High Variation in Inner Retinal Reflectivity Predicts Poor Visual Outcome in Patients With Central Retinal Vein Occlusion: SCORE2 Report 21

Abstract

Purpose: To assess the association of a novel spectral domain optical coherence tomography biomarker with 6-month visual acuity in in the Study of COmparative Treatments for REtinal Vein Occlusion 2.

Methods: Spectral domain optical coherence tomography volume scans were evaluated for inner retinal hyperreflectivity, quantified by optical intensity ratio (OIR) and OIR variation. Baseline visual acuity letter score (VALS), baseline OCT biomarkers, and month 1 OIR were correlated with VALS at month 6. Regression trees, a machine learning technique yielding readily interpretable models, were used to assess for variable interaction.

Results: Only baseline VALS correlated positively with month 6 VALS in multivariate regression. Regression trees detected a novel functional and anatomical interaction in a subgroup. Among patients with a baseline VALS worse than 43, those with an OIR variation at month 1 of more than 0.09 had a mean of 13 fewer letters of vision at 6 months compared with patients with an OIR variation of 0.09 or less.

Conclusions: Baseline VALS was the strongest predictor of month 6 VALS. Regression tree analysis detected an interaction effect, in which higher OIR variation at month 1 predicted worse 6-month VALS in patients with low VALS at baseline. OIR variation may serve as a predictor for poor visual outcome despite treatment of macular edema secondary to retinal vein occlusion in patients with poor vision at baseline.

Translational Relevance: Pixel heterogeneity in three-dimensional OCT data may serve as measure of disruption of the retinal laminations, and this factor may carry visually prognostic value.

Introduction

Retinal vein occlusion (RVO), the second leading cause of vision loss owing to retinal vascular disease,¹ results in vision loss most commonly through the development of macular edema, which can occur in up to 15% of eyes with branch RVO and 30% of eyes with central RVO (CRVO).^2,3 Although intravitreal anti-vascular endothelial growth factor (VEGF) therapy has been commonly used for the treatment of macular edema secondary to RVO,⁴ visual outcomes are heterogenous.^5,6 According to the Study of Comparative Treatments for Retinal Vein Occlusion 2 (SCORE2), a trial that included patients with macular edema due to CRVO or hemiretinal vein occlusion (HRVO), 34.9% of patients in the aflibercept arm and 38.7% of patients in the bevacizumab arm failed to gain three electronic Early Treatment Diabetic Retinopathy (e-ETDRS) lines of vision despite six monthly injections of anti-VEGF.⁷ A variety of clinical and imaging features have been explored as potential prognostic factors for visual acuity (VA) in the treatment of macular edema with anti-VEGF therapy in patients with RVO. 

One hypothesis, driven by a historical understanding of the visual outcome differences between patients with ischemic versus nonischemic CRVO,⁸ is that the identification of markers of retinal ischemia may predict poor visual outcome. Historically, the severity of the relative afferent pupillary defect,⁹ scotoma size on Goldmann perimetry,^5,10 and various indices on full-field electroretinogram^9–11 have been used as markers for retinal ischemia and have been linked to poor visual outcome. Several findings on spectral domain optical coherence tomography (SD-OCT) have been implicated as markers of retinal ischemia in RVO, including enlargement of the foveal avascular zone,^12,13 retinal whitening on en face SD-OCT,^14,15 disruption of the ellipsoid zone (EZ),¹⁶ and disorganization of the retinal layers (DRIL).^17,18 Although many of these have been associated with poor VA on an individual basis,^17–21 a multifactorial analysis by the SCORE2 group suggests that no particular biomarker offers further prognostic ability when evaluated in conjunction with baseline VA.²² 

Recently, the optical intensity ratio (OIR) has been proposed as a potential marker of retinal ischemia.²³ By measuring the reflectivity signal at the pixel level of the inner retina over a three-dimensional area of the OCT cube and normalizing the signal to the mean reflectivity of the retinal pigment epithelium (RPE), both mean reflectivity and the heterogeneity of the reflectivity signal can be compared between individual scans. OIR and OIR variation at 1 month from baseline was associated with poorer visual prognosis at 1 year in a cohort of patients treated with anti-VEGF therapy for macular edema secondary to CRVO.²⁴ However, in further study, baseline VA was still a stronger prognostic factor compared with the measured OIR and OIR variation values.²³ 

The current study aims to apply OIR analysis to data collected in a large, prospective, multicenter clinical trial, the SCORE2 trial, to investigate its potential additional prognostic value in combination with previously assessed SD-OCT biomarkers in this dataset. Using machine learning statistical techniques, the authors sought to evaluate whether a combination of prognostic factors may yield greater predictability than any individual risk factor. 

