July 2019
Volume 60, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2019
Simulating glaucomatous visual fields using a variational autoencoder
Author Affiliations & Notes
  • Samuel I Berchuck
    Statistical Science and Forge, Duke University, Durham, North Carolina, United States
  • Eduardo Bicalho Mariottoni
    Duke University, North Carolina, United States
  • Felipe A Medeiros
    Duke University, North Carolina, United States
  • Footnotes
    Commercial Relationships   Samuel Berchuck, None; Eduardo Mariottoni, None; Felipe Medeiros, None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2019, Vol.60, 1993. doi:https://doi.org/
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Samuel I Berchuck, Eduardo Bicalho Mariottoni, Felipe A Medeiros; Simulating glaucomatous visual fields using a variational autoencoder. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1993. doi: https://doi.org/.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Visual fields (VF) are a common assessment for detecting the existence and severity of glaucoma. VF progression is highly variable, making it difficult to obtain clinically relevant signal. To permit the development of future methods for studying VF data, we introduce a framework for simulating synthetic VFs, spanning the severity of glaucoma, that explores the full distribution of VF progression.

Methods : We analyzed longitudinal VF data acquired as part of an NIH sponsored R01 grant (EY021818, PI: Felipe Medeiros) from 2010 - 2014. The dataset consists 39,719 VFs from 5,600 eyes with glaucoma, or who are suspects or healthy. A variational auto-encoder (VAE), was used to learn the complex spatio-temporal distribution of the VF data, allowing for the simulation of synthetic VF patterns. The VAE is a powerful unsupervised generative machine learning algorithm, because it is not restricted to generating patterns observed in the existing data.

Results : Of the 5,600 patients, 2,169 (39%) have glaucoma, 1,302 (23%) are suspects, and 2,129 (38%) are healthy. The VAE was trained on 92% of the patients in the study and the remaining 8% were used for a test sample. Patients were randomly included in the training and test sets with all three disease designations represented in each. Figure 1 presents the learned distribution of VFs across the spectrum of patterns and reflects the continuum of VF progression. Finally, Figure 2 presents an 8 by 8 grid of VFs that have been generated from the VAE.

Conclusions : The VAE was shown to effectively generate synthetic VFs that resemble what is seen clinically. Furthermore, the patterns generated are not restricted to those seen in the dataset, allowing for the VAE to explore a large domain of VF variability. With further development, the VAE simulation framework can be used to generate benchmark VF data upon which to better understand and validate new methods for glaucoma research.

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.

×