Quantification of Normative Human Choriocapillaris Measures Across the Macula

Abstract

Purpose

To assess macular choriocapillaris (CC) metrics in healthy volunteers (HVs) without ocular disease and demonstrate CC variations in patients with inherited retinal dystrophies (IRDs) using adaptive optics optical coherence tomography angiography (AO-OCTA).

Methods

Twenty-one HVs and three IRD patients were imaged. Macular variation in 20 HVs in CC metrics (CC density, CC diameter, CC tortuosity, void diameter, void area, lobule count, lobule area, and RPE–CC distance) were assessed by imaging a 28° strip of overlapping AO-OCTA volumes (3° × 3°) from the optic nerve head to the temporal macula. En face projections of the CC at 1° intervals across the strip were created and metrics extracted using custom software.

Results

CC density was mostly invariant across the macula at 0.523 ± 0.006, although it was slightly higher in the foveal region. The CC diameter decreased monotonically from 16.1 ± 0.6 µm in the nasal macula (−12°) to 14.2 ± 1.1 µm in the temporal macula (12°). Void measures also decreased from the nasal to the temporal macula: from 16.3 ± 1.1 µm to 14.4 ± 1.3 µm for void diameter and from 1457 ± 398 µm² to 1214 ± 369 µm² for void area. Lobule count peaked in the central macula. RPE–CC distance peaked at the fovea (fovea, 16.8 µm; periphery, ∼13–14 µm). CC tortuosity (1.17 ± 0.01) was constant across the macula. We observed a correlation with age in CC density, void diameter, and void area. AO-OCTA detected subclinical flow dropout in drusen regions. CC changes were observed in IRD patients with characteristics specific to the disease.

Conclusions

Normative AO-OCTA–based CC metrics can provide insight into ocular disease pathology.

The choriocapillaris (CC) is the most anterior (inner) vascular layer of the choroid. It supplies the primary metabolic needs of the photoreceptor (PR)–retinal pigment epithelium (RPE) complex and is the densest capillary network in the human body.^1–3 As such, it plays an important role in the homeostasis of the cells responsible for visual phototransduction, and in the initiation and progression of a broad range of retinal diseases, including age-related macular degeneration (AMD),^4,5 central serous chorioretinopathy,⁶ diabetic retinopathy,⁷ glaucoma,⁸ and ultimately in any retinal disease that affects the photoreceptors and vision. It is also affected in systemic diseases like diabetes,⁹ hypertension,¹⁰ and multiple sclerosis.¹¹

Attempts to image the CC over the years have been challenging. Exogenous dyes are used in angiography for the visualization of choroidal and retinal circulation. Indocyanine green angiography (ICGA) is more effective than fluorescein angiography in imaging the choroid owing to better tissue penetration from its longer absorption and emission wavelengths, as well as near complete binding to protein that prevents diffusion of ICG molecules through fenestrations in the CC endothelial wall.^12–14 The primary difficulty with ICGA lies in its inability to separate the CC signal from the underlying middle (Sattler's) and outermost (Haller's) choroidal layers,¹⁵ although adaptive optics–ICGA (AO-ICGA) techniques have shown some promise.¹⁶ The demonstration,¹⁷ development,¹⁸ and use¹⁹ of optical coherence tomography angiography (OCTA) has enabled visualization and some initial quantification of CC metrics, although care must be taken in depth layer segmentation²⁰ and analysis methodology,²¹ among other considerations.²²

Despite this attention to detail, measurement differences across devices and questions of accuracy persist.²³ For example, CC density has been reported to be as high as 0.70,²⁴ and CC diameter has been reported to be in the range of 20 to 25 µm,^24,25 despite previous histology that has reported a lower range of 15 to 20 µm.²⁶ It is now understood that clinical OCTA inaccuracy (overestimation) in CC vessel structural measurements has to do with the limited lateral resolution of the imaging instrument. Clinical OCTA lateral resolution is typically on the order of 10 to 20 µm, which will produce a blurred image of targets whose dimensions are below this value, even with pixel oversampling. This limitation was further explored by Zhou et al.,²⁷ who increased the input beam diameter and resultant numerical aperture to demonstrate better delineation of CC capillaries, although not necessarily more accurate measures, possibly owing to uncorrected ocular aberrations. There are other causes of inadequate CC visualization in other imaging modalities, including axial resolution (i.e., poorly sectioned CC) in nonconfocal fundus imaging or confocal scanning laser ophthalmology (SLO) imaging, coupled with signal blocked by the strongly waveguided photoreceptors and strongly (wavelength-dependent) absorbing RPE, which generally renders these modalities unable to resolve the CC or choroidal flow for imaging and quantification.²⁸ For clinical OCT, the primary factor limiting full CC visualization and characterization is lateral resolution. The ophthalmology field has turned to other biomarkers that may be more accurately quantified in light of this limitation, including the measurement of flow voids or flow deficits,²⁹ or relative measures.³⁰ A full and accurate characterization of in vivo human CC anatomy has thus been lacking, despite the obvious usefulness of this information for the diagnosis and tracking of retinal disease.

