Noninvasive Imaging of Human Foveal Capillary Network Using Dual-Conjugate Adaptive Optics

Abstract

Purpose.: To demonstrate noninvasive imaging of human foveal capillary networks with a high-resolution, wide-field, dual-conjugate adaptive optics (DCAO) imaging instrument.

Methods.: The foveal capillary networks of five healthy subjects with no previous history of ocular or neurologic disease or surgery were imaged with a novel high-resolution, wide-field DCAO instrument. The foveal avascular zone (FAZ) in each image was defined using a manual procedure. An automated algorithm based on publicly available and custom-written software was used to identify vessels and extract morphologic FAZ and vessel parameters. Capillary densities were calculated in two annular regions of interest (ROIs) outside the FAZ (500 μm and 750 μm outer radius from the foveal center) and in the superior, inferior, temporal, and nasal quadrants within the two ROIs.

Results.: Mean FAZ area was 0.302 ± 0.100 mm², and mean capillary density (length/area) in the inner ROI was 38.0 ± 4.0 mm⁻¹ and 36.4 ± 4.0 mm⁻¹ in the outer ROI. The difference in ROI capillary density was not significant. There was no significant difference in quadrant capillary density within the two ROIs or between quadrants irrespective of ROI.

Conclusions.: The authors have demonstrated a technique for noninvasive imaging and semiautomated detection and analysis of foveal capillaries. In comparison with other studies, their method yielded lower capillary densities than histology but similar results to the current clinical gold standard, fluorescein angiography. The increased field of view of the DCAO instrument opens up new possibilities for high-resolution noninvasive clinical imaging of foveal capillaries.

The central region of the human retina, the macula, differs from the rest of the retina. ¹ The macula is subdivided into three regions—the fovea, the parafovea, and the perifovea—originally defined by Polyak. ² At the very center of the fovea is the foveola, approximately 350 μm in diameter, ³ which has an absence of retinal vasculature and where only photoreceptors and glial cells are present. At the center of the fovea, and including the foveola, is a capillary-free region known as the foveal avascular zone (FAZ). The FAZ is approximately 500 μm in diameter, but there is considerable interindividual variation. ^4,5  

Retinal arterioles that originate from the superior and inferior temporal arterial branches converge toward the FAZ, where they side-branch into vessels of the terminal capillary ring that constitute the FAZ border, bend down and diverge from the FAZ in a deeper layer, and collect into venules interspersed between converging arterioles. Anatomic studies of the macular capillary network have revealed multiple capillary planes, resulting in high capillary densities. ^3,6  

Imaging of retinal capillaries is a difficult task because of their small size (down to approximately 5 μm), low contrast, and arrangement in multiple planes of varying retinal depth. Even good-quality color retinal imaging fails to capture any of the finest capillary details. In fluorescein angiography (FA), the current clinical gold standard for capillary imaging, a contrast agent is injected in the patient's bloodstream to enhance the contrast of the retinal vasculature. This is not performed on healthy eyes because of a small risk for adverse reactions. ^7,8 However, many capillaries are not visualized in FA. An apparent discrepancy was shown between histologically determined capillary densities and the numbers obtained with FA in a comparative histologic and angiographic study of the same capillary networks from retinal whole mounts of macaque retinas. ⁹ Smaller capillaries in the foveal slope and in deeper planes were often missed. This finding has been confirmed by a recent study comparing capillary visualization using high-resolution FA imaging and confocal microscopy in humans. ¹⁰  

The introduction of adaptive optics (AO) in vision research has opened up new possibilities for retinal imaging, ¹¹ allowing high-contrast imaging of retinal features on the order of microns. The main focus of AO in vision research has been imaging of photoreceptors. ^12,13 However, recent publications report on retinal nerve fiber layer (RNFL) capillary imaging in the macaque retina using fluorescein adaptive optics scanning laser ophthalmoscopy (FAO-SLO) ¹⁴ and noninvasive visualization of the human retinal capillary network by mapping leukocyte flow with AO-SLO. ¹⁵  

In this article we present the results of noninvasive imaging of retinal capillaries using a modified version of a recently introduced flood-illumination, dual-conjugate adaptive optics (DCAO) instrument ¹⁶ and compare our findings with those of previous works on capillary visualization. 

