Purpose:
Intraretinal layer segmentation is of paramount importance to monitor the progress of retinal diseases such as the variation of the nerve fiber layer and macular edema formation. The purpose of this study is to develop and validate a method that can automatically segment 10 intraretinal layers in 3-D macular OCT scans.
Methods:
14 macular OCT scans (200 × 200 × 1024 voxels, 6 × 6 × 2 mm3) were obtained from the right eyes of 14 normal subjects using a CirrusTM HD-OCT machine (Carl Zeiss Meditec, Inc., Dublin, CA). The 3-D graph search method hierarchically detected 11 retinal surfaces in 5 multiscale OCT volumes using gradient magnitudes of the OCT volumes. Reference standards were created by averaging in the z-direction manual tracings obtained from 2 retina specialists in 10 randomly selected X-Z images for each OCT scan. The accuracy of computer segmentation results was estimated by comparing to the reference standards in terms of unsigned border positioning error. In addition, the unsigned border positioning errors and absolute layer thickness differences between the computer segmentations and the reference standards were compared to the inter-observer variability between the 2 manual tracings.
Results:
The overall mean unsigned border positioning error of 11 retinal surfaces was 5.75 ± 5.11 µm (2.88 ± 2.55 voxels). While the unsigned border positioning errors of 9 retinal surfaces were significantly smaller than the unsigned border positioning differences between the 2 manual tracings (p < 0.01, 95% CI), those of 2 retinal surfaces were not significantly different (p > 0.81, 95% CI). In 9 out of 10 intraretinal layers, the absolute layer thickness differences between the computer segmentations and the reference standards were significantly smaller than those between the 2 manual tracings (p < 0.01, 95% CI).
Conclusions:
The proposed method is able to automatically segment 10 intraretinal layers from 3-D macular OCT scans and performs comparably to a retina specialist on this dataset in terms of unsigned border positioning error and layer thickness.
Keywords: image processing • imaging/image analysis: clinical • imaging/image analysis: non-clinical