Recent advances in artificial intelligence (AI) have taken the healthcare sector by storm, already evident with the emergence of numerous commercial entities with specific subspecialty niches, for example, aiming to improve the pathologic detection of cancer1 and interpretation of radiology images.2 Within ophthalmology, AI is already augmenting diagnostic imaging capabilities, which may soon lead to deployment of cost-efficient telemedicine screening programs worldwide. Recent studies have shown deep learning models are capable of aiding in the detection and diagnosis of diseases afflicting the posterior segment of the eye with extremely high accuracy.3-17 In April 2018, the US Food and Drug Administration (FDA) granted marketing approval to IDx for the cloud-based IDx-DR, the first AI-based medical device to detect referable diabetic retinopathy (DR) from color fundus photographs, and the first approved instrument to provide a screening decision without clinician input.18
Because of the surging popularity in mainstream media, the terms AI, machine learning, and deep learning have been used interchangeably at times; however, it is important to differentiate and distinguish among them. They can each be viewed as concentric circles, with the largest circle being AI, machine learning being a smaller circle within the subset of AI, and deep learning being the smallest circle within the subset of machine learning.
Deep learning is an increasingly popular and powerful model of machine learning, composed of algorithms that use a cascade of multilayered artificial neural networks – “deep” referring to the number of layers – to independently perform feature extraction from data.19,20 Each successive layer in the network uses the output from the previous layer as input, with the final layer revealing the diagnostic output. Deep learning can be regarded as an improvement on conventional artificial neural networks by creating networks with multiple layers. Learning in this format can be classified as either supervised (classification-based) or unsupervised (pattern analysis-based). The latter represents one of the more fascinating aspects of deep learning, where large datasets are analyzed to discover underlying patterns without the need for feature engineering.
Clinically speaking, instead of researchers hand-coding instructions to an algorithm on what a microaneurysm, hemorrhage, or neovascular frond may look like on a diabetic fundus photograph, rather, they input an image labeled as “severe nonproliferative DR” for example, and with enough labeled data, the computer eventually learns what that is. To train itself, a deep learning neural network must have a variable and large enough dataset available. In the context of ophthalmology, while it is possible that the algorithm independently appreciates the same classical features of DR, it is also possible that it has identified its own pattern recognition of disease beyond the scope of human interpretation. This is referred to as the “black box” of deep learning. Elucidating what exactly the algorithm interprets is the subject of ongoing research.
Deep Learning and Glaucoma
Most of the earlier deep learning initiatives in ophthalmology focused on the detection of referable DR and age-related macular degeneration (AMD). Compared to retinal diseases, however, there have been limited but expanding efforts evaluating the utility of deep learning models in improving the diagnostic accuracy and earlier detection of glaucoma. Unlike DR and AMD, the diagnosis of glaucoma presents a unique challenge to potential automation due to the numerous variables that contribute to the clinician’s decision-making beyond retinal/optic disc-based imaging, such as patient demographics, intraocular pressure (IOP) measurements, corneal thickness measurements, and visual field changes. Furthermore, longitudinal changes over time are also paramount to the accurate diagnosis and management of the disease. Given the multifactorial etiology of glaucoma, groups have been interested in applying deep learning models to analyze these various inputs, including optic disc photographs, optical coherence tomography (OCT) of the nerve and peripapillary retina, and visual field testing, with the goal of improving accurate detection of glaucoma at the earliest possible disease state to prevent irreversible vision loss.
Optic Disc Photographs
Developing deep learning algorithms capable of distinguishing glaucoma suspect and glaucomatous optic discs from normals in the absence of ancillary confirmatory functional testing may prove particularly helpful in augmenting large-scale ocular telemedicine programs, which, in the future, are likely to become increasingly automated.
These efforts have been gradually making progress over the past several years. For example, in 2015, Chen et al reported on a deep learning method for detection of glaucoma based on funduscopic images of the optic disc using the ORIGA and SCES datasets of glaucoma cases.21 They reported area under the receiver operating characteristic curve (AUC) values for each dataset of 0.831 (ORIGA) and 0.887 (SCES), which they found to be better than previously reported models. More recently, in 2018, Li et al used their own deep learning algorithm for the classification of 48,000 optic disc photographs for the detection of referable glaucoma, and reported an AUC of 0.986 with a sensitivity and specificity of 95.6% and 92.0%, respectively.22 The authors noted that coexistence of high or pathologic myopia was the most common cause for false-negative results given that optic nerve evaluation in myopic eyes can be clinically challenging.
