2.1.1.

MR coil design

As stated above, our goal was to design a probe in which the MR coil is designed to fit the fNIRS optodes. A number of criteria had to be taken into account in order to achieve high spatial and temporal resolution. Spatial resolution in fMRI is a function of many parameters, including field of view (FOV), number of voxels across the FOV, sampling bandwidth, number of phase encode lines, and the total amount of time allotted for recording a slice, all of which interact. Moreover, the slice acquisition time determines the number of slices which can be recorded within a given repetition time (thus setting the ultimate temporal resolution of the fMRI acquisition). Therefore, the smaller the number of slices acquired per repetition, the higher the possible temporal resolution of fMRI.

In practice, we can stack multiple acquisition blocks and reduce the FOV in the phase encode direction to decrease the slice acquisition time and, hence, increase the temporal resolution but also preserve spatial coverage.²⁸ As an alternative (but only when available in the facility), we can use the multiband sequence acquisition.^29,30 We limited the imaging volume to a $10\times 10{\mathrm{cm}}^2$ area with a phased array of three RF coils in order to enhance the signal-to-noise ratio (SNR) at high spatial resolution and capture only a focal area (here the prefrontal cortex).^31,32 The RF coils were fixed to a base [Fig. 2(b)] (printed on a 3-D printer with acrylonitrile butadiene styrene (ABS) material, uPrint SE 3D, Stratasys, Eden Prairie, Minnesota), which was shaped to the prefrontal area of the head. Because 3-D printing allows printing of arbitrary shapes, shaping the base of the probe to a high-resolution anatomical image of the specific participant’s head is straightforward; however, for the coil (first generation) used in this study, we physically measured the curvature of the subjects’ forehead and fitted an approximate geometrical form to it in Google SketchUp™. This was done to prove the concept and simplify the procedure. In the future, we will use a standard brain template [i.e., MNI152 (Montreal Neurological Institute)] and the subject’s anatomical scan to maximize the applicability of the probe. The 3-D printed base [Fig. 2(a)] fits to the shape of the scalp over the prefrontal cortex, which results in close and stable contact of the coil and the head. This is important in order to decrease motion artifacts and maximize received signal strength at the cortical surface.

The circular coils had a diameter of $\sim 5\mathrm{cm}$ resulting in a penetration depth of $\sim 2.5\mathrm{cm}$. Because fNIRS is unable to detect signal deeper than $\sim 2\mathrm{cm}$ beneath the skull, this radius was chosen to give the smallest coil diameter which would image the sensitive region of fNIRS at high-fMRI SNR. The receive-only MR coils were tuned to 123.25 MHz to match the center frequency of the Siemens 3T. We used geometrically orthogonal coils in the almost Olympic configuration^31–33 in order to minimize mutual inductance [Fig. 2(b)]. Additional decoupling was accomplished using preamp decoupling (a low-input impedance preamplifier is transformed to a high impedance at the coil terminals with a quarter wave cable to isolate the various elements of an array). The coil was interfaced to the scanner through a four-channel preamplifier interface (Stark Contrast, Erlangen, Germany).

In summary, we developed a phased array with three overlapping RF coils in an equilateral triangle shape, which were mounted on a base shaped to the desired area of the head, here the prefrontal cortex.

2.1.2.

Integration of functional near-infrared spectroscopy probes

The design of the fNIRS probe is of critical importance in the study in order to ensure good contact with the skin (high SNR), avoid motion artifacts, ensure subject comfort, and allow for the desired spatial coverage and depth sensitivity. To achieve these, different criteria (other than those for the MR coil) have been considered. We decided to incorporate at least six source–detector pairs (SD1 to SD6) with different source–detector separations (two detectors with three sources each at 2-, 3-, and 4-cm separation), to assess the depth sensitivity of fNIRS and cover most of the area of interest. In the example study described below, we used exactly six sources and one detector, resulting in six NIRS channels (each source–detector pair is defined as an NIRS channel). The probe is capable of accommodating more channels. The holders for the fNIRS detector and source fibers [Fig. 2(c)] are made with the 3-D printer. The printing material is ABSplus plastic, which is nonferromagnetic and does not cause susceptibility artifacts in the fMRI data. The material is also lightweight, which ensures the subject’s comfort. In addition, we can control the thickness and color of the material used in the probe, which can block or absorb the external light. In this study, we used white ABSplus plastic, and given the size and position of the region of interest (ROI), a quadrant of an oblate spheroid was chosen, with a thickness of 2.4 mm and the following internal dimensions: an equatorial radius of 10.2 cm and a semiminor radius of 7.62 cm.

