High-Density Optical Coherence Tomography Analysis Provides Insights Into Early/Intermediate Age-Related Macular Degeneration Retinal Layer Changes

Abstract

Purpose: To topographically map all of the thickness differences in individual retinal layers between early/intermediate age-related macular degeneration (AMD_early/AMD_int) and normal eyes and to determine interlayer relationships.

Methods: Ninety-six AMD_total (48 AMD_early and 48 AMD_int) and 96 normal eyes from 192 participants were propensity-score matched by age, sex, and refraction. Retrospective optical coherence tomography (OCT) macular cube scans were acquired, and high-density (60 × 60 0.01-mm²) grid thicknesses were custom extracted for comparison between AMD_total and normal eyes corrected for confounding. Resultant “normal differences” underwent cluster, interlayer correlation, and dose–response analyses for the retinal nerve fiber layer (RNFL), ganglion cell layer (GCL), inner plexiform layer (IPL), inner nuclear layer (INL), outer plexiform layer (OPL), outer nuclear layer + Henle's fiber layer (ONL_+HFL), inner and outer segment (IS/OS) thickness, and retinal pigment epithelium (RPE) to Bruch's membrane (BM) thickness.

Results: AMD_total inner retinal clusters demonstrated extensively thinned RNFL, GCL, IPL, and paracentral INL and thickened INL elsewhere, with normal difference means ranging from −8.13 µm (95% confidence interval [CI], −11.12 to −5.13) to 1.58 µm (95% CI, 1.07–2.09) (P < 0.0001 to P < 0.05). Outer retinal clusters displayed thinned paracentral OPL/ONL_+HFL, central IS/OS, and peripheral RPE–BM and thickened central RPE–BM, with means ranging from −1.31 µm (95% CI, −2.06 to −0.55) to 2.99 µm (95% CI, 0.97–5.01] (P < 0.0001 to P <0.05). Effect sizes (−2.56 to 9.93 SD), cluster sizes, and eccentricity effects varied. All interlayer correlations were negligible to moderate regardless of AMD severity. Only the RPE–BM was partly thicker with greater AMD severity (up to 5.44 µm; 95% CI, 4.88–6.00; P < 0.01).

Conclusions: From the early stage, AMD eyes demonstrate thickness differences compared to normal with unique topographies across all retinal layers. Poor interlayer correlations highlight that the outer retina inadequately reflects complete retinal health. The clinical importance of OCT assessment across all individual retinal layers in early/intermediate AMD requires further investigation.

The outer retina has typically been considered the fundamental site of insult in age-related macular degeneration (AMD).^1,2 Compounding evidence has demonstrated that in vivo alterations also occur within inner retinal layers in the early stages of AMD.^3–10 The interlayer relationships are unknown, although various pathophysiological models suggest that outer retinal degeneration may link to inner retinal changes via some cause-and-effect mechanisms.^11,12 Understanding the topography of anatomical differences between AMD and normal eyes and their interlayer relationships could have significant clinical implications, such as identifying where clinicians should be looking for change and answering whether patients with early or intermediate AMD could benefit from complete retinal optical coherence tomography (OCT) evaluation rather than solely focusing on the outer retina. 

Previous topographical analyses of the retina have been defined using the Early Treatment for Diabetic Retinopathy Study (ETDRS) spatial template. These analyses, however, are low density, confining retinal space to nine unequal units (0.79- to 3.36-mm² areal sectors),^3–5 which hinders effective topographical interpretation within and between layers. For example, assessing the effect of eccentricity on retinal layer thicknesses using the ETDRS sectors limits eccentricity to three unequally sized sectors (central, inner, and outer) and assumes perfect concentricity and symmetry, thus inducing the modifiable areal unit problem (MAUP)—statistical bias based on how data are spatially grouped.^13,14 Similarly, comparing retinal thickness differences between layers using the large ETDRS sectors disallows minute adjustments for displacements of Henle's fibers.¹⁵ 

We recently addressed these problems by defining retinal space according to 64 equal units (8 × 8 0.74-mm² areal grids).^16–21 This moderate-density analysis revealed that normal macular gridwise thicknesses could be assigned to spatial groups of statistically similar within-group data and statistically separable between-group data (i.e., clusters, representative of normal macular anatomical topography and less variable than ETDRS sectors).¹⁹ Comparing intermediate AMD to normal eyes using moderate-density analysis revealed extensive topographical differences and potential interlayer relationships across most retinal layers^16,17 that had not been observed in previous ETDRS-based analyses, underlining the benefit of greater spatial (lateral) resolution for OCT thickness analysis. 

There remained some unresolved questions associated with moderate-density analysis.¹⁷ First, topographical patterns of thickness differences were equivocal in the retinal nerve fiber (RNFL) and inner nuclear layer (INL) and at the fovea of inner retinal layers. Correlations were also weak across most interlayer comparisons. It was not clear whether these were true anatomical outcomes or a consequence of incomplete resolution of the MAUP, as clustering was applied prior to comparisons between AMD and normal eyes and therefore assumed topographical differences followed normal macular anatomy. Second, grid size was still spatially coarse relative to the micrometer scale of macular cellular densities.^22–26 Hence, moderate-density analysis may have been incapable of depicting steep gradients of thicknesses and adjusting for lateral displacements of Henle's fibers between the outer and inner retina which range from 0.13 to 0.64 mm.¹⁵ 

Recently, we developed a high-density approach that defines retinal space according to 3600 equal units (60 × 60 0.01-mm² areal grids).²⁷ Greater lateral resolution enables spatially refined topographical interpretation (otherwise masked by lower resolutions) that appropriately represents steep gradients of thicknesses and allows minute adjustments for displacements of Henle's fibers. Modified cluster application after thickness comparisons between AMD and normal eyes will also resolve the MAUP by not making a priori model assumptions about macular topography; instead, clustered thickness differences will represent topographical maps of changes that may occur between AMD and normal eyes. Therefore, we hypothesize that OCT high-density analysis of early/intermediate AMD compared to normal eyes will enable topographical mapping of thickness differences in all individual retinal layers and clarify whether these interlayer differences are related. This could help guide future clinical assessments for the early stages of AMD by identifying where changes may occur and whether OCT evaluation of the outer retina alone can sufficiently depict complete retinal integrity. 

