Evaluation of Retinal Nerve Fiber Layer, Optic Nerve Head, and Macular Thickness Measurements for Glaucoma Detection

Using Optical Coherence Tomography

FELIPE A. MEDEIROS, MD, LINDA M. ZANGWILL, PHD, CHRISTOPHER BOWD, PHD, ROBERTO M. VESSANI, MD, REMO SUSANNA JR, MD, AND ROBERT N. WEINREB, MD

To compare the ability of optical coherence tomography retinal nerve ber layer (RNFL), optic nerve head, and macular thickness parameters to differentiate between healthy eyes and eyes with glaucomatous visual eld loss. DESIGN: Observational case-control study. METHODS: Eighty-eight patients with glaucoma and 78 healthy subjects were included. All patients underwent ONH, RNFL thickness, and macular thickness scans with Stratus OCT during the same visit. ROC curves and sensitivities at xed specicities were calculated for each parameter. A discriminant analysis was performed to develop a linear discriminant function designed to identify and combine the best parameters. This LDF was subsequently tested on an independent sample consisting of 63 eyes of 63 subjects (27 glaucomatous and 36 healthy individuals) from a different geographic area. RESULTS: No statistically signicant difference was found between the areas under the ROC curves (AUC) for the RNFL thickness parameter with the largest AUC (inferior thickness, AUC 0.91) and the ONH parameter with largest AUC (cup/disk area ratio, AUC 0.88) (P .28). The RNFL parameter inferior thickness had a signicantly larger AUC than the macular thickness parameter with largest AUC (inferior outer macular thickness, AUC 0.81) (P .004). A combination of selected RNFL and ONH parameters resulted in the best classication function for glaucoma detection

PURPOSE:

with an AUC of 0.97 when applied to the independent sample. CONCLUSIONS: RNFL and ONH measurements had the best discriminating performance among the several Stratus OCT parameters. A combination of ONH and RNFL parameters improved the diagnostic accuracy for glaucoma detection using this instrument. (Am J Ophthalmol 2005;139:44-55. 2005 by Elsevier Inc. All rights reserved.)

Accepted for publication Aug 26, 2004. From the Hamilton Glaucoma Center and Department of Ophthalmology, University of California, San Diego, California ( F.A.M., L.M.Z., C.B., R.N.W.); and Department of Ophthalmology, University of So Paulo, So Paulo, Brazil (R.M.V., R.S. Jr.). Supported in part by the Foundation for Eye Research (F.A.M.) and NIH Grant EY11008 (L.M.Z.). Inquiries to Felipe A. Medeiros, MD, Hamilton Glaucoma Center, University of California, San Diego, 9500 Gilman Drive, La Jolla, CA 92093-0946; e-mail: fmedeiros@eyecenter.ucsd.edu

optic nerve head (ONH) and retinal nerve ber layer (RNFL) have been reported to precede the development of visual eld loss in glaucoma.13 Detection of ONH and RNFL damage is, therefore, crucial for early diagnosis of glaucoma. Recent attention has also been directed to the role of macular thickness measurements for glaucoma diagnosis. Retinal ganglion cells also are lost in the posterior pole in glaucoma,4,5 where these cells may constitute 30% to 35% of the retinal thickness in the macular region. Optical coherence tomography (OCT) is an optical imaging technique that provides high resolution and reproducible images of the RNFL that discriminate glaucomatous from healthy subjects.6 11 Although OCT has been used, for the most part, to evaluate RNFL thickness, recent improvements in the software also have made possible the evaluation of ONH topography and macular thickness for glaucoma diagnosis and follow-up. A previous investigation demonstrated that OCT ONH measurements correlate well with topographic measurements obtained by confocal scanning laser ophthalmoscopy, another imaging technique that evaluates the ONH.12 Other studies have also shown that OCT macular thickness measurements are signicantly thinner in glaucomatous compared with healthy eyes.5,1315 Although the ability of OCT ONH and
RIGHTS RESERVED. 0002-9394/05/$30.00 doi:10.1016/j.ajo.2004.08.069

HANGES IN THE STRUCTURAL APPEARANCE OF THE

44

2005 BY

ELSEVIER INC. ALL

macular thickness measurements to differentiate glaucomatous from healthy subjects has been reported to be lower than RNFL thickness parameters, no study has yet provided a comparison of these three methods in the same population. Further, it is possible that ONH and macula measurements provide complementary structural information that would increase diagnostic accuracy when combined with RNFL evaluation. The purpose of this study was to compare the ability of OCT RNFL, ONH, and macular thickness parameters to differentiate between healthy eyes and eyes with glaucomatous visual eld loss in one study population. We also investigated whether a combination of these analytical methods improved the accuracy of glaucoma diagnosis by OCT.

METHODS
THIS OBSERVATIONAL CROSS-SECTIONAL STUDY INCLUDED

166 eyes of 166 patients (88 glaucomatous patients and 78 healthy control subjects). Mean age ( SD) of glaucoma patients and healthy individuals was 68 11 years and 65 9 years, respectively (P .09; Students t test). Subjects were evaluated at the Hamilton Glaucoma Center, University of California, San Diego, from April 2002 to January 2004. These patients were included in a prospective longitudinal study designed to evaluate optic nerve structure and visual function in glaucoma (DIGSDiagnostic Innovations in Glaucoma Study). All patients who met the inclusion criteria described were enrolled in the current study. Informed consent was obtained from all participants. The University of California San Diego Human Subjects Committee approved all protocols, and the methods described adhered to the tenets of the Declaration of Helsinki. Each subject underwent a comprehensive ophthalmologic examination including review of medical history, best-corrected visual acuity, slit-lamp biomicroscopy, intraocular pressure (IOP) measurement using Goldmann applanation tonometry, gonioscopy, dilated fundoscopic examination using a 78-diopter lens, stereoscopic optic disk photography, and automated perimetry using 24-2 Swedish Interactive Threshold Algorithm (Carl Zeiss Meditec Inc., Dublin, California, USA). To be included, subjects had to have bestcorrected visual acuity of 20/40 or better, spherical refraction within 5.0 diopters and cylinder correction within 3.0 diopters, and open angles on gonioscopy. Eyes with coexisting retinal disease, uveitis, or nonglaucomatous optic neuropathy were excluded from this investigation. One eye of each patient was randomly selected for inclusion in the study. Normal control eyes had intraocular pressures of 21 mm Hg or less with no history of increased IOP and a normal visual eld result. Normal visual eld was dened as a mean deviation and pattern standard deviation within 95% condence limits, and a Glaucoma Hemield Test VOL. 139, NO. 1 EVALUATION
OF

