EyesFirst Aims to Automatically Detect Chronic Diseases Through Retinal Scans

October 2011
Topics: Diseases, Image Processing, Health Innovation
MITRE's EyesFirst project shows promise for detecting the early stages of chronic diseases such as diabetes, hypertension, and Alzheimer's in a noninvasive way that uses advanced 3-D retinal imaging technology.
Close up of two eyes. One is encircled by red light from machine that is being looked into.

You've heard "The eyes are the mirror of the soul." MITRE researchers are working on a corollary: "The eyes are a window into the state of your health."

In a project called EyesFirst, MITRE's Center for Transforming Health is developing automated methods to detect the early stages of multiple diseases from information contained in 3-D retinal images. When the methods are fully developed, clinicians around the world will be able to use capabilities developed by the EyesFirst project to non-invasively detect conditions such as diabetes, hypertension, cardiovascular disease, cerebrovascular disease, Alzheimer's disease, and Parkinson's disease.

"There's growing evidence that changes in the retina correlate with the onset and progression of certain diseases," says Salim Semy, a lead software systems engineer at MITRE and principal investigator on the project. "The biggest problem with treating these diseases is detecting them early enough. For example, more than 25 percent of diabetic adults are undiagnosed primarily because they haven't been screened for it. This often leads to irreversible vision loss that could have been prevented.

"If we can develop automated methods to screen patients, we can help people head off future medical problems and reduce their medical costs as well."

Drawing on MITRE's Expertise

The EyesFirst project evolved from Semy's interest in eye research as he pursued his master's in biomedical sciences from Boston University's School of Medicine. His research involved work with MIT and the Massachusetts Eye and Ear Infirmary on the design of a retinal prosthetic. During his research, he came across studies on early disease indicators within the retina and saw an opportunity for MITRE to contribute. Semy then proposed EyesFirst as a project under MITRE's internal research and development program.

The project draws on MITRE's expertise in systems engineering, digital signal processing, data analysis, and decision support. Working with Semy is a core team that includes Harry Sleeper, co-principal investigator; Walt Kuklinski, senior principal signal processing engineer; Dave Stein, senior signal processing engineer; and David Smiley, lead software systems engineer.

An Open-Source Research Platform

The basis of EyesFirst is an open-source research platform that will include a public Database of Retinal Images, or DORI. The retinal images are produced using advanced 3-D retinal imaging technology, called optical coherence tomography, or OCT. (See "OCT Scans Retinas Like Radar.") The technician tags the OCT images with the patient's medical history and demographics, and then publishes this information to DORI. Researchers can use the collective dataset in DORI for subsequent analysis in investigations of new retinal predictors of disease.

The platform will use open standards to share the retinal images and metadata with clinical researchers. As an open-source research project, EyesFirst will also make its research results publically available for others to study, use, and improve, including the open-source image processing algorithms. Access to data and research results will assure transparency during all phases of research.

"The open approach brings together clinicians and accelerates their research into chronic disease detection," says Semy.

The Algorithm Challenge

To automate the diagnosis of a disease such as diabetes, researchers must identify which features of the retina give clues to the disease. To accomplish this, the EyesFirst team creates image-processing algorithms that detect retinal changes such as retina thickness, indications of fluid buildup, and the presence of fluid residue called exudates. But before creating those algorithms, the team makes sure the scanned images from the OCT device are free of distortions and artifacts.

Artifacts occur from the way the laser creates an image by scanning the laser back and forth across the retina to build an image. Sometimes the cross-section scans get misaligned because of a time lag from one scan to the next. That's when a heartbeat or slight eye movement puts the next scan out of alignment. A top view of an image with misaligned scans will show, for example, a break in a tiny blood vessel of the retina.

To minimize this effect, the MITRE team applied methods frequently used in radar image analysis to develop a motion artifact correction algorithm that can reliably realign OCT image scans. (See "Motion-Artifact Correction Corrects Image Alignment.")

"In looking for pathology, you need high-quality scans," notes Semy. "We don't want misalignments to be falsely construed as disease."

In the spirit of open-source research, the motion-artifact correction algorithm is going through MITRE's process for an open-source license. The team also created a second algorithm to detect changes in retinal thickness. "This algorithm measures the retina's thickness and compares these measurements against a normal set of values," says Semy. "In diabetes and other diseases, fluid buildup in the retina can make it bulge. Identifying this bulge is important in effectively treating diabetic retinopathy."

OCT imaging produces a cross-sectional view of the retina that shows the retinal layers and the fovea. The fovea is a dip in the retina responsible for sharp, central vision.

As an adjunct to the retinal thickness algorithm, the team is working on an automated classification system that can use the thickness data to determine if the retina is diseased or not.

The team is also developing a third algorithm to detect hard exudates, which are the remains of fluid buildup that the retina reabsorbs. "Even though retinal thickening isn't present, the hard exudates can be detected, which indicates something is still wrong," explains Semy.

Next Steps

Next, the team will start a pilot within MITRE to collect data from employees. "MITRE volunteers will come to our laboratory to register, provide medical history, and get an OCT scan," says Semy. "They will come back every two or three months for another scan. This will help us validate and adjust the process for setting up DORI and collecting data to track retinal changes over longer time periods. Then we'd like to develop DORI outside MITRE and engage ophthalmologic research centers and clinicians to help populate it. We hope our open data provides opportunities for other researchers to use OCT scans for improving preventative health care."

—by David A. Van Cleave


Publication Search