Subjects and Methods

The research followed the tenets of the Declaration of Helsinki. All subjects were informed about the goals, consequences, and protocol of the study and then provided their written informed consent. The study protocol was approved by the ethical review board of the University of Gothenburg. 

Subjects

Five healthy subjects, ranging in age from 30 to 58 years, with clear optical media and no previous history of ocular or neurologic disease or surgery participated in the study. The axial length (AL) of the investigated eye of participating subjects was measured (IOL Master; Carl Zeiss Meditec AG, Jena, Germany). Mean AL for the five healthy subjects was 24.6 ± 0.92 (mean ± SD). Individual retinal scaling factors (RSF; Table 1) were calculated using the formula RSF = IPS × q × P. The parameter IPS is the instrument plate scale (the angle corresponding to a linear extent at the focal plane, inversely proportional to the instrument f-number), q = 0.01306 × (AL − 1.82) is the individual retinal scaling factor, ¹⁷ and P is the pixel size of the retinal imaging camera. Including additional ocular parameters such as corneal curvature and anterior chamber depth has been shown to yield only minor improvement in retinal scaling factor estimation. ¹⁷  

Table 1.

 

Subject Data and Calculation of Individual RSF

Subject	Age (y)	Eye	IPS (°/mm)	P (μm/pixel)	AL (mm)	q (mm/°)	RSF (μm/pixel)
TR	36	R			23.59	0.284	0.904
JT	44	L			25.82	0.313	0.997
PK	32	R	0.43	7.4	24.62	0.298	0.947
HK	30	L			24.37	0.295	0.937
BL	58	L			24.80	0.300	0.954

 RSF = IPS × q × P.

Adaptive Optics Correction

The original DCAO setup is described in detail elsewhere. ¹⁶ A diagram of the modified setup is shown in Figure 1. 

Figure 1.

 

Diagram of the DCAO setup. SLD, super-luminescent diode; DMI, ALPAO 97 actuator DM; DM2, OKO 37 actuator DM; CCD, wavefront sensor camera; CLA, collimating lens array; FS, square field stop; P, pupil conjugate plane; R, retinal conjugate plane; P′, split pupil plane; R′, collapsed retinal image plane of the five retinal beacons. The main components of the AO system are the 835 ± 14 nm SLD, the two deformable mirrors, and the multi-reference wavefront sensor. The main components of the imaging system are the 575 ± 10 nm xenon flash and the retinal camera.

Continuous relatively broadband near-infrared light (834 ± 13 nm) for five retinal beacons is delivered by a super-luminescent diode (Superlum Ltd., Moscow, Russia). The retinal beacons are arranged in a cross on the retina, with a 3.1° separation of the four peripheral beacons from the central beacon. The pupil conjugate deformable mirror (DM) of the modified instrument was a 13.5-mm magnetic DM (ALPAO, Biviers, France) with 97 actuators in a square grid. The second DM, conjugate to a plane just in front of the retina, was a 30-mm piezoelectric DM (OKO Technologies, Delft, The Netherlands) with 37 actuators arranged in a hexagonal grid. Closed-loop correction frequency was approximately 10 Hz. The wavefront sensor allows spatial filtering of the light from the five retinal beacons using only one adjustable iris, and the five Hartmann patterns generated by the retinal beacons are imaged on a single wavefront sensor CCD camera. 

