Comparison of choroidal thickness measurements using semiautomated and manual segmentation methods

Abstract

SIGNIFICANCE This study demonstrated that a semiautomated segmentation method could help inexperienced practitioners to obtain choroidal thickness as good as experienced practitioners. PURPOSE The purpose of this study was to compare choroidal thickness measurements obtained by semiautomated and manual segmentation methods. METHODS Optical coherence tomography images of 37 eyes from 37 healthy young subjects acquired by a spectral-domain optical coherence tomography device were reviewed retrospectively. Two naive examiners measured choroidal thickness using manual and semiautomated methods, whereas two experienced examiners used only the semiautomated method. The semiautomated method referred to a fully automated segmentation program customized based on MATLAB and followed manual verification. After highlighting the inner and outer choroidal boundaries through automated segmentation, examiners reviewed these boundaries in each B-scan and conducted manual revisions if segmentation errors occurred. After selecting points where correct boundary was located, the software used a spline fit to blend the corrected region with the rest of the boundary. All measurements were summarized in a 6-mm Early Treatment Diabetic Retinopathy Study grid. Operation time spent to complete retinal and choroidal segmentation on each eye was recorded. Between-examiner agreements, that is, intraclass correlation coefficient and coefficient of reproducibility (CoR), were calculated among four sets of semiautomated measurements, and within-examiner agreements were comparisons between manual and semiautomated results from the same naive examiners. Eyes with thin or thick choroids were also analyzed separately. RESULTS The between-examiner and within-examiner agreements were excellent with intraclass correlation coefficient of 0.976 or greater. Pairwise within-examiner CoRs ranged from 17.4 to 47.1 μm. Pairwise between-examiner CoRs were between 13.0 and 38.9 μm. Eyes with thin choroid had better agreements than those with thick choroids. On average, naive examiners saved 3 to 5 minutes per eye using the semiautomated method. CONCLUSIONS With the help of a dedicated software, inexperienced practitioners could obtain choroidal thickness measurements with accuracy similar to experienced practitioners. Processing time with the semiautomated method was also reduced.