A-Eye: Deep learning for early detection of eye diseases

This is a close-up photo of an Optical coherence tomography (OCT) machine in use by a clinician performing an eye scan at Moorfields Hospital in London. (Image credit: DeepMind)

There are 285 million people living with some form of sight loss in the world, a number that is expected to triple by 2050. And according to a top London hospital, more than 80-percent of eye diseases can be prevented or cured with treatment if caught early. So it’s with a real sense of urgency that the Moorfields Eye Hospital is endeavoring to speed up and improve the accuracy of detection.

At present, hospitals and clinics rely on human retinal specialists to analyze and interpret optical coherence tomography (OCT) scans, which provide a detailed, 3D image of the back of the eye. But with the sheer number of scans performed, the wait time between review and treatment can be lengthy, especially in situations where any delay might imperil someone’s ability to see. Moorfields alone performs more than 1,000 scans a day from its 30 locations across the U.K. Yet analyzing these scans is the type of repetitive task for which machine learning is well-suited.

The medical center has been working since early 2016 with Alphabet-owned DeepMind to develop AI software that can detect multiple types of eye diseases, like age-related macular degeneration (AMD) and glaucoma. In August, they reached a milestone. By using deep learning, they were able to train software on 14,884 scans and return a diagnosis in just seconds.  

This photo shows DeepMind's software analyzing the scan of a right eye at Moorfields Hospital in London. (Image credit: DeepMind)

DeepMind’s software analyzing the scan of a right eye at Moorfields Hospital in London. (Image credit: DeepMind)

“Not only can [our system] automatically detect the features of eye diseases in seconds, but it can also prioritise patients most in need of urgent care by recommending whether they should be referred for treatment,” DeepMind wrote in a blog post. “This instant triaging process should drastically cut down the time elapsed between the scan and treatment.”

DeepMind’s approach combines two different neural networks. The first, known as the segmentation network, analyzes the OCT scan to provide a map of the different types of eye tissue and features it sees including hemorrhages, lesions or irregular fluid. The second, known as the classification network, analyzes this map to present clinicians with diagnoses and a referral recommendation.

“Crucially, the network expresses this recommendation as a percentage, allowing clinicians to assess the system’s confidence in its analysis,” DeepMind wrote in the blog post.

DeepMind isn’t the only Alphabet-owned entity studying the retina. AI experts from Google and from Verily, Alphabet’s research organization devoted to the life sciences research, are also looking at the human eye. A few weeks ago, Jeff Dean — Google’s head of AI — shared about his obsession with the retina onstage at the Google Cloud Next conference in San Francisco. He and his team are training machine learning models to take a fundus image, which is a picture of the back of the retina, and make an assessment of whether that retina shows symptoms of a disease called diabetic retinopathy.

When AI meets the eye

No matter how smart an AI system becomes, however, there are myriad safety and regulatory issues to comply with in the medical sector.  

“It’s still early days and the next step is for the research to go through clinical trials, as well as regulatory approval, before it can be used in hospitals and other clinical settings,” said Dr. Dominic King, Medical Director at DeepMind Health.

If the technology is approved for general use, Moorfields’ clinicians will be able to use it for free across all of its hospitals for an initial period of five years, according to DeepMind. Moorfields’ hospitals serve 300,000 patients a year.

The good news for other eye care hospitals and clinics around the world, adds King, is that the software can be easily applied to different types of eye scanners, not just the specific model it was trained on at Moorfields. This is especially valuable for health centers in developing countries that may not have retinal specialists handy to analyze the OCT scans.

While Moorfields’ patients represent just a fraction of the 285 million people afflicted with a form of visual impairment, this program is a real start, and possibly a very good example of computer vision helping preserve human vision.