AI in glaucoma: Area ripe for research with unlimited potential
Click Here to Manage Email Alerts
By 2040, worldwide cases of glaucoma are expected to increase 74%, but the potential for artificial intelligence technology to diagnose and manage the disease is vast.
AI has become the state-of-the-art method used for vision tasks such as image classification, and deep learning algorithms have been used successfully in ophthalmology to detect symptoms of diabetic retinopathy, age-related macular degeneration and glaucomatous disease on photographs.
Source: Joel S. Schuman, MD, FACS
“At present, AI is not being used clinically for glaucoma as the clinical use of AI in ophthalmology is limited to diabetic retinopathy detection. With this being said, AI is coming, and there is nothing stopping it,” OSN Glaucoma Board Member Joel S. Schuman, MD, FACS, Elaine Langone Professor of Ophthalmology and vice chair of research in the department of ophthalmology at NYU Langone Medical Center, said. “AI has numerous potentials in glaucoma from detecting disease to predicting and detecting progression. This is an exciting time with the application of AI through a number of different areas, and glaucoma is certainly an area that can benefit greatly from this technology.”
However, some concern has been raised within the ophthalmology field that AI will eventually take the place of clinicians.
“Currently, AI for glaucoma detection and management is only experimental, and there is not a role for it in real-world care,” Michael F. Chiang, MD, director of the National Eye Institute at the NIH, said. “It is important to note that the goal of AI technology is not to replace ophthalmologists, but to develop better technologies to complement the art of medicine. Many ophthalmologists wonder if a machine will replace them, but they should not go into this thinking like that. Instead, they should go in with the mentality of how they can learn about this technology to help make them a better clinician.”
Ocular Surgery News spoke with experts about the potential role of AI to help diagnose and manage glaucoma, how AI may be used to distinguish between glaucoma stages and progression as well as the role that AI will play in glaucoma within the next 5 to 10 years.
Detection
Diagnosing glaucoma is somewhat complex compared with the diagnosis of other ophthalmologic diseases, but data have shown the ability of AI to help detect glaucoma as well as predict disease progression.
“The diagnosis of glaucoma is not straightforward but is based upon many different pieces of data — the disc exam, the visual field and OCT results,” Grace M. Richter, MD, MPH, associate professor of clinical ophthalmology, director of the Glaucoma Fellowship Program and interim director of glaucoma service at Keck Medicine of USC at USC Roski Eye Institute, said. “We also evaluate a patient’s risk factors for glaucoma, such as their central corneal thickness, family history and others. We put all the puzzle pieces together and then make the determination on whether a patient has glaucoma. One can imagine that as the power of AI moves forward, there will be a lot of potential for the technology to put all this data together in a similar manner and help on a wider scale where ophthalmologists are not able to determine whether a patient has glaucoma or is at risk for glaucoma. But we are still several steps away from this.”
Results of a study presented at the 2020 American Glaucoma Society annual meeting, previously reported by OSN, showed two conventional machine learning algorithms and two convolutional neural network models achieved high glaucoma diagnostic performances by reading full retinal nerve fiber layer (RNFL) thickness maps.
Richter and colleagues sought to assess the diagnostic accuracy of multiple machine learning models using full RNFL thickness maps to detect glaucoma among 93 eyes of 69 participants compared with 128 eyes of 128 healthy participants. Results showed that all four machine learning models had area under the curve (AUC) of greater than 0.90 compared with an AUC of 0.76 with mean circumpapillary RNFL thickness.
“In this study, we sought to determine if AI applied to the RNFL thickness map could outperform prior studies that had reported on the diagnostic accuracy of the OCT summary parameters in detecting glaucoma. All four machine learning models did outperform the traditional summary parameters,” Richter said. “These models were based on a relatively small data set of a few hundred patients, but they still improved diagnostic accuracy by about 20%. We can only imagine that by combining big data with AI like this that it will be even more powerful. This project is just scratching the surface of the potential for AI in glaucoma and demonstrates that we are not necessarily using the data that we already have to its fullest potential at this point. AI will allow us to detect glaucoma better with any given technology.”
In a study presented at the 2019 American Academy of Ophthalmology annual meeting, researchers acquired 13,231 visual fields to develop an AI-enabled visualization tool to project outcomes and functional worsening of glaucoma.
The glaucoma radar enabled investigators to consistently verify the functional worsening or no worsening of the t-distributed stochastic neighbor embedding map.
The radar potentially provides clinicians with three layers of advanced visualization of glaucoma knowledge, including global visual functional severity, extent of visual functional loss in hemifields and local patterns of visual functional loss, OSN previously reported.
