Harvard Researcher Uses AI to Tackle Common Cause of Childhood Blindness

by Jamie Beckett

Doctors can prevent one of the most common causes of blindness in young children — but only when they can detect it.

The disease, called retinopathy of prematurity, or ROP, affects the youngest, smallest and most vulnerable infants. These are preterm babies born before 31 weeks who weigh less than 2¾ pounds.

Doctors can treat ROP if they catch it early enough, but there’s no objective way to determine which cases need treatment.

Jayashree Kalpathy-Cramer thinks AI can make a difference. Kalpathy-Cramer is a researcher at the Athinoula A. Martinos Center for Biomedical Imaging at Harvard Medical School and Massachusetts General Hospital. She and her postdoctoral fellow, James Brown, are developing a GPU-accelerated deep learning system that could automatically determine the severity of the disease.

“The main thing about ROP is that it’s preventable,” she said. “This is an area where deep learning can make a real difference.”

Stuck in the 1980s

ROP happens because preterm babies’ eyes aren’t fully developed. Blood vessels that supply the retina grow quickly in the last few weeks before birth. If the process is interrupted, the vessels may stop growing or grow into parts of the eye where they don’t belong.

Although the disease affects only a tiny fraction of premature infants, its effects can last a lifetime. ROP often improves on its own, but severe cases can lead to blindness or eye problems such as crossed eyes, lazy eye, glaucoma and early cataracts.

When doctors screen for ROP, they classify it in order of severity — normal, pre-plus and plus — depending on the condition of the eye’s blood vessels. Plus cases require treatment. Doctors decide the level of ROP by comparing what they see in the retina (or a digital image of it) with a standard photo selected by experts in the 1980s.

Not surprisingly, numerous studies show there’s widespread disagreement among experts about where to draw the line among the three categories.

“That’s where I hoped we could do something better, by leveraging the recent advances in computer vision,” Kalpathy-Cramer said.

Automated Diagnosis 

To do that, Kalpathy-Cramer obtained a dataset of 6,000 images matched with expert diagnoses from the Imaging and Informatics in Retinopathy Consortium, led by Dr. Michael Chiang of the Oregon Health & Science University. Kalpathy-Cramer and Brown used this data to train a deep neural network to differentiate normal, pre-plus and plus images.

Working at the Center for Clinical Data Science operated by Mass General and Brigham and Women’s Hospitals in Boston, Kalpathy-Cramer used an NVIDIA DGX-1 AI supercomputer with various cuDNN-accelerated deep learning frameworks to develop an algorithm for ROP diagnosis.

Next, she’ll test the algorithm on some 100,000 images supplied by Aravind Eye Hospitals and the Banker Retina Clinic and Laser Centre in India. Afterward, her plan is to try it out as a screening method in India.

A pediatric ophthalmologist uses an indirect ophthalmoscope to examine an infant for signs of retinopathy of prematurity (ROP), a condition that can cause childhood blindness.
A pediatric ophthalmologist uses an indirect ophthalmoscope to examine an infant for signs of retinopathy of prematurity (ROP)

Expert Diagnosis Where Experts Are Scarce

Kalpathy-Cramer is especially keen to deploy her method in low- and middle-income countries where access to highly trained ophthalmologists is often lacking. Long term, she wants to develop an inexpensive, portable device that could be used by nurses for an initial screening.

“If we can get our algorithm working well, we think it can really make a difference in reversing preventable blindness worldwide,” Kalpathy-Cramer said.

Update: After this post was published, Kalpathy-Cramer and her colleagues published several papers detailing advances in their work using deep learning to automate diagnosis and prevent childhood blindness. In the latest paper, published on ArXiv in May 2018, researchers describe how they used generative adversarial networks to expand image datasets for training deep learning models to analyze images for ROP and brain tumors. Below are links other recent papers:

* Main image and video for this story are courtesy of the National Eye Institute, National Institutes of Health.