Premium
Error rate of automated choroidal segmentation using swept‐source optical coherence tomography
Author(s) -
Kong Mingui,
Eo Doo Ri,
Han Gyule,
Park Sung Yong,
Ham DonIl
Publication year - 2016
Publication title -
acta ophthalmologica
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.534
H-Index - 87
eISSN - 1755-3768
pISSN - 1755-375X
DOI - 10.1111/aos.12989
Subject(s) - optical coherence tomography , segmentation , frame rate , abnormality , word error rate , retinal , artificial intelligence , ophthalmology , choroid , medicine , computer science , nuclear medicine , retina , physics , optics , psychiatry
Abstract Purpose To investigate the error rate of automated choroidal segmentation and the effect of frame averaging on error rate. Methods A horizontal B scan at the fovea was performed in patients having various retinochoroidal disorders using swept‐source optical coherence tomography ( OCT ) with frame‐averaging technique. Scanned images were classified into four morphological groups: normal from fellow eyes ( N F ), normal from pathologic eyes ( N P ), retinal abnormality (R) and retinochoroidal abnormality ( RC ) group. Choroidal segmentation was automatically performed using built‐in software of a swept‐source OCT device, and the error rate of choroidal segmentation was analysed. Results Qualified images for all four averaging types with different number of averaged frames were acquired in 89 eyes of 77 patients. Images of 12, 20, 24 and 33 eyes were classified as N F , N P , R and RC group, respectively. The choroidal segmentation error was detected in 1–2 images (8.3–16.7%) in the N F group, 3–6 images (15.0–30.0%) in the N P group, 4–8 images (16.7–33.3%) in the R group and 17–19 images (51.5–57.6%) in the RC group. The error rate was significantly higher in RC group than other groups (p < 0.05). Increasing the number of frames for averaging showed no significant effect on the error rate in all groups (p > 0.05). Conclusion Automated choroidal segmentation showed a high error rate in images with choroidal abnormalities, and the averaging effect could not reduce the error rate significantly. Thus, further technological improvement is needed to increase the accuracy of the automated choroidal segmentation.