Improvement of Diagnostic Performance Regarding Retinal Nerve Fiber Layer Defect Using Shifting of the Normative Database According to Vessel Position

Abstract

Purpose.: To evaluate the clinical efficacy of early detection of retinal nerve fiber layer (RNFL) defect in a Korean population using shifting of the normative RNFL thickness database according to vessel position.

Methods.: Retinal nerve fiber layer thickness data of 151 healthy eyes (normative group) and 120 validation subjects (validation group; additional healthy controls plus early glaucoma) were prospectively obtained using spectral-domain optical coherence tomography (SD-OCT) measurement. Clinical profiles and position of major retinal artery peaks (superotemporal, STa; inferotemporal, ITa) were investigated with position of RNFL peaks. Three different criteria for the lower 1% limit were adopted for the validation. Criterion 1 used normative data from the manufacturer. Criterion 2 used data from healthy volunteers. Criterion 3 used four combinations of two subgroups from data for volunteers, which were divided by the median value of STa in the superior region and that of ITa in the inferior region. The κ value was used to determine the diagnostic performance of each criterion (agreement with standard answer).

Results.: Assessment of the validation group using criterion 3 showed greater accuracy than with criterion 1 or criterion 2 (κ = 0.571, 0.774, and 0.979). Although SD-OCT specificity for RNFL defect detection was similar among the criteria (100%, 98.8%, and 98.9%), sensitivity was highest with criterion 3 (42.4%, 72.7%, and 100%) (all values; criteria 1, 2, and 3, respectively).

Conclusions.: Shifting of the lower 1% reference line, according to vessel position, could remarkably improve the diagnostic performance regarding RNFL defect detection with SD-OCT.

Introduction

In glaucoma patients, confirming retinal nerve fiber layer (RNFL) defects is crucial to early detection and management of their quality of life. The earlier the detection, the higher the quality of life. ^1,2 Retinal nerve fiber layer thickness can be measured with optical coherence tomography (OCT). As technology has developed, the resolution of OCT has become better than ever. However, clinically, many cases that have obvious RNFL defects without any “red alert” from the OCT device have been observed. ^3,4 With regard to clinicians' decision about treatment, these cases are more problematic in early glaucoma patients than in moderate to severe patients. 

Retinal nerve fiber layer defects can be confirmed when compared with normative databases. ⁵ Although the current normative data distributed by each manufacturer generally show good sensitivity and specificity for detecting RNFL defects, individual variations in RNFL thicknesses exist and lower the sensitivity. Hood et al. ^6,7 demonstrated that variations in the local peaks of the RNFL correlated with the location of the main temporal branches of the superior and inferior vessels. Because of this variation, RNFL defects using OCT scans demonstrated areas under the receiver operating characteristic curves (AUROCs) of 0.7 to 0.9, which were only fair to good even in advanced glaucoma patients. ^8–11 And other factors like axial length and age can also affect the RNFL thickness. ⁴ However, we need to know how strongly these factors are related to the RNFL peak position. Another problem with the normative data has been the conventional normative databases from the manufacturers, which include 0% to 6% Asians. Several articles have shown that RNFLs in an Asian population are thicker than in a Caucasian population. ^12–14 Therefore, before evaluating the vessel contribution to RNFL peaks, an ethnic normative database should have been established using each OCT device. And finding the most contributable factor is imperative to developing an easier approach in our real practice. 

The aim of this study was to establish a normative database of RNFL thickness in a Korean population and to evaluate the clinical efficacy of early glaucoma detection using shifting of the normative RNFL thickness according to the position of the major retinal artery, the factor that contributes most to the RNFL peak position. 

Methods

One hundred fifty-one healthy eyes of 151 Korean adults, 58 early glaucoma eyes of 58 Korean adults, and 62 age-matched additional healthy control eyes of 62 Korean adults were consecutively enrolled in 2013. This study was approved by the Institutional Review Board of CHA Bundang Medical Center and conducted at CHA Bundang Medical Center. The study conformed to the Declaration of Helsinki, and written informed consent was obtained from all subjects. All subjects underwent a full ophthalmologic examination including best-corrected visual acuity (BCVA), Goldmann applanation tonometry by a glaucoma specialist (SR), central corneal thickness (CCT), axial length (AXL), spherical equivalent refractive error (SER), gonioscopy, visual field test, fundus examination after pupil dilation, fundus photography, red-free photography, and spectral-domain optical coherence tomography (SD-OCT). Central corneal thickness was measured using an ultrasonic pachymeter (UP-1000; Nidek, Gamagori, Japan), and AXL was measured using Echoscan (US-4000; Nidek). Fundus photography and red-free photography were performed with a fundus camera (VX-10i; Kowa, Nagoya, Japan). 

