Purchase this article with an account.
Mark Christopher, Christopher Bowd, James A Proudfoot, Nicole Brye, Akram Belghith, Michael Henry Goldbaum, Jasmin Rezapour, Massimo Antonio Fazio, Christopher A Girkin, C Gustavo De Moraes, Jeffrey M Liebmann, Robert N Weinreb, Linda M Zangwill; Performance of Deep Learning Models to Detect Glaucoma Using Unsegmented Radial and Circle OCT Scans of the Optic Nerve Head. Invest. Ophthalmol. Vis. Sci. 2021;62(8):1014.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To evaluate the accuracy of using unsegmented radial and circle OCT scans of the optic nerve head (ONH) in deep learning (DL) models to detect glaucoma and estimate visual field (VF) mean deviation (MD).
Spectralis ONH radial circle (ONHRC) scans from 192 healthy subjects (330 eyes) and 441 glaucoma patients [DMC1] (712 eyes) provided 2,601 OCTs with 62,424 radial and 7,803 circular B-scans for analysis. The ONH-centered OCTs consisted of 24 equally-spaced radial B-scans and 3 circular B-scans at diameters of 3.5mm, 4.1mm, 4.7mm. VF data consisted of 24-2 testing collected within 180 days of imaging. Subjects were randomly divided into independent training (85%), validation (5%), and test (10%) sets. Individual DL models were trained to distinguish healthy vs. glaucoma eyes and predict VF MD based on unsegmented (1) radial and (2) circular B-scans using Resnet50 models. Diagnostic accuracy was evaluated using area under the receiver operating characteristic curve (AUC) and examined as a function of B-scan type (radial vs. circle), diameter, position, and glaucoma severity. VF estimation was evaluated using R2 and mean absolute error (MAE).
DL models using radial B-scans detected any glaucoma with an AUC (95% CI) of 0.77 (0.63 – 0.87), mild glaucoma (MD >= -6.0 dB) with 0.69 (0.47 – 0.82), and moderate-to-severe glaucoma (MD < -6.0 dB) with 0.86 (0.72 – 0.94). DL models using circular B-scans detected glaucoma with an AUC (95% CI) of 0.84 (0.76 – 0.89), mild glaucoma with 0.75 (0.66 – 0.83), and moderate-to-severe glaucoma with 0.97 (0.94 – 0.98). In detecting glaucoma, circular B-scan diameter had relatively little impact on performance, while radial B-scan position had a larger impact on performance (Figure 1). For predicting VF MD, DL models using circle scans performed better (R2 = 0.83, MAE = 1.8 dB) than models using radial scans (R2 = 0.77, MAE = 1.8 dB).
Circular B-scans outperformed radial in detecting glaucoma and performed comparably in estimating VF damage. However, radial orientation had a substantial impact on glaucoma detection accuracy. DL models that better exploit positional information could help increase accuracy.
This is a 2021 ARVO Annual Meeting abstract.
(A) Illustration of the radial and circular B-scans. (B) AUC as a function of circular B-scan diameter. (C) AUC as a function of B-scan orientation.
This PDF is available to Subscribers Only