IBM developed a method combining deep learning and visual analytics to detect and gauge the severity of diabetic retinopathy from an eye image. The team found their method beat out other published efforts that use deep learning.
People with diabetes may develop diabetic retinopathy, where high blood sugar levels damage blood vessels in the retina. These vessels may swell and leak, or close, stopping blood flow altogether, according to the American Academy of Ophthalmology.
The leading cause of blindness among diabetics, diabetic retinopathy is classified into five levels of severity. If left untreated, it can cause permanent blindness, but early detection is crucial in warding off vision loss.
"To substantially reduce the number of people unnecessarily losing vision from diabetic eye disease, there is a real need for innovation to improve effective screening of those who are at risk to enable early sight-saving treatment,” said Peter van Wijngaarden, principal investigator at the Centre for Eye Research Australia at University of Melbourne, in a statement.
Using more than 35,000 eye images, the IBM team trained the technology to identify different types of lesions and assess the extent of the damage to the retina’s blood vessels, IBM said in the statement. These included micro-aneurysms and hemorrhages, which can indicate the presence and severity of diabetic retinopathy.
The method, which uses deep learning, convolutional neural networks (CNN) and dictionary-based learning, was accurate 86% of the time in classifying the severity of the disease. The scale ranges from no diabetic retinopathy to mild, moderate, severe and proliferative diabetic retinopathy.
IBM’s algorithm analyzes images pixel by pixel and patch by patch, said IBM’s Rahil Garnavi in a blog post. It “learns patterns associated with a particular pathology and disease,” she wrote. As the tech churns through more images, it will get better and better at differentiating between the different levels of severity.
The disease is currently diagnosed using fundus photography of the retina—a clinician manually examines these images to spot lesions. But this is often a time-consuming and subjective process. A computer-aided method that quickly and accurately classifies the severity of the disease could also standardize the interpretation of eye images.