Research aims to help patients at risk for diabetic blindness

Diabetic retinopathy (DR) is one of the leading causes of blindness in the world and affects up to 80 percent of diabetic patients. However, with proper detection, the effects of DR can be often treated and blindness prevented. Computer Science doctoral student Jeremy Benson is looking at effective screening using a widely available software library from Google called Tensorflow Inception Network. He will use the resources at The University of New Mexico Center for Advanced Research Computing (CARC) in his research.

“Diabetic Retinopathy is an especially hot topic for businesses in the medical imaging field. Currently, there's a race to reach FDA-approval for a software system that can accurately screen images from patients at risk. I work at VisionQuest Biomedical, and this is our major project at the moment,” Benson explained.

His research is called Transfer Learning for Diabetic Retinopathy: Screening for at-risk-patients in a clinical setting using Tensorflow’s Neural Networks. By collecting a repository of readily available images, machines can learn to automatically identify objects, or in this case, symptoms of DR.

“One of the biggest challenges in the image processing field, especially when dealing with many images and images of high dimension and resolution is how we store and perform calculations on all of it without running out of memory. CARC has some machines with large GPUs (technology for processing images) and sufficient memory to deal with lots of images at a time. Certain tasks that would take forever on a laptop are possible with CARC systems,” he said.

Specialists who can diagnose DR are not widely available everywhere. That is where this sophisticated technology can help.

“Simply put, there are way more diabetics than there are experts to diagnose their symptoms. A solution to this problem was to introduce tele-medicine systems, or remote cloud-based solutions that allow an expert in one geolocation to view images and grade them. But even then, there aren't enough experts to view all the images. So, we try to help the clinics and doctors triage their patients by using software to screen for cases with this condition. At VisionQuest, we work with clinics in Mexico. A single site can see hundreds of patients, potentially, so reducing the clinic’s workload in which patients need additional tests or follow-ups is key.”

The technology can significantly help medical providers overloaded with patients. But until the technology passes FDA requirements, it’s not ready for all medical professionals to use.

“The main thing here is getting through FDA approval. In Mexico, we have regulatory clearance. Once we are cleared for commercial use in the U.S., this tech will reduce costs for providers. Identifying the patients-at-risk group and those who are OK significantly cuts down on the total number of patients who need to be seen for further tests or exams by a specialist,” Benson said, adding, “Specialists should focus their time on patients with real issues instead of everyone who comes through the door. If we can get the people with vision-threatening issues to the specialist and the people who are healthy back home, it's a win for everybody.”

Benson concluded “It’s great working on these types of projects that make a real impact. The biology and science, it's all really neat!”


Preprocessed fundus images fed to the network. No pathology (left) and advanced pathology (right)