Two-dimensional intrinsic and hyperspectral imaging

Figure 2 shows a typical setup for 2-D hyperspectral imaging of exposed rat cortex. This neuroimaging approach relies on the distinctive absorption spectra of oxy- and deoxyhemoglobin ( ${\mathrm{HbO}}_2$ and $\mathrm{HbR}$ ) as shown in Fig. 1a. The hemoglobin in blood is the most significant absorber in the brain at visible and NIR wavelengths, and functional activity in the brain results in changes in blood flow, blood volume, and oxygenation.²⁶. Blood delivers oxygen to tissues by locally converting ${\mathrm{HbO}}_2$ to $\mathrm{HbR}$ , so changes in the oxygenation state of blood correspond to changes in the relative concentrations of ${\mathrm{HbO}}_2$ and $\mathrm{HbR}$ . Consequently, variations in the amount and oxygenation state of hemoglobin modulate the absorption properties of the brain. By simply shining light into the exposed cortex and taking pictures with a camera, the hemodynamic response to stimulus can be readliy observed. Further, since $\mathrm{HbR}$ and ${\mathrm{HbO}}_2$ have different absorption spectra [see Figs. 1a and 1b], measurements with different wavelengths of light can produce images that are preferentially sensitive to changes in the concentration of either ${\mathrm{HbO}}_2$ or $\mathrm{HbR}$ . At isobestic points, where the $\mathrm{HbR}$ and ${\mathrm{HbO}}_2$ absorption is the same (e.g., $\sim 500$ , 530, 570, and $797\mathrm{nm}$ ), changes in total hemoglobin concentration $(\mathrm{HbT}=\mathrm{HbR}+{\mathrm{HbO}}_2)$ can be measured independently of changes in blood oxygenation.

Fig. 2

A typical setup for camera-based imaging of the exposed cortex. Inset shows typical hemodynamic and calcium-sensitive responses in rat somatosensory cortex to $\sim 1\text{-}\mathrm{mA}$ forepaw stimulus delivered in $3\mathrm{msec}$ pulses at $3\mathrm{Hz}$ for $4\sec$ .

Early approaches to optical imaging of the exposed cortex referred to “intrinsic signal” imaging, and often used illumination at an isobestic point to examine localized changes in tissue $\mathrm{HbT}$ .²⁷ By adding a second wavelength (often $610\mathrm{nm}$ ), researchers were able to investigate oximetric changes in addition to changes in $\mathrm{HbT}$ .^{28, 29} Since ${\mathrm{HbO}}_2$ absorption at $610\mathrm{nm}$ is low (around 16% of the absorption of $\mathrm{HbR}$ ), changes in absorption at $610\mathrm{nm}$ are largely attributed to, and interpreted as, changes in $\mathrm{HbR}$ . However, it should be noted that even though ${\mathrm{HbO}}_2$ absorption at $610\mathrm{nm}$ is lower than that of $\mathrm{HbR}$ , ${\mathrm{HbO}}_2$ changes can interfere with and distort inference of $\mathrm{HbR}$ concentration changes from measurements at $610\mathrm{nm}$ .

To calculate the actual changes in concentration of ${\mathrm{HbO}}_2$ , $\mathrm{HbR}$ , and $\mathrm{HbT}$ , the absorption of the brain must be measured at two or more wavelengths (hyperspectral imaging). In some cases, multiple wavelengths across the visible spectrum are used. This is typically achieved by synchronizing rapid camera acquisition with a filter wheel in front of a white light source, as shown in Fig. 2.^{30, 31} Full spectra of the reflectance changes along a single line on the cortex have also been used to provide a more detailed spectral picture of the changes occurring during functional activation.^{32, 33}

Calculation of hemoglobin concentrations from such measurements is not trivial, as the effects of scattering must be considered, and then data must be combined with the known spectra for ${\mathrm{HbO}}_2$ and $\mathrm{HbR}$ to deduce the concentration changes corresponding to the measurements. In this case, we assume that scattering in the brain does not change significantly during the hemodynamic response. (Note that the following approach is also applied to noninvasive clinical optical topography, as is discussed in Sec. 2). The basis of this approach is the modified Beer-Lambert law,³⁴ which states that:

$I_0$ may be spatially varying and difficult to accurately determine. $G$ is also very difficult to quantify. Therefore, it is common to consider changes in detected signal relative to an initial time point $t=0$ or a reference state:

While imaging absorption changes in superficial tissues appears quite straightforward, the effects of scatter must be carefully considered. Since scattering is wavelength dependent in tissues [an approximate scatter spectrum is shown in Fig. 1a], each measurement wavelength will experience different levels of scattering. The mean path of light through tissue $\left(\mathrm{DPF}x\right)$ will also be altered by the background absorption properties of the tissue, which are also heavily wavelength dependent, owing to the presence of hemoglobin, so $\mathrm{DPF}{x}_{\lambda 1}\ne \mathrm{DPF}{x}_{\lambda 2}$ in almost all cases. Therefore, as given by Eqs. 4, 5, accurate estimates of these wavelength-dependent pathlength factors $\mathrm{DPF}{x}_{\lambda n}$ are required even to calculate differential changes in hemoglobin concentration ( $\mathrm{\Delta }{c}_{{\mathrm{HbO}}_2}$ and $\mathrm{\Delta }{c}_{\mathrm{HbR}}$ ).

Note, however, that if $I\left(t\right)∕I\left(t_0\right)$ is measured at an isobestic point, and $\mathrm{DPF}x$ does not change over time, then $I\left(t\right)∕I\left(t_0\right)$ will be directly proportional to $\mathrm{\Delta }\mathrm{HbT}\left(t\right)$ .

Estimation of this wavelength-dependent pathlength $\mathrm{DPF}{x}_{\lambda }$ is difficult. For exposed-cortex imaging, a Monte Carlo model of light propagation is generally employed.³⁵ Such a model requires estimates of the background absorption and oxygenation level of the brain tissue to predict the likely scattered paths of photons.³¹ For camera imaging, the mean pathlength calculation should simulate a widely distributed illumination, and incorporate the numerical aperture of the camera’s lens.

It has been demonstrated that incorrect estimates of $\mathrm{DPF}{x}_{\lambda }$ can directly impact the correct calculation of ${\mathrm{HbO}}_2$ , $\mathrm{HbR}$ , and $\mathrm{HbT}$ changes in tissue.³⁵ Note also that if only values of $\mathrm{\Delta }{\mathrm{HbO}}_2\left(t\right)$ and $\mathrm{\Delta }\mathrm{HbR}\left(t\right)$ are derived, that it is not possible to infer changes in absolute oxygen saturation $\mathrm{\Delta }{\mathrm{SO}}_2={\mathrm{HbO}}_2\left(t\right)∕\left[{\mathrm{HbO}}_2\right(t)+\mathrm{HbR}(t\left)\right]-{\mathrm{HbO}}_2\left(t_0\right)∕\left[{\mathrm{HbO}}_2\right(t_0)+\mathrm{HbR}(t_0\left)\right]$ without first estimating the baseline absolute concentrations of ${\mathrm{HbO}}_2$ and $\mathrm{HbR}$ .

