Applications of deep learning in detection of glaucoma: A systematic review

Author: Mirzania D, Thompson AC, Muir KW.

Geographical coverage: Not reported

Sector: Service Delivery

Sub-sector: Screening, diagnosis

Equity focus: Not reported

Study population: Patients with glaucoma

Review type: Effectiveness review

Quantitative synthesis method: Narrative synthesis

Qualitative synthesis method: Not applicable

Background

Glaucoma is a leading cause of irreversible blindness worldwide, projected to affect 111.8 million individuals by 2040. Early detection is critical, as more advanced visual field loss and higher intraocular pressure (IOP) at diagnosis increase the risk of vision loss. However, early glaucoma is often asymptomatic, posing challenges to timely diagnosis. Recent advances in artificial intelligence (AI), particularly deep learning (DL), offer promising tools. These systems autonomously learn features from imaging data and may improve diagnostic accuracy beyond traditional methods. Despite these developments, challenges remain, including the lack of a true reference standard and the variability in available datasets.

Objectives

To highlight original studies in which DL algorithms were applied to detect glaucoma and monitor its progression, and to discuss some of the challenges associated with DL algorithm development in glaucoma.

Main findings

Overall, deep learning (DL) algorithms demonstrate significant potential to improve the detection and diagnosis of glaucoma, potentially transforming clinical practice and patient outcomes.

The search identified 1,226 records, of which 29 studies were included. Most studies (n = 21) used colour fundus photography (CFP) to train and test DL models. Other modalities included optical coherence tomography (OCT, n = 5) and standard automated perimetry (SAP, n = 3).

Algorithm performance was primarily assessed using the area under the receiver operating characteristic curve (AUC), where values approaching 1.0 indicated strong discrimination between glaucomatous and non-glaucomatous eyes. Eight studies were excluded due to high risk of bias (n = 4), incomplete outcome reporting (n = 3), or lack of clarity regarding training and testing datasets (n = 1).

Many DL models showed high diagnostic accuracy, with AUC values frequently above 0.90. The integration of fundus photography into telehealth systems presents an opportunity to support glaucoma screening in low-resource settings. DL algorithms could also enable automated diagnosis, enhancing accessibility and efficiency in clinical workflows.

Methodology

A literature search was conducted in PubMed, Embase, and Web of Science on 24 April 2020 to identify studies using CFP, OCT, or SAP that reported the development or training of novel or existing convolutional neural networks (CNNs) for the identification of glaucomatous eyes.

Selected studies were reviewed in detail, and relevant data were extracted. Each study was evaluated for risk of bias to ensure reliable conclusions. Due to heterogeneity in study design, the findings were synthesised narratively.

Applicability / External validity

The review noted concerns regarding the generalisability of DL models in glaucoma detection. Many algorithms were trained and tested on datasets lacking ethnic and racial diversity, which may limit applicability to wider populations. Additionally, variability in imaging devices and data quality influenced model performance. The authors recommended the use of multi-ethnic datasets and validation in diverse clinical settings to enhance external validity.

Geographic focus

The geographic location of the included studies was not reported.

Summary of quality assessment

Confidence in the review’s conclusions is low. The authors limited study inclusion to English-language publications and did not indicate whether multiple reviewers were involved in screening or data extraction. Reference lists of included studies were not checked, which may have excluded relevant literature.