1.

Introduction

Periocular inflammatory and infiltrative conditions are of particular importance given the proximity to the eye, thereby raising the risk for dry eyes, exposure keratopathy, eyelid malposition, blurry vision, eye pain, vision loss, double vision, and rarely, extension to the central nervous system. Diseases that affect the region include thyroid eye disease, rosacea, and eczema, as well as basal cell carcinoma, squamous cell carcinoma, melanoma, and vascular malformations. Clinical diagnosis and management of these diseases are oftentimes dependent on invasive, time-intensive, and costly procedures and testing, as well as immunotherapy, surgical resection, and periocular reconstruction. Non-invasive objective approaches to measuring local inflammation and infiltration are therefore needed to better quantify disease activity, progression, and treatment response.

The periocular region is unique due to a complex ocular adnexal anatomy (Fig. 1).¹ It is comprised of the eyes, including the upper and lower eyelids, eyelashes, and medial and canthal regions, as well as the nose, brow, and midface. The underlying bony framework creates a unique geometric structure with variable height differences, in part affected by the deep orbital bones and structures, in addition to the supraorbital ridge and frontal bone rostrally, nasal bone and malar process of the maxilla medially and inferiorly and zygoma laterally. The soft tissue landmarks are the suprabrow rostrally, medial canthus and nasofacial sulcus medially, nasojugal crease and infraorbital crease inferiorly, and lateral canthus laterally.

Fig. 1

Periocular anatomy. Demarcation of eyelid and canthal structures, in addition to cutaneous folds, creases, and underlying bony framework.

Several imaging techniques are available to image the skin, but these probe a superficial depth (Table 1). For example, reflectance confocal microscopy (RCM) is an optical imaging technique that utilizes a single near-infrared (NIR) wavelength (830 nm) to scan the superficial dermis to a penetration depth of $\sim 200$ to $300\mu \mathrm{m}$.² RCM is limited by its use of a single wavelength and interrogates only up to the papillary dermis or upper reticular dermis. Similarly, multiphoton microscopy can provide cellular-level resolution for skin imaging, but it only penetrates up to the superficial dermis, has a limited field of view, and requires eye-safety protocols when used around the periocular region.³ Ultrasound, particularly at high frequencies, achieves a resolution of subcutaneous lesions comparable to measurements obtained via histology, but it is also limited by a penetration depth of $\sim 1.5\mathrm{mm}$ at 100 Hz.⁴ Optical coherence tomography (OCT) uses an infrared laser to image the skin with $\sim 2\mathrm{mm}$ of penetration depth. Although a large meta-analysis found some evidence for the role of OCT in identifying basal cell carcinoma (BCC) when compared with visual inspection and dermoscopy, no evidence-based guidelines currently exist for using OCT to diagnose or monitor cutaneous inflammatory disease and non-BCC malignancies.^5,6 OCT possesses advantages in terms of fast acquisition, high resolution of skin morphology, and detection of particulate formations, such as vaccine-loaded carriers, nanoshells, and nanoparticles.^7–9 It is however limited in acquiring high-resolution images of non-particulate skin structures, similar to that of RCM.¹⁰

Table 1

Subsurface imaging technology comparison.

Spatial frequency domain imaging (SFDI) is a non-invasive, non-contact wide-field imaging platform that applies patterned NIR illumination to quantify the spatially resolved optical properties of in vivo subsurface tissue and can be performed in the clinical setting.^11–19 It specifically allows for visualization of tissue volumes up to 5 mm beneath the surface. A region of interest (ROI) along a targeted tissue is illuminated over a range of wavelengths, and the illuminations are projected and recorded by a camera to extract the diffuse reflectance at each specific wavelength and frequency. Ultimately, reduced scattering and absorption coefficients are determined by fitting to a known forward model.²⁰ Validation studies were performed by our group using tissue-simulating phantoms with a reported accuracy of $\sim 6\%$ and 3% in absorption and reduced scattering parameters, respectively.²⁰ SFDI can quantify tissue function and structure and has successfully predicted cutaneous flap survival during surgical reconstruction, laser treatment response in port-wine stains, and burn severity.^19,21

