Automatic detection of glaucoma via fundus imaging and artificial intelligence: A review

1. Introduction

• An estimated 112 million people will have glaucoma by 2040 (76); although irreversible, early diagnosis can enable treatment to slow or prevent progression and blindness.Source: §1 · p. 3
• In the UK, around half of community glaucoma patients remain undiagnosed (12), while glaucoma accounts for 25% of hospital eye-service appointments and costs the NHS upwards of £75m yearly (7).Source: §1 · p. 3

The detection of glaucoma by a clinician

• Clinicians assess optic nerve head changes by tracing cup and disc contours, which support diagnosis explanations and patient participation in care decisions.Source: The detection of glaucoma by a clinician · p. 3
• Fundus cameras are relatively low-cost and can show optic nerve head abnormalities before visual-field loss, making portable imaging useful for screening in low-resource settings.Source: The detection of glaucoma by a clinician · p. 4

Detecting glaucoma via AI

• AI-enabled fundus screening could reduce observer error, support large-scale low-cost screening in remote areas, and help refine referrals and reduce false positives in the NHS.Source: Detecting glaucoma via AI · p. 4
• The state of AI frameworks using fundus cameras while providing optic cup and disc contours remains unclear, including their accuracy, suitability, and training, testing, and validation.Source: Detecting glaucoma via AI · p. 4

Overview

• The review aims to clarify AI terminology, summarize the state of segmented-fundus detection methods and their two approaches, and discuss progress and areas needing further work.Source: Overview · p. 5

2.1. Cup to disc ratio

• CDR, especially vCDR and ACDR, is a standard ONH measure of glaucomatous neuropathy, but can mislead with unusual disc or cup size or myopic changes; RDR and hCDR are alternatives (52,14; 33).ACDR provides a two-dimensional measure for assessing structural ONH changes.Source: §2.1 · p. 5

2.2. Neuroretinal rim area ratio

• The NRR is the area between OC and OD margins; healthy rim widths typically follow I>S>N>T (17), but healthy eyes may violate this rule (75), so it is a clinical tool, not a diagnostic test.Source: §2.2 · pp. 5–6

2.3. Disc Damage Likelihood Scale

• DDLS grades glaucoma progression in ten stages while accounting for OD size, improving inter-observer repeatability and agreement with the gold standard versus vertical CDR, but requires time and clinician training (14,71; 73; 14).Source: §2.3 · p. 6

3.1. Fundus image segmentation

• Image segmentation partitions an image into clinically meaningful regions; fundus segmentation finds the optic cup and disc boundaries, whose annotations form the segmented image.Manual segmentation uses expert annotation, while automatic segmentation uses mathematical algorithms, including level-set, threshold-based, and clustering-based approaches.Source: §3.1 · p. 6

3.2. Image features

• Clinically interpretable features have clinical meaning, such as vCDR or NRR area, whereas abstract features are mathematically or statistically derived and may lack clinical interpretability.Source: §3.2 · p. 6

3.3. Probability of glaucoma

• A glaucoma probability estimates the chance of disease for a new image, should reflect prevalence in the population of interest, and should express uncertainty when the image is poorly represented in training.A 99% estimate does not mean certainty; uncertainty estimation is a complex, active research area.Source: §3.3 · pp. 6–7

3.4. Image classification

• Image classification assigns a fundus image to disease categories, such as healthy or glaucomatous, often by thresholding estimated probability; uncertainty in that estimate carries into classification.Classification may be binary or multi-class, including a suspected-glaucoma category.Source: §3.4 · p. 7

3.5. Classifier

• A classifier is the mathematical, statistical, or machine-learning method that estimates a patient's disease category; examples include support vector machines and logistic regression.Source: §3.5 · p. 7

3.6. AI framework

• Two-step frameworks segment the image before classification, whereas one-step frameworks classify end-to-end; segmented outputs can aid interpretation, and explainability is crucial in medicine.One-step deep-learning methods can be computationally intensive and difficult to explain; two-step methods can provide interpretable or abstract features. Explainability need not come at the cost of accuracy.Source: §3.6 · pp. 7–8

3.7. Evaluating the performance of AI

• Performance should be assessed against expert ground truth using multiple measures; sensitivity matters because false negatives can delay glaucoma care, while AUROC measures discrimination across thresholds.Accuracy, sensitivity, specificity, positive predictive value, and negative predictive value capture different aspects of performance; no single measure is sufficient.Source: §3.7 · pp. 8–9
• Internal evaluation tests data from the training cohort, using a train-test split or cross-validation; external evaluation uses a different source to assess generalisability.In k-fold cross-validation, training and testing are repeated across partitions and the reported metric is their average; internal testing alone may not establish generalisability.Source: §3.7 · p. 9

• A comprehensive literature search identified papers for the review, and a table was used to extract relevant information.Source: §4 · p. 9
• Extracted study information included authors, publication year, classification approach, and data characteristics such as sample size, public availability, annotators, and imaging devices.Source: §4 · pp. 9–10
• The review focused on glaucoma classification after segmentation but also recorded segmentation techniques, validation methods, and accuracy, sensitivity, specificity, and AUROC because segmentation can heavily influence classification results.Source: §4 · p. 10