Another important issue to be considered in 2-D imaging of the cortex is the depth sensitivity of the measurement. The detected light has scattered within the tissue, and reports absorption that it has experienced along its path. The majority of detected photons have predominantly sampled the very superficial layers of the cortex, whereas fewer have traveled deeply and managed to return to the surface. This means that the image of the cortex is a superficially weighted sum of signals from different layers. Since the cortex has a complex structure of large superficial vessels and deeper capillary beds, the relative contributions of these compartments to the overall signal can be difficult to interpret. Additionally, since the mean penetration depth of the light will also depend on its wavelength, all wavelengths have not necessarily interrogated exactly the same region of tissue, which can impact the quantitative accuracy of estimates of ${\mathrm{HbO}}_2$ , $\mathrm{HbR}$ , and $\mathrm{HbT}$ .

A typical hemodynamic response in the somatosensory cortex of a rat during a four-second forepaw stimulus ( $3\mathrm{\mu }\mathrm{s}$ pulses at $3\mathrm{Hz}$ ) is shown inset in Fig. 2. The hemodynamic response to stimulus is fairly slow, peaking several seconds after the first stimulus pulse. A large increase in ${\mathrm{HbO}}_2$ is generally seen, along with an increase in $\mathrm{HbT}$ and a decrease in $\mathrm{HbR}$ . The overall timing characteristics of each hemoglobin type are different from each other, and contain information about the underlying oxygen delivery and consumption dynamics.^{19, 36, 37} Some researchers routinely observe a short initial increase in $\mathrm{HbR}$ , attributed to rapid oxygen consumption at the site of neuronal activation.^{29, 31, 33} The origin of this initial dip is somewhat contentious, and it is discussed further in Sec. 2.1.3. The data shown in Fig. 2 are an average response of 400 stimulus repetitions. It should be noted that individual stimulus responses are often affected by physiological noise, such as variations in baseline blood flow to the region and fluctuations in systemic parameters such as blood pressure.

Despite all of these important considerations, 2-D optical imaging of the exposed cortex has provided very significant and well validated observations of functional cortical hemodynamics. Specific examples of applications of these techniques are given at the end of this section.

Two-dimensional voltage sensitive dye imaging

As described before, intrinsic signal and hyperspectral imaging can report the hemodynamic response to functional activation in the brain. This hemodynamic response is generally assumed to spatially correlate with the underlying brain activity corresponding to neurons firing and conducting electrical signals. The interrelation between neuronal and hemodynamic activity is currently an area of intense research.^{18, 38, 39} This neurovascular coupling is thought to be an essential part of normal brain activity and is believed to be disturbed by pathologies such as Alzheimer’s disease and stroke.^{40, 41} Understanding whether it is possible to determine detailed information about neuronal activity from hemodynamic measures is also an important area of research, driven by the fact that clinical fMRI is currently only sensitive to hemodynamic changes.

Voltage sensitive dyes (VSDs) provide a way to optically detect neuronal activity in the in-vivo exposed-cortex. VSDs were first used in 1968⁴² but received renewed attention in 1986 when they were demonstrated for in-vivo mammalian brain imaging by Grinvald and colleagues^{43, 44, 45} studies that predate the first reported intrinsic signal measurements.²⁷ Since then, significant advances in digital detector technology, dye formulation, and staining strategies have led to established use of VSDs to visualize neuronal activity in-vivo ^{2, 24, 46, 47} and also for mapping electrical activity in the heart.^{48, 49}

VSDs are molecules that generally bind across a neuron’s membrane. Changes in membrane potential cause measurable changes in the fluorescence of the dye.⁵⁰ Since a large percentage of the total membrane area is dendritic, VSD signals reflect both spiking and synaptic activity.⁴²

Calcium sensitive dyes (CaSDs) are compounds that increase their fluorescence in response to increases in calcium ion concentration. When cells such as neurons are loaded with CaSDs, it becomes possible to monitor intracellular calcium.⁵¹ Since a calcium influx occurs when a neuron fires an action potential, the calcium dye’s fluorescence can optically report this, providing a second tool with which to optically image neuronal activity.

To image these dyes, a similar system to that shown in Fig. 2 is used, with the filter wheel replaced by an excitation filter of the appropriate wavelength. A long-pass filter is placed in front of the camera to block excitation light and isolate the fluorescence emission. In some cases, a dichroic mirror is used such that the illumination light is reflected onto the brain, and emitted fluorescence passes through the dichroic to the camera (via a secondary emission filter to remove residual excitation light). This is the configuration of a commercially available VSD imaging system sold by Optical Imaging Incorporated (Rehovot, Israel).²

As shown in the inset of Fig. 2, the hemodynamic response to stimulus typically peaks 2 to $4\sec$ after the stimulus begins. The neuronal response is much faster. Measured using VSDs and CaSDs, neuronal responses typically peak within 25 to $100\mathrm{msec}$ , and are usually followed by an inhibitory or refractory period of negative signal, which returns to baseline at around 150 to $300\mathrm{msec}$ after the stimulus starts.⁵² The black trace in Fig. 2 shows the fluorescence changes seen in the somatosensory cortex stained with Oregon Green 488 BAPTA-1, AM calcium sensitive dye (Invitrogen) during the $4\text{-}\sec$ forepaw stimulus ( $3\text{-}\mathrm{\mu }\mathrm{s}$ pulses at $3\mathrm{Hz}$ ). Very fast, individual responses for each stimulation pulse can be seen, in contrast to the very slow hemodynamic response. To capture these responses, a faster camera is required than for hyperspectral or intrinsic signal imaging (preferably $<10\mathrm{ms}$ per frame).

Voltage sensitive dyes typically have quite low contrast (typical $\mathrm{\Delta }F∕F\sim 0.5\%$ ). As a result, they can be affected by changes in hemoglobin absorption due to heart rate, breathing, and of course functional activation. Two methods have been developed to overcome this. 1. New dyes have been developed that excite and emit at higher wavelengths, corresponding to regions where hemoglobin absorption is less significant [see Fig. 1a]. 2. Acquisition of data is often triggered on the heart beat and sometimes breathing of the animal.² Many repetitions are generally averaged, and “blank” acquisitions where no stimuli are presented are interleaved such that the final signal is given by the stimulated – blank data. Voltage sensitive dyes can only be used in animals at present, and are usually topically applied directly to the exposed cortex for 1 to $2\mathrm{h}$ before imaging commences.⁵³ Some VSDs have been shown to quickly photobleach or have phototoxic effects, and so illumination times are typically kept to a minimum. It is also important to note that the change in fluorescence or absorption of VSDs is usually a manifestation of a spectral shift in either the absorption or emission spectrum of the dye.⁵⁴ This means that measurement wavelength selection should not necessarily be based on the peak absorption and emission wavelengths, nor should broadband emission filters spanning the whole emission peak always be selected. The excitation and emission wavelengths should be selected based on areas of the spectrum where the largest change in intensity will result when a spectral shift occurs (e.g., a band on the falling side of the emission spectrum).⁵⁴ Note that selection of wavelengths may therefore determine whether an increase or decrease in absorption or fluorescence is observed in response to an increase in membrane potential.

CaSDs have much higher contrast than VSDs $(\mathrm{\Delta }F∕F\sim 5\%)$ and hence, despite generally exciting and emitting in the visible spectrum, are not strongly affected by small hemodynamic changes (although the small dip after the calcium response in Fig. 2 may be due to the concurrent increase in $\mathrm{HbT}$ and ${\mathrm{HbO}}_2$ absorption). The challenge with $\mathrm{CaSDs}$ is introducing them into the neurons successfully, which to date has required the addition of dimethyl sulfoxide (DMSO) and pluronic acid, followed by direct pressure injection via micropipette into the cortex and incubation of $1∕2$ to $1\mathrm{h}$ .^{3, 55} However, it is possible to achieve fairly uniform loading of neurons within an area of several millimeters of cortex, and to observe responses similar to those seen with voltage sensitive dyes. Unlike VSDs, CaSDs tend to increase the intensity of their fluorescence emission in the presence of calcium, without significant shifts in the spectral position of their emission and excitation bands.