Although SFDI has been successfully applied to in vivo tissues that are relatively flat and unvarying in terms of height (e.g., volar forearm, abdomen, dorsum), measurements on portions of the anatomy that have significant height variations and curvature may result in optical property inaccuracies due to limitations of the height and curvature correction algorithms. Correction algorithms for these issues exist and are used; they have limitations, but developmental efforts continue.²² To address this gap, we characterize the potential geometric impact of neighboring bony (nasal process of the maxilla, nasal bone, supraorbital rim, lateral zygoma) and soft tissue (nasal cartilage and anterior projection, brow ptosis) regions on the accuracy of SFDI measurements by fabricating, imaging, and analyzing a facial phantom comprised of poly(dimethylsiloxane) (PDMS) embedded with titanium dioxide granules (scatterer) and India ink (broadband absorber), to identify robust and accurate ROIs to ultimately image human patients. These PDMS phantoms are developed in the lab given their portability, ability to mimic in vivo biologic conditions, and overall durability.^23–25

2.

Materials and Methodology

2.2.

Imaging

SFDI measurements were obtained using the Reflect RS™ (Modulim, Inc., Irvine, California, United States) to collect calibrated diffuse reflectance images over a $15\times 20{\mathrm{cm}}^2$ field of view. The SFDI light sources consist of eight light-emitting diode sources with eight center wavelengths (470 to 851 nm). For each wavelength, sinusoidal light patterns at five spatial frequencies (0, 0.05, 0.10, 0.15, and 0.2/mm) were projected onto the tissue, following conventional procedures previously used.²⁶ The Modulim Reflect RS device has built-in cross polarizers, which minimize glare (i.e., surface specular reflection).

The facial phantom was affixed to a custom-made cart-based chin rest with dual forehead (FH) apposition to simulate human patient positioning and was imaged in a dark room to minimize the effects of ambient illumination [Figs. 2(a) and 2(b)]. The phantom was imaged via an “en face” and “side profile” (nasal tip over contralateral cheek junction) at each of the nine image acquisitions. The posterior flat surface of the phantom was then imaged 15 times by flipping the phantom and re-affixing the phantom to the cart-based chin rest with dual FH apposition. Each region was imaged three consecutive times using the device. A single acquisition, which includes a single image taken at each of the eight wavelengths at all five spatial frequencies, took $\sim 30\mathrm{s}$ to complete. Using the software accompanying the instrument (Modulim Inc.), data processing of three repetitions for all anatomical locations on a single-facial phantom took less than 10 min. Following the conventional SFDI measurement technique, a flat PDMS-based tissue-simulating reference phantom with known optical properties was measured at each imaging time point under the same lighting conditions as that of the facial phantom. These imaging data were used to calibrate the SFDI data.

Fig. 2

(a) SFDI system consisting of a light source projecting spatially modulated light onto a facial phantom positioned in a chin rest. Facial phantom with delineated periocular anatomy captured (b) “en face” and (c) “side profile” with (d, e) identified ROI, ITQ, INQ, STQ, FH, RLNB, and CEM used to measure ${{\mu }_s}^{\prime }$ and ${\mu }_a$ coefficients. (f) Reduced scattering plots for each measured optical property.

3.

Results

3.1.

Precision and Accuracy in Facial Phantom Measurement

The mean absorption (${\mu }_a$) and reduced scattering (${{\mu }_s}^{\prime }$) coefficients for the back surface of the phantom were 0.0275 and $0.736{\mathrm{mm}}^{-1}$, respectively, and the standard deviations were 0.000577 and $0.00685{\mathrm{mm}}^{-1}$ at the 851 nm wavelength (Table 2). The measured values for the absorption and reduced scattering coefficients reveal tight clustering of the data across each wavelength (Fig. 3).

Table 2

Optical properties of the posterior flat surface of the facial phantom at 851 nm.

μa (mm−1)		μs′ (mm−1)
Mean	Standard deviation	Mean	Standard deviation
0.0275	0.000577	0.736	0.00685

Fig. 3

Optical properties of the posterior flat surface of the facial phantom at all wavelengths.

These values were precise for the back surface of the phantom, as well as for each of the periocular ROIs (INQ, ITQ, STQ) and FH of the facial phantom as repeated measurements of each area resulted in similar values with coefficients of variation for ${\mu }_a$ at 0.054, 0.065, 0.036, and 0.041 (${\mathrm{mm}}^{-1}$) and for ${{\mu }_s}^{\prime }$ at 0.047, 0.043, 0.028, and 0.02 (${\mathrm{mm}}^{-1}$), respectively (Fig. 4).

