Purchase this article with an account.
Nripun Sredar, Kevin Ivers, Hope M Queener, George Zouridakis, Jason Porter; Automated segmentation of lamina cribrosa microarchitecture from in vivo adaptive optics scanning laser ophthalmoscope images. Invest. Ophthalmol. Vis. Sci. 2015;56(7 ):1001. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Previous quantifications of anterior lamina cribrosa surface (ALCS) pores from in vivo adaptive optics scanning laser ophthalmoscope (AOSLO) images were based on the subjective, manual marking of pore boundaries. We sought to develop an automated method to objectively segment and analyze ALCS pores and beams from AOSLO images.
In vivo AOSLO images of ALCS microarchitecture were acquired in 6 normal and 3 experimental glaucoma (EG) non-human primate eyes and 2 normal human eyes. ALCS pores were segmented using an automated hierarchical method that identified pore locations using a statistical region merging method and determined pore boundaries using an active contour level sets method. Pores were manually segmented in the same images and compared with pores segmented using the automated method. The sensitivity and specificity of the automated method was computed on a per-pixel basis dependent on whether pixels were classified as pores or beams (gold standard = manual segmentation). The repeatability of calculating ALCS pore area using automated segmentation was assessed following registration of images taken at 2 different time points (mean separation = 15 ± 7 weeks) in 3 normal eyes. Only pores that were segmented in both sessions were analyzed.
The mean sensitivity and specificity of the automated method for 11 eyes were 80.9 ± 2.9% and 96.8 ± 1.6%. There were no significant differences in ALCS pore area calculated via automated segmentation at 2 different time points in 3 normal eyes (Paired t-test, P>.05). The percentage of pores identified by automated segmentation at corresponding locations in images from both sessions in 3 normal eyes were 89%, 93% and 92% (n=72, 42 and 48 pores).
The quantification of ALCS pores via automated segmentation was comparable to manual segmentation results and was repeatable over time. The decreased sensitivity (relative to specificity) is likely due to errors in automated segmentation of pores in areas of low contrast. Our automated segmentation method can be used to objectively characterize laminar microarchitecture in normal and glaucomatous eyes.
This PDF is available to Subscribers Only