2.1.

Optical Neuroimaging System Specifications

To create a wearable, whole-head imaging system that weighs $<1\mathrm{lb}$, the commonly used 2.5-mm-diameter optical fibers need to be $\sim 10$-fold smaller in diameter. Commercially available $200\text{-}\mu \mathrm{m}$-diameter optical fibers are closest to matching this requirement ($\mathrm{NA}=0.5$, FP200URT, Thorlabs, New Jersey).² However, the major challenge in redesigning HD-DOT with $200\text{-}\mu \mathrm{m}$-diameter fibers is the decrease in SNR that results from reducing the size of the fiber. To determine a detection design that enables the use of $200\text{-}\mu \mathrm{m}$-diameter fibers for HD-DOT, we begin by characterizing the light detection capabilities of APDs and CMOS sensors.

A useful metric for comparing these two detectors is detectivity, which is defined as the NEP divided by the area of the light collection

A detectivity of $\sim 100\mathrm{fW}/\surd \mathrm{Hz}\text{-}{\mathrm{mm}}^2$ is needed for high-quality HD-DOT.² We compared the effective detectivity of six commercially available APDs to the superpixel sCMOS detection (Appendix B). The Hamamatsu C12703-01 had the lowest baseline NEP and the lowest effective detectivity with $200\text{-}\mu \mathrm{m}$-diameter fibers for HD-DOT (Table 2 and Appendix B). For this reason, we compared all the superpixel values with the C12703-01 APD module. Although the Hamamatsu C12703-01 module has the lowest detectivity of the detectors available on the market and is currently used for HD-DOT imaging with 2.5-mm-diameter optical fibers, the detectivity is still $\sim 70$-fold too high for use with $200\text{-}\mu \mathrm{m}$-diameter fibers.² Short source–detector separation measurements that sample the scalp and skull would have high SNR, but the longer separation measurements that sample the brain would have insufficient SNR. To measure these longer separation measurements, we require a detector with lower detectivity at a $200\text{-}\mu \mathrm{m}$-diameter collection.

Intriguingly, CMOS cameras have a 10,000-fold lower NEP at the single pixel level. Although the detectivity of a single pixel is sixfold higher than the APDs due to their small size, the fiber delivering light with 1:1 magnification will be a $200\text{-}\mu \mathrm{m}$-diameter circle on the sensor. Summing pixels within that $200\text{-}\mu \mathrm{m}$-diameter circle to create a “simple-binned” detector is predicted to match the requirements for HD-DOT imaging.

The rest of this section compares the CMOS characteristics to the commercially available APD module with the lowest NEP (Hamamatsu C12703-01) and is organized as follows: Sec. 2.1.1 explains the NEP, detectivity, and DNR specifications of the CMOS camera for individual pixels and simple-binned pixels at a standardized specification bandwidth of 1 Hz. Section 2.1.2 documents the 1-Hz specifications using a superpixel algorithm to sum the pixels while removing noise sources. Section 2.1.3 explores the effect of HD-DOT system parameters such as frame rate and encoding on the “effective” specifications of the superpixel system. Lastly, Sec. 2.1.4 details the specifications for the APD-based HD-DOT imaging system. All specifications discussed as follows are documented in Table 1.

Table 1

HD-DOT system specifications. Theoretical predictions and experimental measurements of specifications for HD-DOT at 830 nm. Effective NEP and detectivity for the 200-μm superpixel HD-DOT are calculated using 6-Hz frame rate, 25 encoding steps, and 1 wavelength. Effective specifications for the APD are calculated using 16 timesteps at 10 Hz. The area for simple-binned and superpixel calculations was determined by multiplying a single-pixel area by the number of pixels (N=697).

2.1.1.

