Purchase this article with an account.
Antoine Rivail, Wolf-Dieter Vogl, Ferdinand Georg Schlanitz, Magdalena Baratsits, Ursula Schmidt-Erfurth, Hrvoje Bogunovic; Prediction of disease conversion in intermediate AMD from longitudinal OCT using self-supervised deep learning. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1355.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The risk of progression to late age-related macular degeneration (AMD) is heterogeneous and difficult to estimate. Current image-based prediction models use specific AMD biomarkers such as drusen, but do not consider the entire morphologic spectrum nor the speed of morphologic change. We aim at predicting individual conversion in eyes with intermediate AMD using longitudinal OCT imaging representation beyond conventional biomarkers.
To encode longitudinal OCT, a deep neural network was trained in a self-supervised way by predicting the appearance of an unseen follow-up B-scan from a sequence of preceding observations (Fig .1). The trained network provides a generic representation that simultaneously encodes the structure of the retina and its evolution over time. By combining the representations of all B-scans, we obtained a representation of an OCT volume at that time-point. The compact representation was used as input to a machine learning classifier to predict whether an individual eye will convert to GA or CNV.
The methodology was trained and evaluated on a longitudinally registered dataset of 57 patients with bilateral intermediate AMD (98 eyes, 424 OCTs, followed over three months intervals for a duration of 3-7 years), with 9 patients converting to GA and 10 to CNV during the study. The classification tasks consisted of predicting whether the eye was going to show evidence of conversion three months after the last visit. Using the novel representation the classification resulted in a ROC AuC of 0.90 (specificity of 0.85 at 0.80 sensitivity) for GA and a ROC AuC of 0.65 for CNV (specificity of 0.49 at 0.80 sensitivity).
Our proposed method effectively utilizes longitudinal imaging by learning a generic compact representation of the aging retina and its evolution over time. Our representations shows promising results in predicting incoming conversion to late AMD and may improve personalized risk assessment in the management of AMD. Similar to previous studies, the onset of GA can be more reliably predicted than imminent CNV. The unbiased unsupervised approach offers the potential to identify previously unknown features in AMD, an entity with largely undisclosed pathogenesis.
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