Image quality assessment of retinal fundus photographs for diabetic retinopathy in the machine learning era: a review

Abstract

This study aimed to evaluate the image quality assessment (IQA) and quality criteria employed in publicly available datasets for diabetic retinopathy (DR). A literature search strategy was used to identify relevant datasets, and 20 datasets were included in the analysis. Out of these, 12 datasets mentioned performing IQA, but only eight specified the quality criteria used. The reported quality criteria varied widely across datasets, and accessing the information was often challenging. The findings highlight the importance of IQA for AI model development while emphasizing the need for clear and accessible reporting of IQA information. The study suggests that automated quality assessments can be a valid alternative to manual labeling and emphasizes the importance of establishing quality standards based on population characteristics, clinical use, and research purposes. In conclusion, image quality assessment is important for AI model development; however, strict data quality standards must not limit data sharing. Given the importance of IQA for developing, validating, and implementing deep learning (DL) algorithms, it’s recommended that this information be reported in a clear, specific, and accessible way whenever possible. Automated quality assessments are a valid alternative to the traditional manual labeling process, and quality standards should be determined according to population characteristics, clinical use, and research purpose.

Introduction

Diabetic retinopathy (DR) is a microvascular complication of diabetes and a leading cause of severe visual loss worldwide [1, 2]. Aiming at the early diagnosis and treatment of this condition, screening programs have been established to routinely analyze retinal photographs and select the cases that need specialized evaluation [3–5]. Recently, artificial intelligence (AI) and deep learning (DL) algorithms have been reported to achieve robust performance in detecting DR from retinal photographs, representing a promising tool to manage the large amount of image data generated by screening programs [2, 6–9]. Despite recent advances in this technology, many factors should be taken into account when developing and deploying an AI system, such as the number of included images [9, 10], patient’s characteristics [10], labels [11–13], and image quality [9, 14, 15].

Several failures in the clinical deployment of AI are a consequence of inadequate model development, the inclusion of only homogeneous populations, and non-diverse data [7, 16]. In ophthalmology models, using exclusively high-quality images can produce higher pre-deployment accuracy results; however, these algorithms fail when they are deployed in real-world settings [17, 18]. Datasets with more diverse demographics, diseases, and image quality are crucial for real-world simulation [16].

During the exam-capturing process, different measures can be taken to deal with low-quality images, including image reacquisition and pharmacological mydriasis [19, 20]. A manual or automatic quality assessment in databases can be performed by addressing image parameters such as focus, image field, artifacts, and illumination.

A quality assessment can lead to more generalizable algorithms regardless of the approach chosen to handle poor-quality images. Nevertheless, there are some obstacles involved in the image quality assessment (IQA): 1) the rating of image quality is subjective, and intergrader variability can be high, especially for images in the middle of the quality scale [21, 22] 2) the criteria for the evaluation of the image quality vary according to the type of diagnosis being made since different diseases have different image features and regions of interest [21].

Since quality control of retinal images influences model performance, fairness, and generalizability of DL algorithms, this paper reviewed publicly available DR datasets’ image quality criteria and assessment.

Methods

In this study, we evaluate the IQA and the quality criteria employed across publicly available datasets. This review compared the datasets employed in PubMed/MEDLINE Artificial Intelligence and Ophthalmology articles, using the search strategy: ((“dataset” [tiab] OR “database” [tiab]) AND (“publicly available” [tiab] OR “free of charge” [tiab] OR “freely accessible” [tiab] OR “publicly accessible” [tiab]) AND (“eye”[tiab] OR “ophthalmology”[tiab] OR “ophthalmology”[MeSH] OR “retina”[tiab])) AND ((“2000”[Date - Publication]: “2021”[Date - Publication])).

From the final cohort of articles, two authors (MBG and LFN) manually reviewed the applied dataset’s websites and any dataset-related documentation/publication available at these websites. These sources were considered “direct sources” of information. For incorporated datasets originating from DR screening programs, we also reviewed the publications describing the programs if the references of these publications were easily accessible (i.e., described on the websites or dataset documentation/publication). These sources were considered “indirect sources” of information. To enhance the article assessment quality, the reviewers selected were retina specialists and experts in ophthalmology and AI research. In order to minimize subjective bias, each reviewer independently evaluated every article. In discordance cases, the reviewers reached a consensus after discussion.

