Three-Dimensional Light Field Fundus Imaging: Automatic Determination of Diagnostically Relevant Optic Nerve Head Parameters

Abstract

Purpose: Morphological changes to the optic nerve head (ONH) can be detected at the early stages of glaucoma. Three-dimensional imaging and analysis may aid in the diagnosis. Light field (LF) fundus cameras can generate three-dimensional (3D) images of optic disc topography from a single shot and are less susceptible to motion artifacts. Here, we introduce a processing method to determine diagnostically relevant ONH parameters automatically and present the results of a subject study performed to validate this method.

Methods: The ONHs of 17 healthy subjects were examined and images were acquired with both an LF fundus camera and by optical coherence tomography (OCT). The LF data were analyzed with a novel algorithm and compared with the results of the OCT study. Depth information was reconstructed, and a model with radial basis functions was used for processing of the 3D point cloud and to provide a finite surface. The peripapillary rising and falling edges were evaluated to determine optic disc and cup contours and finally calculate the parameters.

Results: Nine of the 17 subjects exhibited prominent optic cups. The contours and ONH parameters determined by an analysis of LF 3D imaging largely agreed with the data obtained from OCT. The median disc areas, cup areas, and cup depths differed by 0.17 mm², −0.04 mm², and −0.07 mm, respectively.

Conclusions: The findings presented here suggest the possibility of using LF data to evaluate the ONH.

Translational Relevance: LF data can be used to determine geometric parameters of the ONH and thus may be suitable for future use in glaucoma diagnostics.

Introduction

Glaucoma is a group of diseases with a worldwide impact as a leading cause of irreversible vision loss.¹ Seventy-six million cases of glaucoma were reported in 2020; this number is anticipated to increase to 111.8 million individuals in 2040.² Currently, Europeans represent approximately 10% of those affected by glaucoma.^2,3 Untreated glaucoma can lead to irreversible blindness. Currently, glaucoma is the second leading cause of blindness, with 11.2 million people suffering from bilateral disease.³ Early diagnosis of this condition and critical follow-up are important for the treatment of this disease and the preservation of vision. 

Open-angle glaucoma is the most commonly diagnosed form of this disease. Risk factors for this condition include an increased cup–disc ratio (CDR), CDR asymmetry (comparing the right and left eye), disc hemorrhage, and elevated intraocular pressure (i.e., high-tension glaucoma).⁴ Additional risk factors are older age, family history, and ethnic background. The mechanical stress and strain resulting from an increase in the intraocular pressure push the optic nerve head (ONH) outward and damage the retinal nerve fiber layer (RNFL).¹ These morphological changes begin during the early stages of glaucoma.^5,6 As the disease progresses, affected individuals experience a gradual loss of their visual field that can progress to complete blindness.⁶ Because this gradual progression allows central visual acuity to be preserved until the late stages of the disease, glaucoma is generally diagnosed at a late stage.⁷ Therefore, methods that might be used for early diagnosis are important to support ongoing efforts to preserve vision—in particular, the visual field. 

Methods used to diagnose glaucoma include imaging techniques used to detect morphological changes and functional processes (i.e., visual field testing).^8,9 Established methods used to examine the morphological properties of the ONH include imaging with a fundus camera, optical coherence tomography (OCT), and scanning laser tomography.¹⁰ Diagnostically relevant parameters determined by these methods include the size of the optic disc and the optic cup, as well as the CDR. Another important parameter is the cup depth.^11,12 

Imaging with a fundus camera is a widely used technique that generates two-dimensional (2D) images of the fundus surface.^11,13,14 This is a cost-effective method that can be performed rapidly and requires only a single shot.^13,14 However, one limitation of fundus cameras is that they generate only 2D recordings of the surface rather than a more useful three-dimensional (3D) representation.^13–15 

As a further development, there is the stereoscopic fundus camera, which also allows 3D imaging. There are two different types: the sequential and the simultaneous stereo fundus camera.¹⁶ In the sequential type, two successive images are taken at a slightly different angle. Most modern mydriatic fundus cameras are capable of producing those sequential images by shifting the camera slightly to the left and to the right from the central position. The disadvantages of this method are the undefined and non-reproducible stereo basis (separation between the center of the lenses) because of the manual shift. Consequently, the three-dimensional effect may be inconsistent between different image pairs. Furthermore, the stereoscopic accuracy decreases in the periphery because of the smaller stereoscopic angle.^17,18 In the simultaneous type, the two images are taken in one shot with a defined stereo basis; however, decreasing accuracy in the periphery also results here. In contrast, with the microlens array there is a constant measurement uncertainty over the sensor. 

