In its early stages, glaucoma is predominantly asymptomatic, making early diagnosis difficult. Consequently, the majority of cases are identified in later stages after considerable damage has taken place, thus rendering future interventions ineffective at preserving functional vision. If glaucoma is detected early enough, interventions such as pressure lowering medications or surgery can be used to slow down or halt the progression of disease. Thus, a key strategy to addressing glaucoma worldwide is detecting the disease at its earlier stages and predicting when individuals are at high risk for glaucomatous progression.

Advances in artificial intelligence (AI) and, more specifically, deep learning (DL) provide hope for an earlier and more accurate diagnosis of glaucoma in patients. AI uses computer systems (algorithms) to perform tasks that normally require human intelligence, such as visual perception and natural language processing. Deep learning, a branch of AI, is largely inspired by the structure of the human brain and functions by processing data and creating patterns for use in decision making. There are several methods through which such systems learn. One method, known as supervised learning, occurs when a machine is given a set of input–output pairs and the machine learns characteristics that define the specified output. This is useful in situations where there is a predicted output (ie, mild, moderate, severe glaucoma) for a given input (ie, optic nerve head images) and where a labeled dataset is available to train the AI algorithms. In terms of glaucoma diagnosis, this would mean feeding the machine data from both healthy eyes and eyes suffering from glaucoma. The AI algorithm learns from a process known as backpropagation to optimize its ability to identify the presence or absence of disease indicators.

Unsupervised learning is another branch within DL where the machine is given unlabeled data and categorizes the input into groups that may not necessarily match traditional output classes. For example, the algorithm might cluster optic nerve images that exhibit a characteristic loss of the neuroretinal rim, such as notching, with other similar images and optic nerves with intact neuroretinal rims in another group without labeling them respectively as diseased and healthy. This can be particularly useful to unveil novel patterns or trends that are relevant to glaucoma.

In both unsupervised and supervised learning, a training phase is necessary where the machine is fed the data set (labeled in supervised, unlabeled in unsupervised). Once the algorithm is optimized, the accuracy of the algorithm is defined by how well it predicts a validation data set. Only then is the algorithm allowed to analyze an independent “test” data set. This independent data set must be comprised of images that are different from the initial data set to ensure the results are not biased. In ophthalmology, and more specifically in glaucoma, DL has accurately detected complicated patterns from images and has been shown to have accuracy levels that rival the results from expert human graders.¹ This article will focus on how DL algorithms can facilitate the diagnosis of early-stage glaucoma and identify individuals at high risk for glaucomatous progression.

Detecting Glaucomatous Damage Using Optical Coherence Tomography Images

Optical coherence tomography (OCT) has become widely used in the clinical setting to detect structural changes of the optic nerve and macula, which has been shown to be highly correlated with the diagnosis and monitoring of glaucoma and detecting pathologic glaucomatous changes prior to vision loss.² Identifying retinal nerve fiber layer (RNFL) thinning is one of the first signs of glaucoma and often precedes functional loss.³ As a result, detecting RNFL damage is critical in identifying glaucoma in its early stages. Spectral-domain (SD)-OCT has become the gold standard for objectively quantifying structural damage in glaucoma, because it allows for both a single time point measurement as well as longitudinal quantifications of tissue thickness. This imaging technique is widely used in the clinical setting;⁴ however, the equipment is expensive and not portable, and the images require segmentation so that essential measurements can be extracted. Although various components are automatically segmented by the machine, this process is associated with a high rate of errors; one study found that artifacts in SD-OCT imaging for RNFL thickness was 46.3%,⁵ and another study found that segmentation failure for RNFL thickness was 21%.⁶ These errors can often lead to inaccurate diagnoses and adversely affect management. Manually reviewing each image for potential errors and artifacts can be time consuming and burdensome. To avoid these errors, several studies have attempted to use DL to predict RNFL damage from OCT images without using segmentation and found significantly better results than determining RNFL through conventional methods.^7,8 Marionetti et al found that their DL algorithm that analyzed raw OCT images performed similarly to the conventional method in good-quality images and better than conventional methods in images with artifacts.⁷

