Purchase this article with an account.
Stephanie J. Chiu, Stefanie G. Schuman, Cynthia A. Toth, Joseph A. Izatt, Sina Farsiu; Automatic Segmentation of Retinal Layers in SDOCT Images with Age-Related Macular Degeneration Pathology. Invest. Ophthalmol. Vis. Sci. 2011;52(14):1305.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The segmentation of retinal layers in SDOCT images is crucial for the study and diagnosis of ocular diseases such as AMD. However, manual segmentation is time and labor intensive, and the delineation of pathological layers is subjective. Furthermore, current automatic techniques have difficulty segmenting images containing drusen or geographic atrophy (GA), and only address images of high quality. We have developed an algorithm to automatically segment three retinal boundaries in SDOCT images of varying image quality for patients with AMD.
We extended our general segmentation framework based on graph theory and dynamic programming to segment three retinal boundaries in images with drusen and GA: the vitreous-NFL border and the two boundaries isolating the RPE+drusen complex. Image flattening was performed based on pathology in the RPE, and weights were customized to effectively segment the RPE-choroid despite the lack in intensity contrast. A total of 20 volumes with non-neovascular AMD were selected with 10 containing drusen and 10 including GA. Each set of 10 volumes had 5 good and 5 low quality volumes. 11 B-Scans from each volume were segmented manually by two expert graders and automatically using our software.
We measured the average position of the three retinal boundaries in each B-scan. We calculated the absolute value difference of the average layer boundary positions between the manual and automatic estimates and likewise between the two manual expert graders. The mean and standard deviation of these differences across all B-scans are compared in Table 1.
Our automatic algorithm accurately segmented three retinal boundaries with a mean difference and standard deviation comparable to that of a second grader. The performance was reliable for images containing drusen and GA and for images varying in quality. This automatic approach will reduce time and labor costs and yield an objective evaluation for the study of AMD using real-world, clinical data.
This PDF is available to Subscribers Only