Among the alternatives, the Heidelberg Retina Tomograph (HRT; Heidelberg Engineering, Heidelberg, Germany) utilizes scanning laser technology. The region to be examined is subjected to point-by-point scanning with a weak laser.¹⁹ The confocal scanning algorithm generates a layered 3D image that allows the surface structure of the papilla to be mapped and visualized. However, because of the scan paradigm, this method takes more time to complete, motion artifacts can occur, and it is more expensive and unwieldy than fundus photography.²⁰ Moreover, the operator needs to perform manual image processing and be capable of drawing a precise contour line to define the optic disc margin. Registration errors can corrupt the data.⁹ 

OCT is an imaging technique used to visualize the papilla and surrounding tissue that generates cross-sectional images of the region under examination^21–23 and thus facilitates 3D visualization of the optic disc, optic cup, and retinal layers.²³ It is also possible to use this method to measure the thickness of the RNFL. Point-by-point scanning is performed and data are generated by scanning algorithms. As described for the HRT, OCT is time consuming, and motion artifacts can occur. Additionally, OCT systems are the most expensive of all of the methods previously discussed.²⁰ 

In an effort to advance the state of the art in glaucoma diagnostics with none of the aforementioned limitations, a new method was developed that generates 3D images of the fundus with only one photographic shot via the use of a fundus camera in combination with light field (LF) technology.^24–27 Using this method, the entire LF is recorded with a microlens array that facilitates the reconstruction of depth information and provides a representation of the surface of the papilla.²⁸ A subsequent 3D analysis is also possible. A technical setup that could be used for papilla imaging was presented by Schramm et al.²⁴ In this publication, the authors documented a method for LF imaging of the ONH and showed that relevant depths could be quantified in units of virtual mm (v-mm). However, calibration is needed to compare these results with data from other imaging systems. Similarly, depth information can only be estimated from structures with a defined minimal degree of contrast, including the regions containing the blood vessels as well as the border and structures within the papilla.^24,25 As a result, the data are unevenly distributed. Additional artifacts are frequently overlaid on the data, including mis-estimated depths and frequencies.²⁴ 

Until now, no method is known to diagnose glaucoma based on information that could be obtained from an LF fundus camera. Here, we present an algorithm that uses LF data to calculate geometric parameters of the ONH based on surface features. To verify the accuracy and practical value of this algorithm, we performed a subject study in which we compared calculated papilla parameters generated by LF with data obtained from OCT. Methods used to convert units of v-mm to those used to generate OCT data are also presented. 

Methods

Subjects

A study to evaluate ONH parameters based on LF fundus camera data and to compare the results to findings obtained by OCT was performed on 17 young healthy adults (10 males, 7 female; average age, 29.5 ± 7.1 years; range, 24–46 years). Included were healthy people over 18 years. Exclusion criteria established for this study focused on the need to minimize the influence of light scatter and to obtain images of good quality. Thus, subjects with cataracts, nystagmus, and increased lid-closure reflex were excluded from the study. Further exclusion criteria were myopia less than −5 diopters (D), hyperopia more than +5 D, and astigmatism more than 3 D. In addition, the following were excluded: tropicamide intolerance, pregnancy, or breastfeeding. 

All subjects underwent a preliminary ophthalmologic examination that included determination of visual acuity, objective refraction, tonometry, and slit-lamp microscopy to ensure the absence of any previous diseases and any of the aforementioned exclusion criteria. The preliminary examination required approximately 10 minutes and was performed several days before participation in the study. All 17 subjects met the inclusion criteria and were deemed eligible to participate in the study. The eye with the lower cylindrical refractive error was selected for evaluation. We evaluated six left eyes and 11 right eyes from the 17 enrolled participants. 

Approval for the study was obtained from the ethics committee of the Friedrich Schiller University of Jena (Jena, Germany; processing number 4307-01/15). All procedures were carried out in accordance with the tenets of the Declaration of Helsinki. All the subjects were informed about the procedure before the preliminary ophthalmologic examination and provided written informed consent. 

Study Design, Procedure, and Technical Setup

Subjects were divided into two groups. Data from 14 subjects were used to develop the algorithm. Data from three additional subjects were used as independent test data. The left or right eyes selected for examination as described above were inoculated with two drops of Mydrum ophthalmic solution (Mydriaticum Stulln; Pharma Stulln GmbH, Stulln, Germany). Pupil dilation was confirmed after a 15-minute waiting period, and photographs were taken. To avoid technical and physiological confounding of both recording techniques, participants for algorithm development were randomly divided into two groups. Seven participants were randomly selected to be examined first with OCT (SPECTRALIS; Heidelberg Engineering, Heidelberg, Germany) followed by evaluation using the LF fundus camera as per the setup described by Schramm et al.²⁴ In the other seven subjects, the order of testing was reversed. After completion of both tests, the data were analyzed and the papilla parameters were determined. The procedure used in this study is outlined in Figure 1. 

Figure 1.

 

Flowchart of the study design. Subjects were divided into two groups: development data (n = 14) and test data (n = 3). After selection and preliminary examination, the subjects underwent mydriasis. Participants were randomly assigned to one of two groups for image acquisition. In one group, the fundus was imaged first with OCT and then the LF fundus camera; in the second group, the order of image acquisition was reversed. Data were then exported and analyzed.

Optical Coherence Tomography

The OCT examination was performed using the Heidelberg SPECTRALIS with the glaucoma module (Glaucoma Module Premium Edition, Software Version 6.0) and the ONH scan pattern. The imaging process included 24 radial scans that were centered above the optic disc and three circular scans with distinct radii. Centering was performed automatically with guidance from the anatomic positioning system. OCT imaging required approximately 5 minutes. 