Pixel and simple-binning NEP, DNR, and detectivity at 1 Hz

The single-pixel NEP at 1 Hz (${\mathrm{NEP}}_{\mathrm{pix}}$) is calculated by relating the 1-Hz noise floor of the pixel to the optical power incident upon the pixel (Appendix A). Throughout our calculations, we assume that the primary noise source that contributes to the NEP of CMOS pixels is read-out noise. At 830 nm, the Zyla 5.5 sCMOS ${\mathrm{NEP}}_{\mathrm{pix}}$ is $0.0022\mathrm{fW}/\surd \mathrm{Hz}$, four orders of magnitude lower than the APD NEP of $20\mathrm{fW}/\surd \mathrm{Hz}$. By summing $N$ pixels together, the NEP of the summed pixels (${\mathrm{NEP}}_{\text{sum}}$) is related to the NEP of a single pixel by

Eq. (3)

$${\mathrm{NEP}}_{\text{sum}}(N)=\sqrt{N}{\mathrm{NEP}}_{\mathrm{pix}}.$$

The ${\mathrm{NEP}}_{\text{sum}}$ increases as the square root of $N$ because the variance of the read-out noise increases by a factor of $N$ [Fig. 1(a) and Appendix A]. Using Eq. (3) with $N=697\text{pixels}$ corresponding to the size of the optical fiber on the CMOS chip, the theoretical ${\mathrm{NEP}}_{\text{sum}}=0.058\mathrm{fW}/\surd \mathrm{Hz}$ (Table 1), $\sim 350$-fold lower than the NEP of the C12703-01 APD modules (${\mathrm{NEP}}_{\mathrm{APD}}=20\mathrm{fW}/\surd \mathrm{Hz}$).

Fig. 1

SP-DOT system specifications as a function of the superpixel size. (a) The NEP should increase as the square root of the number of pixels summed if the noise source is limited by read-noise (blue dashed line). Without subtracting off the correlated row noise (red line), the noise floor begins to deviate from the theoretical values at about 90 pixels. The size of the fiber tip on the CMOS sensor imaged at 1:1 magnification is $200\mu \mathrm{m}$, denoted by the vertical dashed line. The (b) effective detectivity ($D_{\mathrm{eff}}$) and (c) DNR that are calculated based on the NEP are subsequently decreased and increased with larger $N$, respectively. After removing the correlated row noise with the superpixel algorithm (green line), the NEP, DNR, and $D_{\mathrm{eff}}$ are within a factor of 2 of the theoretical values. In addition, the DNR satisfies the required specifications for HD-DOT neuroimaging denoted by the green-shaded areas. The Deff value is $4\times$ higher than what is required for HD-DOT but this can be accommodated by a $4\times$ increase in light input.

The 1-Hz detectivity of a pixel ($D_{\mathrm{pix}}$) is defined as the ${\mathrm{NEP}}_{\mathrm{pix}}$ divided by the area of the pixel ($A_{\mathrm{pix}}$). For 830 nm, the Zyla 5.5 sCMOS, $D_{\mathrm{pix}}=52\mathrm{fW}/\surd \mathrm{Hz}\text{-}{\mathrm{mm}}^2$, 24-fold lower than the detectivity of the APD system using $200\text{-}\mu \mathrm{m}$-diameter fibers ($D=1270\mathrm{fW}/\surd \mathrm{Hz}\text{-}{\mathrm{mm}}^2$).² When summing pixels, the area increases linearly with the number of pixels summed. Combining this with, Eq. (3) provides the detectivity of summed pixels ($D_{\text{sum}}$)

Eq. (4)

$$D_{\text{sum}}(N)=\frac{\sqrt{N}{\mathrm{NEP}}_{\mathrm{pix}}}{N{A}_{\mathrm{pix}}}=\frac{D_{\mathrm{pix}}}{\sqrt{N}}.$$

Thus, the detectivity is inversely proportional to the $\sqrt{N}$ [Fig. 1(b)]. Using Eq. (4), the theoretical $D_{\text{sum}}$ for a $200\text{-}\mu \mathrm{m}$-diameter circle of pixels ($N=697$) is $2.0\mathrm{fW}/\surd \mathrm{Hz}\text{-}{\mathrm{mm}}^2$ (Table 1), 636-fold lower than the detectivity of the APD modules with $200\text{-}\mu \mathrm{m}$-diameter fibers ($D_{\mathrm{APD}}=1273\mathrm{fW}/\surd \mathrm{Hz}\text{-}{\mathrm{mm}}^2$).

