AI in retina moving toward real-world practice
Click Here to Manage Email Alerts
Retinal diseases are a major cause of severe vision loss, with numbers constantly increasing due to the aging of the population, the diabetes epidemic and the growing prevalence of high myopia.
At the same time, science is progressing, with higher efficacy and durability of pharmacotherapies, breakthrough therapies recently approved and novel approaches in the pipeline. To meet the challenges of this exponential increase in the number of patients needing diagnosis, monitoring and treatment, a transformative approach is required.
Source: Ursula Schmidt-Erfurth, MD
“The community has understood that artificial intelligence is needed to be able to manage this overwhelming burden and that AI is fit for purpose, with the first pioneering tools available for patient management,” Ursula Schmidt-Erfurth, MD, professor and chair of the department of ophthalmology at the Medical University of Vienna, said.
Retinal imaging, with OCT in the forefront, generates massive volumes of data, the largest and highest resolution of all medical imaging modalities. This makes the retina a favorable ground for digital innovation and artificial intelligence.
“Artificial intelligence can reliably do all the tasks you train it for. It can recognize the clinically relevant biomarkers, not only recognize them as humans do, but detect them, localize them respective to their layer and quantify them in their three-dimensional extension. This is most important for treatment in the sensitive area of the macula. We do have the drugs. We do have the diagnostic hardware, which is OCT. And now we add AI to take full benefit of this combination for the sake of best care of our patients,” Schmidt-Erfurth said.
Monitoring fluid
The pioneering research of Schmidt-Erfurth and her team was pivotal in the development of two AI-based medical devices developed by RetInSight, a spin-off of the Medical University of Vienna: the Fluid Monitor and the GA Monitor.
Fluid is the most important biomarker in the diagnosis of neovascular age-related macular degeneration and in the monitoring of disease activity and anti-VEGF treatment efficacy.
“Human experts can mostly see whether there is fluid, but it takes a huge amount of time to look through the entire OCT data set, and quantifying changes over the follow-up is not feasible at all. We are treating hundreds of thousands of patients on a daily basis and need something that is reliable and precise, that can be used by a mouse click, accessible for everybody, something that speeds up the workflow because we have to take care of so many patients, and undertreatment is a big problem in the real-world setting,” Schmidt-Erfurth said.
The Fluid Monitor has a clear advantage over human evaluation in the ability to characterize fluid subtypes and measure fluid three dimensionally as volume rather than overall thickness of the entire retina.
“What we have been using so far is just a semiquantitative measurement, which is central retinal thickness (CRT), but CRT is just a random cut through the diseased macula. It does not differentiate between different fluid types and cannot correctly measure the amount of fluid, which reflects disease activity. When we correlated CRT with true fluid volumes, we found that the correlation was below 0.5, which is extremely low in neovascular AMD. CRT should not be used for re-treatment indications,” Schmidt-Erfurth said.
The RetInSight Fluid Monitor has gained MDR class 2 approval as Clinical Decision Support Systems for the monitoring of patients with neovascular AMD and can be used in clinical routine in Europe. In the U.S., it is available for investigational use to take measurements of fluid with institutional review board approval and patient consent.
Currently, it is linked to the Spectralis HEYEX 2 platform (Heidelberg Engineering), and within a few months, the Topcon Triton OCT as well as the Cirrus device (Zeiss) will also be able to be analyzed by the tool.
“The algorithm is device independent but has to be adjusted to each device separately, such as CRT,” Schmidt-Erfurth said.
Predicting identification of GA progression
The GA Monitor, also approved for clinical use in Europe, allows complete assessment of disease activity, identifying patients who will progress, not progress or progress slowly and empowering the specialist in making decisions about treatment need and benefit. In contrast to fluid, geographic atrophy (GA) activity cannot be detected on an OCT by clinical assessment.
“Typically, in geographic atrophy the photoreceptors (PR) are thinning out first, and the AI-based OCT can clearly visualize the area of diseased photoreceptors. The retinal pigment epithelium (RPE) loss follows the path of the PR loss in a second step. Patients who have high disease activity and will progress have a large zone of photoreceptors thinning already around the classical GA lesion seen clinically by RPE defects. These patients need treatment, and there is no reason to wait for further and irreversible disease progression and functional loss,” Schmidt-Erfurth said.
It also monitors the effects of therapies, allowing for prompt detection of side effects, such as the formation of intraretinal fluid, inflammation or occlusive disease.
“The potential impact on clinical routine is huge,” Schmidt-Erfurth said. “Now that treatments for GA are available, we have to take care of these large patient populations in a way that is most beneficial for them, avoiding lifelong injections without any benefit and helping the health care systems to manage the socioeconomic burden in a sustainable manner.”
