Cerebral oximetry performance testing with a 3D-printed vascular array phantom

: Cerebral oximetry based on near-infrared spectroscopy represents a unique noninvasive tool for real-time surgical monitoring, yet studies have shown a significant discrepancy in accuracy among commercial systems. Towards the establishment of a standardized method for performance testing, we have studied a solid phantom approach – based on a 3D-printed cerebrovascular module (CVM) incorporating an array of 148 cylindrical channels – that has several advantages over liquid phantoms. Development and characterization of a CVM prototype are described, including high-resolution imaging and spectrophotometry measurements. The CVM was filled with whole bovine blood tuned over an oxygen saturation range of 30-90% and molded-silicone layers simulating extracerebral tissues were used to evaluate penetration depth. Saturation measurement accuracy was assessed in two commercially-available clinical cerebral oximeters. For one oximeter, both neonatal and pediatric sensors showed a high degree of precision, whereas accuracy was strongly dependent on saturation level and extracerebral geometry. The second oximeter showed worse precision, yet greater robustness to variations in extracerebral layers. These results indicate that 3D-printed channel array phantoms represent a promising new approach for standardized testing of clinical oximeters. performance was seen for one of the devices, whereas the other was more robust. Based on comparisons with prior phantom and clinical results, our testing approach provided both absolute quantities and trends that were reasonable. Although improvements in optical property tunability of 3D-printing materials and microchannel print quality will enhance the clinical realism of our approach, this method has the potential to become a powerful tool for elucidating biological factors that impact device performance. Our future work will further validate this approach by testing a greater range of NIRS oximeters and sensors and addressing other confounding biological factors.


Introduction
Tissue oximeters based on near-infrared spectroscopy (NIRS) have become increasingly common for cerebral oxygenation monitoring of premature infants in intensive care units [1,2], as well as during neonatal and pediatric surgeries, particularly cardiac procedures [3,4]. The ability of NIRS-based devices to assess cerebral oxygen saturation enables real-time intervention to mitigate ischemia-reperfusion events or defects in inherited metabolic pathways [5]. In adults, active cerebral monitoring with oximeters can have an important impact during and after cardiovascular surgery and neurosurgery [6]. Effective devices can reduce mortality, transfusion rates and resource consumption while mitigating the risk of side effects that can adversely impact neurodevelopment or cardiac function [7,8].
Despite broad clinical implementation of cerebral oximeters, research has raised several questions regarding the performance of these devices. Inconsistency among commercial systems is a primary issue. For instance, Dix et al. [9] compared measurements from three commercial oximeters in 55 infants and found differences of 10 to 15% in tissue oxygenation (StO 2 ) values. In a study of five commercial oximeters in adults, Bickler et al. [10] found mean bias -defined as a difference between cerebral oximeter readings and the manufacturerspecified weighted arterial (SaO 2 ) and jugular bulb (SvO 2 ) saturation -varied from 3 to 13%. A number of other clinical studies have found significant differences in performance amongst cerebral oximeters [11][12][13]. The lack of standardization in clinical outputs have led some to conclude that cerebral oximeters are ineffective for estimating absolute StO 2 and should be limited to trend monitoring [10].
Differences in performance likely stem from variations in design features such as illumination wavelengths, illumination-collection geometry, component specifications and data processing algorithms. Because these variables impact light-tissue interactions, the quality of detected signals, and how StO 2 is calculated from these signals, they may determine robustness to inter-and intra-patient variations in anatomy (e.g., thickness of skull and subarachnoid space) and physiology (e.g., hematocrit, tissue turbidity). While manufacturers incorporate approaches such as multiple source-detector separation distances and specialized algorithms to minimize interference from extracerebral tissues, differing implementations may lead to diverse outcomes [14,15]. Identifying an effective paradigm for benchtop evaluation that addresses biological factors known to impact performance would represent a major step toward satisfying the need for standardization that many researchers have advocated [13,16].
In vivo testing of cerebral oximeters in healthy human subjects is currently the dominant method for establishing performance; however, this approach has several significant limitations [17]: (a) a lack of a single "gold standard" reference value for StO 2 , as NIRS devices sample arteries, veins and capillaries with a variety of saturation levels; (b) a lack of clear insight into whether devices are sensing the brain or more superficial tissues; (c) a reliance on the assumption that StO 2 measurements in the superficial frontal cortex are equivalent to a weighted mean of arterial and venous saturation levels [17] -which may not be accurate in a wide variety of subjects under a range of physiological conditions [18]; and (d) the fact that ethical considerations preclude healthy subject testing at low saturation or blood concentration levels, or in children and infants. Thus, other approaches may be appropriate for augmenting or replacing volunteer subject clinical trials for validation; such as clinical convenience sample studies [19] and benchtop test methods.
