventing the planar geometry of the simulation, but rather to allow for reasonable estimates of absorption in the brain, as well as the amount of light required to modulate BOLD. To that end, we manually extracted the coordinates of four points for each subject: the upper-right-front corner of the laser headgear, the upper-left-front corner, the upper-right-back corner, and the lower-right-front corner. Collectively, these points allowed for the computation of the position and orientation of the laser aperture. For each voxel in the MRI, we then computed its coordinates in a new two-dimensional space (radius and depth; note the cylindrical symmetry of the light absorption) referenced to the site of light incidence. The absorption at a given voxel was computed as the absorption at the simulation grid point with minimum Euclidean distance to the re-referenced voxel coordinate (nearest-neighbor interpolation). The same procedure was followed for the standard (“Colin 27”) head [28], where we averaged the position and orientation of the laser aperture across subjects to define the site of light incidence on the standard head.
fMRI Preprocessing
Preprocessing of BOLD data was performed with the AfNI software package (Version 17.3.03) [15], scripted in the Matlab programming language (Mathworks, Natick, MA). The scripts are available at jd-lab.org/resources/laser.
The anatomical image was first skull-stripped using the 3dSkullStrip function, whose output was then used to create a brain mask via the 3dAutomask function. The anatomical image and brain mask were then registered to the Talairach atlas using the function @auto_tlrc.
To remove large transient spikes from the functional data, we applied the function 3dDespike to the BOLD series of every voxel and echo. The operations of slice-time correction, motion-correction, and alignment with the Talairach-registered anatomical MRI were then performed via the function align_epi_anat.py. Specifically, motion correction was performed by aligning each volume of the BOLD series to a reference volume (i.e., frame 5) using a 6-parameter rigid-body transformation. Registration of the functional images with the anatomical image was performed by a 12-parameter affine transformation maximizing the Local Pearson Correlation cost function.
A second stage of motion correction was then performed by regressing out the time series of the 6 alignment parameters computed in the first motion correction step (3 translation + 3 rotation), as well as its temporal derivatives. This involved projecting the multivariate BOLD onto the null space of the 12-dimensional time series. To further denoise the BOLD, we also regressed out the first 3 principal components of the white matter BOLD, a customary step based on the premise that the white matter represents a “nuisance” signal [11]. The BOLD was then low-pass filtered at 0.1 Hz, a conventional frequency cutoff for BOLD analysis [11]. The time series of each voxel and echo was z-scored. To adjust for small across-subject differences in the volume number corresponding to tPBM onset, we temporally aligned (via a circular shift) each subject’s multivariate BOLD series so that the onset occurred at frame 215 (exactly 10 minutes).
We performed an additional spatial alignment step to each subject’s BOLD data to enforce consistency in the region of the brain receiving tPBM. From the coordinates of each subject’s headgear (see Interpolating light absorption onto human head), we computed the region-of-support of the interpolated absorption pattern (taken here as the spatial extent containing 99% of the total absorption). We then computed the centroid of this region and circularly shifted each subject’s BOLD by the difference between their centroid and the mean centroid across subjects. Note that this step slightly moved each subject’s anatomy out of the Talairach space, so that the effect of the tPBM could be coherently averaged across subjects.
After the temporal and spatial alignment, we computed the group average of the preprocessed BOLD. The motivation for performing an across-subject average (as opposed to conducting the whole-brain analysis at the level of each individual and then employing a second-level group analysis, as is common in conventional general linear model analyses) was to allow for the resolving of tPBM effects in our effectively single-trial data. In conventional event-related fMRI studies, one is afforded multiple repetitions of an experimental stimulus. Here, however, we were limited to a single application of the tPBM intervention, resulting in a low signal-to-noise ratio (SNR) in the individual BOLD traces. Averaging across subjects afforded us a SNR increase of . Finally, the group average BOLD was smoothed to a FWHM of 8 mm using the AfNI function 3dBlurInMask, yielding the final data that was analyzed for effects of tPBM. In order to obtain variability estimates for the analyzed group-averaged BOLD, we employed bootstrapping (sampling with replacement) along the N = 20 subject dimension. 100 bootstrap replicates of the group-averaged BOLD were generated such that standard deviations of the various BOLD statistics (signal intensity, AR spectra, and percent signal change) could be estimated.
Chow Test for detecting an effect on BOLD
To test for an effect of the tPBM on BOLD, we performed the Chow test, which detects the presence of a structural break in a time series [13]. The test is based on comparing the linear regression coefficients of two segments of a time series, thus providing a general test of significant differences (as opposed to simply testing whether one time series is higher or lower than another). To detect acute effects, we tested for significant differences between the pre-tPBM and peri-tPBM periods. To test for effects outlasting the stimulation, we tested for differences between the pre-tPBM and post-tPBM periods. For each period, we fit an autoregressive (AR) model of the form: where yv(t) is the BOLD signal of voxel v at TR t, avk is the autoregressive coefficient of voxel v at lag k, and ev (t) is a zero-mean additive white Gaussian error term. The AR framework models each BOLD sample as a linear combination of the past p samples and a white noise error term. The rationale behind employing this model for BOLD is that hemodynamics change slowly and are thus temporally redundant, yielding strong autocorrelations in the signal. In other words, any part of the signal that is not predictable from past values is likely noise. The signal captured by the AR model is precisely the predictable signal. Separate tests were conducted at each BOLD echo, and all tests were implemented using the Matlab function chowtest.