New OCTA imaging technologies have provided opportunities to improve CC contrast and resolution. High-speed A-line rates in the mega-Herz (MHz) range suffer less eye motion artifact, which can improve the ability to section the CC leading to higher contrast capillary vessel segments.^31,32 The improved lateral resolution provided by AO via the correction of ocular aberrations coupled with the depth sectioning capabilities of OCT and its accompanying dye-free angiography methodology (AO-OCTA) has also been demonstrated for high-quality CC imaging.³³ This initial high-speed OCTA and AO-OCTA studies provided CC size measurements closer to histology, though some variability across studies persisted. The primary limitation of AO for clinical adoption has been its limited field-of-view (FOV), and, although those previous studies imaged at multiple eccentricities, the proportion of macular surface area imaged in previous AO studies has been relatively small.

The aim of the current study was to demonstrate accurate measures of CC structure with AO across the entire macula in a clinically viable and efficient imaging protocol that can be used to establish normative values in healthy eyes for further validation (e.g., repeatability studies) and comparison with diseased eyes. To that end, we used ultrahigh-speed Fourier domain mode-locked (FDML)–based AO-OCTA^34,35 to visualize the CC across a 28° (∼8.2 mm) span of the macula. Furthermore, we measured CC density, CC diameter, CC tortuosity, void diameter, void area, lobule count, and RPE–CC distance in a cohort of healthy participants and the results were compared with previous histology and imaging studies. We also demonstrate our technique in a limited number of diseased eyes in an early state of progression. Our study marks a further step in the clinical development of AO-OCTA CC imaging and demonstration of AO-based CC biomarkers of disease.

Methods

Approvals, Participants, and Consent

The study protocol was approved by the Institutional Review Board at the U.S. Food and Drug Administration (FDA). The study adhered to the tenets of the Declaration of Helsinki. Twenty-one participants with no clinical diagnosis of eye disease were recruited, consented to participate after potential risks were described, and enrolled in the study. Three additional participants were recruited from the National Institutes of Health National Eye Institute clinical population with inherited retinal dystrophies (IRD), including cone–rod dystrophy (CRD, retinitis pigmentosa GTP-ase regulator), Usher's syndrome 2A (USH2A), and rod–cone dystrophy (RCD, CERKL). All participants underwent a standard eye examination by board-certified ophthalmologists (O.J.S., E.Y.C., or C.C.) before AO imaging.

AO-OCT Imaging Session

The FDA FDML AO system has been described previously.³⁴ The AO-OCT channel of the FDML AO system, which operates at a 3.4-MHz line rate and 12.97 vol/s (512 × 512 lateral pixels), was used in AO-OCTA mode. The system and image postprocessing pipeline were enhanced to include a phase correction approach to compensate buffer variations in the FDML laser. Full characterization of the AO-OCTA capabilities of the ultrahigh-speed system can be found elsewhere.³⁵

For all participants, the following system and scan parameters were used. A single 28° strip across the macular from the rim of the optic nerve head (ONH) to the temporal retina was imaged in 11 discrete locations with a FOV of 3° × 3° and 2.5° separation (∼17% overlap), as shown in Figure 1. The slow scan galvanometer was set to eight repeated B-scans for each lateral location. The effective volume rate was 1.62 vol/s and the fixed voxel separation time was 0.3 ms (for the bidirectional 6.64-kHz B-scan rate). Two AO-OCTA videos, each composed of two volumes, were collected at each location to account for the possibility of blinks during the scan, where 8192 B-scans (2 vol. × 512 pixel × 8 repeats) were acquired and saved in approximately 20 s. Including redirection of gaze and participant breaks, the total time to collect volumes at all locations was approximately 20 minutes. The system focus was set to maximize signal from the PR–RPE complex for best visualization of the closely situated underlying CC.

Figure 1.

Wide-field AO-OCTA CC mapping. Example montages stitched from 11 regions in a 32-year-old healthy volunteer (HV) (3228). (A) Spectralis SLO image. (B) PR and (C) CC layers segmented from AO-OCTA volumes. PR layer is the composite projection of inner segment/outer segment junction and cone outer segment tip bands with thickness of 25 µm and CC layer projection thickness is 10 µm. Colored ROIs in (B) and (C) are magnified in (D). ROIs in (C) are sampled locations where CC quantification was performed. Scale bar, 200 µm. (D) The 3× magnified regions at 11T, 6T, 0, 1N, 4N, and 9N showing sampled regions across macula.

AO-OCTA Methodology

After the AO-OCTA volumes were acquired, they were processed with a custom algorithm that includes standard OCT processing, including k-space mapping, digital dispersion compensation, and fast Fourier transform, among other steps. Both AO-OCT intensity and AO-OCT angiography volumes were flattened to the PR-RPE complex. A speckle variance OCTA methodology was used to separate static reflectance and delineate the CC capillary network.³⁵ The variance in the speckle signal across the four repeat scans in each direction of the bidirectional scan was calculated. Our implementation convolves the temporal variation signal from all combinations of B-scans, yielding a temporal decorrelation range of 0.3 to 1.2 ms for our scan settings. The angiography images in the forward and reverse directions were averaged to produce the final angiography B-scan. The highest quality AO-OCTA volume was selected by visual inspection from the four acquired volumes for further analysis. In several cases (3/21 HVs), volume averaging was performed³⁵ because the OCTA signal from a single volume was insufficient for CC quantification owing to lens opacity or a small pupil. For all other participants, only a single AO-OCTA volume was needed for CC quantification.