Performance testing based on tissue-simulating phantoms provides an alternate approach for generating evidence of device effectiveness, in a manner analogous to standardized image quality test methods for established medical imaging modalities like ultrasound [20]. Phantom-based performance tests for NIRS oximeters can provide a rigorous, reproducible and convenient approach to evaluating performance using well-characterized materials and gold standard referencing of oxygenation levels. Such methods enable well-controlled studies of a wide range of variables such as patient-to-patient variations in tissue morphology (e.g., skull thickness) and physiology (e.g., hematocrit, optical properties) or device design variables (e.g., illumination wavelengths, illumination-collection geometry). Such an approach has the potential to quantitatively elucidate differences in device performance characteristics and robustness to confounding biological factors [21]. In addition to device validation, phantom-based test methods can facilitate early device development, quality control for manufacturing and constancy testing in clinical settings.
Liquid phantoms based on commercial fat emulsions such as Intralipid are often used in NIRS performance testing, as they can be mixed with blood to achieve tissue-relevant optical properties [22,23]. Blood oxygen saturation has been modified to achieve a range of target levels by introduction of oxygen and nitrogen gas [24], or by adding reducing agents such as yeast [18] or sodium dithionite [25]. This type of phantom has been implemented to evaluate performance of cerebral oximeters. Earlier versions of Intralipid-based phantoms for tissue oximetry involved direct immersion of sensors in the liquid medium [26,27]. More recently, investigators have used Intralipid-blood solutions along with solid silicone layers to represent extra-cerebral [18] and superficial skin/fat tissues [28]. While liquid phantoms have provided useful insights into oximeter performance, they also have a number of inherent disadvantages. First, blood/hemoglobin-Intralipid phantoms have been shown to exhibit signal drift over a period of hours due to changes in lipids that affect scattering [29,30]. Second, referencing to StO 2 can pose challenges, as a gold standard approach like CO-oximetry may not be effective for dilute, turbid samples and diffuse-reflectance-based methods (e.g., frequency domain NIRS) may not be sufficiently accurate and/or can be expensive. Third, the development of a specialized system for controlling and monitoring a turbid, blood-based liquid phantomespecially when a stable, non-equilibrium saturation level is desired -can be burdensome due to the specialized equipment, labor and knowledge required.
Solid polymer phantoms containing fluid channels filled with blood or hemoglobin solutions may represent an effective alternative for testing oximeters. Over 20 years ago, Kurth et al. [24] introduced a solid brain phantom for NIRS constructed from polyester resin mixed with TiO 2 and incorporating 500 blood-filled channels with a 0.5 mm inner diameter. While we are not aware of any subsequent studies implementing this type of high-density channel-array approach for NIRS, it may now be possible to generate similar phantoms with less effort and greater flexibility using 3D printing technology. Our lab has pioneered the fabrication of 3D-printed biophotonic phantoms [31], developing models with both idealized and biomimetic vascular geometries for evaluation of hyperspectral and fluorescence imaging systems [31][32][33]. Other groups have also shown success in developing 3D-printed phantoms for different biophotonic applications [34,35]. Channel array phantoms not only enable the use of CO-oximetry for referencing of blood samples but provide a matrix material that is stable during measurements and relatively easy to fabricate, handle, and measure. Furthermore, this approach enables rapid re-design and dissemination of new phantom geometries.
The purpose of this study was to evaluate the potential for improving phantom-based testing of NIRS oximeters using emerging 3D printing technology. Specifically, our goals were: (1) characterize optical properties and channel morphology for a promising printer/material combination; (2) generate a prototype cerebrovascular module (CVM) incorporating an array of blood-filled channels as part of a phantom with removable layers and demonstrate that it produces reasonable values when measured with commercial cerebral oximeters; and (3) perform a parametric evaluation of the impact of extracerebral layers on clinical cerebral oximeter performance. By establishing a proof-of-principle for 3D-printed oximetry phantoms, this work has the potential to provide a foundation for other advances in tissue phantoms for NIRS technology and lead to new standardized test methods.