The single-pixel DNR (${\mathrm{DNR}}_{\mathrm{pix}}$) is defined as the full well of the pixel ($w$) divided by the noise floor ($\sigma$). A single pixel has a ${\mathrm{DNR}}_{\mathrm{pix}}$ of $1.2\times {10}^4$, two orders of magnitude lower than the APD and one order of magnitude too low for HD-DOT imaging. As the full well increases linearly with the number of pixels summed and the noise floor increases as the square root, the ${\mathrm{DNR}}_{\text{sum}}$ increases as the square root of the number of pixels summed [Fig. 1(c)]

Eq. (5)

$${\mathrm{DNR}}_{\text{sum}}(N)=\frac{Nw}{\sqrt{N}\sigma }=\sqrt{N}{\mathrm{DNR}}_{\mathrm{pix}}.$$

Using Eq. (5), the theoretical ${\mathrm{DNR}}_{\text{sum}}$ for a $200\text{-}\mu \mathrm{m}$-diameter circle of pixels on the sCMOS ($N=697$) is $3.2\times {10}^5$ (Table 1), approximately threefold higher than the required DNR for HD-DOT.²

To test Eqs. (3)–(5) experimentally, we measured the specifications for a sCMOS sensor (Zyla 5.5, Andor Technologies) using a $200\text{-}\mu \mathrm{m}$-diameter fiber relayed to the CMOS sensor using $1:1$ magnification (Table 1). By experimentally summing $N=697\text{pixels}$ ($6.5\mu \mathrm{m}\mathrm{w}\mathrm{i}\mathrm{d}\mathrm{t}\mathrm{h}\times 6.5\mu \mathrm{m}\mathrm{h}\mathrm{e}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t}\text{pixels}$) in a circular shape that corresponded to the $200\text{-}\mu \mathrm{m}$-diameter optical fiber, we were able to measure the NEP of summed pixels and therefore calculate the summed pixel detectivity and DNR. The detectivity of the summed pixels is a factor of $\sim 2$ higher than the theoretical values (experimental $D_{\text{sum}}=3.4\mathrm{fW}/\surd \mathrm{Hz}\text{-}{\mathrm{mm}}^2$ versus theoretical $D_{\text{sum}}=2.0\mathrm{fW}/\surd \mathrm{Hz}\text{-}{\mathrm{mm}}^2$). Similarly, the DNR of the summed pixels is a factor of $\sim 2$ lower than the theoretical values (experimental ${\mathrm{DNR}}_{\text{sum}}=1.5\times {10}^5$ versus theoretical ${\mathrm{DNR}}_{\text{sum}}=3.2\times {10}^5$). The discrepancy between theoretical and experimental values of NEP and DNR is exacerbated at large fiber diameters (Fig. 1), limiting the use of fibers to below $\sim 100\mu \mathrm{m}$. To use $200\text{-}\mu \mathrm{m}$-diameter optical fibers, we developed a superpixel referencing algorithm to improve the detectivity and DNR of the CMOS sensor.

2.1.2.

Superpixel NEP, detectivity, and DNR at 1 Hz

Although simple-binning of pixels improves the CMOS sensor’s detectivity and DNR, our experimental measurements of detectivity and DNR were worse than our theoretical predictions (Table 1 and Fig. 1). This mismatch between experimental results and Eqs. (3)–(5) was primarily caused by structured, non-Gaussian noise on the CMOS sensor. In addition to the shot-noise measured across the entire sensor (e.g., as measured in a background image), there is correlated noise within each row of the CMOS sensor caused by a voltage offset after each row readout [Fig. 2(a)]. Upon removal of this row noise, the temporal noise decreases by a factor of 2 [Fig. 2(b)].

