1. Introduction

Monitoring pupillary size and light-response is a key component of the neurologic examination in comatose patients with brain injury from multiple etiologies. Changes in these pupillary parameters can provide important clues to neurologic deterioration, elevated intracranial pressure and herniation [1]. The prognostic and clinical value of pupillary examinations have been well established for traumatic and non-traumatic injuries [2–17] and have been featured in benchmark studies as far back as the early 1980s [3,8]. For example, during transtentorial herniation, pupil size and reactivity are the most sensitive and easily identifiable findings [2]. In a study of functional recovery in patients with traumatic transtentorial herniation, only 3.5% of patients with bilaterally fixed (unresponsive to light) and dilated pupils at admission had a functional recovery [18]. The combination of the pupillary examination and Glasgow Coma Scale (GCS) score provides more accurate prognostic information, and patients with a GCS of 3 with reactive pupils have a 33% survival rate, as opposed to patients with a GCS of 3 and fixed dilated pupils who have no reasonable chance of recovery [9]. Therefore, early and accurate detection of pupillary abnormality can both reduce the time to appropriate interventions, which are frequently life-saving, and avoid frequent computed tomographic scans. Additionally, accurate pupillary examination can help detect transtentorial herniation in some head injury patients with low speed trauma who undergo early herniation [18].

When performing a pupillary examination in patients with traumatic brain injury, the medical staff is looking for abnormalities, such as an asymmetric pupil size (anisocoria), irregular pupil shape, or a sluggish or absent pupillary light response. According to standard care protocols, upon detection of such an abnormality, a physician should be notified immediately and often head CT-scan and continuous intracranial pressure monitoring are subsequently used to aid in definitively identifying the cause of the abnormal pupil. Pupillary evaluation in comatose patients is often performed in a subjective manner by manually opening the patient eyelids, measuring the size of each pupil with a pupil-gauge ruler, and using a flashlight to illuminate the open eye in a darkened room to assess pupil reactivity. The manual method provides diameter evaluation with spatial resolution of ∼1mm and its output regarding the pupil reactivity is binary. This method suffers from large variability both across the nursing staff and within the same examiner [19–22]. Alternatively, pupil size and reactivity can be evaluated using a pupilometer, which is a handheld imaging device that provides measurement standardization and accuracy of ∼0.1mm in the measured pupil diameter [23]. Importantly, evaluation using either of these two manual methods is carried out variably, intermittently every 4–6 hours based on local protocols and on the availability of medical and nursing personnel. However, this frequency of pupillary assessment might be too low to obtain timely alert and prevent irreversible brain damage [5,17]. For example, Cohen et. al have shown [17] that in patients with acute traumatic epidural hematoma a delay of 90 minutes or more between the detection of pupil size asymmetry (anisocoria) and a craniotomy procedure was associated with greater mortality. Further, a delay of 70 minutes or less resulted in no mortality and a good or reasonable outcome.

Here we present a novel method for monitoring pupillary size and reactivity in close eyelid condition, specifically designed to meet the needs of unconscious patients. Our method is based on applying near-IR light through the temple, which penetrates the eye, emitted through the pupil and the closed eyelid, and then imaged by a camera placed in front of the eye. To assess pupil reactivity, we apply short pulses of visible white light that induce pupil contraction through the closed eyelid. Pupil diameter was evaluated using a generative image-analysis model that accounts for the light scattering through the eyelid. The quantitative measurement of pupil diameter before and after the white-light pulse, can lead to a continuous, rather than binary, scale for pupil reactivity. This method, validated in a clinical trial, may lead to automated, accurate and continuous pupillary monitoring that will alert on pupillary changes in real time and thus save delays that postpone neurologic lifesaving procedures.