Adaptive Optics Imaging

Retinal illumination with an effective energy of 120 μJ at the corneal plane is supplied by a 1-ms flash from a Xenon flash lamp (Retinapan 45-II; Nikon, Tokyo, Japan), filtered by a 575 ± 10 nm bandpass filter (FF01–575/15–25; Semrock, Rochester, NY). Hemoglobin has a local absorption peak in the 550- to 600-nm wavelength range, ¹⁸ and the imaging wavelength of our instrument was, therefore, well suited for imaging blood vessels. The flash pupil diameter entering the eye measured approximately 3 mm, and a square field stop limited the illuminated field to approximately 7° × 7° on the retina. A central block placed in a corneal conjugate plane within the flash relay optics was used to reduce corneal reflections. Light reflected from the retina was diverted by a cold mirror and relayed through a pair of photographic objectives (50 mm f/1.8D AF Nikkor; Nikon Imaging Japan Inc., Tokyo, Japan). A field stop, corresponding to a diameter of 6 mm at the eye, was positioned in a pupil conjugate plane between the two objectives. Retinal images, with a diffraction limited resolution of approximately 2 μm, were captured with a 2048 × 2048-pixel, 16-bit monochromatic CCD camera (Pike F-421B; Allied Vision Technologies GmbH, Stadtroda, Germany). 

Image Postprocessing

Camera raw images were postprocessed and analyzed using publicly available and custom-written software (MatLab; MathWorks, Natick, MA). Image magnification was calibrated with respect to individual parameters of AL. 

Uneven flash illumination in raw images was reduced using a four-step procedure (Figs. 2a–e). First, the raw image (Fig. 2a) was spatially filtered to remove high-frequency noise (Fig. 2b). Second, image contrast was enhanced using contrast-limited adaptive histogram equalization (CLAHE) with the built-in MatLab function adapthisteq (Fig. 2c). Third, the CLAHE-filtered images were flat-fielded using a low-pass filtered image (CLAHE-filtered images convolved with a σ = 20 pixels Gaussian kernel) to reduce uneven illumination (Fig. 2d). ¹⁹ The images were finally convolved with a small Gaussian kernel (σ = 1) to smooth the final image but retain local structure (Fig. 2e). 

Figure 2.

 

Montage of image postprocessing steps. (a) Raw image. (b) Spatially filtered image. (c) CLAHE-filtered image. (d) Flat-fielded image. (e) Smoothed image (Gaussian filter [σ = 1]). (f) Spline curve fit to manually traced FAZ border. (g) Output image from vesselness filter. (h) Skeletonized tracing of identified vessels.

Vessel Detection

An outline of the FAZ perimeter (Fig. 2f) was performed by manually selecting points along the FAZ border. A spline curve was fit to the set of perimeter points using a spline function, and the region delimited by the spline curve was then shrunk by five pixels along the entire circumference to ensure that perimeter vessels were not excluded from analysis. The resultant area was used as a mask of the FAZ. The FAZ center of mass was used as the definition of the foveal center. 

Retinal capillaries and larger vessels were automatically detected using a publicly available MatLab implementation of a Hessian-based vesselness filter. ²⁰ The Frangi vesselness filter identifies tubular geometric structures over a specified scale range, chosen as σ = 4 to 12 in steps of 2 in our implementation. The image of identified vessels from the Frangi filter (Fig. 2g) was contrast stretched and converted to a binary image. Finally, the binary image was skeletonized, a procedure that removes pixels on the boundaries of identified vessels but does not allow connected vessel segments to break apart. The end result was a 1-pixel-wide tracing of identified vessels (Fig. 2h) that was used to determine vessel length and calculate vessel densities. 

