Methods

Twenty-seven eyes of 14 healthy subjects from the University of Pittsburgh Medical Center Eye Center were enrolled. Eyes with no history or evidence of intraocular surgery, retinal disease or glaucoma, refractive error less than 8 D and normal-appearing ONHs qualified as normal healthy subjects.

All eyes showed normal comprehensive ocular examination and automated perimetry with glaucoma hemifield test (GHT) within normal limits (Humphrey Field Analyzer, Carl Zeiss Meditec). The peripapillary region was scanned on all eyes using conventional OCT (Stratus OCT) and SD-OCT (Cirrus) at the same visit. All scans were performed through undilated pupils. Institutional review board and ethics committee approval were obtained for the study, and informed consent was obtained from all subjects. This study followed the tenets of the Declaration of Helsinki and was conducted in compliance with the Health Insurance Portability and Accountability Act.

Image acquisition

Stratus OCT

Circular scans centred on the ONH were obtained using the “Fast RNFL Scan” pattern, which performs three 3.4 mm diameter circular scans (with 256 A-scan) around the ONH in a rapid succession (1.92 s). Three fast RNFL scans were obtained from each eye in a single session by one operator. The scan circle was centred on the ONH manually. Images with a signal strength (SS) less than 6 were discarded as poor-quality images as recommended by the manufacturer. RNFL thickness was measured using the Stratus OCT system (software version 5.0). Segmentation failure was defined as obvious deviation of the segmented inner and/or outer RNFL borders from the subjectively perceived borders. Consecutive 15% or cumulative 20% segmentation failure within a given image was considered to be of poor analysis quality and discarded.

Cirrus HD-OCT

3D cube OCT data were obtained using the “Optic Disc Cube 200×200 Scan” pattern, which performed raster scanning in a 6×6 mm square centred on the ONH (total scan time was 1.48 s) consisting of 200 frames of horizontal linear B-scans with 200 A-scan lines per B-scan. Images with SS less than 8 were discarded as poor-quality images. This cut-off is different from Stratus SS because, despite both representing signal strength, they are different measurements. There is no validated consensus about the ideal cut-off for Cirrus SS, but the manufacturer recommends that SS 8 or above should be considered acceptable images. In addition to SS, images with detectable eye motion (larger than one vessel diameter or a major distortion of the ONH) in OCT fundus images were also discarded (fig 2B). Analysis quality criteria were the same as Stratus OCT and based on 3.4 mm circular virtual OCT scan images. Three Optic Disc Cube scans were obtained from each eye in a single session by a single operator (KRS).

RNFL thickness measurements

TD-OCT

Global mean, four quadrants, and 12 clock hour RNFL thickness were measured for each of three fast RNFL scans.

SD-OCT

A virtual circular scan with 3.4 mm diameter (256 sampling along the circle) was automatically resampled on each 3D OCT cube data by Cirrus system software after determining the ONH centre. Two methods for determining the ONH centre were employed on Cirrus images:

• Centred Each Time (CET) method. The ONH centre was detected automatically by the Cirrus system separately for each image. RNFL thickness parameters (global, four quadrants and 12 clock hours) were measured along the virtual circle for each of the three Optic Disc Cube scans (fig 3). This is the method currently implemented in the Cirrus system software.

• Centred Once (CO) method. The ONH centre was automatically detected on the first data cube by the Cirrus system software as described above, then this ONH centre location was imported to the two other data cube sets by registering the OCT fundus image of the first data cube (baseline) to the others (second and third scans). Registration of the OCT fundus images was performed using ImageJ,13 the public-domain image-processing software for bio-science with TurboReg plug-in (Biomedical Imaging Group, Swiss Federal Institute of Technology, Lausanne, Switzerland) (fig 4).14 TurboReg registered two given images (baseline against either second or third scans) by performing x, y translation and torsional rotation based on ONH and retinal vessel information. This approach is similar to the ONH analysis method used by HRT. RNFL thickness parameters (global, four quadrants, and 12 clock hours) were then measured along the virtual circle for each of the three Optic Disc Cube scans sharing one universal ONH centre.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3

Centred Each Time method utilising an automatically detected optic nerve head centre on each of the cube data obtained using spectral domain optical coherence tomography. Slight variation in the disc margin detection leads to a displacement of the resampling circle location that can be observed relative to the vessel branches (blue arrow).

• Download figure
• Open in new tab
• Download powerpoint

Figure 4

Centred Once method importing the automatically detected optic nerve head centre from the reference scan (A) onto the other cube data (B and C) obtained from the same eye. Retinal nerve fibre layer thickness measurements were then obtained on each image along the 3.4 mm diameter circle (white circles on (D) and (E)). (F, G) Images visualising the performance of optical coherence tomography fundus image registration by superimposing both the reference image ((A) in red) and the target image ((D) and (E) in green).

Statistical analysis

Linear mixed effect models1517 were used to analyse the data from the hierarchical study design and compute the components of variance for the following random effects: subjects, eyes within subjects, eyes within subjects by method interaction and scans within eyes within subjects. The statistical models included both a fixed effect for methods and a random interaction effect between methods and eyes within subjects. The square root of the scan variance component (VC) was computed for each method. Reproducibility is conventionally defined as 2.77 times the square root of the scan variance component. This value represents the difference between two measurements that would be needed in order to have 95% confidence that an observed difference is not due solely to measurement error. Because repeated measurements are taken for each eye of each subject, 95% confidence intervals on the ratios of the reproducibilities for each possible pair of methods had to be computed in order to assess differences in reproducibility between methods. When the ratio between methods is equal to one, the reproducibilities are identical. When the 95% confidence interval does not include one, the reproducibilities for the methods are statistically significant at or below the 5% level of significance. More importantly, the 95% confidence intervals indicate the plausible range of the true ratios of reproducibility and thus indicate how precisely (narrow interval) or imprecisely (wide interval) the ratios are known. The intraclass correlation coefficients (ICCs) were also calculated for each parameter and method using the ratios of the appropriate variance components.

