Deep Learning-Based Detection of Reticular Pseudodrusen in Age-Related Macular Degeneration.

Abstract

Background: Reticular pseudodrusen (RPD) signify a critical phenotype driving vision loss in age-related macular degeneration (AMD). This study sought to develop and externally test a deep learning (DL) model to detect RPD on optical coherence tomography (OCT) scans with expert-level performance.

Methods: RPD were manually segmented in 9800 OCT B-scans from individuals enrolled in a multicentre randomised trial. A DL model for instance segmentation of RPD was developed and evaluated against four retinal specialists in an internal test dataset. The primary outcome was the performance of the DL model for detecting RPD in OCT volumes in five external test datasets compared to two retinal specialists.

Results: In an internal test dataset consisting of 250 OCT B-scans, the DL model produced RPD segmentations that had higher agreement with four retinal specialists (Dice similarity coefficient [DSC] = 0.76) than the agreement amongst the specialists (DSC = 0.68; p < 0.001). In the five external test datasets consisting of 1017 eyes from 812 individuals, the DL model detected RPD in OCT volumes with a similar level of performance as two retinal specialists (area under the receiver operator characteristic curve [AUC] = 0.94, 0.95 and 0.96 respectively; p ≥ 0.32).

Conclusions: We present a DL model for automatic detection of RPD with expert-level performance, which could be used to support the clinical management of AMD. This model has been made publicly available to facilitate future research to understand this critical, yet enigmatic, AMD phenotype.