We included in the review publicly available color retinal fundus photos datasets with diabetic retinopathy patients. Datasets containing exclusively other ocular diseases (e.g., glaucoma or age-related macular degeneration) or imaging modalities (e.g., optical coherence tomography and fluorescein angiogram) and non-available images were excluded since image quality criteria may vary across different diseases or imaging modalities.

We reviewed direct and indirect sources by focusing on the following questions: Has any IQA been described, and if so, have the image quality criteria been specified?

Results

The search strategy yielded 130 articles. The first screening included only human and English, Spanish, and Portuguese articles, and the second consisted of title and abstract assessment. After the screening process, 39 articles were eligible for full-text assessment.

We retrieved the study’s included databases from the selected articles, excluding non-available datasets (Fig. 1).

Fig. 1. Flowchart diagram.

Process of articles inclusion and assessment.

From the reviewed articles, we included in this quality assessment analysis the EyePACS, DIARETDB0 and DIARETDB1, E-Ophtha, DRIVE, MESSIDOR-2, IDRiD, ROC, DR1 and DR2, REVIEW, HEI-MED, APTOS, DERIVA, DRIMDB, ROTTERDAM, and ONHSD. From the analysis of the references, we included the SUSTech-SYSU, DDR, and DRiDB datasets.

Among the datasets included in this study, image quality assessment was described in 12 (60%) datasets; however, the specific image quality criteria were specified in only 8 (40%) databases (Fig. 2, Table 1).

Fig. 2. Datasets assessment.

Image quality assessment across 20 publicly available DR datasets.

Table 1.

Image quality criteria across publicly available DR datasets.

Public dataset	Number of images	IQA criteria
Eye Picture Archive Communication System (EyePACS)	88,702	Present (direct source)
Standard Diabetic Retinopathy Database Calibration Level 0 (DIARETDB0)	130	Absent
Standard Diabetic Retinopathy Database Calibration Level 1 (DIARETDB1)	89	Absent
E-Ophtha	463	Present (indirect source)
Digital Retinal Images for Vessel Extraction (DRIVE)	40	Absent
Methods to Evaluate Segmentation and Indexing Techniques in the Field of Retinal Ophthalmology (MESSIDOR-2)	1748	Absent
Indian Diabetic Retinopathy Image Dataset (IDRiD)	516	Absent
Retina Online Challenge (ROC)	100	Present (indirect source)
DR1	1077	Present
DR2	520	Present
DDR	13,673	Present (direct source)
Diabetic Retinopathy Image Dataset (DRiDB)	50	Absent
Retinal Vessel Image Set for Estimation of Width (REVIEW)	16	Absent
The Hamilton Eye Institute Macular Edema Dataset (HEI-MED)	169	Absent
Asia Pacific Tele-Ophthalmology Society(APTOS)	5590	Absent
Digital Extraction from Retinal Images of Veins and Arteries (DERIVA)	50	Present (indirect source)
Diabetic Retinopathy Image Dataset (DRIMDB)	216	Absent
Rotterdam Ophthalmic Data Repository	1120	Absent
SUSTech-SYSU	1219	Present (direct source)
Optic Nerve Head Segmentation Dataset (ONHSD)	99	Absent

Direct source: sources directly related to the dataset (i.e., dataset website, documentation or any publication about the dataset). Indirect source: sources indirectly related to the dataset (i.e., publications describing the screening programs in which the images were acquired).

Apply quality classification criteria

EyePACS (Eye Picture Archive Communication System)

EyePACS is a publicly available fundus image dataset* that contains 88,702 CFP. The images were provided by the Eye Picture Archive Communication System (EyePACS), which is a free platform for retinopathy screening and made available by the California Healthcare Foundation [23].

In addition to DR classification, the EyePACS grading protocol [24] also includes criteria for IQA. Image quality factors considered are focus, illumination, image field definition, and artefacts. After the grading of these items, the image quality is classified in one of the following categories: excellent, good, adequate, insufficient for full interpretation, insufficient for any interpretation, or other. According to the grading protocol, the grader should specify the nature of the image quality problems. Details of the quality classification are described in Tables 2 and 3.

Table 2.

EyePACS quality factors for IQA.