Methods

Study Participants

Study data were obtained from SCORE2, a multicenter, prospective, randomized noninferiority trial of 362 eyes with macular edema secondary to CRVO or HRVO comparing intravitreal bevacizumab with aflibercept (clinicaltrials.gov identifier NCT01969708).⁷ The study was approved by the institutional review boards associated with each center and adhered to the tenets of the Declaration of Helsinki. The study visits were conducted with treatment of either aflibercept or bevacizumab provided on a monthly basis from baseline through month 6 (M06), at which time participants with a protocol-defined good response were a re-randomized within each arm to monthly versus treat-and-extend dosing and participants with a protocol-defined poor or marginal response were switched to an alternative treatment (patients in the aflibercept arm were switched to treatment with dexamethasone implant and participants in the bevacizumab arm were switched to aflibercept).²⁵ After month 12, treatment was per investigator discretion. Inclusion criteria were center-involved macular edema, defined as a central subfield thickness (CST) of 300 µm or more (or ≥320 µm if measured on a Heidelberg Spectralis Machine; Spectralis Heidelberg Engineering, Heidelberg, Germany), and an e-ETDRS VA letter score (VALS) of 19 or greater (approximately 20/400) and 73 or less (approximately 20/40). Eyes with cataract and media opacity were excluded to ensure quality image acquisition. 

Data Collection

All SD-OCT scans were acquired by certified photographers using the SCORE2 reading center (Fundus Photograph Reading Center, University of Wisconsin-Madison) approved protocol with either a Carl Zeiss Meditec Cirrus (Carl Zeiss Meditec, Dublin, CA) or a Heidelberg Spectralis (Spectralis Heidelberg Engineering, Heidelberg, Germany) OCT machine on dilated eyes.²⁵ The Zeiss macular volume scans were 6 mm and composed of 512 A-scans and 128 B-scans, and the Heidelberg scans were 20 × 20° and composed of 512 A-scans and 97 B-scans. Imaging was obtained from one eye of each study participant. SD-OCT scans were evaluated at baseline and month 1 (M01). Best-corrected VALS was measured using the e-ETDRS protocol at baseline, M01, and M06. 

SD-OCT grading performed by the SCORE2 reading center, as described in detail in SCORE2 Report 15, was collected for analysis.²² Qualitative observations recorded included subretinal fluid (present vs, absent), intraretinal fluid (present vs. absent), EZ morphology within the CSF (normal, patchy, or absent), and DRIL both inside and outside the 1-mm foveal centered area. DRIL, as originally established by Sun et al.,²⁶ was defined as the inability to distinguish the boundary between the ganglion cell layer–inner plexiform layer complex from the inner nuclear layer and outer plexiform layer occupying a horizontal extent of 50% or greater (500 mm) of the measured area. The presence of DRIL required that two or more consecutive scans be involved. The CST was calculated from top of the internal limiting membrane to the top of the RPE.²⁷ Subjects with ungradable values were excluded. 

OIR Calculation

An analysis of the OCT data and determination of OIR were semiautomated as per the protocol described by Mehta et al.²⁴ (Fig. 1) and can be reviewed publicly at https://github.com/thenitman/OIRscript. SD-OCT cubes from each patient were extracted in raw format at the M01 visit for analysis. Retinal morphology distortion and widespread macular edema limited the ability to consistently perform OIR analysis on the baseline images and, thus, the baseline OIR data were not included for analysis. The original raw OCT cubes were processed for layer detection. OCT cube data were divided into three 1 × 1-mm cubes at 1 mm temporal, superior, and inferior to the foveal center, from which the OIR was calculated using a custom script that divides the OCT signal intensity of the inner retina (ganglion cell layer–inner plexiform layer complex) by the signal intensity of the RPE for each A-scan. Layer identifications that were not continuous with adjacent A-scan measurement locations were excluded to account for identification failure. A single OIR from a single A-scan was calculated by dividing the mean denoised inner retina A-scan signal intensity by the mean denoised RPE A-scan signal intensity. The process was then repeated on all A-scans within the three OCT cubes and approximately 9000 OIRs were averaged to create a final OIR. The standard deviation of values for the scans obtained was calculated to ascertain the variation in reflectivity measurement. Visual quality control was verified by author N.M. to ensure correct segmentation of the layers of interest. 

