Impact of segmentation density on spectral domain optical coherence tomography assessment in Stargardt disease.

Purpose: Automated spectral domain optical coherence tomography (SD-OCT) segmentation algorithms currently do not perform well in segmenting individual intraretinal layers in eyes with Stargardt disease (STGD). We compared selective B-scan segmentation strategies for generating mean retinal layer thickness and preserved area data from SD-OCT scans in patients with STGD1.
Methods: Forty-five eyes from 40 Stargardt patients were randomly selected from the ongoing Natural History of the Progression of Atrophy Secondary to Stargardt Disease (ProgStar) study. All eyes underwent SD-OCT using a standard macular volume consisting of 1024 × 49 equally spaced B-scans within a 20 × 20 degree field centered on the fovea. All 49 B-scans were segmented manually to quantify total retina, outer nuclear layer (ONL), photoreceptor inner segments, photoreceptor outer segments (OS), and retinal pigment epithelial layer (RPE). Mean thickness and total area were generated using all 49 B-scans (spaced 122 μm apart), 25 B-scans (every other B-scan, spaced 240 μm apart), 17 B-scans (every third scan, 353 μm apart), and 13 B-scans (every fourth scan, 462 μm apart), as well as by using an "adaptive" method where a subset (minimum 25 B-scans) of B-scans that the grader deemed as significantly different from adjacent B-scans were utilized. Mean absolute and percentage errors were calculated for macular thickness and area of different retinal layers for the different B-scan subset selection strategies relative to using all 49 B-scans, which was considered the reference or ground truth.
Results: Mean thickness and area measurements were significantly different for any regularly spaced reduction in B-scan density relative to the ground truth. When an adaptive approach was applied using a minimum of half the scans, the differences relative to ground truth were no longer significantly different. The mean percent differences for the area and thicknesses of the various layers ranged from 0.02 to 33.66 (p < 0.05 for all comparisons) and 0.44 to 7.24 (p > 0.05) respectively.
Conclusion: Manual segmentation of a subset of B-scans using an adaptive strategy can yield thickness and area measurements of retinal sublayers comparable to the reference ground truth derived from using all B-scans in the volume. These results may have implications for increasing the efficiency of SD-OCT grading strategies in clinical trials for STGD and other related macular degenerative disorders.