Limbus Versus Pupil Center for Ocular Alignment Measurement With Corneal Reflexes

Jean-Cyriaque Barry and Andreas Baches

Purpose. To investigate the accuracy of ocular misalignment measurement, using corneal reflections. Methods. Corneal reflex positions were measured relative to two landmarks, the limbus center and the entrance pupil center, using high-resolution digital images for cyclopean gaze angles from 0 to 18.8 (34.04 prism diopters; PD) to the right and to the left in 10 subjects. Distance h from the center of the corneal curvature to each landmark was determined from linear regressions and showed significant differences between both conditions: mean Anmbus was 5.243 mm, and mean /jpupi| was 4.884 mm. From these, limbus- and entrance-pupil-center-related Hirschberg ratios were determined as lT/mm (19.43 PD/mm) and 11.82/mm (20.92 PD/ mm), respectively; and ocular alignment was calculated for both conditions. Simulated angles of strabismus were calculated as the binocular difference between ocular alignment of the right and left eyes (condition 1), and as the monocular difference between the ocular alignment of right eyes and left eyes separately (condition 2). Results. Condition 1: Errors in simulated angles of strabismus were approximately twice as large for entrance pupil center compared with those in limbus-center-based evaluation; in the primary position, the 95% pupil-related confidence interval of the binocular difference was 5.217 (9.1 PD), compared with 3.174 (5.5 PD) for the limbus-related option. Condition 2: Errors were approximately equal. Conclusions. The entrance pupil center is a less reliable landmark than is the limbus center for measuring ocular alignment by using corneal reflections, because of unreliable positions of the entrance pupil center; different mean Hirschberg ratios should be used in both conditions. Invest Ophthalmol Vis Sci. 1997; 38:2597-2607.

A . number of corneal reflex techniques for measuring ocular alignment have been suggested and analyzed.1"21 However, the relative merits of these techniques have, to our knowledge, never been studied experimentally by direct comparison with the same alignment data. Most techniques may, in principle, be derived from equation 1 5 1 9 by adequate substitutions or approximations: (f> = arcsin (x/h) where (1)

From the Department of Ophthalmology, RWTH Aachen University, Germany. Supported in part by DFG (Ef 6/3-2 and Ba 1316/5-3). Corporate support received from Chiron Adatomed Intraocular Lenses, National Instruments, and XOn Software, Munich; Management fntelligenter Technologien, Aachen Kornelimilnster; Boston Polymer Technologies, Weinheim; and eltec, Mainz, is: Germany; Graftek, Mirmande, France; and pro cornea bv, Erbeek, The Netherlands. Submitted for publication November 25, 1996; revised February 12, 1997; accepted June 16, 1997. Proprietary interest category: N. where Reprint requests: J-C Barry, Department of Ophthalmology, Medical Faculty of HR. Eberhardt Karls University, Schleichstr. 12-16, D-72076 Tubingen, Germany.

(f>: rotation angle of the corneal axis relative to the light source, x: displacement of corneal reflex relative to an axially located ocular landmark h: distance of the axially located landmark to the center of the corne; curvature, as illustrated in Figure 1. Equation 1 results from a paraxial approximation, given a cyclopean eye with a spherical cornea. It holds best for eye rotation angles that are not large (less than 20) and in conditions in which the light source is not close to the cornea that is, for distances larger than one third of a meter.19 A simplified relationship uses the Hirschberg ratio (HR) in a linear approximation of equation 1. That
77 = HR*x

(2)

77: is the eye rotation angle calculated with the

Investigative Ophthalmology & Visual Science, November 1997, Vol. 38, No. 12 Copyright Association for Research in Vision and Ophthalmology