The inclusion criteria for both the normative group and the validation group were visual acuity of at least 20/30 and age between 18 and 80 years. Subjects with clinical evidence of retinal disease, neuro-ophthalmologic disease, history of refractive or retinal surgery within 3 months, or closed iridocorneal angle and refractive error more than ±8.0 diopters and ±3.0 diopters of cylinder were excluded from both groups. In the normal group, subjects were additionally excluded if they had intraocular pressure (IOP) higher than 21 mm Hg, glaucomatous optic disc, or abnormal visual field test results (described in the section on visual field tests below). Cases of severe myopic degeneration on the retina or physiologic cupping with large optic disc size were also excluded for the sake of accuracy. Subjects were consecutively enrolled in a validation set if they had mean deviation (MD) values higher than −6 decibels (dB; normal or early damage). One eye from each participant was included randomly in the assessment. 

The validation group was set to compare the RNFL defect-detecting ability of three different comparative criteria for lower 1% limit. Criterion 1 was from the manufacturer's data; criterion 2 was from the established normative data in 151 healthy Korean eyes; and criterion 3 was from shifting of the data of the established normative values in 151 healthy Korean eyes considering each volunteer's vessel position. For criterion 3, the volunteers' normative data were separated into four subgroups, which were combinations of two superotemporal groups and two inferotemporal groups, according to each volunteer's vessel position, for confirmation of RNFL defects. As the parameter contributing the most to the RNFL peak was the position of the major retinal artery (described in the section on statistical analysis), the superotemporal and inferotemporal subgroups were divided by the median angle degree of the superotemporal and inferotemporal major vessel peaks, respectively. These were designated references 1, 2, 3, and 4 (see more in the sections on RNFL measurement and statistical analysis). All three criteria were assessed based on the standard criterion defined as the “early glaucoma state.” The early glaucoma state was defined as that in which subjects showed glaucomatous optic disc contours plus a corresponding abnormal visual field result, or RNFL defects in red-free photography plus a corresponding abnormal visual field result. Using these parameters, validation subjects were divided into two groups: “healthy control,” which should not be detected by the RNFL defect measurements, or “early glaucoma,” which should be detected by the RNFL defect measurements. 

RNFL Measurements and Evaluation of Angles

All scans were done by a highly experienced operator. The peripapillary RNFL thicknesses, with 16 averaged, consecutive circular B-scans (diameter of 3.5 mm, 768 A-scans), were measured using the Spectralis OCT (Heidelberg Engineering GmbH, Heidelberg, Germany). An online tracking system was utilized for eye movement compensation. To avoid the FoDi technology effect, the internal fixation mode was deactivated. The subjects were seated and aligned their heads upright. After both eyes were examined, one eye was included in the study if it was within the 95% range of fovea location (from 0.01 disc diameter [DD] to 0.50 DD inferior to the optic disc center) and showed no asymmetry of fovea location (asymmetry of both eyes > 0.25 DD). ^15,16 The RNFL thickness from the inner margin of the internal limiting membrane to the outer margin of the RNFL layer was automatically segmented using the instrument's built-in software (version 5.4.7.0). All measurements were automatically segmented and reconfirmed one by one. Only the OCT images with image quality scores of more than 22 were used in the analyses. 