(GHT) within normal limits. Normal control eyes also had a healthy appearance of the optic disk and RNFL (no diffuse or focal rim thinning, cupping, optic disk hemorrhage, or RNFL defects), as evaluated by clinical examination. Eyes were classied as glaucomatous if they had repeatable (two consecutive) abnormal visual eld test results, dened as a PSD outside of the 95% normal condence limits or a Glaucoma Hemield Test result outside normal limits, regardless of the appearance of the optic disk. Average MD of the glaucomatous eyes on the visual eld test nearest the imaging date was 4.96 dB. According to the Hodapp-Parrish-Anderson16 grading scale of severity of visual eld defects, 61 patients (69%) were classied as having early visual eld defects, 15 patients (17%) had moderate defects, and 12 patients (14%) had severe visual eld defects. Although the appearance of the optic disk on stereophotographs was not used as an inclusion criterion, the results of stereophotograph assessment were used for comparison with Stratus OCT ONH measurements. Simultaneous stereoscopic optic disk photographs (TRC-SS; Topcon Instrument Corp of America, Paramus, New Jersey, USA) were evaluated by two experienced graders, and each was masked to the subjects identity and to the other test results. The graders visually estimated the horizontal and vertical cup/disk ratios based on the contour of the cup. The mean value of the two graders was used as a nal grading. Subjects underwent ocular imaging with dilated pupils using the commercially available optical coherence tomograph, Stratus OCT (Carl Zeiss Meditec, Dublin, California, USA). All patients had optic nerve head, RNFL thickness, and macular thickness scans obtained during the same visit. OCT employs the principles of low-coherence interferometry and is analogous to ultrasound B-mode imaging but uses light instead of sound to acquire highresolution images of ocular structures. In brief, a lowcoherence near-infrared (840 nm) light beam is directed onto a partially reective mirror (beam splitter) that creates two light beams, a reference and a measurement beam. The measurement beam is directed onto the subjects eye and is reected from intraocular microstructures and tissues according to their distance, thickness, and different reectivity. The reference beam is reected from the reference mirror at a known, variable position. Both beams travel back to the partially reective mirror, recombine, and are transmitted to a photosensitive detector. The pattern of interference is used to provide information regarding distance and thickness of retinal structures. Bidimensional images are created by successive longitudinal scanning in transverse direction. Quality assessment of Stratus OCT scans was evaluated by an experienced examiner masked to the subjects results of the other tests. Good-quality scans had to have focused images from the ocular fundus, an adequate signal-to-noise GLAUCOMA DETECTION USING OCT 45

MEASUREMENTS

FOR

ratio (33 dB for RNFL and macula scans), and the presence of a centered circular ring around the optic disk (for RNFL scans). For macula and ONH scans, the radial scans had to be centered on the fovea and optic disk, respectively. RNFL scans were also evaluated as to the adequacy of the algorithm for detection of the RNFL. Only scans without overt algorithm failure in detecting the retinal borders were included in the study. If one type of scan was classied as unacceptable, the patient was excluded from the study. From an initial group of 189 eligible patients, 23 (12%) had unacceptable Stratus OCT scans and were excluded from further analysis.
RNFL THICKNESS MEASUREMENTS:

The fast RNFL algorithm was used to obtain RNFL thickness measurements with Stratus OCT. Three images were acquired from each subject, with each image consisting of 256 A-scans along a 3.4-mm-diameter circular ring around the optic disk. A mean image was automatically created by the Stratus OCT software. Parapapillary RNFL thickness parameters automatically calculated by existing Stratus OCT software (version 3.1) and evaluated in this study were average thickness (360-degree measure), temporal quadrant thickness (316 degrees to 45 degrees), superior quadrant thickness (46 degrees to 135 degrees), nasal quadrant thickness (136 degrees to 225 degrees), inferior quadrant thickness (226 degrees to 315 degrees), and thickness for each of 12 clock-hour positions with the 3-oclock position as nasal, 6-oclock position as inferior, 9-oclock position as temporal, and 12-oclock position as superior. Other parameters evaluated included superior maximum (Smaxthickest point in the superior quadrant), inferior maximum (Imaxthickest point in the inferior quadrant), and relational parameters such as Imax/Smax, Smax/Imax, Imax/ Tavg (inferior maximum/temporal quadrant thickness), Smax/Navg (Superior maximum/Nasal quadrant thickness), and Max-min (difference between the thickest and thinnest points along the measurement circle).