En face CC images were created from the AO-OCTA volumes by summing depth pixels in a slab approximately 7 to 11 µm thick (5–7 pixels) around the CC peak location approximately 15 µm (10 pixels) below the RPE layer using average intensity projection, where the thickness varied depending on location and OCTA signal strength. The PR en face images were also generated from the corresponding AO-OCT intensity volumes by using averaged intensity projection from the cone inner segment/outer segment junction layer to the cone outer segment tip layer to facilitate montage creation and fovea identification. The en face CC images were then manually stitched together in Photoshop (Adobe, San Jose CA, USA) to create the final montage across the macula, where the best quality and match in overlapped regions was chosen by visual inspection of PR and CC structures. From the en face CC macular montage, 1° × 1° regions of interest (ROIs) were extracted, spaced at approximately 1° increments, accounting for eye length associated lateral magnification, for further extraction of CC metrics (Fig. 1). The fovea location was determined from the PR montage, where the PR density is greatest at the center of the relatively dark pit region (Fig. 1B). The ROIs were allowed to shift laterally a small amount (<0.25°) to avoid shadows from large overlying vessels and other artifacts and where the image quality from visual inspection was greatest.

CC Vascular Methodology

Custom software was developed to quantify CC metrics, which is further described elsewhere.³⁵ The code was comprised of custom elements, standard Matlab commands (MathWorks, Natick, MA, USA), and ImageJ plug-ins and scripts (U.S. National Institutes of Health, Bethesda, MD, USA). Starting from the raw, en face, 1° × 1° CC regions, the ROI image was resampled by 1.5× using bicubic interpolation and binarized using a custom algorithm, which included Gaussian filtering, thresholding, vessel tubular enhancement, and particle filtering. The intensity threshold, vessel thickness, sigma-gauss, and particle filter parameters could be adjusted by a grader with visual inspection (Fig. 2). To aid parameter selection, the software output a color-coded overlay of the binarized map on the CC image, as well as the binary vessel map and its negative (void map) on the CC image to aid visualization of capillary and void regions, respectively (Fig. 2C). Moreover, the software calculated the signal-to-noise ratio from the average OCTA flow signal in the capillary regions and the standard deviation of the OCTA noise signal in the void regions that were also used to guide parameter selection. Once the binarized image was produced, the ImageJ scripts ‘skeletonize (2D/3D)’ and ‘analyze skeleton’ were used to skeletonize the image (Fig. 2C) and calculate further metrics.

Figure 2.

Analysis methodology to extract capillary metrics from AO-OCTA montage. (A) Example AO-OCTA CC montage across 28° for a 38-year-old HV (6581). (B) Corresponding vessel binary map. (C) ROI analysis of the 0° location shown in (A) and (B) (yellow box). Shown are the 1.5× resampled vessel region (3× zoom), the binary map, the color overlay of the map on the vessel ROI, the void overlay showing the pixels without flow, the vessel over showing flow pixels, and the skeleton.

The software quantified capillary density (fraction), capillary diameter (µm), void diameter (µm), void area (µm2), and tortuosity. All measures were corrected for individual axial lengths measured with biometry using the Bennett–Rabbetts eye model.³⁶ Capillary density is the proportion of pixels occupied by vessels in the binary image divided by the total ROI pixels. Capillary diameter was calculated using Matlab commands bwskel() and bwdist(), which skeletonize and apply a Euclidian distance transform to the binary image, respectively, to automatically sizes vessels using the distance between the vessel midline (from the skeleton) and the vessel edge (from the binary mask). Void diameter and void area were calculated by inverting the binary image and sizing the resultant particles using the Matlab function bwconncomp(). Finally, tortuosity was calculated after skeletonization as the ratio of branch length to the Euclidian distance between vessel segment junctions (or nodes), while excluding any branch length less than 15 pixels (∼18 µm) where a tortuosity measure is less accurate or meaningful. More branch and node metrics (e.g., mean distance between junctions) can be extracted from the skeleton image for further analysis.

All CC metrics except RPE-CC distance and lobule count were independently calculated by two graders (D.X.H. and Z.L.) with visual inspection of each ROI while adjusting the intensity threshold, vessel thickness, and particle filter parameters. The Gaussian filter parameter (sigma-gauss) was not changed from its default setting. The software was reset to the default parameters for each ROI, which provided a reasonably good vessel segmentation starting point. From there, the intensity threshold was adjusted depending on the variable signal-to-noise ratio so that the binarized mask matched the actual vessel coverage. Infrequently, the vessel thickness parameter was also adjusted to account for regions with thicker vessels when the threshold parameter was insufficiently capable of creating the correct mask. Finally, individual particles were examined and those judged to be noise and not vessel segments were removed by adjusting the particle filter parameter. The intensity threshold and particle filter parameters were most often adjusted and the vessel diameter parameter was infrequently adjusted. The CC metrics reported below are the average values calculated by the two graders.