LF Fundus Camera

The technical setup of the LF fundus camera included a mydriatic fundus camera (FF450; Carl Zeiss Meditec, AG, Jena, Germany) and an LF imager (R12; Raytrix GmbH, Kiel, Germany) connected via a C-mount adapter as originally described by Schramm et al.²⁴ The field of view of the fundus camera was set on 30° to provide an optimal match with the optical aperture of the fundus camera and LF imager. The study was performed with green light illumination at a wavelength of 520 ± 20 nm. The software associated with the LF sensor (RxLive 5.0.046.0; Raytrix GmbH) was used for image acquisition, depth reconstruction, and exporting data. Fifteen fundus images were taken from each participant in the subject group for algorithm development. The camera and illumination intensity were readjusted after five images for each subject based on individual glare sensitivities. The shutter time and gain of the LF camera were set near saturation to avoid over-exposure of the papilla region. The examination with the LF camera required approximately 30 minutes. Data export was performed with the associated LF software, and papilla parameters were determined by the software-based algorithm. For the independent test subjects, one fundus image was taken. The setup of this part of the study is presented in Figure 2. 

Figure 2.

 

Setup of LF fundus camera, processing software, and algorithm for papilla parameter determination. The LF fundus camera includes a fundus camera with a bandpass filter (520 ± 20 nm) and an LF imager. A personal computer (PC) with LF software is required for data processing. Papilla parameters were calculated with the developed algorithm.

Data Analysis

Papilla Parameters

Diagnostically relevant parameters include the area of the optic disc and optic cup and the optic cup depth. The distance between the ends of Bruch's membrane (opening points indicated by crosses in Fig. 3b) defines the diameter and thus the boundary of the optic disc.^29,30 The optic cup was defined as the region between the inner limiting membrane (ILM) as indicated by the horizontal line connecting the ends of Bruch's membrane (dashed line in Fig. 3b).³⁰ The optic cup depth was then determined as the maximum distance between the bottom of the optic cup and the mean height of its border. Figure 3a shows the defined boundaries and the optic disc parameters based on OCT cross-sectional and top-view fundus images as an example. As noted above, the ratio between the areas is known as the CDR. 

Figure 3.

 

OCT images top (a) and cross-sectional views (b, c) of the papilla. The inner limiting membrane (ILM; yellow), optic disc border (blue), and optic cup border (red) are marked as indicated. In (b), the optic cup depth is marked with a double-headed arrow and is the distance defined by the dashed line connecting the ends of Bruch's membrane (at the crosses) and the deepest point within the optic cup. In (c), an OCT sectional view is shown with the overlay tool used to measure the papilla parameters. The long arrows in all panels mark the position and reading direction of the cross-sectional image.

Some optic disc excavations appear different compared with the example shown in Figure 3 (i.e., young eyes, ONH swelling, ONH drusen, and pseudopapilledema). Therefore, another qualitative parameter is introduced: the presence of an optic cup depth (optic cup extends below reference plane), referred to in the following as the prominent optic cup. Figure 4 shows a cross-sectional OCT image of a subject with a prominent optic cup (Fig. 4a) and a subject without a prominent optic cup (Fig. 4b) for comparison. 

Figure 4.

 

Subjects with (a) prominent optic cup with present cup depth (CD) below Bruch's membrane opening (BMO) and (b) without prominent optic cup without present cup depth. The presence of a prominent optic cup will be used as a qualitative evaluation parameter.

Optical Coherence Tomography

Analysis of OCT data was performed using the glaucoma module software. The boundaries of the optic disc were automatically determined for each radial scan based on detection of Bruch's membrane. The area of the optic disc was calculated automatically and presented in units of square millimeters (mm²). The area and depth of the optic cup were determined manually. The overlay tool of the glaucoma module was used to analyze the various radial scans and provided output in units of micrometers (µm) (Fig. 3c). Two of 24 radial scans in which the base of the optic cup and Bruch's membrane were clearly visible were selected for each subject. We verified that no blood vessels were included in the region covered by the cross-sectional image to eliminate the possibility that these might cover or provide a shadow that would obscure the underlying structures. Of the two scans, one that was more vertically oriented and another that was more horizontally oriented were selected. In each selected scan, the ends of Bruch's membrane were marked graphically by a horizontal line, and the distance between these intersections with the ILM was measured (Fig. 3c). These distances were denoted as the vertical (d_v) and horizontal (d_d) diameters of the optic cup. The following equation for elliptical shapes was used to calculate the area (A_Cup) of the optic cup:  

\begin{eqnarray}{A_{{\rm{Cup}}}} = \frac{1}{4}{d_{\rm{d}}}{d_{\rm{v}}}\pi\end{eqnarray}

(1)

The ratio of the two surface areas is the CDR. 

For each selected scan, the vertical distance between the deepest point of the cup and the horizontal connection line (dashed line in Fig. 3b) was measured and used to determine the cup depth (t_Cup). The maximum cup depth determined in each of the evaluated scans was used in further evaluations. 

LF Fundus Camera

To analyze the LF data, it was first necessary to reconstruct the depth information. The software provided with the LF imager was used for this purpose as well as to export the data obtained. Details regarding the software settings are summarized in the Appendix. 

The point cloud (x, y, z) and the fundus images were exported (Fig. 5, raw data). Data processing was performed using a newly implemented algorithm via MATLAB 2018b (MathWorks, Natick, MA). A flowchart of this algorithm is shown in Figure 5 with data presented in generalized measurement units (MUs). 

Figure 5.