below this line are dened as the disk cup and above this line as the neuroretinal rim. ONH parameters automatically calculated by existing Stratus OCT software (version 3.1) and evaluated in this study were vertically integrated rim area (total volume of rim tissue calculated by multiplying the average of individual rim areas times the circumference of the disk), horizontally integrated rim width (estimate of total rim area calculated by multiplying the average of individual rim widths times the circumference of the disk), disk area, cup area, rim area, cup/disk area ratio (ratio of cup area to disk area), horizontal cup/disk ratio (ratio of the longest horizontal line across the cup to the longest horizontal line across the disk), and vertical cup/disk ratio (ratio of the longest vertical line across the cup to the longest vertical line across the disk).
MACULAR THICKNESS MEASUREMENTS: The Fast Macular Thickness protocol was used to obtain macular thickness measurements with Stratus OCT. The macular scans consist of six radial scans in a spokelike pattern centered on the fovea with each radial scan spaced 30 degrees from one to another. To ll the gaps between the scans, OCT uses interpolation. Stratus OCT software calculates retinal thickness as the distance between the vitreoretinal interface and the junction between the inner and outer segment of photoreceptors, which is just above the retinal pigment epithelium. Three concentric circles divide the macular thickness map into three zones: fovea, inner macula, and outer macula. The inner and outer zones are further divided in four quadrants by two diagonal lines. Thus, a total of nine areas (fovea, superior outer, superior inner, inferior outer, inferior inner, temporal outer, temporal inner, nasal outer, and nasal inner) are available for analysis. For this study, concentric circles with default diameters of 1 mm, 3 mm, and 6 mm were used to divide the macular thickness map. Macular thickness parameters automatically calculated by existing Stratus OCT software (version 3.1) and evaluated in this study were foveal thickness, superior outer macular thickness, inferior outer macular thickness (IOM), temporal outer macular thickness, nasal outer macular thickness, superior inner macular thickness, inferior inner macular thickness, temporal inner macular thickness, and nasal inner macular thickness. Average macular thickness was calculated as the weighted average of the sectoral macular thickness measurements excluding the fovea. STATISTICAL ANALYSIS:

OPTIC NERVE HEAD MEASUREMENTS: The Fast Optical Disk scanning protocol was used to obtain ONH measurements with Stratus OCT. The ONH scan consists of six radial scans in a spoke like pattern centered on the ONH. The OCT interpolates between the scans to provide measurements throughout the ONH. In optic nerve head scans, the device automatically determines the disk margin as the end of the retinal pigment epithelium/choriocapillaris layer. One can manually adjust the demarcation of the edge of the retinal pigment epithelium to improve the outlining of the disk margin. However, to minimize subjectivity, the automatically determined default disk margin was used in this study. A straight line connects the edges of the retinal pigment epithelium/choriocapillaris, and a parallel line is constructed 150 m anteriorly. Structures

Student t tests were used to evaluate optic nerve head, RNFL thickness, and macular thickness measurement differences between glaucomatous and healthy eyes. Results of statistical signicance were also provided after Bonferronis correction based on the number of comparisons within each analysis. Pearsons OPHTHALMOLOGY JANUARY 2005

46

AMERICAN JOURNAL

OF

correlation coefcients were used to assess the correlations between continuous variables. Bland and Altman plots were constructed to assess agreement between optic disk stereophotograph assessment and Stratus OCT ONH measurements.17,18 Receiver operating characteristic (ROC) curves were used to describe the ability to differentiate glaucomatous from healthy eyes of each Stratus OCT software-provided parameter. The ROC curve shows the trade-off between sensitivity and 1-specicity. An area under the ROC curve of 1.0 represents perfect discrimination, whereas an area of 0.5 represents chance discrimination. The method of DeLong and associates19 was used to compare areas under the ROC curve. A discriminant analysis was performed to develop a classication function (linear discriminant function LDF) designed to identify and combine the best Stratus OCT measures to differentiate glaucomatous from normal eyes. A principal component analysis was initially performed to select a reduced set of variables that accounted for most of the variance of the original data set.20,21 The central idea of PCA is to reduce the dimensionality of a data set consisting of a large number of interrelated variables while retaining as much as possible of the variation present in the data set.20 Thirteen principal components explaining 91% of the variance of the data set were selected according to Jollifes criterion.22 After varimax rotation, the variables with highest loadings in each component were selected for further analysis and possible inclusion in the discriminant function.21 LDFs were then constructed using all possible subsets from the reduced set of variables. The bias-corrected area under the ROC curve was used as a measure of the performance of each LDF. Bias correction was performed using 10-fold cross-validation. The model that maximized the bias-corrected ROC curve area was selected as best. To evaluate model stability, 1,000 bootstrap random samples were drawn with replacement from the original sample. The model selection procedure was then applied to each of the 1,000 bootstrap resamples, and the best model was selected in each resample. The frequency of inclusion of the variables in the selected models was reported. Important variables should be included in most of the replications, and the inclusion frequencies may be used as a criterion for the importance of a variable.23 To determine generalizability of the derived Stratus OCT LDF to new patients, the nal LDF was applied to an external independent sample from a different geographic area.24 Patients in this sample were not used in any of the steps of model development. This validation set included 63 eyes of 63 subjects (27 glaucomatous and 36 normals) evaluated at the Glaucoma Center of the University of So Paulo, Brazil. Informed consent was obtained from all subjects, and the appropriate regulatory and ethics committees VOL. 139, NO. 1 EVALUATION
OF

approved all protocols. The inclusion and exclusion criteria were identical to those used for the derivation set. Mean age (SD) of glaucoma patients and healthy subjects was 59 15 years and 56 10 years, respectively (P .36). Average MD of the glaucomatous eyes on the visual eld test nearest the imaging date was 6.58 dB. According to the Hodapp-Parrish-Anderson16 grading scale of severity of visual eld defects, 16 patients (59%) were classied as having early visual eld defects, 4 patients (15%) had moderate defects, and 7 patients (26%) had severe visual eld defects. A P value less than .05 was considered statistically signicant. Statistical analyses were performed using software SPSS v.10.0 (SPSS Inc., Chicago, Illinois, USA) and S-PLUS 2000 (Mathsoft Inc., Seattle, Washington, USA).

