Diabetic retinopathy is a sight-threatening condition that occurs in people with diabetes. The disease results from progressive damage to the small blood vessels in the retina — the light-sensitive layer of tissue located at the back of the eyeball responsible for building images and transmitting them to the brain. Such damage is the outcome of chronically high blood sugar levels in diabetics. The affected blood vessels begin to hemorrhage or leak fluid, which distorts vision. As the condition advances, new irregular blood vessels multiple on the retina’s surface, ultimately resulting in scarring and even eventual blindness. Though treatment can help manage progression, there is no cure for diabetic retinopathy. Early diagnosis is therefore crucial to effectively manage the disease.
Perhaps the easiest and inexpensive means of diagnosis diabetic retinopathy is via analysis of fundus images (photos of the retina and back of the eye). These images are taken noninvasively, adapted for large-scale screening purposes, and sent to ophthalmologists who examine them for microaneurysms, hemorrhages, and bright legions: the three earliest symptoms of the disorder. “It’s important to detect these abnormalities in the early stage,” Isabel Figueiredo of the University of Coimbra said. “If they are detected, doctors can stop the process of the disease.”
During a scientific session at the 9th International Congress on Industrial and Applied Mathematics, currently taking place in Valencia, Spain, Figueiredo presented an automatic screening software to efficiently identify the early signs of diabetic retinopathy. “Our goal is to mathematically analyze ophthalmological images in order to detect microaneurysms, hemorrhages, and bright lesions,” she said. Figueiredo’s work was solicited by Retmarker, a Portugal software application company that handles technological projects related to eye health.
This fundus image depicts the presence of a microaneurysm in the early stages of diabetic retinopathy.
Retmarker contacted Figueiredo and her team at Coimbra directly and asked them to develop a mathematical screening network for diabetic retinopathy detection. The company was not happy with its existing software and wanted something more accurate. Because Figueiredo was already working on image processing, she agreed. In 2013, Retmarker and the University of Coimbra signed a memorandum of understanding that outlined project goals, timelines, etc. Figueiredo and her colleagues divided up the work, and the university signed a license agreement with Retmarker in February 2017. The resulting software is now included in Retmarker’s products.
In addition to detecting microaneurysms, hemorrhages, and bright lesions, Figueiredo’s software also locates the abnormalities in fundus images. “We consider an image like a function, Figueiredo said. “We use analysis techniques to detect these abnormalities.” To establish their screening process, she and her colleagues derived novel detectors for the lesions that were dependent upon lesion-intrinsic properties. Their choice of techniques specifically relies on the analysis of several wavelet bands (derived from the retinal images’ isotropic undecimated wavelet transform decomposition), an appropriate combination of Hessian multiscale analysis, variational segmentation, and cartoon+texture decomposition.
The group based their decision criteria for lesion detection within an image frame on a simple thresholding approach. “Our main contribution was defining the right math,” Figueiredo said, specifically in the identification of accurate descriptors for detecting abnormalities. Microaneurysms appear in fundus images as small dots between 12 and 80 micrometers wide. They are discernable in the images’ wavelength bands, while segmentation methods effectively locate clusters of bright lesions. While adapting image processing techniques, Figueiredo and her team discarded the optic disc and vessel network to avoid making mistakes and inaccurately identifying lesions.
Screening methods detect the existence of hemorrhages—which indicate diabetic retinopathy—in fundus images.
After finalizing the software—which recently won second prize in a U.K.-based competition—Figueiredo and Retmarker conducted blind tests to evaluate its performance. They tested a total of 45,770 fundus images from 11,511 patients, and found a 90-100 percent sensitivity and a 70 percent specificity on a per-patient basis. These results were impressive, as sensitivity and specificity are the two evaluations most commonly used in—and most important to—the medical industry.
Portuguese public hospitals have been utilizing Figueiredo’s software since 2017, and have employed it in over 172,000 patient exams. In December 2018, Retmarker sold the product to Australia, where 2,000 additional patients have now benefitted from the program. Retmarker intends to continue expanding the software’s coverage to other companies. “We are very happy with the success this has achieved so far,” Figueiredo said.
||Lina Sorg is the associate editor of SIAM News.