2. Methods

2.1 Imaging device

To measure pupil size and light-reactivity, we developed an imaging device consisting of a near-IR sensitive camera, a near-IR LED (810nm wavelength), and a white LED, all mounted on a chin-rest (Fig. 1(a) and Appendix). The camera and both LEDs were synchronized using a dedicated controller as illustrated in Fig. 1(b). The pupil was imaged by illuminating the eye through the temple with the near-IR LED and imaging the eye from the front with either an open or a closed eyelid. This method relies on the observation that near-IR light penetrates through the temple into the eyeball and is partially emitted from the eyeball through the pupil and across the eyelid. To facilitate imaging, the selected IR wavelength should be both absorbed by the iris and transmitted by the other components of the eye: the crystalline lens, the cornea, and the eyelid. These conditions, which allow the light to emanate from the eye through the eyelid, are satisfied by near-IR in the range of 750-900nm [24]. Additionally, using near-IR light does not influence the size of the pupil because the retina is insensitive to light at this spectral range. To induce the light pupillary reflex, we illuminated the eye from the front using a pulse of white LED light. This light can be seen even through a closed eyelid, which acts as a red-pass filter with ${\sim} 5\%$ transmission [25]. To quantify the light pupillary reflex, we imaged the pupil in IR before and after the visible light pulse and compared the pupil diameters in both IR measurements. Measuring the light pupillary reflex is an important medical assay on its own right and provides a controlled environment for testing our method. In particular, it tests the capacity of our device to detect changes in pupil diameter as well as expands the range of measured diameters.

 

Fig. 1. The experimental system. (a) The imaging device consists of an near-IR camera, a $810nm$ near-IR LED and a white LED. These components are mounted on a chin-rest, imaging one eye at a time. Here, the IR-LED illuminated the left eye through the temple, and the white LED illuminated the same eye from the front to induce the light pupillary reflex through the closed eyelid. (b) System block diagram. Upon an initiation command from the PC, a programmable controller set the timing of image acquisition and illumination pulses of the two LEDs via hardware trigger signals. Images then transferred from the camera to the PC. (c, d) Measurement protocol. The experiment was performed in a dark room. Each IR image acquisition consisted of a $0.1\textrm{s}$ pulse of the IR-LED, synchronized with a $0.1\textrm{s}$ exposure of the camera (red line). To find the optimal IR-LED position and camera parameters, we first took several images with an open eyelid, until the pupil was clearly seen (left image in d). Subsequently, we imaged the eye in closed eyelid using a three-step automatic acquisition sequence: A first image was taken in IR (d, middle column); the eye was illuminated from the front by a $1s$ pulse of white light (blue line); A second image was taken in IR (d, right column). In some cases, the volunteer was asked to stretch down the eyelid, which reduced both eyelid scattering and absorption (d, bottom row). (d) The pupil is clearly seen as a bright spot on the closed or stretched eyelid.

Download Full Size | PDF

2.2 Measurement protocol

We performed a clinical trial on 39 healthy volunteers in accordance with the Institutional Review Board of the Hadassah Medical Center (S/N HMO-0555-17) and radiation safety requirements (IEC 62471 international standard), and under an ophthalmologist supervision (see Appendix). Each volunteer provided an informed written consent for participating in the trial. For each volunteer we measured both eyes, one eye at a time. The trials were conducted in a dark room, to reduce background illumination and the camera was acquiring an image during the entire IR pulse. The IR-LED was housed in an opaque sleeve contacting the temple, which placed the IR-LED 2cm from the skin.

First, to find the optimal position of the IR-LED on the temple, we recorded a series of IR images of the pupil with an open eyelid, scanning the position of the IR-LED on the temple until the pupil was clearly seen (Fig. 1(d), open eyelid). Second, using the same IR-LED position, we recorded several IR images of the pupil with a closed eyelid, verifying the pupil was indeed visible through the eyelid. Finally, we applied the following three-step protocol (Fig. 1(c),(d)): (1) Took an image in IR using a $0.1\textrm{s}$ illumination pulse to measure pupil size; (2) waited for $0.2\textrm{s}$ and then illuminated the eye from the front with a $1\textrm{s}$ pulse of the white LED to induce the pupillary light reflex; (3) waited for 1ms and took a second image in IR using another $0.1\textrm{s}$ pulse to measure the pupil size after it had contracted. During this sequence, the volunteers were asked to fix their gaze to minimize pupillary motion. The short delay between the white light pulse and the second IR imaging was set to minimize pupillary motion and to avoid imaging the expansion of the pupil expected after the white light turns off.