Optimization of the AR model order
In order to select the AR model order employed in the Chow tests, we performed an optimization routine that spanned all possible model orders from p = 1 to p =10. For every voxel in the brain, we fit an AR model of each order to the voxel’s time series (all time series were normalized by dividing the signal by its sum of squares). At each voxel and order, we employed a 5-fold cross-validation procedure where 80% of the data was used to learn the AR coefficients, while the remaining 20% was employed to measure the resulting residual signal (each fold spanned a contiguous time segment). We conducted this procedure on the class of models described in Eq. (1), and also those including a constant term: where c was also fit by the procedure. At each model order, the power of the residual signal was averaged over all voxels, folds, and echos. We found slightly lower errors with models that did not include the constant, and thus proceeded with only those models. The selected model was chosen by identifying the knee point of the curve of mean residual versus model order, which occurred at p = 4 (Fig S1), matching the value found by [27]. While slightly lower errors were found at higher orders (5 to 7), this selection represented a more parsimonious model that was less likely to produce spurious detection, which is expected to increase with the number of AR coefficients.
Whole-Brain Analysis
Chow tests were conducted independently at each voxel. To correct for multiple comparisons, we controlled for the False Discovery Rate (FDR) at a level of α = 0.05 (N = 86,851 comparisons, one for each voxel in the standard brain). Significant voxels that survived FDR correction were then pruned to retain only those belonging to a cluster of at least 40 voxels (625 mm3), where a cluster was defined as a set of voxels such that each voxel in the cluster was less than 2.5 mm (center-to-center distance) to at least one other voxel in the cluster.
Computation of AR spectrum
To gain insight into the nature of the tPBM effect on the BOLD signal (increase versus decrease, low-versus high-frequency), we computed the spectra of the BOLD captured by the AR models employed in the Chow tests. For each of the three time segments (pre-, peri-, and post-tPBM), we computed the AR spectra according to: where Av (l) is the spectral coefficient at frequency 2πl/N, avk is the fitted AR coefficient of lag k and voxel v (see Eq. 1), and N = 32 is the length of the Fast Fourier Transform. Note that by construction, av0 = 1 [9]. The signal recovered by the AR model may be viewed as the output of a linear filter whose frequency response is given by Av(l) to a white noise input. We thus compared the magnitudes of Av(l) between the three temporal segments of the experiment.
Analysis of BOLD effect size versus echo time
To test whether the effect of tPBM on BOLD exhibited a linear dependence on echo time, we considered voxels exhibiting a significant acute effect at any of the 3 echos. There were N = 258 such voxels. For each of these, we computed the percent signal change (frequency-weighted average of the AR spectral magnitude), then computed the average percent signal change across voxels, and finally regressed the result onto echo time using ordinary least squares. To test for a significant linear dependence of percent signal change on echo time, we compared two linear regression models. The “full” model was specified by: where PSC(t) is the percent signal change at echo time t, b is the slope and bo the y-intercept. The residual error of the full model was contrasted to that given by the following “reduced” model: where a general linear F-test was conducted to evaluate whether the full model significantly reduces the residual over the reduced model.
Logistic regression analysis
To obtain estimates of the amount of light that is required to modulate the BOLD signal, we conducted a classification analysis aimed at predicting whether a given brain region exhibited a significant modulation during tPBM. For this analysis, we resampled the pattern of significant voxels (i.e., results of the Chow tests) onto the anatomical space of the standard head [28] where the absorption was defined. Voxels were then split into two classes: those exhibiting a significant effect (after FDR correction and clustering), and those that did not. Voxels outside of the brain and those having zero absorption (i.e., locations very far from the site of stimulation) were excluded from the analysis. For the analysis of echo 3, we also excluded voxels in the significant clusters residing in the paracentral lobule and cerebellum which were presumed to be related to movement (see Table 1). This led to the following number of voxels in each echo: 3555 significant and 212384 not-significant (Echo 1), 2433 significant and 213506 not-significant (Echo 2), 890 significant and 215049 not-significant (Echo 3). Due to the large imbalance in the sizes of the two classes, we employed a common technique in class-imbalanced learning problems, namely repeating the infrequent examples (i.e., significant voxels) such that the number of examples in each class was equal [38]. We then fit a simple classifier of the form: where a is the absorption, β is the linear term multiplying the absorption, and βo is the threshold which represents the optimal decision boundary between predictions of significant and non-significant voxels. At each echo, learning of the model parameters β and βo was performed by the Matlab function mnrfit [40]. From the logistic regression model, we estimated the absorption required to achieve a BOLD effect with probabilities of 0.5 and 0.95 by setting p(significant) to the desired probability and then solving for a in Eq. (6).
Simulation of BOLD signal changes
In order to demonstrate the combined effects of increased CBF and OEF, we conducted a simple simulation of the BOLD signal:
Baseline values were selected as So = 1 and . We simulated a 40% increase in So, a 10% decrease in T2*, as well as their combined effect.
Acknowledgments
This research was supported by a City University of New York (CUNY) Junior Faculty Research Award to JPD. The authors would like to thank Lazar Fleysher for technical support during the initial development of the experimental setup. We would also like to acknowledge the help of Christian Fong in subject recruitment and experimental assistance, as well as helpful preliminary discussions with Hanli Liu. Contributions: GMD and JPD designed the research, collected the data, analyzed the data, and wrote the manuscript.