The probes were directly integrated into the base onto which the coils were fixed to ensure stability and avoid motion artifacts [Fig. 2(c)]. This integration was achieved by drilling (or printing) holes into the base between the coils. These cylindrical hollow tubes (probe holders) allow optical fibers to be freely adjusted to ensure good contact with the skin. There are two reasons for making the fNIRS holders separate from the base. The first is to make adjustment of the hair below the sources and detector fibers easy. These holes were snug enough that we could place the cylindrical probe holders in them without additional fastening mechanism. Through the holes, hair could be shifted out of the probe area beforehand. The second reason was to improve the quality of the probe. 3-D printed materials have different strengths along different axes. By printing the base and the holders separately, we were able to maximize the strength of each component.

An additional problem in simultaneous fMRI and fNIRS studies is the precise registration of fNIRS optode locations with the MR data to accurately determine the sensitive region of the fNIRS measurements. Normally, localization measurements such as the 10-20 system or vitamin E MR markers are used, with clear limitations. The former is known to be a rough estimate, and the latter lacks accuracy due to the vitamin capsule size ($\sim 10\times 5\times 5\mathrm{mm}$) and the chemical shift image displacement of the lipid contents of the capsule.²⁶ We incorporated MR visible markers into the probe by filling tygon tubing with dilute solution of a gadolinium MR contrast agent (Prohance) in water. The tube’s diameter was $\sim 3\mathrm{mm}$, and it was wrapped around all NIRS optodes to make them visible in both the structural and functional MR images (not shown). We used the marker’s location on the higher structural scans to locate the optodes and registered them (using FSL toolbox img2imgcoord) to the functional scans.

2.2.1.

Magnetic resonance imaging signal-to-noise ratio

For the comparison between the SNR of the 32-channel coil and the multimodal probe, data were acquired from a biomedical informatics research network phantom³⁴ and from a participant lying quietly in the scanner. All MR data were acquired on a Siemens TIM Trio 3 Tesla whole-body MR scanner. Multiband echo planar imaging (EPI) data were obtained (University of Minnesota sequence cmrr_mbep2d_bold R011^29,30) with the following parameters: $\mathrm{TR}/\mathrm{TE}=525/30\mathrm{ms}$, flip angle 66 deg, matrix $96\times 96$ on a $172\times 172\mathrm{mm}$ FOV, multiband $\text{factor}=3$, $24\times 1.8\mathrm{mm}$ slices with no gap. Slices were prescribed tangent to the surface of the head at the location of the receiving fibers. We acquired 720 time points (378 s).

2.2.2.

Validation experiment

For the task-evoked paradigm experiment, functional data from the multimodal probe were acquired for three tasks: (1) a rapid semantic task, (2) a null or control task for the rapid semantic task, and (3) a subtraction task, which was repeated twice. The tasks were partly replicated from earlier studies.¹² The duration of the first, rapid semantic task block was 15 s, during which abstract or concrete words were presented for 1 s with 1 s interstimulus interval, resulting in eight stimuli in one block. The participant had to press the right button when an abstract word preceded a concrete word (e.g., for “Thought”+“Cup” or “Liberty”+“Pen”). Otherwise, the participant had to press the left button. The control task was identical in all regards, except that the right button was pressed only if a specifically determined word combination occurred (e.g., every time “Cup” preceded “Thought”). The semantic and control tasks were separated by a 30 s rest period. The subtraction task consisted of two blocks of 15 s separated by a 20 s rest period, where a pair of numbers was shown on the monitor—a large number followed by a smaller one, which had to be sequentially subtracted from the first number (e.g., 365–13). The participant was reminded each time to breath normally. The entire run of all tasks (30 s rest, 15 s semantic task, 30 s rest, 15 s null semantic task, 30 s rest, two blocks of 15 s subtraction separated by 20 s) lasted 192 s and was repeated five times.

2.3.1.