In addition to reproducibility, the distance between ONH centres identified by the CET and CO methods sets was calculated on the second and third 3D OCT scans. Since the centre location is the same for the baseline cube data for CET and CO, baseline cube data were excluded from this analysis.

Results

The mean age of the 14 healthy volunteers (three males and 11 females) was 37.3 (10.3) years. The mean RNFL thickness measurements were 107.5 (9.9) μm (TD-OCT), 97.4 (8.1) μm (SD-OCT: CET) and 97.4 (8.0) μm (SD-OCT: CO) (table 1). There was one instance of algorithm failure, and this scan was excluded from the study. No scan was excluded due to eye motion.

The mean ONH centre location distance between CET and CO was 1.92 (1.63) pixels, which was equivalent to 58 (49) μm on the retina. For global RNFL thickness measurements, the estimated reproducibility ratio for CET/TD-OCT was 0.370 (95% CI 0.272 to 0.505), which indicates that the reproducibility for CET was better (ie, the variance was smaller) than for TD-OCT (37% of the reproducibility value for TD-OCT). Because the 95% CI does not include one, this indicates that the reproducibilities are statistically significantly different at the 5% level of significance. Similarly, the estimated reproducibility ratios were 0.353 (CI 0.259 to 0.481) for CO/TD-OCT, and 0.953 (CI 0.733 to 1.239) for CET/CO (table 2). For sector RNFL thickness measurements, SD-OCT using CET and CO showed statistically significantly smaller reproducibility ratios than TD-OCT except for 11 o’clock with CET. The reproducibility ratio results imply that, regardless of the ONH centring method, SD-OCT significantly improves RNFL thickness measurement reproducibility (reduces variability) in comparison with conventional TD-OCT. There was no statistically significant difference in reproducibility ratios between CET and CO in any of the global or sectoral measurements.

ICCs for each parameter and method are shown in table 3 along with the corresponding square root of the variance component for scans. ICCs (table 3) from all three methods (TD-OCT, CET and CO) at 11 o’clock were lower than other ICCs. Figure 5 shows that the reproducibility of both CET and CO methods was uniformly and substantially better than TD-OCT, and the differences were statistically significant.

• Download figure
• Open in new tab
• Download powerpoint

Figure 5

Square root of scan variance component (“SD”) plot from table 3. The reproducibility of both Centred Each Time (CET) and Centred Once (CO) methods was uniformly and substantially better than time-domain optical coherence tomography (TD-OCT) and the differences were statistically significant.

Discussion

The present results showed that the RNFL thickness measurement reproducibility was statistically significantly improved over conventional TD-OCT when using SD-OCT. The acquisition of a 3D data set is likely an important factor in the improvement. The OCT segmentation algorithm takes advantage of the three-dimensional neighbouring structure information available in SD-OCT, which is not possible with a two-dimensional conventional TD-OCT scan, making the SD-OCT segmentation more robust. The higher axial resolution of SD-OCT is a likely contributor as well, as segmentation is more reliable with better-defined RNFL borders.

It is interesting to consider the fact that the scanning time of the SD-OCT 3D cube scan is actually slower than conventional TD-OCT yet achieves a lower RNFL measurement variability. Circular resampling on a 3D cube data requires 56.7% of the 3D data in order to obtain the 3.4 mm circle data (114 out of 200 frames (3.42 mm); each frame vertically covers 30 μm on the retina) that take 0.84 s to acquire, while a single circular scan with TD-OCT takes 0.64 s to acquire. Assuming the characteristics of eye motion during scanning are constant, data acquired using a longer scan time should contain more eye motion artefacts. Nevertheless, SD-OCT showed a lower measurement variability than TD-OCT. This may be because the OCT fundus image provides an easy and reliable way to detect significant eye motion (like a saccade) during a scan so that the eye can be rescanned, if necessary, to minimise eye motion artefacts. With TD-OCT there is no systematic way to detect the magnitude of eye motion during scanning, since the corresponding fundus reference image is captured after the actual OCT scan. It is also possible that the SD-OCT segmentation algorithm is more robust than the TD-OCT algorithm because of a higher axial resolution and readily available 3D structural information as discussed above.

We frequently observed apparent variability in the shape of the automatically detected ONH margins using the Cirrus system software. Because the disc centre was defined based on the location of the disc margin, it was surprising that the CET method did not show a greater variability than the CO method. This is likely because the only information that the virtual circular scan resampling needs is the ONH centre location; it does not take into account ONH shape. Actually, only a 58 μm difference in the ONH centre location was observed between the CET and CO methods. Gabriele et al reported that the peripapillary circular scan has ±350 μm of measurement stability margin where the overall RNFL thickness measurement error stays within the expected measurement error range.18 The present result of 58 μm ONH centre location deviation is well within this stability margin. Further investigation is needed to see if the same is true on eyes with glaucoma and other pathologies.

In conclusion, SD-OCT 3D cube data provide significantly lower RNFL thickness measurement reproducibility compared with conventional TD-OCT (eg, Stratus OCT). Since the CO method does not have any advantage in measurement reproducibility over the CET method, the RNFL thickness measurement circle placement method employed by the manufacturer is clinically useful.