When anti-VEGF treatment became available, the rise of OCT usage was enormous and fast, and OCT became an indispensable tool to guide anti-VEGF therapy in millions of individuals. Now that complement inhibitors have become available as therapies for GA, AI-based tools will become equally indispensable at an even larger dimension, she said. Moreover, early detection is essential as the GA process is irreversible and lesions start outside of the fovea, for a long time not causing visual symptoms. Fully equipped automated detection on OCT in a community-based manner will be most helpful to avoid this blinding process in the wide population.
More accurate screening of DR
In the area of diabetic eye disease, an AI algorithm developed by the Wisconsin Reading Center allows for automated identification of patients eligible for enrollment in diabetic retinopathy (DR) clinical trials.
“It was very difficult for clinicians to identify these patients, and screening failure rates were as high as 50%. Our algorithm was built on a small pilot data set. It functions pretty well, and accuracy wise, it is about 75%,” Amitha Domalpally, MD, PhD, research director of the Wisconsin Reading Center, said.
The DRCR Retina Network recently approved the training of the algorithm on its data set to produce a much more robust model to be released this year.
The Wisconsin Reading Center was the first to be established for DR studies nearly 50 years ago when clinical trials were new to ophthalmology. It was pivotal in the development of the ETDRS scale and nowadays provides the reference standard for clinical trials in FDA validation of AI algorithms aimed at population-based DR screening.
Three population screening models, which screen for referable DR, are currently FDA approved: the IDx-DR (Digital Diagnostics), the EyeArt (Eyenuk) and AEYE-DS (AEYE Health).
“Many more are available globally that go through the same rigorous validation process in their own country, including smartphone-based applications,” Domalpally said.
A long evaluation process
The Singapore Eye Lesion Analyser (SELENA+) is another algorithm for DR screening that has proved in studies to be faster and equally accurate compared with human graders.
“The results of the pilot had been published a few years ago, and currently this algorithm is under evaluation for clinical use,” Gemmy Cheung, MD, professor at Duke-NUS Medical School in Singapore, said. “The reason why it is still not fully deployed is that evaluations in phases are required to ensure the accuracy of such algorithm for any national program. We need to be very certain about the positive and negative predictive values and the rates of false positive and false negative in particular. It is therefore necessary to look at large numbers to understand possible eventualities which we may or may not have imagined with a smaller sample size in a research setting.”
In addition, the large number of images to be processed requires a lot of computing power and a large storage capacity, and while this system may free up the manpower for human grading, human IT specialists need to be employed.
“So, there is a little bit of financial calculation that needs to be worked out for the sweet spot of the most cost-effective manner of deploying it,” she said.
Another project that involves AI technology at her university is the use of the Notal OCT Analyzer (Notal Vision) in AMD treatment trials.
“Traditional anatomical biomarkers are central subfield thickness, central retinal thickness, and presence or absence of fluid. But with computing power, we can now analyze fluid at volumetric levels, and we can also classify them according to compartment,” Cheung said.
“Having the ability to assess the whole macular volume scan is particularly valuable in the polypoidal choroidal vasculopathy subtype, as the maximal activity often affects parafoveal regions and not necessarily the central subfoveal area. Having this method to evaluate volumetrically is very important to us. In addition to subretinal fluid and intraretinal fluid, we also pay a lot of attention to pigment epithelium detachments in polypoidal choroidal vasculopathy. We have published several reports on the utility of automated volumetric measurements of retinal fluid compartments at baseline and following treatment,” she said.
Not yet AI, but machine learning
Philip J. Rosenfeld, MD, PhD, professor of ophthalmology at Bascom Palmer Eye Institute, University of Miami, has been involved in the development of OCT algorithms for identifying and quantifying disease characteristics and the different stages of disease progression in AMD. The aim is to predict and prevent further worsening with timely treatment.
“The focus in our field has been on late-stage disease, both wet macular degeneration and geographic atrophy,” he said.
However, there is still a long way to go before what is called “artificial intelligence” will be able to fulfill the promises that the term suggests, he said.
“I don’t like to use the term AI because we are not really using AI. We are using machine learning algorithms to automate the identification of features that we have already identified as being important in predicting disease progression,” he said.
In fact, AI has failed to pick out features that had already been identified by human observation, he said, mentioning as an example the double-layer sign, one of the most predictive features of progression from dry to wet AMD.
“Years ago, we identified it using swept-source OCT angiography, but every attempt of using AI to unveil high-risk characteristics that predict future exudation has failed to pick out these lesions. So, I hold this up as an example of how AI just hasn’t delivered the promise of identifying features that we didn’t otherwise identify based on our own observational skills. I think it will happen, but we are not there yet,” Rosenfeld said.