Examples of the use of VSDs and CaSDs for neuroscience studies are given at the end of this section. Imaging of CaSDs with two-photon microscopy is described in Sec. 2.2

Laser speckle-flow imaging

Another recent development in 2-D optical imaging of the exposed cortex is laser speckle-flow imaging. While hyperspectral and intrinsic imaging provide measurements of the changes in ${\mathrm{HbO}}_2$ and $\mathrm{HbR}$ concentrations, they do not explicitly measure changes in the flow rate of the blood. Flow is a very important parameter if the rate of oxygen delivery (and/or consumption) is to be calculated. Laser Doppler has been widely used to provide point measurements of blood flow in the brain. However, laser speckle-flow imaging of the exposed cortex can provide a spatially resolved image of the flow throughout the vascular network during functional activation.^{56, 57}

Acquiring laser speckle-flow imaging data uses a similar setup to that shown in Fig. 2. However, in this case, the light source is replaced with a divergent laser diode (typically at $780\mathrm{nm}$ ). The camera then simply acquires rapid and high-resolution images of the exposed cortex, which appear to be low-contrast images of the speckle pattern from the laser illumination. The speckle pattern is caused by the coherent laser light scattering within the brain. If red blood cells are moving within the image, they cause the speckle pattern to vary over time. The rate at which the speckles change relates to how fast the red blood cells are moving. Laser Doppler measures this temporal modulation. However, within the camera image, this intensity fluctuation causes spatial blurring. By calculating the spatial uniformity of an area of camera pixels (e.g., $5\times 5$ or $7\times 7$ ), the amount of blurring can be quantified [from the pixel group’s (standard deviation of intensity)/(mean intensity) for each time point]. This speckle contrast can be mathematically related to the correlation time, which gives a measure of the velocity of flow of the red blood cells.⁵⁶ Laser speckle-flow imaging has been performed simultaneously with hyperspectral imaging of the exposed cortex (by adding a second camera, a dichroic, and filters to block the laser light). This has enabled the cerebral metabolic rate of oxygen consumption to be calculated.^{30, 36, 37} Examples of the use of speckle-flow imaging are given next.

Functional imaging for sensory and cognitive processing research

Most regions of the brain’s cortex appear to be dedicated to consistent tasks. Particularly in the sensory cortices, there is even spatial correspondence between, for example, the location of two whiskers on the snout of a rat and the two sites (or barrels) on the somatosensory cortex that respond when each whisker is flicked.^{58, 59} Different parts of the auditory cortex respond to different frequencies of sound.⁶⁰ Studies of these systems are prevalent throughout the functional imaging literature, but optical imaging of the exposed cortex is currently the only technique that can provide sufficient resolution to reveal the intricate nuances of these hemodynamic and neuronal responses.

Figure 3a shows a recent exposed-cortex optical imaging result revealing the spatially distributed cortical response to visual stimulus in anesthetized rat.⁶¹ Cortical imaging data were acquired at $590\mathrm{nm}$ and the results reveal a very clear spatially correlated retinotopic map of the cortical responses to visual stimuli across the entire visual field of one eye. The purpose of this study was to determine whether it is feasible to use rats as a model to evaluate adaptation of visual cortical function in response to retinal degeneration, and for development of associated potential therapies.

Fig. 3

Examples of in-vivo exposed cortex imaging. (a) Visual stimuli were presented in different parts of a rat’s field of view, and the corresponding cortical hemodynamic $\mathrm{HbT}$ responses were mapped (color code indicates corresponding location of visual stimulus). Reproduced with permission from Gias ⁶¹ (b) A cranial window was implanted to allow direct imaging of primate visual cortex. A baseline image of the exposed brain is shown above an intrinsic signal ocular dominance map, acquired by subtracting the response from visual stimulus to the right eye from the left eye. Chronic VSD imaging was also performed, and ocular dominance maps closely resembled intrinsic signals. Time courses of the voltage-sensitive signal (red) showed a more prompt onset that the hemodynamic response (blue). Reproduced with permission from Slovin ²⁴ (c) An optical fiber bundle was used to image the exposed, VSD stained cortex of an awake mouse while it explored its surroundings. Image series shows the mouse’s whisker gradually brushing against an obstacle and the corresponding cortical response. Reproduced with permission from Férézou ⁶⁵ (d) A fiber optic bundle was used to image the exposed cortex of a rat undergoing electrical whisker stimulation during simultaneous acquisition of fMRI at $4.7\mathrm{T}$ . The hemodynamic response shows localized cortical changes in $\mathrm{HbR}$ , ${\mathrm{HbO}}_2$ , and $\mathrm{HbT}$ , which correspond well to oblique-slice fMRI BOLD response (study by Hillman, Devor, DeCrespigny and D'Arceuil). (e) Baseline speckle contrast image showing higher flow as darker contrast and strongly accentuated vessels. Image series from top to bottom show flow prior to induction of spreading depression, $2\min$ after which a wave of cortical flow increase moves across the field of view, and $20\min$ after, when hyperperfusion in the dural vessels is prominent. Plot shows time course of flow for region of interest in cortex (blue) and dural vessel (red). Reproduced with permission from Bolay ¹⁵ (f) (left) Cross sectional $x\text{-}z$ images through the cortical ${\mathrm{HbO}}_2$ response to a $4\text{-}\sec$ forepaw stimulus in rat imaged using laminar optical tomography, (top) transecting a superficial vein, and (bottom) transecting the deeper capillary response. (right) 3-D rendering of arterial, capillary, and venous compartment hemodynamic responses. Ex-vivo two-photon microscopy of vasculature (inset) agrees with compartment discrimination. Reproduced with permission from Hillman ¹⁹

Figure 3b shows an example of a highly complex study of the visual cortex of awake monkeys using both intrinsic signal (hemoglobin) and VSD imaging.²⁴ Three Macaca fascicularis monkeys had ‘windows’ surgically implanted into their skulls over their visual cortex (covering areas V1, V2, and V4). The cortical response to visual stimulus was then imaged repeatedly for up to 1 year using VSD RH-1691 (Optical Imaging Incorporated, Rehovot, Israel) to stain the cortex 1 to 3 times per week (excited at $630\mathrm{nm}\pm 5\mathrm{nm}$ , emission: $650\text{-}\mathrm{nm}$ long-pass). For studies of the visual cortex, it is very common to utilize two equivalent visual stimuli and display the difference between the two responses. This is because each response on its own consists of many different changes in the vasculature and tissue, so by subtracting two or more equivalent states, only the very specific difference between the stimuli responses should be revealed. In this case, Fig. 3b shows intrinsic signal ocular dominance maps ( $605\text{-}\mathrm{nm}$ reflectance) generated by showing the monkey a pattern with either his left or right eye blocked (the baseline image of the exposed cortex is also shown). The difference between these two states therefore reveals the response of one eye as bright areas and the response of the other as dark stripes. The image shows that this part of the visual cortex processes similar areas of the image from each eye in close proximity to each other, interleaving them, rather than processing the full images from the left and right eyes separately. Very similar maps were seen for both VSD and intrinsic signal imaging, although the timing and evolutions of the neuronal and hemodynamic signals were quite different. The data shown here are just the introduction to a very broad study where similar maps were used to investigate the visual cortex’s response to rapid saccadic eye movements. There are numerous other chronic optical imaging studies of monkey visual cortex⁶² and monkey somatosensory cortex.⁶³ Very recent studies have also demonstrated that plasticity of ocular dominance can even be studied in mice using exposed-cortex optical imaging.⁶⁴

Another recent novel approach to exposed-cortex optical imaging is shown in Fig. 3c. While most rodent measurements use anesthetized animals, some studies of native behavior require that the animal be both awake and able to freely move around. Férézou, Bolea, and Petersen developed a technique where a mouse’s voltage sensitive dye-stained exposed somatosensory cortex could be imaged via a flexible fiber bundle while the awake mouse explored its surroundings.⁶⁵ By simultaneously acquiring video of the mouse’s behavior, they were able to directly image the neuronal response in the brain when the mouse’s whisker contacted an object.

