Deep learning software can automatically classify retinal atrophy using fundus autofluorescence (FAF) images, and accurately distinguish between between retinal atrophy that developed from age-related disease and late-onset inherited retinal diseases (IRDs), according to research published in the journal Computers in Biology and Medicine.
The investigation used a deep learning-based algorithm to classify 314 FAF images as either indicative of lesions originating from geographic atrophy (GA), or from Stargardt disease or pseudo-Stargardt multifocal pattern dystrophy (PSPD). Investigators affirm that “an accurate differential diagnosis between GA and late-onset IRDs masquerading as GA on FAF can be performed using a deep learning model, with excellent accuracy and excellent AUC-ROC values.”
The researchers used 2 approaches to train and test the multi-layer deep convolutional neural network (CNN). In the first approach, a ‘hold out’ validation technique involved dividing data from 314 FAF images into 3 subsets for individual algorithm runs. In the second approach, the dataset was split into 10 equal subsets. After image preprocessing, the software categorized each case of atrophy as either degenerative GA or genetically-caused. The CNN achieved what the investigators called “excellent accuracy.”
Subjects underwent FAF at the Ophthalmology Department of the Centre Hospitalier Intercommunal de Créteil, France. For each subject, 1 FAF image was sampled per year during the study period, April 2007 to August 2020. Participants were diagnosed with genetically confirmed IRD with atrophy or GA from age-related macular degeneration (AMD).
“Several IRDs phenocopy GA and accurate diagnosis may be difficult based on fundus appearance alone,” the study explains. “Deep learning applied to FAF imaging may help the clinician’s task, ensure a correct referral to a specialized ocular genetics unit if a genetic cause is suspected, and ultimately assist the clinician when genetic testing is not available.”
Previous studies have applied FAF imaging to distinguish GA from other retinal diseases, such as central serous chorioretinopathy, but the current analysis focuses on atrophy secondary to late-onset Stargardt disease and PSPD, which investigators suggest are more difficult to discern. While a limitation of this study was its retrospective nature, its strength was the use of integrated gradient visualization to evaluate the thoroughness of the classification.
Invasive methods such as fluorescein and indocyanine green angiography provide sufficiently clear distinctions between GA and IRD-related atrophy, although FAF is a non-invasive test, and the American Academy of Ophthalmology currently recommends non-invasive imaging for patients with IRD, according to the study.
Investigators note that rates of atrophy growth, and even risks for resulting vision loss is unique between late-onset Stargardt disease or PSPD, and AMD. “Therefore, refined phenotyping in these cases is essential, to correctly assess the disease course in patients presenting with RPE atrophy,” the study says.
Disclosure: Several study authors declared affiliations with biotech and pharmaceutical companies. Please see the original reference for a full list of disclosures.
Miere A, Capuano V, Kessler A, et al. Deep learning-based classification of retinal atrophy using fundus autofluorescence imaging. Computers in Biology and Medicine. 2021;130(1):104198.doi:10.1016/j.compbiomed.2020.104198.