Purchase this article with an account.
Luis De Sisternes, Gowtham Jonna, Margaret Greven, Theodore Leng, Daniel Rubin; Accuracy, repeatability and reproducibility of a novel approach to quantify individual drusen in spectral-domain optical coherence tomography images. Invest. Ophthalmol. Vis. Sci. 2016;57(12):1619.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Drusen features in Spectral-Domain Optical Coherence Tomography (SD-OCT) have shown utility as biomarkers of age-related macular degeneration (AMD). We introduce a novel automated method to segment and quantify individual drusen in SD-OCT images, and evaluate the accuracy, repeatability and reproducibility of its measurements.
Our method uses a smooth interpolation of the Retinal Pigment Epithelium (RPE) outer boundary to identify regions of substantial lifting in the inner RPE and inner-segment boundaries. Individual drusen are then differentiated by a Watershed transform and morphological refinements in the obtained lift map. Drusen properties are measured automatically from the segmentation results: number, area, volume, extent, slope, height, density and intensity within drusen. We evaluated our approach in 192 early and intermediate AMD eyes from 129 patients, analyzing segmentation accuracy by measuring overlap ratio (OR) with manual markings, and measurement repeatability and reproducibility by computing their mean percentage difference (PD) in repeated scan pairs, acquired twice using 200x200 and 512x128 macular scan protocols. Drusen area and volume repeatability for each protocol and reproducibility across the two protocols were also compared to Cirrus Review software results.
Our segmentation method produced segmentations of high accuracy, showing similar agreement with manual markings (0.72±0.09 OR) to measured intra- and inter-reader variability (0.78±0.09 and 0.77±0.09 OR, respectively). Automated drusen area and volume measurements displayed high repeatability (7.2% and 9.63% PD for 200x200 and 512x128 protocols, respectively) and reproducibility (7.37% PD), and were more stable across protocols than the Cirrus Review software (14.38% PD, with an excellent 3.32% PD in 200x200 protocol but a worse 22.91% PD in 512x128 protocol). Measurements of drusen slope and mean drusen intensity showed significant differences across protocols.
Our method shows accurate segmentation results that are relatively stable in scans using the same protocol and also comparable across protocols, with the exception of drusen slope and mean intensity inside drusen due to intrinsic differences in the images. Cirrus quantification software shows a higher repeatability at the 200x200 protocol than using the 512x128 protocol.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.
This PDF is available to Subscribers Only