From temporal-superior-nasal-inferior-temporal (TSNIT) graphs, each subject's RNFL thickness data at intervals of 5° (a total of 72 values) were extracted using Cool Ruler software (version 1.5, freeware from Fabsoft, Inc.; www.fabsoft.com [in the public domain]). Quantitative comparison between the reconstructed graph and the original graph for the entire group of subjects (n = 271) was performed using ImageJ software (National Institutes of Health; http://rsb.info.nih.gov/ij/index.html [in the public domain]). Area under the graph as a random multiple was calculated using the “analyze particle” function of ImageJ. The mean area difference for the reconstructed graph versus the original graph was 0.12 ± 0.09% (0.11 ± 0.09%, 0.12 ± 0.10%, and 0.13 ± 0.07% for the normative group, age-matched control eyes in the validation group, and early glaucoma eyes in the validation group, respectively; P = 0.96, one-way analysis of variance, ANOVA). The highest point (presented as angle degree) of RNFL thickness was defined as the superotemporal peak (STp) and inferotemporal peak (ITp). The location of the vessel peak in the OCT image was measured using PicPick software (version 3.2.4, freeware from NTeWORKS, www.picpick.org [in the public domain]). The software can draw a horizontal line from the center of the disc as a reference line. The cross points in a 3.5-mm peripapillary circle and the major vessels were observed, and the angles between this reference line and the cross points were measured with PicPick software (angles of superotemporal artery [STa] superotemporal vein [STv], inferotemporal artery [ITa], and inferotemporal vein [ITv]). Figure 1 depicts an example of determining the STa, STv, ITa, and ITv. The major vessel was defined as the thickest vessel heading toward the temporal region originating from the major branch. All subjects' RNFL data were correctly segmented, and no eyes with vague angle direction of blood vessel were detected by the secondary manual confirmation. The center of the disc was determined at the onset of the scanning procedure by an experienced operator. 

Figure 1

 

Fundus images of OCT device and fundus camera determining the angle degree of major retinal artery and vein. Left: The angles α and β indicate the position of the superotemporal (STa) and that of the inferotemporal major retinal artery (ITa) from the horizontal line, respectively. With a setting of 360°, the position of the ITa was considered as 360° minus angle β when it was utilized for analysis. The green circle with arrowhead indicates the scanned line automatically presented by the SD-OCT built-in software. Right: Major retinal artery was confirmed with the fundus camera image. The short red lines highlight the cross regions between the major retinal artery and 3.5-mm peripapillary circle (green circle, same size as the OCT image at left), and the short blue lines highlight the cross regions between the major vein and 3.5-mm peripapillary circle.

Visual Field Tests

Standardized visual field tests were performed using static automated white-on-white threshold perimetry (Swedish Interactive Threshold Algorithm standard 24-2, Humphrey Field Analyzer II; Carl Zeiss Meditec, Inc., Dublin, CA, USA). A reliable visual field was defined by fixation loss < 30%, false-positive error or false-negative error < 20%. An abnormal visual field was defined by one or more of the following criteria: glaucoma hemifield test outside normal limits; pattern standard deviation decreased to P < 0.05; three or more nonedged points in a cluster decreased to P < 0.05, with one of these decreased to P < 0.01. ¹⁷ Abnormal field test results were reconfirmed with a repeat examination within 3 months from the baseline visit. 

Validation Through Graphic Display

All extracted data for RNFL thicknesses were displayed through SigmaPlot software (version 12.0; Systat Software, Inc., San Jose, CA, USA), and all graphs were presented in a spline (smoothed) shape mode. If a subject's RNFL thickness line contacted the reference graph (i.e., lower 1% limits of normative data), even in the smallest region of the graph in any location, it was defined as abnormal. When the RNFL graph was compared with the reference from the manufacturer, the same definition was applied. 

Statistical Analysis

Statistical analyses were performed using the SPSS software (version 21.0; SPSS, Inc., Chicago, IL, USA). χ² test, independent sample t-tests, and one-way ANOVA were utilized to compare demographic characteristics. P values < 0.05 were defined as statistically significant. 

To determine the factor most correlated to the RNFL peak for criterion 3, Pearson's correlation coefficients were calculated to assess the correlation of each parameter with STp and ITp. For criterion 3, the most highly correlated parameter was adopted for the grouping procedure from each region (superotemporal and inferotemporal). To confirm the suitability of the model developed through adoption of these parameters, multiple regression analysis was performed, and the adjusted R ² was calculated (using stepwise selection method). 

Table 1 describes the parameters that were correlated to STp and ITp. The highest Pearson's correlation coefficients with STp and ITp were STa (r = 0.756, P < 0.001) and ITa (r = 0.587, P < 0.001), respectively. For criterion 3, the superior region data in the normal group were divided into two groups according to the median value (major artery angle degree of 71) of the STa calculated from the normal group (n = 151). The same method was used for the inferior region data in the normal group according to the median value of the ITa (major artery angle degree of 287). To ascertain the feasibility of the models that had STa and ITa as the only independent variables in each peak group (STp and ITp), adjusted R ² were calculated from multiple regression analyses. In the STp and ITp model, the adjusted R ² were 0.568 (P < 0.001) and 0.340 (P < 0.001), respectively. 

