Subjects

A retrospective medical records review was conducted for patients who had visited Seoul National University Hospital Glaucoma Clinic between July 2017 and September 2018. All of the subjects underwent a full ophthalmic examination, including measurement of visual acuity, slit-lamp biomicroscopy, gonioscopy, measurement of intraocular pressure (IOP) by Goldmann applanation tonometry, measurement of refractive error with an autorefractor (KR-890; Topcon, Tokyo, Japan), measurement of corneal pachymetry (Pocket II Pachymeter Echo graph; Quantel Medical, Clermont-Ferrand, France) and dilated fundus examination. On maximal pupil dilation, all subjects underwent optic disc stereo photography and red-free fundus photography by digital fundus camera (VX-10; Kowa Company, Tokyo, Japan), SD-OCT scan (Cirrus HD-OCT; Carl Zeiss Meditec, Dublin, California, USA) and swept-source OCTA scan (DRI OCT Triton; Topcon, Tokyo, Japan). The subjects also underwent standard automated perimetry (SAP; Humphrey C24-2 SITA-Standard visual field; Carl Zeiss Meditec, Dublin, California, USA) within 6 months of OCTA scanning. Additionally, subjects’ self-reported presence of diabetes mellitus or systemic hypertension, information on medications used by them and any history of aorta surgery that could affect ocular perfusion were recorded.

The NTG was defined as glaucomatous optic neuropathy (enlarged cupping, neuroretinal rim thinning and RNFL defects), glaucomatous visual field defect, open iridocorneal angle and IOP≤21 mm Hg. The GS were defined as those that have glaucoma suspicious-looking optic disc (enlarged cupping, neuroretinal rim thinning, and RNFL defects) on dilated fundus examination without glaucomatous visual field defect. Glaucomatous visual field defect was defined as a cluster of ≥3 points with p<0.05 on a pattern deviation map in at least one hemifield, including ≥1 point with p<0.01; a pattern SD of p<0.05 or a glaucoma hemifield test result outside the normal limits. Only subjects with reliable visual field test results (fixation loss<20%, false-positive errors<15%, false-negative errors<15%) were included. If both eyes were eligible for the study, the eye with a higher MD of SAP was included and images from left eyes were horizontally flipped for consistent analysis with right eyes. We excluded subjects with (1) ocular conditions known to affect macular thickness or macular vessel density other than glaucoma (subjects with systemic hypertension or diabetes mellitus without retinopathy are not excluded), (2) high myopia (spherical equivalent<−6 dioptres), (3) intraocular pressure (IOP) >21 mm Hg, (4) no or unreliable SAP result within 6 months of OCTA scanning, (5) mean deviation (MD) of SAP <−5.5 dB, (6) unreliable OCT or OCTA scan, (7) a history of ocular trauma, (8) a history of ocular surgery excepting uncomplicated cataract or glaucoma surgery.

OCT and OCTA image acquisition

Macular GCIPL thickness was measured by SD-OCT according to the macular ganglion cell analysis (GCA) protocol. To explain briefly, 200 horizontal B-scans × 200 A-scans within a 6×6×2 mm cube centred on the fovea was performed for automatic identification of the outer boundaries of the RNFL and inner plexiform layer (IPL). A GCIPL thickness map was obtained within an annulus of 1.0 and 1.2 mm inner vertical and horizontal diameter and 4.0 and 4.8 outer vertical and horizontal diameter. The annulus was automatically divided into six sectors (SN, superonasal; SS, superosuperior; ST, superotemporal; IT, inferotemporal; II, inferoinferior; IN, inferonasal) in each of which the average GCIPL thickness was calculated.

The circumpapillary RNFL thickness was measured using the same SD-OCT device with the GCA protocol. Briefly, 200 horizontal B-scans × 200 A-scans within a 6×6×2 mm cube were performed, the built-in software automatically calculating the RNFL thickness of the peripapillary circle centred on the optic disc; 3.46 mm in diameter consisting 256 A-scans. Also, the circumpapillary RNFL thickness in all 12 clock-hour sectors was automatically generated and analysed.