Quality factors	Description
Focus	Is the focus good enough to perform adequate grading of the smaller retinal lesions such as microaneurysms and intraretinal microvascular abnormalities?
Illumination	Is the illumination adequate (not too dark, not too light)? Are there dark areas or washed-out areas that interfere with detailed grading?
Image field definition	Does the primary field include the entire optic nerve head and macula? Are the nasal and temporal fields adequately centered to include at least 80% of the non-overlapping portion of the field (i.e., the nasal portion of the nasal field, temporal portion of the temporal field)?
Artifacts	Are the images sufficiently free of artifacts (such as dust spots, arc defects, and eyelash images) to allow adequate grading?

Table 3.

Description of retinal image quality classification according to EyePACS.

Quality grade	Description
Excellent	There were no problems in any of the image quality factors listed above, and all of the detailed retinopathy questions were gradable.
Good	There were some problems in one or two of the image quality factors, and all of the detailed retinopathy questions were gradable.
Adequate	There were some problems in three or four of the image quality factors listed above, and all of the detailed retinopathy questions were gradable
Insufficient for full interpretation	One or more of the detailed retinopathy questions were marked “cannot grade,” but some of the questions were gradable.
Insufficient for any interpretation	None of the question were gradable
Other	Some other factor(s) in the quality of the images interfered with the grading.

E-Ophtha

E-ophtha is a database designed for scientific research in DR and contains 463 CFP annotated for exudates and microaneurysms. It has been generated from the OPHDIAT Tele-medical network for DR screening, in the framework of the ANR-TECSAN-TELEOPHTA project funded by the French Research Agency (ANR) [25]. In the OPHDIAT screening program, a photograph was considered to be of acceptable quality if (1) the center of the fovea and the retinal vessels were clearly visible and (2) more than two-thirds of the image could be assessed [26].

ROC (Retina Online Challenge)

This dataset consists of 100 CFP, which were used in the Retinopathy Online Challenge (ROC), a multi-year competition for various aspects of DR, with a focus on microaneurysm detection [27]. ROC images were acquired as part of the EyeCheck project, a web-based screening program for Diabetic Retinopathy that started in 2000 in the Netherlands. In this screening program, gradability criteria included focus of the large vessels, visibility of at least one arteriole or capillary in both the upper and the lower half of the image, evenness of illumination, and absence of artefacts [28]. Only gradable images were included in the ROC dataset [27].

DR1 and DR2

DR1 and DR2 are two publicly available datasets containing, respectively, 1077 and 520 CFP. The images were provided by the Department of Ophthalmology, Federal University of Sao Paulo (UNIFESP), and graded according to the presence of DR lesions and the need for referral [29, 30]. These datasets were manually annotated for image quality according to criteria such as the image field and the presence of blurring [31].

HEI-MED (The Hamilton Eye Institute Macular Oedema Dataset)

This dataset consists of 169 CFP, collected as part of a telemedicine network for the diagnosis of DR developed by the Hamilton Eye Institute, the Image Science and Machine Vision Group at ORNL with the collaboration of the Université de Bourgogne. The dataset was created to train, and test algorithms for the detection of exudates and diabetic macular oedema and the images were manually annotated for the presence of exudation areas.

Instead of descriptive quality criteria, an automated quality assessment technique was applied to this dataset. During the capturing process, quality metrics for the acquired images, which varied between 0 and 1, were displayed on the operator’s camera, allowing them a chance to take a new image if necessary. According to the authors, all 169 images were of sufficient quality [32, 33].

DERIVA (Digital Extraction from Retinal Images of Veins and Arteries)

This dataset was created to evaluate an automated method for structural mapping of retinal vessels and consists of 50 CFP, acquired as part of the EyeCheck project. The gradability criteria employed by this screening program was previously described for the ROC dataset [28, 34].

DRIMDB (Diabetic Retinopathy Image Database)

This dataset consists of 216 CFP and was created to evaluate the performance of an IQA approach. All images were classified into three classes by the same expert. Non-retinal images that could have been obtained for several reasons (e.g., wrong focus or patient absence) were graded as outliers. The remaining images were divided into two groups: good and bad-quality images, but the quality criteria for each class are not provided [35].