Methods

Study Population

The study population was recruited through retrospective review of patient records from the Centre for Eye Health (CFEH), Sydney, Australia, from July 12, 2010, to February 20, 2021. CFEH is a referral-based clinic providing advanced diagnostic testing and management of ocular disease by optometrists and ophthalmologists.²⁸ All participants in this study had provided prior written informed consent for research use of their de-identified data in accordance with the tenets of the Declaration of Helsinki, and the study was approved by the Biomedical Human Research Ethics Advisory Panel of the University of New South Wales. 

To cover the breadth of AMD in its early stages, both AMD_early and AMD_int eyes were included and formed the AMD_total group. Eligibility of AMD_total eyes was based on the following criteria: age of ≥ 50 years, clinical diagnosis of AMD_early or AMD_int,²⁹ and no other macular-involving disease or significant structural abnormalities such as vitreomacular traction, or intra-/subretinal deposits, fluid, pigment, or vascular changes. Clinical diagnosis of AMD was based on fundus photography between two or more non-blind investigators according to a modified Beckman Initiative classification.²⁹ Participants 50 to 54 years of age were considered to have AMD if all phenotypic criteria were fulfilled, consistent with other notable studies.^30–33 Phenotypic severity criteria for AMD were 

• • AMD_early—presence of medium drusen (≥63 to <125 µm) without pigmentary abnormalities related to AMD
• • AMD_int—presence of large drusen (≥125 µm) or pigmentary abnormalities related to AMD with at least medium drusen

Eyes with any other macular-involving disease or significant structural abnormalities, reticular pseudodrusen, and/or any signs of late AMD (macular neovascularization, geographic atrophy, or disciform scarring) were excluded from the study. 

Eligibility of normal eyes was based on the following criteria: age of ≥50 years, visual acuity better than 0.1 logMAR (20/25 Snellen) for participants <60 years old or 0.2 logMAR (20/32 Snellen) for participants ≥60 years old, intraocular pressure <22 mmHg in both eyes, and no macular-involving disease or significant structural abnormalities as described above. One eye was selected per participant, and where both eyes were eligible simple randomization was used to select a random eye. 

Propensity-Score Matching

Propensity-score matching, using multivariable logistic regression based on age, sex, and spherical equivalent refraction, was applied first to AMD_early and AMD_int eyes with an equal number of participants per group. AMD_total eyes were then propensity-score matched to an equal number of normal eyes. Fuzzy matching without replacement of propensity scores was performed to randomize the selection of participants and reduce an imbalance of potentially confounding co-variables between groups, hence mitigating selection bias,^34,35 rather than exact matching, which leaves individuals unmatched in a limited sample pool. Match tolerance was increased after each iterative random draw until the maximum number of matches. This resulted in relatively balanced propensity scores among groups (logistic regression predicted probability mean ± SD for AMD_early, 0.49 ± 0.1; AMD_int, 0.51 ± 0.07; AMD_total, 0.52 ± 0.16; normal, 0.43 ± 0.17).³⁶ 

Image Acquisition and Retinal Layer Segmentation

OCT macular cube scans comprised of 61 B-scans spaced 120 µm apart within an area of 8600 × 7167 µm or 30° × 25° were obtained using the SPECTRALIS SD-OCT (Heidelberg Engineering, Heidelberg, Germany). This scanning protocol contained the highest number of B-scans using commercially available SPECTRALIS SD-OCT settings without significantly compromising image quality.³⁷ If multiple scans were available per participant, the earliest scan meeting the inclusion criteria, above 15-dB signal strength, and without significant artifacts was used. Ocular tilt, automatic segmentation, and manual correction were applied to each scan via the HRA/SPECTRALIS Viewing Module 6.9.5.0 (Heidelberg Engineering). 

Participants were randomized into one of two blocks, and each block was assigned to an optometrist (MT or VK) for independent review and manual correction of segmentation where necessary for the RNFL, ganglion cell layer (GCL), inner plexiform layer (IPL), INL, outer plexiform layer (OPL), outer nuclear layer + Henle's fiber layer (ONL_+HFL), inner and outer segment (IS/OS) thickness, and retinal pigment epithelium to Bruch's membrane (RPE–BM) thickness (Fig. 1A, brown insert). The optometrists then reviewed the alternate block, and any further manual correction was resolved through discussion and consensus between them. Consensus for all segmentation was reached after one session of discussion. Blinding of graders to participant disease status during outer retinal segmentation was not possible, as drusen and/or pigmentary changes are obvious, and masking of the outer retina during inner retinal segmentation is yet unavailable in commercial OCT software. Although this may have produced potential bias, manual correction of segmentation has shown excellent repeatability and reproducibility in AMD eyes³⁸ and therefore was regarded as the ground truth for retinal anatomy (as opposed to automatic segmentation) in concordance with other studies.^39–41 Notably, segmentation boundaries were manually corrected to continue through large vasculature to mitigate their effect on thicknesses (Fig. 1A, magenta insert, asterisk). Segmentation was also corrected to combine Henle's fiber layer with the ONL as commonly done in OCT studies due to its inconsistent reflectivity (although anatomically part of the OPL) (Fig. 1A, cyan insert, arrowhead), and to resolve mis-segmentation around drusen (Fig. 1A, cyan insert, dagger). 

