Purchase this article with an account.
Heena Divecha, Vincent Tai, Shwetha Mangalesh, Stephanie J Chiu, Katrina Winter, Cynthia Toth; Comparison of retinal layer thickness measurements from automated segmentation versus with manual correction for optical coherence tomography (OCT) at the fovea in preterm infants in the BabySTEPS study. Invest. Ophthalmol. Vis. Sci. 2020;61(7):3232.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Segmentation of retinal layers is a crucial step in interpreting results from OCT data. Manual segmentation is subjective and tedious for large data sets. Adult segmentation algorithms insufficiently address differences in retinal layers in the developing infant eye. We determine the efficacy of novel automated segmentation of retinal layers from macular OCT in preterm infants.
For 72 preterm infants enrolled in BabySTEPS (STudy of Eye imaging in Premature infantS), we analyzed foveal scans from both eyes acquired from an investigational, handheld, bedside, swept-source OCT device. Scans were not excluded for macular pathology. Foveal scans were segmented using a novel infant-specific automatic segmentation program, Duke Optical Coherence Tomography Retinal Analysis Program (versions 63.9) (Fig 1). Automated thickness measurements were compared to thicknesses after manual correction of segmentation lines by a senior grader (gold standard). We analyzed the difference in location of inner border of the internal limiting membrane, Bruch’s membrane and retinal thickness at the fovea center.
The 72 infants had 824 eye imaging sessions up to 42 weeks postmenstrual age (PMA). Segmentation was successful in 799 (99%) sessions. Fovea was not captured in 17 sessions. The gold standard foveal thickness ranged from 98-619μm with mean (standard deviation) 247.6 (115.6)μm due to developing infant retina and presence of macular edema in over half the infants in our cohort. Automated thickness measurements at the foveal center were in absolute agreement (zero difference) with the gold standard in 67% and within 25μm in 87% scans. The difference between automated and the gold standard was ≤ 5% in 78% of scans and ≤10% in 85% of the gold standard thickness.
The automated thickness measurements from a novel infant segmentation program were comparable to the thickness after manual correction of the segmentation lines in a vast majority of scans. This algorithm may be a useful research and future clinical tool in studies of the developing infant retina.
This is a 2020 ARVO Annual Meeting abstract.
Fig 1: Foveal OCT B-scans of a 41 weeks PMA infant: (A) unsegmented (B) automatic segmentation of 8 retinal layers (note lack of ellipsoid zone (blue) at fovea) (C) manual correction of automatic segmentation.
This PDF is available to Subscribers Only