CVM des
To fabricate t Ltd., Eden Pra following nom and Z-axis -6 and reduced displayed in F optically simi printed mater sphere spectr USA) and the selected show for which the µM, hematoc VeroWhite re We used a assess phanto , USA) to ngle-layer multi-channel array sample was printed and evaluated. Imaging results are presented in Fig. 3 along with a photograph of the sample with channels filled with India ink (Super black, Speedball Art product, Statesville, NC, USA). Imaging with microCT provides visualization of phantom nonuniformities or defects as well as quantification of printed feature sizes for comparison against designed geometry. Samples showed a homogeneous build, as opposed to the internal lattice structure we have seen with a fused deposition modeling printer [31]. Additionally, we evaluated the viability of channels over a range of diameters and identify the smallest useable size. As seen in Fig. 3, channels with nominal diameters greater than 0.6 mm are patent. Cross-sectional images of each channel were evaluated in four different locations along their length. The measured diameter was consistently smaller than the nominal value by approximately 0.1 mm, possibly due to flow of the liquid photopolymer during the solidification process. The CVM was constructed with overall dimensions of 130 x 120 x 50 mm, with the perfused channel-array region limited to a volume of 120 x 80 x 32.5 mm. Channel density was based in part on the total hemoglobin concentration (ctHb) of the neonatal brain, which the literature indicates is between 25 and 115 µM [18,37,47]. The CVM incorporated a total of 148 cylindrical linear channels spaced 4 mm apart in the vertical and horizontal directions, with the top row of channels (top edge) 0.8 mm below the phantom surface. Each channel had a nominal diameter of 0.90 mm, yet the mean measured channel diameter, based on µCT images of the thin multi-channel sample was 0.78 ± 0.12 mm (mean ± standard deviation). We also estimated the mean volume of the phantom channels by injecting water into the CVM and subtracting the volume of the side window regions, as well as by injecting known volumes of water directly into randomly selected individual channels (n = 8) until they were filled. The measured ctHb and channel diameter were 54 ± 0.08 µM and 0.75 ± 0.01 mm for the bulk water measurements; and 53 ± 0.06 µM and 0.74 ± 0.05 mm for the individual channel method. Two 3D-printed lids made of clear resin (120 x 55 x 60 mm) were also attached at the sides of the CVM by screws to minimize air exposure and subsequent reoxygenation of blood. Each lid contained a port (Fig. 2), which was used to attach a bloodfilled syringe to one side and release air from the other. The impact of any boundary effects was investigated by comparing results obtained when the bottom and sides were covered either by aluminum foil or a black cloth. This was performed for two StO 2 levels with two different superficial layer cases (T L1 = 0, 7.5 mm). No significant difference in StO 2 was observed between these two cases, indicating that the phantom was sufficiently large as to produce negligible boundary effects.

Fabrication of Superficial Phantom Layers
To evaluate the impact of extra-cerebral layers on cerebral oximeter accuracy, we fabricated phantom layers to mimic scalp/skull and cerebrospinal fluid (CSF) regions. Figure 4 provides data on optical properties reported in the literature [18,45,48] used as approximate target values for these phantom layers. CSF in the subarachnoid space was modeled as a lowturbidity medium due to the influence of the arachnoid trabeculae [40]. Optical properties for this layer were previously estimated to be µ a = 0.004 mm −1 and µ' s = 0.25 mm −1 at λ = 800 nm (39). Scalp/skull layers were molded from 30 g PDMS (Sylgard 184 Silicone Elastomer, Dow Corning Corp., Washington, DC, USA) mixed with 0.28 g titanium dioxide (TiO 2 , anatase titanium (IV) oxide, Sigma-Aldrich, St. Louis, MO, USA) and 0.01 g India ink. CSF layers were comprised of 30 g PDMS and 0.19 g TiO 2 . PDMS solutions were mixed using a homogenizer (Tissue-Tearor, Model 985370, BioSpec Products Inc., Bartlesville, OK, USA) and the curing agent (Sylgard 184, Dow Corning) was added at a 10:1 ratio of silicone to curing agent, stirred, and then degassed using a vacuum chamber for half an hour (Labfirst Scientific and Industrial Equipment LLC, Cheyenne, WY, USA). After air bubbles were removed, the uncured materials were poured into Teflon molds to produce layers of 1 mm and 2.5 mm thickness. Samples were then cured in an oven for 3-4 hours at 75 °C.
An electronic height gauge device was used to validate PDMS layer thickness (Mitutoyo America Corp., series 570, Aurora, IL, USA), and all samples showed deviation less than 20% from the targeted thickness. PDMS layer optical properties were measured using an integrating sphere spectrophotometer, as described above. Resultant absorption and reduced scattering data for the scalp/skull and CSF layers were in good agreement with the literature [18,40,45,48].    [18], from the slope of the curve). A rms results increased monotonically from 5% to 21% as Layer 1 was increased from 0 to 7.5 mm. The T L2 = 2.5 mm case indicates a nearly complete loss of sensitivity to changes in StO 2 .