Figure 1.

 

Macular volume scans within the SPECTRALIS HRA+OCT Viewing Module were automatically segmented and manually corrected to define the retinal layers in each B-scan (A).¹⁷ Note that automatic segmentation boundaries were manually corrected to continue through large vasculature (magenta insert, asterisk), combine Henle's fiber layer with the ONL (cyan insert, arrowhead), and resolve mis-segmentation around drusen (cyan insert, dagger). Grids that were completely covered, such as by retinal blood vessel shadowing (yellow), were manually excluded. Thicknesses were extracted across 3600 (60 × 60) grids (total 6880 × 6880 µm or 24° × 24° area) centered on the fovea for each retinal layer (B). Gridwise thicknesses were compared between AMD_total and corrected normal eyes, and resultant values were denoted as normal differences (µm) = z_n, where n = grid number from 1 to 3600 in random order. Two-step clustering was performed with AMD_early and AMD_int as independent groups, and the cluster process was reiterated to reduce cluster size by 1 until cluster means were ≥ 1.96 SD separable (C). Results were de-convoluted and presented as (D) graph (left; mean [95% CI]) and topography map (right) formats with legend (middle). Clusters were separated into negative or positive ranks based on magnitude of normal difference (µm) and represented via a gradient color scale: C_{−1, −2, −3, …} = more thinned (darker blue; C_{+1, +2, +3, …} = more thickened (darker green). C₀ indicates within 1.96 SD of zero normal difference (cream). In the topography map, this example shows greater thinned (darker blue) retina toward the peripheral macula and greater thickened (dark green) retina centrally. The black cross denotes the foveal center; the scale at the bottom right and all images are in right-eye format.

High-Density Grid Data Extraction

Thicknesses were custom extracted across 3600 grids (60 × 60 0.01-mm² areal units; 115 × 115 µm or 0.4° × 0.4° sided units) centered on the fovea and totaling an area of 6880 × 6880 µm or 24° × 24° (Fig. 1B). The grid density (i.e., 60 × 60 grids) was selected to maximize coverage of segmented thickness values within the 61 total B-scans of each macular cube. Therefore, each grid contained approximately one B-scan. The automatic real time (i.e., number of image frames averaged per location) was increased from the default nine to 12 to improve the signal-to-noise ratio.³⁷ The total grid area (i.e., 6880 × 6880 µm) was selected because it was commensurate with the commercially available SPECTRALIS SD-OCT total grid area using 64 grids (8 × 8 0.74-mm² areal units) as seen in previous studies.^16,17,19 Note that the total grid area did not cover the entire macular cube scan of 8600 × 7167 µm or 30° × 25°. Custom extraction was performed using code developed in MATLAB 9.9 (MathWorks, Natick, MA, USA) by author DAC. Grids that were completely covered by a retinal blood vessel, distinct peripapillary atrophy, or the optic nerve were manually excluded to minimize potential confounding. Exclusion of these grids was applied to all retinal layers for each participant to mitigate any related potential confounding (e.g., shadowing) (Fig. 1A, yellow insert). 

Corrected Thickness Comparisons

Grid-wise thicknesses for normal eyes were corrected for age, sex, and refraction using the multivariable linear regression model: 

\begin{equation*}{y_n} = a{x_1} + b{x_2} + c{x_3} + k\end{equation*}

where y = corrected thickness (µm); n = grid number from 1 to 3600; x₁ = age (years); x₂ = sex (binary; males = 0, females = 1); x₃ = refraction (spherical equivalent diopters [D]); a, b, and c are coefficients; and k is a constant. 

Refraction was used as a surrogate for axial length, as previous systematic review has demonstrated strong correlation between the two variables.⁴² Gridwise thicknesses were then compared between AMD_total and corrected normal eyes, and the resultant values were denoted as gridwise normal differences:  

\begin{equation*}{z_n} = \left( {{\rm{AM}}{{\rm{D}}_{{\rm{total}}}}{\rm{gri}}{{\rm{d}}_n}{\rm{thickness}}} \right) - {y_n}\end{equation*}

where z = normal difference (µm). For example, grid_n of an AMD eye from a 57-year-old female with −3.00 D refraction would be compared to grid_n of a normal eye corrected to the equivalent of a 57-year-old female with −3.00 D refraction. The rigorous use of corrected thickness for each of the 3600 grids for each retinal layer for each participant was implemented to ensure that gridwise comparisons were not confounded by multiple variables. Correction was applied at the gridwise level (pre-clustering), instead of at the group level (post-clustering) as performed in previous work,^16,17 so as not to assume that all grids within each spatial group (cluster) shared exactly the same regression characteristics. This was done to address the MAUP, which states that statistical bias may be introduced based on how data are spatially grouped.^13,14 

Cluster Analysis

Unsupervised cluster analysis was performed to identify new spatial groupings of retinal thickness changes between AMD_total and normal eyes (i.e., normal differences). Groupings were comprised of normal differences that were statistically similar within clusters and statistically separable between clusters.⁴³ AMD_early and AMD_int were set as independent groups, enabling further quantitative comparison between study groups. 

