Highly accurate and precise automated cup-to-disc ratio quantification for glaucoma screening

ABSTRACT Objective An enlarged cup-to-disc ratio (CDR) is a hallmark of glaucomatous optic neuropathy. Manual assessment of CDR may be inaccurate and time-consuming. Herein we sought to develop and validate a deep-learning-based algorithm to automatically determine CDR from fundus images. Design Algorithm development for estimating CDR using fundus data from a population-based observational study. Participants A total of 184,580 fundus images from the UK Biobank, Drishti_GS, and EyePACS. Main Outcome Measures The area under the receiver operating characteristic curve (AUROC) and coefficient of determination (R 2 ). Methods FastAI and PyTorch libraries were used to train a convolutional neural network-based model on fundus images from the UK Biobank. Models were constructed to determine image gradability (classification analysis) as well as to estimate CDR (regression analysis). The best-performing model was then validated for use in glaucoma screening using a multiethnic dataset from EyePACS and Drishti_GS. Results Our gradability model vgg19_bn achieved an accuracy of 97.13% on a validation set of 16,045 images, with 99.26% precision and AUROC of 96.56%. Using regression analysis, our best-performing model (trained on the vgg19_bn architecture) attained an R 2 of 0.8561 (95% CI: 0.8560-0.8562), while the mean squared error was 0.4714 (95% CI: 0.4712-0.4716) and mean absolute error was 0.5379 (95% CI: 0.5378-0.5380) on a validation set of 12,183 images for determining CDR (0-9.5 scale with a 0.5 interval). The regression point was converted into classification metrics using a tolerance of 2 for 20 classes; the classification metrics achieved an accuracy of 99.35%. The EyePACS dataset (98172 healthy, 3270 glaucoma) was then used to externally validate the model for glaucoma diagnosis, with an accuracy, sensitivity and specificity of 82.49%, 72.02% and 82.83%, respectively. Conclusions Our models were precise in determining image gradability and estimating CDR in a time-efficient manner. Although our AI-derived CDR estimates achieve high accuracy, the CDR threshold for glaucoma screening will vary depending on other clinical parameters. Precis Deep-learning-based models can accurately diagnose and monitor glaucoma progression through automated CDR assessment. However, the CDR threshold for glaucoma screening may vary depending on other clinical parameters.