RESULTS
RNFL THICKNESS MEASUREMENTS:

Table 1 shows mean values of Stratus OCT RNFL parameters in glaucomatous and normal eyes. After Bonferronis correction ( 0.002; 25 comparisons), statistical signicant differences were found for all parameters except thickness at 9-oclock, Imax/Smax, Smax/Tavg, and Smax/Navg. Table 1 also shows ROC curve areas and sensitivities at xed specicities. The 3 Stratus OCT RNFL parameters with largest areas under the ROC curves were inferior thickness (0.91), average thickness (0.91), and inferior maximum (0.90). There were no statistically signicant differences in the ROC curve areas for these parameters (P .05 for all comparisons).

EXTERNAL VALIDATION ANALYSIS:

Table 2 shows mean values of Stratus OCT ONH parameters in glaucomatous and normal eyes. After Bonferronis correction ( 0.006; 8 comparisons), statistical signicant differences were found for all parameters except disk area. Table 2 also shows ROC curve areas and sensitivities at xed specicities. The 3 Stratus OCT ONH parameters with largest areas under the ROC curves were cup/disk area ratio (0.88), vertical cup/disk ratio (0.88), and HIRW (0.88). Vertical and horizontal cup/disk ratio measurements obtained by the Stratus OCT also were compared with those obtained by stereophotograph assessment. There was no statistically signicant difference between mean Stratus OCT vertical cup/disk ratio and stereophotograph vertical cup/disk ratio [0.59 0.20 vs 0.58 0.24; P .59, paired t test). There was a statistically signicant correlation between the two measurements (r .87; P .001). Figure 1 shows a Bland and Altman plot of the agreement in vertical cup/disk ratio between Stratus OCT and stereophotograph assessment. The difference (stereophotograph vertical cup/disk ratioStratus OCT vertical cup/disk ratio) was plotted against the average of the two measurements. Although no signicant xed bias was observed, a GLAUCOMA DETECTION USING OCT 47

OPTIC NERVE HEAD MEASUREMENTS:

MEASUREMENTS

FOR

TABLE 1. Mean ( SD) Values of Stratus OCT Retinal Nerve Fiber Layer Parameters With Areas Under the Receiver Operating Characteristic (ROC) Curves and Sensitivities (Sn) at Fixed Specicities (Sp)
Glaucoma (n 88) Normal (n 78) Sn/Sp (Sp 95%) Sn/Sp (Sp 80%)

Parameter

P*

ROC (SE)

Inferior thickness (m) Average thickness (m) Imax (m) Thickness at 6-oclock (m) Thickness at 7-oclock (m) Max-min (m) Superior thickness (m) Smax (m) Thickness at 11-oclock (m) Imax/Tavg Thickness at 5-oclock (m) Nasal thickness (m) Thickness at 3-oclock (m) Thickness at 4-oclock (m) Thickness at 2-oclock (m) Thickness at 1-oclock (m) Thickness at 12-oclock (m) Smax/Tavg Thickness at 10-oclock (m) Imax/Smax Thickness at 8-oclock (m) Temporal thickness (m) Thickness at 9-oclock (m) Smax/Navg Smax/Imax

84.6 23.2 74.2 13.3 112.3 30.5 90.2 29.6 87.7 32.0 94.0 25.1 92.2 22.5 120.4 26.6 96.1 29.5 1.92 0.54 75.8 23.5 59.5 15.8 49.5 12.5 59.1 16.8 69.9 23.1 86.0 26.0 94.4 27.4 2.08 0.61 69.0 20.6 0.97 0.31 60.7 17.9 60.5 15.2 51.8 13.6 2.16 1.01 1.15 0.39

123.8 16.5 96.5 9.90 159.3 21.1 133.8 24.5 132.4 21.2 124.5 18.6 118.6 16.0 149.8 19.5 124.9 19.0 2.43 0.51 105.2 25.5 76.1 19.6 62.2 19.2 75.2 19.9 91.1 25.2 109.2 21.9 121.8 26.7 2.28 0.48 80.9 18.4 1.08 0.17 69.7 16.2 67.6 13.1 52.4 11.2 2.10 0.63 0.95 0.15

.0001 .0001 .0001 .0001 .0001 .0001 .0001 .0001 .0001 .0001 .0001 .0001 .0001 .0001 .0001 .0001 .0001 .019 .0001 .004 .0001 .002 .760 .634 .0001

0.91 (0.02) 0.91 (0.02) 0.90 (0.03) 0.87 (0.03) 0.87 (0.03) 0.85 (0.03) 0.83 (0.03) 0.81 (0.03) 0.78 (0.04) 0.76 (0.04) 0.80 (0.03) 0.76 (0.04) 0.70 (0.04) 0.74 (0.04) 0.74 (0.04) 0.75 (0.04) 0.76 (0.04) 0.64 (0.04) 0.68 (0.04) 0.65 (0.04) 0.67 (0.04) 0.65 (0.04) 0.51 (0.05) 0.48 (0.05) 0.35 (0.04)

65/96 71/95 63/95 56/95 64/95 55/95 52/96 50/95 47/95 38/95 34/95 13/95 10/96 22/95 21/95 39/95 31/95 27/96 26/96 33/95 24/95 22/95 11/95 5/95 7/95

89/80 86/80 85/80 78/80 81/80 81/80 73/81 68/80 68/80 50/81 63/80 61/80 34/81 55/81 52/80 50/80 58/80 43/80 46/81 44/81 49/81 38/81 22/82 17/80 16/80

TABLE 2. Mean ( SD) Values of Stratus OCT Optic Nerve Head Parameters With Areas Under the Receiver Operating Characteristic (ROC) Curves and Sensitivities (Sn) at Fixed Specicities (Sp)
Glaucoma (n 88) Normal (n 78) Sn/Sp (Sp 95%) Sn/Sp (Sp 80%)

Parameter

P*

ROC (SE)

Cup/disk:area ratio Vertical C/D ratio HIRW (mm2) Rim area (mm2) VIRA (mm2) Horizontal C/D ratio Cup area (mm2) Disk area (mm2)