He is confident that machine learning algorithms today, and AI in the future, hold great potential for improving the quality and efficiency of patient care. They will help avoid misdiagnosis and provide non-retina specialists with the ability to identify patients at risk and then refer those patients to experts.
“They will allow us to make better use of our time and provide information to patients that predict future visual acuity benefits from a particular treatment. Patients will benefit, physicians will benefit and payers will benefit because there will be more efficient use of expert opinion and care. The whole health care system will benefit from machine learning algorithms and artificial intelligence. But right now, we’re not at the point of relying entirely on AI to decide who should get treated. There is no replacement for human intelligence and the importance of having a conversation with patients to decide whether a treatment should be given,” he said.
Caveats to consider
There are also caveats that need to be appreciated when using algorithms in clinical practice, Rosenfeld said.
“A lot of groups that develop AI algorithms think that just because they train the algorithm on given images from a particular disease, then they can broadly apply that algorithm to everyone with that disease. That’s just not the case. We have been developing machine learning algorithms to identify specific features, and we are constantly having to modify them because we find cases where they fail,” he said.
Another limitation is that algorithms are currently, for the most part, instrument dependent and technology dependent.
“One of the things we are doing now is trying to develop algorithms that work on both the swept-source and the spectral-domain technologies, that are instrument agnostic. If we develop an OCT algorithm, it should be applicable to whatever OCT technology we are using,” Rosenfeld said.
The advantage of AI is not just to reduce human labor but to enable work that is just not humanly possible to do, such as delineating the fluid at every scan, Cheung said.
“We are talking about 60 scans at least in a macular volume and doing that at every visit for every single patient. Even the most diligent fellow would find that extremely hard. And the algorithm is very accurate. It is less subjected to inter-visit variability,” she said.
However, in the process of translating AI to clinical practice, there are still questions that need to be answered.
“There are concerns whether the training data set is representative enough for where it is going to be deployed and whether it is applicable to another population or scans that are performed by another instrument,” Cheung said. “As clinicians, since there are multiple concurrent efforts developing at the same time, we can be a little lost about which one we should we go for and, as new data comes in, how often these algorithms should be updated.”
Because algorithms are trained to perform a specific task, another question is whether the algorithm would be able to detect warning signs that are not related to the condition in question.
“Back in diabetic screening, if there is, for example, a melanoma, would the AI algorithm be able to pick it out? Although it’s going to be a very rare event, we don’t want to miss even one single case,” she said.
AI is a dynamic field, and this is just the beginning, Schmidt-Erfurth said. Algorithms will constantly be refined, and a lot of new indications and new therapeutic strategies will be reached by AI.
“Our path is to perform extensive internal and external validation on study data and real-world data in a manner which aligns to regulatory requirements. This is how, over more than 10 years of work, we were able to provide the first approved AI-based tools,” she said.
At every step of this development, however, documentation, quality control and standardization are key.
“AI is the new kid on the block and has to undergo standardization and quality control. There is a big hype about AI, and there are more and more papers and groups that present their own algorithms. But to have an algorithm certified as CDS system, we have to bring this through regulatory approval, and this process has to start from the beginning. It is a very complex, systematic process, and very few groups are able to do this,” she said.
Overcoming bias, improving access
Preventing and overcoming bias in algorithms are other challenges, according to Domalpally. Because AI only knows what it is trained on, biases are an issue, and the applicability of algorithms to populations of different ethnicities and different countries, as well as to different imaging systems, will require long processes of standardization and head-to-head comparison.
Retinas from white individuals typically exhibit lighter pigmentation compared with those from Asian or African descent, she said. Because most of the clinical trials have a large majority of white population, human graders have been trained for many years on retinal images that belong to the white population and may not have the same confidence in detecting disease features in pigmented retinas.
“This problem has come to light now with AI, but as reading center staff, we have seen this for many decades. We complain about AI and its bias, but AI just learns from humans. It’s the humans who have the bias,” she said.
Grader training in racially diverse data sets is needed, but for now it is important to consider that an algorithm trained on a predominantly white population will not function in a pigmented retina.
“A U.S. FDA-approved algorithm may not work in another country,” she said.
However, AI has great potential to improve access to care and reduce health disparities. Patients with diabetes often do not reach the ophthalmologist until the eye is greatly compromised. AI-based screening through simple tabletop or smartphone-based cameras could help detect referable diabetic eye disease at the primary care provider’s office or in endocrinology clinics.
“Studies have shown that even with the diagnosis, not everybody goes to the ophthalmologist. But without the diagnosis, nobody will go,” Domalpally said.
The impact can be huge in countries with a large population, such as India and China, where smartphone-enabled imaging is well established.