Functional imaging to investigate the mechanisms of activation

The functional imaging studies described above exploit the fact that colocalized neuronal and hemodynamic responses can be detected in the brain in response to stimuli.²⁶ However, very little is understood about how and why these functional responses actually occur.^{40, 41, 66} Understanding the origins and characteristics of these responses, and the interplay between neuronal and hemodynamic processes, is key to developing an improved picture of healthy and impaired brain function. It is also an essential part of developing and improving all functional brain imaging modalities that rely on the contrast provided by hemodynamic changes such as fMRI. Since exposed cortex optical imaging can provide such a close view of the working brain, as well as access for invasive measurements such as electrophysiology, it has become an established tool for the study of basic functional mechanisms.

A wealth of recent studies have investigated so-called neurovascular coupling (NVC) relationships using optical methods. ^{18, 31, 38, 39, 58, 67, 68, 69} By quantifying the hemodynamic response (via optical imaging and/or laser Doppler) simultaneously with the neuronal response (via electrophysiology), these studies sought to identify a parametric relationship that could be used to predict one response from the other. Devor demonstrated in rat that within a single whisker barrel, the neuronal response to different amplitudes of tactile whisker stimulus saturated before the hemodynamic response.¹⁸ More recently, subtleties of the spatiotemporal evolution of both responses have shown that the NVC relationship is highly complex.^{58, 70} A recent study also demonstrated that there can be significant differences between the responses of anesthetized and awake rodents.²³

An alternative to attempting to infer neuronal responses from BOLD-type measurements is to instead locate sites where oxygen consumption is increasing by determining the cerebral metabolic rate of oxygen consumption $\left({\mathrm{CMRO}}_2\right)$ .^{71, 72} In principle, these regions should spatially correlate more closely with neuronal activity than the BOLD response, which can be significant even in distant draining veins.⁷³ ${\mathrm{CMRO}}_2$ is a parameter that can only be calculated by considering the amount of oxygen that is delivered to and extracted by the tissue. Although 2-D hyperspectral imaging can provide changes in ${\mathrm{HbO}}_2$ , $\mathrm{HbR}$ , and $\mathrm{HbT}$ concentration, the data represent discrete snapshots of the current oxygenation state of the blood, and not the rate at which the blood is being replaced. It is therefore necessary to also measure blood flow, to determine the rate of delivery of oxygen to the tissue. Other estimates are also required such as the variations originating in the venous compartment. Both laser Doppler and more recently speckle-flow imaging have allowed simultaneous measurement of oxygenation, blood volume, and flow dynamics in the exposed cortex,^{30, 36, 37} thereby allowing direct calculation of ${\mathrm{CMRO}}_2$ . These calculations are based on models of the hemodynamic response to stimulus. ^{32, 74, 75, 76, 77, 78} Interestingly, in-vivo optical imaging and microscopy of the cortex are also allowing the vascular mechanisms of the hemodynamic response underlying these vascular models to be investigated directly.^{19, 79, 80}

Another approach to comparing exposed-cortex optical imaging findings to fMRI is to perform fMRI studies and optical measurements on the same subjects in response to the same stimulus. Several recent studies have demonstrated simultaneous optical imaging of the exposed cortex during fMRI acquisition.⁸¹ An example of such a study is shown in Fig. 3d, where an MRI-compatible fiber optic endoscope was configured to allow synchronized fMRI and functional optical imaging at two wavelengths (570 and $610\mathrm{nm}$ ) during electrical whisker stimulus of varying amplitudes.

Similar studies have been performed in humans scheduled to undergo open cortex brain surgery^{21, 82, 83}: functional responses to voluntary motion of the tongue were recorded in the language area of the brain prior to surgery using fMRI, and then compared to optical imaging results obtained during surgery, where the awake patient was asked to repeat the same stimulus paradigm.⁸³ These results were compared to electrocortical stimulation maps acquired by directly stimulating the cortex and questioning the patient about the resulting involuntary tongue movements.

A final optical imaging result of significant importance, and which has prompted much debate, was the discovery of the so-called “initial dip.”³³ In some studies, the hemodynamic response to stimulus is seen to begin with an increase in $\mathrm{HbR}$ , prior to a subsequent decrease. This finding was interpreted as evidence that neurons initially use up local oxygen more quickly than it is being supplied by the gradually increasing blood flow, an important finding for modeling the mechanisms of the hemodynamic response. It was proposed that this early signal could also allow improved spatial localization to the neuronal response. This initial HbR rise was subsequently also seen in fMRI results as an initial dip in the BOLD signal,^{84, 85} and in many other optical imaging studies.^{29, 37, 86} However, the dip is controversial because it is not seen in all studies.⁸⁷ Reasons proposed for this discrepancy have ranged from differences between species, anesthesia, and paradigms, to measurement techniques and errors in spectroscopic analysis of optical signals (especially the wavelength dependence of pathlength⁸⁸). An excellent summary of the initial dip can be found in Ref. 89.

Functional imaging to investigate pathologies and treatments

A further use of exposed cortex imaging is to directly investigate the manifestation, characteristics, and treatment responses of diseases and disorders. An example of one such study is shown in Fig. 3e, which used speckle-flow imaging to investigate migraine in a mouse model.¹⁵ Cortical spreading depression (CSD) is a slow moving wave of neuronal and hemodynamic activity that spreads across the cortex following a small insult to the cortex, such as a pin prick.⁹⁰ CSD is thought to be the cause of the visual disturbances that often precede a migraine headache. This study demonstrated that speckle-flow imaging could not only see the wave of CSD as a transient increase in parenchymal blood flow, but also revealed a prominent subsequent increase in blood flow in the middle meningeal artery (MMA), a vessel that traverses the dura. Such a dural blood flow increase is consistent with the delayed onset of a debilitating migraine headache. Supporting measurements were then made to conclude that CSDs can trigger events leading to headache, identifying new pathways that could potentially be targeted for the development of treatments for migraine suppression.

