Adding more evidence to the efficacy of artificial intelligence in vision testing, a new study reports its deep learning algorithm did just as well, if not better, than eye care providers in detecting and referring glaucomatous optic neuropathy.

The study used fundus images taken from screening programs, studies and a glaucoma clinic. The algorithm was trained using a retrospective dataset of 86,618 images that were assessed for glaucomatous optic nerve head features and referable glaucomatous optic neuropathy—defined as an optic nerve head appearance that was worrisome enough to justify referral—by 43 graders. The algorithm then compared three data sets: set A included fundus images reviewed by glaucoma specialists; set B were images from a teleretinal screening program; and set C were images from a glaucoma clinic.

The study evaluated the algorithm through the area under the receiver operating characteristic curve (AUC), sensitivity and specificity for referable glaucomatous optic neuropathy and glaucomatous optic nerve head features.

The algorithm performed best when identifying glaucoma in set A, but was less successful with iamges in sets B and C. The investigators suggested the decreases in performance could be explained by the differences in the reference standard and patient populations.

Nonetheless the algorithm—which was developed using eye care providers’ grades based on fundus photographs alone—maintained good performance and its prediction rate was “fairly well correlated” with the results of a full glaucoma workup, the researchers said.

The investigators also noted the algorithm showed significantly higher sensitivity than seven of 10 graders not involved in determining the reference standard, including two of three glaucoma specialists, and showed higher specificity than three graders (including one glaucoma specialist), while remaining comparable with others. The researchers reported the most crucial features related to referable glaucomatous optic neuropathy by both glaucoma specialists and the algorithm were presence of vertical cup-to-disc ratio of 0.7 or more, neuroretinal rim notching, retinal nerve fiber layer defect and bared circumlinear vessels. 

Phene S, Dunn RC, Hammel N, et al. Deep learning and glaucoma specialists: the relative importance of optic disc features to predict glaucoma referral in fundus photographs. Ophthalmology. September 24, 2019. [Epub ahead of print].