Cluster number and size were not predefined to avoid violating statistical separability of cluster means and also inducing the MAUP^13,14 via a priori assumption that spatial patterns of change were uniform across all retinal layers despite their varying thickness distributions.⁴⁴ Instead, two-step clustering was selected for its proven robustness against many other cluster algorithms⁴⁵ and was applied to random ordered grids⁴⁶ using the log-likelihood method⁴⁷ to generate a suitable cluster number for each individual retinal layer while considering the lowest Bayesian information criterion, intracluster similarity, and intercluster separability.⁴⁸ Subsequently within each retinal layer, cluster means were compared and, if the separability criterion of ≥1.96 SD was not met, the cluster number was reduced and the cluster process reiterated until the criterion was satisfied (Fig. 1C). Note that by definition, the criterion of ≥1.96 SD ensured statistical difference between cluster means within each layer and is equivalent to d′ = 1.96 from signal detection theory⁴⁹ and 95th-percentile normal distribution. 

Final clusters for each retinal layer were deconvoluted to generate mean and confidence intervals and were displayed as graphs and spatial topography maps (Fig. 1D). Clusters were labeled as negative or positive and ranked based on the magnitude of normal difference (µm), represented via a gradient color scale for ease of interpretation: C_{−1, −2, −3, …} = more thinned (darker blue); C_{+1, +2, +3, …} = more thickened (darker green). C₀ indicates within 1.96 SD of zero normal difference (cream). Normal differences (µm) were also expressed in SD units (i.e., Z-scores calculated using mean and distribution of the sampled corrected normal population). Note that normal differences (SD units) are not statistically equivalent to the SD of normal differences (µm), as the latter is calculated using means and distributions of both sampled AMD and corrected normal populations.⁵⁰ Cluster normal differences (µm) were compared against eccentricity (mm) to determine potential effects of eccentricity in each layer. The RNFL was excluded from eccentricity comparisons due to its nasal to temporal (rather than concentric) distribution.⁵¹ 

Interlayer Correlation Analysis

To determine whether normal differences between retinal layers were related, gridwise correlations for AMD_total eyes were performed for all combinations of the GCL, IPL, INL, OPL, ONL_+HFL, IS/OS, and RPE–BM. The RNFL was excluded due to its relatively complex topographic localization with other retinal layers.⁵² Correlations between the inner retina (namely, the GCL or IPL) versus the outer retina (namely, the ONL_+HFL, IS/OS, or RPE–BM) were adjusted according to the displacement function of ganglion cells from their cone inner segments averaged from principal meridians as described by Drasdo et al.¹⁵ Specifically, average eccentricity of each outer retinal grid was plotted using an Akima spline, and subsequent displacement (along the same angular plane from the fovea) of each corresponding inner retinal grid was interpolated. Each outer retinal grid (Fig. 2A) was then correlated against the corresponding displaced inner retinal grids (Fig. 2B). In cases where some outer retinal grids corresponded to more than one inner retinal grid, these inner retinal grids were averaged. Thus, some outer retinal grids remained without a corresponding inner retinal grid (396/3600) (Fig. 2B, gray) and were excluded from interlayer comparisons that adjusted for displacement. Other retinal grids (544/3600) (Figs. 2A, 2B, white) were not adjusted for displacement, as their eccentricities lay beyond the displacement function.¹⁵ Adjustments for the OPL and INL were not performed due to lack of available formulae in the literature. Results were converted to absolute correlations (|r|) to focus on strength of the relationships.⁵³ 

Figure 2.

 

Comparisons between the outer and inner retinal grids according to the displacement function of Drasdo et al.¹⁵ Corresponding grids between the (A) outer retina (RPE–BM, IS/OS, or ONL_+HFL) and (B) inner retina (IPL or GCL) are indicated by color in the columns and rows. Inner retinal grids that did not have a corresponding outer retinal grid are gray. Retinal grids that were unadjusted for displacement are white. The middle black cross denotes the foveal center.

Dose–Response Analyses

Dose–response relationships were explored to address possible cause and effect between (1) AMD and retinal thickness normal differences,^54,55 and (2) AMD interlayer normal differences. First, cluster analysis was repeated for and then compared between AMD_early and AMD_int eyes. Cluster normal differences were calculated using AMD_total clusters, as these were assigned in consideration of statistically similar normal differences between disease severities. Second, interlayer correlation analysis was also repeated for and then compared between AMD_early and AMD_int eyes. Resultant values were denoted as severity differences. 