Figure 1.

 

Visual representation of the OIR semiautomated algorithmic analysis. (A) Sample en face OCT scan with colored overlays demonstrating locations of the temporal, superior, and inferior extracted cubes for OIR measurement. (B) Sample foveal B-scan flattened to the level of the RPE is shown with an extracted raw A-scan to the right corresponding to the thin blue overlay line. Red and blue arrows highlight areas of the A-scan that were algorithmically identified (using peak identification) to correspond with the ganglion cell layer–inner plexiform layer complex and the RPE, respectively. This process is repeated for all A-scans within the extracted cube. (C) Tracings of A-scan pixel intensities are displayed with relative brightness shown on the y-axis and location along the A-scan (vertically down the B-scan) on the x-axis. Red and blue arrows again highlight target areas. (D) Placement of the red and blue locations back onto the original image confirms proper layer identification. (E) Histogram representation of distribution of OIRs obtained by dividing the pixel intensity of the inner to the RPE. The majority of OIRs are centered around 0.8 but range from 0.6 to 1.1.

Statistical Analyses

Categorical values were converted to numeric values and continuous values were log-transformed to permit statistical analysis. Baseline VALS, baseline and M01 OCT grades, and M01 OIR values (globally and of the temporal retinal cube) were assessed for correlation among themselves and then were regressed against M06 VALS in both univariate and multivariate fashion. To account for multiple statistical testing and to guard against type I errors, a P value with Bonferroni correction of less than 0.005 was considered statistically significant. Statistical analysis was performed using python scripting. Regression trees, a machine learning technique yielding readily interpretable models, were used to assess for variable interaction. To guard against overfit, a maximum tree depth of two was used and each leaf of the tree had to contain at least 10% of the data. 

Results

SD-OCT scans were available for analysis for 361 of the 362 participants at baseline and 357 participants at M01. The mean patient age was 68.9 ± 12.0 years at randomization. The mean VALS improved from 50.3 ± 15.2 at baseline to 69.3 ± 17.5 at M06. The mean CST decreased from 665.0 ± 223.2 mm at baseline to 260.6 ± 103.7 mm at M06. 

Correlations between prognostic indicators²² were assessed and are presented in Supplemental Table S1. The presence of DRIL, subretinal fluid, EZ defects, and increased CST were all relatively correlated with each other at baseline and M01. An increased OIR was correlated with increased OIR variation. Increased temporal OIR variation was highly correlated with the presence of DRIL outside the 1-mm foveal center. 

Results of regressions using M06 VALS as the response variable are presented in Table. Baseline VALS was highly correlated (r² = 29%; P < 0.001) with M06 VALS. Baseline DRIL outside the foveal central zone (P = 0.001), increased OIR mean (P = 0.002), and increased OIR variation of the temporal cube (P < 0.001) were all negatively associated with M06 VALS, although r² values were all less than 2%. Only baseline VALS correlated positively with M06 VALS in multivariate regression (P < 0.001). 

Table.

 

Results of Regressions Using the Level of VA at M06 (M06 VALS) as the Response Variable

 

Regression tree analysis (a simple machine learning model that can identify optimal cut-off values within a dataset) was performed on the same predictor candidates (baseline SD-OCT characteristics, baseline VALS, and M01 OIR) to M06 VALS. Baseline EZ defects were excluded owing to a high number of ungradable images (75.1%). The regression tree analysis included 259 eyes. A novel interaction between baseline VALS and M01 OIR was revealed (Fig. 2). Patients with a baseline VALS of less than 43.5 letters (approximately 20/125) and an OIR variation of more than 0.086 (unitless) had an average of 13 fewer letters of vision at M06 (48.1 letters or approximately 20/120) compared with patients with OIR variation of 0.086 or less (61 letters or approximately 20/63). 

Figure 2.

