Abstract

Purpose : Drusen deposits that form between the retinal pigment epithelium (RPE) and Bruch's membrane (BM) are a significant risk factor for age-related macular degeneration (AMD). In this work, a novel way of automatically segmenting the RPE-BM complex in OCT images from multiple devices using a convolutional neural network is proposed.

Methods : A total of 15744 B-scans (Bioptigen SD-OCT, NC, USA) from 269 early AMD patients and 115 normal subjects, selected from a publicly available dataset, were used in this study (split on subject level in 70% train, 10% validation and 20% test). For each B-scan, the inner limiting membrane (ILM), RPE, and BM have been manually segmented (Fig. 1). A convolutional neural network (U-Net) with generalized Dice loss function was trained to segment the three retinal layers. Data augmentation was used, including geometric and intensity transformations. Postprocessing was applied to the output of the U-Net, removing the small connected components and using a minimal cost path function to ensure connectivity within each layer. To test the segmentation algorithm, OCT data from three other devices (3D-2000 SD-OCT, Topcon Corporation, Japan; Spectralis SD-OCT, Heidelberg Engineering, Germany; MERLIN SS-OCT, Imagine Eyes, France) were used. The segmentation performance was assessed using Dice score and visual inspection by an experienced AMD specialist.

Results : The Dice score for the Bioptigen test set was 0.97. Visual assessment confirmed that the predictions of the model in the B-scans from the four different devices followed the retinal layers accurately. The predicted RPE-BM complex was color coded displayed as a function of the RPE-BM distance (abnormal thickness indicated in red), thus allowing the location and analysis of the drusen deposits (Fig. 2).

Conclusions : The proposed approach is a first step towards the automation of soft drusen quantification and follow up, as it enables the automatic assessment of the RPE-BM complex in healthy and AMD data from four different OCT devices.

This is a 2020 Imaging in the Eye Conference abstract.

 

B-scan from the Bioptigen test set with the manually segmented retinal layers (ILM, BM, and RPE) overlayed (A); annotated mask (B); prediction with the RPE-BM complex highlighted (C).

 

Model predictions and highlighted RPE-BM complex on each dataset: Bioptigen (A); Spectralis (B); Topcon (C); MERLIN (D). Different acquisition settings were considered between devices.