July 2019
Volume 60, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2019
Predicting retinal thickness from fundus images across modalities using deep learning
Author Affiliations & Notes
  • Olle Gottfrid Holmberg
    Helmholtz Zentrum München - German Research Center for Environmental Health, Institute of Computational Biology, Neuherberg, Munich, Germany, Neuherberg, Germany
  • Karsten Ulrich Kortuem
    Ludwig-Maximilian-University, University Eye Hospital Munich, Munich, Bayern, Germany
    Moorfields Eye Hospital, United Kingdom
  • Niklas Köhler
    Helmholtz Zentrum München - German Research Center for Environmental Health, Institute of Computational Biology, Neuherberg, Munich, Germany, Neuherberg, Germany
  • Fabian Theis
    Helmholtz Zentrum München - German Research Center for Environmental Health, Institute of Computational Biology, Neuherberg, Munich, Germany, Neuherberg, Germany
  • Footnotes
    Commercial Relationships   Olle Holmberg, Heidelberg Engineering (R); Karsten Kortuem, Bayer (F), Bayer (R), Heidelberg Engineering (R), Novartis (F), Novartis (R), Zeiss (R); Niklas Köhler, None; Fabian Theis, None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2019, Vol.60, 1519. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Olle Gottfrid Holmberg, Karsten Ulrich Kortuem, Niklas Köhler, Fabian Theis; Predicting retinal thickness from fundus images across modalities using deep learning. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1519.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Retinal thickness is an important marker for visual acuity in every retinal disease. This information can be derived best from an optical coherence tomography (OCT) scan with multiple retinal layer segmentation tools. However, access to OCT imaging devices is limited. The aim of this study is to develop a deep learning based method that is able to infer the thickness information directly from the widely available fundus infrared (IR) imaging data thus enriching existing IR observations.

Methods : We used the OCT and fundus IR scans acquired from standard medical examinations of a large specialized clinic for Ophthalmology. 677 patients were used for training, 98 patients as a separate holdout for validation and 128 patients for testing and reporting results. To generate accurate high-resolution ground truth thickness maps from the OCT images we developed an Unet-based segmentation model, which we trained on public available OCT segmentation datasets and refined with a transfer learning approach onto our clinical data. The resulting high-resolution thickness maps were coregistered with the fundus images, used to train thickness regression networks.

Results : Overall, the mean deviation of all predicted thickness maps from the ground truth is 12.7 % with a standard deviation of 8.0 %. Furthermore, plotting the results of our algorithms reveals that captures complex thickness patterns like diabetic macular edema as well as pathological thickness distortions originating from epiretinal membranes, cystitis macular edema, and choroidal neovascularisations.

Conclusions : Using modern deep learning (DL) techniques, we are able to infer thickness information directly from IR fundus images vastly simplifying the process of acquiring retinal thickness information. This algorithm could be the basis of a valuable tool in pre-screenings especially in areas with limited access to OCT imaging devices.

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

×
×

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.

×