How iCare ILLUME Uses AI to Catch Diabetic Retinopathy Sooner

Spend a day inside a diabetic eye screening service, and you’ll see two things: lots of very busy people, and an “unavoidable” bottleneck of images waiting to be graded. 

Diabetic retinopathy (DR) screening takes approximately twenty minutes, from imaging to grading, meaning a single grader might work through anywhere between fifty and two hundred images in a shift. 

It is careful, technical, repetitive work, and the queue only ever seems to grow. In some services the wait between a patient being screened and their result arriving sits at around three weeks.

Demand is growing, but budgets and capacity aren’t. 

With an estimated six million people in the UK now living with diabetes, the highest number on record, accurate, rapid DR screening is becoming mandatory.

This is exactly the problem iCare set out to solve with their ILLUME software. Part of a wider portfolio of iCare ophthalmic technology, ILLUME combines retinal imaging with AI-supported analysis to make diabetic eye screening faster and more efficient.. By factoring a clinically validated AI engine into the screening process, ILLUME is able to grade retinal images in under a minute, flagging the patients who need a closer look, while allowing human graders to work through significantly more images in a shift. 

This isn’t about replacing the clinician; it’s about improving efficiency across an already stretched healthcare system. Today, we’ll be talking about iCare’s ILLUME, and how it’s reshaping the way we screen for DR.

Suffering from success: the DR screening story

Around 22% of people with diabetes have some degree of diabetic retinopathy, and for 6% of them, it will be vision-threatening.

Caught in time, it is one of the most treatable causes of sight loss there is, which is exactly why early detection is so important. Treatments such as laser photocoagulation and intravitreal anti-VEGF injections work well, but they work best before a patient ever notices a change in their vision, meaning the sooner you can get a diagnosis, the better.

The NHS Diabetic Eye Screening Programme has offered annual digital screening to everyone with diabetes since 2003, and it’s widely considered a successful programme. A 2021 study showed that, for the first time in at least five decades, diabetic retinopathy stopped being the leading cause of certifiable blindness among working-age adults in England.

This is great news, of course, but it’s not without its drawbacks: the NHS is not equipped to handle the growing demand for screening, at least not with existing tech. There are a rising number of patients and a finite number of trained graders; something has to give. 

AI is being used to protect the DR screening process and ensure that its success continues well into the next five decades. 

How does iCare ILLUME support with DR screening?

The iCare ILLUME AI solution screening layer that sits on top of the iCare DRSplus fundus camera and the RETCAD AI software developed by Thirona. 

The DRSplus captures TrueColor confocal images across a 45-degree field, and because it is confocal, it can produce a gradable image through cataract and other media opacities, with pupils as small as 2.5mm, needing no dilating drops. For each eye, the operator takes two images, one centred on the macula and one on the optic disc. Unlike retinal OCT, which provides cross-sectional imaging of retinal structures and is invaluable for investigating conditions such as diabetic macular oedema, the DRSplus is designed for rapid fundus photography and DR screening.

RETCAD’s AI then gets to work. Before it scores anything, it runs a quality check on each image, flagging any that are too poor to read so the operator can retake them there and then before the patient is sent home (much better than discovering the problem weeks later!). 

If the pictures are adequate, it will assess them, returning a severity score for each eye on a scale from 0.00 to 5.00, mapped to the international ICDR grades. The patient’s overall result is the higher score of their two eyes.

While a human grader often assigns a whole number, RETCAD gives you a more accurate decimal reading. For example, where RETCAD might give you a “2.31”, a human grader might simply elect to give you a “2”. This specificity is designed to give clinicians more information on how borderline a case is. 

Alongside the number, the software produces heatmaps that pick out suspicious areas, one tuned to bright lesions such as exudates, and another to red lesions such as microaneurysms and haemorrhages. It presents the result on a red-to-green scale that can be used to more easily explain issues to a patient. 

An example output of iCare ILLUME, helping clinicians explain DR to their patients. 

The whole analysis comes back to the operator’s browser in well under a minute, with nothing to install, and the report can be stored, printed or pushed into the patient’s electronic record. Images are pseudonymised before they leave the practice and handled in a GDPR-compliant, region-specific Amazon Web Services environment over an encrypted connection.

In short, iCare ILLUME screens images in seconds rather than minutes, meaning you can screen more patients per day, potentially improving the vision of many people currently stuck on a waiting list. 

How accurate is iCare ILLUME?

On the publicly available Messidor-2 benchmark, RETCAD achieved an AUC of 0.981 for identifying referable retinopathy. A real-world study at the Ramón y Cajal University Hospital in Madrid, where 3,189 patients attending routine screening had their images assessed by both RETCAD and a panel of expert human graders, found that sensitivity came in at 90.5% and specificity at 97.1%. 

AI was shown to be ideal for initial triage, as it only missed six referable cases across more than 6,000 images, none of which were severe.

Of the 274 images RETCAD flagged as referable, two trained human graders reviewed the same photographs and reached different conclusions. One marked 86 as referable. The other marked 150. 

AI is capable of noticing DR significantly earlier, and it does not grow fatigued over time. It will make the same call at 4pm as it does at 9am. 

When RETCAD was evaluated specifically on iCare DRSplus images, it returned an AUC of 0.989 with 97.5% sensitivity and 92.9% specificity. In that same dataset, it recorded 100% sensitivity for vision-threatening proliferative disease. No sight-threatening case was missed.

How ILLUME is enabling clinicians, not competing with them

The initial reaction among graders, understandably, tends to be negative; that this is the beginning of the end of their careers. However, in reality, it’s more likely the beginning of something brand new and far, far more efficient. 

As the diabetic population grows, the same workforce simply cannot grade its way through the high volumes by hand, and the time freed by automating the routine “no disease” reads is time that can go back into screening more patients and looking after the ones in front of you.

The safeguards that protect patients today carry straight over. In current practice, ten per cent of the images a grader passes as showing no disease are independently re-checked; that same audit process is applied to the AI’s negatives, so nothing rests on the software’s word alone. 

In addition, RETCAD is a “locked” AI, meaning it can’t learn or change its behaviour in the field, making its performance fixed and known. And it never touches the part of screening that needs a person. Someone still must operate the camera, put an anxious patient at ease and explain what comes next. The AI reads the image; it does not run the clinic.

What it changes most, for the patient, is time. A result that might once have taken three weeks can be ready in days, which means those with serious disease are treated far sooner. 

For your ophthalmology department, ILLUME helps you create a cleaner, better-triaged referral flow: the people who genuinely need specialist treatment get it sooner, and those who don’t get the good news faster.

RETCAD scores retinopathy severity on the same R0 to R3 scale NHS services use, it can be deployed in the cloud or onto a hospital’s own server, and it works with the DRSplus or any fundus image taken at a 45-degree field. It is already being used by NHS trusts and by the private providers who screen on behalf of NHS England.

Want to see the difference for yourself?

While DR screening is technical and high stakes, it’s repetitive enough that it becomes a fantastic candidate for partial automation. 

It’s incredibly difficult for a screener to remain focused and engaged across a long shift. Handing the initial reading to a validated AI does not lower the standard of care, it often improves it. 

If your department is stretched thin, or you’re feeling the pain of missed diagnoses, perhaps it’s worth trialling something new. 

To learn more about iCare ILLUME and how it might fit into your screening programme, reach out to our specialists today. We can organise a demo, placing the tech directly into your workflow so you can see how it handles screening in the real world. 

Either give us a call on 0121 458 6800, or get in touch by clicking here.