AI in Eye Care: Past, Present, and Future

With a growing global population and its increasing number of elderly and myopic patients, screening for sight threatening conditions is becoming more vital. In ophthalmology and optometry, artificial intelligence has the power to make this process cheaper, easier, faster and more effective. AI is reshaping the way professionals diagnose, manage, and treat eye diseases (most notably diabetic eye disease) (1). With deep learning algorithms, predictive analytics, and image processing capabilities, AI is improving patient outcomes while streamlining workflows for clinicians (2).

But how did AI get to this point, and where will it go next?

A brief history

In the 1820s, Charles Babbage developed his mechanical “Difference Engine,” a machine designed to automatically calculate equations. While Babbage’s device is recognized as a predecessor to the modern computer, it wasn’t until over 100 years later, with the advent of electronic computing and the microprocessor, when artificial intelligence (AI) really began to evolve.

The concept of neural networks was first introduced in the early 1940s during World War II when Alan Turing and his team helped to crack the Nazi Enigma code (3), with Turing’s systems eventually leading to the first real computers. A decade later, in 1950 Turing posed the question, “Can machines think?” (4). His“Turing Test” proposed that if a person communicated with a computer but was unaware they were conversing with a machine, then this was artificial intelligence in action.

The term “artificial intelligence” itself was coined by American computer scientist, John McCarthy, at the Dartmouth Conference in New Hampshire in 1956 (5). Chess-playing computer algorithms were developed in the 1960s; these were “rule-based systems” that could assist with decision making (6). The first industrial assembly-line robot was introduced by General Motors in 1961, and then, in 1964, Joseph Wizenbaum developed a natural language processing computer program called ELIZA, which simulated human conversation (but didn’t pass the Turing Test) (7).

By the 1990s, IBM’s Deep Blue was able to beat world chess champion Garry Kasparov in a widely publicized event in 1997. Apple launched Siri in 2011, allowing users to speak with their smartphones. In 2014, a chatbot named Eugene Goostman was able to beat the Turing Test by convincing a panel of judges they were talking to a human (8). And in 2016, the AlphaGo program beat Lee Sodol, the world champion Go player. (Go is much more complex and difficult than chess.) To win, AlphaGo discovered a new and unique set of moves that no (human) player had ever tried before.

Most recently, of course, we have seen the arrival of OpenAI’s ChatGPT, alongside China’s rival lower-cost AI model, Deep Seek.

AI in eye care

AI’s role in ophthalmology dates back to the 1980s with the development of early automated screening methods for diabetic retinopathy (DR) (9). Then in the early 2000s, machine learning algorithms were introduced to analyse retinal images, paving the way for AI-based diagnostic tools (10).

A landmark moment came with the development of convolutional neural networks (CNNs), which enabled AI systems to process vast amounts of imaging data. In 2016, Google’s DeepMind collaborated with Moorfields Eye Hospital in London to develop an AI system capable of detecting eye diseases – such as diabetic retinopathy and age-related macular degeneration (AMD) – with expert-level accuracy (11).

Current eye care applications: AI-powered diagnostics and screening

Eye care practitioners (ECPs) now have access to various AI-based systems that can aid them in preventing vision loss in their patients, including:

• IDx-DR – an autonomous system used for detecting diabetic retinopathy (DR). This system is now known as LumeticsCore, and automatically analyzes digital fundus images to detect DR and determine the severity of the disease.

• Eyenuk EyeArt – a highly sensitive tool for determining stages of DR and macular edema. An EyeArt screening service (EyeScreen) also exists as part of the DESP (diabetic eye screening program) in many areas, replacing or assisting clinicians in monitoring the disease.

• RetinAI Discovery – a comprehensive platform that has a strong set of ophthalmology tools, but is also linked to pharma and life sciences in general. The optical coherence tomography (OCT) AI helps clinicians to interpret retinal conditions and spot subtle lesions. (One thing I particularly like is the geographic atrophy (GA) progression predictor. [Figure 1]). The system is also useful in accelerating clinical trials of new retinal treatments (13).

• Altris AI - a cloud-based OCT interpretation tool for detecting a large array of retinal lesions, aiding in detecting pathologies more accurately. One element of significant interest is the glaucoma risk assessment module (Figure 2). With the significant issues of false positive referrals for glaucoma (partly due to issues with normative databases on OCTs), this may help to reduce the burden on secondary care.

• AEYE – DS – produced by AEYE Health, this is another DR screening tool for clinicians.

• RetinaLyze is a Danish based company that has a well-established customer base for screening of fundus images, for diabetes, AMD and glaucoma. It has now incorporated OCT analysis.

• RetinSight has some interesting developments, including a GA monitor and fluid monitor to help understand the effectiveness of therapies. The company also has a range of solutions to help detect and monitor wet AMD, retinal vein occlusions and diabetic macular oedema. Figure 3 shows an example of RetinSight’s colour fundus disease screening software. The left image is the “clean” image without the algorithm processing. The right image shows an example of what the algorithm has located, including diabetic lesions.