Results for the Oximeter A -ped sensor (Fig. 9) were very similar to those for the neonatal sensor, except that for the T L1 = 1 mm series, three low readings at high saturation levels resulted in greater A rms for this case than expected. It is also worth noting that results for the T L1 = 2.5 mm case showed slightly greater sensitivity than the Oximeter A -neo sensor did, which may indicate that the medium sensor was designed to address larger superficial layers.   Results for the Oximeter B neonatal sensor (Fig. 10) were markedly different than those for Oximeter A with either small or medium sensors. Oximeter B showed worse precision than Oximeter A (about 2-3 times larger), which may have obscured trends due to layer thickness. Yet increasing T L1 or T L2 had comparatively minor impact on slope/sensitivity and bias. The T L2 = 2.5 mm case, which produced very poor sensitivity for Oximeter A, produced somewhat increased bias for Oximeter B, but there was minimal impact on slope and any trends due to T L1 were not clear.
To better understand the results provided for larger superficial layer thicknesses, we also fabricated a 4-cm-thick slab of scalp/skull-mimicking material -representing the limiting case -and measured it with each sensor and device. Oximeter A -neo and -ped sensors provided a mean StO 2 of 82.9 ± 1.4% and 74.9 ± 1.0%, respectively, whereas Oximeter B with a neonatal sensor provided a mean StO 2 of 99.9 ± 0.4%. It is also worth noting that a prior study [28] measured a superficial-tissue-mimicking slab with slightly different properties (μ a = 0.059 cm −1 , μ s ' = 5 cm −1 at 692 nm) using the adult sensor of Oximeter B, and found a value of 78%; using the same sensor and our phantom slab, we measured an StO 2 of 73%.

Discussion
The modular, 3D-printed phantom approach developed and tested in this study has the potential to facilitate evaluation and inter-comparison of NIRS-based cerebral oximeters, including patient-specific factors that impact performance. In this section, we address three main issues: (1) evaluation of the validity of our phantom-based test method through comparisons with the literature, (2) insights provided by our results into the confounding effect of extra-cerebral tissue layers and (3) considerations of 3D-printed phantom quality.

Comparisons with liquid phantom studies
One approach to evaluating the suitability of our test method involved comparisons to prior studies implementing a more well-established liquid phantom [18,28,29]. In two of these articles, measurements were performed with a 2.5 mm-thick turbid silicone layer representing extra-cerebral tissues above mixtures of erythrocytes and Intralipid. Comparisons of test results are provided for both oximeters in Fig. 11. While these prior studies have not provided measured μ a spectra for liquid phantoms, based on their composition, we estimate that their μ a values are in the lower end of the range presented in the literature for cerebral tissue, whereas our CVM prototype is closer to the higher end of this range. We compared results for the Oximeter A -neo sensor as tested with our solid phantom and T L1 = 2.5 mm, T L2 = 0 mm, to prior results from a liquid phantom with ctHb = 47.5 and 75 μM a similar superficial layer [29]. These two data sets have a variety of features in common, including large positive bias at low-to-moderate StO 2 levels and a small amount of negative bias at high StO 2 levels. The liquid phantom measurements showed a linear regression slope of 0.49, whereas for our solid phantom a value of 0.63 was found. Although no A rms data was provided in the prior study, based on published graphical data we approximate a value of 16.2% (interpolated for 54 μM). Our solid phantom measurements produced A rms values of 7.5% and 8.9% for Oximeter Aneo and ped sensors, respectively. Overall, liquid and solid phantoms provided similar trends, with a minor to moderate level of discrepancy in quantitative values.