“In India, even 15 to 20 years ago, they had very well-established field screening programs for cataract and diabetic retinopathy. Adding AI is almost like giving steroids to those programs,” Domalpally said.
- References:
- Banerjee I, et al. Sci Rep. 2020;doi:10.1038/s41598-020-72359-y.
- Chu Z, et al. Biomed Opt Express. 2022;doi:10.1364/BOE.449314.
- Coulibaly LM, et al. Eye (Lond). 2023;doi:10.1038/s41433-022-02154-8.
- Coulibaly LM, et al. Ophthalmol Retina. 2023;doi:10.1016/j.oret.2023.05.004.
- Daich Varela M, et al. Graefes Arch Clin Exp Ophthalmol. 2023;doi:10.1007/s00417-023-06052-x.
- Domalpally A, et al. Lancet Digit Health. 2021;doi:10.1016/S2589-7500(21)00140-0.
- Gerendas BS, et al. Retina. 2022;doi:10.1097/IAE.0000000000003557.
- Ipp E, et al. JAMA Netw Open. 2021;doi:10.1001/jamanetworkopen.2021.34254.
- Jampol LM, et al. JAMA Ophthalmol. 2020;doi:10.1001/jamaophthalmol.2019.4782.
- Jiang X, et al. Transl Vis Sci Technol. 2021;doi:10.1167/tvst.10.4.11.
- Keenan TDL, et al. Am J Ophthalmol. 2021;doi:10.1016/j.ajo.2020.12.012.
- Keenan TDL, et al. Ophthalmology. 2021;doi:10.1016/j.ophtha.2020.06.038.
- Kihara Y, et al. Ophthalmol Sci. 2022;doi:10.1016/j.xops.2022.100197.
- Loo J, et al. Ophthalmol Sci. 2023;doi:10.1016/j.xops.2023.100292.
- Mai J, et al. Sci Rep. 2023;doi:10.1038/s41598-023-34139-2.
- Mares V, et al. Br J Ophthalmol. 2023;doi:10.1136/bjo-2022-323014.
- Pawloff M, et al. Eye (Lond). 2023;doi:10.1038/s41433-023-02615-8.
- Pawloff M, et al. Retina. 2022;doi:10.1097/IAE.0000000000003385.
- Pramil V, et al. Ophthalmol Retina. 2023;doi:10.1016/j.oret.2022.08.007.
- Reiter GS, et al. Retina. 2021;doi:10.1097/IAE.0000000000003023.
- Reiter GS, et al. Ther Adv Ophthalmol. 2022;doi:10.1177/25158414221083363.
- Riedl S, et al. Ophthalmol Retina. 2022;doi:10.1016/j.oret.2022.05.030.
- Rim TH, et al. Br J Ophthalmol. 2021;doi:10.1136/bjophthalmol-2020-316984.
- Rivail A, et al. Biomed Opt Express. 2023;doi:10.1364/BOE.487206.
- Schlegl T, et al. Ophthalmology. 2018;doi:10.1016/j.ophtha.2017.10.031.
- Schmidt-Erfurth U, et al. Invest Ophthalmol Vis Sci. 2018;doi:10.1167/iovs.18-24106.
- Schmidt-Erfurth U, et al. Prog Retin Eye Res. 2018;doi:10.1016/j.preteyeres.2018.07.004.
- Schmidt-Erfurth U, et al. Prog Retin Eye Res. 2022;doi:10.1016/j.preteyeres.2021.100972.
- Vogl WD, et al. Ophthalmol Retina. 2023;doi:10.1016/j.oret.2022.08.003.
- Waldstein SM, et al. JAMA Ophthalmol. 2020;doi:10.1001/jamaophthalmol.2020.1376.
- Zafar S, et al. Curr Diab Rep. 2022;doi:10.1007/s11892-022-01467-y.
- Zhou H, et al. Ophthalmol Retina. 2020;doi:10.1016/j.oret.2019.09.012.
- For more information:
- Gemmy Cheung, MD, of Singapore Eye Research Institute, Singapore National Eye Centre, Singapore, can be reached at gemmy.cheung.c.m@singhealth.com.sg.
- Amitha Domalpally, MD, PhD, of Wisconsin Reading Center, Madison, Wisconsin, can be reached at domalpally@wisc.edu.
- Philip J. Rosenfeld, MD, PhD, of Bascom Palmer Eye Institute, Miami, Florida, can be reached at prosenfeld@med.miami.edu.
- Ursula Schmidt-Erfurth, MD, of the Medical University Vienna, Department of Ophthalmology, Vienna, Austria, can be reached at ursula.schmidt-erfurth@meduniwien.ac.at.
Click here to read the Point/Counter to this Cover Story.
Collapse
Click Here to Manage Email Alerts