Statistical Analysis

Statistical analyses were performed using Prism 9.2 (GraphPad, San Diego, CA, USA), SPSS Statistics 25 (IBM Corp., Armonk, NY, USA), and Excel 2108 (Microsoft, Redmond, WA, USA). Significance was considered as P < 0.05. Sex was dummy coded for regression modeling (i.e., males = 0 and females = 1). Normality was tested using the D'Agostino–Pearson test. Comparisons were performed using each participant's data as a single unit of observation rather than each grid where possible, the former representing unpaired or unrelated observations. Single comparisons between continuous variables (including cluster normal differences) were then performed using unpaired Student's t-tests or Mann–Whitney U tests depending on normality. Comparisons between related continuous variables (i.e., within-layer cluster comparisons) were performed using a mixed-effects, repeated-measures model with non-assumption of sphericity (equal variability of differences; Geisser–Greenhouse correction) and Holm–Sidak's multiple comparisons test. Comparisons between categorical variables were performed using Fisher's exact test. Comparisons between paired proportions (i.e., cluster area asymmetry comparisons) were performed using McNemar's test. Multiple single comparisons of unpaired observations were performed using unpaired Student's t-tests or Mann–Whitney U tests without adjustment. No statistical adjustment was performed, as each comparison was considered individually important⁵⁶; instead, these results were considered contextually among all other results. Normal differences in SD units were interpreted according to Cohen's effect sizes: ≥0.2 = small, ≥0.5 = medium, and ≥0.8 = large.⁵⁷ Correlational analyses were performed using Pearson's correlation coefficient.⁵⁷ Correlation strength were interpreted according to Schober et al.,⁵⁸ such that |r| < 0.1 = negligible, |r| < 0.4 = weak, |r| < 0.7 = moderate, |r| < 0.9 = strong, and |r| ≥ 0.9 = very strong. Correlation coefficients were transformed using Fisher's Z-transformation for comparisons, then back-transformed.⁵⁹ 

Results

Participant Demographics

One eye from each of a total of 192 individual participants was used in this study: 96 AMD_total eyes (48 AMD_early, 48 AMD_int) and 96 normal eyes. Expectedly, there were no significant differences regarding age, sex, or spherical equivalent refraction among any groups following propensity-score matching (Table). 

Table.

 

AMD and Normal Participant Demographics, Including Separate AMD_early and AMD_int Groups

 

Topographical Differences in the Inner Retina Via Cluster Analysis

Gridwise thicknesses of each layer were compared between AMD_total and corrected normal eyes, and resultant normal differences were clustered. The cluster means of each individual retinal layer were established to be separable by ≥1.96 SD and further statistically confirmed using a mixed-effects, repeated-measures model and Holm–Sidak's multiple comparisons test (P < 0.0001 for all). For the RNFL, grids were assigned into three clusters that were all thinned relative to normal (C_{−3, −2, −1}) (Fig. 3A). There was extensively thinned RNFL inferiorly and partly superiorly in AMD_total eyes (C₋₃ mean, −8.13 µm; 95% CI, −11.12 to −5.13 and C₋₂ mean, −3.47 µm; 95% CI, −4.91 to −2.03) (P < 0.0001 for both) (Figs. 3A, 3B) that occupied half (49.7%) of the macula scan area. The asymmetry was confirmed with spatial division of C_{−3, −2} into halves (Supplementary Fig. S1A), whereby C_{−3, −2} had more extensive coverage in the inferior than superior half (34.7% vs. 15%, respectively; P < 0.0001). The remaining scan area (C₋₁, 50.4%) was slightly thinned (−0.77 µm; 95% CI, −1.46 to −0.07; P < 0.01). Large effect sizes (Z-scores calculated using the mean and distribution of the sampled corrected normal population) were observed in C₋₃ and C₋₂ (−1.62 SD and −1.6 SD, respectively), whereas C₋₁ had medium effect size (−0.53 SD). Exact cluster sizes and normal differences are provided in Supplementary Table S1. 

Figure 3.

 

Cluster analysis for AMD_total eyes in the inner retina. Normal differences (µm) for each cluster are presented as graphs (mean [95% CI]) for the (A) RNFL, (C) GCL, (E) IPL, and (G) INL, with significance above each data point from unpaired Student's t-tests or Mann–Whitney U tests: *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. Normal differences in SD units are included below the x-axis. Normal differences for each cluster are then presented as topography maps for the (B) RNFL, (D) GCL, (F) IPL, and (H) INL. Presentation is as described for Figure 1D.

For the GCL, grids were assigned into three clusters, again mostly exhibiting thinned normal differences (C_{−2, −1}) (Fig. 3C). The thinned GCL clusters were located paracentrally and at the peripheral macula more so superiorly and temporally (C₋₂ mean, −2.49 µm; 95% CI, −3.5 to 1.47; P < 0.0001 and C₋₁ mean, −0.67 µm; 95% CI, −1.2 to −0.14; P < 0.01) (Figs. 3C, 3D), accommodating 78.1% of the macular scan area. Large effect sizes were observed in these thinned clusters (C_−2, −1.34 SD; C₋₁, −1.06 SD). Asymmetry was confirmed with division of C_{−2, −1} into quadrants (Supplementary Fig. S1B), with more coverage in the superior than inferior (23% vs. 16.2%, respectively; P < 0.0001) and temporal than nasal quadrants (23.1% vs. 17.9%, respectively; P < 0.0001). The linear regression slope of the GCL cluster normal difference (µm) versus eccentricity (µm) was significant (C_{−2, −1, 0}, β = 0.71; 95% CI, 0.67–0.75; P < 0.0001), implying thinned GCL (which began paracentrally) was lessened with increasing eccentricity. There was no GCL normal difference centrally or at parts of the peripheral macula, particularly inferiorly and nasally (C₀, 0.5; 95% CI, −0.08 to 1.07; P = 0.09). 

