Purchase this article with an account.
Joelle Hallak, Maximilian Pfau, Luis De Sisternes, Rawan Allozi Rupnow, Joseph Baker, Theodore Leng, Daniel L. Rubin; Fully-automated prediction of anti-VEGF treatment using imaging biomarkers in neovascular age-related macular degeneration. Invest. Ophthalmol. Vis. Sci. 2020;61(7):4235.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To predict the number of anti-VEGF injections during a pro re nata or treat and extend treatment based on sets of optical coherence tomography raster scans in neovascular age-related macular degeneration in a real-world setting.
Spectral-domain optical coherence tomography (SD-OCT) volume scans (19 B-scans across 20°x15°) form the Department of Ophthalmology and Visual Sciences, University of Illinois at Chicago were segmented with a custom deep-learning-based analysis pipeline. Retinal thickness and reflectivity values were extracted for the central and the four inner ETDRS subfields for six retinal layers (inner retina, outer nuclear layer, inner segments [IS], outer segments [OS], retinal pigment epithelium-drusen complex [RPEDC] and the choroid), as well as drusen related features (drusen area, average thickness, drusen slope). Random forest regression was used to predict the number of anti-VEGF injections. Variable importance (varimp) scores were generated and ranked for each feature using the mobforest package in R.
A total of 66 eyes of 46 patients with a median age of 77.2 years (58.6, 93.4) and follow-up time of 3.4 years (0.1, 8.1) were included in this analysis. Our fully-automated pipeline included 245 imaging features. To address multicollinearity, features were selected such that all pairwise correlations were less than 0.8. Features were reduced to 102, and that set was included in the random forest. Selected features included the mean of the maximum OS and IS intensity projection for the superior-subfield (varimp 4.07 and 2.62, respectively), and the standard deviation of the IS thickness for the superior-subfield (varimp 3.04). Additionally, features related to the mean of the outer nuclear layer thickness in the superior- and temporal-inner-subfield were also selected (varimp 1.21 and 1.15, respectively).
The proposed fully-automated pipeline allows for selecting SD-OCT imaging features to predict future anti-VEGF treatments in a real-world setting. These models may help to identify patients in need of upcoming long-lasting anti-VEGF treatment variants (e.g. port delivery system, gene therapy).
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only