Diabetic retinopathy: technology improves screening to save lives
Medical advancements – including fundus cameras and artificial intelligence – offer hope for patients who may otherwise lose their sight (or even their lives) to diabetes.
Diabetic retinopathy (sometimes called diabetic eye disease or DR) remains a leading cause of blindness worldwide. It occurs when changes in blood glucose levels cause changes in retinal blood vessels. The vessels may swell up (macular oedema) and leak fluid into the rear of the eye. Another form of the disease occurs when a lack of blood supply causes abnormal blood vessels to grow on the surface of the retina. If the diabetes and the retinopathy are left untreated it can lead to blindness.
It is estimated that at least 10% of diabetic patients in South Africa have sight threatening diabetic retinopathy and the number of diabetics with any form of retinopathy may be closer to 40%.
‘Numerous studies have shown the beneficial impact of early detection,’ says Dr Marli Conradie an endocrinologist at Mediclinic Durbanville. However, it is not feasible, taking into account the global epidemic of diabetes mellitus, that all patients are screened by an eye specialist.
Dr Conradie adds that fundus photography has revolutionised screening programmes as it makes possible to screen patients at the site of clinical contact and then send the retinal photographs to someone with the knowledge to interpret them. ‘Artificial intelligence is the next step, where the photographs will then be interpreted by a computer-generated programme and further decisions can then be made based on this,’ she adds.
‘The importance of accurate screening for diabetic retinopathy cannot be overemphasised,’ says Dr Stephen Cook, an ophthalmologist at the Eye Centre in East London. He explains that the detection of any retinopathy may indicate the patient is at a higher risk for kidney failure, strokes and heart attacks.
‘Diabetic retinopathy (DR) is an important biomarker (biological sign) for future risk of complications of diabetes and it is cost effective to detect,’ he says. ‘The presence of any DR is correlated with systemic health risks (twice the risk of coronary artery and stroke events). The extent of the retinopathy is gradable. Severe disease correlates with the risk of severe systemic disease and blindness,’ he says.
‘All doctors are trained to use hand-held ophthalmoscopes to examine the retina but it requires some skill,’ says Dr Cook. ‘It is also time-consuming because the pupil has to be dilated. With diabetic patient numbers increasing at unprecedented rates, fewer doctors are routinely examining the retina. This task has fallen to optometrists – and that’s only if a patient is even able to have regular eye exams.
‘In the interim, fundus cameras (a specialised low-power microscope with an attached camera) have become really powerful and we are benefiting from the advent of non-mydriatic (allowing doctors to get a picture of the retina without dilating the pupil) digital fundus photography. This has made screening for DR much more accessible, sensitive and specific than was possible with the direct ophthalmoscope,’ he adds.
In other words, trained graders who are not necessarily from a medical background can now collect the information needed to diagnose and grade the condition and relieve some of the medical fraternity’s burden.
‘In Scotland, they already have a very well-established automatic grading system. What they’ve found is they’ve been able to screen nearly 40% more patients than they could before they implemented the system and they’ve done this without increasing their staff numbers.’
Google is also working on an artificial intelligence model, based on this ethos. In a paper published in JAMA in December 2016, Google researchers demonstrated the potential of a deep learning (using large data for decision-making) algorithm that used retinal photographs to interpret signs of diabetic eye disease in patients.
The database consisted of over 100 000 retinal images (images of the back of the eye) that were graded by a panel of over 50 licensed ophthalmologists as a way to ‘train’ the database. They then ‘tested’ the accuracy of the trained database with two datasets of fundus images. The results showed that the algorithm was as accurate as an ophthalmologist in detecting the disease.
As Dr Cook says, ‘if one can integrate Google’s artificial intelligence with the fundus camera (already proven to be a cost-effective diagnostic method providing high-quality images), you’ve got a very powerful machine that can greatly increase the number of people being diagnosed by any one clinic.’
The Google algorithms have the potential to fine-tune their ability to learn and improve diagnostic performance. For instance, current technology relies on 2D photographs, but work is underway to use 3D images – Optical Coherence Topography (OCT) – to diagnose a wider range of eye diseases. OCT allows the algorithms to work with more subtle signs of disease than regular 2D images might show. However Dr Cook explains that this may add unnecessary expense to the diagnostic model.
Once disease is detected, people need to be informed of the increased risk of severe complications while these can still be reversed by tight control of blood sugar, blood pressure and cholesterol. Once more severe retinopathy is detected, referral to an ophthalmologist is appropriate because treatment would go beyond normal diabetic management.
While ophthalmologists in South Africa are already using fundus cameras fairly widely in private practice, they have not yet adopted deep learning algorithms for diagnosing and grading the disease.
‘Google’s new Artifical Intelligence (AI) technology could have an extremely favourable impact on the detection of diabetic eye disease in areas where there is no access to any form of specialised care,’ Dr Conradie says. ‘However, one must always be wary of taking away the clinical judgement of an expert in interpreting special investigations as there is usually more to medicine than meets the eye. Sometimes other problems are detected during an examination: something that an algorithm may not necessarily pick up,’ says Dr Conradie.
Dr Cook adds that if a patient is offered a diagnosis by one practitioner, benefiting from a human touch, rather than through a telehealth model, they tend to adhere to their treatments more vigorously.
‘In South Africa, we will have to deal with the sheer volume of diabetic patients that will increasingly come through our practices and reach the many undiagnosed diabetics who are missing the most important years of early diabetes management. The key will be to integrate these technologies in ways that are of most benefit to our public,’ he concludes.