2597

2598

Investigative Ophthalmology & Visual Science, November 1997, Vol. 38, No. 12

varies with ocular components and with the landmark chosen. Influence of various parameters on the HR was reference examined previously.1213151819 A dependence of HR on the corneal radius may be predicted from Figure 1, in that the distance h depends on the position of the center of curvature of the anterior corneal surface, and hence on its corneal radius r. Because of peripheral flattening of the cornea, the values for the corneal radius rand the distance h increase; consequently, the HR decreases in the periphery.21 light source There is widespread agreement that, within the limits of measurement accuracy, a single HR of 21 to 22 PD/mm may suffice3'4'6'12"16'18'19 to measure ocular alignment for clinical and scientific purposes, regardless of individual factors. However, large variations of FIGURE l. Sketch of how equation 1 (see Methods) can be the individual HR of the order of up to 100% relative derived in the paraxial approximation. A small light source to the mean HR have to be expected.5121318 positioned sufficiently far away is imaged by the cornea (apIn securing higher accuracy, little attention has proximated as a spherical surface) as a virtual image, the been paid to the role of h or, in other words, to the anterior (Purkinje I) corneal reflex. It is situated on the axis proper choice of the reference plane and landmark. connecting the light source and the center of the curvature of the cornea, hitting the cornea normally. The corneal reflex Most often, the corneal reflex position was measured relative to the entrance pupil center. The entrance displacement x is the projection of the separation between pupil is the virtual, enlarged image of the physical the central landmark and the corneal reflex onto the plane pupil, which appears a little closer to the observer normal to the observer's imaging axis, which is assumed to than the physical pupil.23 coincide with the light source direction. The distance x depends on the eye rotation angle 4> and on the axial position For a quick, clinical estimation of corneal reflex of the chosen landmark relative to the center of curvature; position (for example, with the Hirschberg test), the that is, on the distance, h, as stated in equation 1. entrance pupil center is probably the only viable landmark relative to which the corneal reflex asymmetry between both eyes may be judged, in that there is Equation 2 results from a series expansion22 of usually a bright, easily observable contrast between the arcsin(x) and truncation after the linear term, and entrance pupil and the corneal reflex. The entrance from the approximation that the cyclopean gaze angle pupil center is also used as a landmark for ocular P equals the true eye rotation angle of a real eye that alignment measurement in recently developed digital is decentered by half the interpupillary distance. Both imaging devices for refractive screening.24"26 approximations have adverse effects: The approximation arcsin(x) x means that the calculated eye rotaThe limbus margins27 or the center of the limbus tion angle will be underestimated. The cyclopean eye were used as alternative landmarks, for instance, beapproximation results in an overestimation of the calcause the margins of the entrance pupil were more culated eye rotation angle, and because both approxidifficult to locate in the presence of a dark iris than mations are of the same order (for instance, ~2% for was the limbus with visible light.3'6 This was combined P = 20 [36.4 PD]), they compensate each other, with computation of eye alignment from ratios of disexplaining to some extent the good results obtained tancesfor instance, the ratio of corneal reflex diswith this approximation. placement to the width of the limbus.17'20 No calibraThe HR can be gained empirically (HR^pj,.) from tion for measurement of distances in images would linear fits of gaze angle P versus measured corneal then be necessary; consequently, the examination disdisplacement x It may also be calculated (HRcaic) from tance did not have to be controlled. equation 1 by setting x = 1 mm for any appropriate The movement of the limbus or of its center with chosen or given distance h and dividing the resulting eye rotation is readily described through the movevalue </>HR by 1 mm.5 ment of the limbus in the normal plane of the observThe determining parameter for the theoretically er's direction of gaze as h sine of the eye rotation derived HRcaic is the distance h that is the separation angle. The entrance pupil center, however, may unbetween the center of anterior corneal curvature and dergo a prismatic effect with eye rotation, because of the reference plane in which the landmark lies, reits refraction through the cornea. Hence, its motion garding which the decentration x of the corneal reflex in the normal, plane of the observer's imaging axis may be specified (Fig. 1). This implies that the HR would be more difficult to predict.

Accuracy of Ocular Alignment Measurement

cnnwms

2599 fixation target was used to generate a bright corneal reflex in both eyes. Two additional units, positioned 8.8 cm to the left and right of the central light source, were used to generate additional Purkinje images for further studies. For each gaze direction, three pictures were taken and stored. The horizontal positions of Lhe corneal reflex were measured in 10 young, orthotropic subjects for all gaze directions from 0 (0 PD) to 18.8 to the right and to the left (Fig. 3). Informed consent was obtained from all subjects, and the studies were conducted in accordance with the tenets of the Declaration of Helsinki. The video signal was digitized (512X512 pixels, 8bit gray level depth) with two frame grabbers (Matrox PIP 1024, Dorval, Canada) and was stored in a desktop computer. The digitized images were transferred to a digital image workstation and were evaluated interactively with NIH Image 1.58 software (National Institutes of Health, Bethesda, MD). The imaging system was calibrated with a millimeter grid, and the pixelto-millimeter ratio was determined as 21.5 pixels/mm,
left pupil margin right pupil margin

A
pmr of *ye wilhPurlunje 1 reflex f

FIGURE 2. Design of apparatus used to record corneal reflex positions in different gaze directions up to 18.8 of cyclopean gaze angle /? to the right (positivesign) and left (negative sign). The distance of the central fixation target to the subject's pupillary plane was adjusted to 50 cm.

Therefore, the use of the limbus center as a landmark might offer advantages over the use of the entrance pupil centermore so, because the latter is known to vary significantly between subjects28 and is prone to unpredictable asymmetric shifts with pupil constriction,29 occurring with changes of accommodation, vergence, ambient light levels, or neurovegetative ton us. Therefore, the object of this study was to examine to what extent the choice of the limbus or entrance pupil center as a landmark influences the experimentally found HR, and which of the two choices leads to a more accurate measurement of the ocular alignment. METHODS Experimental In the design of the experiments, care was taken to minimize nonsystematic errors for analysis of systematic differences between different measurement conditions and evaluation procedures. Therefore, a head and chin rest, similar to that used for slit lamp examinations, was used to control head position and distance to the recording apparatus. Subjects were instructed to keep their heads still throughout the measurements in all gaze directions. However, the head was not completely immobilized (for instance, with a bite bar or a strap). All values measured or calculated in degrees () were converted into prism diopters (PD) by the equation PD = 100 tan. The apparatus (Fig. 2) consisted of a horizontal fixation bar located 50 cm from the eyes. Fixation targets were positioned at cyclopean gaze angles of 18.8 (34.04 PD), 10 (17.63 PD), 5 (8.75 PD), and 0 (0 PD). Two charge-coupled device cameras (SDT 3440, Seitner Datentechnik, Seefeld, Germany), one for each eye, equipped with extension tubes and lenses, were used for imaging the anterior eye segments at a 50-cm distance. A photo flash unit (Tumax A16) mounted 10 cm vertically under the central 0

left limbus

right limbus

FIGURE 3. Horizontal positions P of ocular structures involved in the measurement of the horizontal corneal reflex decentrations relative to the limbus center (xiimbus) and to the entrance pupil center (*pUpj|). The horizontal entrance pupil center position (centerpiipi|) was calculated as [P(right pupil margin) + P(left pupil margin)]/2. The horizontal limbus center position (centeriimb) was calculated as [P(right limbus) + P(left limbus) ]/2. Consequently, in the right eye, *puPii = P(corneal reflex) center pil|jill and x\;in\JUS = P(corneal reflex) center| imblis . In the left eye, ,\pUpj| = [P(corneal reflex) centerpupil], and x\imbUi = [P(corneal reflex) centeriimbus,}. These definitions were chosen to have a positive sign of x for a nasal shift of the corneal reflex, consistent with a corresponding positive angle K, and vice versa for a temporal shift.

