Editor’s Note: Ophthalmology Times is pleased to introduce a blog series on artificial intelligence (AI), called “Innovations in Ophthalmology.” The first installment in this series provides an overview of AI in medicine. Look for future blogs that will delve into more specific speclatiies within ophthalmology.
Artificial intelligence (AI) has made its way into medicine. Once an AI system gains the ability to recognize patterns or markers of a disease, it can become a tool for automated diagnosis. AI systems are already available or in development for the detection of multiple ophthalmic diseases, including diabetic retinopathy (DR), age-related macular degeneration (AMD), and glaucoma.
The earliest forms of medical AI were simple automated detectors, designed to recognize a defined set of disease features that were programmed into the system. A limitation of these early systems is that they will recognize only patients who express features that are included in the defined program.1
The most advanced iteration of medical AI teaches itself the features of disease by analyzing a representative set of images from people with and without the disease, and possibly across various stages of disease. During the learning phase, the system performs multiple rounds of analysis, assessment, and re-analysis until each image can be faithfully identified. In contrast to simple automated detectors, AI systems that self-teach (called deep learning with convolutional neural networks) are unconstrained in the number of disease features that they may identify.1
How AI benefits patients and physicians
Automated AI systems have multiple benefits that will advance clinical practice. By design, an AI system is a tool for automated diagnosis, which can reduce burden in a care setting with limited physician resources. An AI system can perform its function anywhere, which means that data can be collected remotely and sent to an AI center for analysis. This is basically an AI version of telemedicine, where you can provide a diagnosis and other expert guidance without the need for travel (patient or physician), increasing access to care for patients who live in remote or difficult-to-reach areas, and reduces the burden on patients, care givers and physicians.
AI systems can quickly process a large amount of data, which means that you can analyze a full series of closely spaced scans that cover a wide area of retina from a single patient. This increases the probability of identifying early-stage disease that may only show features in small isolated areas. Improving the ability to detect early-stage disease is significant, because outcomes are often best when treatment is initiated at the beginning of a disease course, before the tissue has been damaged beyond repair.
In addition to advancing clinical practice, deep learning AI systems might contribute to an improved understanding of disease mechanisms. Deep learning AI has the potential to identify previously unknown patterns of disease that would improve our understanding of the pathogenesis of disease, and provide additional markers for diagnosis, staging, and prognosis.1
AI in practice and development
Medical AI systems are now available or in development for the detection of a number of ophthalmic diseases, including DR, wet AMD, cataract, and glaucoma. AI systems are tested for their ability to accurately detect a disease, and this is typically assessed with measures called sensitivity and specificity. Sensitivity is a measure of how well the system catches all positive cases of the disease, and specificity is a measure of how well the system avoids false positives. For each measure, the higher the value (on a scale of 0.0% to 100%), the better the accuracy.
The most developed AI systems in ophthalmology are those that are designed to detect Diabetic Retinopathy. Google Brain is a deep learning AI research team that created a system to identify patients with DR and DME based solely on the analysis of retina fundus photos. The accuracy of Google Brain’s AI system was evaluated with two test runs that use fundus photos from patients that had already been diagnosed by expert physicians (The EyePACS-1 data set and Messidor-2 data set). Depending on how the analysis was performed (whether focused on sensitivity or specificity), Google Brain’s AI system had sensitivity values of 97.5% and 96.1% in each practice set, and specificity values of 98.1% and 98.5%.2
Based on the success of these and other tests, Google Brain announced that they were working with the Aravind Eye Care System in India to integrate their AI system as part of an initiative to increase access to global DR care.3
IDx is an AI company working on separate deep learning systems for the detection of multiple ophthalmic diseases, including DR, AMD, and glaucoma. The IDx-DR system is designed for use in a primary care setting, and provides results within a minute of submitting fundus photos. Any patient with a positive diagnosis of DR receives a corresponding referral to an ophthalmologist.