Similar studies using both speckle flow imaging and hypersepctral imaging of the exposed cortex have examined the effects of middle cerebral artery occlusion as a model of stroke.^{13, 91} Both the acute stage and chronic manifestations of stroke can be studied. Function of the area affected by a stroke can be evaluated after days or even weeks of recovery.^{92, 93} The effects of established and experimental treatments can be tested in these models of human stroke, and even compared to the fMRI manifestation of the same outcomes.⁹⁴

It is also possible to study epilepsy in animal models using exposed-cortex intrinsic signal optical imaging⁹⁵ and VSDs.⁹⁶ The hemodynamic response to intrasurgical cortical stimulation during epilepsy surgeries on humans has also been investigated using optical imaging.²²

Intrinsic fluorophore imaging

Several recent studies have shown compelling evidence that it is possible to measure concentration changes in intrinsic flavoproteins (FAD) in the brain from their intrinsic fluorescence in vivo. ^{97, 98, 99, 100} A system almost identical to that used for VSD imaging is employed, but no contrast agent is added to the brain. Excitation is between 450 and $490\mathrm{nm}$ and emission is between 500 and $550\mathrm{nm}$ . FAD provides a measure of metabolic changes in cells (neurons and possibly glia), and has a distinct time course compared to shorter neuronal and longer hemodynamic responses. FAD imaging was recently used to explore visual plasticity in mice in a similar way to the studies of the visual cortex described before,¹⁰¹ as well as plasticity in the auditory cortex¹⁰² and epileptic activity.¹⁰³ Since flavoprotein fluorescence requires visible light for its detection, it is unlikely that it will be possible to detect these signals noninvasively (see Fig. 1).

Fast scattering signal

Another source of intrinsic optical signal in the brain is the fast scattering signal. Historically, this signal has been the source of much debate, particularly related to the difficulties in detecting it noninvasively,^{104, 105} although the signal is clearly observable in isolated nerves.¹⁰⁶ Recently, exposed cortex measurements during whisker stimulation in rats have provided compelling data revealing localized signals corresponding to regions of neuronal activity.¹⁰⁷ Changes in the dark-field reflectance of $660\text{-}\mathrm{nm}$ light occurred within $50\mathrm{ms}$ of stimulation and correspond well with electrophysiological recordings. A $4\text{-}\mathrm{mm}\text{-}\mathrm{diam}$ imaging probe was placed in direct contact with the dura over the exposed somatosensory cortex. The probe consisted of an outer ring of optical fibers delivering illumination light from $660\text{-}\mathrm{nm}$ light emitting diodes, and a central fiber optic imaging conduit transmitting remitted light to a sensitive charge-coupled device (CCD) camera. Several possible mechanisms for the origin of the fast scattering signal have been proposed, including reorientation of membrane proteins and swelling of neuronal dendrites due to water influx in response to ionic currents during depolarization.¹⁰⁷

Note that the possibility that fast scattering changes in the brain may distort and interfere with VSD intensity signals has not yet been explored. It is also still not fully understood whether significant scattering changes occur in the brain as part of the (slower) hemodynamic response. Some researchers have tried to compensate for scattering changes by including a scattering spectrum term in Eq. 2, as if it represents a third chromophore.³³ Care should be taken with this approach, since it does not completely account for scattering effects, and incorporating a third chromophore (which may not in fact be changing) may adversely affect the fit results of the other two chromophores ( ${\mathrm{HbO}}_2$ and $\mathrm{HbR}$ ).

Optical coherence tomography of functional brain activation

Optical coherence tomography (OCT) is a well established technique for high resolution depth-resolved imaging of structures such as the eye.¹⁰⁸ OCT overcomes the effects of light scattering to obtain high resolution depth-resolved images by using coherence to isolate light that has been directly backscattered. This light therefore has the shortest possible pathlength through the tissue. OCT is also typically performed using NIR wavelengths $>1\mathrm{\mu }\mathrm{m}$ . As a result, OCT signals carry very little information about the hemoglobin absorption properties of tissue [Eq. 1 and Fig. 1] and cannot be used to measure fluorescence contrast. The contrast in OCT is therefore generally due to refractive index mismatches, and not due to changes in absorption corresponding to hemoglobin concentration or oxygenation. Nevertheless, OCT has recently been applied to brain imaging to explore depth-resolved changes in the cortex of cats and rats during functional stimulus.^{109, 110, 111}

To date, the functional OCT changes reported have followed the slower time scale of the hemodynamic response. Both increases and decreases in OCT contrast are seen in close proximity, at times when $\mathrm{HbT}$ should be increasing. It is possible that OCT could also detect fast neuronal scattering signals. However, the slower signals observed to date are hypothesized to correspond to changes in the speed and distribution of scattering red blood cells, and changes in diameter and tone of pial vessels and capillaries.

Laminar optical tomography

While 2-D optical imaging of the exposed cortex has found many very valuable applications, there remains concern over the fact that images are a 2-D projection of a combination of signals from superficial vessels and deeper layers of the cortex. The vascular structure of the cortex is organized with large arteries and veins on the surface, and then arteriolar and venous branches diving almost straight down to deeper capillary beds. These capillary beds supply the active neurons, which themselves are also organized in specific layers within the cortex. Therefore, not only are neuronal and vascular signals likely to be quite heterogeneous as a function of depth, but the laminar variations in neuronal and hemodynamic responses are of great interest. Laminar optical tomography (LOT) is a recently developed imaging technique that can add a depth dimension to hyperspectral imaging of the exposed cortex.^{19, 112, 113}

LOT utilizes a combination of scanning microscopy instrumentation and optical tomographic image reconstruction techniques to acquire noncontact measurements of the cortex at more than 40 frames per second. It can image to depths of $>2\mathrm{mm}$ with 100 to $200\text{-}\mathrm{\mu }\mathrm{m}$ resolution. A recent result is shown in Fig. 3f, where LOT was used to image the exposed-cortex hemodynamic response to rat forepaw stimulus at 473 and $532\mathrm{nm}$ . The depth-resolved images were validated by comparison to the vascular architecture, and microsvascular density determined via two-photon imaging of fluorescent vascular casts from the same animals. A spatiotemporal delineation technique was also applied that exploited the distinct differences in the temporal evolution of the functional signal in the different vascular compartments: arterioles, capillaries, and veins. Figure 3f shows two cross sectional $(x-z)$ images through the ${\mathrm{HbO}}_2$ cortical response, one transecting a superficial vein, and one transecting the deeper capillary response. A 3-D rendering of the three vascular compartments, delineated via their distinctive time courses, is also shown, revealing superficial arterioles and venules, and deeper capillary responses. A two-photon image of a vascular cast from the same animal is shown alongside, with the corresponding arterioles and veins highlighted. The temporal signatures of each different vascular compartment were cross-validated with in-vivo two-photon imaging and provided new insights into some of the basic assumptions of the hemodynamic models for fMRI interpretation described before.¹⁹

Since LOT detects scattered light, it is highly sensitive to blood oxygenation and $\mathrm{HbT}$ changes and is also suitable for fluorescence imaging. As a result, LOT in now being developed to allow simultaneous depth-resolved imaging of hemodynamics and VSDs or CaSDs. Depth-resolved imaging of VSDs and CaSDs would allow the laminar properties of neuronal responses to be studied in detail. Currently, such measurements can only be made using highly invasive depth-resolved electrode arrays, although demonstration of a gradient index lens-based system also recently showed promising results imaging VSDs in mouse cortex to depths of $150\mathrm{\mu }\mathrm{m}$ .¹¹⁴

Functional processing

A recent example of in-vivo two-photon microscopy of brain function was the use of calcium-sensitive dyes to look at visual responses in cats and rats.^{3, 127} Data were acquired using a custom-built two-photon microscope, and used Oregon Green 488 BAPTA-1 AM dye (Invitrogen) excited at $800\mathrm{nm}$ to provide calcium-sensitive fluorescent contrast. The dye was pressure injected into the cortex as described before for 2-D imaging of CaSDs. The complex cortical patterns of the response to visual stimuli of different orientations [similar to the ocular dominance maps in Fig. 3b] were found to occur even on the scale of discrete individual neurons.