Fig. 2

Correlated row noise. (a) Due to the readout structure of the sCMOS cameras, pixels within the same row have a correlated offset value after each readout. This can be demonstrated by plotting the values of two pixels as a function of time, which are either within the same column (pixels a and b) or same row (pixels c and d). The values of pixels a and b are not temporally correlated ($r=0.1$). The values of pixels c and d are highly correlated over time ($r=0.7$). (b) This correlated row noise can be removed by subtracting an average of multiple pixels within a row, reducing the temporal noise by $2\times$.

We calculated and removed the correlated row noise using a “superpixel algorithm.” The CMOS image of each fiber tip [Fig. 3(a)] was segmented into a core, buffer, and reference region using a system-specific mask [Fig. 3(b)]. Although the fiber tips required segmentation, the fiber tip locations and sizes did not change from experiment to experiment and thus the same segmentation mask was used run to run and day to day. The pixels segmented as the core regions are all pixels within a $100\text{-}\mu \mathrm{m}$-radius circle around a manually designated center pixel. This method based on distances ensured that the same number of pixels was summed for all fiber tips ($N=697\text{pixels}$). The core region corresponds to the image of the fiber tip, the reference region provides measurements for determining the correlated row noise, and the buffer region serves to avoid any optical or electronic crosstalk from the core pixels into the reference region.

Fig. 3

Superpixel algorithm. (a) A single raw sCMOS image of 10 detector fiber tips is background subtracted. (b) The image is partitioned into super pixels around each fiber tip and the values of the reference pixels are used for calculating the correlated row noise. (c) After subtracting the row noise, all core values are summed to create the final superpixel value per detector.

The row noise was removed using the following algorithm. Per row ($j$), the row noise ($R_j$) is defined by averaging the pixel values in the reference region of that row. To remove the row noise, we needed to scale the row noise by the number of pixels ($N$) in that row within the core. Thus, the scaled row noise is $N*R_j$. The superpixel value per row of the core ($S{P}_j$) is

Eq. (6)

$$S{P}_j=\sum _{i=1}^N(C_i)-N{R}_j.$$

Note that because the fiber tip is circular, the number of pixels summed ($N$) for the core varies per row. The total superpixel value for a single fiber tip is then the sum of all $j$ rows of the core

Eq. (7)

$$S{P}_{\text{total}}=\sum _{j=1}^{J}S{P}_j.$$

After applying the superpixel algorithm, a single CMOS image of 10 fiber tips is condensed into 10 superpixel values [Fig. 3(c)].

After removing these noise structures from the CMOS images, the noise [Fig. 2(b)], and therefore the NEP and detectivity of the $200\text{-}\mu \mathrm{m}$-diameter circle superpixel are decreased by $\sim 2\times$ compared with simply summing pixels. The DNR increased approximately twofold using the superpixel algorithm. Now, the DNR ($2.6\times {10}^5$) and detectivity ($D=2.4\mathrm{fW}/\surd \mathrm{Hz}\text{-}{\mathrm{mm}}^2$) are comparable with the same specifications of the APD using 2.5-mm-diameter optical fibers with 50% packing fractions, and closely match our theoretical predictions. Thus, with the superpixel summing algorithm, the DNR and detectivity for sCMOS detectors are promising for use in HD-DOT imaging when quoted at a 1-Hz bandwidth.

2.2.

Superpixel DOT System

To establish the feasibility of SP-DOT, we developed a 24-source and 28-detector system using a Zyla 5.5 sCMOS. The sources and detectors are located over the visual cortex to test the system performance with retinotopic mapping. This retinotopy system provides 264 usable source–detector measurements with a cap that weighs $<1\mathrm{lbs}$.

