Automated Analysis of Binocular Alignment Using an Infrared Camera and Selective Wavelength Filter

Abstract

Purpose.: We present a computerized method of measuring binocular alignment using a selective wavelength filter and an infrared camera, and validate the efficacy of automated image analysis compared to the gold standard prism and alternate cover test (PCT).

Methods.: A prospective observational pilot study was performed on 30 subjects with intermittent exotropia, 30 subjects with esotropia, and 30 orthotropic subjects who were able to cooperate with the PCT. Two independent ophthalmologists examined the angle of deviation using the PCT. Full-face images were obtained with an infrared camera while a selective wavelength filter was placed in front of either eye. Images were analyzed using the 3D Strabismus Photo Analyzer. Interobserver variability, test–retest reliability and correlation between the angles of deviation were determined for both methods.

Results.: The 95% limit of agreement of interobserver variability was ±4.8 prism diopters (PD) for the PCT and ±4.3 PD for the selective wavelength filter analysis. The 95% limit of agreement of test–retest reliability between the PCT and selective wavelength filter analysis was ±8.5 PD. Results of the PCT and selective wavelength filter analysis showed a strong positive correlation (R = 0.900, P < 0.001).

Conclusions.: Infrared images acquired with a selective wavelength filter can detect the latent components of strabismus, and automated image analysis showed excellent agreement with the standard PCT. This automated method is an accurate and reliable tool for measuring ocular deviation with minimal observer dependency.

Introduction

The prism and alternate cover test (PCT) is the gold standard test for measuring binocular alignment, including the latent components of strabismus, such as intermittent strabismus, phoria, or dissociated deviation, which is not possible with the Hirschberg or Krimsky methods. There have been a few trials to interpret the Hirschberg test or Krimsky test objectively using photographic and videographic methods by estimating the displacement of the corneal light reflex to obtain the angle of ocular misalignment. ^1–10 Few attempts have been made to measure and quantify automatically the dynamic and latent properties of strabismus associated with disruption of fusion. ^11,12 Techniques that can analyze the PCT automatically require minimal operator interpretation, and this can improve the reliability of manual measurements, as well as being efficient and more applicable in clinical practice. 

We have developed previously a computerized software that automatically quantifies the angle of strabismus from photographs based on a biometric 3-dimensional (3D) eye model with good reproducibility and minimal interobserver variability. ¹⁰ This software shows excellent agreement with the Krimsky test, but cannot measure the latent component of strabismus, resulting in less correlation with the PCT. ¹⁰ To overcome this limitation, we developed an occluder made with a unique filter that blocks the subject's view and all visible light, but selectively transmits infrared light with wavelengths above 720 nm. Thus, photographs taken under infrared light with the selective wavelength filter in front of the eye visualize the details of the eye completely behind the occluder, while blocking the subject's view. These infrared images may reveal the latent components of strabismus that are manifest only after disruption of fusion. While the recently reported software can detect only manifest strabismus, the inclusion of the selective wavelength filter allows automatic quantification of these dynamic properties of strabismus. ¹⁰  

In our study, we evaluated the efficacy of automated image analysis using images acquired with a selective wavelength filter and an infrared video camera to estimate binocular alignment, and compared the results with the standard PCT. 

Methods

Participants

Consecutive subjects with exotropia, esotropia, and normal orthotropic controls were selected prospectively in this pilot study. Subjects unable to perform the PCT, with incomitant strabismus, horizontal deviations >50 prism diopters (PD), vertical deviations >5 PD, ocular disease other than strabismus, systemic disorders, and subjects with extreme biometric proportions, such as high refractive errors >6.00 diopters (D), were excluded. The angle of strabismus was measured with the PCT. Full-face images were taken with an infrared camera during the PCT with and without an occluder in front of either eye. Image analysis was performed with the 3D Strabismus Photo Analyzer (Seoul National University R&DB Foundation, Seoul, Republic of Korea). Approval to conduct this prospective study was obtained from the Institutional Review Board (IRB) of Seoul National University Bundang Hospital, and adheres to the tenets of the Declaration of Helsinki. 

Assessment of Ocular Alignment

Two independent ophthalmologists performed the PCT. A standard set of plastic prisms (Richmond Products, Inc., Albuquerque, NM) was used. Subjects fixated on accommodative targets at 1/3 m, and the examiners placed prisms before one eye and alternately occluded the eyes. The prism power was increased gradually until the direction of refixation was reversed, and then reduced until no further refixation movement was seen. The prism magnitude that neutralized the deviation, or the median value of consecutive prisms between which refixation movements reversed direction, was recorded. Prisms were not stacked or split. The average value measured by the two independent ophthalmologists was noted. 