Table 1

 

Correlation Between Retinal Nerve Fiber Layer Peaks and Clinical Profiles in the Normal Group

	STp		ITp
	Pearson's Correlation Coefficient	P Value	Pearson's Correlation Coefficient	P Value
Age, y	0.382	<0.001	−0.034	0.678
Sex	0.065	0.425	0.119	0.144
IOP, mm Hg	0.147	0.072	−0.095	0.246
AXL, mm	−0.305	<0.001	0.175	0.032
SE, D	0.408	0.005	−0.351	0.018
CCT, μm	0.058	0.481	−0.086	0.293
STa, deg	0.756	<0.001	−0.39	<0.001
STv, deg	0.343	<0.001	−0.171	0.035
ITa, deg	−0.338	<0.001	0.587	<0.001
ITv, deg	−0.378	<0.001	0.365	<0.001

As a next step, the lower 1% limit of each group's RNFL data (STa < 71, STa ≥ 71, ITa < 287, and ITa ≥ 287] was calculated. Four new references were applicable through the combinations of each subgroup (reference 1 [STa < 71 + ITa < 287], reference 2 [STa ≥ 71 + ITa < 287], reference 3 [STa < 71 + ITa ≥ 287], and reference 4 [STa ≥ 71 + ITa ≥ 287]). These new reference groups were used for criterion 3 and were applied to each subject according to the subject's own STa and ITa values. Although the correlation between RNFL peaks and the peak location of vessel was high (Table 1) and the model including STa + STv and ITa + ITv showed the highest adjusted R ² value (Table 2), only the model including STa and ITa was utilized to avoid excessive sample size. 

Table 2

 

Model Fitness of the Superotemporal and Inferotemporal Retinal Nerve Fiber Layer Peak Compared by Adjusted R ² Values (Multiple Regression Analysis, Stepwise Selection Method)

		Entered Variables	B (SE)	P Value	Adjusted R 2
STp	Model 1	STa	0.733 (0.052)	<0.001	0.568
	Model 2	STa	0.694 (0.053)	<0.001	0.583
		STv	0.123 (0.049)	0.012
ITp	Model 1	ITa	0.509 (0.058)	<0.001	0.34
	Model 2	ITa	0.472 (0.055)	<0.001	0.414
		ITv	0.199 (0.045)	<0.001

 The variables included in the analysis were age, sex, intraocular pressure, axial length, central corneal thickness, STa, STv, ITa, and ITv. B, predicted regression coefficient; SE, standard error; R ², determination coefficient.

For comparison among each criterion, κ values were extracted to evaluate the agreement of three criteria with the early glaucoma state. (The κ statistic was calculated from a 2-by-2 table in which the result for each criterion corresponds to that for early glaucoma diagnosis. Kappa values between 0.0 and 0.2 indicated slight agreement; those between 0.21 and 0.40 indicated fair agreement; those between 0.41 and 0.60 indicated moderate agreement; those between 0.61 and 0.80 indicated good agreement; and those between 0.81 and 1.00 indicated almost perfect agreement.) With estimated κ values of 0.4 and 0.8 for criterion 1 and criterion 3 from a pilot study, a minimal number of 56 subjects was required to detect a 0.4 difference in κ values above positive ratings of 0.5 with 90% statistical power. ¹⁸  

Results

Table 3 shows the demographic features of 151 healthy volunteers (normal group) and 120 subjects for validation (validation group). The mean age of the normal group was slightly younger than that of the validation group. The mean values of BCVA, MD, and visual field index (VFI) in the normal group were higher than those in the validation group. Forty-eight percent of the normal group and 51% of the validation group were left eyes that were enrolled in the analysis (P = 0.767). Male/female percentages, IOP, AXL, SER, and CCT measurements were also similar in the two groups. In the validation group, parameters for the healthy control group and the early glaucoma group including age, sex, IOP, AXL, and CCT were similar, but MD and VFI were not (Table 4). 