The RPE–CC distance was calculated manually by a single expert grader (Z.L.) in ImageJ by marking the axial pixel values corresponding with the en face views of the RPE and CC layers in the AO-OCT and AO-OCTA volumes, respectively, across the macula for each location (2.5° separation). The CC lobules were counted manual with a custom ImageJ cell counting plugin using a multipoint feature (Supplementary Fig. S2). The full CC montage was imported into the software and the location of the nexus of segments in each lobule was marked as the center of the lobule to record the coordinates of each lobule across the horizontal macular strip. Lobules were counted by two graders, first by a junior grader (K.K.) and then by a senior grader (Z.L.) upon reviewing the junior grader's counts. Lobule density across the macula was examined by grouping the lobule counts in 3° segments (nasal to temporal) centered on the fovea. The average lobule area was calculated with the assumption of continuous lobule structure across the macula by dividing the total montage area by the number of lobules counted. This approximation is expected to be reasonably accurate for healthy volunteers¹⁸ and approximately match manual subjective demarcation of the lobule regions.

Statistical Analyses

All statistical analyses were performed in Microsoft Excel with the Real Statistics Resource Pack and Analysis ToolPak add-ins, which include data analysis tools for statistical and engineering analysis. General statistical analyses included the calculation of mean, standard deviation, correlation coefficient, and 95% confidence limits. Paired t-tests were used to analyze some measures (e.g., drusen vs. non-drusen). Regression analyses with Pearson (R_P) or Spearman (R_S) correlation were used to assess the linear relationship between CC metrics (CC density, CC diameter, void diameter, void area, RPE–CC distance, tortuosity, and lobule area) and eccentricity or age. Lin's concordance coefficient (LCC) was used to assess inter-rater variability between the graders for CC metrics.³⁷ The threshold P value for statistical significance was set to 0.05.

Results

Participants

The demographic and clinical characteristics of the participants who underwent AO-OCT imaging are shown in Table 1. Of the 21 HVs, CC images from one volunteer were of insufficient quality to be included in the analysis. Three HVs (ages 38.7, 58.6, and 62.5 years) had some evidence of subclinical drusen deposits (very small drusen <30 µm in diameter), and for these cases the ophthalmologist (O.J.S.) reviewed the clinical images and found no significant evidence of clinical disease to warrant exclusion from the normative dataset. Of the three participants with retinal degeneration who had imaging attempted, we were able to collect images from two. The third subject with RCD, the oldest of the IRD patient (53.6 years old), had extremely poor vision (visual acuity = hand motion), was unable to fixate, and so no usable AO images could be collected.

Table 1.

Demographics and Eye Exam (Abbreviated) Data for Participants With Usable AO Images

Parameter	HV (n = 20)	IRD (n = 2)
Age, years (mean ± SD)	43.2 ± 13.4	44.3 ± 2.0
Range	27.7–72.3	46.0–48.8
Sex (count, %)
Female	7 (35%)	0 (0%)
Male	13 (65%)	2 (100%)
Race (count, %)
Asian	4 (19%)	0 (0%)
Black	2 (10%)	0 (0%)
White	15 (71%)	2 (100%)
Eye imaged with AO (%)
OD	60%	100%
OS	40%	0%
Axial length, mm
Mean ± SD	24.1 ± 1.0	24.2 ± 0.6
Minimum	22.1	23.8
Maximum	26.4	24.7
BCVA Snellen ≤20/25 (%)	100%	50%

BCVA, best-corrected visual acuity; OD, oculus dexter; OS, oculus sinister.

Normal CC Measures Across the Macula

We observed a continuous CC capillary network without any large regions of flow dropout for all HVs with an example shown in Figure 1. AO-OCTA macular montages (PR and CC depth sections) from all study participants are included in the Supplemental Material. There was reasonably good agreement between the two graders for the CC metric analysis, with a LCC of 0.70 (95% confidence interval [CI], 0.65–0.74), 0.91 (95% CI, 0.90–0.93), 0.69 (95% CI, 0.65–0.72), 0.69 (95% CI, 0.65–0.73), and 0.72 (95% CI, 0.68–0.76), for vessel density, vessel diameter, void diameter, void area, and tortuosity (Supplementary Fig. S3).

The mean values for the CC metrics for all HVs across the macula are shown in Figure 3. In the rim region of the ONH (>10° N), the vascular structure changes. In particular, the vessels become larger, as observed in Figure 3B. There are also differences observed in the foveal region, including very slightly greater density (Fig. 3A), smaller diameter (Fig. 3B), and correspondingly smaller void diameter/area (Figs. 3C and 3D). Otherwise, any difference observed in the CC measures across the macula was continuous.

Figure 3.

CC morphological measures across the macula. Shown are mean CC metrics for 20 HVs including (A) vessel density, (B) vessel diameter (µm), (C) void diameter (µm), (D) void area (µm²), (E) RPE–CC distance (µm), and (F) lobule count. AO montage images used to extract (A–D) are sampled from 25 ROIs spaced by 1° across the macula from 12°N to 12°T. For (E), the RPE–CC distance was measured from the summed A-line profiles for each volume separated by 2.5°. For (F), lobules were counted across the AO montage images and grouped in 3° ROIs. Upper and lower confidence limits (solid lines), and the Pearson's linear regression fit to the data (dashed line) are shown.