Analysis of the IPL grids led to three assigned clusters (C_{−2, −1, 0}) (Fig. 3E). There was extensively thinned IPL covering 70% of the macular scan area (C₋₂, −1.55 µm; 95% CI, −1.95 to −1.16; P < 0.0001 and C₋₁, −0.75 µm; 95% CI, −1.11 to −0.39; P < 0.01) (Figs. 3E, 3F) with large effect sizes (C₋₂, −2.33 SD; C₋₁, −1.22 SD). Scattered parts of the IPL showed no normal difference, more so toward the peripheral macula and inferiorly (C₀, 0.18 µm; 95% CI, −0.28 to 0.64; P = 0.44; 30% macular scan area). Correspondingly, there was more coverage of C_{−2, −1} in the superior than inferior quadrant (22.7% vs. 15.8%, respectively; P < 0.0001). These results were supported by the linear regression slope of IPL cluster normal difference versus eccentricity, which was significant (C_{−2, −1, 0}, β = 0.08; 95% CI, 0.04 to 0.12; P < 0.001), implying lesser thinned IPL with increasing eccentricity. 

Analysis of the INL grids led to four assigned clusters (C_{−2, −1, 0, +1}) (Fig. 3G). The two thinned clusters (C_−2, −1.66 µm; 95% CI, −2.16 to −1.16; P < 0.0001 and C₋₁, −0.59 µm; 95% CI, −0.99 to −0.19; P < 0.05) (Figs. 3G, 3H) were situated paracentrally and comprised a majority of the macular scan area (55.2%) with large effect sizes (C_−2, −2.56 SD; C₋₁, −1.04 SD). The remaining clusters showed either no normal difference (C₀, 0.3; 95% CI, −0.07 to 0.68; P = 0.44) or thickened INL (C₊₁, 1.58 µm; 95% CI, 1.07–2.09; P < 0.0001; 8.9% macular scan area) at the central and peripheral macula. The thickened INL (C₊₁) had a large effect size (2.25 SD). The linear regression slope of INL clusters was significant (C_{−2, −1, 0, +1}, β = 0.55; 95% CI, 0.51–0.58; P < 0.0001) and implied lesser thinned and greater thickened INL with increasing eccentricity. 

Topographical Differences in the Outer Retina Via Cluster Analysis

For the OPL and ONL_+HFL, analyses of each individual layer revealed an inverted topographical pattern of the other (Supplementary Fig. S2). This suggested that disorganization of the HFL could potentially be creating artifactual interrelated cluster results in the OPL and ONL_+HFL. Subsequently, the two layers were combined into the OPL/ONL_+HFL, and two clusters were assigned (C_{−1, 0}) (Fig. 4A). There was thinned OPL/ONL_+HFL scattered paracentrally and peripherally, mostly superiorly and inferiorly (C₋₁, −1.31 µm; 95% CI, −2.06 to −0.55; P < 0.05; 68.3% macular scan area) (Figs. 4A, 4B) with large effect size (C₋₁, −1.87 SD). The remaining area (31.7% macular scan area) showed non-significant normal difference (C₀, 0.37 µm; 95% CI, −0.47 to 1.18; P = 0.58). The linear regression slope was significant (C_{−1, 0}, β = −0.26; 95% CI, −0.33 to −0.2; P < 0.0001). This implied greater thinned OPL/ONL_+HFL with increasing eccentricity. 

Figure 4.

 

Cluster analysis for AMD_total eyes in the outer retina. Normal differences (µm) for each cluster are presented as graphs (mean [95% CI]) for the (A) OPL/ONL_+HFL, (C) IS/OS, and (E) RPE–BM and as topography maps for the (B) OPL/ONL_+HFL, (D) IS/OS, and (F) RPE–BM. Presentation is as described in Figure 1D.

For the IS/OS, grids were assigned to two clusters (C_{−1, 0}) (Fig. 4C). Only the central cluster was significant showing thinned IS/OS (C₋₁, −1.25 µm; 95% CI, −2.1 to −0.41; P < 0.001; 3.6% macular scan area) (Figs. 4C, 4D) with large effect size (C₋₁, −1.47 SD). The remaining non-central cluster showed non-significant normal differences (C₀, 0.19 µm; 95% CI, −0.38 to 0.75; P = 0.52). The linear regression slope was significant (C_{−1, 0}, β = 1.99; 95% CI, 1.96 to 2.01; P < 0.0001) and implied lesser thinned IS/OS with greater eccentricity. 

Finally, RPE−BM grids were assigned to three clusters (C_{−1, +1, +2}) (Fig. 4E), with thickened RPE−BM centrally (C_+1, 0.62 µm; 95% CI, 0.29–0.96; P < 0.01 and C₊₂, 2.99 µm; 95% CI, 0.97–5.01; P < 0.0001; 41.3% scan area) (Figs. 4E, 4F) with large effect sizes (C₊₁, 2.54 SD; C₊₂, 9.93 SD), and thinned RPE−BM peripherally (C₋₁, −0.28 µm; 95% CI, −0.54 to −0.04; P < 0.01; 58.8% scan area) with large effect size (C₋₁, −1.05 SD). An eccentricity effect was supported by the linear regression slope, which was significant (C_{−1, +1, +2}, β = −0.52; 95% CI, −0.54 to −0.49; P < 0.0001) and implied greater thinned and lesser thickened RPE−BM with increasing eccentricity. 

Interlayer Correlation Analysis

To then determine whether normal differences between different retinal layers were related, gridwise correlations for AMD_total eyes were performed for all combinations of the GCL, IPL, INL, OPL, ONL_+HFL, IS/OS, and RPE–BM. The OPL and ONL_+HFL were analyzed individually because interlayer correlation analysis compares gridwise thicknesses and should be unaffected by artifactual interrelated clustering, unlike cluster analysis. Correlations between the GCL or IPL versus the ONL_+HFL, IS/OS, or RPE–BM included adjustment for displacement of Henle's fibers. 

