Purchase this article with an account.
Katrina Mears, Li Zhang, Kyungmoo Lee, Xiayu Xu, Milan Sonka, Michael Abramoff; Automated Segmentation of Sub-Retinal Layers in Age-related Macular Degeneration. Invest. Ophthalmol. Vis. Sci. 2013;54(15):5531.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
In age-related macular degeneration (AMD), sub-retinal layers may be affected by accumulations of extracellular material (drusen) or fluids-associated abnormalities (sub-retinal fluid, pigment epithelial detachment). We report a fully automated method to identify the sub-retinal layers from macula-centered clinical spectral-domain optical coherence tomography (SD-OCT) scans with both dry and wet AMD.
23 patients with clinically significant AMD underwent SD-OCT imaging (Zeiss Cirrus, 200x200x1024 voxels, 6.0x6.0x2.0mm3, voxel size of 30.0x30.0x2.0µm3). 23 macula-centered volumetric scans were obtained from 23 eyes of these patients, meanwhile, 11 pairs of repeats scans of the same eye were also included in this study. The method consisted of two main steps: abnormalities detection and surface segmentation. In the abnormalities detection step, initial segmentation was performed using our previously reported intra-retinal layer segmentation method. However, segmentation errors occasionally occurred in the sub-retinal region in the vicinity of drusen and/or fluid-filled cysts (Fig. 1.B). In order to refine the segmentation result, a 3D supervised voxel-classification method was introduced to detect abnormal regions. Features were extracted based on initial segmentation and 3D probability maps created by assigning a likelihood value between 0 and 255, showing the spatial positions of abnormal regions (Fig. 1.C). In the surface segmentation step, a probability-constraint graph-search method was introduced to segment sub-retinal layers (Fig. 1.D), where the 3D probability maps were treated as cost function prior constraints. The thickness maps were computed as Euclidian distances between pairs of sub-retinal surfaces for both the initial segmentation and the proposed segmentation approaches(Figs. 1.E to F).
The results were evaluated by a retinal specialist (MDA) showing that the proposed method improved the sub-retinal layer segmentation in 21 out of 23 cases on the data set. Furthermore, reproducibility test for the average layer thickness of the sub-retinal layers was performed on the proposed method as well. The root mean square coefficient of variation (CV) was only 8.1% (Fig. 2).
In this study, we have developed a fully automated method to segment the sub-retinal layers. Our method produced marked improvement compared to our previous segmentation and showed high reproducibility.
This PDF is available to Subscribers Only