2.1.

Module Design

An advanced 10-layer rigid-flexible PCB technology was adopted as the basis for the HD-DOT modules. Two patterns of module were implemented: dual-hexagon (dual-hex) and triple-hexagon (tri-hex) (Fig. 1). Each module consists of a flexible chain of either 2 or 3 regular hexagonal units. Each hexagonal section is based on the Gowerlabs LUMO module and is equipped with 3 dual-wavelength light-emitting diode (LED) sources and 4 silicon photodiode detectors. Each LED package provides two wavelengths of light at 735 and 850 nm, and each silicon photodiode has a photosensitive area of $2.77\times 2.77{\mathrm{mm}}^2$. Each LED is positioned at a corner of the hexagon to form an equilateral triangle, and three (of four) photodiodes are placed at the remaining three corners. The final photodiode is positioned at the center of the hexagonal unit. This layout generates source–detector separations (SDS) of 10 mm (9 channels) and 20 mm (3 channels) within each hexagonal unit. In each hexagonal unit, the LEDs are controlled by an LED driver and each photodiode is connected to 1 of 4 channels of an analog-to-digital converter (ADC) which records the data in real time. In this system, when illuminated, each LED provides an optical power of approximately 8 mW. A local micro-controller unit provides control for each module. It regulates LED source illumination, fetches optical measurements from the ADC, and writes data into a serial bus.

Fig. 1

The 10-layer rigid-flexible HD-DOT modules of two different designs: (a) dual-hex module and (b) tri-hex module. (a) A 3D diagram of a dual-hexagon module, each module contains four connector tabs, which include two female connectors and two male connectors (back side). (b) A photograph of the triple-hexagon module, demonstrating its flexibility. The triple-hexagon has the same design as the dual-hexagon module, except for the additional hexagonal board added in the center.

The dual-hex module yields 48 DOT channels per wavelength over an area of $\sim 9.4{\mathrm{cm}}^2$ and the tri-hex module can produce 108 DOT channels per wavelength over an area of $\sim 14.1{\mathrm{cm}}^2$. To avoid the need for separate cabling, board-to-board connectors are built into each DOT module and positioned on a dedicated connector tab. In each module, there are four connector tabs, with each housing one male or female connector. A flexible section is present between each connector tab and the adjacent rigid hexagonal unit to provide additional flexibility and articulation of the connector. The length of this flexible section is set as 3 mm, which is sufficient to permit an appropriate bend radius on the basis recommended by IPC-2223 flex-rigid design rules and the minimum radius of curvature expected in the target population (28 week gestational age preterm infant).⁵² The total length of female and male connector tabs are 10 and 8 mm, respectively, which results in a 12-mm cross-module distance when a male and female connectors of adjacent DOT modules are connected. The flexible section between each two hexagonal units in both double-/triple- hexagonal modules is correspondingly set at 12 mm so as to maintain the symmetry of the imaging array when combining multiple modules.

2.2.

Combining Multiple Modules

As the dual-hex and tri-hex modules can be connected together via the in-built board-to-board connectors, the system can be configured to provide a vast range of possible layouts. Figure 2 provides some examples of possible configurations. It is noteworthy that if a given connector tab is unused as part of a (permanent) array layout, those tabs can easily be removed. As a first demonstration of this system, we used 3 dual-hex and 2 tri-hex modules with a 2-3-2-3-2 layout configuration as shown in Fig. 3. This module configuration provides coverage of approximately $60{\mathrm{cm}}^2$, which is sufficient to ensure the sensorimotor cortex of the term-age neonate can be fully covered. This configuration provides 36 source and 48 detector locations and can generate 1728 logical DOT channels per wavelength.

Fig. 2

A conceptual diagram of arbitrary module combinations, such as (a) an imaging band, (b) a block, and (c) a patch. Combining different numbers of dual-hex and tri-hex modules, a vast range of possible layouts can be generated via in-built board-to-board connectors.

Fig. 3

The 12-hex DOT imaging array built using 3 dual-hex and 2 tri-hex modules, with encapsulation. This imaging array adopts a 2-3-2-3-2 layout that can provide appropriate coverage for the motor cortex of neonates. The system configuration of control unit (LUMO hub), PC, and cabling are also depicted (not to scale).