Measurements using Oximeter B and our solid phantom with T L1 = 2.5 mm, T L2 = 0 mm were compared with liquid phantom data for ctHb values of 45 and 70 μM Fig. 10(b) (19). While liquid phantom results were qualitatively similar to our Oximeter A data -with a large positive bias at low-to-moderate StO 2 levels and a small degree of negative bias at high StO 2 levels -our solid phantom results for Oximeter B showed less bias at high and low StO 2 values, but greater variability (worse precision). The sensitivity of liquid phantom measurements showed linear regression slopes of 0.41 (for ctHb = 45 μM) and 0.54 (for ctHb = 70 μM) (19), whereas our results indicated a slope closer to unity (0.6). Although no A rms data was provided for liquid phantoms, our estimate based on published graphical data is 15.5%. Our solid phantom measurements indicated a greater level of accuracy, with an A rms of 9%. The source of this discrepancy between liquid and solid phantom results for Oximeter B remains unclear. While the source of the variability in Oximeter B data is also not known, our findings are consistent with a prior study which found same-site reproducibility for this device to be 4.6% vs. 2.9% for Oximeter A [52]. Fig. 11. Comparison of our solid phantom measurements with selected data from a prior liquid phantom study (19,28). ctHb in our channel array phantom was 54 µM and in the cited research were (a) 47 Several clinical studies have demonstrated a lack of agreement between commercial cerebral oximeters however, little data was found in the literature on accuracy in neonates and children for the two devices studied. Schneider et al. [53] performed a clinical study of neonates which compared the two devices in the current study. Results showed relatively little difference, with StO 2 of 73.3%/74.2% (left/right hemisphere) for Oximeter A vs. 76.6%/77.8% for Oximeter B. These results agree with our findings that at high StO 2 values, relatively little difference in accuracy was seen between Oximeter A and B. Additionally, a clinical study of neonates by Hessel et al. [54] indicated that the Oximeter A neonate sensor produced a positive bias relative to a neonatal INVOS sensor at low saturations and negative bias at high saturations. Given that a neonatal INVOS sensor was shown by Kleiser et al. [29] to exhibit a consistent positive bias of 10-20%, but a nearly optimal sensitivity (slope close to 1.0), the results from Hessel et al. [54] are in qualitative agreement with our findings. It is also worth noting that in results with our phantom and the aforementioned liquid phantom [29], local bias tended to be greatest at StO 2 values of 40% or less. This would seem to indicate that the poorest performance occurs in an StO 2 range most relevant to unhealthy patients. Such an outcome may be attributable to the fact that safety concerns limit the range of StO 2 values over which clinical testing of cerebral oximeters can be performed in healthy human subjects [10].

Impact of extra-cerebral layers
Numerous prior studies have concluded that cerebral oximeter readings are influenced by extra-cerebral layers [14,55,56]. Our phantom-based test results provide insights into the quantitative impact of superficial layers on device performance. As layers simulating the scalp and skull and CSF-filled subarachnoid space were increased in thickness, there was a major increase in positive bias for Oximeter A neonatal sensor, with A rms increasing from 5% for no superficial layers to 21% for T L1 = 7.5 mm, T L2 = 1 mm. When a CSF-simulating layer was present, the turbid layers showed much less sensitivity to changes in S R O 2 . Oximeter B shows less variation to changes in layer thickness. While the change in sensitivity/slope with thickness is minimal for Oximeter B, minor degradation in accuracy and precision are apparent.
Some of this discrepancy in sensitivity to superficial layers between devices may be attributable to differences in sensor reflectivity. The tissue-contacting surface of Oximeter B appears black; however, the Oximeter A probes are light in color. Spectrophotometry indicated that the diffuse reflectance of both Oximeter A probes was approximately 58% at 800 nm, whereas the Oximeter B probe showed a value of about 5%; all three probes showed minimal variation across the 700-900 nm range. Highly absorbing sensor surfaces should minimize the re-entry of reflected photons into the medium but have little effect on deeper propagating photons. More reflective sensors would enable exiting photons to re-enter the tissue and continue propagating in superficial regions -possibly via repeated reflections off the sensor surface -thus increasing detection of shallow-penetrating light. As a result, Oximeter A was likely more sensitive to the properties of superficial layers.
It is also noteworthy that for Oximeter A neonatal and pediatric sensors, the StO 2 value measured in the bulk non-perfused phantom slab was similar to the levels approached as the superficial layer was increased in thickness. This relationship may provide insight into the device's penetration depth and the degree to which displayed values may in many cases represent some weighted version of StO 2 in superficial and deeper layers. The StO 2 values displayed by both oximeters during slab measurements also raise questions about the utility of these devices in low perfusion conditions. The differences in these StO 2 values may be attributable to the use of different algorithms for each device/sensor combination.