Magnetic resonance imaging signal-to-noise ratio

MRI data were processed in FSL FEAT [FMRIB Expert Analysis Tool, v5.0.7 (Ref. 35), Oxford University, United Kingdom).³⁶ For the comparison of the SNR between the coils, standard preprocessing steps were applied to the data: temporal high-pass filter (cut-off $\text{frequency}=0.01\mathrm{Hz}$ to remove very slow instrumental drifts) and motion correction (FSL FLIRT^37,38). In the beginning of the temporal time course, 20 volumes were discarded to allow for T1 relaxation. In order to calculate the SNR, an ROI was chosen, and the mean intensity of the signals divided by the mean standard deviation of the signals at each time point was calculated over a sphere of 7 mm ($\sim 59\mathrm{voxels}$). In addition, the mean intensity of the signals divided by the mean standard deviation over the entire time course over the same region was calculated to estimate temporal SNR. SNR and temporal SNR are separate measures of the coils’ spatial and temporal performance, which we can compare between coils. An additional preprocessing step for the temporal standard deviation was to smooth voxels with a 1.8 mm kernel. Because of the geometry of the multimodal probe, the SNR advantage relative to the 32-channel coil will be greatest at the scalp. For the human participant, the anatomical voxel used was visually selected in the center of the most sensitive area of the multimodal probe (centered $\sim 2\mathrm{cm}$ from the surface of the head), to make the best estimate of the effective SNR. This is close to the maximum depth of penetration for fNIRS, and as the B1 sensitivity of the multimodal coil is much less uniform than that of the 32-channel coil in the ROI used, this comparison will tend to favor the 32-channel coil. We were unable to make comparisons with only a subset of the 32-channel coils due to the algorithmically selected coils (including remote ones) with multiband.

2.3.2.

Validation experiment

For the fMRI data of the task-evoked study, each of the five fMRI slice stacks was preprocessed as described above. In addition, data were prewhitened and the five preprocessed fMRI slice stacks were concatenated in the slice direction (fslmerge) prior to a general linear model analysis. The stimulation protocol was convolved with the hemodynamic response function. The significance threshold at the voxel level was set to $z=2.3$ with a cluster threshold of $p=0.05$ (applied to the clusters remaining after the local voxel threshold). The fMRI data were also coregistered to the paricipant’s structural scan, where fNIRS markers are visible. Moreover, we calculated fNIRS sensitivity profiles below these source–detector pairs in order to analyze the spatial overlap between fMRI and fNIRS data. The photon probability densities were calculated using an updated version of a locally written program to calculate the Banana path using the models described in Ref. 39. The original version of this software was used in the analyses in Ref. 23. Briefly, the user specifies the coordinates on the neuroimaging informatics technology initiative anatomic image where the source and detector fibers are placed. The software finds the tangent plane to the head at each fiber location and constructs a plane which intersects the line between the two points and is perpendicular to each of the tangent planes. The line between the source and detector is the $x$ axis, and the line perpendicular to the $x$ axis lying in the constructed plane is the $z$ axis. The probability density is then calculated using Eq. (13) from Ref. 39.

Each pair of raw fNIRS time courses (690- and 830-nm data) was converted into two time courses representing oxy- ($\mathrm{\Delta }[\mathrm{HbO}]$) and deoxy-concentration change ($\mathrm{\Delta }[\mathrm{Hb}]$) using the modified Beer–Lambert law^40,41 in MATLAB^® (The Mathworks, Natick, Massachusetts). For the differential path length factors, we used published values of 6.51 at 690 nm and 5.86 at 830 nm.⁴² The low-frequency component (0.01 to 0.2 Hz) of $\mathrm{\Delta }[\mathrm{HbO}]$ and $\mathrm{\Delta }[\mathrm{Hb}]$ was obtained by using a zero-delay Fourier-domain band-pass filter in MATLAB^®. The low cutoff was chosen to adapt the frequency region in which the activation was expected. The fNIRS data corresponding to the five fMRI scans were extracted into separate time courses. The individual fNIRS time courses were then detrended, despiked, motion corrected,⁴³ and normalized (divided by their standard deviation). Folding averages (an average over the identical periods of a task paradigm) for all task periods were calculated by averaging over all five repetitions of the task. We compared the average time course to the task paradigm by convolving the task paradigm with the hemodynamic response function.

3.1.1.

Probe functional magnetic resonance imaging signal-to-noise ratio test—phantom

For the phantom data (Fig. 3, left panels), we found an SNR of $366.33\pm 2.73$ for the multimodal probe compared to an SNR of $259.43\pm 4.64$ for the 32-channel coil, an improvement of 41%. The temporal SNR was 248.75 for the multimodal probe and 116.68 for the 32-channel coil, a 113% improvement.

3.1.2.

Probe functional magnetic resonance imaging signal-to-noise ratio test—human data

When the probe was tested on a human subject (Fig. 3, right panel), we found an SNR of $61.74\pm 3.73$ for the multimodal probe and an SNR of $29.70\pm 1.87$ for the 32-channel coil in the same cortical region, a 108% improvement. The temporal SNR was 57.85 for the multimodal probe and 28.55 for the 32-channel coil, a 103% improvement.