 

Flowchart documenting the proposed algorithm to calculate papilla parameters. Input includes reconstructed point cloud data from the LF fundus camera. Processing steps include preprocessing, surface reconstruction, and parameter determination. Determinations of the boundaries of the optic disc, optic cup, and neuroretinal rim are based on the rising and falling edges as shown. Results include the optic disc area, optic cup area, optic cup depth, and CDR.

The fix values discussed in the sections to follow were determined empirically based on findings obtained from the 14 healthy volunteers enrolled in our study. These data points are the first to be available that compare the results of a LF imaging system with those obtained using OCT. Values were chosen to create the best matches with OCT data based on the optic disc and optic cup boundaries. 

Preprocessing was performed to estimate the central coordinate and reduce artifacts. It was important to have some knowledge of the position of the papilla to facilitate further data processing. Thus, the central coordinate of the papilla was initially estimated from the original point cloud data. For this purpose, the data were cut in the z-direction with a slice thickness of 0.25 MU; the median position of each cut was determined, and the average of the mean coordinates was used as the estimated center of the papilla. 

Artifact reduction was performed to address the superposition of disturbances. For this purpose, tilt correction was performed by calculating a linear fit and then subtracting tilt and offset from the raw z-data. Because of increased noise in the periphery of the papilla, the edge region was truncated. The width of the truncated area was 12.5% of the total data diameter; thus, truncation reduced the dataset by approximately 25%. Tilt correction was done in the first step to avoid generating slanted cutting edges leading to reduced depth information in the periphery of the papilla. 

The next step in data processing was surface reconstruction to address the non-uniform data distribution. This was performed using a mathematical model with radial basis functions implemented in MATLAB as described by Chirokov.³¹ The following empirically determined implementation settings were used: Function, multiquadric; Sigma, 0.0796; and RBFsmooth, 2. For superior processing, the data were mapped as an equidistant grid with steps of 0.01 MU. This provided a smooth surface with less noise to optimize parameter determination. 

For a more precise determination of the boundaries of the optic disc and cup, the estimated center position was updated; the lowest point of the reconstructed surface was used as the new center position. Original values were then converted into polar coordinates; the center of the papilla was identified as the central coordinate. 

To obtain ONH parameters, the boundaries of the optic disc and optic cup were determined from the reconstructed papilla surface by analyzing the curve progression. The rising and falling edges of the optic disc and cup wall were evaluated with each arc beginning from the center of the optic disc and extending to the outer papilla. All arcs covered 1° segments of the papilla surface, and the full papilla was covered by 360 sections. In the implementation presented, index units (idx, where 1 idx = 0.01 MU) were used. 

Surface structures of the ONH exhibited different expression patterns, shapes, and wall slopes. A selection of these results is shown in Figure 6. The determination of the optic disc boundaries was thus carried out using case discrimination, which is a method that systematically considers a variety of different shapes. For boundary determination, findings were categorized into one of three groups based on cup appearance. The positions of the papilla and cup boundaries assigned to the respective angular sections were based on empirically determined weights based on LF fundus camera data. The parameters may vary when examining reconstructed surfaces using other imaging systems. The calculation steps described in the section to follow refer to an angular section of 1°. 

Figure 6.

 

Different optic cup shapes presented on OCT cross-sectional images. (a) Papilla without a prominent optic cup. (b) Papilla with prominent optic cup and a rising edge only in the curve progression. (c) Papilla with a prominent optic cup and a rising edge on the left side, and a rising and then falling edge on the right side.

First, the positions of the rising and falling edges were determined. A rising edge was defined at a slope within the range of 1/4 to 2/3 of the maximum gradient. A falling edge was characterized as a section with a gradient smaller than –0.2 MU/idx. Each sample was defined by the presence or absence of a prominent cup and the relative positions of the rising or falling edges: 

• • No prominent optic cup (Fig. 6a)—If the minimum and maximum height differed by less than 1 MU, only the optic disc position was determined. Due to the overall weak height profile in individuals without a prominent optic cup, the optic disc boundary was positioned at a characteristic local minimum (or maximum if there is no local minimum) at a distance from the central point.
• • Prominent optic cup identified (Figs. 6b, 6c)  
  • ○ Rising and falling edges (Fig. 6c with right edge of the papilla)—The rising edge represents the optic cup wall. The edge plateaus in the peripapillary region and is followed by a falling edge. The optic disc boundary (Disc) was positioned on the falling edge:  
    \begin{eqnarray}Disc\left( i \right) = \frac{2}{3}{\rm{id}}{{\rm{x}}_{{\rm{fall\;}}1}} + \frac{1}{3}{\rm{id}}{{\rm{x}}_{{\rm{fall\;end}}}}.\end{eqnarray}

    (2)
  • ○ Rising edge only (Fig. 6b, right and left edges; Fig. 6c with left edge of the papilla)—To determine the optic disc boundary position, an auxiliary quantity was determined that characterizes the mean rise after the initial rising edge. The auxiliary position denotes the idx at which this mean rise added with 0.2 MU/idx was undercut. The optic disc position was then calculated:  
    \begin{eqnarray}Disc\left( i \right) = \frac{1}{4}{\rm{id}}{{\rm{x}}_{{\rm{aux}}}} + \frac{3}{4}{\rm{id}}{{\rm{x}}_{{\rm{rise\;end}}}}.\end{eqnarray}

    (3)

If the optic disc boundary position was determined from the first third of data, the findings were checked to determine whether another characteristic surface rise occurred afterward. If this was the case, the optic disc position was re-examined according to the newly detected rise. 

