Abstract
Purpose :
(1) To automatically identify, segment and quantify optic disc drusen (ODD) regions in optical coherence tomography (OCT) images of the optic nerve head (ONH); (2) To use such segmentations to discriminate ODD (i.e. benign pseudo-papilledema) from true papilledema due to raised intracranial pressure.
Methods :
This was a cross-sectional comparative study including patients with confirmed ODD (23 eyes), papilledema due to high intracranial pressure (12 eyes), and healthy controls (10 eyes). Volume raster scans of the ONHs, each consisting of up to 97 Bscans, were acquired using Spectralis OCT, then processed with adaptive compensation to enhance tissue contrast and visibility (including ODD regions). A deep learning algorithm (DRUNET; Devalla et al., Biomed Opt Express, 2018) was then trained on 2,807 Bscans (from 31 eyes), and tested on 1,211 Bscans (from 14 eyes), in order to differentiate ODD regions and conglomerates (when present) from all other ONH tissues. Both training and testing sets were balanced with respect to the explored conditions. The performance of our algorithm was then assessed (against manual segmentations of ODD regions) using the generalised dice coefficient. Finally, for each eye in the testing set, we computed a drusen score, defined as the total volume of ODD regions and conglomerates.
Results :
Our algorithm successfully segmented ODD regions and conglomerates in all ODD eyes from the testing set with a generalised dice coefficient of 0.86. In those eyes, our ODD volume predictions matched the ground truth volumes relatively well (Figure A-B). In all papilledema and healthy eyes from the testing set, the drusen score was zero (i.e. no ODD), indicating that our algorithm had the advantage to also discriminate ODD from true papilledema (Figure C-D).
Conclusions :
We propose a new deep learning system able to automatically identify, segment and quantify ODD regions in OCT images of the ONH, discriminating in the same time patients with benign pseudo-papiledema from those with true papilledema due to raised intracranial pressure. Further longitudinal studies should establish if this method has the ability to automatically monitor volume changes in these optic nerve lesions.
This is a 2020 ARVO Annual Meeting abstract.