Functional mechanisms

The advent of functional two-photon microscopy has allowed the mechanisms of neurovascular coupling to be studied in vivo in unprecedented detail. Whereas 2-D optical imaging of the cortex, as described earlier, has enabled parametric comparisons between electrophysiology and the hemodynamic response,¹⁸ two-photon imaging enables the interrelations between individual neurons, glial cells, and vessels to be examined in intricate detail.

Both confocal microscopy¹²⁸ and two-photon microscopy⁷⁹ have been used to examine the dynamics of red blood cells flowing through capillaries in the exposed cortex. In Ref. 79 and more recently in Refs. 116, 129, the capillary response to stimulation and hypercapnia have been examined using fluorescent plasma tracers such as dextran conjugated fluorescein or Texas dextran red (Invitrogen). These high molecular weight dyes cannot pass the blood brain barrier or stain red blood cells, and so provide strong contrast to vessels. Typical in-vivo image stacks of fluorescein-perfused vessels in rat brain are shown in Fig. 4c. The dark stripes and dots are moving red blood cells whose dynamics can be quantified from their motion in sequential images or line scans.¹²⁶ The flow and dilation dynamics of pial arterioles and veins was also recently investigated using two-photon microscopy.¹⁹

Figure 4d (left) shows simultaneously acquired in-vivo rat brain measurements of neurons labeled with Oregon Green 488 BAPTA-1 AM dye and vessels containing Texas dextran red. Both color channels were acquired with two-photon excitation at $800\mathrm{nm}$ and 550-long-pass dichroic and emission filters to separate the two emission bands onto different photomultiplier detectors. Figure 4d (right) shows similar data acquired in a transgenic mouse, engineered such that only certain neurons in the brain express green fluorescent protein.^{130, 131} Capillaries perfused with Texas dextran red can be seen to be weaving amidst the webs of dendrites.

2-D optical imaging of hemodynamic and CaSD changes can also be acquired prior to imaging the same animal with two-photon microscopy. Figure 4f illustrates 2-D imaging of the calcium-sensitive Oregon Green 488 BAPTA-1 AM response in the somatosensory cortex of a rat during forepaw stimulus ( $1.5\mathrm{mA}$ , $3\text{-}\mathrm{\mu }\mathrm{s}$ pulses at $3\mathrm{Hz}$ for $4\sec$ ), followed by two-photon imaging of individual neurons at a depth of $315\mathrm{nm}$ in the cortex responding to the same forepaw stimulus. The two-photon fluorescence of the CaSDs increased by up to 17% during stimulation.¹²⁶

Such direct and simultaneous observations of both neuronal and hemodynamic responses and structures in in-vivo systems enable the intricate mechanisms underlying neurovascular coupling relations to be closely examined and compared to macroscopic-scale measurements and even fMRI results.

Two recent studies used similar two-photon imaging techniques to present compelling evidence supporting the role of astrocytes as active mediators of blood flow control.^{80, 132} Astrocytes make up a large fraction of the cells in the brain, and have been shown to be both metabolically active and capable of transmitting signals between neurons and blood vessels.^{133, 134} In Ref. 132, astrocytes in mouse brain were labeled with calcium-sensitive rhod-2 AM, caged ${\mathrm{Ca}}^{2+}$ , and DMNP-EDTA AM, while the vasculature was perfused with dextran conjugated fluorescein. Two-photon imaging of astrocytic endfeet contacting arteries was performed in vivo at $825\mathrm{nm}$ , with spectrally separated detectors for rhod-2 AM and fluorescein. A second laser $\left(355\mathrm{nm}\right)$ was then focused onto an endfoot and pulsed to cause photolysis of the DMNP-EDTA, triggering intracellular uncaging of calcium in the endfeet. This forced increase in calcium caused the endfoot to dilate. In a subsequent study, a CaSD (fluo-4 AM) and astrocyte-specific dye SR101 were used to demonstrate that similar calcium increases in astrocytic endfeet could be observed during prolonged whisker stimulation in vivo.⁸⁰

Functional imaging to investigate pathologies and treatments

In-vivo two-photon imaging of intravascular dextran-conjugated fluorescein in rats was recently used to investigate the effects of stroke on blood flow dynamics, as well as the recovery of flow following two treatments: hemodilution and injection of a recombinant tissue-type plasminogen activator.¹³⁵ The study also created the strokes using a second laser delivering targeted high fluence ultra-short NIR pulses to cause vessel rupture, intravascular clotting, or extravasation of blood components depending on the laser power.

Alzheimer’s disease research has been greatly impacted by the availability of fluorescent markers that bind to beta amyloid plaques in the brain.^{5, 136, 137} The so-called Pittsburg compound B [2-(4’-methylaminophenyl)-6-hydroxy-benzothiazole, PIB] can be used as a clinical contrast agent for PET imaging and is showing promise as a diagnostic tool in humans.¹³⁸ However, since PIB is fluorescent, it can also be used as a two-photon microscopy contrast agent for exposed cortex imaging.¹¹⁷ Tg2576 APP transgenic mice develop beta amyloid plaques over the course of their lives, and are assumed to provide a realistic model of the progression of Alzheimer’s disease. Since Alzheimer’s in humans can currently only be confirmed postmortem, these mouse models offer an opportunity to investigate the early effects of the disease, as well as learn about its progression and to allow development of earlier treatment and diagnostic methodologies.

Mice with chronically implanted cranial windows can be repeatedly imaged with two-photon microscopy for up to a year. This has allowed two-photon microscopy to be used to evaluate the effect of developing amyloid- $\beta$ plaques on the structure and function of surrounding neurons and blood vessels. Figure 4e shows three-channel two-photon image stacks of the brain of a 21 to 24 month old transgenic mouse with beta amyloid deposits shown in blue, blood vessels (perfused with Texas dextran red) shown in red, and neurons expressing green fluorescent protein in green. The amyloid binding dye (methoxy-XO4, similar to PIB) was injected intraperitoneally the day before imaging. The neurons were labeled by transfection from an adeno-associated virus injected into the brain 2 to 3 weeks prior to imaging.¹⁴ Extensions of this work to noninvasive animal and perhaps even human imaging are being explored.¹³⁹

Direct topography. Direct topography is the most widespread approach to clinical optical brain imaging and is often referred to as near-infrared spectroscopy (NIRS). In its simplest form, NIRS uses light source and detector pairs positioned on the head of infants or adults. Light that passes through the skull scatters within the cortex of the brain [as illustrated in Figs. 1b, 5b, 5f]. Despite significant attenuation, variations in absorption due to changes in local $\mathrm{HbT}$ , ${\mathrm{HbO}}_2$ , and $\mathrm{HbR}$ concentration result in detectable modulations in the intensity of the emerging light. If an array of sources and detectors is used, the changes measured between each source-detector pair can be approximately mapped or interpolated to the underlying cortex, and a low-resolution $(>5\mathrm{mm})$ 2-D topographical map of cortical activation can be produced.^{152, 154, 173} When many overlapping and multidistance topographic measurements are acquired, it is possible to use mathematical models of light propagation to create a depth-resolved reconstruction of functional activation.^{174, 175, 176} While more complex, this approach can overcome several of the limitations of direct topography, as described below.