This system is controlled as per the commercial platform developed by Gowerlabs Ltd. The modules are connected to a base station (“hub”) for control and communication (via embedded serial bus) and to supply power using a single lightweight cable. The hub itself is connected to a laptop PC via USB (see Fig. 3). The system operates at a full frame rate of 10 Hz.

2.3.

Shielding and Encapsulation

For this neonate-specific HD-DOT system, it was critical to achieve an encapsulation that is lightweight, soft, and comfortable for the infant. It is also necessary that each external component is readily cleanable and/or disposable, to permit application in the clinical environment. Thus, a silicone rubber casing was designed to fit around each hexagonal unit in any orientation (Fig. 4). The black opaque silicone was created using a custom-designed injection moulding tool, with optically clear silicone windows subsequently cast in over the location of each optical component, to provide a permanent bond and a seamless contact surface. The finished individual casing can be stretched over each unit and removed/replaced for disinfection or disposal when necessary. Each casing also contains an external groove all the way around the outside edge to allow the module to be fixed into a stretchable head cap such as an EasyCap (Brain Products, Germany). Each casing is designed with a curved bottom surface so as to provide optimum coupling to the neonatal scalp.

Fig. 4

Encapsulation of the individual dual-hexagon and triple-hexagon modules. (a) A cross-sectional diagram of an encapsulated module, and (b) a photograph of a triple-hexagon module with complete encapsulation connected into and EasyCap.

After the encapsulation of the individual modules (either dual-hex or tri-hex), an array can be formed by connecting the modules together with the desired layout via the in-built board-to-board connectors. Plastic insulation was then applied over the connection joints to support further mechanical strength and robustness. After that, the DOT array was mounted into a fabric EasyCap [as per Fig. 4(b)] into which specific holes had been cut to accommodate the silicone encapsulation.

2.4.

Phantom Design

We began by creating the main body of the phantom. A homogeneous epoxy resin optical phantom was created based on the 39-week infant scalp surface from the 4D neonatal head model developed by Brigadoi et al.⁵³ The two-component epoxy resin (Crystal Clear 206, Smooth-On Inc.) was mixed with titanium dioxide particles (Superwhite, Tiranti Ltd., UK) and an NIR dye (Projet 900NP, ICI, UK) in order to obtain a reduced scattering coefficient ${{\mu }_{\mathrm{s}}}^{\prime }=1{\mathrm{mm}}^{-1}$ and an absorption coefficient of ${\mu }_{\mathrm{a}}=0.02{\mathrm{mm}}^{-1}$ at 800 nm.^54,55 After casting, a 3D-printed jig was used to allow precision drilling of two 8-mm-diameter holes into the phantom. The holes were positioned symmetrically about the mid-line of the phantom head and drilled to a depth of 73 mm into the bottom surface of the phantom, such that the bottom of the resulting holes coincided with the surfaces of the left and right pre-central gyri, as determined using the gray-matter (GM) surface mesh of the same 39-week infant head model.

The thermochromic dye used for the dynamic phantom was the “blue-orange” thermochromic powder dye (Changzhou Hali Chemical Co., Ltd., China), with an activation temperature of 43°C. To create solid targets that could be cast in any desired shape, the thermochromic powder was mixed with the epoxy resin (EpoxAcast™ 690), to a concentration of $25\mathrm{mg}{\mathrm{cm}}^{-3}$.

Prior to selecting the dye for construction of the phantom, its attenuating properties were examined. Semi-circular disk-shaped samples of 3-mm thickness were cast using a powder concentration of $25\mathrm{mg}{\mathrm{cm}}^{-3}$ as already stated. A spectrometer (FLAME, Ocean Optics Inc.) was used to perform transmission measurements through the disk at room temperature and at 50°C to evaluate the changes in absorption spectra before and after heating to above the activation temperature. Light from a tungsten halogen light source (HL-2000, Ocean Optics Inc.) was coupled to the surface of the disk using an optical fibre, with a second fibre returning collected light transmitted through the disk and returning it to the spectrometer. The disk was positioned vertically to allow on a hot plate, which was used to control the temperature of the sample. Another optical fibre carried the emerging light to the spectrometer. Figure 5 shows the change in absorbance (OD) as a function of wavelength for the thermochromic sample. Heating causes a decrease in absorbance that is most prominent at around 600 nm (i.e., in the orange/red region of the visible spectrum, as expected). However, a measurable change is still apparent in the NIR spectrum, and this dye was selected for construction of the dynamic phantom target.