To relay the light from the detector fibers to the sCMOS chip, we designed dedicated detector boxes. The SMA end of the detector fibers screws into female SMA adapters on the custom-designed light-tight boxes. Inside the box, 0.75-m length $200\text{-}\mu \mathrm{m}$-diameter fibers ($\mathrm{NA}=0.5$, FP200URT, Thorlabs, New Jersey) transmit light from the detector fibers to a dedicated aluminum block containing the cleaved fiber tips [Fig. 4(a)]. A $4f$ relay (consisting of two Nikon 50 mm $f/1.2$ Nikkor Ai-S Manual lenses) provides a $1:1$ magnification of the cleaved fiber tips onto the Zyla 5.5 sCMOS chip (Andor, Belfast, UK). To take advantage of the horizontal readout speed of the sCMOS camera, the 28 detector fiber tips were positioned horizontally in the fiber array. The fibers were separated by $200\mu \mathrm{m}$ of aluminum to minimize optical crosstalk on the sCMOS chip, resulting in rows of 23 detectors.

Fig. 4

Superpixel HD-DOT system components. (a) The detector fibers terminate in an aluminum fiber array holder with the fibers separated by $200\mu \mathrm{m}$ (top). To take advantage of the readout speed of an sCMOS, the fibers are in rows of 36. The fiber array holder is imaged with a 1:1 magnification via a two-lens system (bottom). (b) The cap has bent fiber tips protruding through it that can slide in and out of the cap to match the curvature of the head. The source fibers have on average 2 mm of space to allow for the beam to expand for ANSI compliance. The detector fibers are flush with the scalp. Rubber stripping is used on the outside of the cap to provide a force toward the center of the cap. (c) The superpixel imaging cap consists of 24 sources and 28 detectors that cover the visual cortex. Due to the low weight of the fibers, the subject can support the cap and move around freely.

To optimize the amount of light reaching the head, we used laser diode sources (50 mW, 830 nm HL8338MG, 690 nm HL6750MG Thorlabs, New Jersey). The current delivered to the lasers was modulated by dedicated, high-bandwidth (20 MHz) digital input/output lines (PCI-6534; National Instruments, Austin, Texas). A single-source box contains 32 laser diode source channels and electronics in one standard 19-in. rack unit ($\text{height}=1\mathrm{RU}$ or 1.75 in.). To focus as much light as possible onto the $200\text{-}\mu \mathrm{m}$-diameter fibers (NA 0.5, FP200URT, Thorlabs, New Jersey), a custom-designed, three-dimensional (3-D) printed laser coupler was attached to the source box containing a 2.5-mm ball lens (N-BK7, Edmunds Optics) placed between the diode and fiber SMA tip (SMA905, 10230A, Thorlabs, New Jersey). The laser coupler transmitted 60% of the laser light into the fiber.

To improve the fiber coupling, ergonomics, and comfort of the imaging system cap, we bent the fibers into a 90-deg curve just before entering the cap [Fig. 4(b)]. Both source and detector fibers had an SMA905 tip on one end and a bent tip on the other to sit in the cap. All right-angle tips were custom-made and polished. To maintain rigidity within the cap, a clear plastic sleeve was glued around the fiber tip that passed through the cap. To provide a safe and comfortable surface that sits against the scalp, biocompatible, translucent epoxy (UV10MED, Masterbond, New Jersey) covered the tip of the fiber [Fig. 4(b)]. An average of 2 mm of the clear epoxy was placed on the tip of the fibers to allow for the light to expand to a beam diameter of 2 mm once entering the skin ($0.2\mathrm{mW}/{\mathrm{mm}}^2$). Detector fibers had the minimum amount of epoxy needed to sufficiently cover the tip, averaging $<0.2\mathrm{mm}$. In total, the retinotopy SP-DOT system consisted of 76, 5-m-long optical fibers with an SMA connector on one end and a 90-deg curve on the other.