SUSTech-SYSU

This dataset contains 1219 CFPs collected from the Department of Ophthalmology, Gaoyao People’s Hospital and Zhongshan Ophthalmic Center, Sun Yat-sen University. The database was created to help in the development and validation of exudate detection algorithms and DR classification algorithms. In addition to exudate annotations, DR classification and bounding boxes for the optic disc and fovea location are also provided. The dataset underwent a quality control stage, and images that were “too blurry” or “of extremely large-area lesions” were excluded [36].

DDR

DDR is a publicly available dataset containing 13,673 CFP, collected from 147 hospitals in China. The images have image-level annotations of DR severity, pixel-level annotations of DR lesions, and bounding-box annotations automatically generated from pixel-level annotations.

In this dataset, the image acquisition was performed taking into account parameters such as optic disc and fovea location, focus, and exposure. The detailed principles are described in Table 4. All images were graded according to DR severity and image quality and were divided into six classes: no DR, mild nonproliferative DR, moderate nonproliferative DR, severe nonproliferative DR, proliferative DR, and ungradable. Images “with a blurring degree of more than 70% and without clearly visible lesions” were considered ungradable [37].

Table 4.

Principles applied during the acquisition of images of the DDR dataset.

Principle	Description
1	The center of a single fundus image should be located at the middle point of the line connecting the optic disc and the fovea. The central fovea or optic disc center should be horizontal in the single fundus image.
2	Fundus photography requires accurate focus. In healthy humans, the upper disc surface, retinal main blood vessels, retinal nerve fiber layer, macular fovea and other structures should be clearly identifiable. Additionally, lesions should be clear and legible for patients with DR.
3	The fundus image should be moderately exposed. The interface between the optic disc and the optic cup, the small blood vessels on the surface of the optic disc, and the normal retinal nerve fiber layer are clearly distinguishable.
4	When taking a fundus photograph, the patient should have their eye opened wide and should not blink. Hair and eyelashes are not allowed in the field at the moment of capturing the image.

DIARETDB0 and DIARETDB1 (Standard Diabetic Retinopathy Database Calibration Level 0 and 1)

DIARETDB0 and DIARETDB1 are two publicly available datasets created to evaluate methods for automatic detection of DR. These databases contain, respectively, 130 and 89 CFP taken in the Kuopio University Hospital, which were annotated for DR retinal findings, such as exudates, hemorrhages, and microaneurysms. An automated method for evaluating the partial image quality was applied to the DIARETDB0 dataset. A binary mask for every fundus image is provided, representing whether a certain location is applicable for the analysis or not. Information on the IQA is not provided for the DIARETDB1 dataset [38–40].

IDRiD (Indian Diabetic Retinopathy Image Dataset)

IDRiD is a dataset containing 516 CFP, which is available as a part of “Diabetic Retinopathy: Segmentation and Grading Challenge,” organized in conjunction with IEEE International Symposium on Biomedical Imaging (ISBI-2018), Washington D.C. The dataset provides 1) pixel-level annotations for DR lesions and the optic disc, 2) image-level severity grading of DR and macular oedema, and 3) optic disc and fovea center location [41].

According to the authors, “experts verified that all images are of adequate quality” [41], and the dataset “did not include non-gradable images” [42], but further details about IQA criteria are not provided.

APTOS (Asia Pacific Tele-Ophthalmology Society)

This dataset, released by the Asia Pacific Tele-Ophthalmology Society, contains 5590 publicly available CFPs. The authors mention that the images in this dataset may have artefacts, be out of focus, underexposed, or overexposed, but information on the IQA is not provided. **

Don’t apply quality classification criteria

DRIVE (Digital Retinal Images for Vessel Extraction)

This dataset was established to enable comparative studies on segmentation of blood vessels in retinal images. It consists of 40 CFPs, taken from a DR screening program in the Netherlands, that contain manual segmentation of the retinal vasculature. Information on the IQA is not provided on direct sources. Further details about the screening program are not provided, and thus indirect sources couldn’t be assessed [43].

MESSIDOR-2 (Methods to Evaluate Segmentation and Indexing Techniques in the Field of Retinal Ophthalmology)

The messidor-2 dataset consists of 1748 CFP and was established after the merge of the “Messidor-Original” database, created to facilitate studies on the computer-assisted diagnosis of diabetic retinopathy, with the “Messidor-Extension” database. Information on the IQA is not provided for this dataset [44].

