Automated Quantification of Macular Ellipsoid Zone Intensity in Glaucoma Patients: the Method and its Comparison with Manual Quantification

The macular ellipsoid zone intensity (mEZi) is a known marker of disease severity in a number of diverse ocular diseases. The purpose of this study was to establish an automated method (AM) for mEZi quantification and to compare the method’s performance with that of a manual method (MM) for glaucoma patients and healthy controls. Seventy-one (71) mild-to-moderate glaucoma patients, 71 severe-glaucoma patients, and 51 controls were enrolled. Both calibration (n = 160) and validation (n = 33) image sets were compiled. The correlation of AM to MM quantification was assessed by Deming regression for the calibration set, and a compensation formula was generated. Then, for each image in the validation set, the compensated AM quantification was compared with the mean of five repetitive MM quantifications. The AM quantification of the calibration set was found to be linearly correlated with MM in the normal-to-severe-stage glaucoma patients (R2 = 0.914). The validation set’s compensated AM quantification produced R2 = 0.991, and the relationship between the 2 quantifications was AM = 1.004(MM) + 0.139. In the validation set, the compensated AM quantification fell within MM quantification’s 95% confidence interval in 96.9% of the images. An AM for mEZi quantification was calibrated and validated relative to MM quantification for both glaucoma patients and healthy controls.


Validation of AM versus MM Quantification. The distribution of mEZi (by MM) in validation study was
comparable with that of calibration study. The mean ± SD of MM quantification in the calibration set (4.08 ± 0.42 in normal group; 3.14 ± 0.43 in mild-to-moderate glaucoma group; 2.36 ± 0.52 in severe glaucoma group) was not significantly different from the mean ± SD of MM quantification (4.12 ± 0.45 in normal group; 3.13 ± 0.38  in mild-to-moderate glaucoma group; 2.36 ± 0.28 in severe glaucoma group) used in the present validation set (Mann-Whitney test [P = 0.991, 0.893 and 0.446, respectively]). In Fig. 2A, the compensated AM quantification is plotted versus the mean of the five MM quantifications for each of the 33 image sets. The correlation produced R 2 = 0.991 (P < 0.001). The relationship between the AM (abscissa) and MM quantifications (ordinate) was AM = 1.004(MM) + 0.139, as established by Deming regression. Figure 2B shows the Bland-Altman plot, with an average difference of 0.15 mEZi (AM > MM), the limits of agreement extending from −0.07 to + 0.37. The AM quantification fell within the 95% CI of MM counts for 32 of the 33 images sets (96.9%). The purpose of this paper was to present an automated algorithm and software application for mEZi quantification and to assess the method's performance relative to a previously published manual method with SD-OCT images taken from normal and glaucomatous eyes of differing disease stages. The proposed method's results proved to be highly correlated to those of manual mEZi measurement for an initial calibration of 160 images. It also showed a performance similar to that of an experienced ophthalmologist for a validation set of 33 images counted five times each, its results falling within the ophthalmologist's 95% CI for 32 of the 33 images. Since both sets of images demonstrated mEZi quantification that ranged from normal to end-stage glaucoma, the obtained results indicated that the proposed method met the requirements laid out at the beginning of the work: an automated mEZi calculator that is applicable to all levels of glaucomatous damage as well as computationally robust and efficient.

