Abstract
Purpose :
In order to utilize the abundance of data generated by spectral domain optical coherence tomography (SD-OCT) and to facilitate the advances in diagnosis, tools are needed to perform downstream analysis in a prompt manner. Such analyses include generation of retinal layer thickness maps and OCT enface images. Here we characterize an automated multi-retinal layer segmentation algorithm (MLS) for fast and reliable quantification of seven intra-retinal layer boundaries in retinal OCT images.
Methods :
Both normal subjects and subjects with abnormalities (e.g. age-related macular degeneration(AMD), diabetic retinopathy(DR), vitreo-retinal interface abnormalities (VRI)) were imaged by prototype SD-OCT system with 2.9 mm scan depth. The data set consisted of SD-OCT volumes acquired using angiography scan patterns. The scan types included: 3x3mm, 6x6mm, HD 6x6mm over 2mm depth, 8x8mm, HD 8x8mm, and 12x12mm over 2.9 mm depth. The segmentation of datasets (separating seven retinal layer boundaries (ILM, outer RNFL, outer IPL, outer INL, outer OPL, IS/OS, RPE) over five specified B-scan per OCT volume) were performed by 1) two human graders using EdgeSelect, a manual segmentation method, and 2) MLS algorithm. Both segmentation results were evaluated by a clinician based on the clinical acceptance.
Results :
A total of 290 B-scans were evaluated for each category (Normal: 75 B-scans; AMD:55 B-scans; DR:45 B-scans; VRI:55 B-scans; other abnormalities: 60 B-scans). Correlation plots and Bland Altman plots were generated and corresponding parameters (correlation coefficient and 95% confidence interval (±1.96 SD) of Bland Altman plots were shown in both figure 1 (manual vs manual) and figure 2 (average manual vs MLS)) were reported for each layer boundary across 290 B-scans.
Conclusions :
Two graders showed strong correlation and good agreement using EdgeSelect, the manual segmentation method. A high correlation and good agreement between average manual segmentation and MLS were also demonstrated, indicating the results from MLS is as accurate as the results from human graders.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.