5.1 Papers included; 5.2 Characteristics of identified papers

• Of 1,018 identified papers, 28 met eligibility criteria; 11 used logical rules and 17 used machine-learning or statistical modelling frameworks.Most exclusions concerned one-step frameworks without segmentation, segmentation-only studies, or studies using other imaging modalities.Source: §5.1 · p. 10, §5.2 · p. 10

5.3 Logical rule-based AI frameworks for glaucoma detection from segmented images

• Dutta et al.’s vCDR-only detection rule achieved 90% accuracy on 10 images, with one misclassification: an image with vCDR 0.6 labeled healthy despite ophthalmologists’ glaucomatous ground truth (24).Ahmad et al. reported 92% accuracy, 93% sensitivity, and 88% specificity (3); Soorya et al. reported 97% accuracy, 96.5% sensitivity, and 98% specificity (50). Both tested on datasets from one source.Source: §5.3 · p. 11

5.4. Machine learning/statistical modelling – based AI frameworks for glaucoma detection from the segmented image

• SVMs were the most popular classifier, used in 9 of 17 papers; one LME framework achieved 99.7% AUROC, 100% sensitivity, and 98.3% specificity internally, then 91% AUROC on external RIM-ONE V3 validation.The FFBPNN and DTDNN each reported 100% accuracy, sensitivity, and specificity.Source: §5.4 · pp. 12–13
• Every reviewed machine-learning framework used a CDR variant; 10 papers used clinically interpretable features and 7 used novel spatial, spectral, texture, or colour features.Combining structural and textural features improved performance in one study, while another found vCDR plus an abstract texture-energy feature performed best.Source: §5.4 · p. 14

5.5. Approaches to segmentation

• Optic-cup segmentation is harder than optic-disc segmentation because of low contrast between the cup and neighbouring disc, and few papers focused on cup segmentation.Reviewed segmentation approaches ranged from colour-intensity and texture thresholding to fully convolutional neural networks.Source: §5.5 · pp. 14–15

5.6. Glaucoma disease groups

• Of the 28 papers, 24 (85%) used binary classification and 4 used three classes; one multi-class study reported precision of 0.76 and recall of 0.72.The four studies handled suspected glaucoma differently; the cited study’s multi-class performance was inferior to its binary results.Source: §5.6 · p. 15

5.7. Approaches to validation of methods and the reporting of performance metrics

• Nine of 11 rule-based papers used one dataset, while machine-learning papers varied in validation; only one used external validation and four used multiple databases.Six rule-based papers reported accuracy, sensitivity, and specificity; 14 machine-learning papers reported all three metrics.Source: §5.7 · pp. 15–16

5.8. Databases used for development and testing of the AI frameworks reviewed

• DRISHTI and HRF were each used seven times; DRISHTI contains 101 images, while HRF contains 45, including healthy, glaucomatous, and diabetic-retinopathy images.Messidor had 100 glaucoma-annotated images, and ORIGA contained 650 images.Source: §5.8.1 · p. 16

5.8.1. Publicly available datasets

• RIM-ONE V1, V2, and V3 contain 169, 455, and 159 images, respectively, with five, one, and two expert references for V1, V2, and V3.V1 images came from three Spanish hospitals; V3 images came from one hospital and were captured with a stereo fundus camera.Source: §5.8.1 · p. 17
• GlaucomaDB contains 120 images (85 healthy and 35 glaucomatous), with ground truths from two ophthalmologists, and requires author permission for research access.The images came from a larger local-hospital database; its region or country was not disclosed.Source: §5.8.1 · p. 17
• Only 50 of HEI MED’s 169 images were annotated for glaucoma detection: 30 healthy and 19 glaucomatous, with annotations from one ophthalmologist.The dataset was used by one framework and is available on GitHub.Source: §5.8.1 · p. 17
• DRIONS comprises 110 images (95 healthy and 15 glaucomatous), excludes cataract cases, includes only Caucasian subjects, and is available to download online.Its images were digitised from slides at 600 x 400 pixels.Source: §5.8.1 · p. 17
• DIARETDB H, developed primarily for diabetic retinopathy research, contains 89 images (81 healthy and 8 glaucomatous) with annotations from four medical experts and is publicly downloadable.Images were captured at one Finnish hospital using a 50-degree field-of-view camera with varying imaging settings.Source: §5.8.1 · pp. 17–18
• DRIVE contains 40 images (34 healthy and 6 glaucomatous) annotated by two ophthalmologists, but despite being described as public, it is not easily accessible online.Images were acquired with a Canon CR5 camera at 768 by 584 pixels.Source: §5.8.1 · p. 18

5.8.2. Private datasets

• Fifteen private databases were popular, but restricted access and sparse cohort or imaging details limited assessment of framework robustness and generalisability.Many papers reported little beyond dataset size, and some database information was difficult to obtain without contacting authors.Source: §5.8.2 · p. 18