0.55 0.19 0.70 0.14 1.23 0.28 1.02 0.42 0.17 0.16 0.76 0.16 1.31 0.60 2.34 0.47

0.26 0.14 0.45 0.16 1.69 0.28 1.74 0.45 0.48 0.34 0.50 0.19 0.61 0.39 2.35 0.51

.0001 .0001 .0001 .0001 .0001 .0001 .0001 .847

0.88 (0.03) 0.88 (0.03) 0.88 (0.03) 0.88 (0.03) 0.87 (0.03) 0.86 (0.03) 0.84 (0.03) 0.51 (0.05)

69/95 65/95 55/95 51/95 58/95 59/95 50/95 6/95

80/80 81/80 77/80 81/80 82/80 74/80 74/80 19/80

*HIRW horizontal integrated rim width; VIRA vertical integrated rim area.

statistically signicant proportional bias was detected (r .34; P .001). For lower values of vertical cup/disk ratio, Stratus OCT measurements tended to be higher than stereophotograph measurements; whereas for higher values of vertical cup/disk ratio, Stratus OCT measurements tended to be lower than stereophotograph measurements. For the horizontal cup/disk ratio, Stratus OCT measurements were signicantly larger than stereophotograph-based subjective assessment (0.64 0.22 vs 0.54 0.22; P .001; 48 AMERICAN JOURNAL
OF

paired t test). A signicant correlation was obtained between the two measures (r .84; P .001). The Bland and Altman plot (Figure 2) showed the presence of xed bias, but no proportional bias was detected (r .01; P .87).
THICKNESS MEASUREMENTS: Table 3 shows mean values of Stratus OCT macular thickness parameters in glaucomatous and normal eyes. After Bonferronis correction ( 0.005; 10 comparisons), MACULAR

OPHTHALMOLOGY

JANUARY 2005

FIGURE 1. Bland and Altman plot of the agreement on vertical cup/disk ratio between stereophotograph assessment and Stratus OCT measurements. The difference (stereophotograph Stratus OCT) is plotted vs the average (stereophotograph Stratus OCT)/2. The existence of proportional bias is indicated by the signicant slope of the line regressing the difference on the average (r .34; P < .001). The regression line is shown with 95% individual condence limits bands.

statistically signicant differences were found for all parameters except NIM thickness and foveal thickness. Table 3 also shows ROC curve areas and sensitivities at xed specicities. The 3 Stratus OCT macular thickness parameters with the largest areas under the ROC curves were IOM thickness (0.81), Macular thickness average (0.75), and TOM thickness (0.75). There were no statistically signicant differences in the ROC curve areas for these parameters (P .05 for all comparisons).
COMPARISON OF RNFL, ONH, AND MACULAR THICKNESS MEASUREMENTS: No statistically signicant differ-

ence was found between the areas under the ROC curves (AUC) for the RNFL thickness parameter with largest AUC (inferior thickness; AUC 0.91) and the ONH parameter with largest AUC (cup/disk area ratio; AUC 0.88) (P .28). The RNFL parameter inferior thickness had a signicantly larger AUC than the macular thickness parameter with largest AUC (IOM; AUC 0.81) (P .004). The AUC of the ONH parameter cup/disk area ratio was higher than that of the macular thickness parameter IOM, but the difference did not reach statistical signicance (P .09). VOL. 139, NO. 1 EVALUATION
OF

We evaluated whether a combination of Stratus OCT measures improved the discrimination between glaucoma and healthy subjects. The best discriminant function resulting from the combination of Stratus OCT parameters had the following formula: LDF 3.023 (2.659 cup/disk area ratio) (0.035 average thickness) (0.013 thickness at 7-oclock) (0.011 thickness at 11-oclock) (0.031 thickness at 9-oclock) This LDF had an AUC of 0.97 (SE 0.01) with an estimated bias of 0.011. The AUC of the LDF was signicantly larger than that of the single Stratus OCT software-provided parameter with largest AUC (inferior thickness) (0.97 vs 0.91; P .012). Figure 3 shows the ROC curves for the two parameters of RNFL, ONH, and macular analyses with largest AUCs and also for the Stratus OCT LDF. For specicity at 95%, the LDF had a sensitivity of 90% (cut-off of 0.284). For specicity at 81%, the LDF had a sensitivity of 94% (cut-off of 0.683). Although there were several LDFs that were competitive with the above formula, the variables included in the nal LDF had the highest frequencies of inclusion in the models selected in the bootstrap samples. Figure 4 illusGLAUCOMA DETECTION USING OCT 49

MEASUREMENTS

FOR

FIGURE 2. Bland and Altman plot of the agreement on horizontal cup/disk ratio between stereophotograph assessment and Stratus OCT measurements. The difference (stereophotograph Stratus OCT) is plotted vs the average (stereophotograph Stratus OCT)/2. The continuous line represents the mean difference and 95% limits of agreement. The existence of xed bias is indicated by the signicant deviation from zero of the mean difference between stereophoto and Stratus OCT measurements.

TABLE 3. Mean ( SD) Values of Stratus OCT Macular Thickness Parameters With Areas Under the Receiver Operating Characteristic (ROC) Curves and Sensitivities (Sn) at Fixed Specicities (Sp)
Glaucoma (n 88) Normal (n 78) Sn/Sp (Sp 95%) Sn/Sp (Sp 80%)

Parameter

P*

ROC (SE)

Inferior outer macula (m) Macula average (m) Temporal outer macula (m) Superior outer macula (m) Nasal outer macula (m) Temporal inner macula (m) Inferior inner macula (m) Superior inner macula (m) Nasal inner macula (m) Fovea (m)

