Image-Driven AI Power in Glaucoma

In this article, Dr Yvette Wang delves into major imaging modalities in glaucoma, and the current applications of artificial intelligence available to assist with effective screening and management of this sight-threatening disease.

As the leading cause of irreversible blindness worldwide, glaucoma affects approximately 95 million people, with open-angle glaucoma accounting for 65 million cases and angleclosure glaucoma for 30 million.¹ The global prevalence of glaucoma among individuals aged 40–80 years is 3.54%, and projections indicate that by 2040, the number of cases will rise to 111.8 million, disproportionately impacting populations in Africa and Asia.²

One of the biggest challenges with glaucoma is its insidious nature; it often goes undetected until significant and irreversible vision loss has occurred. Population-based surveys reveal that one in 40 adults over the age of 40 suffer from glaucoma-induced visual function loss.³

Unfortunately, no treatments currently exist to restore vision lost to glaucoma, making early detection and treatment crucial. Despite this, around 50% of glaucoma cases in developed countries remain undiagnosed, underscoring the urgent need for effective screening and intervention strategies.⁴

Diagnosing and monitoring glaucoma is a complex and costly process that relies heavily on the expertise of clinicians. Typically, a multimodal imaging approach is used, including colour fundus photography (CFP), visual field (VF) tests, and optical coherence tomography (OCT). Each of these techniques plays a vital role in identifying and monitoring the disease, from detecting optic nerve head changes with CFP to assessing retinal nerve fibre layer (RNFL) loss with OCT and evaluating VF defects.

In recent years, artificial intelligence (AI) algorithms have significantly advanced our understanding and management of many eye diseases, including glaucoma. One of AI’s most substantial benefits is its ability to process images rapidly and efficiently. AI can analyse vast numbers of ocular images in a short time, facilitating high-throughput screening and population-level analysis. By integrating AI with ocular images, this approach enhances glaucoma diagnosis, supports ophthalmologists in making swift clinical decisions, and improves glaucoma screening efforts (Figure 1).

COLOUR FUNDUS PHOTOGRAPHY

CFP is a cornerstone in glaucoma detection due to its low-cost, non-invasive, and quick imaging capabilities.⁵ These features make CFP particularly valuable in low-resource settings. Key indicators of glaucomatous optic neuropathy can be detected on CFP, such as cupping of the optic nerve head (ONH), asymmetry in the cup-to-disc ratio (CDR), and thinning of the neuro-retinal rim⁶ (Figure 2). These structural abnormalities often appear before functional impairment, making early screening possible.⁷

Despite its advantages, traditional human interpretation of CFP faces challenges in glaucoma detection. The anatomy of the ONH can vary significantly among individuals, complicating the identification of glaucomatous damage. Studies have shown that even experts only moderately agree on the detection of ONH damage from fundus photographs.⁸ The variability in the ONH area, which can differ fivefold between individuals, further complicates the establishment of a definitive CDR threshold for pathological cupping.⁹

AI technology holds potential to transform the landscape of glaucoma detection using CFP by addressing these challenges. AI algorithms excel at segmenting and detecting the ONH, optic disc, and RNFL, enabling precise analysis of fundus photographs. Deep learning (DL) algorithms are particularly adept at identifying characteristic fundus lesions of glaucoma, such as RNFL defects, ONH injury, and optic disc haemorrhages, even for some subtle pathological changes that may be hard for humans to detect.

Early AI applications in glaucoma focussed on straightforward tasks, such as algorithms that quickly located the optic disc region.¹⁰

These algorithms were usually trained to detect glaucoma-like discs based on vertical CDR values of 0.7 or 0.8, defined through clinical experience.¹¹ Almost 20 studies have reported robust performance in identifying glaucomatous optic neuropathy and suspect glaucoma using CFP, with a high area under the curve (AUC) of around 0.95 for diagnostic accuracy, having comparable specificity to eye care providers.^11,12 Additionally, DL algorithms can predict the characteristics of other imaging modalities based on CFP. Some of them can predict objective and quantitative OCT measurements, such as RNFL thickness, and some can even forecast the future development of visual field defects in glaucoma suspects.^13,14

OPTICAL COHERENCE TOMOGRAPHY

Introduced over 30 years ago, OCT is a noninvasive imaging technique that provides detailed cross-sectional views of the ONH and retina.¹⁵ It plays a crucial role in diagnosing and monitoring glaucoma by identifying morphological changes, particularly in the RNFL. Recent advancements in OCT technology have significantly enhanced its capabilities, enabling faster and more precise imaging. Anterior segment OCT (AS-OCT) assesses the anterior chamber and iris, while posterior OCT offers detailed and reproducible measurements of ONH parameters, including peripapillary RNFL thickness, macular parameters, optic disc parameters, and ONH cube scans.^16,17

AI integration with OCT has revolutionised the detection and management of glaucoma. Different types of OCT images, including 2D B-scans, 3D volumetric scans, AS-OCT, and OCT-angiography (OCTA) images, have been used to develop AI algorithms for glaucoma diagnosis. These AI models have demonstrated high diagnostic accuracy, with AUC scores ranging from 0.78 to 0.99.¹⁸

These scores highlight the effectiveness of AI in distinguishing glaucomatous eyes from normal eyes and in predicting RNFL thickness and different glaucoma stages.

