Graph-based segmentation of corneal epithelium and endothelium in Optical Coherence Tomography images

Abstract

Purpose : To report on a novel robust graph-based segmentation method for the Epithelial and the Endothelial surfaces of abnormal cornea images obtained using high definition optical coherence tomography (HD-OCT).

Methods : Thirty-six patients were imaged using HD-OCT (Envisu R2210, Bioptigen, Buffalo Grove, IL, USA). These included patients with different pathologies: Dry eye (6 eyes), Keratoconus (6 eyes), Fuchs Dystrophy (6 eyes), Corneal Graft Rejection (6 eyes), Stem Cells Deficiency (6 eyes) and normal controls (6 eyes). All the images were manually segmented by two expert graders. The graph-based automated segmentation algorithm is based on a two stages approach. A directed graph was constructed using image pixels as the graph vertices. The Epithelial and the Endothelial boundaries were initially segmented using an edge energy between graph vertices based on the normalized gradient of the images. The initial segmentation of the Endothelium was fitted to a polynomial and the fit was used to define an additional directional energy term. A new edge energy that includes the gradient energy, the directional energy, and a penalty term is then used to refine the segmentation of the Endothelium. Examples of the segmentation results are shown in Fig. 1.

Results : The algorithm was able to successfully segment the Epithelial and the Endothelial surfaces of all included HD-OCT images. The inter-operator error between the two manual operators was 1.86±1.60 pixels for the Epithelial surface, 4.08±6.64 for the Endothelial surface and 2.91±4.85 pixels for both surfaces. The mean segmentation error between the graph-based segmentation and the manual operators was 2.04±1.52 pixels for the Epithelial surface, 3.05±3.51 pixels for the Endothelial surface and 2.53±2.71 pixels for both surfaces (Fig. 2). The segmentation errors for the graph-based segmentation algorithm errors were significantly less than the inter-operator errors (Wilcoxon Rank Sum Test; significance level = 0.05; p < 0.0001).

Conclusions : The proposed graph-based segmentation algorithm adapted for corneal HD-OCT images was objectively comparable to the manual operator segmentation in segmenting normal as well as pathological corneas.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Fig. 1 Examples of the graph-based segmentation results

 

Fig. 2 Segmentation error comparison