The mean CC density did not change across the maculae of the HVs (Fig. 3A and Supplementary Table S1) ($R_S^2=0.046;P=0.30$), maintaining a consistent value of 0.51 to 0.53 (mean, 0.523 ± 0.006 for all ROIs and HVs) from 12° temporal to the nasal rim. The mean CC diameter decreased across the macula (Fig. 3B and Supplementary Table S2) ($R_S^2=0.83$; P < 0.05) from 16.1 ± 0.6 µm at the nasal rim to 14.2 ± 1.1 µm at 12°T (mean, 14.9 ± 0.4 µm across the macula for all ROIs and HVs). In a corresponding manner, both the void diameter and area also decreased across the macula from nasal rim to temporal regions. The void diameter decreased across the macula (Fig. 3C and Supplementary Table S3) ($R_S^2=0.54;P<0.05$) from 16.3 ± 1.1 µm at the nasal rim to 14.4 ± 1.3 µm at 12° T (mean, 15.6 ± 0.5 µm across the macula for all ROIs and HVs). With a fixed vessel density and both decreasing vessel diameter and void diameter, it can be inferred that the number of capillary segments increased from the nasal rim to the temporal macula. The void area decreased across the macula (Fig. 3D and Supplementary Table S4) ($R_S^2=0.31;P<0.05$) from 1457 ± 398 µm² at the nasal rim to 1214 ± 369 µm² at 12° T (mean, 1478 ± 190 µm across the macula for all ROIs and HVs). The void area exhibited greater variability across the macula than the other measures. Tortuosity did not change across the macula (Supplementary Table S5) ($R_S^2=0.0004;P=0.92$), maintaining a constant value of 1.16 to 1.18 across the macula (mean, 1.17 ± 0.01 across the macula for all ROIs and HVs).

The RPE–CC distance peaked at the fovea and displayed a linear relationship with eccentricity in the nasal and temporal sides of the macula (Fig. 3E and Supplementary Table S6) (N: $R_P^2=0.97$, P < 0.05; T: $R_P^2=0.97,P<0.$05). The number of lobules measured in the 3° × 28° horizontal macular strip did not exhibit a linear relationship (Fig. 3F and Supplementary Table S7). Generally, the number of lobules was greater near the fovea and higher on the temporal side compared with the nasal side (mean, 37 ± 9 for the total count across the macula for all HVs). The individual participant values for all CC morphological measures are shown in Supplementary Figure S1, where the trends described above can be observed.

Effect of Aging on CC

We observed a negative correlation between age and CC vessel density (Fig. 4A) ($R_S^2=0.296;P<0.$05), with CC vessel density decreasing from approximately 0.55 at age 20 to less than 0.50 by age 70. There is also a slight but significant increase in void diameter (Fig. 4C) ($R_S^2=0.448;P<0.$05) and void area (Fig. 4D) ($R_S^2=0.633;P<0.$05), with the void diameter increasing from approximately 14 µm at age 20 to approximately 17 µm at age 70. We did not observe any change with age in CC vessel diameter (Fig. 4B) ($R_S^2=0.007$; P = 0.72) or tortuosity ($R_S^2=0.013;P=0.$63). There was an overall slight increase in the RPE–CC distance for values summed across the macula with age that did not reach statistical significance in our cohort (Fig. 4E) ($R_S^2=0.011;P=0.$14). Likewise, there was no change in lobule count ($R_S^2=0.028;P=0.$48), or lobule area (Fig. 4F) ($R_S^2=0.026;P=0.$50) with age. From these results, we infer that a loss of vessel segments occurs with age, but that vessel diameter and the overall lobule structure does not change with age.

Figures 4.

Effect of aging on CC measures. Shown are the CC values as a function of age for 20 HVs including (A) vessel density, (B) vessel diameter (µm), (C) void diameter (µm), (D) void area (µm²), (E) RPE-CC distance (µm), and (F) lobule area (mm²). Each datapoint is the mean value for that HV across all macular locations. The Pearson's linear regression fit to all the data is shown as a dashed line.

Effect of Drusen Deposits and Disease on CC

Three participants had evidence of small hard drusen deposits that can be associated with normal aging but may also be an indication of early areas of dry AMD. On the clinical OCT cross-sections, the drusen lesions were small (<100 µm in diameter), sub-RPE, and hyporeflective in multiple locations. One subject was further examined with AO-OCT with an additional 11 scans along the vertical midline in addition to the standard horizontal scan set (Fig. 5 and Supplementary Fig. S5). The extent of the drusen was seen clearest with en face projection images extracted at the depth of the RPE, where drusen was observed as multiple small hyporeflective regions across the macula (Supplementary Fig. S4B), with an apparent relatively intact overlying PR mosaic (Supplementary Fig. S4A). There were more drusen lesions and they extended further in eccentricity in the superior–inferior direction compared with the nasal–temporal direction for this subject. Note also that individual RPE cells were not resolved in the montage because only single AO-OCTA volumes were analyzed. The PR was generally elevated in these regions, as seen on the cross-sectional AO-OCT scans (Supplementary Fig. S4B). The CC network in the matched drusen regions (Supplementary Fig. S4C) was not obviously disrupted in a few regions (arrows), whereas in most regions there were fewer capillary segments beneath the drusen regions (arrowheads). This result provides some evidence that the drusen does not completely block, scatter, or distort the light passing to the CC that returns to make up the OCT signal for that layer.