2600

Investigative Ophthalmology & Visual Science, November 1997, Vol. 38, No. 12

with no significant bias between central and peripheral portions of the image. The examiner took the x,y coordinates of the lateral limbus margins, the entrance pupil margins, and the corneal reflex in each eye. As illustrated in Figure 3, the quantities calculated from this data were the horizontal limbus diameter (LD), the positions of the limbus center and the entrance pupil center, and the horizontal distance of the corneal reflex relative to the entrance pupil center (xpupil) and the limbus center (limbus) A nasal decentration x carried a positive sign, and a temporal decentration x carried a negative sign. Three examiners participated in the evaluation of the image data. There were six subjects with light irides and four subjects with dark irides. On average, the localization of the horizontal position of the nasal and temporal limbus appeared subjectively a little more difficult than the localization of the nasal and temporal entrance pupil margins, because of the presence of a gray-level transition zone from the iris to the sclera that represented the position of the limbus, whereas there was a sharp contrast between entrance pupil and iris, unless the subject had a dark iris. Error levels associated with image evaluation were assessed in preliminary trials, and the typical discrepancy in the measurement of the corneal reflex position between examiners and between reevaluations of the same image by the same examiner was 0 or 1 pixels. The overall reproducibility of the measurement of the distance between ocular landmark and corneal reflex in the study with all 10 subjects was assessed in the following way: In the binocular pictures taken in the primary position, the mean and standard deviation of the difference of measured corneal reflex position x, within each set of three images taken of one eye in one gaze direction, was calculated in all 10 subjects for entrance pupil center and limbus center, separately. The standard deviation, SD(x), of the corneal reflex position x in a set of three images served as an indicator of its reproducibility in each subject. Therefore, the 20 standard deviations, SD(x), obtained for all right and left eyes and each landmark (entrance pupil-limbus center), were averaged and their standard deviations calculated: mean of SD(xHm. blls) = 0.048 mm with a SD of SD(xlimbus) = 0.031 mm, and mean of SD(Xpupii) = 0.044 mm with a SD of SD(Xplipii) = 0.031 mm. There was no statistically significant difference between the mean values of pupiland limbus-based measurement of the corneal reflex position. Therefore, the reproducibility of corneal reflex position was the same for pupil-based and limbusbased measurement. The slightly smaller value of the mean of SD(xpupji), compared with SD(xumbus) is in keeping with the examiners' subjective impressions that the margins of the entrance pupils were easier to

locate than the somewhat less well defined limbi: In general, entrance pupil margins had a sharp dark contrast with the iris, whereas the limbus could be described as a pixel region with a transition from one gray level to another. In the latter instance, the examiner had to delimit the middle of the transition zone to mark the limbus position, which appeared less well denned than the entrance pupil margins. The same checks were executed for secondary gaze positions and for monocular images; and again, no significant difference between pupil- and limbusbased measurements were observed in the reproducibility of the corneal reflex position. The 0.05-mm magnitude of the mean reproducibility roughly equaled the pixel size; therefore, it can be concluded that the determination of the corneal reflex position was limited by the minimum pixel (1pixel) error in this setup, for either condition. Bear in mind that measurement of the corneal reflex position involved three measurements, each of which could be affected by, for instance, variations of fixation, and by the pixel error (error inherent to digitized images) that could accompany the determination of the nasal and temporal entrance pupil margins or of the nasal and temporal limbus and of the corneal reflex position. This could have entailed the maximum pixel error (3 pixels), but the current data indicate that the reproducibility was far better than that. In the workup of the patients, the keratometer readings and their objective autorefractions (RK-1; Canon Europe, Amstelveen, The Netherlands) were taken. The keratometer readings were converted to give the corneal curvature in the horizontal meridian. Most subjects were moderately myopic (mean spherical equivalent [SE] approximately 2 diopters [D]; minimum SE -5.6 D, maximum SE +1.7 D). The largest measured astigmatic component was 2.5 D. Data Evaluation In the first step of the data analysis, the measured displacements x were plotted against the sine of the gaze angle (5 to determine the experimental value of /*and the offset xo(fi = 0) compared with the relationship x = h'sin(P) + xo (derived from equation 1). This was done for each subject, for the entrance pupil center and the limbus center reference, and in both eyes separately. The values of h in the 10 subjects for each reference center were plotted against the measured corneal radius in the horizontal meridian for right eyes and left eyes separately (Figs. 4A, 4B, 5A, 5B). Mean corneal radius in the right eyes (r = 8.06 0.26 mm [SD]) and in the left eyes (r = 8.07 0.30 mm) were slighdy above the average adult value of r = 7.8 mm, in keeping with the slight predominance of myopic subjects with larger globes participating in the study. From this data, a mean hpUpi\ = 4.884 mm