Overall interlayer mean correlations were negligible-to-moderate strength (0.09 ± 0.09 to 0.46 ± 0.19). This was regardless of the number of statistically significant grids associated with correlations (458/3600 to 2987/3600 grids, P < 0.05). Correlations were strongest in the GCL versus IPL (|r| mean ± SD, 0.46 ± 0.19; 3141/3600 grids significant), GCL versus INL (|r| mean ± SD, 0.24 ± 0.14; 2009/3600 grids significant), INL versus OPL (|r| mean ± SD, 0.22 ± 0.14; 1712/3600 grids significant), and OPL versus ONL_+HFL (|r| mean ± SD, 0.22 ± 0.19; 1352/3600 grids significant) (Fig. 5). 

Figure 5.

 

Interlayer correlation analysis for AMD_total eyes. The strength of correlation is presented in red (left) cells, where |r| is expressed as mean ± SD and pictorially represented on a shaded gradient scale: |r| = 0.1 to |r| = 0.5 to |r| > 0.9 (light orange to red); |r| < 0.1 (white). The number of statistically significant grids is presented in green (right) cells, expressed as counts and pictorially represented as light green shading: P < 0.05. Note that there are 3600 total grids but 3204 grids after adjustment for the displacement of Henle's fibers for comparisons between the GCL or IPL versus the ONL_+HFL, IS/OS, or RPE–BM.15 Excluded grids are shown in gray.

Dose–Response Analyses

Finally, dose–response analyses were performed to determine if topographical differences and interlayer correlations were associated with AMD severity. First, cluster normal differences (µm) were recalculated for AMD_early and AMD_int eyes as separate groups (Supplementary Table S2). All clusters for AMD_early and AMD_int eyes showed consistent direction of significant normal differences; for example, a cluster that demonstrated significantly thinned retina for AMD_early was also thinned for AMD_int, albeit with different magnitudes. Normal differences of AMD_early were then subtracted from AMD_int eyes, and resultant values were denoted as severity differences (µm). Comparing AMD_int and AMD_early eyes, severity differences were positive in most clusters of all layers (Supplementary Table S2), suggesting mostly thicker retina with greater AMD severity. This, however, only reached significance in the RPE–BM centrally (C₊₁, 1.07 µm; 95% CI, 0.97–1.18; P < 0.01 and C₊₂, 5.44 µm; 95% CI, 4.88–6.00; P < 0.01) (Figs. 6A, 6B). Second, interlayer correlations were recalculated for AMD_early and AMD_int eyes as separate groups (Supplementary Table S3). All correlations between AMD_int and AMD_early eyes were non-significant, highlighting that interlayer relationships were not affected by AMD severity. 

Figure 6.

 

Dose–response cluster analysis comparing AMD_int and AMD_early eyes in the RPE–BM. Severity differences (µm) for each cluster are presented as (A) a graph (mean [95% CI]) and (B) a topography map. Presentation is as described in Figure 1C. Clusters are as described in Figures 4E and 4F. Cluster severity differences are represented via a gradient color scale: more thickened, darker teal; no significant difference, gray. No other retinal layers demonstrated significance severity differences.

Discussion

OCT high-density thickness analysis revealed anatomical differences between early/intermediate AMD and normal eyes with unique topography and large effect sizes across all retinal layers. Interestingly, differences between retinal layers were not strongly related or affected by disease severity, suggesting that it is unlikely that there is a localized cause and effect between outer retinal to inner retinal changes. These results imply that thickness differences across all retinal layers are established from early AMD, and that clinical OCT assessment of the outer retina alone does not sufficiently depict complete retinal integrity in early/intermediate AMD. 

Anatomical Differences Present Across All Retinal Layers from Early AMD

AMD eyes displayed concurrent anatomical topographical differences from normal across all retinal layers, present in early and intermediate AMD. In the inner retina, previous studies comparing early and/or intermediate AMD to normal eyes across global or low-density areas have reported mostly thinned GCL and IPL,^3–10,60,61 in agreement with our current results, possibly signifying reduced density of ganglion cells and their synapses with INL neurons.⁶² Expectedly, studies have also reported reduced peripapillary RNFL thicknesses in early/intermediate AMD,^10,60 although, puzzlingly, many studies suggest that the macular RNFL is unaltered,^3,5–7 possibly due to sparse distribution at the macula.^{23,51,63–66} Recently, our moderate-density analysis highlighted potential changes in the macular RNFL of intermediate AMD eyes, although with an indistinguishable topographical pattern.¹⁷ The greater (56×) lateral resolution and modified cluster application afforded by the current study confirms extensively thinned macular RNFL in both early and intermediate AMD eyes which now corresponds to the concomitantly thinned underlying GCL. 

Previous studies have also reported thinned INL via global or low-density retinal assessments in early and/or intermediate AMD compared to normal eyes,^3–5 possibly signifying reduced densities of bipolar, horizontal, amacrine, and/or Müller cells.⁶⁷ Our spatially refined topographical analysis confirms that the thinned INL is mostly paracentral, but it also revealed that there was less thinned and greater thickened INL with increasing eccentricity which may be associated with changes in INL neurons and/or glia (Müller) cells with comparable topography.^24,67–69 Thickening in the INL may be further corroborated by other studies describing inner retinal remodeling in outer retinal degenerations, such as outgrowth of rod bipolar dendrites⁷⁰ and horizontal and amacrine cell neurites⁷¹ and upregulation of Müller cells.^72–74 Thickened INL centrally, on the other hand, was peculiar considering that there should be an almost non-existent cellular population within ∼1 mm eccentricity,^67,69 and it may represent distortion of retinal layers from underlying drusenoid elevations rather than true cellular proliferation or hypertrophy. 