Figure 1(d) demonstrates that both the pupil and its contraction following white light exposure are clearly seen through the closed eyelid. Comparing the images with open and closed eyelids, we see that the eyelid both absorbs and scatters the IR light emanating from the pupil. We noticed that both these effects are reduced when the eyelid was stretched downwards by the volunteer (Fig. 1(d), bottom). Hence, for some of the volunteers the experimental protocol was performed both with a relaxed closed eyelid and with a stretched closed eyelid. Overall, we managed to image pupils through closed eyelids in $38$ of the $39$ volunteers. For $34$ volunteers we managed to obtain at least one image pair taken before and after a white light pulse, in which the pupil was clearly seen. Our analysis includes 242 such image pairs.

2.3 Data analysis

Due to IR scattering through the eyelid, the pupil image has no distinct boundaries, as shown by the intensity profiles in Figs. 2(a)-(e). Hence, the main challenge in extracting the pupil diameter from the acquired images is formulating a consistent and meaningful definition of this diameter. To this end, we formulated a generative model that approximates the process of diffusive light scattering from the unscattered pupillary disk (Fig. 1(d)) through the eyelid, as illustrated in Figs. 2(f)-(g). This approach is consistent with the diffusive nature of light propagation in tissue and particularly in skin [26–28], as well as with the fact that the typical eyelid thickness of 4–5mm [29,30] is larger than the diffusive length of light in skin. Additionally, light emanating from the pupil towards the inner side of the eyelid is already diffused due to its passage through tissues. Therefore, light diffusion through the eyelid can be modeled as a convolution of the pupillary disk with a two-dimensional Gaussian scattering kernel.

 

Fig. 2. Pupil diameter measurement. (a-e) Extraction of pupil intensity profiles from the acquired images. The center of the pupil was manually located on the IR images taken before (a) and after (c) the white light pulse. Scale-bars are $5\textrm{mm}$. (b, d) The pupil intensity profiles were calculated by averaging each image intensity across 10 rays covering a ${20^{ {\circ} \; }}$ section (orange lines in (a)) to avoid the eyelashes. The linear background profiles due to the stronger IR illumination coming from the temple side (dashed lines) were calculated and subtracted from the measured intensity, resulting in the pupil profiles in (e). (f) Generative model. The model consists of four parameters and generates two pupillary model images. A disk of diameter ${D_1}$ and intensity A (top left image) was convolved with a Gaussian scattering kernel of standard deviation $\sigma$ (middle image), resulting in the first model image (top left). This image modeled the pupil before the white light exposure. A second disk of diameters ${D_2}$ and same intensity A was convolved with the same Gaussian to form the second model image, which described the pupil after the white light exposure (bottom row). The common A and $\sigma $ parameters reduce the number of free parameters to four per each pair of consecutive images. (g) The one dimensional profiles of the pupil model images. The model profiles are not Gaussians due to the convolution process. These profiles were used to calculate a cost function, which was minimized in a gradient descent process to obtain the best fit of the free parameters $({{D_1},{D_2},\; \sigma ,A} )$.

