Purpose
<br /> Manual segmentation of choroidal boundaries is time consuming, tedious and may have inter observer variation. We developed an automated algorithm to segment choroid and provide choroidal volume measurements. Present study aims to validate the automated algorithm in comparison to expert manual segmentation.
Methods
Three sets of 97 spectral domain optical coherence tomography (SD-OCT) scans, obtained using enhanced depth imaging, were analyzed. Automated segmentation algorithm to detect the upper and lower boundaries of the choroid, was based on structural similarity, adaptive Hessian analysis and tensor voting. Single observer performed manual segmentation twice, in a masked fashion, to obtain intra-observer repeatability. Automated algorithm was validated using 291 SD-OCT scans from three 97-scan datasets as well as individual sets. Cross correlation coefficient and Dice coefficient was used to correlate manual and automated segmentation.
Results
Overall correlation coefficient and Dice coefficient between manual segmentation was 99.77± 0.19% and 96.73 ± 1.39% respectively. Overall correlation coefficient and Dice coefficient between average manual segmentation was 99.49 ± 0.41% and 93.25 ± 3.07% respectively. Correlation coefficients and absolute differences in choroidal volume between manual and automated segmentation for three sets are shown as Table 1 and 2 respectively.
Conclusions
Automated algorithm was in close agreement with manual segmentation with average correlation coefficient of about 99% and mean Dice coefficient of about 93%.