Purchase this article with an account.
Bianca S Gerendas, Amir Sadeghipour, Martin Michl, Felix Goldbach, Georgios Mylonas, Thomas Alten, Oliver Leingang, Stefan Sacu, Hrvoje Bogunovic, Ursula Schmidt-Erfurth; Deep learning-based automated fluid quantification in clinical routine OCT images in neovascular AMD over 5 years. Invest. Ophthalmol. Vis. Sci. 2021;62(8):122.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To demonstrate the benefit of objectively quantifying intra- (IRF) and subretinal fluid (SRF) in real-world clinical routine OCT monitoring in eyes with nAMD using an automated deep learning-based algorithm.
Data from five databases of the Vienna Imaging Biomarker Eye Study (VIBES) registry from 2007-2018 (electronic patient records, treatment database and 2 OCT devices) were analyzed using the Vienna Fluid Monitor. Matching all entries and filtering for active nAMD by baseline (BSL) OCT of suitable quality for automated IRF, SRF and CST segmentation led to inclusion of 1127 eyes. Visual acuity (VA) and OCT data at BSL, month (M) 1-3 and years (Y) 1-5, age, gender and number of treatments were included in the analysis. Subanalyses compared the performance of the algorithm to manual analysis of the Vienna Reading Center in a subset of 20% of eyes.
Mean CST was 358µm at BSL and decreased to 280-303µm over the entire follow-up. IRF/SRF volumes were highest at BSL (IRF: 22/77/107nl in 1/3/6mm area; SRF 14/86/263nl in the 1/3/6mm area). IRF decreased to a mean of 4-5nl at M1-M3 in the 1mm area and increased to 11nl at Y1 and to 16nl at Y5. SRF decreased to a mean of 3-5nl at M1-M3 in the 1mm area and remained below 7nl until Y5. IRF was the strongest parameter to reflect the course of visual acuity over time (Figure). When compared to manual expert analysis the Vienna Fluid Monitor found large amounts of fluid (IRF/SRF=39/23nl 1mm; 186/350nl total volume) in reading center determined presence and almost no fluid (IRF/SRF=2/1nl 1mm; 16/21nl total volume) in determined absence of both IRF and SRF (accuracy/sensitivity/specificity ~0.80).
Deep learning-based automated fluid quantification in clinical routine images is well-suited to objectively, reliably and rapidly measure treatment response and may guide clinical management in nAMD. The Vienna Fluid Monitor introduces expertise at reading-center level to a clinical routine setting while saving valuable time of the examiner. Automated volume measurements of retinal fluid compartments in a real-world dataset over a period of 5 years suggests IRF volume as the most practical parameter for treatment decisions.
This is a 2021 ARVO Annual Meeting abstract.
Relative picture of VA vs. fluid vs. CST. When IRF is decreasing during M1-M3, VA is increasing, later IRF increases with VA decrease; the decrease in CST/SRF is only visible during M1.
This PDF is available to Subscribers Only