In cases in which a prominent optic cup was identified, the optic cup boundary was then determined. The position of the rising edge was redefined within the optic disc position using a threshold value of 50% of the maximum gradient. An auxiliary position was also determined that reflects a height relationship. For this purpose, a position was determined at which the z-value was 0.75 MU smaller than at the optic disc boundary. The position of the optic cup boundary (Cup) was then calculated based on the positions of the rising edge and the auxiliary variable as follows:  

\begin{eqnarray}{\rm{Cup}}\left( i \right) = \frac{1}{3}{\rm{id}}{{\rm{x}}_{{\rm{rise\;cup\;}}1}} + \frac{1}{3}{\rm{id}}{{\rm{x}}_{{\rm{rise\;cup\;end}}}} + \frac{1}{3}{\rm{id}}{{\rm{x}}_{{\rm{aux}}}}. \quad \end{eqnarray}

(4)

The optic cup measurements were considered to be valid if positions were determined on at least 25% of the angle sections. 

Due to the separate consideration of the angle sections, there was a possibility that jumps in the boundary positions of neighboring angle sections might be observed. Therefore, a subsequent postprocessing smoothing operation was performed. A median filter of length 10 (corresponds to 10°) was used to eliminate extreme positions, followed by a mean filter of length 10. In addition, to reduce larger protrusions, a generalized ellipse was calculated from the coordinates of the boundary position of the optic disc and optic cup. The final positions of the optic disc and optic cup contours were then composed based 1/5 on the determined positions and 4/5 on the generalized ellipse, thus permitting individual details of the contour shape to be preserved. After transforming index positions back into coordinates, the contours were projected onto the individual fundus images for visualization. 

Finally, the geometric papilla parameters A_Disc, A_Cup, CDR, and t_Cup were calculated based on the determined optic cup and optic disc boundaries. These values correspond to the parameters calculated from OCT data. 

The surface areas of the optic disc and optic cup were determined as projections onto the x-y plane and based on the determined boundaries. The CDR was calculated as the ratio of the two surface areas. 

The maximum optic cup depth (t_{cup max}) was calculated from the data points within the cup (z_{within cup}) with the average height from the optic cup contour (z_{boundary cup}) serving as a reference. The maximum optic cup depth was then calculated as the distance between the mean height and the deepest point within the optic cup:  

\begin{eqnarray}\!\!\!\!\!\!\!\! {t_{{\rm{cup\;max}}}} = {\rm{mean}}\left( {{z_{{\rm{boundary\;cup}}}}} \right) - {\rm{min}}\left( {{z_{{\rm{within\;cup}}}}} \right). \end{eqnarray}

(5)

 

In order to compare OCT and LF data, the LF data needed to be scaled in the x-y plane and in the z-direction because of differences in magnification. To calculate the scaling factor, both the fundus image sections and the resolution of the imaging systems were considered. Several characteristic landmarks such as vessel crossings near the optic disc were identified, and the distances in pixels were determined manually. This approach corresponds to a metric calibration and is usually used to compare different imaging systems.^32,33 Furthermore, the resolution of the imaging systems was considered. In our system, there are different magnifications in the x-, y-, and z-directions. The x- and y-directions correspond to the orientation perpendicular to the optical axis and have the same magnifications. The z-direction corresponds to the direction with the optical axis and has a different magnification. After consultation with the manufacturer of the LF system, this can be approximated via a quadratic relationship between lateral and axial magnification.³⁴ This is due to the size of the microlenses and their distance to the sensor in our setup. 

For evaluation of the study results, the papilla parameters determined from the LF fundus camera data were multiplied with the scaling factors of the corresponding direction. The scaling factor in the x- and y-directions was 0.56; the scaling factor in the z-direction was 0.35. 

Explorative Analysis

As a qualitative parameter, the number of correctly determined optic cups was evaluated. The evaluation of OCT cross-sectional images provides a ground truth based on the classification in Figure 4. Incorrectly identified optic cups were classified into either false-positive or false-negative groups. Results are presented in a fourfold table. The geometric ONH parameters of the development dataset were compared quantitatively via descriptive analysis for the two imaging systems. The optic disc area, optic cup area, CDR, and optic cup depth were evaluated and compared, with the latter data presented as a violin plot; the median and mean, span, and distribution shape of the violin plots were determined. The mean values were determined from the results of the individual LF recordings. ONH parameters of optic cups designated as false positives were excluded from the analysis. The study participants were then divided into those with and those without a prominent optic cup. Data from the independent test data are presented separately for optic cup depth. 

Results

Qualitative Analysis of Correctly Identified Optic Cups

Fifteen images were collected from a single eye from each of the 14 study participants of the development dataset. Seven of the participants showed a prominent optic cup. Only nine of the 210 images were mismatched with regard to the correctly identified optic cup. The findings shown in the Table include the number of correctly and incorrectly assigned optic cups. Of the subjects exhibiting prominent optic cups, there were only two mismatches overall (1.9%) which were detected in two images from a single subject. Of the subjects without prominent optic cups, we identified seven mismatches (6.6%), which arose from data collected from four subjects. For the three subjects of the test data, one image was taken with each system. Two of the additional subjects showed a prominent optic cup, which was exhibited with both imaging systems and their analysis. 