Asaoka et al additionally reported on a fundus photography-based deep learning algorithm for the detection of glaucoma, albeit using a smaller number of images.23 In their study, a training dataset of 1,364 color fundus photographs with glaucomatous appearance and 1,768 without glaucoma appearance was used. Notably, the testing dataset was subclassified into 4 groups: 30 eyes of 30 highly myopic patients with glaucoma, 22 eyes of 22 highly myopic patients without glaucoma, 34 eyes of 34 non-highly myopic patients with glaucoma, and 28 eyes of 28 non-highly myopic patients without glaucoma. The investigators found that the algorithm could diagnose glaucoma from the photos as accurately as or better than those diagnosed by an ophthalmologist, with an AUC of 0.954, sensitivity of 95.0%, and specificity of 70.3%. That the algorithm performed well even in the setting of high myopia is encouraging.
Visulytix is an AI company developing machine learning clinical decision support systems. The company’s Pegasus platform autonomously screens for glaucoma through optic disc assessment of color fundus images or OCT volumes. Recently, Seo et al presented the results of a study comparing diagnostic accuracy of optic disc image assessment by glaucoma specialist evaluation versus an automated, deep learning-based decision support tool (Pegasus-disc) in 186 optic disc images from 186 patients.24 Of these images, 81 (43.5%) met reference standard criteria for manifest glaucoma. The AUC of the algorithm for detection of manifest glaucoma was 0.840, with sensitivity of 77.8% and specificity of 81.9%. This compared favorably to the 2 glaucoma specialists, which demonstrated an AUC of .870, with sensitivity of 77.8% and specificity of 93.3%. The study concluded that Pegasus-disc was equivalent to the consensus of 2 human experts in the detection of glaucoma, and, in some cases, may have higher sensitivity. Given the encouraging initial results, the study is now being expanded to include up to 400 subjects.
To implement effective population-screening programs (ie, diabetic vision screening), concurrent assessment for related vision-threatening conditions would be mandatory. Ting et al aimed to accomplish this with their deep-learning algorithm applied to multiethnic cohorts of diabetic patients, which also was trained to identify referable glaucoma.8 In the primary validation dataset (n=71,896 images), the AUC of the algorithm for referable glaucoma was 0.942, with sensitivity of 96.4% and specificity of 87.2%. A particular strength of this study was the use of diverse retinal images of varying quality from different camera types and in representative screening populations of varying ethnicities, which was unique given that publicly available fundus image sets (ie, Messidor-2) consist largely of homogenous white individuals.
Optical Coherence Tomography
With the promising results from deep learning interpretation of fundus photography, efforts have quickly expanded toward OCT analysis, given its widespread adoption and integration into ophthalmic clinical practice.10,12-16 Notably, clinicians are increasingly utilizing OCT in the management of glaucoma, especially as the quality and resolution of the technology continues to improve. In turn, OCT has added new diagnostic parameters to aid in the earlier detection and management of glaucoma, which lends itself to potential integration with deep learning.
Girard et al developed a deep learning model which “digitally stains” (ie, highlights) OCT images of the optic nerve head based on tissue layer.25 In their study, 100 total images from 40 healthy subjects and 60 with glaucoma were used to train a deep learning algorithm to digitally stain 6 layers of the optic nerve head: retinal nerve fiber layer (RNFL) + prelamina, all other retinal layers, retinal pigment epithelium, choroid, peripapillary sclera, and lamina cribrosa. Given that these tissues demonstrate significant anatomic changes in the glaucomatous state, the investigators advocated that automated identification and measurement of these layers may offer a framework to potentially improve glaucoma management.
Muhammad et al utilized a hybrid deep learning method combined with a single wide-field OCT protocol to distinguish eyes previously classified as either healthy suspects (n=47) or mild glaucoma (n=57) based on retinal nerve fiber layer thickness measurements.26 The reported an accuracy that ranged from 63.7% to 93.1%, depending on the input map. Overall, their findings outperformed standard OCT and visual field clinical metrics in distinguishing eyes that were healthy from those with early glaucoma. A separate group from the Singapore Eye Research Institute (SERI) employed their own hybrid approach, training a deep-learning system in the detection of glaucoma based on both OCT and fundus images.27 In their study, the deep learning system was trained on paired 3D OCT volumes together with fundus images (n=168 eyes). Notably, the combination of OCT and fundus images demonstrated a higher AUC and specificity compared to both methods alone, which the investigators hypothesized could be due to a possible synergistic effect.