Selective Wavelength Filter and Infrared Image Analysis

We developed an occluder with the HOYA R72 filter (Kenko Tokina Co., Ltd., Tokyo, Japan, Fig. 1A) that blocks all light with wavelengths less than 720 nm. This includes all light in the visible spectrum, so the occluder completely blocks the subject's view. The subject's head was stabilized, and the subject looked at a fixation target of red light at 1/3 m distance. Full-face infrared images were obtained with a video camera (HDR-HC3; Sony Electronics, Inc., Tokyo, Japan) at a distance of 1 m from the subject (Figs. 1B–D, Supplementary Movie S1). Average luminance of the room was set to 30 lux. The super night shot mode was selected for the activation of the infrared light illumination mounted originally in the camera to reduce the noise from the visible light. Two independent examiners performed the selective wavelength filter analysis. Infrared images were captured with the occluder in front of either eye for 2 to 3 seconds. While the subject's view still was occluded, the 2-dimensional (2D) infrared image clearly showed the limbus, pupil, and the corneal light reflex of the occluded eye, even during spectacle wearing. These images were preprocessed, registered, and analyzed automatically with the 3D Strabismus Photo Analyzer, of which the detailed mechanisms and procedures were introduced in our previous study. ¹⁰ The algorithms of the software were modified slightly from its prototype, and the angle kappa of each eye was estimated semiautomatically during fixation while the opposite eye was occluded. Measurements were taken with the selective wavelength filter in front of either eye and the mean value was noted. The results were displayed initially in degrees and converted simultaneously into PD by the following equation: PD = 100 × tan (ø°). 

Figure 1

 

(A) The selective wavelength filter made with the HOYA R72 filter. (B) Under normal luminance conditions of visible light, the filter blocks light in the visible spectrum, occluding the subject's view and concealing the image of the occluded eye. (C) Under infrared light, the subject's view still is occluded, but the detailed image of the limbus, pupil, and the corneal light reflex are revealed through the selective wavelength filter. (D) The image clarities of the limbus, pupil and the corneal light reflex are not reduced by spectacle wear.

Main Outcome Measures

The interobserver variability was calculated between the two independent examiners for the PCT and selective wavelength filter analysis. The test–retest reliability, concordance correlation coefficients, and Pearson correlation coefficients between the angles of deviation measured by different methods were analyzed. 

Statistical Analysis

Statistical analyses were performed using SPSS for Windows (Ver. 18.0, Statistical Package for the Social Sciences; SPSS, Inc., Chicago, IL) and the MedCalc statistical packages (Ver. 11.3, MedCalc Statistical Software, Ostend, Belgium). Agreement between measurements was represented in Bland-Altman plots and concordance correlation coefficients. Pearson's χ² test, likelihood ratios, and one-way ANOVA were used to compare characteristics between groups. Pearson's correlation coefficient was obtained to determine the strength of the linear relationship between each method. P values of <0.05 were considered statistically significant. 

Results

Subject Characteristics

Of 90 subjects included in this study, 30 had intermittent exotropia, 30 had esotropia, and 30 were orthotropic. There were no significant differences in age (P = 0.117 by one-way ANOVA) and sex among the subgroups (P = 0.070 by likelihood ratio, see Table). 

Table

 

The Mean Angle of Deviation Measured by Two Independent Examiners With the PCT and With the Selective Wavelength Filter Analysis

Type of Deviation	Angle of Deviation
	Exotropia, n = 30	Esotropia, n = 30	Orthotropia, n = 30	P Value
Age, mean y ± SD (range)	9.3 ± 3.6 (2–20)	5.7 ± 4.2 (1–17)	6.9 ± 10.5 (1–34)	0.117
Male (%)	15 (50%)	14 (47%)	22 (73%)	0.070
PCT, mean PD ± SD (range)	28.2 ± 9.2 (16–50)	25.8 ± 14.4 (8–50)	0
Selective wavelength filter analysis, mean PD ± SD (range)	28.1 ± 9.4 (16–50)	25.0 ± 13.6 (8–52)	0.5 ± 0.8 (0–3.0)

Interobserver Variability

Interobserver variability of each method was determined between the results of two independent examiners. The Bland-Altman plots for the PCT and selective wavelength filter analysis showed consistent variability across the graph, except for selective wavelength filter test measurements with an average angle of ≥40 PD, which showed a deviation toward a negative slope between 2 examiners. The half-width of the 95% limit of agreement was ±4.8 PD for the PCT and ±4.3 PD for the selective wavelength filter analysis (Figs. 2A, 2B). There was no overall tendency for the values of one examiner to be higher or lower than the values of the other. 