REVIEW (Retinal Vessel Image Set for Estimation of Width)

This dataset consists of 16 CFP, and it was made available online by the Department of Computing and Informatics at the University of Lincoln, Lincoln, UK. The dataset was created for the evaluation of retinal vessel diameter measurement algorithms and contains images with manually annotated vessel edges. Information on the IQA is not provided for this dataset [45].

ROTTERDAM (Rotterdam Ophthalmic Data Repository)

This dataset includes 1120 CFP captured as part of the DR screening program of the Rotterdam Eye Hospital in the Netherlands and was created to evaluate the accuracy of fundus image registration methods. Information on the IQA is not provided for this dataset [46].

ONHSD (Optic Nerve Head Segmentation Dataset)

This database contains 99 CFPs taken from patients attending the diabetic screening program at City Hospital - Birmingham. The dataset was created to evaluate an algorithm for the localization and segmentation of the optic nerve head. The authors mention that there is considerable quality variation in the images, but information on the IQA is not provided [47].

DRiDB (Diabetic Retinopathy Image Database)

This dataset consists of 50 CFPs taken in a university hospital in Zagreb. Experts were asked to annotate areas containing DR findings and mark the blood vessels, optic disc, and macula. The authors mention that the images in this dataset contain a varying amount of noise, but information on the IQA is not provided [48].

Discussion

Image quality assessment is important for AI model development; however, strict data quality standards must not limit dataset development. The focus on quality may have the unintended consequence of discouraging less-funded institutions from contributing data because of concerns that their images are not of sufficient quality. In this work, we review 20 publicly available DR datasets and the described image quality assessment of retinal fundus photographs. Our findings show that 12 (60%) datasets mention having performed some quality control on the images, and 8 (40%) datasets specified the quality criteria employed in the IQA. Moreover, there is a lack of consensus on the criteria of image quality, and this information is often not easily accessible (i.e., available on indirect sources). Surprisingly, even on the DRIMDB, a dataset developed for quality assessment purposes, the criteria employed to classify the images as “good” or “bad” quality are not provided [35].

Public datasets and collaborative research

Sharing more diverse and representative data is needed in order to promote better AI development. Publicly available datasets represent a valuable alternative to deal with the high costs and the legal barriers often involved with data access [10, 44]. Data sharing is essential to promote collaborative research, bias assessment, reproducibility, and generalizability [49].

Our results show that, among publicly available DR datasets, the information on the image quality assessment is often missing, and some datasets assess image quality, but the quality criteria are not published. This is the case for IDRiD, which included only gradable images, and DRIMDB, which included good and bad-quality images, and provides the respective quality labels. For the databases that mention the image quality criteria, this information was often vague (SUSTech-SYSU, DR1, and DR2 datasets) or difficult to access (E-Ophtha, ROC, and DERIVA datasets). EyePACS provides a detailed grading protocol and specific quality criteria, but it’s not clear if the images with insufficient quality were excluded from the dataset.

Importance of image quality assessment

The image quality should not be the main consideration for dataset creation and sharing; nevertheless, IQA plays a role in the development of machine learning models [50].

When discussing IQA, it’s important to differentiate “image quality” from “image gradability,” terms often used interchangeably in the literature. While “quality” is overall image-based, gradability is regarding a specific disease grading. According to IRIS Reading Center, a teleophthalmology group, gradability considers clinical and anatomical features that need to be present so that the diagnosis of a specific disease can be made. Therefore, high-quality images are not always gradable. For example, a retinal image can have good focus and adequate illumination (high quality) but may be considerably ungradable for diabetic retinopathy if the center of the retina (i.e., macula) is not included in the image [51]. In the EyePACs’ IQA, the criteria “illumination” is image-based, while the criteria “image field definition” takes into account anatomical features like the optic disc and the macula.

Automated techniques can be applied to determine the quality of the retinal images, which was the case for the DIARETDB0 and the HEI-MED datasets. This approach, also known as objective IQA, can be time and cost-saving, especially when dealing with large amounts of image data [52]. Despite the potential benefits of the objective IQA, there are some challenges that limit advances in this research field. Currently, IQA algorithms are developed and tested on specific fundus image databases, and the lack of a benchmark dataset for quality assessment purposes makes it difficult to compare the performance of different models [52–54]. Therefore, datasets containing information about the quality assessment, including the image quality criteria, are crucial for the development and evaluation of automated diagnostic models and improvements in automated IQA techniques.

