'AI beats humans at detecting cause of blindness in babies'

An algorithm that uses artificial intelligence (AI) can automatically and more accurately diagnose a potentially devastating cause of childhood blindness than most expert physicians, a study has found.

The finding published in the journal JAMA Ophthalmology could help prevent blindness in more babies with the disease, called retinopathy of prematurity, or ROP.

The algorithm accurately diagnosed the condition in images of infant eyes 91 per cent of the time, researchers said.

On the other hand, a team of eight physicians with ROP expertise who examined the same images had an average accuracy rate of 82 per cent, they said.

"There is a huge shortage of ophthalmologists who are trained and willing to diagnose ROP. This creates enormous gaps in care, even in the US, and sadly leads too many children around the world to go undiagnosed," said Michael Chiang, a professor at Oregon Health & Science University (OHSU) in the US.

"This algorithm distills the knowledge of ophthalmologists who are skilled at identifying ROP and puts it into a mathematical model so clinicians who may not have that same wealth of experience can still help babies receive a timely, accurate diagnosis," said Jayashree Kalpathy-Cramer from the Massachusetts General Hospital.

Retinopathy of prematurity is caused by abnormal blood vessel growth near the retina, the light-sensitive portion in the back of an eye.

The condition is common in premature babies and is the leading cause of childhood blindness globally.

The new algorithm specifically uses deep learning, a form of AI that mimics how humans perceive the world through vision, including identifying objects.

The researchers combined two existing AI models to create the algorithm, while developing extensive reference standards to train it.

They first trained the algorithm to identify retinal vessels in more than 5,000 pictures taken during infant visits to an ophthalmologist.

Next, they trained it to differentiate between healthy and diseased vessels.

They compared the algorithm's accuracy with that of trained experts who viewed the same images and discovered it performed better than most of the expert physicians.