July 2018
Volume 59, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2018
Deep Learning Convolutional Neural Network for the Classification and Segmentation of In Vivo Confocal Microscopy Images
Author Affiliations & Notes
  • Pedram Hamrah
    Cornea/Ophthalmology, NEEC, Tufts Medical Center, Tufts University, Boston, Massachusetts, United States
    Center for Translational Ocular Immunology, Tufts Medical Center, Tufts University School of Medicine, Boston, Massachusetts, United States
  • Dilruba Koseoglu
    Center for Translational Ocular Immunology, Tufts Medical Center, Tufts University School of Medicine, Boston, Massachusetts, United States
    Cornea/Ophthalmology, NEEC, Tufts Medical Center, Tufts University, Boston, Massachusetts, United States
  • Ilya Kovler
    RSIP Vision, Jerusalem, Israel
  • Avi Ben Cohen
    RSIP Vision, Jerusalem, Israel
  • Ron Soferman
    RSIP Vision, Jerusalem, Israel
Investigative Ophthalmology & Visual Science July 2018, Vol.59, 1733. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Pedram Hamrah, Dilruba Koseoglu, Ilya Kovler, Avi Ben Cohen, Ron Soferman; Deep Learning Convolutional Neural Network for the Classification and Segmentation of In Vivo Confocal Microscopy Images. Invest. Ophthalmol. Vis. Sci. 2018;59(9):1733.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : In vivo confocal microscopy (IVCM) is a non-invasive imaging tool that allows visualization of corneal layers at a cellular level. It provides valuable information about immune dendritiform cells (DCs) that aid with the diagnosis of a variety of inflammatory corneal pathologies. Currently analyses of these images require manual image selection. The purpose of this study is to show utilization of a deep learning system (U-Net convolutional neural network (CNN)) for classification of corneal layer images as well as DC detection and segmentation, therefore increasing diagnostic accuracy

Methods : Classification CNN was trained on 1540 images consisting of 178 endothelial layer, 441 epithelial layer, 510 subbasal nerve layer and 411 stromal keratocyte images. The method was tested on a total of 610 images comprising of 100 endothelial layer, 99 epithelial layer, 281 subbasal nerves layer and 130 stromal keratocyte images. Sensitivy for layer Segmentation NN was trained on 208 images (6170 DCs) and tested on 25 images (280 DCs).

Results : Classification: Our results show an accuracy of 97% for corneal layers on the testing data set. Sensivity and specificity for layer detection was 97% and 98% or higher for all layers, with a AUC of 95% or higher for all layers. Detection: The sensitivity rate for DC detection was calculated as 85% (239 true positive detections from 280 DCs with 48 false positive and 41 false negative detections). Segmentation: The sensitivity for DC segmentation is 0.60 with a specificity of 0.99, mean segmentation AUC is 0.81.

Conclusions : IVCM is a beneficial diagnostic tool that enables the diagnosis of corneal pathologies that are clinically vague. Our results show that deep learning can be utilized in the selection and the segmentation of IVCM images.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.

 

×
×

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.

×