Another important structural parameter in determining glaucoma involves changes in the optic nerve head (ONH). Because SD-OCT often is used to quantify ONH damage, it is associated with the same drawbacks listed above regarding segmentation. Several studies have revealed success in applying DL algorithms to the detection of important parameters of ONH damage.^9,10

Detecting Glaucoma Through Fundus Photography

Obtaining quality OCT images requires expensive hardware as well as skilled personnel, which is not always possible in resource-strained areas. In contrast, fundus photography is a universally available diagnostic modality that is more suitable for screening, because the technology is more cost effective and easier to use. Studies have attempted to determine the utility of this approach in diagnosing and monitoring the progression of glaucoma in various clinical settings.¹¹ Within the past decade, several technologies utilizing smartphones to create portable nonmydriatic fundus cameras have been validated in the clinical setting for use in low-resource areas.¹²

Fundus photography is useful in detecting glaucoma, because it allows the clinician to appreciate structural characteristics such as the optic nerve, peripapillary atrophy, and vessel characteristics. More specifically, the components of interest include assessing disc size, cup-to-disc ratio, and neuroretinal rim integrity, all of which are centered around the optic disc. Just as it was important to segment the components of interest in OCT imaging, localization of the essential features of the optic disc from fundus photos is the focus of DL algorithms used in diagnosing glaucoma.

Of note, the interpretation of fundus photographs for glaucoma diagnosis is vulnerable to high clinician interobserver and intraobserver variability.^13,14 Several studies have demonstrated that DL algorithms can identify the structural damage of the optic nerve from fundus photographs with results comparable to expert human graders.^15-19 Although these findings are promising, given the significant variability of human interpretation, Medeiros et al trained a DL algorithm using the objective measurements of RNFL thickness obtained by SD-OCT to classify disc photos for glaucomatous damage.¹⁶ This approach to analyzing fundus photographs may be a reasonable substitute for SD-OCT in resource-strained settings and may address variability issues noted with human graders in any clinical setting.

Detecting and Predicting Glaucomatous Progression

It is difficult to definitively detect glaucomatous progression in the clinical setting, due to the high test–retest variability seen in automated visual field testing, SD-OCT images, and fundus photographs. Another promising application for DL algorithms is predicting the rate of glaucomatous progression. Although the majority of studies have focused on detecting referable glaucoma, a few studies have focused on predicting future visual field changes^20,21 as well as incorporating SD-OCT data to detect progression in glaucomatous patients.^22-24 These DL algorithms require validation in larger cohorts and in diverse populations to understand the true utility and predictive capacity of these findings.

Limitations

Before widespread use of AI, and more specifically DL algorithms, can be implemented in the clinical setting, several key factors need to be considered. One barrier that hinders the widespread use of DL for glaucoma is the lack of a universal, standard definition for the disease, which contributes to the high interobserver and intraobserver variability seen in glaucoma diagnosis.²⁵ At present, glaucoma is diagnosed based on a combination of patient risk factors obtained through history (age, diabetes status, ethnicity, medications), tonometry, structural characteristics (angle anatomy, optic nerve damage, RNFL thinning) and functional characteristics (visual field defects). It is difficult to combine these heterogeneous metrics to create a universal criterion for diagnosis and progression of disease.

To correctly determine the accuracy of a DL algorithm, there is a need for a gold standard, also known as the ground truth, to compare it to. However, with the lack of an objective definition of glaucoma, this becomes difficult.

In addition, the majority of studies comparing the performance of AI in diagnosing and managing glaucoma to conventional methods have been done on patients that have been recruited from glaucoma clinics. As a result, AI algorithms may not be generalizable to patients recruited from the general population.²⁶ Therefore, special efforts need to be made to ensure that the training and testing data sets are representative of all demographics to allow the DL algorithm to be applicable to diverse populations.