Fig. 4

Facial phantom optical property scatter plot. The distribution of measurements for the back of the phantom, “side profile,” and “en face” for ROIs at 851 nm.

These values were also accurate in that the periocular ROIs and FH of the facial phantom were similar in value to that of the back surface of the phantom, with a maximum mean difference of 0.003 to 0.031 ${\mathrm{mm}}^{-1}$ for ${\mu }_{\mathrm{a}}$ and ${{\mu }_s}^{\prime }$ respectively, which are not significant (Table 3). The optical properties for each ROI at all eight measured wavelengths (470 to 851 nm) are also provided (Table S1 in the Supplementary Material).

Table 3

Facial phantom ROI optical properties at 851 nm. Two sample t tests with 99.9% confidence intervals are presented.

μa (mm−1)					μs′ (mm−1)
ROI	“en face”	Back	Difference	Percent difference	ROI	“en face”	Back	Difference	Percent difference
ITQ	0.03 (0.029, 0.031)	0.028 (0.028, 0.029)	−0.002 (0.001, 0.003)	6.9 (4.2, 9.6)	ITQ	0.785 (0.766, 0.805)	0.746 (0.737, 0.754)	−0.04 (0.023, 0.056)	5.6 (4.1, 7.2)
INQ	0.027 (0.024, 0.03)	0.027 (0.027, 0.028)	0 (−0.003, 0.003)	0 (−10.3, 11.4)	INQ	0.729 (0.704, 0.755)	0.735 (0.728, 0.742)	0.005 (−0.028, 0.017)	−0.7 (−3.7, 2.4)
STQ	0.028 (0.026, 0.029)	0.027 (0.027, 0.028)	0 (−0.001, 0.002)	4.2 (1.6, 7)	STQ	0.761 (0.739, 0.783)	0.73 (0.725, 0.735)	−0.031 (−0.011, 0.051)	4.2 (1.6, 7)
FH	0.03 (0.028, 0.032)	0.028 (0.027, 0.029)	−0.003 (0.001, 0.004)	9.9 (6.5, 13.4)	FH	0.765 (0.735, 0.795)	0.734 (0.726, 0.743)	−0.031 (0.004, 0.057)	4.2 (2.5, 5.9)
CEM	0.021 (0.019, 0.024)	0.028 (0.028, 0.029)	−0.007 (−0.009, −0.005)	−23.2 (−29.8, −15.9)	CEM	0.889 (0.853, 0.919)	0.883 (0.875, 0.892)	0.005 (−0.028, 0.038)	2.5 (−1.5, 6.7)
RLNB	0.02 (0.014, 0.026)	0.029 (0.028, 0.029)	−0.009 (−0.015, −0.003)	−29.5 (−47.8, −4.6)	RLNB	0.747 (0.638, 0.857)	0.88 (0.867, 0.892)	−0.132 (−0.239, −0.025)	−12.2 (−23.1, 0.1)

However, ${{\mu }_s}^{\prime }$ of the RLNB was artificially decreased at $0.645{\mathrm{mm}}^{-1}$ with an increased spread of data [$0.068{\mathrm{mm}}^{-1}$ standard deviation ($\sigma$)], compared with ITQ ($0.79{\mathrm{mm}}^{-1}$ mean, $0.036\sigma$), INQ ($0.73{\mathrm{mm}}^{-1}$ mean, $0.03{\mathrm{mm}}^{-1}$ $\sigma$), STQ ($0.761{\mathrm{mm}}^{-1}$ mean, $0.02\sigma$), and FH ($0.765{\mathrm{mm}}^{-1}$ mean, $0.015{\mathrm{mm}}^{-1}$), whereas the CEM ($0.754{\mathrm{mm}}^{-1}$ mean, $0.15\sigma$) was minimally impacted (Figs. 5 and 6).

Fig. 5

Box and whisker plots demonstrating the artifactual decrease in ${\mu }_a$ for both the RLNB and CEM regions and the artifactual decrease in ${{\mu }_s}^{\prime }$ for RLNB.