Table.

 

Fourfold Table to Show Correct and Incorrect Identification of Optic Cups From 210 Images Collected From 14 Study Participants

 

Quantitative Analysis of Papilla Parameters

Our processing method was used to identify diagnostically relevant papilla parameters, including optic disc area, optic cup area, CDR, and optic cup depth. These parameters determined by each of the two imaging systems were compared. Using OCT, the optic disc area determined for all subjects ranged from 1.26 to 2.6 mm², which are values that correspond to the sizes of small to normal-sized discs.^5,35 The data presented in violin plots in Figure 7 include optic disc areas of the subjects based on the presence or absence of prominent optic cups as determined by each of the two imaging systems. This presentation reveals several specific differences between the two groups. We note that the span of the data points was smaller for those with compared to those without a prominent optic cup. 

Figure 7.

 

Violin plots used to compare the optic disc areas determined by the different imaging systems. (a) Subjects with a prominent optic cup (n = 7 subjects, 103 LF images, 7 OCT datasets), (b) Subjects without a prominent optic cup (n = 7 subjects, 98 LF images, 7 OCT datasets). Findings from subjects with a prominent optic cup show similar median and mean values and similar data point spans (OCT vs. LF camera). By contrast, findings from subjects without a prominent cup exhibit a larger median and mean value and span compared to those collected via OCT.

The overall ranges of optic disc areas determined from LF fundus and OCT images that feature a prominent optic cup were similar to one another (Fig. 7a). The median value of the LF data distribution was only 0.17 mm² higher than that calculated using OCT data. Mean values were similar for the imaging systems. Mean standard deviation between the results of the different LF images was 0.16 mm². 

Larger differences in optic disc areas were observed in the group without prominent optic cups. As shown in Figure 7b, the median value determined from the LF data was 0.48 mm² larger than that calculated from OCT data. The mean value for the LF data was higher, as well. The main distribution of the LF data was more narrowband and with a higher value than in OCT data. Mean standard deviation between the results of the different LF images was 0.24 mm². 

Further analysis was performed to evaluate parameters based on images that indicated prominent optic cups. Of note, cup area, cup depth, and CDR would be effectively zero for all participants without optic cups; if evaluated together, this would result in an inappropriate skew of the data distribution. 

The manually determined optic cup area from OCT data resulted in optic cup areas ranging from 0.17 to 0.86 mm². Figure 8a presents a violin plot of the optic cup areas determined for each of the two imaging systems. Although the overall shape of the data distribution was similar in both systems, the mean and median values determined from LF camera data were somewhat lower than those calculated using the OCT system; the median value was 0.04 mm² smaller. The mean standard deviation between the results of the different LF images was 0.06 mm². 

Figure 8.

 

(a) Violin plot for the comparison of the optic cup areas determined using different imaging systems. Data shown are from subjects with prominent optic cups (n = 7 subjects, 103 LF images, 7 OCT datasets). Shape of the distribution are similar for the imaging systems but median, mean, and span are smaller for LF data. (b) Violin plot for the comparison of the CDRs calculated using data from different imaging systems. All data pertain to subjects with prominent optic cups (n = 7 subjects, 103 LF images, 6 OCT datasets). The mean and median values and range of the LF data are smaller than those determined using OCT data.

CDRs were then calculated based on the surface areas of the optic discs and optic cups. These ratios ranged from 0.09 to 0.46 based on images obtained from subjects with prominent optic cups (Fig. 8b). Although the overall range of CDRs determined from the LF data was somewhat lower than that calculated from the OCT data, both the mean and median values were lower in the former group; the median value was 0.07 smaller. The mean standard deviation between the results of the different LF images was 0.04. 

Figure 9.

 

(a) Violin plot for the comparison of optic cup depths based on data from two imaging systems. Shown are findings from subjects with prominent optic cups (n = 7 subjects, 103 LF images, 7 OCT datasets). Both the span and median values determined from LF data are smaller than those determined using OCT. (b) Relationship between measured cup depth by OCT and LF camera in each subject (a linear approximation, where x is the cup depth in mm from OCT data; y is the cup depth in mm from LF camera data; n = 15 LF images per subject from 7 subjects). Additional test data from two subjects with one LF image each are marked with black asterisks.

One new development was that these two imaging systems could be used to determine optic cup depth (Fig. 9a). Optic cup depths between 0.1 and 0.48 mm were determined using OCT; a somewhat smaller span was calculated from the LF data. Although the shapes of the data distributions remained similar to one another, the median value for optic cup depth determined from the LF data was 0.07 mm smaller than that identified by OCT data. The shapes of the data distributions were similar. Mean standard deviation between the results of the different LF images was 0.027 mm. The data in Figure 9b include optic cup depths determined for each subject using both imaging systems with a linear approximation (R² = 71.77%; P < 0.05) as shown. The mean values were determined from the results of the individual LF images. The results of the two test datasets (subjects with prominent optic cup) are plotted as black asterisks in Figure 9b. Optic cup depths determined using OCT were 0.11 and 0.49 mm. The additional data cover the model span well. 