Accuracy of Ocular Alignment Measurement

2601 and hpUp\\ (r = 7.8 mm) = 4.854 mm, with HRpupii (r = 7.8 mm) = 11.89%nm (21.05 PD/mm), with both HRs falling in the range of reported HRs. 612 " 151819 The calculated eye rotation angles ?7REpupii and r)Xs. pupii (where RE = right eye and LE = left eye) were computed, using HRpupil mean, and equation 2. In the same manner, ?7RE Umbus and ?7LE umbus were computed using HRiimbus mean. Two ways of interpreting eye alignment data may be used clinically to measure ocular misalignment. These two options were simulated 30 with our data. Condition 1, binocular approach: The angle of strabismus was obtained by subtracting the value of 7 (/?) 7 for any gaze direction ft from the value 7 (/5 = 0) of 7 the fellow eye in the primary position. This should be a good estimate of the manifest ocular misalignment under binocular viewing conditions. However, differences in the offsets XQ of the corneal reflex in the primary position in the right and left eye could lead right eyes, pupil center

right eyes, limbus center

CO

X)

A
5.8

7.4

7.6

7.8

8.2

8.4

8.6

corneal radius r / mm left eyes, limbus center

7.6

7.8

8.2

corneal radius r / mm
FIGURE 4. Plots of the distance humbu:i (obtained from linear fits of measured Xiimbus versus sin[/5]) against the horizontal corneal radius r (A) in the right eye, and (B) in the left eye, for the limbus center condition.

7.4

7.6

7.8

8.2

8.4

8.6

corneal radius r / mm left eyes, pupil center

(right eye: 4.840 0.320 mm; left eye: 4.927 0.347 mm) and a mean hyimhus = 5.243 mm (right eye: 5.233 0.290 mm; left eye: 5.253 0.308 mm) were derived, and linear fits were obtained, one for right and one for left eyes. The HRs were calculated as (180/TT)'arcsin (1 mtn/A)/l mm, as explained above. The differences of the mean HRpupii and HR|imbus were found to be statistically significant for right eyes and left eyes, separately. After averaging between right and left eyes, we found HR p u p i l mean = 11.82/mm (20.92 PD/mm), and HR, imbus mean = l T / m m (19.43 PD/mm). These HRs were used in the rest of the study to calculate 7 7 from observed corneal reflex positions. For comparison, the linear fits of Figures 4A, 4B, 5A, and 5B were used to calculate the HR for the average corneal radius in adults, r = 7.8 mm. This resulted in hVimbus (r = 7.8 mm) = 5.165 mm, and HR,imbus ( r = 7.8 mm) = 11.16/mm (19.73 PD/mm),

7.4

7.6

7.8

8.2

8.4

8.6

corneal radius r / mm
FIGURE 5. Plots of die distance hpupu (obtained from linear fits of measured Xplipii versus sin[/?]) against the horizontal corneal radius r (A) in the right eye, and (B) in the left eye, for the entrance pupil center condition.

2602

Investigative Ophthalmology & Visual Science, November 1997, Vol. 38, No. 12 condition [\) - iimbus, ngnt e y e s

to errors in this approach. The differences nocC/3) = !7REG0) " ViAP = 0), Ai7LEbinoc(/3) = - rjRZ(P = 0), with 0 = 5 (8.75 PD), 10 (17.63 PD), and 18.8 (34,04 PD) were used to simulate angles of strabismus of the same size. They were computed in the right and left eyes of the 10 subjects, together with their mean values and SDs. These followed a normal distribution, and the confidence intervals for the measured angle of strabismus were calculated as twice the standard deviation. Condition 2, monocular approach: The angle of strabismus was obtained by subtracting the value of rj(P) for any gaze angle 0 from the value r)(0 = 0) of the same eye in the primary position. This equals, or at least comes close to, the orthoptic simultaneous prism and cover test procedure.31 The differences, AT)nKMIOC(0) = v(P) ~ l(/3 = 0), with 0 = 5 (8.75 PD), 10 (17.63 PD), and 18.8 (34.04 PD), should equal angles of strabismus of the same size. The advantage of this approach is that the offset #o of the corneal reflex in the primary position is automatically eliminated that is, the angle K32 or any other individual offset does not bias the measurement. The simulated angles of strabismus A?7monoc(/3) were computed in the right and left eyes of the 10 subjects with their mean values and SDs. Using the same procedure as for condition 1, assessment was made to determine whether the simulated angles of strabismus could be treated as normally distributed. The data were found to be normally distributed for bodi conditions, and the confidence intervals for the differences between calculated and expected angles of strabismus were calculated as the mean difference plus twice the SDs33 to assess the error levels. Finally, to illustrate the variability of the position of the entrance pupil center relative to the limbal center, the horizontal distances of the entrance pupil center to the limbus center, centerpupil-centeriil1lbus, were plotted against the sine of the gaze angle /3 for all 10 subjects in the right and left eyes. RESULTS Four twin box plots (Figs. 6A, 6B 7A, 7B, 8A, 8B, 9A, 9B) allow for a comparison of accuracy of misalignment measurement with entrance pupil and limbus center reference. In the box plots, the ordinate indicates the difference between measured and expected angle of strabismus: Boxes show 25% and 75% confidence intervals, and T bars indicate 95% confidence intervals; data points outside the 95% confidence interval are shown separately. The boxes represent the seven simulated angles of strabismus, from right (-18.8, -34.04 PD) to left (+18.8, +34.04 PD). Note that the central 0 box refers to the differences of measured angles in the primary position, below which

condition (1) - limbus, left eyes

FIGURES 6 TO 9. Box plots (for Figures 6 to 9) of differences between calculated and expected angles of strabismus for binocular condition 1 and monocular condition 2. Central bar indicates median, upper and lower box borders refer to 25% to 75% confidence interval limits, and T-bars refer to 95% confidence intervals. Data points outside these limits are shown as individual points. FIGURE 6, Condition 1: box plots of the difference between calculated eye rotation angles i7iilllbus and 77iiinblis (/? = 0) of the fellow eye, (A) right eye misaligned and (B) left eye misaligned, for six simulated angles of strabismus calculated from limbus center reference data.