Morphologic Analysis

Morphologic FAZ and capillary parameters were analyzed within specified regions of interest (ROIs) and ROI quadrants (Fig. 3) using custom-written code in MatLab. Capillary length and density were analyzed in the nasal, inferior, temporal, and superior quadrants of two annular ROIs outside the FAZ. The inner ROI (ROI₅₀₀) was defined from the edge of the FAZ to a radius of 500 μm from the foveal center. The outer ROI (ROI₇₅₀) was defined from the edge of ROI₅₀₀ to a radius corresponding to the histologic definition of the foveal radius of 750 μm from the foveal center, ³ excluding pixels within the FAZ if the FAZ extended outside the inner ROI. The total running length, L, in millimeters of identified capillary segments within an ROI was obtained by summing all pixels in the binary skeletonized image. Vessel density was defined as L/A, where A is the area of the ROI in square millimeters. All measures were adjusted for individual retinal scaling. A data and predictive analytics tool (IBM SPSS Statistics 18.0.2 for Mac; SPSS Inc., Chicago, IL) was used for all statistical analyses (see Results). Differences in mean capillary densities were investigated using one-way ANOVA. Analysis of confounding factors such as age or axial length was not performed because of the small sample size. Post hoc analysis using the Tukey HSD and Games-Howell tests was performed to verify the normal distribution of data. Findings with error P < 0.05 were considered statistically significant. 

Figure 3.

 

Schematic drawing of the two ROIs and corresponding quadrants (N, nasal; I, inferior; T, temporal; S, superior) used in capillary density analysis in a right eye. The quadrants were horizontally mirrored to maintain nasotemporal classification for left eyes.

Results

Camera raw and post-processed image grayscale distributions for all subjects are shown in Figure 4. The postprocessing algorithm reduced the variability in image illumination and processed images from all eyes without algorithm failure, yielding a mean histogram value of postprocessed images of 152.7 ± 1.0 (mean ± SD). A montage of unprocessed, processed, and identified vessel images of the five healthy subjects is shown in Figure 5. Images of skeletonized vessels, including FAZ (central black area), are shown in Figure 6. Typical processing times (on a 2.8-GHz Intel [Santa Clara, CA] Core i7 iMac [Apple, Cupertino, CA] running 64-bit MatLab R2010a) are 2 to 3 seconds for image postprocessing, approximately 2 minutes for the manual FAZ contour selection, and 17 seconds for vessel detection, skeletonization, and analysis. 

Figure 4.

 

Grayscale distribution of camera raw and postprocessed images. The mean gray scale value of postprocessed images was 152.7 ± 1.0 (mean ± SD).

Figure 5.

 

Unprocessed images (left), processed images (middle), and image with identified capillaries (right). Images are scaled according to individual retinal scaling factor. Image labels indicate subject, age, and eye.

Figure 6.

 

Montage showing images of skeletonized identified capillaries and FAZ (black central area) of subjects TR, JT, PK, HK, and BL (top to bottom). Images are scaled according to individual retinal scaling factor. Image labels indicate subject, age, and eye.

Extracted FAZ parameters, including mean ± SD, for the five healthy subjects are listed in Table 2. FAZ area was 0.302 ± 0.100 mm², and FAZ perimeter length was 3369 ± 769 μm. FAZ equivalent diameter (the diameter of a circle with the same area as the FAZ) was 612 ± 106 μm. 

Table 2.

 

Descriptive FAZ Parameters Calculated from Manual Tracing of the TCR

	TR	JT	PK	HK	BL	Mean	SD
Area, mm2	0.406	0.200	0.354	0.189	0.359	0.302	0.100
Perimeter length, μm	3898	3030	3008	2498	4409	3369	769
Major axis length, μm	830	580	692	602	816	704	116
Minor axis length, μm	651	488	660	413	604	563	108
Equivalent diameter, μm*	719	505	671	491	676	612	106

*  The diameter of a circle with the same area as the FAZ.

Mean capillary density was 38.0 ± 4.0 mm⁻¹ in ROI₅₀₀ and 36.4 ± 4.0 mm⁻¹ in ROI₇₅₀, but this difference was not significant (F(1,38) = 1.238; P = 0.273) (Table 3). Mean quadrant densities in ROI₅₀₀ ranged from 35.5 ± 3.2 mm⁻¹ in the superior quadrant to 40.0 ± 7.2 mm⁻¹ in the inferior quadrant and from 34.9 ± 4.1 mm⁻¹ in the nasal quadrant to 38.6 ± 6.3 mm⁻¹ in the inferior quadrant of ROI₇₅₀ (Table 3). The highest densities were found in the inferior quadrants of both ROIs. A box plot of quadrant capillary density distributions, including outliers, is shown in Figure 7. 

