Automated Segmentation of Pigment Epithelial Detachment from SD-OCT

Abstract

Purpose : Pigment epithelial detachment (PED) is a prominent feature of exudative age-related macular degeneration (AMD), which often involves disrupted Bruch’s membrane (BM) and detached retinal pigment epithelium (RPE). We report a fully automated method to identify PED on clinical spectral-domain optical coherence tomography images (SD-OCT) by segmenting three related surfaces: upper border of RPE, lower border of RPE, and BM.

Methods : 70 patients with exudative AMD underwent SD-OCT imaging (Zeiss Cirrus, 512x128x1024 voxels, 6.01x6.05x2.0mm³, voxel size of 11.74x47.24x1.96µm³). 70 macula-centered volumetric scans were obtained. First, we segmented the OCT volumetric image using our previously reported intra-retinal layer segmentation approach (the Iowa Reference Algorithm). However, segmentation inaccuracies occasionally occurred around PED due to tissue loss and/or disordered sub-retinal layer. Therefore in the second step, we introduced a shape-prior based Graph-Cut---Graph-Search algorithm to segment the PED-related surfaces. The shape-prior models were derived from initial segmentations and were implemented as the regional constraints of PED shape likelihood. The automated segmentation results were compared with manually segmented lower and upper borders of RPE and BM. Absolute and relative differences were used as validation indices for evaluating the proposed method. We have created an elevation map of PED by identifying the local height of PED along each A-scan over the entire macula.

Results : The agreement between our novel automated and the manual segmentations was high: average relative difference was -1.70 ± 6.63 µm and average absolute difference was 5.01 ± 4.65 µm.

Conclusions : In this study, we show a fully automated method to segment PED in patients with exudative AMD. Our new method has markedly outperformed our previous segmentation in robustness and has potential to improve the diagnosis and management of exudative AMD.

This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.

 

Figure 1. Pigment epithelial detachment segmentation: (a) original B-scan; (b) 3D visualization of the automated PED segmentation; (c) example B-scan with automatically segmented PED-related surfaces: upper border of RPE (red), lower border of RPE (yellow) and BM (green); (d) same B-scan with manually segmented PED-related surfaces.