a manifest strabismus may not be distinguished from alignment data occurring in orthotropic subjects. In other words, 0 data delimit the threshold for detection of a manifest misalignment in the primary position.30'34'35 Condition 1 simulated binocular angles of strabismus: Figures 6A and 6B show box plots of the difference between expected and simulated angles of strabismus (limbus center), for (A) right eyes and for (B) left eyes. Figures 7A and 7B show the box plots of the difference between expected and simulated angles of strabismus (entrance pupil center), for (A) right eyes and for (B) left eyes. In the primary position, the maximum 95% confidence intervals for right and left eyes were limbus, 3.174 (5.5 PD), and entrance pupil, 5.217 (9.1 PD). Clearly, errors were smaller for the limbus option, for all simulated angles of strabismus. The 95% confidence levels at 18.8 of simulated angle of strabismus were, at most, 26.4% (limbus) and 37.0% (entrance pupil) of the amount of the simulated angles of strabismus.

Accuracy of Ocular Alignment Measurement condition (1) - pupil, right eyes

2603

condition (1) - pupil, left eyes

B
FIGURE 7. Condition ]: box plots of the difference between calculated eye rotation angles 77pupi| and 77piipi| {0 = 0) of the fellow eye, (A) right eye misaligned and (B) left eye misaligned, for six simulated angles of strabismus calculated from entrance pupil center reference data.

trance pupil and limbus center positions, plotted against the sine of the gaze angle, varied for different gaze angles. In most subjects, the differences decreased when looking to the left (negative sines). For some of the plots, satisfactory linear regressions could be obtained. At approximately an 11 (19.43 PD) gaze angle to the left, entrance pupil and limbus centers of the right eyes coincided, on average, within a range of ~0.25 mm. The average distance between entrance pupil center and limbus center in the primary position is indicated by the differences at sin(0): On average, the entrance pupil centers were decentered ~0.18 to 0.23 mm nasally, in keeping with independent data.28 The difference between both increased with a more temporal position of the eye, because the projected distances between the entrance pupil and limbus centers appeared larger when viewed obliquely, in that the entrance pupil plane is slightly behind the limbus plane (Figs. 4, 5). The data for left eyes (not shown in Fig. 10) suggest similar conclusions. Note, however, that in one subject (top plot, black triangles), the difference was nearly constant and linearity was limited in some of the plots. This could be because of the entrance pupil movement; the limbus movement should be linear with the sine of the gaze angle, in keeping with the eye model in Figure 1. However, it could be speculated that in some
condition (2) - limbus, right eyes

Condition 2 simulated monocular angles of strabismus: Figures 8A and 8B show box plots of the difference between expected and simulated angles of strabismus (Hrnbus center), for (A) right eyes and for (B) left eyes. Figures 9A and 9B show the box plots of the difference between expected and simulated angles of strabismus (entrance pupil center), for (A) right eyes and for (B) left eyes. In the primary position, the maximum 95% confidence intervals for right and left eyes were limbus, 1.07 (1.9 PD), and entrance pupil, 1.3 (2.3 PD). Errors were slightly smaller for the limbus option for all simulated angles of strabismus. The 95% confidence levels were, at most, 19.3% (limbus) and 21.9% (entrance pupil) of the amount of the simulated angles of strabismus. It could be speculated that with a higher camera resolution, systematic differences between pupil- and limbus-based measurement reproducibility could have been noted, if they were present. However, such error levels would, on average, have had to be smaller than or equal to at least 1 pixel. Consequendy, any systematic difference between pupil-and limbus-based mean corneal reflex distances that was markedly larger than 1 pixel must be caused by genuine differences between these conditions, in that the image evaluation was done in the same images. Figure 10 (right eyes only) reveals that the en-

condition (2) - limbus, left eyes

FIGURE 8. Condition 2: box plots of the difference between calculated eye rotation angles rj|imbus and 77timl)lls (0 = 0) of the same eye, (A) right eye misaligned and (B) left eye misaligned, for six simulated angles of strabismus calculated from limbus center reference data.

2604 condition (2) - pupil, right eyes

Investigative Ophthalmology 8c Visual Science, November 1997, Vol. 38, No. 12

condition (2) - pupil, left eyes

FIGURE 9. Condition 2: box plots of the difference between calculated eye rotation angles i7pupil and J7pupi, (/3 = 0) of the same eye, (A) right eye misaligned and (B) left eye misaligned, for six simulated angles of strabismus calculated from entrance pupil center reference data.

ter is a much less reliable landmark than the limbus center, when used to measure the relative position of the corneal reflex. Unpredictable shifts of the entrance pupil center29 and variable anterior chamber depths3 contribute to this unsystematic error. To examine whether varying pupil size could have been the reason for the lower precision of eye alignment measurement, horizontal entrance pupil and limbus diameter variations were analyzed in the same way as the reproducibility of corneal reflex position (see Methods): For each eye, the mean diameter of the limbus and of the entrance pupil, and its standard deviation, SD (diameter), were determined from a triplet of images in the primary position. Then the mean SD (diameter) and standard deviation of SD (diameter) of the 20 individual standard deviations of the diameters were computed. Mean standard deviations indicated a significant difference (P < 0.0001, paired Student's Rest) between entrance pupil and limbus diameter variation: Mean SD(pupil diameter) = 0.456 mm, with a standard deviation of SD (pupil diameter) = 0.154 mm, mean SD (limbus diameter) = 0.132 mm, with a standard deviation of SD (pupil diameter) = 0.137 mm. Note that the mean SD(pupil diameter) was much larger than the pixel error. This indicated that the reproducibility of diameter measurements was equal for the entrance pupil and for the limbus.