Table 3

 

Demographic Features of Normal Group and Validation Group

	Normal Group, n = 151		Validation Group, n = 120		P Value*
	Mean (Range)	SD	Mean (Range)	SD
Age, y	47.03 (20 to 80)	12.85	50.62 (19 to 76)	13.56	0.027
Sex, % female	51	-	57	-	0.391
BCVA, Snellen	0.97 (0.8 to 1.0)	0.08	0.94 (0.8 to 1.0)	0.1	0.025
IOP, mm Hg	15.30 (10 to 20)	2.68	14.90 (10 to 22)	2.58	0.21
AXL, mm	23.93 (21.15 to 27.90)	1.26	23.90 (21.60 to 27.90)	1.37	0.814
SE, D	−1.39 (−7.50 to +3.00)	2.64	−1.05 (−8.00 to +2.00)	2.95	0.573
CCT, μm	546.66 (435 to 621)	32.77	544.30 (480 to 614)	26.24	0.531
MD, dB	−0.74 (−4.16 to 2.22)	1.29	−1.11 (−5.65 to 2.11)	1.55	0.039
VFI, %	99.05 (95 to 100)	1.09	98.64 (88 to 100)	1.86	0.032

*   Independent samples t-test except for sex variable (χ² test).

Table 4

 

Demographic Features of Two Groups (Early Glaucoma Group and Age-Matched Healthy Control Group) Within the Validation Group

	Healthy Control, n = 62		Early Glaucoma, n = 58		P Value*
	Mean (Range)	SD	Mean (Range)	SD
Age, y	52.05 (20 to 70)	11.54	53.31 (19 to 76)	15.51	0.64
Sex, % female	55	-	59	-	0.715
BCVA, Snellen	0.94 (0.8 to 1.0)	0.1	0.95 (0.8 to 1.0)	0.09	0.935
IOP, mm Hg	14.81 (10 to 20)	2.56	15.00 (11 to 22)	2.62	0.683
AXL, mm	24.02 (22.29 to 27.90)	1.2	23.73 (21.60 to 27.49)	1.57	0.28
SE, D	−1.15 (−7.00 to +1.00)	2.56	−0.94 (−8.00 to +2.00)	3.37	0.81
CCT, μm	547.76 (496 to 614)	24.6	539.92 (480 to 585)	27.81	0.118
MD, dB	−0.66 (−3.64 to 2.11)	1.34	−1.60 (−5.65 to 0.50)	1.62	0.001
VFI, %	99.10 (96 to 100)	1	98.14 (88 to 100)	2.38	0.006

*   Independent samples t-test except for sex variable (χ² test).

Comparison of RNFL Defect-Detecting Performance

Figure 2 shows comparisons of the lower 1% limit from three different normative data sets. The green line (criterion 1) indicates that the data from the manufacturer had the lowest values. The blue line (criterion 2) indicates the data from healthy Korean volunteers, showing increased values at most angle locations. The red line (criterion 3) indicates the data after a recombination, considering the STa and ITa locations, and demonstrates a greater increase than the blue line data. The graph at the bottom of Figure 2 indicates that nasalized (red dot line, reference 2) and temporalized (red dashed line, reference 3) references were higher than the reference for the total population (criterion 2). These results also suggested that the nasalization or temporalization of the major artery definitely contributed to the angle and the height of the RNFL peak. 

Figure 2

 

The lower 1% limit of retinal nerve fiber layer (RNFL) thickness according to each criterion. The green line indicates the lower 1% limit of the manufacturer's data (criterion 1). The blue line indicates data from a healthy Korean population (criterion 2). The red line shows regrouping of data after shifting, based on each subject's vessel position (criterion 3). Top left: Reference (Ref.) 1 of criterion 3, which demonstrates the lower 1% limit of subjects with STa < 71 and ITa < 287. Top right: Ref. 2, which demonstrates subjects with STa ≥ 71 and ITa < 287. Mid left: Ref. 3, which demonstrates subjects with STa < 71 and ITa ≥ 287. Mid right: Ref. 4, which demonstrates subjects with STa ≥ 71 and ITa ≥ 287. Bottom: The more nasalized group (Ref. 2) and more temporalized group (Ref. 3) show higher levels of the lower 1% limit than the other groups.