The imaging cap determines the spatial location of the source and detector fibers and maintains the coupling between the fibers and the scalp [Fig. 4(c)]. To provide the general shape of the cap, the outermost layer of the cap is composed of rigid thermoplastic molded to an anatomically correct adult-sized head model. A layer of soft neoprene on the inside of the thermoplastic provides comfort and supports the fibers. The array of sources and detectors is placed in two interlaced rectangular arrays with first-through third-nearest neighbor separations of 1.3, 3.0, and 3.9 cm, respectively. Fibers are kept perpendicular and pressed inward toward the scalp by rigid spacers and rubber stripping on the outside of the cap. To improve coupling of the fibers with the scalp, the fiber tips extend $\sim 7\mathrm{mm}$ through the interior of the cap to facilitate combing of the fibers through the hair. Lastly, the cap is held in place against the head with hook-and-loop straps across the forehead of the subject.

2.3.

Superpixel DOT Acquisition and Algorithm

To successfully use the Andor sCMOS for HD-DOT neuroimaging, the system utilizes spatial and temporal encoding, noise reduction algorithms, and superpixel summing. Individual camera images are 2560 pixels wide by 175 pixels tall [Fig. 3(a)]. This asymmetric region of interest is used to permit faster frame rates. Sets of 25 images (24 source positions and one dark frame) are collected at 6 Hz resulting in a camera frame rate of 150 Hz. The image exposure length is, therefore, 6.7 ms and the laser exposure time with a 75% duty cycle is 5 ms. The light collected by the detector fiber is spread across the CMOS chip as a 31-pixel diameter circle [Fig. 3(b)]. The superpixel algorithm converts individual pixel counts to source–detector pair light levels.

2.4.

Subjects and Stimulus Protocol

The research was approved by the Human Research Protection Office of the Washington University School of Medicine, and informed consent was obtained from all participants before scanning. We recruited five healthy adult subjects (four females and one male). Subjects sat in an adjustable chair facing a 19-in. liquid crystal display at a 90-cm viewing distance.³ We positioned the SP-DOT imaging cap such that the optode array was centered circumferentially on the back of the head with the center of the bottom row of fibers on the inion.

All presented visual stimuli were radial, reversing, black-and-white grids (10-Hz reversal) on a 50% gray background. The paradigm consisted of an angularly swept radial grid wedge that rotated at $10\mathrm{deg}/\mathrm{s}$ to complete a sweep of the entire visual field every 36 s. The grid rotated clockwise three times per stimulus run. The stimulus began with the grid at the bottom of the screen, resulting in a left visual field stimulus at 10 s and a right visual field stimulus at 35 s accounting for the hemodynamic delay. Gray screens were presented for the 30 s before and 15 s after the complete sweep sequence to allow for the visual cortex to return to baseline. The total length of one run with three wedge rotations was, therefore, just under 3 min.

2.5.

Functional Data Analysis

Changes in absorption coefficient were calculated by normalizing each source–detector pair measurement by their mean and log transformation. The SNR in the heart rate frequency band was calculated as the average power in the pulse frequencies (0.5 to 1.5 Hz) normalized by the total power in the higher frequency noise band (2 to 3 Hz). The first-nearest neighbor (1-nn) pulse SNR per source or detector is calculated as the mean pulse SNR of all the 1-nn measurements of that particular source or detector.

Log-ratio measurement channels were bandpass filtered to 0.008 to 0.2 Hz. Superficial and systemic hemodynamic artifacts were removed by regressing the averaged 1-nn measurements from all other measurements. First, second, and third nearest-neighbor measurements that exhibited a standard deviation $<2.5\%$ of the mean signal were used for the image reconstruction. Volumetric reconstructions of absorption coefficients at 830 nm were obtained using a single sensitivity matrix based on the segmented head model of the MNI atlas.^12,13 Sensitivity matrices were inverted using a Tikhonov regularization constant of $\lambda =0.01$ and a spatially variant regularization parameter of $\beta =0.1$.²

3.1.