205 16 216 13 202 15 220 18 235 18 246 18 254 19 259 21 264 21 202 25

224 16 231 15 218 18 236 17 247 18 257 19 265 19 269 19 269 21 201 28

.0001 .0001 .0001 .0001 .0001 .0001 .0001 .002 .12 .83

0.81 (0.03) 0.75 (0.04) 0.75 (0.04) 0.73 (0.04) 0.68 (0.04) 0.67 (0.04) 0.65 (0.04) 0.63 (0.04) 0.55 (0.05) 0.47 (0.05)

47/95 35/95 32/95 36/95 21/95 22/95 26/96 18/95 14/95 6/95

73/82 50/80 51/80 48/80 39/80 42/80 34/80 34/80 31/80 18/80

*Macula average thickness is calculated from the weighted average of all sectors excluding the fovea. SE standard error.

trates the frequency of inclusion of the variables in the nal models selected in the 1,000 bootstrap replications. The variables included in the nal LDF were selected in 59% to 88% of the nal models, whereas the other variables were included in no more than 32% of the models. The full model containing all 13 variables had a bias-corrected AUC of 0.96. 50 AMERICAN JOURNAL
OF

When applied to the independent validation sample (n 63), the LDF had an AUC of 0.97 (SE 0.02). Figure 5 shows the ROC curves for the LDF when applied to the independent sample and for the two parameters with largest AUC for RNFL (average thickness; AUC 0.93 and RNFL thickness at 6-oclock; AUC 0.92), ONH (VIRA; AUC 0.92 and cup/disk OPHTHALMOLOGY JANUARY 2005

FIGURE 3. Receiver operating characteristic (ROC) curves of the two parameters with largest areas under the ROC curves from the Stratus OCT retinal nerve ber layer (inferior thickness and average thickness), optic nerve head (cup/disk area ratio and vertical cup/disk ratio), macular thickness analysis (inferior outer macular thickness [IOM] and macula average), and of the linear discriminant function (LDF) obtained from the combination of selected parameters.

area ratio; AUC 0.91), and macular thickness analyses (IOM; AUC 0.79 and SOM; AUC 0.78) on the independent sample.

DISCUSSION
THE ANALYSIS OF STRATUS OCT SOFTWARE-PROVIDED PA-

rameters showed that parapapillary RNFL measures and ONH topographic parameters had the highest power to discriminate glaucomatous from healthy eyes. Areas under the ROC curves and sensitivities at moderate and high specicities were similar for the best parameters from each of these two methods of analysis. We also found that a combination of selected RNFL and ONH parameters in a linear discriminant function resulted in further improvement of the diagnostic accuracy of OCT. The ROC curve areas for the Stratus OCT RNFL measurements were similar to those obtained with the previous versions of this technology. The areas under the ROC curves for the earlier OCT models have been VOL. 139, NO. 1 EVALUATION
OF

reported to range from 0.79 to 0.94, depending on the parameter and characteristics of the population evaluated.9 11,15,2527 In studies evaluating the diagnostic ability of several OCT parameters, the RNFL thickness in the inferior region often had the best performance to discriminate healthy eyes from eyes with early to moderate glaucoma with sensitivities between 67% and 79% for specicities 90%.9,11,26 In our study, the parameter inferior thickness also had the highest area under the ROC curve, with sensitivity of 65% for specicity at 95%. The parameter average thickness also had a similar performance. The RNFL thicknesses at 7-oclock and at 11-oclock were included as RNFL parameters in our discriminant function, along with RNFL thickness at 9-oclock and average thickness. The high discriminating ability of RNFL thickness at 7- and 11-oclock is readily understandable, as these variables represent the inferior temporal and superior temporal sectors of the optic disk, respectively, which are the sectors most commonly affected in glaucoma.28 The inclusion of these variables may also be related GLAUCOMA DETECTION USING OCT 51

MEASUREMENTS

FOR

FIGURE 4. Frequency of inclusion of the 13 different variables in the best models selected in each of the 1,000 bootstrap replications.

to the detection of localized nerve ber layer defects that are most commonly seen in these sectors.29 The average RNFL thickness is a measurement of global thickness of the RNFL and, therefore, is presumably important in the differentiation of glaucoma from healthy eyes. The inclusion of the RNFL thickness at 9-oclock in the discriminant function is more difcult to explain. This variable corresponds to the thickness at the most temporal sector, in the region of the papilomacular bundle. No signicant difference was observed in the mean RNFL thickness values at 9-oclock between glaucomatous and healthy eyes. This is in agreement with previous studies demonstrating that the RNFL is usually preserved in the region of the papilomacular bundle until late in the course of the disease. It is well known that RNFL thickness can vary widely among healthy subjects limiting the usefulness of absolute thickness values to separate glaucomatous from healthy subjects. Conversely, the evaluation of the modulation of the RNFL thickness may provide a useful tool for detection of relative loss of nerve bers in glaucoma.30 The modulation represents the difference between the thickest and the thinnest parts of the RNFL around the optic disk. The RNFL thickness difference between the 9-oclock sector and the 7- and 11-oclock sectors could provide an indication of the modulation of RNFL thickness around the optic disk. The negative sign of this variable in the LDF formula indicates that, all other variables being equal, 52 AMERICAN JOURNAL
OF

a subject with thicker RNFL at 9-oclock (and lower modulation) will have a higher chance of having glaucomatous visual eld loss than a subject with thinner RNFL in this sector (and higher modulation). In a recent study, Nouri-Mahdavi27 and associates found that the thickness at 7-, 10-, and 11-oclock provided the best combination of OCT RNFL parameters to discriminate patients with glaucomatous visual eld loss from healthy subjects. In their study, the thickness at 10-oclock had a positive correlation with the presence of glaucomatous visual eld loss, that is, higher values indicated a higher chance of glaucoma. Thus, it is possible that the RNFL thickness at 10-oclock in their discriminant function had the same role as the RNFL thickness at 9-oclock in our study. This is supported by the high correlation existing between these two variables (r .77, P .001 in our study). Stratus OCT ONH parameters also performed well for glaucoma detection in our study. Areas under the ROC curves were similar for all ONH parameters except disk area. The parameter cup/disk area ratio had the highest sensitivity with specicity at 95%, and this parameter was also included in the nal LDF developed in our study. Highly signicant correlations were found between Stratus OCT and stereophotograph assessment of vertical and horizontal cup/disk ratios. However, important disagreements were detected between these two methods when Bland and Altman plots (Figures 1 and 2) were analyzed, OPHTHALMOLOGY JANUARY 2005

