Purchase this article with an account.
LeAnn Mendoza, Mark Christopher, Nicole Brye, James A Proudfoot, Akram Belghith, Christopher Bowd, Jasmin Rezapour, Massimo Antonio Fazio, Michael Henry Goldbaum, Robert N Weinreb, Christopher A Girkin, Jeffrey M Liebmann, C Gustavo De Moraes, Linda M Zangwill; Deep Learning Predicts Demographic and Clinical Characteristics from Optic Nerve Head OCT Circle and Radial Scans. Invest. Ophthalmol. Vis. Sci. 2021;62(8):2120.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop deep learning (DL) models for predicting age, sex, race, diabetes diagnosis, hypertension, cardiovascular disease, and axial length from Spectralis optical coherence tomography (OCT) retinal nerve fiber layer (RNFL) circle and radial scans.
Spectralis OCT circle and radial scans of the optic nerve head (ONH) acquired from healthy subjects, glaucoma suspects and glaucoma patients from the Diagnostic Imaging in Glaucoma Study (DIGS) and African Descent and Glaucoma Evaluation Study (ADAGES) were randomly assigned to training (85%), validation (5%) and testing (10%) datasets by patient. DL models were trained on unsegmented circle and radial scans to predict age, sex, race, diabetes diagnosis, hypertension, cardiovascular disease (CVD), and axial length. The circle scan dataset included 52,552 individual B-scans from 1,772 patients. The radial dataset included 111,456 individual B-scans from 730 patients.
The DL models of the best circle and radial scans predicted age with a mean absolute error (MAE (95% CI)) within 5.4 (4.9, 5.9) years and 5.1 (4.5, 5.8) years, respectively and a R2 (95% CI) of 0.73 (0.67, 0.78) and 0.64 (0.49, 0.76), respectively. For Axial length, the circle scan model had a MAE of 0.7 (0.6, 0.9) mm and an R2 of 0.3 (0.2, 0.4), and the radial scan model had a MAE of 0.8 (0.7, 1.0) mm and an R2 of 0.4 (0.2, 0.5). Accuracies (AUROC (95% CI)) for predicting sex in the best circle and radial scans was 0.72 (0.65, 0.79) and 0.68 (0.59, 0.77), respectively. The AUROC for predicting race in the best circle and radial scans was excellent, both 0.96, with 95% CIs of (0.92, 0.98) and (0.91,0.99), respectively. The AUROC for predicting diabetes diagnosis in the best circle and radial scans was 0.65 (0.52, 0.77) and 0.76 (0.64,0.85), respectively. The AUROC for predicting hypertension in circle scans was 0.71(0.59, 0.81) and in radial scans was 0.64 (0.54, 0.73). The AUROC for predicting CVD diagnosis in circle scans was 0.56 (0.47, 0.65) and in radial scans was 0.54 (0.48,0.62).
These results suggest that there are indicators of demographic and clinical characteristics embedded within OCT images showing that DL can infer these characteristics from images of the ONH. With the exception of predictions of diabetes, the DL models of circle scans predicted clinical and demographic features as well or better than radial scans.
This is a 2021 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only