6.1 Three key findings of this review

• The review covered 28 papers published from 2011 to March 2021, finding promising optic-cup/disc segmentation and two downstream approaches: logical rules and machine-learning/statistical models.Research activity grew from 2011 to 2019, then fewer papers appeared in 2020.Source: §6 · p. 18, §6.1 · p. 18
• Rule-based frameworks were limited by simplistic rules: ten internally validated single-dataset studies reported 83–97% accuracy, while one externally tested paper reported sensitivity/specificity pairs of 93/92% and 87/87%.Thresholds varied across the 11 papers, limiting understanding of performance across data sources; the internal-validation results should be interpreted cautiously.Source: §6.1 · pp. 18–19
• Machine-learning/statistical frameworks reported 85.1–100% accuracy, predominantly through internal validation; some results were comparable to one-step deep learning, but same-dataset comparisons remain necessary.Li et al. (46) reported an accuracy of 0.986.Source: §6.1 · p. 19, §6.2 · p. 20
• Two-step frameworks make intermediate decisions explainable through optic-cup/disc contours and may require 100 times fewer training data than deep learning.MacCormick et al. (49) used approximately 300 training images and achieved AUROC 99.6% in internal validation and 91.0% in external validation; Li et al. (46) used 30,000 images. The two-step approach can support clinician–patient decision-making (19).Source: §6.1 · p. 19

6.2 Three key unresolved issues of current knowledge and potential areas for future studies

• Future work should compare one- and two-step frameworks on the same datasets, compare OCT with fundus imaging, and assess uncertainty and additional patient data sources.One-step deep learning reduces segmentation and feature-engineering effort but is less interpretable and requires many annotated images; additional data could simulate clinical workflows and help explain AI outputs.Source: §6.2 · p. 20

6.2.2 There is a need to keep building and sharing suitable datasets

• Building suitable datasets requires addressing missing optic-nerve-head annotations and access barriers in landmark cohorts, while including glaucoma suspects and longitudinal data.The review names UKGTS (29), OHTS (32), and NICOLA (51) as unused resources; pixel-level annotations require expert effort, and dataset access may require applications, payment, and protocols.Source: §6.2.2 · pp. 20–21
• Datasets need fuller reporting, reliable multi-grader ground truth, and greater diversity in devices, patient ethnicities, and image conditions.Only 3 of 12 publicly available databases had more than two annotators; single-grader labels risk bias, while homogeneous datasets may impair generalisability.Source: §6.2.2 · p. 21
• High-quality reporting guidance should be promoted because missing dataset and classifier details—including model structure and hyperparameter tuning—hinder reproducibility.The review calls for support of EQUATOR, an initiative that provides reporting guidance for medical studies involving AI.Source: §6.2.2 · pp. 21–22

7. Method of Literature Search

• The literature search covered PubMed, Scopus, Web of Science, and Medline (OVID) through 2021.The search strategies are detailed in supplementary file 1.Source: §7 · p. 22

Search terms

• Search strategies combined terms for glaucoma, fundus or retinal images, classification or diagnosis, automated detection or prediction, and segmentation.The database-specific query syntax varied across the four search strategies.Source: Search terms · p. 22

Eligibility criteria

• Eligible papers classified glaucoma using segmented optic nerve head fundus images; segmentation-only and non-segmented classification studies were excluded.Included papers also required online full text in English.Source: Eligibility criteria · p. 22

• The British Council for Prevention of Blindness provided PhD studentship funding to Dr. Gabriela Czanner, Dr. Silvester Czanner, Professor Colin E. Willoughby, and Dr. Srinivasan Kavitha.Source: §8 · p. 22

A Vision impairment and blindness [Internet]. [cited 2021 Nov 14]. Available from: https://www.who.int/news-room/fact-sheets/detail/blindness-and-visual-impairment

• The cited material identifies NHS Digital’s *Hospital Outpatient Activity – 2014-15* report, cited on 14 December 2021.The entry provides a URL for the report; no findings from it are included in the supplied passage.Source: A Vision impairment and blindness… · p. 28

D Spiegelhalter D. Should We Trust Algorithms? Harv Data Sci Rev [Internet]. 2020 Jan 31 [cited 2021 Dec 17];2(1). Available from: https://hdsr.mitpress.mit.edu/pub/56lnenzj

• The D subsection passage lists Leslie’s guide to responsible AI design and implementation in the public sector [E].The entry is titled “Understanding artificial intelligence ethics and safety.”Source: D Spiegelhalter D. Should We Trust… · p. 28

F Reporting guidelines | The EQUATOR Network [Internet]. [cited 2021 Nov 14]. Available from: https://www.equator-network.org/reporting-guidelines/

• The F subsection passage lists the European Glaucoma Society’s 2020 fifth-edition terminology and guidelines [G].Source: F Reporting guidelines | The EQUATOR… · p. 28
• The F subsection passage also lists the DIARETDB1 standard diabetic retinopathy database, identified as calibration level 1 [H].Source: F Reporting guidelines | The EQUATOR… · p. 28