Purchase this article with an account.
Luis de Sisternes, Theodore Leng, Qiang Chen, Jeffrey Ma, Vibha Mahendra, Daniel Rubin; Automated Drusen Segmentation and Quantification from SD-OCT Images to Predict AMD Progression. Invest. Ophthalmol. Vis. Sci. 2013;54(15):4154.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Characterization of non-exudative (dry) age-related macular degeneration (AMD) is currently based on the observation of very limited drusen characteristics in color fundus photographs, which has limited value in predicting patients whose AMD will progress. We propose tools to automatically segment and quantify drusen according to a series of novel features from spectral domain optical coherence tomography (SD-OCT) images, taking advantage of this imaging technique’s ability to resolve structures in the depth axis. This novel quantification can be informative in risk of progression to the wet form of AMD.
We obtained 2146 longitudinal SD-OCT scans acquired over a 5 year time interval from 350 eyes of 178 patients with AMD. We developed an iterative three-dimensional algorithm to automatically segment drusen in the scans. The automated segmentation results were compared to manual outlines drawn twice by two expert readers, resulting in differences similar to the ones observed among and within the readers. A series of novel quantitative features describing drusen size, shape, and reflectivity characteristics were automatically extracted, as well as features describing the evolution of drusen properties in longitudinal SD-OCT scans from the same patient. The most informative features in terms of correlation with future wet AMD status (dry or wet) were identified using Lasso regression.
Lasso regression analysis provided a set of informative drusen features to predict future progression to wet AMD. The top four predictive features were higher values in maximum drusen height and mean volume occupied per drusen, and lower reflectivity values inside drusen regions and number of independently identified drusen. The features selected as most informative also varied between imminent or long term progression prediction. Progression risk also increased with age, presence of wet AMD in the contralateral eye, and it was higher for female patients.
We have developed a fully automated method for drusen segmentation and quantification in SD-OCT images. Extracted quantitative features proved useful in developing models to predict the progression of dry-to-wet AMD. The results indicate that these novel quantitative features can be used to characterize the AMD disease process, identifying a subgroup of patients with increased risk to develop wet AMD.
This PDF is available to Subscribers Only