Table 5 shows the agreement of the three criteria with the standard diagnosis of early glaucoma eyes. Kappa values for criterion 3 (κ = 0.979, P < 0.001) were highest among the three criteria. However, criterion 2 (κ = 0.774, P < 0.001) also showed higher κ values than criterion 1 (κ = 0.517, P < 0.001), suggesting that the particular normative database for the Korean population aids early detection of RNFL defects. This aspect was also seen in the sensitivity and specificity check. The sensitivity and specificity for each criterion, compared with the standard criterion for early glaucoma state diagnosis, are also listed in Table 5. Although criteria 1 and 2 showed good discriminatory ability, criterion 3 yielded the highest results. 

Table 5

 

Kappa Values, Sensitivity, and Specificity of Each Criterion for Detecting Early Glaucoma State

	κ Value (95% CI)	P Value	Sensitivity (95% CI)	Specificity (95% CI)
Criterion 1	0.517 (0.343–0.691)	<0.001	0.424 (0.272–0.592)	1.000 (0.958–1.000)
Criterion 2	0.774 (0.643–0.905)	<0.001	0.727 (0.557–0.849)	0.988 (0.938–0.998)
Criterion 3	0.979 (0.938–1.000)	<0.001	1.000 (0.896–1.000)	0.989 (0.938–0.998)

 CI, confidence interval.

Two representative cases showing a remarkable ability to detect early glaucoma are displayed in Figures 3 and 4. A localized RNFL defect (defined as angular widths of 10°–30°) and a wedge RNFL defect (defined as angular widths < 10°) that could not be detected by criteria 1 and 2 were detected by criterion 3. There was one false-positive case using criterion 3. The false-positive case had a RNFL defect located at the superonasal region (135°), which in general correlated less well with the glaucomatous defect. 

Figure 3

 

Case 12 in the validation group is a representative case that shows the detection ability of criterion 3 in early glaucoma diagnosis. The red area indicates the lower 1% limit of retinal nerve fiber layer thickness according to each of the data points. Top left: Red-free photography shows a wedge-shaped RNFL defect in the superotemporal region. The short red arrows indicate the margin of the defect. Bottom left: Pattern deviation and global indices of visual field tests demonstrate borderline defects (mean deviation of −2.88 dB) in the inferotemporal quadrant. Top right: Criterion 1 and (mid right) criterion 2 cannot detect any RNFL defect by the lower 1% limit of RNFL thickness. Bottom right: Criterion 3 data are shown. Only criterion 3 detects the glaucomatous RNFL defect (long red arrow).

Figure 4

 

Case 60 in the validation group is a more definite RNFL defect case detected only by criterion 3. The red area indicates the lower 1% limit of retinal nerve fiber layer thickness according to each data point. Top left: Red-free photography shows one wedge-shaped RNFL defect in the superotemporal region (short red arrows) and one localized RNFL defect in the inferotemporal region (short blue arrows). Bottom left: Pattern deviation and global indices of visual field test demonstrate borderline defects (mean deviation of −2.45 dB) in both superior and inferior regions. Top right: Criterion 1 and (mid right) criterion 2 cannot detect any RNFL defect by the lower 1% limit of RNFL thickness. Bottom right: Only criterion 3 detects both RNFL defects (the long red arrow depicts the wedge defect, and the long blue arrow depicts the localized defect).

Discussion

Optical coherence tomography is a highly developed and advanced technique that can measure RNFL thickness in healthy and glaucomatous eyes. However, OCT does not always provide informative data because of its low sensitivity with respect to glaucoma detection. Budenz et al. ⁸ reported that even in eyes with moderate to advanced glaucoma, time-domain OCT (Stratus OCT) showed sensitivity of only 68%, which was judged by the lower 1% limit of its normative database. The sensitivity decreased to 61% in early glaucomatous subjects, although it increased to 65% using SD-OCT. ¹¹ Despite the improvement in coherence, the sensitivity of OCT devices was not optimal especially in early glaucoma patients. 