Discussion

This study aimed to compare the results of the processed LF fundus data with findings obtained using OCT to determine diagnostically important geometric parameters associated with papillae in the human eye. We enrolled 17 healthy subjects and collected fundus images using both imaging systems. LF camera data were analyzed with a novel processing algorithm, and the OCT data were processed using glaucoma module software from the manufacturer. Analyzed and calculated parameters included optic disc area, optic cup area, CDR, and maximum optic cup depth. Our analysis revealed that nine of the original 17 participants exhibited a prominent optic cup. All geometric parameters were assessed in this cohort with findings that were in good agreement with those obtained from the OCT data. As a means of understanding the differences between these two imaging systems, we present both the overall setup and the processing algorithm. 

Optic cup depth was examined in the nine participants with prominent optic cups (Fig. 9). Results from the two imaging systems were similar to one another, with depths ranging from 0 to 0.5 mm, although the range and median were smaller when calculated from the LF data. The depths for the independent test datasets fit well with the previous results. Different imaging systems with different illumination systems have been used for data acquisition, and different depths of the fundus are detected by a reflection based on different wavelengths of light. The LF system features green light and evaluates the surface of the fundus.²⁴ OCT data reveal distinct retinal layers based on the coherence of the signal. During surface reconstruction, LF data are smoothed. Although this is critical to reduce artifacts, it can result in an underestimation of the maximum optic cup depth. Optic cup depth is also measured with respect to the mean height of the optic cup boundary. In cases of prominent optic cups or in patients diagnosed with glaucoma, the edges of the optic cup are pronounced and steep. If the optic cup surface is underestimated and the mean height of the optic cup boundary is too low, the cup depth will be underestimated. A comparison of LF with OCT findings from each individual revealed a linear trend. Our results also revealed a different linear approximation compared to that reported by Schramm et al.²⁴ This may be because different scaling factors were used to transform v-mm units provided by the LF camera into true mm units. Furthermore, participants in the Schramm et al.²⁴ study exhibited more prominent optic cups compared to the results presented here. It is somewhat difficult to analyze results from smaller optic cups with the methods that are currently available, and the percentage error as well as the standard deviation are higher than those determined for more prominent optic cups. 

The existence of prominent cups and their correct identification by LF camera methods was determined as a qualitative parameter. Most of the images of the development data collected from the 14 participants were categorized correctly (Table), and the false-negative rate was lower than the rate of false positives. For the independent test data, all three subjects were classified correctly. We concluded that the identification of a prominent optic cup by LF methods was an overall reliable conclusion, and we detected only a few incorrect assignments, data points, or artifacts capable of influencing surface reconstruction. The categorization of images that exhibited prominent optic cups was based on the height differences between the deepest point of the cup and the edge of the neuroretinal rim. Subjects with prominent optic cups typically exhibit small neuroretinal rims and deeper optic cups. By contrast, participants who do not have prominent optic cups exhibit tissue bulges around the papillae as well as neuroretinal rims and RNFLs that are thicker in general. If one only examines the surface of the fundus, it becomes difficult to differentiate between images with a high neuroretinal rim as opposed to a deeper optic cup. The OCT imaging system provides an interpretation based on the properties of the tissue layers below the fundus surface, most notably Bruch's membrane. 

Optic disc areas were analyzed from participants both with and without prominent optic cups (Fig. 7). We observed several notable differences between the two groups. For example, whereas the data spans observed in the group of subjects with prominent optic cups are similar to one another, the median value calculated from the LF data is higher than that observed using OCT. The median optic disc area is overestimated by LF data by 0.17 mm². As discussed previously, LF data represent the fundus surface, and raw data are overlaid with artifacts that are reduced during processing. The papilla edges are also smoothed slightly during this processing step, which results in a flatter ONH. By contrast, the determination of the optic disc boundary from OCT data is based on the position of the opening point of Bruch's membrane and not on any features associated with the fundus surface. Furthermore, shadows on deeper retinal layers and Bruch's membrane may be detected if vessels are located appropriately within the cross-sectional scan. This is an important area of uncertainty also for this imaging system. The differences between imaging systems as well as the standard deviation were larger in our evaluation of subjects who did not exhibit prominent optic cups. Both the range and median calculated from the LF data were higher than in the OCT data; likewise, the optic disc area was overestimated. Characteristics of this subject group included the small variation in height within the papilla and a partially pronounced neuroretinal rim. Detection of rising and falling edges of the papilla was also more difficult in this cohort compared to subjects with a more prominent optic cup. There was a higher uncertainty in data processing, which is reflected in a larger standard deviation. Papilla boundaries were determined for different cross-sections. As is apparent, differences in several cross-sections add up to larger deviations when calculating two-dimensional parameters such as area. Alternatively, one-dimensional parameters (i.e., the horizontal, vertical, maximum, and minimum diameters) can be used for evaluation. In future work, optic disc area may also be used to divide the ONH into micro, macro, and normal optic discs based on their size. This may be helpful to initially differentiate between glaucomatous and macro discs. Subsequently, follow-up examinations are very important for further diagnosis and therapy. 

