Deep Learning Model Detects Sea Fan Neovascularization With High Accuracy

scanning electron micrograph of a sickle cell red blood cell
Technology could increase access to screening for sickle cell retinopathy.

Researchers have developed a deep learning system capable of detecting sea fan neovascularization from color ultra-widefield fundus photographs (UWF-FPs) with high sensitivity and specificity, according to a report published in JAMA Ophthalmology. With further validation, this deep learning system may facilitate automated screening of asymptomatic patients with proliferative sickle cell retinopathy, the report shows. 

The model employs a convolutional neural network that was trained based on the largest database to date of UWF-FPs portraying adults with sickle cell hemoglobinopathy. This is the first time machine learning has been directed to the challenge of classifying this type of peripheral retinal neovascularization, the researchers noted. Investigators believe this technology may help to identify patients who most merit referral to a retinal specialist for evaluation and possible treatment of the vision-threatening.

“We developed a deep learning system to detect sea fan neovascularization from color UWF-FPs with high sensitivity and specificity compared with retinal specialist reference standard grades,” investigators explain. This convolutional neural network (CNN) detected the absence or presence of sea fans in individual UWF-FPs with 97.0% accuracy, 97.4% sensitivity, and 97.0% specificity. 

This cross-sectional analysis included 1182 UWF-FPs of 190 nondiabetic adults who received care for sickle cell hemoglobinopathy (SCH) or sickle cell retinopathy at the Wilmer Eye Institute of Johns Hopkins University in Baltimore, Maryland, from January 1, 2012 to January 30, 2019. Those with sickle cell trait were included, because they are also at risk for proliferative sickle cell retinopathy (PSR). Two masked retinal specialists graded the fundus photographs, as well as 898 fluorescein angiography (FA) images that represented 76.0% of the total study population. A third masked retinal specialist graded any images with indeterminate status.

To find retinal specialist graders’ sensitivity for the test set, the category of sea fan neovascularization incidence was combined with the category for indeterminate presence. One primary grader’s sensitivity achieved 100% with specificity of 98.0%, while the second masked grader’s sensitivity reached 94.9% with specificity of 97.0%.

Investigators noted that PSR affects underserved communities to a greater extent than the average population, and a “robust automated system” may increase access to both screening and retinal specialist referrals. “Integration of color UWF-FPs into nonophthalmic medical practices may help reduce the burden of separate in-person ophthalmology visits for asymptomatic patients with SCH,” according to the researchers.

With earlier detection of sea fan neovascularization, physicians could offer patients treatment options, such as prophylactic scatter laser photocoagulation, which may halve the rates of PSR-associated vision loss and vitreous hemorrhage compared with observation alone, according to the researchers. Prior research shows that, untreated, 17% of these patients developed severe vision loss in approximately 7 years.

Limitations of the study included the presence of eyelids at edges, and a degree of reduced resolution in the UWF-FP periphery. Also, the analysis focused on an adult-only population, as well as on sea fans, leaving out other ocular effects of PSR.

Disclosure: Several study authors declared affiliations with the biotech or pharmaceutical industries. Please see the original reference for a full list of authors’ disclosures.


Cai S, Parker F, Urias MG, Goldberg MF, Hagar GD, Scott AW. Deep learning detection of sea fan neovascularization from ultra-widefield color fundus photographs of patients with sickle cell hemoglobinopathy. JAMA Ophthalmology. Published online December 30, 2020. doi:10.1001/jamaophthalmol.2020.5900.