Fig. 5

The change in absorbance as a function of wavelength between room temperature (22°C) and 50°C (above the activation temperature of 43°C) for a 3-mm-thick thermochromic sample, with a dye concentration of $25\mathrm{mg}{\mathrm{cm}}^{-3}$. (a) The spectral range 400 to 950 nm, and (b) 700 to 950 nm.

The targets for the final phantom were created by casting the dye and epoxy resin mixture in a cylindrical polyethylene tube. The samples were left to cure at room temperature for 24 h. After curing and extraction, each sample was machined and refined into an 8-mm-diameter cylinder with a height of 8 mm. A recess of 5-mm diameter was drilled into the end of each target cylinder in order to host a $20\text{-}\mathrm{\Omega }$ surface mount resistor and a $4.7\text{-}\mathrm{k\Omega }$ bead thermistor. The resistor dissipates power which yields a rise in temperature, monitored via the thermistor. After the electronic component insertion, the recess was carefully filled with the same dye-resin and left to cure. Two 8-mm-diameter rods of 77 mm in length were also created using a similar method as already mentioned, but using the resin mixture to match the optical properties of the main volume of the head phantom. A small hole was drilled through the center of each rod to act as a cable route. Two thermochromic targets were then glued to the top of each rod using a thin layer of epoxy. This resulted in two cylindrical rods, 8 mm in diameter and 85 mm in length, each with an 8-mm-long thermochromic tip. The phantom and target construction are shown in Fig. 6. An automated control circuit, based on a negative feedback loop that maintains the target temperature just above the transition temperature, was used to control the current through the target circuit, as described previously by Hebden et al.⁴⁸

Fig. 6

The thermochromic rod at (a) room temperature and (b) above the activation temperature. (c) A diagram of the internal design of the thermochromic rod and (d) an illustration of the phantom and rod.

2.5.

Phantom Imaging Experiment

This flexible HD-DOT system was used to yield 3D images of the infant dynamic phantom. The array shown in Fig. 3 was placed on the head surface above the two targets, centered over the apex and covering symmetrically the left and right sensorimotor cortices. Cranial landmarks, which were built into the phantom mould, were used as reference points to align the array.

A heating power of 1.4 W was delivered to the rods leading to a change in the optical properties of the target that became evident after approximately 60 s. The target activation was controlled using a USB relay, itself controlled via serial port using Matlab (The MathWorks, Inc.). After a 20-s rest period at the beginning of the recording, the target control circuit was switched on for a period of 180 s, which was followed by 300 s of cooling during which the targets were allowed to return to room temperature. This cycle was repeated 10 times for each target. The experimental setup is shown in Fig. 7.

Fig. 7

The experimental setup used to control phantom activation experiment. The control box is adopted from Ref. 48 to control the temperature of the thermochromic rod. The power source of the control box is regulated by predefined Matlab script via a USB COM port and the USB relay.

2.6.

Photogrammetric Registration

A photogrammetric method was used to obtain a three-dimensional model of the phantom head surface with the HD-DOT cap in place. A video was recorded while panning around the phantom at three different heights using the camera of an X-series iPhone (Apple, Inc., California). The movie file was imported directly into Metashape⁵⁶ to create a 3D model, which was exported in the .PLY format. To identify the exact location of sources and detectors, each module was tagged with markers at the points directly above the three sources. Images of the resulting mesh model of the array and phantom are shown in Fig. 8. The cranial landmarks and the source marker locations were determined from the point cloud using the software CloudCompare (Open Source Project⁵⁷). As the dimensions of the module encapsulations are known precisely, the locations of every optode could be determined without further approximation. The resulting 3D array positions are also shown in Fig. 8.

Fig. 8

(a) Frontal, lateral, and superior views of the imaging array in place on the dynamic phantom. (b) The corresponding optode positions registered to the anatomical atlas model. The black circles on subpanel (b-3) indicate the rod positions. (c) The 3D channel layout as viewed from above.

