Purchase this article with an account.
Christopher Bowd, Akram Belghith, Linda M Zangwill, Michael Henry Goldbaum, Mark Christopher, Elham Ghahari, Huiyuan Hou, Sasan Moghimi, Rafaella Penteado, Robert N Weinreb; Diagnostic accuracy of OCTA vessel density and OCT tissue thickness measurements using machine learning. Invest. Ophthalmol. Vis. Sci. 2019;60(9):3044. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To compare gradient boosting classifier model (GBM) analysis of OCTA-measured vessel density (VD) and OCT-measured tissue thickness to software generated parameters for classifying healthy and glaucomatous eyes.
100 eyes of 51 healthy participants and 264 glaucomatous eyes of 185 patients from the Diagnostic Innovations in Glaucoma Study who underwent OCTA and OCT (OptoVue Avanti with AngioVue software) imaging of the macula and ONH were enrolled in this observational cross-sectional study. GBMs were trained independently on 1) all exportable macula VD measurements, 2) all exportable peripapillary VD measurements, 3) all exportable macula thickness (GCC) measurements, and 4) all exportable peripapillary thickness measurements from 69 healthy eyes and 188 glaucoma eyes. The discrimination power of GBM measurements was assessed using an independent dataset composed of 31 healthy eyes and 76 glaucomatous eyes. GBM VD classifications were compared to superficial whole image macular VD, whole image capillary density (CD, with large vessel removal) and peripapillary CD classifications and GBM tissue thickness classifications were compared to global GCC thickness and global peripapillary RNFL thickness by comparing ROC curve areas (AUROCs).
The glaucoma test group was significantly older [mean (95% CI) 71.9 (68.7, 75.1) and 59.5 (53.4, 65.6) years, respectively, P < 0.001) and had worse visual field MD [-6.9 (-8.5, -5.2) dB and -0.53 (-0.2, 0.8) dB, respectively, P < 0.001) compared to the healthy test group. AUROCs with associated 95% CIs for the investigated GBM analyses and the instrument provided parameters are shown in the figures below. GBM peripapillary VD results were similar to or better than instrument provided VD parameter results and GBM RNFL thickness results were better than instrument provided thickness parameter results. Peripapillary VD-based GBM delivered the best classification performance (all p < 0.05). Classification performance of macular VD-based and macular thickness-based GBMs was worse than classification performance of peripapillary VD-based and RNFL thickness-based GBM models (all P < 0.05).
Machine learning gradient boosting models improve glaucoma diagnostic accuracy of OCTA measurements compared to individual instrument provided parameters. Such techniques could be incorporated into instrument software to improve clinical usefulness.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only