In the past 10 years, the diagnostic performance of OCT devices for glaucoma detection with various methods converged to approximately between 0.7 and 0.9. However, most studies were conducted in populations with varying degrees of glaucomatous damage. ^4,8–11 Kanamori et al. ¹⁹ conducted a study using OCT 2000 showing that the AUROC for the inferior quadrant was highest for the early glaucoma group (AUROC = 0.863), but the sensitivity was less than 67% (specificity setting of more than 90%) for average RNFL thickness. Nouri-Mahdavi et al. ²⁰ reported the highest AUROC in the superior quadrant for OCT 2000 as 0.840 in their study of 59 early glaucoma eyes; however, the sensitivity was less than 77% (specificity at 90%) for average RNFL. In 2012, Lisboa et al. ²¹ reported the highest sensitivity of Spectralis OCT as 47.6% (specificity at 95%), even though the largest AUROC in a preperimetric glaucoma population with the SD-OCT was 0.88. Recently, a study by Kim and colleagues ²² showed that the AUROC of the best parameters in macular ganglion cell–inner plexiform layer (GCIPL) thickness was 0.823 in a preperimetric glaucoma population; the sensitivity was only 58.7% (specificity at 81.5%). In our study, Spectralis OCT showed an AUROC of 0.859 (detailed data not shown) in early glaucoma subjects (mean MD of −1.11 dB) when all seven RNFL thickness values (average RNFL thickness and quandrant RNFL thicknesses of superotemporal, temporal, inferotemporal, inferonasal, nasal, and superonasal region) were considered in a multiple logistic regression model. And even though the AUROC value was similar to the results from previous studies using other devices, each RNFL thickness had extremely poor AUROC values (between 0.153 and 0.362). This difference may be due to the racial difference and our subjects' earlier degree of damage compared to subjects in other studies. Our study sought to show Spectralis OCT results according to a new normative database based on vessel position, which agreed with an early glaucoma diagnosis. As multiple logistic regression analysis is not always feasible in clinical practice, considering all seven RNFL values may be inefficient. 

Studies on more damaged glaucomatous eyes have shown higher diagnostic performance. Lu et al. ¹⁰ conducted a study with 99 variously damaged glaucoma eyes that showed the highest AUROC (0.92) using the combination of overall average and inferior and superior quadrant RNFL thicknesses and a sensitivity of 73% (specificity at 95%). Leung et al. ²³ reported the diagnostic ability of Cirrus HD-OCT (variously damaged glaucoma eyes of average MD at −8.66 dB) as AUROC and sensitivity of 0.962 and 91% (specificity at 95%), respectively (P > 0.10). Because the purpose of our study was to confirm diagnostic performance for early detection of RNFL defects considering the factor most contributing to RNFL peaks (in our study, vessel positions), subjects with advanced glaucomatous damage, which could be discriminated easily, were not appropriate for enrollment. 

As shown in Table 1, although we utilized the peak location of the artery (STa and ITa) as the only parameter for discriminating shifting of the RNFL peak and the RNFL normative reference, there were other variables that correlated with the RNFL peaks (STp and ITp). The peak location of the vein was the second highest. In terms of model fitness, the STp model including both STa and STv showed the highest adjusted R ² in the analysis using a multiple regression model (Table 2). The ITp model including both ITa and ITv also showed the highest adjusted R ². This aspect was also reported by Hood et al. ^6,7 However, the reason for choosing the peak location of the artery was that more variables required a greater sample size. In this vein, we have sensed that interactive database construction with more data could be a great help in providing a more accurate model for the RNFL peak (which might be the next goal). 

Many parameters can affect the normative database of RNFL thickness. In our study, criterion 3 data showed remarkably higher κ values than we expected. We believe this is due to the use of the Korean normative database, because criterion 2 data also showed higher κ values than criterion 1 data. Many studies have also implied that ethnic variability could affect RNFL defect detection. ^12,13,24 Moreover, myopic eyes had more temporalized RNFL peaks. The relationship between myopia (or long AXL) and RNFL peak location has been described in many articles. ^24–26 Yamashita et al. ²⁵ reported that the angles between the superotemporal and inferotemporal peak RNFL positions were highly correlated with the artery angle (r = 0.92, P < 0.001) and were correlated negatively with the AXL (r = −0.49, −0.38; P < 0.01). Kang et al. ²⁶ showed that subgroups with higher AXL had more temporalized RNFL peak positions. Chung and Park ²⁷ also concluded that increased AXL appeared to increase the separation of the RNFL peaks. Interestingly, an age correlation with the RNFL peak location of the superior region was not found in the inferior region. In older-aged subjects, the STp was more nasalized (P < 0.001), and it is assumed that age correlated with RNFL thickness in the superior region, as previously reported. ^24,28–30 However the correlation faded away in a multiple regression analysis including other factors like AXL and CCT, implying its colinearity with these factors. 