A similar synergistic effect was observed in a separate study by Zangwill et al where they trained a deep learning model based on scanning laser ophthalmoscopy (SLO) images and RNFL thickness maps extracted from spectral-domain OCT cube acquisitions amongst a cohort of 375 subjects (538 eyes) with glaucomatous visual field damage and 254 subjects (444 eyes) without visual field damage.28 The investigators found that for detecting confirmed visual field damage from OCT images, a combined SLO and RNFL thickness deep-learning model achieved the highest AUC (0.92) out of the 3 models tested (SLO only model with AUC=0.81, and RNFL thickness only model with AUC=0.85).
Visual Field Testing
With respect to visual fields, Asaoka et al compared a deep-learning method (feed-forward neural network) with other machine learning methods to differentiate fields of preperimetric open-angle glaucoma patients (defined as eyes with a glaucomatous optic disc or fundus appearance, or both, and an apparently normal visual field) from those of healthy eyes.29 In total, 171 preperimetric glaucoma 30-2 visual fields from 51 open-angle glaucoma patients were analyzed with 108 30-2 visual fields from 87 healthy patients. The investigators reported an AUC of 0.926 with the deep-learning algorithm, which was significantly greater than other machine-learning methods employed in the study.
While most studies using deep learning have focused on image classification toward confirmatory and more definitive diagnosis of disease, there have been limited investigations into the potential of deep learning in predicting future findings or treatment outcomes. This field of predictive medicine may be particularly interesting in the management of glaucoma, as many currently unknown factors contribute to the rate or severity of glaucoma progression. If predictive deep-learning models could accurately forecast visual field progression, this could positively augment clinical decision making, allowing the glaucoma specialist to know which patients to treat more or less aggressively.
Most recently, Wen et al trained a deep-learning artificial neural network on consecutive Humphrey 24-2 visual fields (HVF) from 1998 to 2018 from an institutional database to determine whether they could forecast a future 24-2 HVF finding.30 A total of 32,443 24-2 HVFs were extracted, resulting in over 1.7 million perimetry points with hundredth-decimal precision. The investigators found that the deep-learning models successfully predicted progressive visual-field loss in glaucomatous eyes up to 5.5 years in advance, with a correlation of 0.92 between the model predicted mean deviations and actual mean deviations on future HVFs.
The Future of Deep Learning in Glaucoma
Deep learning has shown substantial promise to date in automated image analysis and classification of fundus photographs, OCT images, and visual fields. Moving forward, this technology appears poised to enable earlier glaucoma detection and diagnosis. However, additional testing and research is required to further clinically validate this technology, including the lack of prospective clinical trials. Looking further into the future, deep learning may eventually enable precision medicine for the optimal management of glaucoma. GP
- PathAI. https://www.pathai.com . Accessed June 17, 2018.
- Enlitic. http://www.enlitic.com . Accessed June 17, 2018.
- Abràmoff MD, Lou Y, Erginay A, et al. Improved automated detection of diabetic retinopathy on a publicly available dataset through integration of deep learning. Invest Ophthalmol Vis Sci. 2016;57(13):5200-5206.
- Abràmoff MD, Folk JC, Han DP, et al. Automated analysis of retinal images for detection of referable diabetic retinopathy. JAMA Ophthalmol. 2013;131(3):351-357.
- Squirrell DM, Talbot JF. Screening for diabetic retinopathy. J R Soc Med. 2003;96(6):273-276.
- Gulshan V, Peng L, Coram M, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA. 2016;316(22):2402-2410.
- Gargeya R, Leng T. Automated identification of diabetic retinopathy using deep learning. Ophthalmology. 2017;124(7):962-969.
- 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.
- Burlina PM, Joshi N, Pekala M, Pacheco KD, Freund DE, Bressler NM. Automated grading of age-related macular degeneration from color fundus images using deep convolutional neural networks. JAMA Ophthalmol. 2017;135(11):1170-1176.
- Lee CS, Tyring AJ, Deruyter NP, Wu Y, Rokem A, Lee AY. Deep-learning based, automated segmentation of macular edema in optical coherence tomography. Biomed Opt Express. 2017;8(7):3440-3448.
- Schlegl T, Waldstein SM, Bogunovic H, et al. Fully automated detection and quantification of macular fluid in OCT using deep learning. Ophthalmology. 2018;125(4):549-558.
- Fang L, Cunefare D, Wang C, Guymer RH, Li S, Farsiu S. Automatic segmentation of nine retinal layer boundaries in OCT images of non-exudative AMD patients using deep learning and graph search. Biomed Opt Express. 2017;8(5):2732-2744.
