Deep Learning for Identifying Eyes at Risk for Glaucomatous Optic Neuropathy
By Lynda Seminara
Selected By: Stephen D. McLeod, MD
Journal Highlights
Ophthalmology, December 2019
Download PDF
Phene et al. developed an algorithm based on deep learning and tested its effectiveness for photographic features of the optic nerve head (OHN) that would prompt referral for further evaluation of glaucomatous optic neuropathy (referable GON). They found that a deep learning algorithm trained solely on fundus images has greater sensitivity than eye care providers for detecting referable GON; specificity was comparable for the two methods of detection.
The fundus images used in this research were obtained from screening programs, published studies, and a glaucoma clinic. The algorithm was trained using 86,618 images that also were graded by eye care providers for glaucomatous ONH features and referable GON. Of the 43 graders, 14 were fellowship-trained glaucoma specialists, 26 were comprehensive ophthalmologists, and three were optometrists.
The algorithm was validated using three datasets: 1) Dataset A included 1,205 images (one per patient; 18.1% referable) adjudicated by panels of glaucoma specialists; 2) dataset B consisted of 9,642 images (one per patient; 9.2% referable) from a diabetic teleretinal screening program; and 3) dataset C comprised 346 images (one per patient; 81.7% referable) from a glaucoma clinic. Outcome measures were area under the receiver operating characteristic curve (AUC), sensitivity, and specificity for referable GON and glaucomatous ONH features.
The algorithm’s AUC for referable GON was 0.945 in dataset A (95% confidence interval [CI], 0.929-0.960), 0.855 in dataset B (95% CI, 0.841-0.870), and 0.881 in dataset C (95% CI, 0.838-0.918). AUCs for glaucomatous ONH features ranged from 0.661 to 0.973. The sensitivity of the algorithm was significantly higher than that of seven of 10 graders not involved in determining the reference standard, including two of three glaucoma specialists. The specificity of the algorithm exceeded that of three graders (including one glaucoma specialist) and was comparable to that of other graders. The algorithm performed favorably across independent datasets. According to specialists and the algorithm, crucial features of referable GON were vertical cup-to-disc ratio ≥0.7, notching of the neuroretinal rim, abnormality of the retinal nerve fiber layer, and baring of the circumlinear vessels.
The authors suggested that algorithms such as this one may improve the effectiveness of glaucoma screening in settings without clinicians who can interpret ONH features.
The original article can be found here.