FIGURE 5. Receiver operating characteristic (ROC) curve of the linear discriminant function when applied to the independent sample and of the two parameters with the largest areas under the ROC curves for RNFL (average thickness and RNFL thickness at 6 oclock), ONH (vertically integrated rim area [VIRA] and cup/disk area ratio), and macular thickness analyses (inferior outer macula thickness [IOM] and superior outer macula thickness [SOM]) on the independent sample.

indicating the low utility of correlation coefcients to assess agreement between methods of measurement, an issue that has already been extensively acknowledged in the literature.17,18,31 The limited agreement with stereophotographic assessment does not preclude the use of Stratus OCT ONH parameters for glaucoma diagnosis. Stereophotographic assessment of cup/disk ratio is a subjective measure and has a large interobserver variability.32 In contrast, Stratus OCT ONH assessment provides objective measures of optic disk topography using an automated process of optic disk edge detection and cup delimitation, and a recent study has found these measures to be highly reproducible.33 In fact, the discriminating abilities of Stratus OCT ONH parameters were similar to those of the best RNFL measures, and the combination of an ONH parameter (cup/disk area ratio) with RNFL thickness parameters resulted in the best discriminant function for glaucoma detection in our study. The utility of the topographical evaluation of the ONH with OCT for glaucoma diagnosis still needs further VOL. 139, NO. 1 EVALUATION
OF

evaluation. As the automatic algorithm for detection of the disk margin is based on the determination of the end of the retinal pigment epithelium/choriocapillaris layer, it is possible that the evaluation of the disk margin will be inuenced by changes in these layers such as with progressive parapapillary atrophy in glaucoma.34 A manual algorithm for disk margin determination is also available. However, we did not use the manual algorithm to avoid introducing a subjective component to our analysis. Furthermore, the manual and automatic algorithms have been demonstrated to have comparable performance in a previous work.12 Stratus OCT macular thickness measurements had a limited ability to differentiate glaucomatous from healthy eyes in our investigation. Although mean macular thickness parameters of glaucomatous eyes were signicantly lower than that of normal control eyes, the macular thickness parameter with largest AUC had a sensitivity of only 47% for a specicity set at 95%, with an AUC of 0.81. This agrees with a previous investigation where a maxiGLAUCOMA DETECTION USING OCT 53

MEASUREMENTS

FOR

mum ROC curve area of 0.77 for macular thickness parameters was obtained for the discrimination between early glaucoma and normal subjects, whereas peripapillary RNFL thickness parameters had maximum ROC curve area of 0.94 in the same situation.15 In contrast to RNFL and ONH measures, macular thickness parameters were not included in the nal LDF developed in our study. The low frequency of inclusion of macular thickness parameters when the model building process was replicated in the bootstrap samples provides an indication that these variables had at most a weak inuence on the discrimination between glaucomatous and healthy eyes when several Stratus OCT parameters were combined. However, it should be noted that we have only investigated the current macular thickness parameters provided by the standard Stratus OCT printout, and it is possible that advances in the software designed to extract data from the macular area would improve detection of retinal ganglion cell loss in the posterior pole. Our study has limitations. Although the generalizability of the discriminant function combining several Stratus OCT parameters was good when applied to an independent population, the size of the validation sample was relatively small. However, the sample size of the validation group provided 79% power to detect a decrease in performance of 0.05 in the ROC curve area in the validation sample compared with the performance in the development sample.35 The validation sample contained a higher proportion of moderate and advanced cases compared with the development sample. However, even when patients with moderate and advanced glaucomatous visual eld loss were excluded from the validation sample, the LDF still performed well, with an area under the ROC curve of 0.96. Another limitation of our study was that the inclusion criteria for normal subjects required a normal optic nerve appearance at the clinical examination. This was required to avoid the inclusion of subjects with glaucomatous optic neuropathy but normal visual elds in the control group. It might be argued that these inclusion criteria could have overestimated the diagnostic accuracy of OCT parameters, especially of ONH parameters. However, this is a limitation common to case-control studies of this type, and no practical solution to this problem is available at this time. In conclusion, RNFL and ONH measurements had the best performance for glaucoma detection among the several Stratus OCT parameters examined in our study. A combination of ONH and RNFL parameters seems to be promising for glaucoma diagnosis using OCT.

REFERENCES
1. Harwerth RS, Carter-Dawson L, Shen F, Smith EL III, Crawford ML. Ganglion cell losses underlying visual eld defects from experimental glaucoma. Invest Ophthalmol Vis Sci 1999;40:22422250.