Figure 5.

Example AO-OCTA montage of drusen deposits. (A) AO-OCT volumes for a 58-year-old HV (6289) were collected along vertical and horizontal strips spanning the macula shown overlaid on a Spectralis SLO image. (B) Analysis from 48 ROIs (24 each for drusen and non-drusen regions) of the segmented CC layer. (C) We observed lower CC density and higher void diameter but no difference in CC diameter in the drusen regions compared with the non-drusen regions.

From examination of the RPE montage (Supplementary Fig. S4B) by a single grader (Z.L.), we selected 24 regions that contain drusen and 24 regions that were predominantly free of drusen for further quantitative analysis (Supplementary Fig. 5B). These 48 regions were passed through the CC analysis software described elsewhere in this article. We found that the CC diameter for both drusen and non-drusen regions was the same (16 µm; P = 0.71), but higher than the average macula diameter for HVs (14.9 µm) (Fig. 3). The vessel density was significantly lower for drusen regions (0.44 ± 0.03) compared with non-drusen regions (0.51 ± 0.02; P < 0.001), where the latter was similar to that found across the macula for HVs. The void diameter was higher (23.3 ± 2.8 µm) for the drusen regions compared with the non-drusen regions (18.5 ± 1.7 µm; P < 0.001), where the latter was higher but within the CIs for comparable HVs.

We collected AO-OCTA montages across the macula in a similar manner as HVs in two additional patients, one with CRD (Fig. 6) and one with USH2A (Fig. 7). CRD is characterized by pigmented deposits, atrophy (PR–RPE disruption), and lesions in the central macula where cones are affected before rods (the sequence is reverse in RCD). These hallmarks can be seen in the clinical OCT image (Fig. 6D) and especially in the AO-OCT image of the PR layer (Fig. 6A), where individual reticular pseudo-drusen deposits are resolved. From visual inspection, the CC capillaries are preserved in the periphery but show a significantly reduced number of segments and smaller diameters, as well as increased void diameter and area in the macula (Figs. 6B and 6C).

Figure 6.

Example AO-OCTA montages in CRD. In a 48-year-old patient with CRD (AORD_004), montages were assembled from AO-OCT axial projections for (A) the PR layer and (B) the CC layer. (C) A 2× zoom of two regions outside and inside the macula lesion. (D) Clinical OCT scan through the fovea.

Figure 7.

Example of CC montage in USH2A. In a 46-year-old patient with USH2A (AORD_007), montages were assembled from AO-OCT axial projections for (A) the PR layer, and (B) the CC layer. (C) A 1.5× zoom of three regions in the temporal, foveal, and nasal regions. (D) Clinical OCT scan through the fovea.

In USH2A, rods are affected first and so the pathology proceeds from more rod-dense areas in the periphery initially and later to involve more central areas of the retina. The clinical OCT image for this patient shows the loss of PR–RPE layer in the periphery (Fig. 7D) and the AO-OCT image shows preserved cones in the central macula and hyperreflective lesions at the level of the photoreceptors in the periphery (Fig. 7A). The CC network is predominantly intact across the entire macula, although lobule organization may be disrupted in the temporal macula and on the nasal side near the disc (Figs. 7B and 7C). Also, the AO-OCTA signal from the CC network is weaker in the margin between the central region where healthy cones exist and the peripheral region with significant rod loss. The weaker signal may be related to an AO focus change across the transition zone, resulting in a suboptimal OCT signal.

Discussion

The CC is an extremely important vascular network for visual function, providing nutrient and waste exchange for the extensive metabolic demands of the photoreceptors. Variations in vascular morphology may provide clues with respect to AMD and IRD progression and may also serve as a biomarker for early disease initiation because metabolism and function are so closely linked within the neurovascular unit.

Our results reveal that the normal CC has smaller diameter vessels and smaller void regions near the fovea, which serves to supply the more densely packed PRs at that location. Overall, however, the CC density does not change significantly across the macula. CC diameter, void diameter, and void area all decrease from the nasal to the temporal macula, whereas the lobule density peaks in the central macula. This likely arises from early stages of the retinal vascular network development originating at the optic nerve head and growing into the periphery.¹ Evidence that supports this hypothesis is that, with aging, we found no change in the CC diameter or lobule area (Fig. 4B). However, with aging, the CC density decreases and voids increase, indicating a progressive, age-related remodeling of the CC network.

The RPE–CC distance is significantly greater at the fovea and decreases monotonically for more peripheral locations, roughly in an equal manner on the nasal and temporal sides of the macula. The peak at the fovea does not follow the same shape as the increased outer segment length associated with increased cone PR packing in the foveola. However, it does roughly follow an inverse relationship with rod PR density and so the smaller distance in the periphery may be related to overall perfusion demands of the rod photoreceptors. Moreover, our RPE-CC distance measurements align with post-mortem histology on human eyes by Curcio et al.,³⁸ who found the CC and Bruch's membrane (BM) together were 2.0 to 2.5 µm thicker at the fovea (11.5 µm) compared with 3° N/3° T (9.3 µm) with no change in RPE thickness over the same region. Our measurements from the fovea to 5° are less than this (∼1.3–1.8 µm), but because we measured RPE-CC distance from the center of the bands on AO-OCT, it is likely we did not fully capture layer thickness changes. Our overall RPE–CC distance values (14–17 µm) are smaller than those reported with OCTA,³⁹ but similar to those of Kurokawa et al.³³ using AO-OCTA. Our values also match Curcio et al.,³⁸ considering their reported BM thickness and one-half RPE and CC thickness values, which corresponds with our center-to-center layer measurements. Therefore, we conclude that the RPE–CC distance variation across the macula is entirely from changes in BM and CC thickness. Curcio et al.³⁸ and others^33,39 also found increased RPE–CC distance with age, similar to our findings (Fig. 4E), although our trendline did not attain significance for our relatively small cohort.

