June 2021
Volume 62, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2021
Types of identified errors of Endothelium Cells segmentation from Corneal Specular Microscopy using a U-Net Network
Author Affiliations & Notes
  • Fernanda Szeremeta Abib
    Pontificia Universidade Catolica do Parana, Curitiba, PR, Brazil
  • Fernando Cesar Abib
    Anatomy, Universidade Federal do Parana, Curitiba, PR, Brazil
    Cornea, Prof. Fernando Abib Eye Clinic, Curitiba, PR, Brazil
  • Alceu de Souza Britto Jr
    Pontificia Universidade Catolica do Parana, Curitiba, PR, Brazil
  • Footnotes
    Commercial Relationships   Fernanda Abib, None; Fernando Abib, None; Alceu Britto Jr, None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2021, Vol.62, 2155. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Fernanda Szeremeta Abib, Fernando Cesar Abib, Alceu de Souza Britto Jr; Types of identified errors of Endothelium Cells segmentation from Corneal Specular Microscopy using a U-Net Network. Invest. Ophthalmol. Vis. Sci. 2021;62(8):2155.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : The traditional segmentation methods for Endothelial Cells in Non-Contact Corneal Specular Microscopy do not identify perfectly these cells, presenting distorted morphometric results, as demonstrated in ARVO 2016 by Ursulino, Hida, Holzchuh & Abib. The purpose of this cross-sectional study is to know the errors in determining the contours of Endothelial Cells (EC) generated by a U-Net network.

Methods : Images from two different contexts were used to train the U-Net: 1) 30 images from the ISBI Challenge: Segmentation of neuronal structures in EM stacks had their labels preprocessed 2) 78 Corneal Specular Microscopy images obtained automatically by Specular Microscopy Tomey EM-4000 in Prof Fernando Abib Eye Clinic, Brazil. A Data Augmentation algorithm using Keras library was used to increase the number of images to train the network. The EC training images were cropped to fit the size of 256x256 pixels, and their respective labels were binarized. The training labels from the ISBI Challenge were pre-processed. The images used to train the network were different from the ones used to test. The metric used to evaluate the training was the binary accuracy. With the endothelial images (30) of the tests performed by the U-Net, the tracing errors were classified as recommended by Ursulino, Hida, Holzchuh & Abib (ARVO 2016):
(Figure 1)
Type I - Non-counted cells;
Type II - Cell cluster;
Type III - Split cell.

Results : During the training phase, the binary accuracy obtained was of 93.14%. Type of Errors by U-Net identifying Endothelial Cells:
(Figure 2)
Type I - No case (0%) of cells not considered in the identified cells area;
Type II - 27 images (90%) didn’t have clusters, but the contours in some part of the images were determined in light gray and rarely partially traced; 3 images (10%) had clusters defined only with two cells;
Type III - 4 images (13.3%) presented only one split cell, the other 26 images (86,6%) did not present it;
In all 30 images (100%), in some part of the reticulum, the outline of the cells was in light gray.

Conclusions : The U-Net network showed a small number of errors in determining the contours of the endothelial cells. This technology is promising for clinical use requiring a very high number of images for the network training.

This is a 2021 ARVO Annual Meeting abstract.

 

 

×
×

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.

×