Figure 2

 

The Bland-Altman plots for the PCTs and selective wavelength filter tests. The Bland-Altman plots showed consistent variability across the graph, except for selective wavelength filter test measurements with an average angle of ≥40 PD, which showed a deviation toward a negative slope between 2 examiners. The half-width of the 95% limit of agreement was ±4.8 PD for the PCT (A) and ±4.3 PD for the selective wavelength filter analysis (B). There was no overall tendency for the values of one examiner to be higher or lower than the values of the other.

Test–Retest Reliability Between Methods of Angular Measurement

Test–retest reliabilities between each pair of methods represented on the Bland-Altman plots showed consistent variability across the graph for both tests. The half-width of the 95% limit of agreement was ±8.5 PD for the PCT versus selective wavelength filter analysis. There was no overall tendency for any specific test result to be higher or lower than another test value. Concordance correlation coefficients between each pair of methods were 0.96 (95% confidence interval 0.94–0.97) for the PCT versus selective wavelength filter analysis. 

There was no significant difference between the angles measured with the filter in front of the right eye and the left eye, both in exotropia (P = 0.620) and esotropia (P = 0.082, by paired t-test). Regarding the laterality of fixation dominance, 19 patients (63.3%) with exotropia and 18 patients (60%) with esotropia had a dominantly fixating eye. Similarly, there was no significant difference in the angles measured with the filter in front of the dominant and nondominant eyes, both in exotropia (P = 0.650) and esotropia (P = 0.442). 

Correlations Between Methods of Angular Measurement

The PCT and the selective wavelength filter analysis showed a strong positive correlation (R = 0.900, P < 0.001), and there was no significant difference according to the type of deviation (exotropia R = 0.900, P < 0.001; esotropia R = 0.904, P < 0.001; Figs. 3A, 3B). 

Figure 3

 

Scatter plots and correlation between the selective wavelength filter analysis, and the PCT. The selective wavelength filter analysis showed an excellent correlation with the PCT in both exotropia ([A], R = 0.900, P < 0.001) and esotropia ([B], R = 0.904, P < 0.001).

Discussion

In our study, we developed an automated method to quantify binocular alignment, including the latent components of strabismus, using an infrared camera and a selective wavelength filter. Image analysis was performed with the 3D Strabismus Photo Analyzer, and the results showed excellent agreement with the PCT, which is the most common test for measuring binocular alignment. The PCT is performed manually and is measured subjectively by individual examiners, thus the reliability of those measurements depends on the proficiency of examiners and measurement errors. ^13–15 Even the most experienced ophthalmologists applying the PCT have shown inconsistent responses up to 6 to 12 PD. ¹³ In addition, this variability also depends on patient factors, such as alertness and anxiety level at the time of examination, which can vary throughout the day, even within minutes. ¹⁶ These factors all lead to the inherent variability of clinical measurements and an isolated assessment of the magnitude of deviation can vary in a single patient. ¹⁷ Our automated method showed excellent agreement with the standard PCT, and test–retest reliability and interobserver variability also were comparable. Therefore, this method can be considered an accurate and reliable tool for measuring ocular deviation, which does not rely on the ability of the examiner. 