Superpixel DOT Algorithm and Specifications

The superpixel algorithm leverages the millions of pixels on a single CMOS chip to increase the DNR and decrease the effective detectivity. The full set of specifications for SP detection is shown in Table 1.

3.2.

Superpixel DOT Data Quality

The SP-DOT system and cap produce high-quality data, as measured by a light fall-off curve and high SNR calculated from pulse physiology. The light fall-off curve, which measures the source–detector measurement light levels as a function of their distance, demonstrates the desired log-linear fall off, characteristic of light propagation through biological tissue [Fig. 5(a)]. A cap that does not couple the fibers well to the head would have a shallow, flatter fall-off, and larger variations within similar source–detector separations.

Fig. 5

SP-DOT data quality. (a) Data from a healthy volunteer (red measurements) show the expected log-linear decay of light levels with source–detector distance. The spread in the nearest neighbor pairs is less than two orders of magnitude, indicating that there is good coupling between the scalp and all fibers. The first, second, and third nearest-neighbor pairs have $\mathrm{SNR}>20\mathrm{dB}$ based on the intensity values compared with the noise floor (blue measurements). (b) Qualitatively, individual detector channels have high enough SNR to show the pulse waveform in the time domain and the frequency domain at $\sim 1\mathrm{Hz}$. (c) A map of this pulse SNR for all source and detector fibers ensures that all fibers are well coupled to the head with $\mathrm{SNR}>10\mathrm{dB}$.

Each individual detector has sufficient SNR to measure the pulse waveform in both the temporal and frequency domains [Fig. 5(b)]. The SNR of this physiological signal for each detector should have an SNR greater than the experimentally derived threshold of 10 dB. We calculate the SNR by dividing the maximum power in the pulse frequencies by the average power in the higher frequency noise. These values, plotted for each source and detector for both first- and second-nearest neighbors, demonstrate a satisfactory cap-fit [Fig. 5(c)].

3.3.

Visual Activations

For the quadrant-arranged visual stimulations, the images acquired with the SP-DOT system demonstrate neural activations that are both temporally and spatially resolved. The left and right visual cortices are spatially resolved in all participants ($n=5$) by block averaging only three stimulus repetitions per subject [Fig. 6(a)]. Note that all activations are flipped up-down and left-right due to the anatomy of the visual system. When the stimulus is in the bottom right hand corner of the visual field, a positive change in the absorption coefficient is seen in the left visual cortex [Fig. 6(a), top row]. Similarly, when the stimulus is in the bottom left-hand corner of the visual field, a positive change is seen in the right visual cortex [Fig. 6(a), bottom row]. All activation volumes are normalized to their respective maximum values and thresholded to only show values $>40\%$ of the maximum value.

Fig. 6

In-vivo retinotopic mapping. Subjects viewed a rotating, wedge-shaped checkerboard to stimulate the visual cortex. (a) Left and right block-averaged ($n=3$ blocks) visual activations can be separately resolved in all five subjects and in the group average ($n=5$ subjects). (b) Binary mask of the voxels with a value above 40% of the maximum. The overlap map shows the number of subjects with an activated voxel for either the left or right activation, showing moderate overlap incidence between subjects.

Binary masks of the voxels with a change $>40\%$ show symmetrical activations from left and right stimuli [Fig. 6(b)]. Similar spatial results are seen in all five subjects with intersubject variability, as seen with previous HD-DOT systems.³ Although an incidence map of the activation overlap shows variability across individual subjects, the activations are generally colocalized [Fig. 6(b)]. Intersubject comparisons might be improved by ensuring consistent cap placement with better cap design and acquiring subject-specific coregistration of the data to anatomical images. The localization differences measured could also be due to variability in the functional architecture of each subject’s brain.^2,13,14