In their AO-OCT study, Kurokawa et al.³³ also found a decrease in capillary density with a very similar slope compared with our findings (−0.0010 for the current study vs. −0.0013 for Kurokawa et al.), where density decreases from approximately 0.55 at 25 years to less than 0.50 after 75 years. With OCTA, Sacconi et al.³⁰ found a similar decrease with aging in perfusion density in the macula (fovea centered), although with higher values than obtained from the AO studies (decreasing from 0.767 to 0.702 for a 1-mm circle and 0.771 to 0.718 for a 3-mm annulus for participants across the second to eighth decades of life). Although OCTA lateral resolution makes these absolute values suspect, the same negative correlation with aging was observed. Similarly, using histological techniques on 95 human eyes, Ramrattan et al.⁴⁰ found a higher slope with the density decreasing from 0.67 to 0.50 over the same age range. Ramrattan et al. also found a much smaller capillary diameter in their study, as well as an aging effect, decreasing from 9.0 µm to 7.4 µm from 25 to 75 years. The smaller diameter was not found in other histological studies,^26,41 and it is possible that tissue shrinkage or other processing artifacts (related to changes in vessel compliance with age) are responsible for this discrepancy.

Our CC measurements are compared with previous imaging and histology studies in Table 2. In general, we found favorable agreement both with previous imaging and previous histology, except in a few specific instances where methods used in previous measurements were not deemed to produce the accuracy of the current AO-OCTA technique. CC density measurements are extremely sensitive, not only to device lateral resolution but also to binarization algorithm thresholding, and although our results compared favorably with those of Kurokawa et al.,³³ who also used AO-OCTA, the 0.91 density value reported by Dai et al.⁹ is likely a large overestimation, given the inability of OCTA to well resolve individual CC segments or voids. The capillary diameter values from our study, Kurokawa et al.,³³ and Jung et al.¹⁶ match previous histology by Olver,²⁶ while the Marsh-Armstrong et al.³² values of approximately 10 µm may be a slight underestimation owing to different thresholding for binarization. Void diameter values range from 10 to 25 µm for the studies listed. Jung et al.¹⁶ reanalyzed histology presented in the papers by Olver,²⁶ Yoneya et al.,⁴³ and Zouache et al.⁴⁴ to compare to their own AO-ICG measurements of void area and found the histological values (127–401 µm²) to be similar to their result of 313 ± 103 µm². Our results are significantly higher at 1481 ± 473 µm². We performed our own reanalysis of 50 voids in the histology images of Olver and Zouache et al. and found slightly higher values of 256 µm² and 513 µm², respectively. Although these values are closer to our measurements, there remains a significant discrepancy that further investigation is needed to explain.

Table 2.

Comparison of CC Metrics in Healthy Eyes With Previous Imaging and Histology Studies

					Measure
	Study and Subject Parameters				CC Density		CC Diameter (µm)		Void Diameter (µm)		Void Area (µm2)		Lobule Size (µm)		RPE-CC Distance (µm)
Study	Method	n	Age, Years	Eccentricity (°)	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Mean	SD
Current	AO-OCTA	20	28 to 72	−12 to 12	0.52	0.01	14.9	0.5	15.6	0.6	1478	198	443	55	15.1	1.1
Kurokawa et al.33	AO-OCTA	9	26 to 68	1 to 7	0.53	0.06	17.4	1.1							17.5	1.6
Jung et al16	AO-ICG	15	32 ± 9	fovea			17.0	2.9			313	103
Marsh-Armstrong et al32	OCTA	4	NA	−9 to 3			10.0	0.6	19.6	1.4	749	110
Zhou et al27	OCTA	4	28 to 54	fovea					24.3	0.9					19.2	1.1
Dai et al8	OCTA	16	NA	−17 to 17	0.91	0.03					1651	280
Fryczkowski et al40	Histology	36	NA	PP			20 to 35						400 × 650
Fryczkowski et al41	Histology	10	20 to 68	PP									515 × 450
Olver et al26	Histology	1	NA	PP			16 to 20		9.8a	2.9a	127a	59a	200 to 300
											256b	158b
Yoneya et al43	Histology	9	NA	PP					14.3a	4.9a	234a	166a	600 to 800
Zouache et al44	Histology	6	NA	PP					17.6a	7.4a	401a	391a
											513b	380b

^aValues derived from reanalysis of published histology images by Jung et al.¹⁶

^bValues derived from reanalysis of published histology images (current).