Download Full Size | PDF

Implementing this approach at the single-image level requires three fit parameters: The pupil diameter D, the intensity of the unscattered pupillary disk, A, as would have been measured without scattering, and the scattering parameter of the eyelid $\sigma $, which represents the standard deviation of the Gaussian scattering kernel. These three parameters would be used to generate a model-image by starting with an image of a disk with diameter D and uniform intensity A and convolving this image with the Gaussian kernel. However, using these three parameters, $({D,\; A,\; \sigma } )$, to fit the model image to a single data image is under-constrained due to parameter coupling. For example, decreasing D might be compensated by increasing $\sigma $, resulting in an equivalently good fit. To circumvent this issue, we used the fact that thanks to the short delay between the acquisitions of consecutive IR images, pupillary motion between the two images is small. Hence, we assume that for such pairs the disk intensity A and scattering $\sigma $ are identical. Therefore, rather than modeling a single image using three parameters, we modeled a pair of consecutive images using four parameters: one diameter per image $({D_1},\; {D_2})$, and common A and $\sigma $. The constraints added by the four-parameter formulation of the fitting problem reduced the parameter coupling compared with the three-parameter case. For each pair of images, we used this generative model to fit the four parameters via gradient descent optimization (Fig. 2 and Appendix). Briefly, we fitted the intensity profiles of the generated images to the profiles of the measured images after performing background subtraction. The confidence intervals of the fitted values were evaluated using bootstrapping.

2.4 Validation of the pupillary reflex in closed eyelids using ultrasound

To verify that white light illumination in closed eyelids induces the light pupillary reflex, we performed ultrasound measurements on five of the volunteers using a designated ocular ultrasound device (Ellex ultrasound system), operating with a 20MHz probe in B-scan mode. The volunteers were tested in closed eyelids and in a dark room. Their pupil was measured before and during white light illumination, equivalent to the illumination used in the IR-based device. The volunteers were asked to fix their gaze during the scan. The pupil was identified and the probe was positioned such that the scan plane was aligned with the pupil diameter.

3. Results

The measured pupil diameters, ${D_1}$ and ${D_2}$, before and after the white light pulse, respectively (Fig. 3(a)), are within the normal pupil diameter range of $2 - 7\textrm{mm}$. In $98.7\%$ of the cases, pupil diameter decreased after the white light pulse (${D_2} < {D_1}$), consistent with the diameter reduction expected due to the light response reflex. The histogram of the diameter ratio ${D_2}/{D_1}$ is plotted in Fig. 3(b), where the mean ratio over all measurements was $0.63 \pm 0.13$ (mean ± standard deviation). We see no statistically significant effect of the eyelid condition (normal vs. stretched) on the measured diameters or diameter ratio. The distribution of the scattering parameter $\sigma $ is plotted in Fig. 3(c), with a mean over all measurements of $1.19 \pm 0.28\textrm{mm}$. Additionally, there is a slight, though statistically significant, decrease in $\sigma $ in the stretched ($1.15 \pm 0.32\textrm{mm}$) vs. normal ($1.22 \pm 0.6\textrm{mm}$) eyelid condition (KS ${p_{\textrm{val}}} = 0.0024$ and KS distance of 0.25). A decrease in $\sigma $ is consistent with the fact that the stretched eyelid is thinner and, hence, expected to scatter the IR light less.

 

Fig. 3. Pupil size measurements. (a) Measured pupil diameters in $242$ image pairs, showing ${D_2}$, the diameter after the white light illumination vs. the ${D_1}$ diameter before white light illumination. Each point represents an image pair taken for one eye either with a normally closed eyelid ($N = 165$, blue circles) or stretched closed eyelid ($N = 77$, yellow squares). Whiskers represent $95\%$ confidence intervals for each measurement. In $98.7\%$ of the measurements the pupil contracted upon white light exposure, ${D_2} < {D_1}$. (b) Pupil diameter ratio ${D_2}/{D_1}$ for the data in (a). The mean ratio for the images with normally closed eyelid was $0.62 \pm 0.13$ (mean ${\pm} $ standard deviation), the the mean ratio with stretched closed eyelid was $0.65 \pm 0.13$, with no statistically significant difference between the ratios in both conditions (using both t-test and KS-test). (c) The distributions of the scattering parameter $\sigma $ for both eyelid conditions, with a mean of $1.19 \pm 0.28\textrm{mm}$ (mean ${\pm} $ standard deviation), across all $242$ measurements The value of $\sigma $ in stretched closed eyelids was $1.15 \pm 0.32\textrm{mm}$, compared with $1.22 \pm 0.6\textrm{mm}$ in normally closed eyelid. Despite its small magnitude, this difference was statistically significant (KS ${p_{\textrm{val}}} = 0.0024$ and KS distance of $0.25$). (d) The distributions of the single-sided confidence intervals (CI) of all $484$ images ($242$ pairs), plotted for both closed eyelid conditions. There was no statistically significant difference between the two distributions (KS ${p_{\textrm{val}}} = 0.26$). The overall mean CI value was $0.15 \pm 0.11\textrm{mm}$ (mean ± standard deviation), and $97\%$ of all measurements had CI<0.4 mm.