“Glaucoma radar can be used for personalized monitoring of the visual functional loss of eyes with glaucoma. The trajectory of glaucoma worsening of different patients could be kept on the screen until the next visit of the patient, thus allowing the clinician to see the trajectory of visual worsening at all times. Trajectories of the visual functional worsening of different patients can be simply compared on the glaucoma radar, thus allowing the clinician to identify which patients are worsening fast,” Siamak Yousefi, PhD, lead developer of the tool and co-author of the study, said in the previous report.
Management
AI has also shown potential in guiding management of glaucoma.
AI deep learning technology has taken off for research applications and will be used clinically in the future, Linda Zangwill, PhD, professor of ophthalmology, director of clinical research and director of the Imaging Data Evaluation and Analysis Center at the Hamilton Glaucoma Center of the Viterbi Family Department of Ophthalmology at UC San Diego, said.
“Right now, there is not a lot available in terms of commercial software, and that is what people like myself are working on,” Zangwill said. “AI can improve the accuracy and consistency of glaucoma management, and it can provide decision support to automate the interpretation of images and visual fields to improve glaucoma management. The technology also has a role in improving the efficiency of glaucoma clinical trials for new therapeutics. A lot of this is not yet available in clinic, but it is coming very soon.”
Zangwill and colleagues are examining the use of fundus photographs to determine whether a patient has glaucoma.
“For glaucoma management, we predicted visual field damage from the OCT optic nerve head images, which is important because visual field testing is variable, and patients do not like it,” Zangwill said. “If we could predict the visual field from the structure of the OCT image using AI, then we could potentially effectively tailor the frequency of visual field to each patient. Perhaps we can tell the patient that we can skip visual field testing for one office visit because predictions from the OCT suggest their visual field has not changed. Conversely, for patients with deteriorated visual field predicted by OCT, we can schedule visual field testing sooner. With AI, we will be able to tailor management of glaucoma more effectively.”
Felipe A. Medeiros, MD, PhD, director of the clinical research unit at Duke Eye Center, and colleagues found that an OCT-trained deep learning algorithm’s predictions of RNFL thicknesses from fundus photographs effectively predicted future development of visual field defects in eyes of individuals suspected to have glaucoma.
In the retrospective cohort study, researchers used a machine-to-machine (M2M) OCT-trained deep learning algorithm to predict RNFL thickness measurements from fundus photographs. They assessed the efficacy of the M2M-predicted RNFL thickness to detect glaucomatous damage before the appearance of visual field defects and evaluated the deep learning predictions of RNFL thickness and longitudinal changes to predict the risk for glaucoma conversion.
As reported in OSN, study results showed the M2M mean baseline predicted RNFL thickness from fundus photographs was 88.7 µm for converters and 92.1 µm for non-converters. Moreover, the mean rate of change for predicted RNFL thickness was faster for converters at –1.02 µm per year compared with –0.67 µm per year for non-converters (P < .001).
“The performance was very similar to that of measurements obtained by much more sophisticated devices such as OCT,” Medeiros previously told OSN. “Therefore, the use of AI enables one to get predictions from photographs which rival those obtained from OCT in terms of detecting glaucomatous damage in suspects. This is an outstanding result given that photographs are a much cheaper and easily available test.”
Other research by Schuman and colleagues is focusing on using the information from OCT to detect glaucoma, predict progression and detect when glaucoma progresses.
“We are also looking at ways to decrease the variability of glaucoma testing,” Schuman said. “For example, we found that we could decrease testing variability or at least have lower variability. We predicted what the visual field performance would be using OCT and AI. There is a potential to not only have less variability, but maybe to someday do away with visual field testing altogether, which would be very nice for our patients.”
Schuman said he is most looking forward to seeing data from global research groups on transitioning from disc photos to RNFL thickness.
“This is really fascinating work,” he said. “We know that there is information in a disc photo that we are not able to interpret as human beings but computer algorithms are able to analyze and interpret. AI holds the potential to unlock power in those photos.”
Challenges
While AI holds great promise in glaucoma detection and management, some potential limitations and challenges remain, for example, a lack of ground truth in terms of diagnosing glaucoma.
“To train and validate AI systems, we need ground truth diagnosis,” Chiang said. “The challenge here is what the computer scientists call ‘garbage in, garbage out.’ If we do not have ground truth, then we cannot develop a good system.”
Chiang said there are several challenges to obtaining ground truth in glaucoma.
“For one, there is often significant variability among experts in terms of making a glaucoma diagnosis, which makes it very hard to build automated systems. A second challenge is what I call a lack of consensus about the process of diagnosis, which means that of all the different modalities for making a diagnosis, such as optic disc morphology, OCTs and visual fields, there is lack of agreement about the diagnostic process — no one consensus exists,” Chiang said.