• Thirona RetCAD offers a range of customized or large-scale screening options for color fundus images. The company has projects aimed at improving DR screening in populations where diabetes is highly prevalent, but DR screening is limited, especially in remote areas. Its system appears highly tailorable.

Current eye care applications: AI within hardware and software platforms

There are, of course, numerous manufacturers of imaging devices across numerous modalities, and many of these companies have proprietary software for viewing and manipulating the images, as well as electronic medical records (EMRs) or platforms designed to integrate with EMRs:

• Zeiss Forum brings together results from visual fields, OCT, colour fundus and other data. Not only is this a form of ERM, but it uses AI to analyse all the data for clinical diagnosis of glaucoma and progression analysis. Using DICOM, it can import data primarily from Zeiss products. You can use non-DICOM interfacing. There is limited third-party device and data compatibility. Retina workplace assists with other retinal conditions.

• Topcon Harmony again stores and connects data from multiple diagnostic devices via DICOM and non-DICOM integrations. It is device-agnostic, so multiple devices from differing manufacturers can be connected. This allows AI data analysis and progression analysis to take place without the need for one specific range of devices.

• Visionix Nexus can again bring in multi modalities and is device-agnostic. Depending on the data you have, it can screen for a range of conditions or assess and monitor diagnosed conditions and provide a range of reports.

• Heidelberg Heyex is one of the most well-known data integration and AI systems. It is optimized for Heidelberg OCT, perimetry and other devices, but there is some limited third-party DICOM integration.

AI is also being used in the planning of surgical procedures. Companies like ForSight Robotics offer robotic assistance in various ophthalmic surgery settings, such as retinal detachment surgery (14). Other systems being developed include Verily from Alcon, which uses ML tools for planning and optimizing cataract surgery and premium IOL planning. Heru’s OphthAI assists with planning and performing retinal procedures with real-time guidance and tissue identification.

Beyond that, AI could help with streamlining patient waiting lists in secondary care and with telemedicine. Clinicians could provide remote assessment of patients with AI assisting with image analysis.

In the future, AI could further integrate multiple diagnostic and surgery systems, including head-up displays in microscopes. It could assist in planning personalized treatment plans for patients, with tailored therapies based on deep learning and post-therapeutic feedback loops from large datasets. Predictive analytics and prognostic systems could help identify who is most at risk and requires urgent intervention.

Limitations and risks

• GDPR – As large datasets are used in deep learning, there are many concerns about remaining within the auspices of GDPR. There is the potential that data security and privacy rules could be infringed by using such datasets.

• Study size and spread – It’s important that substantial and diverse cohorts are used in datasets so as to avoid issues with incorrect approaches to specific ethnicities with decision making, particularly in diseases like glaucoma where prevalence varies dramatically through ageing across gender and specific populations.

• Reporting and matching – There are calls for new reporting standards to ensure reporting of study results is improved.

• Regulatory issues – When AI becomes more advanced and starts to be more autonomous in making diagnoses, the class of software and hardware will change. This will lead to much more expensive, long winded and detailed verification assessments with the FDA, CE, MHRA, etc. This may hinder development of systems, or it could even alter the way AI is used.

• Legal issues – If AI continues to take over more and more responsibilities and tasks from clinicians, who will be responsible for errors in judgement? Will the supplier of the AI be responsible, or will it always sit in the lap of the clinician? As the autonomy of AI in ophthalmology increases, will legal arguments tend toward the AI being held viewed as a responsible entity?

• Bad actors – This is the most concerning area of AI. Many key opinion leaders in AI worry that bad actors could create new diseases and rogue proteins leading to pandemics. Could this creep into ophthalmology? And what about hacking? Imagine terrible backlogs and clinical errors occurring after major hacking bringing down entire health service systems.

• Losing control – As AI develops in complexity and autonomy, it may find patterns and connections that we simply don’t know exist. For example, AI can detect a patient’s gender simply from analysing their retinal images. What could this mean for the future?

Conclusion

Enhanced diagnostics will hopefully improve patient outcomes and reduce late-stage diseases occurring in a patient’s lifetime. Waiting lists for assessments, follow-up and treatments will likely be reduced. One hopes that the scope of practice will also increase for many clinicians as AI helps them work in new areas of the field.

Risks do exist and there are some workers who may lose their jobs. AI is unique in that it will not require new types of jobs to support it as it develops (unlike in the Industrial Revolution, for example). Many ECPs and related staff may have to retrain as AI does more of the tasks they were trained to perform.

However, AI will also save lives, reduce sight loss, and make many people’s lives easier and better. And maybe ECPs will even be able to take more holidays in the future – with AI helping them decide where to go!