2.7.

Data Pre-Processing

The data processing consisted of the following steps. First, channels were rejected if their coefficient of variation exceeded 8%.⁵⁸ The data were then converted to absorbance and low-pass filtered. Given the phantom activation cycle lasted 500 s, the low-pass filter was set to 0.01 Hz and implemented using a finite impulse response digital filter.

A detrended block-average was performed to obtain relative changes in absorbance for 500-s-long epochs starting from 20 s before the beginning of each heating cycle. The average of all the 20-s periods of data centered on the time at which the maximum change was observed (200-s post-onset) was used to reconstruct DOT images relative to the baseline period. Three-dimensional images showing the changes in the optical absorption coefficient due to the activation of the targets were thus obtained for both wavelengths.

Channel-wise changes in “hemoglobin concentration” were also obtained by treating the dynamic changes recovered from the phantom as if they were human physiology by applying the modified Beer–Lambert law,⁵⁹ with a $DPF=6$ for both wavelengths⁶⁰ to the block-averaged absorbance data. This was performed to examine how closely the phantom mimicked the types of signal morphology typically observed in a functional neuroimaging experiment.

2.8.

Image Reconstruction

Image reconstruction was undertaken in a two-layer mesh model of the infant head. This model included a volume mesh with extracerebral and brain tissue layers, and separate scalp and GM surface meshes, which were derived from the Brigadoi atlas⁵³ for the 39-week post-menstral age point. With the exception of the target, the phantom optical properties are homogenous; a two-layer model was used simply to enable the extraction of images at the level of the GM. Both layers of mesh model were modeled with identical optical properties. The mould in which the phantom head was cast was based on the same dataset used to create our meshes, so they are expected to provide a near-perfect model. Once the mesh was registered to the phantom’s cranial landmarks and optode positions derived from the photogrammetry model, TOAST++⁶¹ was used to compute a forward model for each wavelength and a zeroth-order Tikhonov regularized reconstruction was performed with a regularization hyperparameter of $1\times {10}^{-3}$. Images were reconstructed in the full mesh volume and for visualization purposes, image values at the level of the GM surface were also extracted.

3.1.

System Performance and Data Quality

The use of 10-layer rigid-flexible PCB technology allows this HD-DOT system to exhibit an ultra-compact circuit layout and a miniaturized profile, and consequently a very lightweight construction. The side of each rigid hexagonal unit is 10-mm long, and the area of modules are $9.4{\mathrm{cm}}^2$ for the dual-hex module and $14.1{\mathrm{cm}}^2$ for the tri-hex module. The weight of the dual-hex and tri-hex modules are 8 and 12 g respectively, and the total weight of the imaging array (combing with 3 dual-hex and 2 tri-hex modules and encapsulations) is 70 g. The array configuration shown in Fig. 3 provides coverage of approximately $60{\mathrm{cm}}^2$, which is sufficient to cover the entire sensorimotor cortex of the term-age neonate, as shown in Fig. 8.

Figure 9(a) shows the optical intensity measurements as a function of SDS. The noise floor is measured to be $1.16\times {10}^{-5}$ (arbitrary intensity units), which given the upper measurement limit is 2.5 (arbitrary intensity units) implies a measured dynamic range of 106.6 dB. Figure 9(b) shows the histogram of all possible channels and the accepted channels (good channels) generated on the dynamic phantom as a function of SDS. There are 1728 DOT channels in total per wavelength with 717 good channels per wavelength.

Fig. 9

(a) The channel optical intensity measurements as a function of their SDS. Channels that passed out pre-processing thresholds are depicted in red. (b) Histograms of all possible channels (blue) and accepted channels (red) generated on the dynamic phantom as a function of the SDS.

3.2.

Channel-wise Dynamic Responses

The response to the phantom target activation in the case of the left and right activations is shown in Fig. 10 for the two wavelengths used by the system. The channels that exhibited a response show clear changes in optical density with the latency to peak of approximately 160 to 180 s from the beginning of the activation (heating phase). The changes in absorbance are clear in several channels at 735 nm. Consistent changes are not apparent at 850 nm in this channel-wise view.