of the cases, the limbus itself was seen through the peripheral portions of the cornea, and hence could be treated as an entrance limbus itself. The movement of its center would then be more difficult to predict. DISCUSSION Brodie et al3'6 favored the use of the limbus as a landmark, partly for practical reasons, pardy because they suspected that this choice would reduce error levels caused by varying depths of the anterior chamber.3 In the same study, a nominal value for h was computed, but an experimental dependence on the corneal radius was not established, and a single HR was advocated regardless of pupil- or limbus-based measurement, confirming previous similar reasoning.19 In the eye model shown in Figure 1, the difference between the entrance pupil center and the limbus center as a landmark and the resulting behavior may be explained readily: The entrance pupil center is, on average, ~0.4 mm axially behind the limbus center, and is nasally shifted by approximately 0.25 mm. This can be adjusted for by using appropriate HRs. However, there is some variation between right and left eyes, and the scatter of the difference is sometimes large. This suggests that the entrance pupil cen-

-.3 -.3 -.2

-.1

.1

sin (gaze angle) FIGURE 10. Plot of the horizontal distance between entrance pupil center and limbus center, centerpiipii centerijlllbll5! versus sin(/?) in the right eyes. In this plot, a linear behavior points to a constant ratio of hpUpn to hy,mhm. The slopes of the graphs are proportional to individual ratios of h. The differences change approximately linearly with sin(/3), but the positions and slopes of the curves vary significantly between individuals, and in some instances, nonlinear behavior is evident.

Accuracy of Ocular Alignment Measurement The variability of the entrance pupil diameter, however, was threefold, compared with the limbus diameter variability and compared with the measurement reproducibility of the diameters. Similar results were obtained for the other gaze positions. This means that the most probable explanation for the difference of precision between limbus and entrance pupil measurement of eye alignment in binocular images is given by the larger variations of entrance pupil size, and hence by the associated shifts of the entrance pupil center without an ocular rotation. By contrast, at least in our model (Fig. 1), the limbus center cannot shift in a similar way, because the limbus is a rigid structure. However, the simple anatomic model shown in Figure 1 does not represent all of the real effects that could be expected: Coincidentally, the effects of neglected corneal asphericity and the linearization of the sine in equation 1 compensate each other to some extent, such that the validity of equation 2 is extended to larger angles. Exact ray tracing analysis with software introduced elsewhere38 and modified to model aspheric corneas, showed that when a spherical shape was taken into account, the limbus must be assumed to be a structure imaged through the peripheral portions of the cornea, lying ~0.5 mm behind the corneal surface. This means that the visible limbus is an "entrance limbus" analogous to the entrance pupil. The study's maximum error in eye alignment measurement is tangible in the binocular approach (Figs. 6A, 6B, 7A, 7B): Here, the errors of entrance pupil data became approximately twice as large as those of limbus data. This is of considerable clinical relevance, because most measurements of eye alignment are recorded relative to the entrance pupil center as binocular difference of ocular alignment between right and left eyes, in that this does not require pictures with one eye covered. In the monocular approach (Figs. 8A, 8B, 9A, 9B), any offset in terms of entrance pupil center shift, angle K or other parameter, whatever its cause, was cancelled, because only the difference of the angle r) between two gaze directions was computed. This reduced the error levels such that differences between entrance pupil and limbus center measurements became small, if not negligible. However, this measurement procedure, which requires covering the right eye and the left eye, often cannot be applied in infants, toddlers, and preschool children, in whom screening with an objective ocular alignment technique is most useful. It is of clinical and practical interest to note that when coaxial infrared lighting screening techniques24'25 are used, the entrance pupil margins are easily identifiable structures with simple thresholding algorithms in digitized images. This is because the

2605 dominant image portions are the fundus reflexes against which the corneal reflex appears as a single bright reflex and the entrance pupil margins as sharp boundaries of the fundus reflex. The limbal region is imaged with less contrast under these lighting conditions, which may lead to neglect of the limbus center as a landmark. This explains, possibly together with a limited resolution of the CCD camera used in the apparatus, the limited accuracy of 8 to 10 PD (4.6 to 5.7) for detection of ocular misalignment reported26 elsewhere. In the current study, despite subjective impression, the limbus could be located with precision interactively. We also showed elsewhere that image processing may be used to automate the detection of limbus structures in digitized images36 and with nearly the same precision as is possible in detection of the entrance pupil margins. These results prove that the accuracy of measurement of ocular misalignment depends decisively on the choice of the reference landmark. Different HRs should be used, depending on this choice, because the mean limbus value was 1.5 PD/mm lower than the entrance pupil value in this study. Normative values for h, the distance that, together with the corneal radius and limbus diameter, governs the HR, were established for limbus and entrance pupil center landmarks for an average human adult eye with corneal radius r= 7.8 mm. Most data in the literature 512 " 141819 refer to the entrance pupil center only or do not take into account systematic differences between pupiland limbus-based measurement. For specific populations, in which the corneal radius may be measured, the HR may be determined from our data by interpolation or extrapolation of h, unless the HR is calculated through plots of measured corneal reflex shift versus gaze angle. Using this alternative for evaluation of the present alignment data led to the same systematic difference between pupil- and limbus-based HRs. We conclude that, as a rule, limbus-based HRs are 1 to 2 PD/mm lower in humans than are pupil-based HRs. In animals, similar systematic effects can be predicted. After the current study, which was conducted solely on control subjects, to minimize errors that would impede the accurate assessment of misalignment measurements, a comparison of methods could be done in strabismic subjects with orthoptic control data. However, care must be taken to avoid subjectrelated, time-dependent fluctuations of the angle of strabismus when different techniques are compared. Because random errors would be larger than those in this study, the number of strabismic subjects would have to be much higher to ascertain the same amount of systematic difference. Because of the design of the current study, it was possible to examine differences between techniques with a relatively low number of subjects.