The optic cup area was also determined for participants exhibiting prominent optic cups (Fig. 8a). The median values are somewhat smaller based on LF data compared to OCT, which resulted in an underestimation of the overall area. These values were smoothed to reduce artifacts when the LF data were processed, and the optic cup boundary was identified at an inappropriate distance inside the cup based on surface features alone. As noted above, the analysis of OCT data uses the position of Bruch's membrane to determine the relevant boundary. The challenges involved in evaluating the optic disc area also reflect on the calculation of the area of the optic cup. 

The CDR is a dimensionless calculated ratio of the optic cup area to the area of the optic disc (Fig. 8b). Any deviations in the determination of area will of course be reflected in the calculation of this ratio. As discussed above, the area of the optic disc is frequently overestimated and the area of the optic cup is underestimated using LF data; thus, the values for CDR are most likely underestimated, as well. 

Although OCT technology was used as a reference system in this study, this system also has several limitations. It is critical to recognize that OCT is a scanning-based imaging system and can incorporate movement artifacts that need to be corrected.²⁰ Likewise, shadows below vessels can obscure critical findings and cover the opening of Bruch's membrane. This can create significant problems in and around the ONH, as there are many shadows in this region resulting from the central retinal artery and vein and their branches. The dimensions of these shadows vary and can depend on the course of the vessel. One cannot rely on adequate detection of the optic disc and optic cup diameter in every B-scan. To optimize image acquisition, the operator positions the scan center as centrally as possible within the papilla to ensure that both horizontal and vertical scans represent the minimum and maximum optic disc diameters. Given the elliptical shape of the optic disc, an eccentric position can lead to an incorrect determination and underestimation of the maximum diameter. This becomes a significant problem when calculating parameters because the optic cup is frequently positioned eccentrically in the papilla. Likewise, the cup depth can only be determined precisely if the B-scan passes through its deepest point; if this is not the case, the optic cup depth will be underestimated. 

The LF fundus camera and novel processing methods are also subject to several critical limitations. The LF fundus camera is not calibrated metrically but to a corresponding virtual depth scale. For conversion to metric calibration, one would need to use a special eye model with a measurement scale with a high likelihood that aberrations would ensue. Furthermore, the fundus camera and the LF imager are not fully aligned with one another; the F-numbers of the microlens array and the optical systems do not yet completely match one another. This means that lateral resolution and depth resolution are not fully exploited. A new optical design with matching F-numbers and smaller fields of view will be needed to increase depth resolution. Reconstructed data were sampled irregularly and non-equidistantly, a notable issue in areas with comparatively high and low data density. High data density regions include areas with high contrast such as the edge of the optic disc and along the vessels.^24,25 Additionally, the number of data points obtained varied strongly among images and among subjects which resulted in misestimates of the reconstructed depth information. Depending on the individual anatomy, papilla structures were often superimposed, thus limiting the accuracy of any further analysis. Limitations associated with the use of a fundus camera include wavefront errors, reflections, and scattered light. The susceptibility to stray light and misalignment is increased for the light field fundus camera used. Critically scattering media from the anterior (cornea) and medial (lens) segments of the eye can cause back reflections and backscatter into the viewing system if the alignment is not ideal. Reflections can then occur in several microlenses, which makes reconstruction difficult or can lead to the artifacts with strong frequency components. Correct adjustment to precisely separate the observation and illumination beam path is particularly important to achieve good image quality.^24,25,36 

The presented processing method was adapted to accommodate uneven and noisy data such as that obtained from the LF fundus camera. Although smoothing methods were used to address this problem, this process may eliminate individual details, such as gradual increases in the height of the optic cup wall of the papilla. Furthermore, the slope of the cup wall will be flattened and thus not representative of its true anatomy. Further development of this algorithm might focus on refining the evaluations of the optic disc and optic cup borders, which may require more accurate modeling to be useful for the diagnosis of glaucoma. Other optic cup forms might be evaluated by OCT to create images that can be used to evaluate the anatomical diversity of ONH shapes. The geometric papilla parameters determined to date are based on the analysis of OCT data. Future developments might include new parameters specifically adapted to LF data processing and properties. 

We enrolled 17 healthy young participants to examine the validity of the newly developed processing algorithm as part of a preliminary study. Of the selected subjects, three subjects were used as independent test data. During analysis, it became apparent that only nine of our initial 17 subjects exhibited a prominent optic cup. No glaucoma patients took part in the study; therefore, no statistical evaluations nor any assessments of the role of this technology in disease diagnosis are possible at this time. 

In summary, LF technology facilitates the acquisition of 3D images of the fundus with a single photographic shot. The algorithm presented permitted us to analyze the data and calculate diagnostically relevant parameters of the ONH, including optic cup area, optic disc area, CDR, and the maximum optic cup depth. The study highlighted relevant trends and revealed that the calculated parameters were similar to those identified by the OCT reference system. 

Acknowledgments

The authors thank the Federal Ministry of Education and Research for the funding (13GW0331B to FKZ). The authors thank the Open Access Publication Fund of the Technische Universität Ilmenau for support for the publication costs. 

Disclosure: L. Wegert, None; S. Schramm, None; A. Dietzel, None; D. Link, None; S. Klee, None 

Appendix

The Table A1 provides the parameters of the Raytrix software (RxLive 5.0.046.0; Raytrix GmbH, Kiel, Germany). The parameter values used are listed with a short description. 

Table A1.

 

Processing Parameters of the Raytrix Software With the Values Used (Structure Is Oriented to the Software)