Algorithm for retinal segmentation in OCT imaging may save time, money

An automatic algorithm that provides accurate, reproducible segmentation of three retinal boundaries may reduce the time burden and labor costs of imaging drusen and geographic atrophy, a study found.

A general segmentation framework derived from graph theory and dynamic programming was used to segment spectral-domain optical coherence tomography images of eyes with dry age-related macular degeneration. To evaluative accuracy, layer thickness measurements from two graders were compared with outcomes from 220 B-scans in 20 patients.

To determine reproducibility, automatic layer volumes were also compared; these were generated from 0° vs. 90° scans in five volumes with drusen.

For automatic vs. manual segmentation, mean differences in measured thicknesses of the total retina and retinal pigment epithelium plus drusen complex (RPEDC) layers were 4.2 ± 2.8 µm and 3.2 ±2.6 µm, respectively. Mean differences in total retina and RPEDC volumes were 0.28 ± 0.28% and 1.60 ± 1.57%, respectively.

Moreover, average time per image was 1.7 seconds for the automatic method vs. 3.5 minutes for manual segmentation.

Of note, the algorithm was slightly less accurate in cases with geographic atrophy and drusen because of the different morphology of the RPEDC in this type of pathology, the study authors said. In the presence of subretinal drusenoid deposits, the algorithm also had a tendency to segment the retinal pigment epithelium, rather than the RPEDC. Consequently, a limitation is the need for human review to check for unanticipated errors.