Abstract
Purpose :
Retinal pigment epithelial detachment (PED) is a frequent and clinically important abnormality in eyes. Automated segmentation of Outer Retinal Pigment Epithelium (ORPE) and Inner Choroid (IC) are of relevance for the identification and quantification of PED (Fig.1a). Our purpose is to describe a fully automated method to simultaneously segment three PED-related surfaces: Internal Limiting Membrane (ILM), ORPE and IC, from spectral-domain optical coherence tomography (SD-OCT) volume cubes.
Methods :
The study inlcuded 15 SD-OCT (Heidelberg Spectralis) volume cubes from 15 eyes of 15 subjects (one eye per patient) with the presence of submacular vascularized/fibrovascular PED. Each OCT volume scan consisted of a macular cube of 1024 x 37 x 496 voxels with an average physical size of 5.76 mm x 4.44 mm x 1.92 mm. We segmented the OCT volumes using a novel graph search with truncated convex priors (Shah et al., MICCAI 2015). The method iteratively searches for the best segmentation solution in a sub volume of the image while enforcing truncated convex priors for surface smoothness and surface separation constraints. Segmentation results from our new method were compared with results obtained from the graph search method (Li et al., PAMI, 2006).
Results :
Segmentation of the various layers are shown in Figure 1b and 1c. The severity of segmentation errors was determined by averaging the vertical difference between the expert manual tracings and automated segmentations for all volumes. The quantitative comparison of the segmentation performance is summarized in Figure 2. The proposed method significantly lowered the error for ILM (p < 0.002), ORPE (p < 0.002) and IC (p < 0.002) compared to the graph search method. The proposed method with an average computation time of 746 seconds is much faster than the graph search method with an average computation time of 6082 seconds.
Conclusions :
The proposed method improves the segmentation accuracy and efficiency for the retinal surfaces ORPE and IC, which are crucial for automated identification and quantification of PEDs.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.