Identifying glaucomatous optic nerve head features and glaucoma risk in fundus images at eye-care provider levels of accuracy using deep learning algorithms

Abstract

Purpose : To apply deep learning (DL) to create and validate an algorithm to automatically predict the risk of glaucoma and detect glaucomatous optic nerve head (ONH) features from color fundus photographs; to determine the relative importance of these features in the overall glaucoma risk assessment by glaucoma specialists; and to compare the performance of the algorithm with eye-care providers.

Methods : A DL algorithm was trained using a retrospective development dataset of 58,033 retinal images, graded for glaucomatous ONH features, glaucoma risk and gradability, by a panel of 41 graders (13 glaucoma specialists, 25 ophthalmologists, and 3 optometrists). The resultant algorithm was validated on a dataset consisting of 1,205 images graded by a rotating panel of 3 out of 12 fellowship-trained glaucoma specialists using adjudication as the reference standard. We evaluated the relative importance of the different ONH features by running logistic regression analyses on the validation dataset and considering coefficients and ranking order of binarized features compared to overall referral.

Results : Validation dataset had 1,205 images (1,205 patients; mean age, 56.7 years; 53.2% women; prevalence of referable glaucoma risk in gradable images: 218/1171 [19%]). For detecting referable glaucoma risk, the algorithm had an AUC of 0.940 (95% CI, 0.923-0.955). When comparing to eye-care providers, the algorithm’s ROC curve was superior on a subset of 411 images (see Figure). For glaucomatous ONH features the algorithm had AUCs ranging between 0.608 - 0.977. The presence of vertical C/D Ratio >= 0.7, neuroretinal rim notching, RNFL defect, or bared circumlinear vessels contribute the most to glaucoma risk evaluation by eye-care providers (see Table).

Conclusions : A DL algorithm trained on fundus photos alone can detect referable glaucomatous risk with higher sensitivity and at least comparable specificity to eye-care providers. Certain ONH features have greater influence on overall glaucoma risk assessment by glaucoma specialists.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

The algorithm is illustrated as a blue line, with 10 individual graders indicated by colored dots: glaucoma-specialists (blue), ophthalmologists (red), and optometrists (green). For each image, the reference standard was determined by a different set of three glaucoma specialists in an adjudication panel