Diagnosing Infant Eye Diseases ASAP, Using Generative AI
$2.2 million NIH-funded Stevens project will deploy AI with a twist to help doctors detect vision disorders in prematurely born infants
When infants are born prematurely, a host of serious health issues can occur. One of those issues is retinopathy of prematurity (ROP): an eye disorder and loss of vision that can quickly progress.
If it’s caught immediately, however, ROP can be treated and even reversed.
And Stevens has just been awarded nearly $2.2 million by the National Institutes of Health (NIH) to develop a new, AI-powered way of speeding up those diagnoses.
“ROP is an under-studied problem already, but particularly with regard to the use of AI,” says biomedical engineering department chair Jennifer Kang-Mieler, one of two Stevens principal investigators on the project.
“We want to change this, helping infants recover their vision if possible. And Stevens is the perfect place to do this research.”
My (data) generation
The NIH Multi-PI R01 award to PIs Kang-Mieler and Yu Gan will also involve close collaboration with both Oregon Health & Science University (OHSU) and the University of Illinois-Chicago (UIC).
“Dr. Peter Campbell at OHSU and Dr. Paul Chan at UIC are the world’s leading experts on ROP, with some of the best image databanks,” notes Kang-Mieler. “They have graciously agreed to work with us.”
To build a diagnostic tool, Kang-Mieler and Gan will do something remarkable: create their own expert dataset of what ROP pathology should look like in medical imaging, in order to train an AI how to find it most accurately.
“We are doing this because even the very best data on ROP is limited by the relatively small number of cases,” explains Kang-Mieler. “For some ophthalmological diseases, there are hundreds of thousands of approved images being studied by AI. However, that’s not the case for this disorder.”
teach an AI model to generate the best-possible “synthetic” images of human ROP by employing generative AI image-translation techniques.
To surmount that challenge, Kang-Mieler and Gan will work with Stevens post-doctoral fellows and graduate and undergraduate student teams to collect animal ROP images and process them, extracting the most useful predictive features from the images, then using them toThis new, synthetic image data can then be used to train neural networks that will assist ophthalmologists in spotting ROP or tracking the progression of the disease.
“This is a new approach to address data scarcity in medical generative AI,” notes Gan. “Our lab handled image translation in super resolution and virtual staining, and now we are translating from animal to human.”
“We’re not sure if anyone else is doing something quite like this right now, generating accurate animal-to-human data to predict visual disorders.”
“One strength of our approach is that we have good data on the progression of ROP through the lifetime of animals, from its emergence to development,” adds Kang-Mieler. “That has been captured pretty robustly in the imagery. If we can successfully utilize this information for human health prediction, it could benefit thousands of prematurely born infants.”
How will the researchers make certain their synthetically generated data is accurate and useful?
“We will carefully validate all our imagery and processes through a series of checkpoints,” explains Kang-Mieler, “both by using human ophthalmological experts — who will be asked to judge whether various images accurately depict ROP pathology or absence, without knowing if those images are real or generated by AI — and also through algorithmic methods.”
The work will also build knowledge for the broader medical research community, add the researchers, since any ROP imagery datasets the Stevens team generates will be available to prove useful for other researchers working on similar ocular disorders.
NIH’s initial funding for the project will extend through summer 2028.