The OCTA scans were obtained using the 1050 nm swept-source OCTA device within a 4.5×4.5 mm cube centred on the fovea. The built-in software (IMAGEnet6) was used to generate images for improved detection sensitivity of blood flow and reduced motion artefacts without compromise of axial resolution. Images with a quality score <60, poor clarity, motion artefacts (discontinuities of vessels), locally decreased image intensity due to media opacities (such as floaters) and segmentation errors were regarded as unreliable data, and as such, were excluded from the analysis.

OCTA image processing

The OCTA scan image of the superficial vascular plexus excepting the internal limiting membrane (ie, an en-face slab image from the inner border of the RNFL layer to 15.6 µm from the boundary between the IPL and the inner nuclear layer, obtained according to the manufacturer’s instructions) was exported in grayscale using the built-in software (figure 1A). Then, the images were imported into customised software (ImageJ; National Institutes of Health, Bethesda, Maryland, USA) for vessel density analysis. First, a non-local means (NLM) denoising filter and Hessian filter were applied to the grayscale images in order to reduce the background noise and sharpen the blood vessels. Second, we used an autothreshold algorithm to convert to a binary image and then the image was skeletonised for calculation of vessel density (figure 1B). To minimise the effect of the low-vessel- density areas around large vessels after skeletonisation, area occupied by large vessels was excluded from the analysis (figure 1C). The total vessel length per unit area calculated based on the skeletonised image was defined as ‘superficial microvessel density (SMD)’. The analysis was performed within (1) a 6-sector grid: an annulus of 1.0 and 4.0 mm inner and outer diameter divided by six sectors (figure 1C), (2) a 4-sector grid: an annulus of 1.0 and 3.0 mm inner and outer diameter divided by four sectors, which is the same as the grid used in the Early Treatment Diabetic Retinopathy Study (ETDRS; not shown in figure).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Representative case. (A–C) Postprocessing of the macular OCT angiography scan image of representative case. (A) Original OCT angiography scan image showing decreased perfusion in inferior macular area. (B) Binarised and skeletonised image from (A). (C) Sectoral SMD was calculated within (1) annulus of 1.0 and 4.0 mm inner and outer diameter divided by six sectors (as shown in C), (2) and annulus of 1.0 and 3.0 mm inner and outer diameter divided by four sectors. The area occupied by large vessels (indicated as blue area) was excluded from the analysis. (D) Disc photo, (E) RNFL photo, (F) circumpapillary RNFL deviation map and (G) macular ganglion cell-inner plexiform layer deviation map of spectral domain-OCT, (H) pattern deviation from standard automated perimetry of the patient. OCT, optical coherence tomography; RNFL, retinal nerve fibre layer; SMD, superficial microvessel density.

One author (J-SK) performed all of the measurements on all eyes while blinded to each patient’s clinical information. In order to assess the intraobserver reproducibility, the measurements were performed once again on 30 randomly chosen eyes. For the purpose of assessing the interobserver reproducibility, another author (SUB) performed the same measurements on another 30 randomly chosen eyes.

Statistical analysis

The Statistical Package for the Social Sciences V.25.0 for Windows (SPSS, Chicago, Illinois, USA) and the R packages pROC V.1.13 were used for the statistical analyses,16 the level of statistical significance having been set at p<0.05. The data obtained are described herein, as appropriate, in terms of their mean±SD values, median frequencies (number of cases) and percentages. The Shapiro-Wilk test, independent t-test, Mann-Whitney test and χ² test were used to compare the variables of each group. The intraobserver and interobserver reproducibilities were assessed by intraclass correlation coefficient (ICC). Pearson’s correlation coefficient was used to measure the strength of the relationship between paired data. Specifically, the level of correlation was assessed by coefficient value (r) as follows: 0.10–0.29=weak, 0.30–0.49=moderate and over 0.50=strong. The area under the receiver operating characteristic (AUROC) curve was used to evaluate diagnostic accuracy. To evaluate the diagnostic accuracy of the combination of variables, binary logistic regression analysis was applied to generate the predictive values of each potential index and then the ROC analysis was able to merge these predictive values to obtain a combined ROC curve. The DeLong test was used to compare the diagnostic abilities of the variables.