Eye conditions caused by diabetes (diabetic retinopathy) and deterioration of the macula (age-related macular degeneration) are diagnosed by looking at the back of the eye. These diseases are the main causes of blindness in developed countries and, according to the World Health Organization, are preventable 80 per cent of the time.
Artificial-intelligence algorithms have long been used to support the everyday work of ophthalmologists. Certain common illnesses such as glaucoma, diabetic retinopathy, macular degeneration, corneal conditions and cataracts are diagnosed and monitored based on millions of data points provided by digital cameras.
Ophthalmic instruments are constantly improving and can now generate massive volumes of complex diagnostic images. While the data contained in these images exceeds human analytical capabilities, they’re driving new developments and applications of AI.
Deep-learning models, which have been “trained” using huge volumes of images, have already revealed links between phenomena. Using a single image of the back of the eye, these models can determine a patient’s gender, age, smoking history, blood pressure and cardiovascular risk factors.
Now, innovations in ophthalmic image analysis are showing promise for a new area of precision medicine. Specifically, treatment of age-related macular degeneration could be guided and customized based on predictions of an individual’s response and the risk that a mild condition could deteriorate into a more severe form of the disease. Recent breakthroughs are establishing correlations between how patients respond to treatment and which layers of the retina have been affected by their condition.
Other algorithms will soon be able to help with real-time diagnoses and decision-making, making it easier to perform mass screening and provide medical care for diabetic retinopathy at the most appropriate stage of the disease. The theoretical possibility of evaluating up to 260 million images per day using an automated system is one of the most promising avenues of further study.
Limited in scope
Advances in ophthalmological AI nevertheless still depend on annotated databases that are limited in scope and don’t represent the population groups for which they are intended (in terms of genetics, race, prevalence of the disease and image quality). Furthermore, they are assessed based on varying standards.
Current data showing comparable performance between humans and learning algorithms still have to be adapted for clinical and real-life situations. As it stands, some algorithms even go so far as falsely detecting new illness.
For AI to live up to its potential for enhancing ocular healthcare, we must promote and support the creation of shared annotated databases and ensure that the information in these databases is applicable to the population group for which it is intended.
Source: University of Montreal