 

Visual representation of regression tree analysis of baseline SD-OCT factors, baseline VA, M01 OIR scores to M06 VALS demonstrating relationship between baseline VA, and temporal OIR variation. There were 259 eyes with complete records of disorganization of the inner retina layers both inside and outside the foveal center, intraretinal fluid presence, subretinal fluid presence, OIR mean and variation values both globally and in the temporal retina, and baseline VA. The average M06 vision for the entire cohort was 70.4 letters (top row, center box). The prognostic value with the greatest distinguishing ability was noted to be baseline VA at a cut-off of 43.5 letters. Patients with baseline a VA above that value were noted to have an average M06 VA of 75.2 letters (199 eyes, middle row, right box); however, patients with baseline VA lower than 43.5 letters were found to have a M06 vision of 54.1 letters (60 eyes, middle row, left box). Within patients with baseline VA lower than 43.5 letters, those with a temporal OIR variation of greater than 0.086 were found to have a M06 VA of 48.1 letters (26 eyes, bottom row, far left box) versus 61 letters in a cohort of patients with lower OIR (34 eyes, bottom row, middle left box). In patients with baseline VA of greater than 45.5 letters, patients with baseline VA between 43.5 letters and 58.5 letters were found to have an average M06 VA of 71.2 letters (95 eyes, bottom row, middle right box). Patients with baseline VA of greater than 58.5 letters were found to have an average M06 VA of 79.0 letters (104 eyes, bottom row, far right box).

The average M06 vision for the entire cohort was 70.4 letters (approximately 20/40). The prognostic variable with the greatest distinguishing ability was noted to be baseline VA at a cut-off of 43.5 letters. This was determined empirically through the nature of regression tree analysis, not a priori. Patients with a baseline VA above that value were noted to have an average M06 VA of 75.2 letters (approximately 20/32); however, patients with a baseline VA of lower than 43.5 letters (approximately 20/125) were found to have an average M06 vision of 54 letters (approximately 20/80). A relationship was noted between M01 temporal OIR variation and baseline VA only in the cohort of patients with a baseline VA of lower than 43.5 letters. Within that group, those with a M01 temporal OIR variation greater than 0.086 were found to have a M06 VA of 48.1 letters (approximately 20/125) versus 61 letters (approximately 20/63) in the cohort of patients with a lower M01 temporal OIR variation. In patients with a baseline VALS of better than 43.5 (approximately 20/125), baseline SD-OCT parameters were not helpful in further prognostication. The M06 VA was predicted to be 79.1 letters (approximately 20/25) versus 71.2 letters (approximately 20/40) when baseline vision was above or below 58.5 (approximately 20/70) letters, but no structural–functional relationship was observed. 

An increased r² of the regression tree model (35.5%) compared with multivariate regression of OIR and baseline VALS (32.3%) suggests an increased worth of temporal OIR when used in a regression tree rather than a simple regression model. Both models had a slightly higher r²value than regression of baseline VALS to M06 VALS alone (r² = 29%). 

Discussion

An OIR analysis was developed in an attempt to add a quantitative predictive biomarker to the arsenal of OCT features currently being explored. Although noted to be correlated with poor final acuity in patients in RVO, the OIR itself has not yet been found to provide any additional prognostic value beyond what can be accomplished with VA alone.²³ Similarly, the relationship between other SD-OCT features and VA has been inconsistent, possibly because of the lack of robust statistical models, retrospective analyses, and the small sample size. We sought to investigate the relationship between OIR, multiple SD-OCT features, and VA in 362 eyes with macular edema secondary to CRVO or HRVO in the prospective, multicenter, phase III SCORE2 trial comparing intravitreal bevacizumab with aflibercept. Specifically, we used a machine learning model designed to reveal interaction effects between variables that cannot be ascertained by linear regression alone. 

Weak negative univariate correlations were noted between M06 VA and DRIL outside the foveal center, and between M06 VA and OIR mean and variation. These prognostic implications were minimal when compared with the prognostic impact of the baseline VA and were lost when analyzed in a multivariate fashion with baseline VA included. Clearly, presenting vision has paramount prognostic importance in the treatment of RVO with macular edema with anti-VEGF therapy in the first 6 months. However, subtle interaction effects between baseline VA and presenting SD-OCT features may be present that are not able to be appreciated with linear regression. 