- El Tanboly A, Ismail M, Shalaby A, et al. A computer-aided diagnostic system for detecting diabetic retinopathy in optical coherence tomography images. Med Phys. 2017;44(3):914-923.
- Prahs P, Radeck V, Mayer C, et al. OCT-based deep learning algorithm for the evaluation of treatment indication with anti-vascular endothelial growth factor medications. Graefes Arch Clin Exp Ophthalmol. 2018;256(1):91-98.
- Lee CS, Baughman DM, Lee AY. Deep learning is effective for classifying normal versus age-related macular degeneration OCT images. Ophthalmol Retina. 2017;1(4):322-327.
- Schmidt-Erfurth U, Bogunovic H, Sadeghipour A, et al. Machine learning to analyze the prognostic value of current imaging biomarkers in neovascular age-related macular degeneration. Ophthalmol Retina. 2018;2(1):24-30.
- Bogunovic H, Waldstein SM, Schlegl T, et al. Prediction of anti-VEGF treatment requirements in neovascular AMD using a machine learning approach. Invest Ophthalmol Vis Sci. 2017;58(7):3240-3248.
- IDx. https://www.eyediagnosis.net . Accessed June 17, 2018.
- Shen D, Wu G, Suk HI. Deep learning in medical image analysis. Annu Rev Biomed Eng. 2017;19:221-248.
- Litjens G, Kooi T, Bejnordi BE, et al. A survey on deep learning in medical image analysis. Med Image Anal. 2017;42:60-88.
- Chen X, Xu Y, Wong DWK, Wong TY, Liu J. Glaucoma detection based on deep convolutional neural network. Conf Proc IEEE Eng Med Biol Soc. 2015;2015:715-718.
- 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 Mar 2. [Epub ahead of print]
- Asaoka R, Shibata N, Murata H, Tanito M. Construction of a deep learning algorithm to automatically diagnose glaucoma using a fundus photograph. Presented at: Association for Research in Vision and Ophthalmology (ARVO) 2018 Annual Conference; April 29 – May 3, 2018; Honolulu, HI. http://www.arvo.org/annual-meeting/program/online-planner . Accessed June 17, 2018.
- Seo E, Jaccard N, Trikha S, Pasquale LR, Song BJ. Automated evaluation of optic disc images for manifest glaucoma detection using a deep-learning, neural network-based algorithm. Presented at: Association for Research in Vision and Ophthalmology (ARVO) 2018 Annual Conference; April 29 – May 3, 2018; Honolulu, HI. http://www.arvo.org/annual-meeting/program/online-planner . Accessed June 17, 2018.
- Devalla SK, Mari JM, Tun TA, et al. A device-independent deep learning approach to digitally stain optical coherence tomography images of the optic nerve head. Presented at: Association for Research in Vision and Ophthalmology (ARVO) 2018 Annual Conference; April 29 – May 3, 2018; Honolulu, HI. http://www.arvo.org/annual-meeting/program/online-planner . Accessed June 17, 2018.
- Muhammad H, Fuchs TJ, De Cuir N, et al. Hybrid deep learning on single wide-field optical coherence tomography scans accurately classifies glaucoma suspects. J Glaucoma. 2017;26(12):1086-1094.
- Tjio G, Tham YC, Li S, et al. A novel deep learning approach for glaucoma detection from 3D OCT volumes and fundus images. Presented at: Association for Research in Vision and Ophthalmology (ARVO) 2018 Annual Conference; April 29 – May 3, 2018; Honolulu, HI. http://www.arvo.org/annual-meeting/program/online-planner . Accessed June 17, 2018.
- Christopher M, Belghith A, Bowd C, Goldbaum MH, Weinreb RN, Zangwill LM. Deep learning approaches can detect glaucomatous functional loss better than standard SD-OCT retinal nerve fiber layer thickness. Presented at: Association for Research in Vision and Ophthalmology (ARVO) 2018 Annual Conference; April 29 – May 3, 2018; Honolulu, HI. http://www.arvo.org/annual-meeting/program/online-planner . Accessed June 17, 2018.
- Asaoka R, Murata H, Iwase A, Araie M. Detecting preperimetric glaucoma with standard automated perimetry using a deep learning classifier. Ophthalmology. 2016;123(9):1974-1980.
- Wen JC, Lee CS, Keane PA, et al. Forecasting future Humphrey visual fields using deep learning. Preprint. https://www.biorxiv.org/content/early/2018/04/02/293621.article-info . Accessed June 17, 2018.