July 2019
Volume 60, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2019
Diagnostic Assessment of RNFL Segmentation using a Hybrid Deep Learning Approach
Author Affiliations & Notes
  • Jonathan D Oakley
    Voxeleron LLC, Pleasanton, California, United States
  • Suria S Mannil
    Byers Eye Institute, Stanford University, Palo Alto, California, United States
  • Daniel B Russakoff
    Voxeleron LLC, Pleasanton, California, United States
  • Robert Chang
    Byers Eye Institute, Stanford University, Palo Alto, California, United States
  • Footnotes
    Commercial Relationships   Jonathan Oakley, Voxeleron LLC (E), Voxeleron LLC (P); Suria Mannil, Santen (F); Daniel Russakoff, Voxeleron LLC (E), Voxeleron LLC (P); Robert Chang, Carl Zeiss Meditec (F), Santen (F)
  • Footnotes
    Support  Research to Prevent Blindness, Inc., and National Eye Institute (P30-EY026877), Stanford CIGH (Center for Innovation in Global Health)
Investigative Ophthalmology & Visual Science July 2019, Vol.60, 5599. doi:https://doi.org/
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Jonathan D Oakley, Suria S Mannil, Daniel B Russakoff, Robert Chang; Diagnostic Assessment of RNFL Segmentation using a Hybrid Deep Learning Approach. Invest. Ophthalmol. Vis. Sci. 2019;60(9):5599. doi: https://doi.org/.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : To combine a deep learning based optic nerve head (ONH) segmentation algorithm with volumetric layer segmentation and apply to retinal nerve fiber layer (RNFL) analysis using optical coherence tomography (OCT) data. Performance is then assessed based on confirmed glaucomatous and normal subjects.

Methods : OCT volumes of the ONH comprising 275 glaucoma and 237 controls were collected at the Stanford Univ. Dept. of Ophthalmology. Clinical designations were based on the presence of structural and/or visual field defects. 10 cases were excluded based on retinal tissue being outside of the scan’s axial field of view. Each had their ONH manually traced as viewed in the en face. A deep convolutional neural network using a U-Net architecture [Ronneberger 2015] comprising three sampling layers, dropout and batch normalization was subsequently trained to identify the ONH. 10-fold cross validation was used and the probabilistic output for each test fold was thresholded to generate the final mask. This result was then used to constrain an existing algorithm whose segmented layers includes the inner limiting membrane (ILM) and RNFL’s posterior boundary (Orion, Voxeleron, CA). The extracted peripapillary RNFL thickness was centered on the mask for the final analysis (Figure 1). Dice coefficients, defined as twice the sum of the intersecting mask pixels divided by the total mask pixels in both, were used to assess the final accuracy of the masks relative to the ground truth. Receiver operating characteristic (ROC) curves were used to gauge overall diagnostic efficacy.

Results : Dice coefficients for the deep learning based segmentation were excellent, averaging 0.96 (STD= 0.03) over all volumes (1 being exact overlap). When applied in conjunction with the layer segmentation software the RNFL area under the curves (AUCs) were 0.8, 0.92, 0.83 and 0.96 for temporal, superior, nasal and inferior quadrants, respectively; and 0.96 for the average (Figure 2).

Conclusions : Deep learning offers state of the art methods for object detection in images. Combined with an existing macula segmentation algorithm extends utility of the method to the ONH offering very good diagnostic performance.

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

 

Figure 1 – Example input en face (left) and resultant RNFL thickness map and ONH mask (right - in microns).

Figure 1 – Example input en face (left) and resultant RNFL thickness map and ONH mask (right - in microns).

 

Figure 2 – ROC curves for all RNFL summary parameters (N = 502).

Figure 2 – ROC curves for all RNFL summary parameters (N = 502).

×
×

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.

×