A regression tree analysis is a simple machine learning model in which the dataset is divided into two groups at varying cut-off values of the predictor variables until maximal discriminating ability is reached, by minimizing the residual sum of squares. Two new branches are then created and the new nodes (leaves) are assessed on the remaining variables until the predetermined number of branches and nodes are achieved. To help avoid overfit and bias, a maximum tree depth of two was used and each leaf of the tree was required to contain at least 10% of the data. The model can then relay which factors have proved to be prognostic and display the relationship between the variables. In contrast with popular deep learning techniques, the results of regression trees are readily interpretable. First, the segregation of the data based on prognostic variables can be represented visually. This hierarchical output of predictive variables can reveal the relative importance of each input variable in the model. Second, the segregation or partitions of the data can be queried further and verified. 

Visual representation of regression tree analysis of baseline SD-OCT factors, baseline VALS, and M01 OIR scores to M06 VALS is shown in Figure 2. A relationship was noted between M01 temporal OIR variation and baseline VA only in the cohort of patients with a baseline VA of lower than 43.5 letters. Within that group, those with a M01 temporal OIR variation of greater than 0.086 were found to have a M06 VA of 48.1 letters versus 61 letters in a cohort of patients with a lower M01 temporal OIR. In simpler terms, patients who presented with poor vision and had heterogenous signal in the temporal retina had poorer vision than those who had poor vision on entering the study but relatively homogenous signal in the temporal retina. It is important to note that baseline SD-OCT features such as DRIL, intraretinal fluid, subretinal fluid, and CST were included in the analysis but not noted to add any predictive value to the model. 

OIR variation provided distinguishing ability in a regression tree analysis, but the OIR mean did not. The OIR mean or inner retinal hyper-reflectivity may be a transient manifestation of ischemia in RVO that does not carry prognostic import. In contrast, increased OIR variation may be indicative more severe retinal disorganization and may be representative an ischemic threshold in which the treatment of macular edema with anti-VEGF therapy fails to result in a visual improvement. Nevertheless, these potential biomarkers remain exploratory and further validation is required before suggesting representation of a biologic phenomenon. 

Moreover, only OIR and OIR variation in the temporal retinal cube were found to be prognostically valuable in any analysis. This finding is most likely due to the inclusion of structural anatomy in the superior and inferior cubes, such as inner retinal hemorrhages and blood vessels that artifactually influenced OIR measurements. The temporal retina may also be more susceptible to ischemia owing to distance from the central retinal artery, and thus more likely to manifest prognostic signal. 

This incremental advancement in our understanding of the ability to predict M06 VA is consistent with prior reports from the SCORE2 team, with a few differences attributable to statistical technique and sample and data acquisition. The SCORE2 report 14 used quantitative and qualitative analysis of EZ defects and noted that at all visits VA was better in eyes without an EZ defect compared with those with an EZ defect.¹⁶ Quantitative data of EZ loss were not available for this study. Owing to a large number of qualitatively ungradable images, EZ analysis was excluded from our regression tree. EZ loss or attenuation may hold more value later in the treatment course, given the difficulty with EZ assessment early in the disease course, perhaps owing to significant shadowing of the EZ from intraretinal thickening and fluid. Improvements in OCT fluid resolution with subsequently better assessment of EZ may help to define the role of photoreceptor imaging in vision prediction clearly. 

This difficulty analyzing images with disruption owing to edema also precluded our ability to perform an OIR analysis on the baseline images. An OIR analysis was performed on the M01 rather than baseline images to minimize algorithm failure, and this factor may predispose to a look-ahead bias. In contrast, it may be valuable in a clinical setting to have the ability to use OIR variation as a biomarker after the first treatment with anti-VEGF to further clarify the predicted visual outcome in those patients with poor vision at baseline. 

Recently, Mimouni et al.¹⁷ and Babiuch et al.²⁸ demonstrated that DRIL at baseline and after anti-VEGF therapy was associated with worse VA in treatment-naïve eyes with macular edema secondary to RVO. However, in the SCORE2 report 15, a multivariate analysis using comprehensive SD-OCT imaging markers along with baseline age and VALS, DRIL was not associated with lower vision outcomes.²² The grading of DRIL is difficult and may explain why studies have not found a consistent association between DRIL and VA. Babiuch et al.²² admittedly reported poor interobserver correlations in DRIL, highlighting the difficulty in measuring DRIL accurately in RVO. 