2. Sommer A, Katz J, Quigley HA, et al. Clinically detectable nerve ber atrophy precedes the onset of glaucomatous eld loss. Arch Ophthalmol 1991;109:77 83. 3. Kass MA, Heuer DK, Higginbotham EJ, et al. The Ocular Hypertension Treatment Study: a randomized trial determines that topical ocular hypotensive medication delays or prevents the onset of primary open-angle glaucoma. Arch Ophthalmol 2002;120:701713; discussion 829 830. 4. Zeimer R, Asrani S, Zou S, Quigley H, Jampel H. Quantitative detection of glaucomatous damage at the posterior pole by retinal thickness mapping. A pilot study. Ophthalmology 1998;105:224 231. 5. Greeneld DS, Bagga H, Knighton RW. Macular thickness changes in glaucomatous optic neuropathy detected using optical coherence tomography. Arch Ophthalmol 2003;121: 41 46. 6. Schuman JS, Pedut-Kloizman T, Hertzmark E, et al. Reproducibility of nerve ber layer thickness measurements using optical coherence tomography. Ophthalmology 1996;103: 1889 1898. 7. Schuman JS, Hee MR, Puliato CA, et al. Quantication of nerve ber layer thickness in normal and glaucomatous eyes using optical coherence tomography. Arch Ophthalmol 1995;113:586 596. 8. Bowd C, Weinreb RN, Williams JM, Zangwill LM. The retinal nerve ber layer thickness in ocular hypertensive, normal, and glaucomatous eyes with optical coherence tomography. Arch Ophthalmol 2000;118:2226. 9. Zangwill LM, Bowd C, Berry CC, et al. Discriminating between normal and glaucomatous eyes using the Heidelberg Retina Tomograph, GDx Nerve Fiber Analyzer, and Optical Coherence Tomograph. Arch Ophthalmol 2001;119:985 993. 10. Williams ZY, Schuman JS, Gamell L, et al. Optical coherence tomography measurement of nerve ber layer thickness and the likelihood of a visual eld defect. Am J Ophthalmol 2002;134:538 546. 11. Bowd C, Zangwill LM, Berry CC, et al. Detecting early glaucoma by assessment of retinal nerve ber layer thickness and visual function. Invest Ophthalmol Vis Sci 2001;42: 19932003. 12. Schuman JS, Wollstein G, Farra T, et al. Comparison of optic nerve head measurements obtained by optical coherence tomography and confocal scanning laser ophthalmoscopy. Am J Ophthalmol 2003;135:504 512. 13. Giovannini A, Amato G, Mariotti C. The macular thickness and volume in glaucoma: an analysis in normal and glaucomatous eyes using OCT. Acta Ophthalmol Scand Suppl 2002;236:34 36. 14. Lederer DE, Schuman JS, Hertzmark E, et al. Analysis of macular volume in normal and glaucomatous eyes using optical coherence tomography. Am J Ophthalmol 2003;135: 838 843. 15. Guedes V, Schuman JS, Hertzmark E, et al. Optical coherence tomography measurement of macular and nerve ber layer thickness in normal and glaucomatous human eyes. Ophthalmology 2003;110:177189. 16. Hodapp E, Parrish RK, II Anderson DR. Clinical decisions in glaucoma. St Louis: Mosby-Year Book, 1993.
OF

54

AMERICAN JOURNAL

OPHTHALMOLOGY

JANUARY 2005

17. Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986;1:307310. 18. Bland JM, Altman DG. Measuring agreement in method comparison studies. Stat Methods Med Res 1999;8:135160. 19. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 1988;44:837 845. 20. Jolliffe IT. Principal component analysis. New York: SpringerVerlag, 2002. 21. Daling JR, Tamura H. Use of orthogonal factors for selection of variables in a regression equationan illustration. Appl Stat 1970;19:260 268. 22. Jolliffe IT. Discarding variables in a principal component analysis. Appl Stat 1972;21:160 173. 23. Sauerbrei W. The use of resampling methods to simplify regression models in medical statistics. Appl Statist 1999;48: 313329. 24. Bleeker SE, Moll HA, Steyerberg EW, et al. External validation is necessary in prediction research: a clinical example. J Clin Epidemiol 2003;56:826 832. 25. Greaney MJ, Hoffman DC, Garway-Heath DF, Nakla M, Coleman AL, Caprioli J. Comparison of optic nerve imaging methods to distinguish normal eyes from those with glaucoma. Invest Ophthalmol Vis Sci 2002;43:140 145. 26. Kanamori A, Nakamura M, Escano MF, Seya R, Maeda H, Negi A. Evaluation of the glaucomatous damage on retinal nerve ber layer thickness measured by optical coherence tomography. Am J Ophthalmol 2003;135:513520. 27. Nouri-Mahdavi K, Hoffman D, Tannenbaum DP, Law SK, Caprioli J. Identifying early glaucoma with optical

28.

29.

30.

31.

32.

33.

34.

35.

coherence tomography. Am J Ophthalmol 2004;137: 228 235. Jonas JB, Gusek GC, Naumann GO. Optic disc morphometry in chronic primary open-angle glaucoma. I. Morphometric intrapapillary characteristics. Graefes Arch Clin Exp Ophthalmol 1988;226:522530. Jonas JB, Schiro D. Localised wedge shaped defects of the retinal nerve bre layer in glaucoma. Br J Ophthalmol 1994;78:285290. Xu L, Chen PP, Chen YY, Takahashi Y, Wang L, Mills RP. Quantitative nerve ber layer measurement using scanning laser polarimetry and modulation parameters in the detection of glaucoma. J Glaucoma 1998;7:270 277. Ludbrook J. Statistical techniques for comparing measurers and methods of measurement: a critical review. Clin Exp Pharmacol Physiol 2002;29:527536. Tielsch JM, Katz J, Quigley HA, Miller NR, Sommer A. Intraobserver and interobserver agreement in measurement of optic disc characteristics. Ophthalmology 1988;95:350 356. Paunescu LA, Schuman JS, Price LL, et al. Reproducibility of nerve ber thickness, macular thickness, and optic nerve head measurements using Stratus OCT. Invest Ophthalmol Vis Sci 2004;45:1716 1724. Lai E, Wollstein G, Price LL, et al. Optical coherence tomography disc assessment in optic nerves with peripapillary atrophy. Ophthalmic Surg Lasers Imaging 2003; 34:498 504. Steyerberg EW, Bleeker SE, Moll HA, Grobbee DE, Moons KG. Internal and external validation of predictive models: a simulation study of bias and precision in small samples. J Clin Epidemiol 2003;56:441 447.

VOL. 139, NO. 1

EVALUATION

OF

MEASUREMENTS

FOR

GLAUCOMA DETECTION USING OCT

55