Purchase this article with an account.
Alec Arshavsky, Francesco LaRocca, Joseph Izatt, Anthony Kuo, Sina Farsiu; Automatic Segmentation of Corneal Lamellar Graft Surfaces from Low-Contrast Optical Coherence Tomography Images. Invest. Ophthalmol. Vis. Sci. 2013;54(15):5530.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Endothelial keratoplasty (EK) is a widely used form of corneal transplantation for corneal endothelial disease. By using OCT to examine cut donor tissue for EK, characterization of the lamellar cut is possible. We present methods to automatically identify the intrastromal lamellar dissection (ILD) and the endothelial surfaces of precut donor tissue. We previously developed an algorithm for segmentation of corneal layers in normal eyes (LaRocca, et al. BOE, 2011). However, the quality of many OCT images from donor tissue in storage is exceptionally low (Fig 1). Moreover, LaRocca’s algorithm was not designed to identify the ILD surface. Thus, in this work, we developed a dedicated algorithm for these low-quality images.
We introduce a novel, more robust implementation of our graph theory and dynamic programming (GTDP) segmentation framework. We first attain a pilot estimate of the target surfaces by classic GTDP. Next, we find erroneously segmented surfaces based on morphologic a priori assumptions about grafts and iteratively refine the reliable fragments of each segmented surface. We achieve the final estimate of the surfaces by finding the shortest path from a graph of these reliable fragments. We tested our new algorithm on a total of 250 B-scans (each of size 512×1000) from 5 donor eyes. An expert corneal surgeon evaluated and reported the percentage of correctly segmented surfaces not requiring manual correction. The proportion of correctly segmented surfaces between the new and previous LaRocca algorithm was compared using Fisher’s exact test.
Our new automatic algorithm correctly segmented the ILD in 83.6% of images compared to 4.8% by the LaRocca algorithm (p<0.0001). Our new automatic algorithm correctly segmented the endothelial layer in 66.0% of images compared to 0.0% by the LaRocca algorithm (p<0.0001).
The new automatic algorithm accurately segmented the ILD and the endothelial surfaces in the majority of OCT images from donor corneas and was significantly better than the prior algorithm. This robust automated approach may significantly reduce the resources and time necessary to conduct large-scale screening of precut corneal donor tissue.
This PDF is available to Subscribers Only