Refinement of Choroid/Sclera Interface using Just-Enough-Interaction

Abstract

Purpose : Our group has previously developed a method for graph-based segmentation of the choroid-sclera interface (CSI) in 3D optical coherence tomography (OCT) images. However, the method may lead to inaccurate segmentation in some locations, despite its generally good performance. In this study, we have improved the method by proposing a user-guided approach to segment the CSI. This approach uses a just-enough-interaction (JEI) paradigm, which allows highly efficient minimal (just-enough) user interaction to refine the automated segmentation.

Methods : 20 subjects underwent SD-OCT imaging (Topcon, 512x128x885 voxels, 6.0x6.0x2.3mm³, voxel size of 11.72x46.88x2.60µm³). The 3D OCT volumetric images were first automatically segmented by our previously reported graph-based choroidal layer segmentation. The residual graph (a graph representation of the surface segmentation and cost image) was stored for further refinement. We then performed our proposed JEI approach: the user can correct the entire inaccurate region by approximately indicating only a few correct locations in the region. A polygon line is created for each region and used to locally modify the cost function for the CSI. The JEI refinement results and our previously automated segmentation results were both compared to the manual annotations of the CSI and their average absolute differences are reported in µm.

Results : As shown in Figure 1, the automatically determined surface result of the CSI was attracted by the vessel walls in the choroid. In the analyzed 20 subjects, the JEI refinement of the CSI significantly outperformed our previously reported automated CSI segmentation. The average absolute difference to the manual segmentation decreased substantially from 13.43 ± 7.87 µm (automated) to 4.10 ± 2.35 µm (JEI).

Conclusions : We have reported a user-guided approach to segment the CSI using the just-enough-interaction paradigm. Our new method has the potential to improve success rates of the quantitative OCT image analysis by including expert input in a highly efficient manner.

This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.

 

Figure 1. An example of just-enough-interaction approach to surface segmentation: (a) original B-scan; (b) initial automated segmentation of choroid-sclera interface with a local segmentation inaccuracy; (c) user roughly indicates the correct location using 2 markers (arrows); (d) refined surface segmentation of choroid-sclera interface.