Comparison of AM versus MM mEZi Quantification for Different Stages of Glaucoma. Supplementary
In specific performance, the AM took less than 1 minute to obtain a final mEZi value for one SD-OCT image, according to the test system used in the present study (MacBook Pro laptop, 2.2 GHz Core2 Duo, 3GB RAM; Apple, Cupertino, CA, USA), at least. Meanwhile, manual mEZi calculation took at least 5 to 10 minutes for complete analysis of the same image. Thus, analysis of a large number of images or a wide range of areas by the MM can be time consuming and prone to error. We anticipate that the proposed automated mEZi analysis method will find application to larger and more diverse datasets and will foster additional research of photoreceptor change in glaucoma as well as further evaluation of its clinical significance.
It should be noted that the application of the method proposed in this paper can be expanded to other diseases. EZ intensity is known to be a marker of severity in many diverse retinal pathologies including, among www.nature.com/scientificreports www.nature.com/scientificreports/ others, age-related macular degeneration and inflammatory diseases [8][9][10][11] . Due to disruption of the outer retinal layer by lesions such as drusen, reticular pseudodrusen, or hyperreflective foci in retinal diseases 10 , however, the overall analysis protocol might require modification along with further validation. Additionally, AM analysis is easily modifiable for quantification of the intensities of outer retinal layers other than the EZ, or for calculation of intensities using other retinal layers as reference values. The process of expanding the software in these directions is already underway.
The present study's findings should be interpreted in the light of its limitations. First, the sampling number differed between the two methods: each meridian consisted of 20 retinal segments and one central segment in the MM, but there was a total of 220 samples for the AM. However, EZ intensity reduction during glaucoma-stage advancement has been proposed to occur in the form of an overall retinal change rather than as a focal pathologic change 16 . Thus, the two methods' average values would have shown, regardless of the number of analysis samples, high agreement. Second, mEZi quantification by the AM showed a failure rate of 6.28%. In the OCT scans, the retinal layers were not always parallel to the image's horizontal plane, and moreover, had an irregularly curved shape. In eyes with a highly curved posterior structure, the AM often failed to crop the area of interest. Figure 4 shows an example of OCT scans for which the AM failed to quantify mEZi. Future modifications of the AM to minimize the rate of analysis failure are planned. Third, whereas compensation of automated mEZi quantification was required for the fairest comparison of the AM performance with the MM's in the validation study, it is not believed that  www.nature.com/scientificreports www.nature.com/scientificreports/ automated counts should be compensated in actual scientific applications. Doing so would imply that the manual mEZi analysis method is actually the gold standard, and thus, that it is more accurate than the present study's AM, which is difficult, if not impossible, to determine at this point. Accordingly, future research evaluating the accuracies of the two methods by comparison of molecular changes in photoreceptors will be necessary. Fourth, the performance of the described methods depends, as for any other image processing algorithm, on the quality of the input images. We were careful to include only OCT images with no artifacts and good image quality. Our results might not be directly applicable to lower-quality OCT images obtained in an actual clinical setting, therefore.
In summary, this study presented a novel method for automated mEZi quantification based on SD-OCT images. The calibration and validation data indicate that its results are similar to those of manual mEZi calculation through a wide range -normal to end-stage -of glaucomatous damage. We expect that this automated mEZi quantification method will find application in various studies in the field of ophthalmology.