The major benefits of using infrared images and a selective wavelength filter are as follows. Firstly, this method allows measuring the latent component of strabismus that only manifests after disruption of fusion, such as intermittent strabismus or dissociated deviation. The selective wavelength filter in front of the eye occludes the patient's view completely, and at the same time the 2D infrared photograph clearly shows the dynamic movements and details of the eye behind the occluder. Ocular deviations together with the details of the limbus, pupil, and the corneal light reflex behind the selective wavelength filter are visible with an infrared camera. Schiavi et al. also performed automated measurement of the angle of strabismus after dissociating the two eyes with an occluder using an infrared camera and image analyzer. ¹⁸ However, the position of the eye behind the occluder was obscured and latent components of strabismus could not be measured, which is the major difference from our method. ¹⁸ Secondly, the resolution and quality of the infrared images were not much affected by spectacle wear, and this is another major improvement from the previous version, which could not measure subjects wearing spectacles. ¹⁰ Finally, this method is simple and noninvasive. It requires minimal interpretation and can be implemented easily in normal clinical practice compared to eye tracking methods, such as scleral search coils or spectacle-mounted liquid crystal shutters. ^11,12 It only takes 2 or 3 seconds to acquire a measurement, which is an advantage over the PCT, especially in children with limited cooperation. Every subject was tested successfully with this method except for children under 3 years of age, who could not tolerate the few seconds with an occluder placed in front of their eyes. Although not all children under 3 years of age were testable, we could acquire measurements successfully in a substantial number of subjects who did not cooperate with the PCT. An infrared video camera and selective filter also are inexpensive, and with the help of the image analysis software, this method can be applied in clinical practice. 

The Bland-Altman plots for the selective filter test did not show consistent variability across the graph. Particularly in patients with an angle of ≥40 PD, there was a deviation toward a negative slope between 2 examiners. However, all values in this range lay within the 95% limits of agreement between 2 measurements (Fig. 2B). As for the outliers, there were 3 outliers in the difference between PCT measurements and 1 outlier in the difference between selective filter tests, of whom all patients were esotropic. Regarding the previous report on the interobserver variability of the PCT in esotropic patients, the outliers (7.0, 8.5, 13 PD) lay at the boundaries of the 95% limits of agreement on a difference between 2 measurements; ±4.7 PD for angles of 10 to 20 PD and ±11.7 PD for angles greater than 20 PD, comparably. ¹³ In contrast, the only outlier in the difference between selective filter tests was a patient measured with the smallest angle of deviation (8 PD by the PCT). We presume that slightly asymmetric kappa angles between both eyes may be the cause of measurement errors by the 3D Strabismus Photo Analyzer in such small angles of deviation. ¹⁰  

Linear regression slopes for exotropia (B = 0.993) and esotropia (B = 0.853) were different (Figs. 3A, 3B). The smaller value for esotropia can be interpreted that in esotropic patients, the absolute measurements with the selective filter and software were relatively smaller than the measurements with the PCT. Although the reason is not clear, it may be partly explained by the characteristics of the fixation target during measurements. During the PCT, subjects fixated on small accommodative objects at 1/3 meter. In contrast, small light targets were used during the selective filter analysis, because the luminance of the room was darkened. Light is sufficiently interesting to attract the patient's attention, but accommodation of the eye has little practical effect in changing the nonfocused condition of the stimulus on the retina, resulting in relaxation of accommodation. ¹⁹ Thus, light may induce less accommodation compared to small objects, resulting in reduction of accommodative convergence and the angle of esotropia. Conversely, for exotropia patients, less accommodation may induce a larger angle of exodeviation with the selective filter. Therefore, the difference in accommodative convergence induced by fixation targets may account for the difference between esotropia and exotropia patients. 

One of the major limitations of this method is that the software is based on normative ophthalmic biometry. ¹⁰ Therefore, subjects who have extreme proportions falling out of the normal variation, such as high refractive errors, nanophthalmos, or pathologic myopia, cannot be examined using this program. In addition, the results are less accurate when it comes to large ranges of angular deviations of 50 PD or more, because of the exponential function during conversion between degrees and prism diopters. This method cannot be applied to patients with nystagmus, since the average position of the eye and corneal light reflex during oscillation cannot be determined with the current software. However, crude measurements are possible for now, with randomly captured images during oscillation. Torsional components also are not measurable, since the 3D Strabismus Photo Analyzer measures the position of the eye in horizontal and vertical vectors. Only the magnitude of horizontal and vertical strabismus can be measured separately. 

In conclusion, infrared images acquired with a selective wavelength filter can be used to measure binocular alignment objectively with the 3D Strabismus Photo Analyzer. This is an accurate and reliable tool for measuring ocular deviation, and shows excellent agreement with the standard PCT. 

Supplementary Materials

mov S1 

Acknowledgments

Supported by grants from the Korea Healthcare Industry Development Institute, R&D Project, Ministry for Health, Welfare & Family Affairs, Republic of Korea (A110793). 

Disclosure: H.K. Yang, None; J.-M. Seo, None; J.-M. Hwang, None; K.G. Kim, None