Fig. 6

Mean optical property distributions. The RLNB demonstrates artificially decreased values in ${\mu }_a$ ($0.02{\mathrm{mm}}^{-1}$ mean) and ${{\mu }_s}^{\prime }$ ($0.645{\mathrm{mm}}^{-1}$ mean), with increased standard deviations (0.0038 and $0.068{\mathrm{mm}}^{-1}$, respectively), whereas the CEM demonstrates artificially decreased values in ${\mu }_a$ ($0.02{\mathrm{mm}}^{-1}$ mean) with a tight fit ($0.001{\mathrm{mm}}^{-1}$ $\sigma$), compared with the ${\mu }_a$ of the FH ($0.03{\mathrm{mm}}^{-1}$ mean, $0.001{\mathrm{mm}}^{-1}$ $\sigma$) and ITQ ($0.03{\mathrm{mm}}^{-1}$ mean, $0.002{\mathrm{mm}}^{-1}$ $\sigma$) at 851 nm.

On the contrary, the ${\mu }_a$ for both CEM ($0.021{\mathrm{mm}}^{-1}$ mean, $0.0009{\mathrm{mm}}^{-1}$ $\sigma$) and RLNB ($0.02{\mathrm{mm}}^{-1}$ mean, $0.0038{\mathrm{mm}}^{-1}$ $\sigma$) was artifactually lower compared with ITQ ($0.03{\mathrm{mm}}^{-1}$ mean, $0.002{\mathrm{mm}}^{-1}$ $\sigma$), INQ ($0.027{\mathrm{mm}}^{-1}$ mean, $0.002{\mathrm{mm}}^{-1}$ $\sigma$), STQ ($0.028{\mathrm{mm}}^{-1}$ mean; $0.001{\mathrm{mm}}^{-1}$ $\sigma$), and FH ($0.03{\mathrm{mm}}^{-1}$ mean, $0.001{\mathrm{mm}}^{-1}$ $\sigma$) with increased distribution of data for RLNB ($0.0038{\mathrm{mm}}^{-1}$ $\sigma$) (Figs. 5 and 6).

4.

Discussion

By measuring the wavelength-specific absorption and reduced scattering coefficients as a function of spatial frequency, SFDI can quantify tissue optical properties^18,19 that will enable differentiation of normal tissue from tissue undergoing pathologic transformation. Its ability to penetrate up to 5 mm helps differentiate it from OCT, ultrasound, and RCM, which can only penetrate up to 2 mm in depth and are further limited in the ability to acquire high-resolution images of non-particulate skin structures.^7–10 SFDI has demonstrated utility in evaluating burn severity, tissue transfer flap viability following surgical reconstruction, laser treatment efficacy in patients with port-wine stains, and changes in cerebral blood flow following cardiac arrest.^{12,14,15,17,19} Ongoing studies have enabled quantification of oxy and deoxyhemoglobin concentration as well as oxygen saturation,²⁷ characterized in vivo tissue hemodynamics, and identified subsurface vascular changes in patients before they are clinically obvious. SFDI therefore has the potential to quantify tissue inflammation and hemodynamics and define which metrics correlate with disease activity. By longitudinally following changes in tissue optical properties, we expect to develop a quantitative SFDI-derived metric to study the kinetics of periocular disease evolution and its response to therapy.

The periocular region, however, possesses a unique geometric landscape that may complicate SFDI imaging within this anatomic region. Although well-established studies have demonstrated SFDI’s utility for in vivo imaging of relatively flat anatomic regions, the notable curvature and height heterogeneity of the face may result in measurement errors of optical properties due to existing algorithm limitations. Optimization of these correction algorithms involves calibrating the light source intensity with light during surface profile acquisition and implementing a Lambertian reflectance model-based correction, but the process has yet to be finalized and is confined by height variations up to 3 cm and tilt angles within $\pm 40\mathrm{deg}$. Furthermore, the correction algorithm is dependent upon the phase profilometry data quality.²²

We therefore outline an approach to characterize the impact of the unique periocular anatomy using a fabricated PDMS facial phantom due to its biologic optical similarity to that of human tissue to explore the absorption and reduced scattering coefficients along the periocular region using SFDI. The optical property measurements of each periocular quadrant, i.e., inferior temporal, inferior nasal, and superior temporal, as well as the FH, were reproducible within each cohort, similar in magnitude to each other, and that of the flat back surface of the facial phantom (i.e., no height variability), whereas the CEM and RLNB yielded artifactual data with notable spread in measured values for the RLNB. The minimal variability noted between the “en face” and posterior surface of the phantom may, in part, be due to scattering potentially being greater along the rougher “en face,” or surface, side.