In the outer retina, previous studies using global or low-density spatial assessments of early and/or intermediate AMD eyes compared to normal have reported thinned OPL, ONL, and/or IS/OS^3,5,75–77 and thickened RPE–BM.^3,5,78–80 Our preceding moderate-density analysis indicated thinned OPL and ONL_+HFL, no significant IS/OS differences, and thickened RPE–BM centrally.¹⁷ Accordingly, our current high-density analyses upheld these findings and revealed some additional details. The OPL/ONL_+HFL and RPE–BM were more greatly thinned with increasing eccentricity, which may relate to rod susceptibility in AMD^24,25,81 reflected in structural^2,82–87 and functional^88–93 measures. The centrally thinned IS/OS could ambiguously be true outer segment loss, although this would not explain why non-central areas are unaffected, or a mechanical disorientation of photoreceptor segments from underlying drusenoid elevations, as has been described with adaptive-optics OCT.⁸⁴ 

Unexpected Cause-and-Effect Relationships

A series of studies including this one have revealed neuronal, synaptic, and even vascular differences across the inner and outer retina of AMD compared to normal eyes.^16,17,94 However, the relational implications of these findings are unclear. Using dose–response analysis, this study sought to clarify whether there may be cause and effect between AMD and thickness normal differences across all retinal neuronal and synaptic layers. We observed that, from early to intermediate AMD, the central RPE–BM was thicker, expectedly reflecting increased drusen load. Additionally, the majority of clusters within all retinal layers were less thinned and more thickened, which interestingly may reinforce potential inner retinal remodeling in AMD as described above,^68,72–74 such as cellular hyperactivity and membrane hyperpermeability,⁹⁵ although none (except the RPE–BM) reached statistical significance. Nevertheless, the consistent direction of differences from normal early and intermediate AMD eyes reaffirms that topographical thickness differences are mostly established from early AMD and strengthen the likelihood that AMD is a causative factor of these differences rather than some unmeasured variable or random chance. 

We also used dose–response analysis to assess interlayer normal differences which subsequently were not strongly related in early/intermediate AMD or affected by AMD severity.¹⁷ This suggested that there is unlikely a localized cause-and-effect relationship between progressive outer retinal changes (such as thickened RPE–BM from accumulating drusen) and post-photoreceptor alterations (such as extensively thinned RNFL, GCL, IPL, and paracentral INL), as proposed in the anterograde transsynaptic¹¹ and ischemia postreceptoral degeneration¹² pathophysiological models of AMD. Alternatively, interlayer relationships in AMD may still exist locally via degenerative bioactive molecules^96,97 or some other mechanisms undetectable via OCT thickness analyses or, more broadly, via systemic inflammatory pathways.^98–101 Nonetheless, the lack of interlayer relationships in this study intimates that the extensive anatomical differences across all retinal layers in early/intermediate AMD cannot be assessed via OCT analysis of the outer retina alone. Greater focus on the inner retina may thus be warranted in the future clinically, as well as in the development of therapies for AMD such as optogenetics and prosthetic devices that rely upon sustained post-photoreceptoral health for intervention success.⁹⁵ 

Limitations

The primary limitation of this study relates to inferences drawn from OCT thickness data that do not capture specifically which cellular and synaptic processes^67,102 nor physiological mechanisms may be altered in disease. Further development of our work with more comprehensive co-variable correction such as axial length, use of longitudinal data, and comparison to other measures of retinal integrity such as a high-magnification or adaptive-optics OCT could illuminate the relationship between inner and outer retinal changes in the early stages of AMD. 

Additionally, although we demonstrated statistically significant anatomical topographical differences across all retinal layers, it is unclear whether these differences are clinically important. There are no established guidelines that define clinically important thickness differences for each individual retinal layer, particularly in the context of manual segmentation. To enable transparency of effect sizes and additionally account for varying thicknesses across the macula in each retinal layer, we reported thicknesses in both micrometers and SD units. Our dose–response analysis also helped to answer whether thickness differences within and between retinal layers in early/intermediate AMD are isolated phenomena or demonstrate any clinically detectable cause and effect. Future works will need to explore the clinical importance of thickness differences across all retinal layers. 

Conclusions

Anatomical differences between early/intermediate AMD and normal eyes display unique topography with large effect sizes and are mostly established from the early stage. Interlayer differences, though, are not strongly related regardless of AMD severity. These results elucidate where changes may occur and emphasize that complete retinal integrity cannot be assessed via OCT of the outer retina alone in early/intermediate AMD. 

Acknowledgments

The authors thank Janelle Tong (Centre for Eye Health, Sydney, Australia) for methodological advice and Vincent Khou (Centre for Eye Health, Sydney, Australia) for retinal layer segmentation. 

Supported, in part, by grants from the National Health and Medical Research Council of Australia (NHMRC; 1186915 to MK and DAC and 1174385 to LNS) and by the Rebecca Cooper Foundation. MT is supported by an Australian Research Training Program scholarship. Guide Dogs NSW/ACT provides support for the Centre for Eye Health (the clinic of recruitment) and salary support for MK. 

Disclosure: M. Trinh, None; M. Kalloniatis, None; D. Alonso-Caneiro, None; L. Nivison-Smith, None