Author: Tommaso Tibaldi
Co-authors: Paolo Caselgrandi, Jonathan Oakley, Enrico Borrelli, Francesco Bandello, Michele Reibaldi
AbstractA comparison between two automated algorithms to segment the inner and outer retinal thicknesses in eyes with neovascular age-related macular degeneration
Authors: Tommaso Tibaldi, MD;1 Paolo Caselgrandi, MD;2 Jonathan Oakley, PhD;2 Enrico Borrelli, MD;3 Francesco Bandello, MD;3 Michele Reibaldi, MD.1
1 Department of Ophthalmology, University of Turin, Italy
2 Voxeleron LLC, San Francisco, California, United States
3 IRCCS San Raffaele Scientific Institute, Milan, Italy
Purpose: Manual segmentation of retinal layers from structural optical coherence tomography (OCT) may be time consuming and strictly dependent on graders’ performance. Conversely, automated segmentation allows for a rapid and reliable definition and quantification of retinal layers, even in diseased eyes. The aim of this study was to compare two algorithms for automated segmentation in eyes with neovascular age-related macular degeneration (AMD).
Settings: Retrospective study conducted at Medical Retina Unit at “Città della Salute e della Scienza” Hospital, University of Turin, Turin.
Methods: In this IRB-approved retrospective analysis, we collected data from 50 eyes with neovascular AMD from 50 patients who had structural OCT obtained using a Heidelberg Spectralis HRA+OCT (Heidelberg Engineering, Heidelberg, Germany) device. Each set of OCT scan consisted of 19 B-scans covering approximately 5.5x4.5-mm area centered on the fovea. The neurosensory retinal thickness within the circle of the ETDRS-grid centered over the fovea was measured using both the Spectralis built-in software (Heidelberg engineering software version 220.127.116.11) and automated OCT layer segmentation Orion software (Voxeleron, version 3.0.0). For each algorithm, measurements were automatically averaged across each of the following subfields: the central fovea subfield within the inner 1-mm-diameter circle, the inner circle subfield between the inner and middle 3-mm-diameter circles, and the outer circle subfield between the middle and outer 6-mm-diameter circles. Furthermore, the two circles were divided in four subfields (temporal, superior, inferior, and nasal). The two algorithms automatically identified and calculated the distance from the inner limiting membrane to the posterior of the inner nuclear layer, yielding a combined thickness of the inner retina. The outer retina was identified and measured as the combination of the outer plexiform and outer nuclear layers. Quantitative metrics obtained with the two algorithms were compared.
Results: In our study cohort of neovascular AMD eyes, the mean difference in inner retinal thickness between the two algorithms was 8.97 μm in the central fovea; 22.09 μm, 28.73 μm, 23.92 μm, and 25.11 μm in the inner circle (superior, temporal, inferior and nasal regions, respectively); and 16.02 μm, 18.47 μm, 17.58 μm, and 29.50 μm in the outer circle (superior, temporal, inferior and nasal regions, respectively). The mean difference in outer retinal thickness was 1.28 μm in the central fovea; 3.90 μm, 3.79 μm, 3.55 μm, and 2.80 μm in the inner circle (superior, temporal, inferior and nasal regions, respectively); and 4.92 μm, 0.16 μm, 4.15 μm, and 13.52 μm in the outer circle (superior, temporal, inferior and nasal regions, respectively). R-squared ranged from 0.80 for the inner retinal thickness in the superior region of the inner circle to 0.02 for the outer retinal thickness in the inferior region of the outer circle.
Conclusions: The automated segmentation provided by the two software was generally consistent, although differences suggest that they use different landmarks for segmentation and may not be used interchangeably.