The errors inherent to the RLNB ROI are likely due to significant height and curvature variations, in addition to inherent scattering properties (areas where light is more likely to be redirected), that the surface correction algorithm cannot overcome via correction due to phase unwrapping abnormalities. The CEM, on the contrary, exhibits curvature variations that may have implications in localized areas of increased scattering along a smaller scale than RLNB, specifically along the region where the upper and lower eyelid margins meet. Further, the surface curvature was found to be concave along the RLNB and CEM, whereas the FH, ITQ, INQ, and STQ were found to be convex using a 3D facial scanner and curvature map software platform. Concave surfaces may predispose to error using SFDI as there is the likelihood of secondary illumination. The Reflect RS™ SFDI analysis software (MI Analyze) employs a profile-based correction algorithm to account for varying tissue height and angle with respect to the imager.²⁸ The model is able to account for variations in height within 2 cm from the target location and angles of $\sim \pm 30\mathrm{deg}$ to the plane of illumination. The correction algorithm was validated on tilted phantoms and phantoms having different radii of curvature. However, it does not account for the secondary illumination of the tissue by light reflected from surrounding tissue structures within the illumination field of the device.

Practically, the CEM is a suboptimal ROI when imaging patients. This region of the face, in a human subject measurement context, is a thin tissue layer overlying a fluid-filled globe. The boundary conditions of the computational modeling are considerably different than the real case, and thus, the process for determining the optical properties of this volume of tissue would not be expected to produce high fidelity, accurate optical properties. A potential option moving forward would be to interrogate a multi-material phantom with integrated dynamic flow.

Finally, light source positioning (“en face” versus “side profile”) and/or patient positioning do not appear to impact SFDI measurement along the ITQ, INQ, STQ, and FH, further suggesting that the geometry of the selected ROIs along the periocular region has acceptably little influence on SFDI measurements. Collectively, these results suggest that corresponding measurements of human subjects should yield accurate optical property values.

Overall, our findings suggest that the curvature of each chosen ROI most likely impacts absorption, whereas the height variation impacts scattering. All of these findings bear out at all measured wavelengths, with data depicted for 851 nm. These outcomes emphasize the need to carefully choose the ROI along the periocular profile, given its topography and adjacent structures (e.g., nasal profile or supraorbital rim) to minimally affect biologic SFDI measurements. Specifically, ITQ, INQ, STQ, and FH were similar to that of the back flat surface of the facial phantom, suggesting that these ROI are ideal regions to conduct human SFDI analyses. These results, i.e., the impact of curvature, are likely applicable to other tissue surfaces and are not limited to skin-like textures.

Our approach is not without limitations, however. A single facial phantom for a youthful rendering of a female face was used. We therefore did not study the impact of natural variations that can occur within the face particularly along the superior sulcus or brow or evaluate the potential impact of aging on the underlying ligamentous and bony framework. Furthermore, we only assessed the impact of curvature and height variation and will need to further study the roles of underlying proptosis as well as orbital orientation on either side of the face. Finally, human skin is a complex tissue structure, which necessitates the study of a multi-layer region, which is not accounted for via this technique.

5.

Conclusions

To our knowledge, we describe the first characterization of SFDI performance along the periocular facial region using a facial phantom with known optical properties. By creating an internal control (the posterior flat surface of the facial phantom), the various ROIs around the eyes and adjacent structures were able to be compared with each other to assess precision, reproducibility, and accuracy. The periocular regions, in particular, ITQ, INQ, STQ, and FH, were found to be regions in which optical properties can be determined with high confidence. We have thus characterized the potential geometric impact of neighboring bony and soft tissue regions on the ability of SFDI measurements to accurately determine the optical properties of a fabricated PDMS facial phantom toward identifying robust and accurate ROIs. Based on these results, we cautiously aim to conduct characterization studies of control patients, with the ultimate goal of using the technology to characterize structural and functional changes around the periocular region to aid disease diagnosis and therapeutic response.