In clinical settings, OCT machines automatically segment structures of interest to generate quantitative measures, such as RNFL thickness. Traditional machine learning methods use these segmentationbased features to analyse OCT data and identify glaucoma. However, DL models have surpassed traditional methods by analysing raw OCT data without the need for segmentation. This approach reduces segmentation errors, improving accuracy while decreasing manual workload.¹⁹

Advanced DL algorithms, including 3D models, have further enhanced glaucoma detection.²⁰ For instance, 3D DL algorithms combined with OCT scans can more accurately detect glaucomatous ONH degeneration. These models achieve high diagnostic performance, with AUC scores as high as 0.94, by leveraging detailed volumetric OCT scans of the ONH.²¹ AI models using OCTA can identify key features for glaucoma, such as inferior temporal vessel density, achieving 83.9% accuracy in glaucoma detection.²²

In addition, AI systems can accurately detect major features in AS-OCT, such as anterior chamber angles. They can also process image data and investigate the relationship between dynamic iris changes and primary angle-closure glaucoma. By automatically extracting features, such as thick peripheral iris roll, plateau iris, and expanded lens vault in the anterior segment, AI can reduce errors caused by manual recognition and segmentation.²³ This capability allows AI to simulate static and dynamic goniometry at a level comparable to that of a clinician.²⁴

While OCT is a more expensive technology compared to CFP, its precision and detailed structural insights make it invaluable for glaucoma detection. AI enhances OCT’s efficiency by automating and refining the analysis of ocular features, making it a complementary tool that boosts both diagnostic accuracy and labour efficiency.

VISUAL FIELD ANALYSIS

VF tests are essential for mapping a glaucoma patient’s field of vision, identifying deficits characteristic of glaucoma. These tests, particularly standard automated perimetry, are widely regarded as the gold standard for diagnosing glaucoma, tracking disease progression, and guiding treatment decisions. However, VF tests can be influenced by subjective factors, such as patient attention and cooperation, potentially impacting the accuracy of disease progression assessments.

The need for more efficient and accurate data processing has grown alongside the volume of diagnostic data. DL models have been developed to analyse VFs collected from various healthcare facilities, using total deviation plots, mean deviation values, and pattern deviation probability plots. These models are trained to detect patterns that may not be immediately apparent to clinicians.

Numerous AI algorithms exist to analyse VF, by recognising loss patterns relevant to the clinical diagnosis of glaucoma (such as arcuate defects) and detecting VF progression.²⁵ At distinguishing glaucomatous from non-glaucomatous VFs, DL was significantly more accurate than ophthalmology residents, attending ophthalmologists, and glaucoma experts.²⁶ This may be due to DL algorithms being able to detect patterns between adjacent and distant test points that clinicians might overlook.

In detecting VF progression, DL algorithms exceeded the performance of traditional metrics (e.g., mean deviation slope, permutation of pointwise linear regression, Collaborative Initial Glaucoma Treatment Study scoring, and Advanced Glaucoma Intervention Study scoring).²⁷ Researchers further classified central VF patterns in glaucoma, indicating that specific subtypes with nasal defects were linked to more severe total central loss in the future.²⁸

Future opportunities for AI in visual field testing include training neural networks to identify optic discs associated with manifest VF loss across different disc sizes, aligning with current treatment strategies aimed at slowing disease progression. As AI continues to advance, it holds the potential to enhance the accuracy and efficiency of glaucoma diagnosis and monitoring, ultimately improving patient outcomes.

INSIGHTS TOWARDS CHALLENGES AND THE FUTURE

Unlike AI for diabetic retinopathy, no AI product for glaucoma has yet received United States Food and Drug Administration (FDA) approval. One major challenge is the lack of a universally accepted gold standard for defining the presence and progression of glaucoma. Variations in ONH characteristics across different populations make it difficult to standardise the CDR threshold, complicating the training and validation of AI models. Moreover, AI models need to be tested with data from real-world clinical practice, not just controlled studies, as they must account for variations such as high myopia and other optic neuropathies that can mimic glaucoma.

To address these challenges, the field is fostering innovations in AI tools. Global competitions, like the Retinal Fundus Glaucoma Challenge (REFUGE), organised by the Medical Image Computing and Computer- Assisted Intervention (MICCAI) Society, set new standards for evaluating AI models in glaucoma diagnosis.²⁹ Such competitions attract teams worldwide, promoting the development of robust AI algorithms.