Download Full Size | PDF

To compare the variability of the scattering parameter $\sigma $ within individual volunteers and its variability across all the volunteers, we considered the $19$ volunteers for which we had more than $5$ measurements. The average standard deviation of $\sigma $ across these volunteers was $0.19\textrm{mm}$, lower than the standard deviation of $\sigma $ across the entire data set, which was $0.28\textrm{mm}$. This observation is consistent with the notion that $\sigma $ is a person-specific feature. The accuracy of the measurement was evaluated by calculating the confidence intervals (CI) of ${D_1}$ and ${D_2}$ using bootstrapping (Appendix). The CI distribution (Fig. 3(d)) shows that $97\%$ of the measurements have $\textrm{CI} < 0.4\textrm{mm}$, with no statistically significant difference between the two eyelid conditions (KS ${p_{\textrm{val}}} = 0.27$ and KS distance of 0.096).

Figure 4 shows representative ultrasound scans before and during white light illumination, indicating the pupillary reflex is induced when the eyelid is closed. The mean diameter ratio, equivalent to ${D_2}/{D_1}$ in the IR measurement, was $0.7 \pm 0.06$ (mean ± standard deviation, for $N = 5$), consistent with the IR measurements.

 

Fig. 4. Ultrasound verification of pupillary light response. The pupil diameter in five of the volunteers was measured in closed eyelids using ocular ultrasound, before (left image) and during white light illumination (right image). The ultrasound images show a cross section of the eye, with the cornea and pupil marked in yellow arrows. The pupil is the gap between the two sides of the iris shown in the cross section. Here, the pupil contracted from a diameter of 6.49 mm in the dark, to 4.85 mm during white light exposure (ratio of 0.75).

Download Full Size | PDF

4. Discussion

We presented a device capable of measuring pupil diameter and light reactivity through closed eyelids. This technology is based on frontal imaging of near-IR light that was transmitted through the temple, and on inducing the pupillary response through the closed eyelid using visible light. The resulting IR image pairs, taken before and after the visible light pulse, were analyzed using a generative, Gaussian-scattering model, with typical confidence intervals of 0.4mm. Hence, the spatial resolution of the pupil diameter measurement is estimated to be 0.4 mm and its temporal resolution is limited by the overall duration of the IR and visible light pulses, which is <1.5s overall.

This technology can be implemented as an automated bedside and wearable device for continuous, accurate and standardized monitoring of pupillary parameters. It combines pupil diameter and pupillary reactivity measurements, and improves upon both the temporal (4–6 hours) and spatial (∼1 mm) resolution of the currently common manual procedure. We estimate that using such a device for pupil evaluation every 5–15 minutes is expected to provide sufficient alert for medical intervention and may provide new insight on the dynamics of brain injury. Moreover, the pupil diameter ratio ${D_2}/{D_1}$ constitutes a graded, rather than a binary scale for pupil reactivity. Such a scale may be used for further study of the dynamics of the pupillary parameters, for example, by taking a series of diameter measurement and by establishing their relation with other medical parameters. Moreover, a bedside device that can modulate the intensity of wavelength of the visible light pulse and simultaneously measure both eyes, could provide further clinically relevant information on the consensual pupillary response. In conclusion, this technology may aid in saving medical staff resources and in providing earlier alert for neurologic deterioration and potential brain damage in comatose patients.