Methods
This study was approved by the Seoul National University Hospital Institutional Review Board and faithfully adhered to the tenets of the Declaration of Helsinki. All participants provided their written informed consent.
Overview of study design. First, a calibration study was performed in order to compare mEZi for the two methods (AM vs. MM) with an initial group of 160 image sets. These images had been obtained from normal (n = 40), mild-to-moderate glaucoma (n = 60), and severe glaucoma (n = 60) eyes so as to achieve a balance among the damage levels. Next, a validation study was conducted for comparison of AM with MM quantification, specifically by determining whether or not AM quantification fell within the range of repeated MM quantifications. For this, MM quantification was repeated five times for a separate group of 33 image sets representing a similarly broad range of glaucomatous damage (11 images per group). The AM quantification was compared with both the mean MM quantification and the range of MM re-quantification variability (95% confidence interval [CI]). Finally, in order to demonstrate the overall automated mEZi quantification strategy, SD-OCT image acquisition, AM mEZi measurement, and AM mEZi topographic maps were generated for normal and glaucomatous eyes.
Study subjects. All of the study subjects had been examined between January 2015 and December 2018 at the Seoul National University Hospital Glaucoma Clinic in Seoul, Korea.
For inclusion in the study, subjects had to be 40 to 65 years old, have a spherical refraction that was greater than −6 diopters (D) and less than 3 D, an open anterior chamber angle, and show reliable visual field (VF) test results. The exclusion criteria were: (1) a history of intraocular surgery (except uncomplicated cataract surgery) or retinal laser photocoagulation; (2) any neurological and/or systemic diseases potentially affecting the retinal structure and/or function or VF results. Additionally, any cases of suspicious retinal lesions potentially affecting the outer retinal layer, such as in cases of inflammatory conditions or hereditary or degenerative retinal diseases, were excluded.
Glaucomatous eyes were defined based on the characteristic optic disc appearance (localized or diffuse neuroretinal rim thinning/notching) on stereo disc photography (SDP), red-free fundus imaging of the presence of retinal nerve fiber layer (RNFL) defect in the corresponding region, and the presence of VF defect corresponding to structural change. Optic disc signs on SDP and RNFL changes on red-free imaging were independently evaluated by two glaucoma specialists (AH and KHP) masked to all non-relevant clinical data. Discrepancies between them were resolved by consensus. Based on a reliable (false-positives/negatives <15%, fixation losses <15%) Humphrey Visual Field (HVF) result obtained within 3 months of SD-OCT imaging, the glaucoma patients were divided into two groups: mild-to-moderate glaucoma (VF mean deviation [MD] ≥ −12 dB) and severe glaucoma (VF MD < −12 dB).
The normal controls showed intraocular pressure (IOP) less than or equal to 21 mm Hg, had no history of IOP elevation, no glaucomatous optic disc appearance, no RNFL defect, and normal HVF results. Normal HVF results were defined as an MD and pattern standard deviation within the 95% confidence limits and a glaucoma hemifield test result within the normal limits. If both eyes were eligible, one was selected randomly.
Imaging of outer retinal layer. All of the subjects underwent SD-OCT confocal scanning laser ophthalmoscopy (Spectralis HRA + OCT; Heidelberg Engineering, V Heidelberg, Germany) using the eye-tracking feature (TruTrack; Heidelberg Engineering, Heidelberg, Germany). All of the images were obtained through dilated pupils by a single experienced examiner. The 19 horizontal and 19 vertical line scans of 9-mm length were obtained in the high-resolution setting, and 128 frames were averaged (Fig. 5). Presence of foveal bulge, foveal depression, and inner-retinal-layer thinning, all on SD-OCT, was considered to confirm the foveal area. For inclusion, all of the images were reviewed for non-centered scans or artifacts, and all had a signal quality >20 dB.
Manual mEZi quantification method. The details of the methodology for manual quantification of mEZi have already been reported 16 . Briefly, mEZi intensity was determined as the ratio of the second reflective band to the first (i.e., the EZ/ELM [external limiting membrane] ratio) to account for the variation of OCT scan brightness. Logarithmic-transformed B-scans of each eye of each participant were rendered in the tagged image file format (TIFF).
EZ is known to be less distinct according to eccentricity 17 . For the purposes of consistent and accurate mEZi measurements, then, we analyzed the central macular area (total 4000 µm-length) only. The mEZi (4000-µm of 11 horizontal and 11 vertical retinal scans) was averaged, each meridian comprising 21 retinal segments. The mEZi was measured based on the highest EZ band intensity value as divided by the highest ELM band intensity value of relevant SD-OCT imaging (Fig. 6). All measurements were performed using the public-domain NIH Image program (ImageJ 1.48 v, Wayne Rasband, National Institutes of Health, Bethesda, MD, USA). To avoid retinal vessels'  www.nature.com/scientificreports www.nature.com/scientificreports/ (n = 11), and severe glaucoma (n = 11) eyes. The MM mEZi quantification was performed on all of the 33 images sets by the same operator (YKK) at five separate times separated by at least 3 days. The correlation between AM and MM quantification was assessed by Deming regression, as outlined above. The AM quantification for each image set was then compensated by using the equation resulting from the linear regression of the difference versus average mEZi quantification for the calibration set. The number of images for which compensated AM quantification was outside of the 95% CI of MM quantification was recorded.
Data analysis. Comparison of mEZi distribution in each method (i.e. calibration and validation) was performed by Mann-Whitney test. Inter-method and inter-observer comparisons were performed using Deming regression and Bland-Altman plots. Deming regression is an errors-in-variables model that finds the line offering the best fit for a two-dimensional dataset 18 . It differs from simple linear regression in that it calculates for observation errors on both the x-and y-axes. Deming regression assesses inter-method linearity according to statistical significance based on the CI for the slope not containing 1 and the CI for the intercept not including 0 18 . Bland-Altman plots are presented below for visual representation of the inter-method agreement. Statistical analysis was performed using the MedCalc software (version 12.1.3.0, Mariakerke, Belgium); a P value less than 0.05 was considered statistically significant.

Data availability
The dataset generated during the current study is available in the Supplementary Material. Figure 7. Flowchart of the automated macular EZ intensity (mEZi) quantification method. By thresholding processes, the software eliminates the inner retinal layer and extracts the outer retinal layers including the ELM and EZ. The mEZi was determined by calculating the ratio of the EZ/ELM intensity peaks.