Table 3.

 

Capillary Densities in ROI and ROI Quadrants

Density (mm−1)	TR	JT	PK	HK	BL	Mean	SD
ROI500	41.4	32.2	42.2	36.5	37.8	38.0	4.0
ROI750	38.9	29.7	39.5	38.1	35.6	36.4	4.0
Q500 nas	38.8	33.0	41.6	40.0	39.1	38.5	3.3
Q500 inf	45.0	29.6	48.2	39.0	38.2	40.0	7.1
Q500 temp	45.6	31.8	40.6	35.9	36.2	38.0	5.3
Q500 sup	36.3	34.5	38.6	30.5	37.4	35.5	3.2
Q750 nas	36.3	28.8	40.0	35.5	34.1	35.0	4.1
Q750 inf	43.3	28.5	42.8	42.1	36.2	38.6	6.3
Q750 temp	41.0	30.4	38.4	38.1	36.9	37.0	4.0
Q750 sup	35.2	31.2	36.6	36.8	35.3	35.0	2.3

 Subscript numbers indicate ROI outer radius. Q, quadrant; nas, nasal; inf, inferior; temp, temporal; sup, superior.

Figure 7.

 

Box plot of capillary densities in the nasal, inferior, temporal, and superior ROI₅₀₀ and ROI₇₅₀ quadrants. The bottom and top of the box are the lower and upper quartiles, respectively. The horizontal line in the box is the median. Whiskers represent the lowest and highest values that are neither suspect (*) nor very suspect (○) outliers.

One-way ANOVA of quadrant densities was performed under the assumption that the variances between groups were equal. There was no significant difference in capillary density between inner ROI quadrants (F(3,16) = 0.718; P = 0.556) or outer ROI quadrants (F(3,16) = 0.785; P = 0.519) or in the analysis including all quadrants (F(7,32) = 0.810; P = 0.585). Post hoc analysis of the data using the Tukey HSD and Games-Howell statistics confirmed the initial assumption of equal variance. Analysis of quadrant capillary density obtained by pooling data from ROI₅₀₀ and ROI₇₅₀ (Fig. 8) did not show any statistically significant differences between quadrants (F(3,16) = 1.286; P = 0.313). 

Figure 8.

 

Box plot of capillary densities in the combined nasal, inferior, temporal, and superior quadrants of ROI₅₀₀ and ROI₇₅₀. The bottom and top of the box are the lower and upper quartiles, respectively. The horizontal line in the box is the median. Whiskers represent the lowest and highest values that are neither suspect (*) nor very suspect (○) outliers.

Discussion

Mean FAZ area of 0.302 ± 0.100 mm² and mean equivalent diameter of 612 ± 106 μm obtained with our method were similar to results from previous studies. ^5,15,21,22 Capillary density data were similar to the data of Tam et al., ¹⁵ who report on capillary densities (length/area) extracted from images obtained with AO-SLO in an ROI surrounding the FAZ of approximately the same size as our combined ROI₅₀₀ and ROI₇₅₀. Average capillary densities of Tam et al. ¹⁵ were 30.3, 31.5, 30.7, and 34.0 mm⁻¹ in the nasal, inferior, temporal, and superior quadrants. The combined ROI₅₀₀ and ROI₇₅₀ mean densities in the present study of 36.2, 38.9, 37.2, and 35.1 mm⁻¹ are approximately 16% higher. 