Effective AI approaches for glaucoma classification in the future may need to integrate multimodal data, combining both structural and functional information. For instance, combining VF data with OCT scans results in superior diagnostic performance, similar to the comprehensive approach used by eye specialists.³⁰ Collecting comprehensive clinical data from multiple perspectives enhances AI model performance and reliability, facilitating earlier detection and better understanding of disease trends.

A critical aspect of integrating AI into clinical practice is interpretability. In glaucoma detection, AI models can highlight significant areas in fundus images or OCT scans, helping clinicians understand AI decisions and validate their accuracy (Figure 3). Incorporating interpretability frameworks into AI systems is crucial for their acceptance and effective utilisation in clinical settings.

AI technologies can detect subtle changes and patterns that may not be immediately apparent to human observers, providing a higher level of diagnostic precision. As AI continues to evolve, it holds the promise of revolutionising glaucoma care, making early detection more accessible, diagnostics more accurate, and treatment strategies more personalised. The symphony between AI algorithms and clinical expertise will pave the way for a new era in the management of this vision-threatening disease, ultimately safeguarding the sight of millions worldwide.

Dr Yvette Wang is an ophthalmologist and the current senior clinical advisor at Optain Health. She holds a PhD from the Zhongshan Ophthalmic Center, China’s leading institution for ophthalmology.

Dr Wang’s primary research focus revolves around the integration of artificial intelligence in ophthalmology and the practical implementation of AI-driven medical devices. Her scholarly endeavours have yielded multiple academic publications in prestigious journals such as Nature npj Digital Medicine, Stroke, and Journal of Medical Internet Research, amassing nearly 400 citations. Furthermore, she has delivered both poster and oral presentations at international conferences.

Over five years in clinical practice, Dr Wang has honed her expertise in the diagnosis and treatment of prevalent eye diseases, including cataract, glaucoma, retinopathy, and ocular trauma. Her pioneering work in this area underscores her hope and commitment to leveraging cutting-edge technology for the betterment of patient outcomes.

This article was sponsored by Optain.