Instrumentation. Noninvasive optical brain imaging instrumentation may be continuous wave (cw), frequency domain (FD), or time domain (TD). cw is the most common approach and simply measures the intensity of light emerging from the head at two or more wavelengths. Data acquisition is often frequency encoded (up to $\sim \mathrm{kHz}$ ), such that multiple wavelengths^{177, 178} or multiple sources ^{149, 179, 180, 181} can be measured simultaneously. FD and TD instruments are substantially more complex and expensive, and can measure both a cw component and the time taken for light to pass through the tissue.¹⁸² FD instruments achieve this by modulating their illumination at around $100\mathrm{MHz}$ and measuring the phase delay of the emerging light.^{152, 183} TD instruments use a pulsed laser and measure the temporal distribution of the photons emerging from the tissue. ^{145, 184, 185, 186, 187, 188} The main purpose of this temporal information is to provide a measure of the scattering properties of the head, as illustrated in Fig. 6 .

Fig. 6

Time-domain, frequency-domain, and continuous wave measurement sensitivities to absorption and scatter. Schematic shows simplistically how absorbing and scattering structures will change TD and FD data, but that in transmission, their effect on cw data will be indistinguishable.

Optical fibers are generally used to transmit light to and from the head (see Fig. 5). Direct placement of small sources and detectors on the head has also been demonstrated.^{166, 189}

Calculation of ${\mathit{HbO}}_{\mathit{2}}$ and HbR concentrations. The change in intensity measured between a topographic source and detector (the attenuation) will generally reflect a change in the absorption of underlying tissue [Eq. 1]. The basic approach for deriving changes in ${\mathrm{HbO}}_2$ and $\mathrm{HbR}$ concentration from changes in attenuation was outlined in Sec. 1.1.2 [Eqs. 1, 5]. Determining $I_0$ and $G$ for noninvasive measurements is particularly challenging, since the amount of light coupled into and out of the scalp for every source and detector position would need to be known. Differential measurements relative to time $t=t_0$ [Eq. 3] should cancel these source- and detector-specific scaling factors, although it must be assumed that these factors stay constant for the duration of the measurement (which can be challenging if the subject moves and/or hair is obstructing probe placement).

As with exposed cortex imaging, if measurements of the change in attenuation at multiple NIR wavelengths are to be converted into changes in hemoglobin concentrations, the effects of scattering must be accounted for. If Eqs. 4, 5 are to be used for “direct” topography, an estimate of $\mathrm{DPF}x$ for each wavelength and measurement pair is required.

FD and TD systems can be used to measure $\mathrm{DPF}x$ directly, and as an adjunct to cw imaging can provide $\mathrm{DPF}x$ for a given wavelength by measuring the average time taken for light to pass from the source to detector, and converting to distance based on the speed of light in tissue.^{190, 191} However, if a subset of $\mathrm{DPF}x$ measurements (or generalized values of DPF) are to be used, it should be noted that $\mathrm{DPF}x$ could vary depending on age, gender, and even position on the head, owing to the complex structures of the underlying scalp, skull, and convoluted human cortex.

Simulations for calculation of $\mathrm{DPF}x$ can become increasingly complex if efforts are made to incorporate the curvature of the head, the heterogeneous structures of the scalp, skull, and cortex, the presence of cortical convolutions and sulci, the effects of CSF [see Fig. 1b], and even the effects of isotropic scattering in the brain. ^{88, 192, 193, 194, 195, 196, 197, 198, 199} Estimates of the absorbing and scattering optical properties of each tissue type,^{200, 201} as well as accurate knowledge of the source-detector separations between fibers on the head, are also required. If available, an MRI scan of the head can be used to build a customized model of light propagation for each subject, allowing constrained analysis that accommodates the exact structure of the brain, skull, and scalp underlying each detector.²⁰² However, while this is feasible in studies where fMRI is simultaneously acquired, it is prohibitive for other applications of optical imaging.

Although TD and FD systems rarely have a large number of multiwavelength channels,^{161, 185} they have the advantage that they can be used to deduce the scattering properties of the head, which can then be incorporated into more complex calculations of hemoglobin concentrations.^{201, 203} An FD approach is to measure between several source-detector separations and, under simplifying assumptions of homogeneity, fit for scattering and absorption based on a linearized model of light propagation versus distance.¹⁸³ Such approaches do not require estimates of $\mathrm{DPF}x$ .

In general, for direct topography, simplified models and scaling factors are applied to account for $\mathrm{DPF}x$ and also partial volume effects, which correct for the fact that only a fraction of the detected light actually sampled the cortex. ^{34, 153, 159, 183, 204} Several studies have explored the importance of accurate calculation/measurement of $\mathrm{DPF}x$ and its impact on the determination of hemoglobin concentrations for optical topography.^{198, 205, 206} It has been shown that in many cases, useful functional information can be obtained even when simplifying assumptions are applied to analysis and interpretation. However, it is therefore important to consider the possible impact of systematic errors resulting from this analysis for all NIRS measurements, particularly when imaging the adult head, where many confounding effects can obscure cortical signals and impact correct interpretation.

Extracerebral effects and physiological noise. The effects of blood flow changes in the scalp, and changes in the brain that are not directly related to the cortical stimulus response, must also be considered. The analysis approaches described above assumed that changes in signal were solely due to changes in the cortex of the brain. However, it has also been shown that breathing $(\sim 0.16\mathrm{Hz})$ , heart rate $(\sim 1\mathrm{Hz})$ , and the “vasomotor” or Mayer wave signal $(\sim 0.1\mathrm{Hz})$ ^{207, 208, 209, 210} can cause large changes in measured hemodynamic signals. Substantial repetition of a stimulus and averaging should reduce the effects of unrelated physiological variations in blood flow in the scalp and brain. However, acquisition of so many repetitions is not always possible with adult or infant subjects. In addition, some responses may not average out, or may intrinsically correlate or become entrained with the stimulus period,²⁰⁸ and can affect both brain blood flow and the rich blood supply to the scalp. Recent evidence has suggested that certain stimuli actually invoke a systemic response (possibly through heart rate or blood pressure changes) that could manifest as a change in scalp or even baseline cortical blood flow that correlates with the stimulus period.^{153, 211}

Several measurement, data analysis, and filtering approaches have been proposed to reduce the influence of hemodynamic activity unrelated to the true cortical response. One approach is to simultaneously acquire measurements between close and more distant pairs of sources and detectors, assuming that the closer detector is predominantly detecting signal from the scalp. This scalp signal can then be approximately removed from the signal from the more distant detector, assuming that a similar superficial contribution is affecting both measurements. A similar approach can be achieved with TD data, where the photons arriving at a detector earlier are likely to have traveled more superficially, and therefore contain information about the scalp signal. The benefit of using TD is that later photons that have traveled more deeply between the same source detector travel through the same region of the scalp as the early photons, so cancellation should be improved relative to the multidistance approach.^{187, 212} An extension of this is reconstructed topography, which creates a depth-resolved map that should be able to spatially separate superficial signals from deeper cortex signals, if suitable measurements are acquired.

Spatiotemporal filtering and modeling have also been proposed as a way to remove these systemic signals.^{213, 214} While rapidly becoming more sophisticated, implementation of these data processing techniques can be complex and sometimes subjective. Extensive processing could result in distortion of true cortical signals, particularly when attempting to separate scalp and cortical signals that may both have elements correlating with the stimulus response.²¹¹

Practical measurements

A final difficulty with acquiring functional optical imaging data is reliable placement of sources and detectors on the head. The primary problem is to be able to securely position multiple probes in a pattern such that their interfiber distances are known, and they stay stable during the course of the measurement. Hair on the head is a major challenge, as it can significantly attenuate light entering and leaving the head, and can cause measurement instability over time if the probes shift. Technical advancements have been made in positioning optical fibers and probes onto the head, and relatively stable configurations have been devised.¹⁸⁰ However, further work will be required before it becomes possible to routinely, rapidly, and repeatably position optical topography fibers on the head.