In a study on macaque monkeys, Weinhaus et al. ⁹ presented capillary densities (length/area) obtained from both FA imaging and histologic tissue preparations. They show that visibility of capillaries in FA images depends strongly on capillary diameter and retinal depth and that large-diameter capillaries are better visualized than small-diameter capillaries. Nearly all capillaries are visualized near the border of the FAZ, but the percentage of visualized capillaries falls to <50% by 600 μm eccentricity and <40% by 900 μm eccentricity compared with histologically determined densities. More than 50% of the superficial capillaries, but< 20% of the deep capillaries, were visualized with FA out to an eccentricity of 1200 μm. However, the mean densities reported by Weinhaus et al. ⁹ (0.3–0.9 mm⁻¹) are >40 times lower than our findings. 

Mendis et al. ¹⁰ recently published a comparative high-resolution FA imaging and confocal microscopy capillary visualization study on young, healthy human subjects. The single-layered region of capillaries around the FAZ could be clearly identified in all subjects. The authors analyzed capillary densities in the multilayered capillary network of two parafoveal ROIs. The retinal areas occupied by the capillary network in the two ROIs were approximately 41% for superficial capillaries and 23% for deep capillaries. Even though subjects were young and healthy to optimize FA image quality, high-resolution FA densities were <61% of the histologic densities. The authors suggest that interpretation of capillary detail using FA imaging is significantly limited despite recent technological advances. 

Total processing time for the automated image after processing and vessel identification and analysis was typically <20 seconds. However, the manual selection of terminal capillary ring vessels that define the FAZ, a procedure that takes approximately 2 minutes, introduced an element of subjectivity to the procedure. This was a deliberate compromise, and the development of an automated FAZ detection algorithm will be part of future work. An automated algorithm will also considerably shorten the overall processing time, and a further reduction is likely to be accomplished through an implementation of the algorithm in an optimized C++ program. 

The main limitation of the proposed method is that only subjects with clear ocular media yield good quality images. This is, however, generally not an issue in healthy subjects or in subjects in preclinical stages of retinal disease who are usually excluded from FA imaging. Our noninvasive method can potentially provide an alternative to FA in these cases. 

An intraindividual comparison to FA is lacking. Given that the topic of this work is normal foveal vasculature and given that FA is not performed on healthy eyes, we have failed to obtain comparative material from patients with retinal disease who have undergone FA imaging and have an unaffected central macular region. This topic will be addressed in future work on clinical imaging. 

Another issue is the need to obtain better images of the capillary network by increasing image contrast and suppressing image noise. All images used in the present analysis are single exposures, and, though they are of sufficient quality for analysis, an averaging of multiple images in addition to the current postprocessing algorithm can increase the signal-to-noise ratio even further. However, averaging multiple images will not affect the performance of the flat-fielding technique used to reduce image illumination variability. 

There is a need to further explore the potential to visualize the deeper capillary network in the parafoveal retina, which high-resolution FA fails to do. ¹⁰ Although the reported method is subject to image degradation from light scatter, it is not limited by the presence of choroidal fluorescence as in FA imaging. ¹⁰ Possible advantages or disadvantages compared with present methodologies must be investigated in future work on clinical imaging, but the results of the present study indicate its potential as a noninvasive complement to high-resolution FA imaging. 

In conclusion, the high resolution and increase in corrected FOV offered by a DCAO instrument opens up new possibilities for AO retinal imaging. The results of the present study show that analysis of foveal capillary images obtained with our noninvasive method provides capillary densities comparable to those obtained with AO-SLO and FA imaging. High-resolution AO imaging of the foveal capillary network is of particular importance for diseases affecting the central retina, such as diabetes, and will potentially allow for improved diagnosis through the visualization of early signs of disease. 

Footnotes

 Supported by the Marcus and Amalia Wallenberg Memorial Fund, the De Blindas Vänner i Göteborg Foundation, the KMA Foundation, and the Edwin Jordan Foundation for Ophthalmic Research.

Footnotes

 Disclosure: Z. Popovic, P; P. Knutsson, None; J. Thaung, P; M. Owner-Petersen, P; J. Sjöstrand, None