Our lobule size measurements were similar to the studies by Fryczkowski⁴¹ and co-authors,⁴² who made careful observations of lobules from postmortem analysis. Those histological studies also reported that lobules were absent from the central macula, whereas our measurements (Fig. 3F and Supplementary Fig. S2) indicate a relatively uniform distribution across the macula. Finally, our measurements of RPE–CC distance (16.8 ± 1.2 µm at the fovea decreasing to ∼13–14 µm at 12.5° N and T) are slightly lower than those previously reported using AO, possibly owing to differences in location measured or subject variability. Our RPE–CC distance measurements were significantly lower than OCTA measurements, possibly owing to differences in axial resolution, depth of focus, and segmentation known to exist between AO-OCT and OCT.⁴⁵

The CC is profoundly impacted in a wide variety of retinal diseases, particularly AMD and IRDs that affect the outer retina. The CC is also linked to the blood supply of the optic nerve head and CC dropout and dysfunction are implicated in glaucoma as well.⁴⁶ In early CRD, the CC exhibits specific regional variation with the vessels in the central macular appearing thinner and the voids larger compared with the peripheral macula (Fig. 6). The normal lobule organization of the CC is also clearly disrupted in the central macula in CRD. The CC appears to be less affected by early USH2A and there is a notable lack of correspondence between vascular changes and the hyperreflective lesions observed at approximately 8 to 10° T (Fig. 7). There are also areas in the central macula where the CC segments appear to be blurred or less distinct, possibly owing to disease-related scattering changes in the overlying PR layer. The weaker signal may also be related to suboptimal AO correction across the disease transition zone. It should be noted that the results presented in this paper are relatively early disease stages. In the later stages of many retinal diseases, for example, late-onset retinal degeneration, we have observed larger patches of vascular dysfunction (e.g., flow drop-out) that encompasses nearly the entire CC layer.⁴⁷ It is intriguing that, in the eyes with small drusen, we observed a clear deficit in CC density and a concomitant increase in void diameter in the regions associated with drusen compared with nearby areas that appeared to be normal, whereas the overall CC diameter was not different in drusen affected or unaffected areas but higher than in the healthy volunteers without drusen (Fig. 5). Determining whether the areas without observable drusen have CC alterations that may be indicative of early AMD pathology will require more investigation.

One important consideration in applying our methodology to diseased eyes is the potential need for more accurate segmentation in the presence of CC layer thinning and pathology (e.g., cysts, disruptions, neovascular growth, and accumulated drusen deposits) in the overlying PR–RPE complex. This is analogous to the challenge retinal degenerative diseases posed to retinal layer segmentation algorithms previously developed.⁴⁸ Several groups have developed CC segmentation algorithms to aid quantitation,^49–51 having some success automatically identifying Bruch's membrane in the presence of neovascular growth and other abnormalities, but less success automatically identifying the lower boundary of the CC.⁵² CC thinning may cause encroachment of Sattler's layer into the axial region occupied by the CC leading to inaccurate metrics, which could be mitigated by a more sophisticated segmentation algorithm. Further AO imaging in more disease types will help to confirm the potential of our CC biomarkers and provide higher-resolution data for CC segmentation algorithm development.

Large vessel shadows and other signal artifacts also necessitate careful manual ROI selection for accurate quantification, which may introduce some bias. Although the horizontal strip mitigates this factor by avoiding large vessels, our regions were not completely free of overlying vessel artifacts. Our robust algorithm partially handles this with vessel diameter filtering and manual grader checks. The source of some of the artifact is OCTA decorrelation tails, and algorithms that have mitigated their appearance in clinical OCTA images^53,54 may help to produce more accurate results in AO-OCTA volumes as well.

Some limitations of the current study include its relatively small size of 20 participants (with quantified metrics), although the trends we described are clear even in our smaller cohort. We have also not extracted any functional measurements of the CC network from the AO-OCTA volumes. Some recently demonstrated split detection AO SLO,^55,56 OCTA,⁵⁷ or Doppler imaging⁵⁸ techniques may be better suited for quantifying absolute or relative flow rates, assuming difficulties accessing the CC layer with those modes are overcome. In terms of an imaging protocol suitable for clinical use, a strip across the macula may reveal less information than one that equally covers the superior–inferior range. In anticipation of this, we have configured our imaging system to rapidly collect a mosaic of 5 × 5 AO-OCTA volumes covering a macular region approximately 13° × 13° (3.8 mm × 3.8 mm) in as little as 8 minutes (example shown in Supplementary Fig. S5), which is comparable with current clinical OCTA scans and well-tolerated by most patients. Further improvements to streamline the protocol (auto-mosaicking) and increase the data saving speed will further enhance the potential for clinical translation.

This study represents an immense collection of normative data on CC structure that can be used to study vascular dysfunction that accompanies many ocular diseases. High-speed AO-OCTA conveys several key advantages, primarily improved lateral resolution that allows visualization of capillary segments in the dense CC, and a scanning speed that allows large spans of the CC to be rapidly mapped. We further expect our AO measures to be robust with respect to various vessel segmentation algorithms because the underlying data are generated with these characteristics (high lateral resolution and speed) as well as clear signal discrimination from the CC layer. Overall, our study demonstrates that AO is an extremely useful tool for quantification of the CC capillary network that can provide insight into ocular disease pathology.