The glaucoma community needs to work together to develop standards for what that ground truth is for glaucoma diagnosis. This would be a step forward in advancing how to build AI systems to diagnose glaucoma, Chiang said.
Looking ahead
The research landscape for AI in glaucoma care is vast, and the technology for glaucoma diagnosis and management could manifest in many ways.
“Within the next 5 to 10 years, my hope is that researchers will develop systems that can be used clinically to identify patients who are at risk for glaucoma as well as at risk for rapid progression of disease,” Chiang said. “Another important area for future research is for AI systems to help reduce health care disparities in this country. Medically underserved patients often do not receive the glaucoma screening or management that they need, and some of these patient communities are also at high risk for glaucoma.”
Schuman said he expects that AI will provide more access to glaucoma evaluation to identify more people who have the disease.
“AI may also reduce the number of people who present to the clinician for glaucoma assessment. Instead, there may be an option for remote analysis by a device that is able to capture an image that may be analyzed by the AI algorithm to diagnose the glaucoma,” Schuman said. “This would offer the ability for glaucoma evaluation to more people in the community. It may even be at a lower cost for both the patient and the health care system. In addition, we will be able to determine the probability of disease progression or which patients will progress quickly vs. those who will not. We can imagine if we were able to predict which patients will progress slowly, then that patient may be scheduled for longer intervals in between follow-up.”
Although much work is still needed, some of the innovations are right around the corner, Schuman said.
“Having the ability to determine who will be diagnosed with glaucoma vs. those who will not is a fairly straightforward application for AI,” he said. “It could be independent of the clinician’s office, perhaps. One could imagine that the technology could exist at a primary care physician’s office or maybe even at the drug store where individuals can get tested and determine whether or not it is likely that glaucoma is present and whether or not it’s a good idea to see an ophthalmologist for it.”
Richter said that AI algorithms could become a part of OCT or visual field packages to help ophthalmologists determine whether a patient has glaucoma.
“Far more powerful than that, or perhaps in the more distant future, is the use of AI in glaucoma screening in the community,” Richter said. “Many patients do not have access to an ophthalmologist, and many do not have a reason to have a regular eye exam, especially if they do not already wear glasses. Glaucoma often goes undiagnosed until it is very advanced. It is often called the ‘silent thief of sight’ because most do not present with symptoms until it is very advanced. If we had AI that could detect glaucoma from something as simple as an iPhone ‘selfie’ of the retina, for example, then there is potentially a great amount of blindness that could be prevented with AI. This could be huge and would be one way that AI could manifest in the glaucoma world and make a huge impact.”
- References:
- Christopher M, et al. Ophthalmology. 2020;doi:10.1016/j.ophtha.2019.09.036.
- Jammal AA, et al. Am J Ophthalmol. 2020;doi:10.1016/j.ajo.2019.11.006.
- Linnehan R. AI-enabled radar a new tool for diagnosing, monitoring glaucoma. www.healio.com/news/ophthalmology/20191203/aienabled-radar-a-new-tool-for-diagnosing-monitoring-glaucoma. Published Dec. 5, 2019.
- Linnehan R. Deep learning algorithm’s prediction of RNFL thickness gauges risk for glaucoma conversion. www.healio.com/news/ophthalmology/20210316/deep-learning-algorithms-prediction-of-rnfl-thickness-gauges-risk-for-glaucoma-conversion. Published March 16, 2021.
- Linnehan R. Four AI models show high diagnostic accuracy for glaucoma. www.healio.com/news/ophthalmology/20200309/four-ai-models-show-high-diagnostic-accuracy-for-glaucoma. Published March 9, 2020.
- For more information:
- Michael F. Chiang, MD, can be reached at National Eye Institute, Building 31, Room 6A03, 31 Center Drive, MSC 2510 Bethesda, MD 20892; email: michael.chiang@nih.gov.
- Grace M. Richter, MD, MPH, can be reached at USC Roski Eye Institute, Keck Medicine of USC, 1450 San Pablo St., Suite 4700, Los Angeles, CA 90033; email: grace.richter@med.usc.edu.
- Joel S. Schuman, MD, FACS, can be reached at NYU Langone Health, NYU Grossman School of Medicine, 222 East 41st St., Suite 468, New York, NY 10017; email: joel.schuman@nyu.edu.
- Linda Zangwill, PhD, can be reached at Shiley Eye Institute, University of California, San Diego, 9500 Gilman Drive, La Jolla, CA 92093; email: lzangwill@health.ucsd.edu.
Click here to read the Point/Counter to this Cover Story.
Collapse
Read more about
Click Here to Manage Email Alerts