Image quality criteria standardization

When reported, the image quality criteria vary widely across databases. While in the EyePACS dataset, the images were graded according to four criteria and classified into six quality classes, other datasets employed fewer criteria and quality classes. The definitions for each criteria also vary across datasets. In the EyePACS dataset, the focus of an image was considered good enough if small lesions such as microaneurysms and intraretinal microvascular abnormalities could be properly assessed. On the other hand, in the E-Ophta dataset, the image was considered of acceptable quality if the center of the fovea and the retinal vessels were clearly visible.

The standardization of image quality criteria is challenging across different datasets and for different facilities. Further research with manual or automated quality standards and model performance analysis according to various specifications is needed.

Automatic quality assessment

To address the lack of standards and laborious processes in quality assessments, AI algorithms have been developed and are promising to automatically qualify an image’s gradability. The deep-learning AutoMorph is a publicly available pipeline that uses retinal vasculature morphology on fundus photographs through image preprocessing, quality grading, anatomical segmentation, and morphological feature measurement [55]. The open-source Nderitu et al. algorithm detects laterality, retinal field, retinal presence, and gradability [56].

Recommendations

We recommend diversity in datasets, including image quality, for fairness and generalizability in AI development. Data quality is important during model development; however, it should not be a deterrent to dataset creation and sharing. Retinal fundus photos quality standardization is limited, and automated processes for quality assessment are a viable alternative.

Given the importance of quality assessment when applied in fundus photos and machine learning in the DR context, we propose that, whenever possible, information on the IQA and specific grading used for labeling be reported together with other metadata.

In algorithms, we recommend the report of imaging quality criteria definitions and the parameter to include or exclude images for the model development, to avoid and address biases. The image quality criteria should be clear and as specific as possible, which contributes to greater reproducibility of the methodology.

Implications and limitations

This is the first study to review publicly available DR datasets with a focus on the quality assessment and the quality criteria for retinal images. With this review, we hope to reinforce the importance of IQA and the need to address image quality for better algorithms. Moreover, its significance extends beyond the evaluation of DR grading alone. It sheds light on the inherent risks of biases within AI and the intricate tradeoff between model performance metrics and the effective deployment of AI systems in real-world scenarios. We also hope this study can encourage similar research involving other diseases, imaging modalities, and medical specialties.

The Flexner Report is an example of the unintended consequence of applying strict quality standards for medical schools that led to the closure of almost all Black medical schools in America at the turn of the twentieth century [57]. Quality assessment is important in imaging exams, but the construct of quality is manifold. The collection technique and the equipment characteristics, among others, contribute to the image quality. More studies are needed to address quality assessment standards, which should vary according to the research objectives, clinical use, and deployment setting.

This study has three main limitations. Firstly, there were only 20 datasets included in the review, and the study was also limited to a specific ocular disease and imaging modality. Secondly, the search was restricted to sources (e.g., websites, documentation, and publications) related to the datasets and/or screening programs. Third, we focused on the dataset’s quality assessment report, with further studies evaluating models needed.

Conclusion

Image quality information is crucial not only for the development and validation of diagnostic DL algorithms but also for the improvement of automated IQA techniques. Publicly available DR datasets are valuable research tools. Feedback during the data curation, modeling, and deployment contributes to improvement in data quality assessment. Assessing image quality during model development in a clear and transparent way can improve the value of machine learning models, decrease biases, and contribute to advances in this research field.

Author contributions

MBG: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Writing-original draft, Writing-review & editing. LFN: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Writing-original draft, Writing-review & editing. DF: Writing-review & editing. HF: Writing-review & editing. EK: Writing-review & editing. FKM: Writing-review & editing. CVSR: Writing-review & editing. MM: Writing-review & editing. LAC: Conceptualization, Methodology, Supervision, Writing-review & editing. PK: Supervision, Writing-review & editing. RBJ: Conceptualization, Methodology, Supervision, Writing-review & editing

Competing interests

PAK has acted as a consultant for Google, DeepMind, Roche, Novartis, Apellis, and BitFount and is an equity owner in Big Picture Medical. He has received speaker fees from Heidelberg Engineering, Topcon, Allergan, and Bayer. EK: Sanro Health: Owner, Alimera sciences: Consultant, Genentech: consultant (in past six months, not current). The other authors declare no competing interests.