References

1. Jayaram H, Kolko M, Friedman DS, Gazzard G, Glaucoma: Now and beyond. Lancet 2023; 402(10414): 1788-801. doi: 10.1016/S0140-6736(23)01289-8.
2. Tham YC, Li X, Cheng CY, et al., Global prevalence of glaucoma and projections of glaucoma burden through 2040: A systematic review and meta-analysis. Ophthalmology 2014; 121(11): 2081-90. doi: 10.1016/j. ophtha.2014.05.013.
3. Quigley HA, Glaucoma. Lancet 2011; 377(9774): 1367- 77. doi: 10.1016/S0140-6736(10)61423-7.
4. Heijl A, Bengtsson B, Oskarsdottir SE, Prevalence and severity of undetected manifest glaucoma: Results from the early manifest glaucoma trial screening. Ophthalmology 2013; 120(8): 1541-5. doi: 10.1016/j.ophtha.2013.01.043.
5. Zhu Y, Salowe R, O’Brien JM, et al., Advancing glaucoma care: Integrating artificial intelligence in diagnosis, management, and progression detection. Bioengineering (Basel) 2024; 11(2). doi.org/10.3390/ bioengineering11020122.
6. Prum BE, Jr., Rosenberg LF, Gedde SJ, et al., Primary open-angle glaucoma preferred practice pattern guidelines. Ophthalmology 2016; 123(1): 41-111. doi: 10.1016/j. ophtha.2015.10.053.
7. Swaminathan SS, Jammal AA, Berchuck SI, Medeiros FA, Rapid initial OCT RNFL thinning is predictive of faster visual field loss during extended follow-up in glaucoma. Am J Ophthalmol 2021; 229: 100-7. doi: 10.1016/j. ajo.2021.03.019.
8. Varma R, Steinmann WC, Scott IU, Expert agreement in evaluating the optic disc for glaucoma. Ophthalmology 1992; 99(2): 215-21. doi: 10.1016/s0161-6420(92)31990-6.
9. Ting DSW, Pasquale LR, Peng L, et al., Artificial intelligence and deep learning in ophthalmology. Br J Ophthalmol 2019; 103(2): 167-75. doi: 10.1136/ bjophthalmol-2018-313173.
10. Wang R, Zheng L, Xiong C, et al., Retinal optic disc localization using convergence tracking of blood vessels. Multimedia Tools and Applications 2017; 76: 23309-31. doi. org/10.1007/s11042-016-4146-z.
11. Li Z, He Y, He M, Efficacy of a deep learning system for detecting glaucomatous optic neuropathy based on color fundus photographs. Ophthalmology 2018; 125(8): 1199- 206. doi: 10.1016/j.ophtha.2018.01.023.
12. Ting DSW, Cheung CY, Lim G, et al., Development and validation of a deep learning system for diabetic retinopathy and related eye diseases using retinal images from multiethnic populations with diabetes. JAMA 2017; 318(22): 2211-23. doi: 10.1001/jama.2017.18152.
13. Lee T, Jammal AA, Mariottoni EB, Medeiros FA. Predicting glaucoma development with longitudinal deep learning predictions from fundus photographs. Am J Ophthalmol 2021; 225: 86-94. doi: 10.1016/j.ajo.2020.12.031.
14. Medeiros FA, Jammal AA, Thompson AC, From machine to machine: An OCT-trained deep learning algorithm for objective quantification of glaucomatous damage in fundus photographs. Ophthalmology 2019; 126(4): 513-21. doi: 10.1016/j.ophtha.2018.12.033.
15. Bussel, II, Wollstein G, Schuman JS, OCT for glaucoma diagnosis, screening and detection of glaucoma progression. Br J Ophthalmol 2014; 98 Suppl 2(Suppl 2): ii15-9. doi: 10.1136/bjophthalmol-2013-304326.
16. Mwanza JC, Oakley JD, Budenz DL, Anderson DR, Ability of cirrus HD-OCT optic nerve head parameters to discriminate normal from glaucomatous eyes. Ophthalmology 2011; 118(2): 241-8.e1. doi: 10.1016/j. ophtha.2010.06.036.
17. Ran AR, Tham CC, Chan PP, et al., Deep learning in glaucoma with optical coherence tomography: a review. Eye (Lond) 2021; 35(1): 188-201. doi: 10.1038/s41433-020- 01244-9.
18. Aggarwal R, Sounderajah V, Martin G, et al., Diagnostic accuracy of deep learning in medical imaging: a systematic review and meta-analysis. NPJ Digit Med 2021; 4(1): 65. doi. org/10.1038/s41746-021-00438-z.
19. Miki A, Kumoi M, Usui S, et al., Prevalence and associated factors of segmentation errors in the peripapillary retinal nerve fiber layer and macular ganglion cell complex in spectral-domain optical coherence tomography images. J Glaucoma 2017; 26(11): 995-1000. doi: 10.1097/IJG.0000000000000771.
20. Zhang L, Tang L, Xia M, Cao G, The application of artificial intelligence in glaucoma diagnosis and prediction. Front Cell Dev Biol 2023; 11: 1173094. doi: 10.3389/ fcell.2023.1173094.
21. Maetschke S, Antony B, Garnavi R, et al., A feature agnostic approach for glaucoma detection in OCT volumes. PLoS One 2019; 14(7): e0219126. doi: 10.1371/journal. pone.0219126.
22. Kooner KS, Angirekula A, Treacher AH, et al., Glaucoma diagnosis through the integration of optical coherence tomography/angiography and machine learning diagnostic models. Clin Ophthalmol 2022; 16: 2685-97. doi: 10.2147/ OPTH.S367722.
23. Hao L, Hu Y, Xu Y, et al., Dynamic analysis of iris changes and a deep learning system for automated angle-closure classification based on AS-OCT videos. Eye Vis (Lond) 2022; 9(1): 41. doi: 10.1186/s40662-022-00314-1.
24. Li F, Yang Y, Sun X, et al., Digital gonioscopy based on three-dimensional anterior-segment OCT: An international multicenter study. Ophthalmology 2022; 129(1): 45-53. doi: 10.1016/j.ophtha.2021.09.018.
25. Elze T, Pasquale LR, Bex PJ, et al., Patterns of functional vision loss in glaucoma determined with archetypal analysis. J R Soc Interface 2015; 12(103). doi: 10.1098/ rsif.2014.1118.
26. Li F, Wang Z, Qu G, et al. Automatic differentiation of glaucoma visual field from non-glaucoma visual field using deep convolutional neural network. BMC Med Imaging 2018; 18(1): 35. doi: 10.1186/s12880-018-0273-5.
27. Wang M, Shen LQ, Pasquale LR, et al. An artificial intelligence approach to detect visual field progression in glaucoma based on spatial pattern analysis. Invest Ophthalmol Vis Sci 2019; 60(1): 365-75. doi: 10.1167/ iovs.18-25568.
28. Wang M, Tichelaar J, Pasquale LR, et al. Characterization of central visual field loss in end-stage glaucoma by unsupervised artificial intelligence. JAMA Ophthalmol 2020; 138(2): 190-8. doi: 10.1001/jamaophthalmol.2019.5413.
29. Orlando JI, Fu H, Barbosa Breda J, et al., REFUGE challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Med Image Anal 2020; 59: 101570.
30. Xiong J, Li F, Song D, et al., Multimodal machine learning using visual fields and peripapillary circular OCT scans in detection of glaucomatous optic neuropathy.