July 2019
Volume 60, Issue 9
Free
ARVO Annual Meeting Abstract  |   July 2019
Deep Learning Segmentation-Free Assessment of Retinal Nerve Fiber Layer in OCT Scans
Author Affiliations & Notes
  • Eduardo Bicalho Mariottoni
    Duke university, Durham, North Carolina, United States
  • Alessandro A Jammal
    Duke university, Durham, North Carolina, United States
  • Carla Urata
    Duke university, Durham, North Carolina, United States
  • Atalie C. Thompson
    Duke university, Durham, North Carolina, United States
  • Felipe A Medeiros
    Duke university, Durham, North Carolina, United States
  • Footnotes
    Commercial Relationships   Eduardo Mariottoni, None; Alessandro Jammal, None; Carla Urata, None; Atalie Thompson, None; Felipe Medeiros, Allergan (C), Allergan (F), Bausch&Lomb (F), Carl Zeiss Meditec (C), Carl Zeiss Meditec (F), Heidelberg Engineering (F), Merck (F), nGoggle Inc (F), Novartis (C), Reichert (C), Reichert (R), Sensimed (C), Topcon (C)
  • Footnotes
    Support  National Institutes of Health/National Eye Institute grant EY027651, EY025056 and EY021818
Investigative Ophthalmology & Visual Science July 2019, Vol.60, 1508. doi:https://doi.org/
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Eduardo Bicalho Mariottoni, Alessandro A Jammal, Carla Urata, Atalie C. Thompson, Felipe A Medeiros; Deep Learning Segmentation-Free Assessment of Retinal Nerve Fiber Layer in OCT Scans. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1508. doi: https://doi.org/.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : To develop a segmentation-free algorithm for assessing the thickness of the retinal nerve fiber layer (RNFL) on spectral-domain optical coherence tomography (SDOCT) scans and to investigate its performance in cases where the conventional automated segmentation algorithm fails

Methods : A deep learning (DL) convolutional neural network was initially trained to predict RNFL thickness measurements from 22,715 raw B-scan SDOCT images from 1,331 eyes of 701 subjects. Only images without automated segmentation errors according to human graders were used for this training. The dataset was split into 80% training and 20% validation set, and the automated segmentation curves were removed from the raw B-scan images before they were used to train and test the algorithm. After the DL algorithm learned to predict the RNFL thickness measurement, the output was compared to the original values extracted from the SDOCT report that had been generated by conventional automated segmentation. A separate dataset of 743 B-scan images that had previously been classified as containing segmentation errors was then used to test whether the DL segmentation-free algorithm or the conventional segmentation method provided RNFL estimates that better correlated with standard automated perimetry (SAP) measurements

Results : In images that had been classified as being free of segmentation errors according to human graders, the DL algorithm was able to successfully predict global RNFL thickness measurements that were highly correlated with the global RNFL obtained by automated segmentation (r=0.987, p<0.001; mean absolute error=2.05µm). In the 743 images that were classified as having segmentation errors from conventional automated segmentation, the DL segmentation-free algorithm was able to provide RNFL thickness estimates that showed a significantly higher correlation with the SAP mean deviation (MD) than the estimates provided by conventional automated segmentation (r=0.555 vs. r=0.297, p<0.001)

Conclusions : In SDOCT scans where the conventional automated segmentation had failed, a DL segmentation-free algorithm was able to provide assessments of RNFL thickness that were more strongly correlated with glaucomatous visual field measurements than those provided by the conventional automated method

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.

×