Another critical concern is that the validity of the DL algorithms to detect glaucoma is reliant on learning from and analyzing quality images. Efforts are under way to ensure that quality images are obtained and used more routinely, in hopes of moving away from expert graders to investigate each data point individually.²⁷

Conclusion

AI has enormous potential for diagnosing and monitoring glaucoma. Enhanced ability to detect earlier stages of glaucoma may help maintain quality of life for patients. Given that glaucoma is diagnosed by combining medical history, physical exam, visual field tests, fundus photographs, and OCT scans, an ideal AI application will incorporate several, if not all, of these components to create a reliable prediction of disease and progression. The majority of past and current research has centered around using AI to detect referable glaucoma using visual field testing, SD-OCT, and fundus images. Although there has been less research on utilizing AI to detect anterior chamber changes associated with glaucoma,²⁸ forecasting progression,^20,21 and monitoring for progression of glaucoma,^22,23 these areas are ripe for future research. There is little risk that the clinician will be excluded from these processes given the overall complexity of glaucoma diagnosis and monitoring, as outlined in this article. However, AI offers great hope for automating the interpretation of the various metrics we use in clinic so that the clinician can walk the final mile more efficiently and toward enhanced patient care and improved outcomes.

References

1. Jammal AA, Thompson AC, Mariottoni EB, et al. Human versus machine: comparing a deep learning algorithm to human gradings for detecting glaucoma on fundus photographs. Am J Ophthalmol. 2020;211:123-131. doi:10.1016/j.ajo.2019.11.006
2. Schuman JS, Hee MR, Puliafito CA, et al. Quantification of nerve fiber layer thickness in normal and glaucomatous eyes using optical coherence tomography. Arch Ophthalmol. 1995;113(5):586-596. doi:10.1001/archopht.1995.01100050054031
3. Wollstein G, Kagemann L, Bilonick RA, et al. Retinal nerve fibre layer and visual function loss in glaucoma: the tipping point [published correction appears in Br J Ophthalmol. 2013 Aug;97(8):1088]. Br J Ophthalmol. 2012;96(1):47-52. doi:10.1136/bjo.2010.196907
4. Tatham AJ, Medeiros FA. Detecting structural progression in glaucoma with optical coherence tomography. Ophthalmology. 2017;124(12S):S57-S65. doi:10.1016/j.ophtha.2017.07.015
5. Liu S, Yu M, Weinreb RN, Lai G, Lam DS, Leung CK. Frequency-doubling technology perimetry for detection of the development of visual field defects in glaucoma suspect eyes: a prospective study. JAMA Ophthalmol. 2014;132(1):77-83. doi:10.1001/jamaophthalmol.2013.5511
6. 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
7. Mariottoni EB, Jammal AA, Urata CN, et al. Quantification of retinal nerve fibre layer thickness on optical coherence tomography with a deep learning segmentation-free approach. Sci Rep. 2020;10(1):402. Published 2020 Jan 15. doi:10.1038/s41598-019-57196-y
8. Thompson AC, Jammal AA, Berchuck SI, Mariottoni EB, Medeiros FA. Assessment of a segmentation-free deep learning algorithm for diagnosing glaucoma from optical coherence tomography scans. JAMA Ophthalmol. 2020;138(4):333-339. doi:10.1001/jamaophthalmol.2019.5983
9. Devalla SK, Chin KS, Mari JM, et al. A Deep Learning Approach to Digitally Stain Optical Coherence Tomography Images of the Optic Nerve Head. Invest Ophthalmol Vis Sci. 2018;59(1):63-74. doi:10.1167/iovs.17-22617
10. Maetschke S, Antony B, Ishikawa H, Wollstein G, Schuman J, Garnavi R. A feature agnostic approach for glaucoma detection in OCT volumes. PLoS One. 2019;14(7):e0219126. Published 2019 Jul 1. doi:10.1371/journal.pone.0219126
