Spectral domain optical coherence tomography (SDOCT) has become an important diagnostic imaging modality in clinical ophthalmology [1
] for the examination of both the retina [1
] and cornea [2
]. SDOCT imaging can provide information about the curvature and thickness of different layers in the cornea, which are important for clinical procedures such as refractive surgery. To determine the required anatomical parameters, corneal layer boundaries must be reliably and reproducibly segmented. A corneal segmentation error of a several micrometers can result in significant changes in the derived clinical parameters [22
]. Unfortunately, the large volume of data generated from imaging in settings such as busy clinics or large-scale clinical trials makes manual segmentation both impractical and costly for the analysis of corneal SDOCT images.
To address this issue, several different approaches for segmenting corneal layer boundaries have been proposed with varying levels of success. One of the earliest reports by Li et al.
proposed a combined fast active contour (FAC) and second-order polynomial fitting algorithm for automatic corneal segmentation [12
]. Eichel et al.
implemented a semi-automatic segmentation method by utilizing Enhanced Intelligent Scissors (EIS), a user interactive segmentation method that requires minimal user input, and an energy minimizing spline [15
]. Despite the successful demonstrated accuracy in segmenting high-quality corneal images (e.g. Fig. 1.a), none of these techniques have demonstrated sufficient accuracy for fully automatically segmenting low-SNR images or those corrupted by different sources of artifacts (e.g. Figs. 1.b-d,
), which are inevitable in large scale clinical imaging.
Fig. 2 An example low-SNR corneal image (same OCT data as in Fig. 1.c) in which key regions and different types of imaging artifacts are labeled. Since SNR decreases with depth in SDOCT images, the regions of high and low-SNR also change. Some features, such (more ...)
In a closely related problem, utilizing graph theory in retinal SDOCT segmentation has been proven successful [17
]. The hybrid graph theory and dynamic programming retinal segmentation approach introduced by Chiu et al.
has been shown to be especially flexible for handling different sources of artifacts [21
In this paper, we use a customization of the hybrid graph theory and dynamic programming framework introduced by Chiu et al. for a fundamentally novel application that deals with unique imaging artifacts in corneal SDOCT images. Our method automatically segments three corneal layer interfaces: the epithelium-air interface (an interface between air and the tear film on the epithelium), the epithelium-Bowman’s layer interface, and the endothelium-aqueous interface as shown in . This robust segmentation method is capable of handling varying degrees of SNR and artifacts in corneal images. Illustrative examples of images used in our study are shown in
. Note that while many SDOCT corneal images have high SNR (), low-SNR images with artifacts are not uncommon in a clinical setting (). The anatomical features of interest, the key regions, and different types of imaging artifacts often seen in corneal SDOCT images are labeled in . Variable SNR in SDOCT corneal images results from differences in patient alignment, corneal hydration, tear film status, and patient motion during imaging. A central saturation artifact at the corneal apex and lower SNR in the periphery results from the telecentric (parallel) scan pattern used by most SDOCT corneal imagers. The horizontal line artifacts seen in result from interaction of the central saturation artifact with the DC subtraction algorithm applied to all SDOCT A-scans. All corneal images considered in this study were obtained using OCT systems manufactured by Bioptigen, Inc. and processed with Bioptigen software (InVivoVue Clinic v1.2), although similar artifacts are observed in corneal images obtained by instruments from other vendors.
Fig. 1 Corneal images of varying SNR and artifacts used in this study. (a) Corneal image with minimal artifacts. (b) Corneal image with prominent central and horizontal artifacts (see for the visual description and annotation of these artifacts). (c) (more ...)
The organization of this paper is as follows: Section 2 discusses layer segmentation using graph theory and dynamic programming, Section 3 shows an implementation of our robust algorithm for segmenting three corneal layer boundaries in images of varying quality and artifacts, Section 4 compares our automated results against expert manual segmentation, and concluding remarks are given in Section 5.