2606

Investigative Ophthalmology & Visual Science, November 1997, Vol. 38, No. 12 pass the critical value of 4 pixels per millimeter; some perform well even if only one camera is used to image both eyes at a time. However, care must be taken not to overestimate the spatial and temporal resolution that can realistically be expected with some CCD cameras: For instance, in the interlace mode, 39 half-images are typically read out at 50 to 60 Hz, which can cause motion artifacts (shifted half-images of moving objects), especially in eye-tracking applications; and asynchronous pixel readout or aliasing may cause errors of at least 1 pixel width.39 The maximum spatial resolution is also limited by the amount of CCD sencor elements per row and number of rows, signal-to-noise ratio, bandwidth of the analog video signal, and the frame grabber resolution. Such effects limit the benefit of any subpixel evaluation technique. 16 These problems may be avoided by using appropriate imaging devicesfor instance, analog full-frame-transfer CCD cameras and cameras with a synchronized pixel transfer, or CCD cameras with a digital output, which do not require a frame grabber. Therefore, in addition to the proper choice of the landmark for corneal reflex position, we recommend that the design of recording devices for screening, clinical, and research purposes be guided by the desired measuring accuracy or threshold for detection and that the hardware be carefully selected and tested to ensure that those criteria are met. Key Words accuracy, angle of strabismus, corneal reflex, instrumentation, paraxial ray tracing analysis References 1. Paliaga GP, Paliaga A, Decarli A. Evaluation of ocular rotation by measurement of corneal light reflex movements. Clin Ocul. 1983:499-502. 2. Rolling G. Zur photographischen Dokumentation des schwankenden Schielwinkels. Fortschr Ophthalmol. 1983;79:480-482. 3. Brodie SE. Photographic calibration of the Hirschberg test. Invest Ophthalmol Vis Sci. 1987;28:736-742. 4. Griffin JR, McLin LN, Schor CM. Photographic method for Bruckner and Hirschberg testing. Optom Vis Sci. 1989;66:474-479. 5. Quick MW, Boothe RG. A photographic technique for measuring horizontal and vertical eye alignment throughout the field of gaze. Invest Ophthalmol Vis Sci. 1992;33:234-246. 6. DeRespinis PA, Naidu E, Brodie S. Calibration of Hirschberg test photographs under clinical conditions. Ophthalmology. 1989; 96:944-949. 7. Paliaga GP. Validity, sensitivity and specificity of the (Hirschberg) corneal light reflection test in 390 patients. Binocular Vision. 1989; 4:59. 8. Campos EC, Orciuolo M, Schiavi C. A new automatic deviometer. International Ophthalmology. 1989; 13:291-295.

Other studies concluded that h was usually independent of age after 3 years of age and that the statistical correlation of the HR with the corneal radius was too weak19 to be relevant. Although this may be true, the systematic yet small difference in HR related to the choice of the landmark should be taken into account, because it would improve reported insufficient thresholds for ocular misalignment detection. 26 The highest accuracy of ocular alignment measurement was obtained with the monocular approach, in which the choice of limbus or entrance pupil center did not make a difference. However, such a measurement strategy depends on the cooperation of the subject. In conclusion, we recommend the use of the limbus center for screening and measuring applications with corneal reflexes, whenever recorded images are used. The threshold for detection of ocular misalignment in the binocular approach, the 95% confidence interval, was 3.174 (5.5 PD), compared with nearly twice that error in the widely used entrance pupil center evaluation. To increase the precision of ocular alignment measurement further, measurement with corneal and posterior crystalline lens reflexes from three light sources30'34'35'37 could be used. Using this methodology and the alignment data of this study, the threshold for detection of ocular misalignment was 2.1 (3.7 PD). If the corneal reflex position is simply estimated during a clinical examination, without recording, we speculate that the limbus option probably does not improve the accuracy of the eye alignment assessment: The errors in the estimation of the asymmetry of the corneal reflex position, especially regarding the limbus center, are probably much larger than the potential increase in accuracy indicated by the results of the current study. A similar effect can be expected if the spatial resolution of the recorded images is lowthat is, on the same order as the inherent errors of binocular corneal reflex position measurement: If the limbus-center-related threshold in the primary gaze is taken from the current data as 3.174 (5.5 PD), and a typical HR is ~21 PD/mm, then a critical limit must be expected for a resolution of ~0.25 mm/pixel, or 4 pixels per millimeter as a hardware magnification factor (compared with 21.5 pixels per millimeter in this study, with one camera per eye). To image spatial differences in the corneal reflex position of 0.25 mm reliably, the required pixel density must be higher than this limit, depending on the type of hardware used.39 Fortunately, 35-mm photographs or slides have much better resolution than 0.25 mm, 39 and most of the currently available low-cost analog CCD cameras and digitizing frame-grabbers frequently used for eye alignment measurement perform well enough to sur-