These findings align well with prior studies that have addressed the issue of oximeter penetration depth. While we are not aware of any prior phantom study which addressed the impact of superficial layers on cerebral oximetry, Nasseri et al. [28] studied how adipose tissue thickness changes estimates of muscle StO 2 using adult probes and a liquid phantom with superficial layers. This study showed an increase in positive bias at lower StO 2 values and greater negative bias at very high StO 2 with superficial layer thickness; trends that are similar to those seen for Oximeter A in Fig. 8. This represents a decrease in sensitivity, in a manner analogous to the effect of decreasing ctHb in cerebral phantoms seen in Kleiser et al. [18]. It is also worth noting that Nasseri et al. [28] showed minimal impact on accuracy for superficial layer thicknesses less than 5 mm. The greater sensitivity to superficial layer thickness in our study is likely due to a shallower penetration depth of neonatal/pediatric sensors compared to adult sensors. Clinical evidence of reduced oximeter sensitivity with increasing adipose thickness was provided by Wang et al. [57], who reported that the change in saturation resulting from an exercise test was shown to decrease as adipose thickness varied from 2.5 to 17 mm.
The results of our phantom measurements with superficial layers also is in good agreement with theoretical work by Okada et al. [40].This study concluded that skull and CSF thickness impact NIRS measurements, with CSF layers causing a shunting effect of photons through the relatively non-scattering subarachnoid space and large layer thicknesses significantly reduce the ability to accurately measure cerebral StO 2 . Young et al. [58] also found that extra-cerebral regions may create optical channeling of NIR signals from source to detector. These findings are consistent with results in Figs. 8 and 9, which indicate a strong impact of CSF thickness on sensitivity to changes in blood oxygenation. For a CSF layer thickness of 2.5 mm, the neonatal sensor is essentially completely insensitive to changes in the CVM, and the medium sensor is only moderately sensitive to these changes.

Evaluation of phantom morphology and optical properties
Overall, the 3D printed CVM was shown to provide morphology and optical properties that were sufficiently bio-realistic as to enable testing results that were at least moderately consistent with prior clinical and phantom studies. While the printer has a nominal resolution on the order of 200 μm, the minimum channel size that is printable was much larger, as seen in Fig. 3. Identifying a suitable channel size involved printing a sample with a range of different channel diameters and using micro-CT imaging to identify the smallest channels that were consistently patent. We previously found that to ensure patency, diameters of approximately 0.8-1 mm were required in biomimetic 3D curved channels [32,33]. This limitation highlights the need for printer quality testing of samples over a range of channel diameters. Resolution limitations of current 3D printers may also complicate development of CVMs for testing oximeters at lower ctHb levels, such as the 25 μM value implemented by Kleiser et al. [18] While ideally, the same design would be used with smaller diameter channels, alternate approaches such as dilution of blood may be required to achieve low ctHb levels when such channels cannot be fabricated.
The optical properties provided by the proprietary material are moderately biologically realistic, given the wide range of μ a and μ s ' values that have been measured in cerebral tissues. One of the complicating factors in identifying a suitable set of optical properties is the irregular morphology of the brain along with the difference in optical properties between white and gray matter regions. Given the limited penetration depth of cerebral NIRS, gray matter is likely the most relevant cerebral tissue; however, light that penetrates this relatively thin (2-5 mm) layer, will encounter white matter which can have scattering coefficients several times higher than gray matter. Another limiting factor for our phantom was the lack of ability to modify the optical properties of the printed material. As a result, the μ s ' in our phantom had a slightly sharper spectral decay than desired, falling from about 11 cm −1 at 700 nm to 6 cm −1 at 900 nm. Ideally, a more modest decrease from approximately 8 cm −1 to 6 cm −1 would have been achieved. In future publications, we will explore the use of alternate methods to provide greater flexibility for custom materials that achieve target optical properties.

Conclusion
The results of this study provide strong evidence that modular, 3D-printed channel array phantoms are a promising test method for cerebral oximetry. In many cases -particularly those involving thin extracerebral layers -phantom StO 2 measurements collected using clinical devices showed excellent agreement with reference CO-oximeter values. For thicker superficial layers, significant degradation in performance was seen for one of the devices, whereas the other was more robust. Based on comparisons with prior phantom and clinical results, our testing approach provided both absolute quantities and trends that were reasonable. Although improvements in optical property tunability of 3D-printing materials and microchannel print quality will enhance the clinical realism of our approach, this method has the potential to become a powerful tool for elucidating biological factors that impact device performance. Our future work will further validate this approach by testing a greater range of NIRS oximeters and sensors and addressing other confounding biological factors.

Disclosures
The mention of commercial products, their sources, or their use in connection with material reported herein is not to be construed as either an actual or implied endorsement of such products by the Department of Health and Human Services, Food and Drug Administration. The authors declare that there are no conflicts of interest related to this article.