June 2017
Volume 58, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2017
Evaluation of a Segmentation Method to Estimate Corneal Endothelium Parameters
Author Affiliations & Notes
  • Juan Pedro Vigueras-Guillén
    Rotterdam Ophthalmic Institute, Rotterdam Eye Hospital, Rotterdam, Netherlands
    Quantitative Imaging Group, Delft University of Technology, Delft, Netherlands
  • Jeroen van Rooij
    Rotterdam Eye Hospital, Rotterdam, Netherlands
  • Hans G Lemij
    Rotterdam Eye Hospital, Rotterdam, Netherlands
  • Angela Engel
    Rotterdam Ophthalmic Institute, Rotterdam Eye Hospital, Rotterdam, Netherlands
  • Lucas J. van Vliet
    Quantitative Imaging Group, Delft University of Technology, Delft, Netherlands
  • Koenraad Arndt Vermeer
    Rotterdam Ophthalmic Institute, Rotterdam Eye Hospital, Rotterdam, Netherlands
  • Footnotes
    Commercial Relationships   Juan Vigueras-Guillén, None; Jeroen van Rooij, None; Hans Lemij, None; Angela Engel, None; Lucas J. van Vliet, None; Koenraad Vermeer, None
  • Footnotes
    Support  The Netherlands Organisation for Health Research and Development, TopZorg grants 842005002, 842005004 and 842005007
Investigative Ophthalmology & Visual Science June 2017, Vol.58, 677. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Juan Pedro Vigueras-Guillén, Jeroen van Rooij, Hans G Lemij, Angela Engel, Lucas J. van Vliet, Koenraad Arndt Vermeer; Evaluation of a Segmentation Method to Estimate Corneal Endothelium Parameters. Invest. Ophthalmol. Vis. Sci. 2017;58(8):677.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose : Parameters (or biomarkers) to assess the health status of the corneal endothelium, such as cell density (ECD), polymegethism (variation in cell size, CV) and pleomorphism (% of hexagonal cells, HEX), are derived from a segmentation of the endothelial cells. The accuracy of state-of-the-art methods to estimate these parameters is limited. We evaluate a new framework to automatically estimate such biomarkers in images obtained with a Topcon SP-1P specular microscope.

Methods : Thirty central cornea images from thirty eyes with advanced glaucoma (POAG, mean age 65.5) were captured. All images and Topcon data analyses were exported. An expert annotated the cells to create the ground truth. In the proposed method, segmentation was first done by a stochastic watershed method (Selig et al., BMC Medical Imaging 15:13, 2015), such that each cell was comprised of one or more fragments. A support vector machine inferred the likelihood for all combinations of two and three adjacent fragments to form a cell, based on fourteen features regarding shape, size, and intensity of the fragments. The fragments were merged iteratively, starting with the most likely combination, until no cases with a likelihood above 0.5 remained. Based on the resulting segmentation, the biomarkers were calculated.

Results : An illustrative example of both our and the built-in Topcon method is shown in Fig. 1. Both methods show some inaccuracy in following the exact cell edges. In contrast to the Topcon method, ours segments the whole image. In Fig. 2 we display the error of the estimated biomarkers with respect to the ground truth. To compare the two methods, a paired t-test was performed on each biomarker’s error, showing significantly better accuracy for our method for ECD (p=2.2x10-5), and comparable accuracy for CV (p=0.07) and HEX (p=0.28).

Conclusions : Inaccuracies in the segmentation of the cell boundaries have little direct impact on the biomarkers’ accuracy. However, they do cause errors while merging cell fragments, resulting in over- and undersegmentation, which adversely affect the accuracy. A better initial segmentation would therefore result in more accurate biomarkers. The proposed method significantly improves the estimation of ECD, which is the clinically most important parameter.

This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.

 

Example of a segmentation result by Topcon’s (C) and our method (D).

Example of a segmentation result by Topcon’s (C) and our method (D).

 

Error in ECD (top), CV (middle), and HEX (bottom) for Topcon's (blue) and our method (red).

Error in ECD (top), CV (middle), and HEX (bottom) for Topcon's (blue) and our method (red).

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×