Accuracy of Ocular Alignment Measurement 9. Ruttum MS, Shimshak KJ. Photographic measure25. ment of the angle of strabismus. In: Strabismus and Ocular Molility Disorders. Campos EC, ed. Houndmills, Macmillan Press; 1990:155-160. 26. 10. Hatsukawa Y, Otori B, Patel B, Hogg C, Fells P. The use of the corneal reflex test through glasses. In: Strabismus and Ocular Motility Disorders. Campos EC, ed. Houndmills, Macmillan Press; 1990:161-166. 27. 11. Miller JM, Simons K. Sub-pixel imaging permits automated Hirschberg test measurement of small strabis28. mic deviations with a standard video camera. Invest Ophthalmol Vis Sci. 1991;32:1239. 29. 12. Eskridge JB, Perrigin D, Leach NE. The Hirschberg Test: Correlation with Corneal Radius and Axial Length. Optom Vis Sci. 1990;67:243-247. 30. 13. Carter AJ, Roth N. Axial length and the Hirschberg test. American Journal of Optometry and Physiologic Optics. 1978;55:361-364. 14. Eskridge B, Wick B, Perrigin D. The Hirschberg test: A double-masked clinical evaluation. Am J Optom Phys- 31. iolOpt. 1988; 65:745-750. 15. Wick B, London R. The Hirschberg test: Analysis from 32. birth to age 5. J Am Optom Assoc. 1980;51:1009-1010. 16. Miller JM, Mellinger M, Greivenkamp J, Simons K. 33. Videographic Hirschberg measurement of simulated strabismic deviations. Invest Ophthalmol Vis Sci. 1993; 34:3220-3229. 34. 17. Williams SM. The rulerless Hirschberg method. Invest Ophthalmol Vis Sci. 1994;35:S2202. 18. Riddell PM, Hainline L, Abramov I. Calibration of the Hirschberg test in human infants. Invest Ophthalmol 35. Vis Sci. 1994;35:538-543. 19. Jones R, Eskridge JB. The Hirschberg-TestA ReEvaluation. Am J Optom. 1970;105-114. 20. Weekers R, Gilson M, Joachim M. Measure des petits 36. angles de strabisme au moyen d'une technique photographique. Ophthalmologica 1963;146;217-226. 21. Brodie SE. Corneal topography and the Hirschberg 37. test. Appl Opt. 1992;31:3627-3632. 22. Bronstein IN, Semendjajew KA. Tables of series expansions. In: Taschenbuch der Mathematik, 19th ed. Leipzig: Harri Deutsch; 1979:85. 23. Bennett AG, Rabbetts RB. Clinical Visual Optics. 2nd ed. 38. London: Butterworths; 1989:17,257-259,278-279. 24. Angi MR, Bergamo L, Bisantis C. The binocular videorefractoscope for visual screening in infancy. GerJOph39. thalmoll99S;2:l82-18S.

2607

Uozato H, Saishin M, Guyton DL. Refractive Assessment of infants with infrared video-refractor PR-1000. Invest Ophthalmol Vis Sci. 1991;32:S1238. Guinta A, Peck G, Hui J, Howland H. Horizontal and vertical tropias and phorias in a normal population of infants and children. ARVO Abstracts. Invest Ophthalmol Vis Sci. 1995;36:S953. Graham PA, Naylor EJ. A photographic method of measuring the angle of squint. Br J Ophthalmol. 1957; 61:4-7. Wyatt HJ. The form of the human pupil. Vision Res. 1995; 35:2021-2036. Charlier JR, Behague M, Buquet C. Shift of the pupil center with constriction. Invest Ophthalmol Vis Sci. 1994;35:S1278. Barry JC, Effert R, Kaupp A, Burhoff A. Measurement of ocular alignment with photographic Purkinje I and IV reflection pattern evaluation. Invest Ophthalmol Vis Sci. 1994; 35:4219-4235. von Noorden GK. Binocular Vision and Ocular Motility. 4th ed. St. Louis: Mosby; 1990:170-174. Bennett AG, Rabbetts RB. The schematic eyeangles kappa, alpha. Clinical Visual Optics. 2nd ed. London: Butterworths; 1989:264-266. Bland JM, Altmann DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986:307-310. Barry JC, Effert R, Kaupp A. Objective measurement of small angles of strabismus in infants and children with photographic Purkinje reflection pattern evaluation. Ophthalmology. 1992;99:320-328. Barry JC, Effert R, Reim M, Meyer-Ebrecht D. Computational principles in Purkinje I and IV reflection pattern evaluation for the assessment of ocular alignment. Invest Ophthalviol Vis Sci. 1994;35:4205-4218. Barry JC, Pongs UM, Hillen W. An algorithm for Purkinje image I and IV and limbal center localization. Computers in Biology and Medicine. 1997 (in press). Barry JC, Backes A, Pongs UM, Kirschkamp T, Dunne MCM. Improved computing scheme for measurement of ocular alignment with Purkinje image I and IV reflection patterns. OphthalmolPhysiol Opt. 1997 (in press). Barry JC, Branmann K, Dunne MCM. Study of catopuic properties of eyes with misaligned surfaces by exact ray tracing. Invest Ophthalmol Vis Sci. 1997; 38:1476-1484. Inglis A. Elements of television picture quality. Video Engineering. New York: McGraw-Hill; 1993; 19-60.