Researchers from the National Eye Institute and the National Center for Biotechnology Information showed that a deep learning algorithm could detect geographic atrophy (GA) accurately, and concluded that it may be useful in identifying central GA as well.

The team evaluated 59,812 color fundus photos of 4,582 participants of the AREDS study and developed a deep learning model to detect the presence of GA, plus two additional models to detect central GA in different scenarios. They first trained the deep learning model to predict GA presence in a wide population of eyes, ranging from no disease to advanced AMD. They trained the other two models to predict central GA presence from the same population and from the subset of eyes with GA, respectively. They then compared model performance with that of 88 retina specialists.

The study authors found that the deep learning models held their own against human counterparts in many key measures of disease detection, reporting the comparative findings below: 

	GA detection model	Retina specialists	Central GA detection model	Retina specialists
Accuracy	0.965	0.975	0.966	0.990
Sensitivity	0.692	0.588	0.763	0.448
Specificity	0.978	0.982	0.971	0.993
Precision	0.584	0.368	0.394	0.296

The research team concluded that its “deep learning model demonstrated high accuracy for the automated detection of GA” and demonstrated non-inferiority to human retina specialists.

Keenan TD, Dharssi S, Peng Y, et al. A deep learning approach for automated detection of geographic atrophy from color fundus photographs. Ophthalmol. June 11, 2019. [Epub ahead of print].