Purpose:
Spectral Domain Optical Coherence Tomography enables depth-resolved imaging of the cornea which is critical for the quantification of corneal layer thickness and curvature. Manual segmentation of these layers is often a time-consuming and subjective process. In addition, corneal images taken in the clinic often have low SNR and different types of artifacts. A fully automated and robust method to segment and quantify corneal layers would reduce labor and time costs and provide an objective evaluation of the corneal structures of interest.
Methods:
We extended our general segmentation framework based on graph theory and dynamic programming to handle corneal images. We take advantage of a priori anatomical information such as the normal range of epithelium, Bowman’s layer, and stroma thickness. Our robust algorithm detects critically low-SNR regions, and removes possible segmentation artifacts by interpolation results from the more reliable neighboring high-SNR regions. We tested the algorithm on a total of 20 B-scans from 10 normal adult subjects which were segmented manually by a cornea specialist and automatically using our software. To estimate inter-expert-observer variability, the same 20 B-scans were also graded manually by a trained non-expert.
Results:
We measured 3 layer boundaries on corneal images. We calculated the absolute value difference of the layer boundary between the manual and automatic estimates and likewise between the two manual expert graders. The mean and standard deviation of these differences across all B-scans are compared in Table 1.
Conclusions:
The automatic algorithm accurately segmented 3 corneal layer boundaries, with consistent results better or equal to the observed inter-expert-variability. This robust, automated approach may significantly reduce the resources and time necessary to conduct large-scale ophthalmic studies.
Keywords: cornea: clinical science • image processing • anterior segment