Fig. 10

The change in absorbance (optical density) as a result of the activation of the left hemisphere target for the (a) 735-nm and (b) 850-nm wavelengths, and the right hemisphere target for the (c) 735-nm and (d) 850-nm wavelengths, respectively. The heating period is indicated by the gray-shaded area.

Changes in hemoglobin concentration were considered to examine how well the thermochromic targets can mimic functional hemodynamic responses. Figure 11 shows the simulated hemodynamic response functions (HRFs) displayed on a representative 2D channel layout. The responses exhibit a similar morphology to real HRFs, with the simulated HbO increasing and HbR decreasing. Unlike real HRFs, the simulated HbR response shows a larger magnitude change than HbO. Figure 11 also demonstrates clear, localized changes were obtained for the right-hemisphere activations over the expected locations. (Note: to provide a better fit to the paper format and a more clear display, we decided to cut out the left-hemisphere activations from Fig. 11 and show only one hemisphere.)

Fig. 11

Right hemisphere activation responses in a representative 2D array. It is noteworthy that a high number of channels which exhibit clear responses in the location correspondent to the right rod position show increase in HbO, decrease in HbR as per functional HRF in human brain.

3.3.

Volume and Gray Matter Image Results

Three-dimensional images of the change in absorption coefficient at the two wavelengths used by the system were reconstructed in the two-layer neonatal head mesh. Figure 12 shows a sagittal, axial, and coronal view, and the GM surface extraction of the phantom targets. The slices selected for the orthogonal views correspond to the coordinates of the node exhibiting the largest decrease in absorption coefficient at either wavelength in each activation. As is evident in Fig. 12, the peak changes occur with good spatial correspondence to the expected location of the thermochromic targets, over the pre-central gyrus of each hemisphere. The changes obtained at 735 nm are noticeably larger than those at 850 nm, which is consistent with the absorbance changes shown in Fig. 10. To quantify the localization error associated with these images, we computed the offset of the true center of the absorption change in the rod from the center of mass of the spatial distribution of absorption change. The center of mass was computed for each image of the change in absorption coefficient at each wavelength in each hemisphere. First, the image values where thresholded at 50% of the maximum absolute value found across the whole volume. The remaining image values at each node were then multiplied by the Voronoi volume associated with that node and a weighted sum, using the Voronoi volume-corrected absorption change values as a weight, was computed to find the coordinate of the center of mass. The Euclidean separation between the center of mass and the center of the superior face of the rod was computed, as was the Euclidean separation between the center of mass and the center of the rod target volume. The results of this analysis are shown in Table 1. Using the images shown in Fig. 12, the FWHM of the perturbation in $x$ (left-right), $y$ (anterior-posterior), and $z$ (inferior-superior) were also determined at 735 nm. The resulting values were 12.59, 12.04, and 10.52 mm, respectively, for the left hemisphere target, and 12.14, 16.38, and 17.68 mm for the right hemisphere target. The true dimensions of the thermochromic target in these dimensions were 8, 8, and 8 mm, respectively. Because of the multiple distinct regions of change in absorption coefficient in the 850-nm images, we did not attempt to determine FWHM values. In addition, given the relatively low signal-to noise ratio (SNR), the center of mass values at 850 nm are likely unreliable; as such, we consider the 3D Euclidean localization error to range from 3 to 6 mm based on the values obtained from the wavelength at 735 nm.

Fig. 12

Volumetric and GM -surface images of the change in absorption coefficient at 735 nm (${\lambda }_1$) and 850 nm (${\lambda }_2$) displayed for (a), (b) the left and (c), (d) the right hemisphere activations. Images are an average of a time window of between 190- and 210-s post-activation onset, and reconstructed relative to the mean of the 20-s pre-activation. Each image is thresholded at 50% of the maximum absolute signal deviation found in the volume. The sub-panels in each row show the volume images in (i) sagittal section, (ii) transverse section, and (iii) coronal section, and (iv) an extraction of the GM surface.

Table 1

Euclidean offset distances between the upper surface of the rod and the center-of-mass of the volumetric image and the middle of the rod and the center-of-mass of the volumetric images for the left and right hemisphere targets at 735 nm, which showed the best contrast.