Researchers Find Early Success Spotting Diabetes Related Eye Damage With Algorithms


One of the most devastating long-term effects of diabetes is a condition known as diabetic retinopathy. Diabetic retinopathy affects both Type 1 and Type 2 diabetics, gradually deteriorating their vision and, if left untreated, resulting in full vision loss. The CDC reports that diabetic retinopathy is the leading cause of blindness in US adults, estimating that 40 to 45 percent of diabetics have already developed some stage of the condition. An estimated 80 percent of diabetics will experience some level of vision loss associated with diabetic retinopathy during the first 10 years of being diagnosed with diabetes. Doctors can slow or even halt diabetic retinopathy-related vision loss if it is caught early, but this has proven difficult because early stages of the disease are often asymptomatic.

Currently, the process for diagnosing early-stage diabetic retinopathy involves examining digital pictures of the eye, specifically the fundus, in search of lesions associated with the disease. This screening is a manual and time-intensive process that leads to low screening rates and delayed diagnoses. To improve upon the diagnosis process, the California HealthCare Foundation teamed up with Kaggle, a website that organizes contests for data scientists, and together they launched a contest asking engineers to develop algorithms that could automate the fundus review process by identifying lesions with software, and with minimal manual effort. The contest ran through the summer, promising $50,000 to the winner, and $20,000 and $10,000 to the second and third place teams, respectively. To support the effort, Kaggle uploaded thousands of fundus images, some with lesions and others without, so that contestants would have a standard data source to target with their algorithms.

After five months of work, the results of the challenge have been impressive. The winning team, led by UK statistician Benjamin Graham from the University of Warwick, developed an algorithm that pinpointed lesions in fundus images and agreed with doctor’s independent findings 85 percent of the time. This is especially impressive considering that doctors only agree with each other’s fundus findings 84 percent of the time. Kaggle founder Anthony Goldbloom says that deep-learning algorithms have real potential in other imaging-related applications, explaining, “We’ve been able to watch the spread of deep learning on Kaggle. It’s crushing every image-recognition contest we run.”

Organizers from the California HealthCare Foundation are now teaming up with researchers from University of California, Berkeley to begin evaluating next steps for rolling the algorithm out across the state’s clinics.

Enjoy HIStalk Connect? Sign up for update alerts, or follow us at @HIStalkConnect.

↑ Back to top

Founding Sponsors

Platinum Sponsors