Interestingly, temporal OIR variation being associated with DRIL outside the foveal center in our study suggests that the concept of computationally analyzing pixel heterogeneity may be used as a method to semiautomatically detect disorganization of the retinal laminations. Because the algorithm analyzes the inner retina in a horizontal fashion, an unperturbed retina would theoretically demonstrate a homogenous signal (low OIR variation) within a single retinal lamination. Ischemia may result in inner retinal layer hyper-reflectivity without layer disruption (high OIR and low OIR variation). Further ischemia may totally obliterate the homogeneity in the inner retina (both a high OIR and a high OIR variation) (Fig. 3). The presence of DRIL and high OIR variation may be complimentary surrogate markers of ischemia. However, given the different measurement techniques and the different areas of the retina analyzed (central vs. paracentral), they should not expected to always corelate. 

Figure 3.

 

(Top) One-month foveal B-scan images with red and blue overlays showing layer identification for OIR measurement. The subject on the left presented with 22 letters of vision, which improved to only 40 letters. Note the diffuse hyperreflectivity and retinal lamination disruption in the temporal retina that is persistent at 1 month after presentation. In contrast, the subject on the right demonstrated an improvement of vision from a baseline of 22 letters to 64 letters. Despite baseline poor vision and EZ loss present after a single anti-VEGF injection, the temporal retina demonstrates minimal hyperreflectivity and retinal lamination disruption. (Bottom) Subject on the left demonstrated high OIR and high OIR variation (0.99, 0.31). Note the broad distribution of OIRs through the temporal cube. The subject on the right demonstrated less elevated OIR and OIR variation (0.62, 0.07). The histogram reveals a tighter distribution of OIRs.

The current study identified an interaction effect with certain limitations that may limit external validity. As discussed previously, EZ assessments were excluded owing to a large number of qualitatively ungradable images and OIR analysis was performed on M01 images rather than baseline images. Finally, the regression tree results require validation in external datasets to ensure the reports are not the result of overfit to a unique dataset resulting in a type I error. Specifically, external validation to assess the predictive ability (sensitivity, specificity, and area under the curve) using the OIR variation cut-off value identified by the regression tree analysis is required. External validation will be necessary to see if the prognostic value of OIR as determined by this study is specific to the SCORE2 dataset or if the knowledge gained can be generalized to patients with CRVO with macular edema outside of this particular clinical trial setting. We expect a synergy between regression tree analysis and deep learning techniques in validating prognostic biomarkers. 

In conclusion, baseline VA is the most important variable in predicting future vision in patients treated with anti-VEGF therapy for macular edema secondary to CRVO or HRVO. Regression tree analysis revealed that high OIR variation is a mild negative influencing factor on M06 VA in patients with a poor baseline VA. High OIR variation may be an OCT predictor for poor visual outcome despite treatment of macular edema with anti-VEGF agents. Specifically, it may provide prognostic guidance to physician and patients in patients who have poor vision at baseline and manifest high OIR variation at 1 month despite treatment with anti-VEGF agents for macular edema secondary to HRVO or CRVO. 

Acknowledgments

Meeting Presentation: International Retinal Imaging Symposium, 2022. 

Disclosure: N. Mehta, Eyepoint Pharmaceuticals (F); S. Patil, None; V. Modi, None; R. Vardi, None; K. Liu, None; R.P. Singh, Apellis (F), Graybug (F), Alcon (C), Novartis (C), Asceplix (C), Bausch + Lomb (C), Genentech (C), Roche (C), Regeneron Pharmaceuticals (C), Zeiss (C), Gyroscope (C); D. Sarraf, Amgen (C), Endogena Therapeutics (C), Novartis (C), Genentech (C), Visionix/Optovue (C), Optovue/Visionix (S), Topcon (S), Heidelberg (S), Amgen (F), Bayer (F), Boehringer (F), Genentech (F), Iveric Bio (F), Novartis (F), Regeneron (F); N.L. Oden, None; P.C. VanVeldhuisen, None; I.U. Scott, Regeneron (C), F. Hoffmann-La Roche AG (C), Novartis (C); M.S. Ip, Alimera (C), Allergan (C), Amgen (C), Apellis (C), Clearside (C), Novartis (C), OCCURX (C), Regeneron (C), Biogen (F), Genentech (F), IVERIC Bio (F), Cell Lineage Therapeutics (F), REGENXBIO (F); B.A. Blodi, None; Y. Modi, Alimera (C), Allergan (C), Genentech (C), Théa (C), Zeiss (C)