In February 2018, IDx announced that the results of a pivotal multicenter study showed that when the system was incorporated into a typical everyday clinical setting, the IDx-DR accurately diagnosed moderate to severe DR, including DME, among 898 patients with diabetes (sensitivity of 91% and specificity of 84%).4 In April 2018, the FDA authorized marketing of IDx-DR for the autonomous detection of “more than mild” DR in adults with diabetes.5
Deep learning AI approaches are also in development that could improve the care of patients with wet AMD. These systems are being used to identify anatomic OCT-based features that could predict the timing and extent of disease progression, or which patients will require extensive anti-VEGF treatment after the initiation phase.6,7 Another deep learning AI system has been shown to accurately characterize the pattern of intraretinal fluid in patients with wet AMD or retinal vein occlusion (RVO), with the ability to localize, quantify, and distinguish between intraretinal cysts and subretinal fluid.8
AI systems are being developed and validated for the automatic diagnosis and characterization of other ophthalmic diseases, beyond those that affect the retina, including the following conditions that affect the anterior segment: AI to diagnose and grade (location, density, and opacity) cataract in pediatric patients, based on an analysis of slit-lamp images9; AI to diagnose glaucoma in adolescent or adult patients, based on measurement of the visual field and thickness of the retinal nerve fiber layer (on OCT)10; AI to diagnose keratoconus, based on Scheimpflug tonometry that provides measures of corneal curvature, thickness, opacities.11
An especially exciting development in the field of ophthalmology AI came with the report of a system developed as part of a collaboration between Moorfields Eye Hospital in London and another Google AI team, DeepMind. These teams created an AI system that combines two DLS with the ability to detect 50 ophthalmic diseases based on analysis of three-dimensional OCT data. The first DLS uses the raw OCT data to create a tissue map, and then the second DLS analyzes the tissue map for potential markers of disease.12,13
The DeepMind system was validated in a study that showed it was 94% sensitive, catching most positive cases of each disease. In fact, DeepMind performed as well or better than human clinical experts (retina specialists and optometrists with medical retina training), depending on who the experts were and how much additional information they had to work with (e.g. fundus images, patient medical histories). What’s also impressive is that the system gives more than just a yes-or-no diagnosis, but provides multiple levels of actionable information. For instance, the system provides probabilities for multiple similar diseases in addition to the top pick. The system also provides an accompanying recommendation on urgency of referral, with options of ‘observation only’, ‘routine’, ‘semi-urgent’, and ‘urgent’.12,13
Perhaps what’s most intriguing though, is that the system gives insight into how the diagnosis was made. Until now, most DLS systems have operated within a ‘black box’, where the images are loaded and the answer comes out the other end, and beyond validation studies, you have to trust the result. In contrast, the DeepMind system is giving information along the path, for those who need to see the inner workings, almost like a proof in math class.12,13
Deep learning AI is already advancing eye care, with the first system being FDA-approved in April 2018 for marketing to primary care centers. This system, the IDx-DR, is for the diagnosis of DR in patients with diabetes, but there is ongoing development and validation of additional systems that will advance the care of patients with other ophthalmic diseases, including those that affect the anterior segment.
Rohit Varma, MD, MPH
Dr. Varma is the founding director of the USC Roski Eye Institute and a glaucoma specialist. He also holds the Grace and Emery Beardsley chair in Ophthalmology, and is a professor of preventive medicine at the Keck School of Medicine of University of Southern California, Los Angeles.
1. Roach L. Artificial intelligence. EyeNet Magazine. November 2017;77-83.
2. 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;316:2402-2410.
3. Simonite T. Google’s AI eye doctor gets ready to go to work in India. Wired Magazine. June 8, 2017. www.wired.com/2017/06/googles-ai-eye-doctor-gets-ready-go-work-india/. Accessed April 19, 2018.
4. van der Heijden AA, Abramoff MD, Verbraak F, et al. Validation of automated screening for referable diabetic retinopathy with the IDx-DR device in the Hoorn Diabetes Care System. Acta Ophthalmol. 2018;96:63-68.
5. FDA permits marketing of IDx-DR for automated detection of diabetic retinopathy in primary care [press release]. IDx. April 12, 2018. www.eyediagnosis.net/single-post/2018/04/12/FDA-permits-marketing-of-IDx-DR-for-automated-detection-of-diabetic-retinopathy-in-primary-care. Accessed April 19, 2018.
6. BogunoviÄ H, Montuoro A, Baratsits M, et al. Machine learning of the progression of intermediate age-related macular degeneration based on OCT imaging. Invest Ophthalmol Vis Sci. 2017;58:BIO141-BIO150.
7. BogunoviÄ H, Waldstein SM, Schlelgl T, et al. Prediction of anti-VEGF treatment requirements in neovascular AMD using a machine learning approach. Invest Ophthalmol Vis Sci. 2017;58:3240-3248.
8. Schlegl T, Waldstein SM, Bogunovic H, et al. Fully automated detection and quantification of macular fluid in OCT using deep learning. Ophthalmology. 2018;125:549-558.
9. Jiu X, Jiang J, Zhang K, et al. Localization and diagnosis framework for pediatric cataracts based on slit-lamp images using deep features of a convolutional neural network. PLoS One. 2017;12:e0168606.
10. Kim SJ, Cho KJ, Oh S. Development of machine learning models for diagnosis of glaucoma. PLoS One. 2017;12:e0177726.
11. Ruiz Hidalgo I, Rozema JJ, Saad A, et al. Validation of an objective keratoconus detection system implemented in a Scheimpflug tomographer and comparison with other methods. Cornea. 2017;36:689-695.
12. Ross C. Opening the ‘black box,’ Google DeepMind AI system diagnoses eye diseases and shows its work. STAT Magazine. August 13, 2017. https://www.statnews.com/2018/08/13/google-deepmind-ai-diagnoses-eye-diseases/. Accessed August 14, 2018.
13. De Fauw J, Ledsam JR, Romera-Paredes B, et al. Clinically applicable deep learning for diagnosis and referral in retinal disease. Nat Med. Epub ahead of print Aug 13 2018. doi: 10.1038/s41591-018-0107-6.