11. Tan NYQ, Friedman DS, Stalmans I, Ahmed IIK, Sng CCA. Glaucoma screening: where are we and where do we need to go?. Curr Opin Ophthalmol. 2020;31(2):91-100. doi:10.1097/ICU.0000000000000649
12. Prathiba V, Rajalakshmi R, Arulmalar S, et al. Accuracy of the smartphone-based nonmydriatic retinal camera in the detection of sight-threatening diabetic retinopathy. Indian J Ophthalmol. 2020;68(Suppl 1):S42-S46. doi:10.4103/ijo.IJO_1937_19
13. Tielsch JM, Katz J, Quigley HA, Miller NR, Sommer A. Intraobserver and interobserver agreement in measurement of optic disc characteristics. Ophthalmology. 1988;95(3):350-356. doi:10.1016/s0161-6420(88)33177-5
14. Jampel HD, Friedman D, Quigley H, et al. Agreement among glaucoma specialists in assessing progressive disc changes from photographs in open-angle glaucoma patients. Am J Ophthalmol. 2009;147(1):39-44.e1. doi:10.1016/j.ajo.2008.07.023
15. Li Z, He Y, Keel S, Meng W, Chang RT, He M. Efficacy of a deep learning system for detecting glaucomatous optic neuropathy based on color fundus photographs. Ophthalmology. 2018;125(8):1199-1206. doi:10.1016/j.ophtha.2018.01.023
16. 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):513-521. doi:10.1016/j.ophtha.2018.12.033
17. Thompson AC, Jammal AA, Medeiros FA. A deep learning algorithm to quantify neuroretinal rim loss from optic disc photographs. Am J Ophthalmol. 2019;201:9-18. doi:10.1016/j.ajo.2019.01.011
18. Christopher M, Belghith A, Bowd C, et al. Performance of deep learning architectures and transfer learning for detecting glaucomatous optic neuropathy in fundus photographs. Sci Rep. 2018;8(1):16685. Published 2018 Nov 12. doi:10.1038/s41598-018-35044-9
19. 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):2211-2223. doi:10.1001/jama.2017.18152
20. Berchuck SI, Mukherjee S, Medeiros FA. Estimating rates of progression and predicting future visual fields in glaucoma using a deep variational autoencoder. Sci Rep. 2019;9(1):18113. Published 2019 Dec 2. doi:10.1038/s41598-019-54653-6
21. Park K, Kim J, Lee J. Visual field prediction using recurrent neural network [published correction appears in Sci Rep. 2019 Sep 6;9(1):13173]. Sci Rep. 2019;9(1):8385. Published 2019 Jun 10. doi:10.1038/s41598-019-44852-6
22. Yousefi S, Kiwaki T, Zheng Y, et al. Detection of longitudinal visual field progression in glaucoma using machine learning. Am J Ophthalmol. 2018;193:71-79. doi:10.1016/j.ajo.2018.06.007
23. Christopher M, Belghith A, Weinreb RN, et al. Retinal nerve fiber layer features identified by unsupervised machine learning on optical coherence tomography scans predict glaucoma progression. Invest Ophthalmol Vis Sci. 2018;59(7):2748-2756. doi:10.1167/iovs.17-23387
24. Brigatti L, Nouri-Mahdavi K, Weitzman M, Caprioli J. Automatic detection of glaucomatous visual field progression with neural networks. Arch Ophthalmol. 1997;115(6):725-728. doi:10.1001/archopht.1997.01100150727005
25. Blumberg DM, De Moraes CG, Liebmann JM, et al. Technology and the glaucoma suspect. Invest Ophthalmol Vis Sci. 2016;57(9):OCT80-OCT85. doi:10.1167/iovs.15-18931
26. Campbell CG, Ting DSW, Keane PA, Foster PJ. The potential application of artificial intelligence for diagnosis and management of glaucoma in adults. Br Med Bull. 2020;134(1):21-33. doi:10.1093/bmb/ldaa012
27. Li Z, Guo C, Nie D, et al. Deep learning from "passive feeding" to "selective eating" of real-world data. NPJ Digit Med. 2020;3:143. doi:10.1038/s41746-020-00350-y
28. Fu H, Baskaran M, Xu Y, et al. A deep learning system for automated angle-closure detection in anterior segment optical coherence tomography images. Am J Ophthalmol. 2019;203:37-45. doi:10.1016/j.ajo.2019.02.028