Reconstructed topography. Direct topography is the simplest way to analyze topographic data, but relies on many assumptions and simplifications. An alternative approach is to reconstruct data onto a 3-D rendering of the head. Although analysis becomes more complex, by creating a depth-resolved image of the measured absorption changes, some of the confounds described earlier can be overcome.

Image reconstruction.

Instead of requiring a pathlength estimate, reconstruction effectively attempts to recreate the measured data by simulating alterations in the absorption distribution within the head.^{187, 215, 216} When the simulated data and the measured data match, the absorption distribution is assumed to be correct. Reconstructing still generally requires an estimate of the baseline absorbing and scattering optical properties of the head, and an accurate model of light propagation, but can help if there are depth-dependent variations in the response, including effects from the scalp. Partial volume effects, particularly where one wavelength of light effectively probes a different region, than a second wavelength, can also be accommodated in reconstructed topography.

For reconstructed topography to be effective, it is necessary to measure signals between overlapping sources and detectors, and generally at several source-detector distances [see Fig. 5b]. These measurements provide a set of mutually consistent data for which there should be only one possible corresponding differential absorption distribution. This in turn should also lead to improved $x\text{-}y$ resolution.^{175, 176, 179}

Modeling of light propagation in the human head typically uses either Monte Carlo modeling or the diffusion approximation (DA) to the radiative transfer equation [Eq. 6] and a linearizing assumption such as the Born [Eq. 7] or Rytov approximations.

Analytical solutions to the DA for simple shapes such as slabs and semi-infinite planes exist.¹⁹⁷ More complex approaches include applying the DA via finite element or finite difference methods, where heterogeneous structures can be incorporated into the simulation. Monte Carlo modeling can also be used to simulate light propagation in the head, although for large tissue volumes, it can be prohibitively slow. Once the model has been used to generate the Jacobian $J_{n,m}\left(r\right)$ (otherwise known as the spatial sensitivity matrix, weight matrix, or photon measurement density function), Eq. 7 can be solved for $\mathrm{\Delta }{\mu }_a$ using a linear inversion technique such as Tikhonov regularization, or more complex nonlinear iterative algorithms.²¹⁷

Plots of $J_{n,m}\left(r\right)$ [Fig. 5f] reveal the varying paths of light traveling from source to detector, and show how only some of the detected light has sampled the cortex. The areas of highest sensitivity are directly below the source and detector, revealing why changes in coupling and the scalp have such a significant effect on signal amplitude. The banana shape of the light path is characteristic of diffuse imaging, and is maintained as source-detector distances get larger or smaller, such that the center region gets deeper or shallower, respectively.

Implementation of reconstructed topography can be achieved with various levels of complexity. The simplest approach assumes that changes in absorption are occurring in a homogenous volume of known optical properties. This is an approximation, since the baseline structures of the skull and brain will affect the shape of $J_{n,m}\left(r\right)$ [as shown in Fig. 5f] in a way that cannot be accounted for in a homogenous model.^{198, 218, 219} The effects of CSF have also been widely considered, since the diffusion approximation is only valid for scattering regions where ${\mu }_s^{\prime }\gg {\mu }_a$ , which is not likely to be true for large thicknesses of CSF.^{196, 199} Several papers further suggested that light piping could occur in the CSF layer, significantly reducing penetration of light into the cortex. However, more recent publications have questioned the magnitude of the effect once the complex 3-D nature of the CSF layer, trabeculi, vessels, and refractive index differences are taken into account.^{193, 194, 217} By incorporating a priori structural information into the image reconstruction scheme, constraints can be applied to help improve the localization of activation to the cortex, although care must be taken to account for changes in the data that may originate from extra-cerebral contributions.²⁰²

Data acquisition. As NIR light sources and detectors become increasingly affordable, it has become possible to construct NIRS instruments with large numbers of multiwavelength source and detector channels. While larger arrays allow simultaneous mapping of several regions of the cortex simultaneously, they also present significant challenges for fast data acquisition.¹⁸⁰ Since the human heart beat causes substantial variations in NIRS signals,²²⁰ aliasing can be a significant problem unless measurements from all source-detectors pairs can be acquired at a rate of at least $2\mathrm{Hz}$ . Systems that utilize frequency encoding to allow all sources to be illuminated simultaneously can acquire data in parallel at high frame rates.^{152, 181} However, the attenuation difference between a source and detector separated by 2.5 and $4.25\mathrm{cm}$ can be up to two orders of magnitude.¹⁷⁹ For reconstructed topography, these multidistance measurements are vital, and detectors must have excellent dynamic range and noise characteristics to be able to simultaneously distinguish between frequency-encoded signals of such varying amplitudes. Development of faster systems, with better dynamic range or highly advanced and rapid multiplexing schemes is therefore ongoing.^{176, 179}

Exogenous dyes for noninvasive imaging. Exogenous contrast agents can also be used for noninvasive brain imaging. Indocyanine green (ICG) is an FDA-approved dye commonly used for liver and cardiac function tests.²⁴⁴ ICG has strong NIR absorption and fluorescence, and does not cross the blood brain barrier, and so has also found several optical brain imaging applications. A bolus of ICG can be used to quantify blood flow in the brain, by observing the rate of arrival and accumulation of the dye with NIRS measurements.²⁴⁵ Changes in the dynamics of a dye bolus in patients with perfusion abnormalities such as thrombolitic stroke have also been studied using NIRS.²⁴⁶ ICG has been used to simultaneously determine cerebral blood volume and total circulating blood volume in neonates undergoing blood transfusions.²⁴⁷ An interesting recent study reported that it was possible to also observe the fluorescence of ICG in the brain using noninvasive NIRS.²⁴⁸ This result is important as it supports the possibility that more specifically targeted long-wavelength dyes could feasibly be measured noninvasively in the brain. Such molecular probes may have sufficient specificity to allow localization or even diagnosis of diseases such as brain cancer and Alzheimer’s.^{139, 249}

Noninvasive neuronal imaging via fast scattering. Since it is not possible to use voltage-sensitive dyes in humans, researchers are highly motivated to isolate a strong source of contrast that can directly report changes in neuronal activity. Observation of the fast scattering signal in exposed rodent cortex was described in Sec. 2.1.4. Noninvasive measurement of this signal is highly challenging, as scattering changes are difficult to detect accurately and separate from the effects of absorption. Recent studies of the fast scattering signal have shown compelling results, and compared findings to EEG and fMRI.^{161, 250, 251} However, it is still questioned whether the fast signal is sufficiently robust for routine measurement.^{104, 105}

Diffuse correlation spectroscopy. A relatively recent addition to optical topography has been the development of diffuse correlation spectroscopy (DCS), which allows noninvasive measurement of cortical blood flow. The theoretical basis of this technique is similar to that of speckle-flow imaging described in Sec. 2.1.2, but DCS measures the temporal autocorrelation of laser light with a long coherence length that has scattered through tissue containing moving red blood cells. For human imaging, DCS acquires signals through single mode fibers placed on the scalp.^{252, 253, 254} Promising results in humans have been reported,²⁵⁵ and it has been shown that it is possible to extend the technique to allow 3-D imaging.²⁵⁶