A major limitation of the present study was that the data were not obtained from direct use of the manufacturer's software, which is not normally provided. Because we used data points at 5° intervals, detailed resolution could also have been diminished. This may cause bias and could reduce quality of detection. However, from qualitative (morphologic) and quantitative comparisons of all 271 cases, we confirmed that none were different from the original graph with the known reproducibility of OCT taken into consideration. The mean difference for the recon graph versus the original graph was 0.12 ± 0.09%, which was so low as to be almost negligible. Budenz et al. ³¹ suggested a difference of 9.5 μm in average RNFL thickness measured by the Stratus OCT, and Leung et al. ³² reported an intervisit reproducibility coefficient of average RNFL thickness measured by Cirrus HD-OCT as 4.8 μm. According to these firm results, we can estimate that the maximum reproducibility error of OCT is approximately 0.23% to 0.90% in area under the graph (presuming an average RNFL thickness of 100 μm), and we found that the graph reconstruction in our study was very accurate. With future use of the current Spectralis OCT software, the sectoral data for RNFL thickness could also be analyzed in terms of color code and compared with other reported data. 

Future studies in which the RNFL is imaged and regionalized relative to the anatomically determined axis between Bruch's membrane opening and the fovea ³³ are necessary to adequately address the importance of inconsistent and uncorrected torsion. The FoDi technology provides an axis that is a line from the fovea to the center of the disc. This technology certainly has some advantages in both clinical and theoretical respects. First, a torsional eye movement can be automatically detected and adjusted even in a baseline scan. We have confirmed that with the FoDi, a predicted horizontal axis was provided even in a severely torsioned eye at a baseline scan. According to the manufacturer, this is a novel process based on its normative database. Second, in a serial exam, the FoDi can align each scan to assist comparison at the same position on the TSNIT graphs. Third, the FoDi technology allows the researcher to compare the Spectralis data with those of other OCT devices, as well described by Patel and colleagues. ^34,35 However, at the same time, this technology has a couple of shortcomings. First, it is not open to the public how torsional eye movement can be detected and adjusted by the FoDi even in a baseline scan. The influence of the FoDi on other factors such as the location or course of the retinal vessels and proper disc centration still needs to be investigated, ^25,36 because at baseline there are no reference data to help make the decision of “wrong alignment” except for the manufacturer's normative database, which does not include enough Asian population. Second, despite the FoDi activation, variation in RNFL thickness measures still exists. We think that the major reason for the variation is improper fovea disc alignment by the FoDi technology as described by Valverde-Megias et al. ³⁷ in 2013. Practically, correcting mislocated FoDi is not easy. Thus at present there is not enough evidence to allow complete certainty as to its usefulness for research on torsion-related parameters. Also we must acknowledge that the present study did not focus on detection of an accurate fovea-to-disc axis but instead on the correlation between the position of the retinal vessel and that of the RNFL peak. 

One important advantage of considering vessel parameter as a major variable in a clinical setting is that its use could be more feasible and more efficient for detecting RNFL defects than use of other parameters like AXL or optic disc parameters. We can simply determine the angle degree of the vessel and take it into account for the subdivision of a new reference line (which can be pre-established), whereas other factors require additional examination tools or expensive devices for measurement. Also, if we had adjusted the AXL or magnification parameter, which is an established independent factor, this could simultaneously interfere with the analysis. Adjusting a factor for improvement of diagnostic performance is still debatable. ^26,36 In addition, although scan centration error should always be avoided, on the presumption of its effect on RNFL measurements, disparity between the position of the RNFL peak and that of the major retinal vessel should be minimal since we have utilized the same image. 

In conclusion, from our study results, we suggest that clinicians use new built-in software to provide the best use of the major vessel position for early glaucoma detection. As more data are obtained, additional variables such as AXL and the location of the retinal major vein could be included in the algorithm. This could provide a better detection system for glaucoma using OCT technology. Additional investigations using other OCT devices are also needed to determine device variability. ^30–32,38  

Acknowledgments

Disclosure: S. Rho, None; Y. Sung, None; T. Kang, None; N.R. Kim, None; C.Y. Kim, None