July 2019
Volume 60, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2019
A 3D Deep Learning System for Detecting Referrable Glaucoma Using Full OCT Macular Cube Scans
Author Affiliations & Notes
  • Daniel B Russakoff
    Voxeleron LLC, Pleasanton, California, United States
  • Suria S Mannil
    Byers Eye Institute, Stanford University, Palo Alto, California, United States
  • Jonathan D Oakley
    Voxeleron LLC, Pleasanton, California, United States
  • Robert Chang
    Byers Eye Institute, Stanford University, Palo Alto, California, United States
  • Footnotes
    Commercial Relationships   Daniel Russakoff, Voxeleron LLC (E), Voxeleron LLC (P); Suria Mannil, Santen (F); Jonathan Oakley, Voxeleron LLC (E), Voxeleron LLC (P); Robert Chang, Carl Zeiss Meditec (F), Santen (F)
  • Footnotes
    Support  Robert Chang 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, 1474. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Daniel B Russakoff, Suria S Mannil, Jonathan D Oakley, Robert Chang; A 3D Deep Learning System for Detecting Referrable Glaucoma Using Full OCT Macular Cube Scans. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1474.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : To develop and validate a deep learning system using full 3D optical coherence tomography (OCT) volumetric macular cube data for an automated binary glaucomatous optic neuropathy (GON) classification defined as referral or non-referral.

Methods : OCT macula cube scans from Cirrus HD-OCT (Carl Zeiss Meditec) were taken from a total of 282 eyes from 150 patients at the Stanford Univ. Dept. of Ophthalmology. Each scan was labelled as “Glaucoma Refer” or “Glaucoma Not Refer”, according to the criteria of glaucomatous cupping with visual field defect and need for treatment vs. normal. No glaucoma suspects were included. A novel convolutional neural network (CNN), gNet3D, was used to train a binary classification of “referral” vs. “non-referral.” To reduce data variance, each volume was homogenized using a cropping from the ILM to a fixed offset (390 microns) below Bruch’s membrane and resampled to a uniform size (128^3) using automated layer segmentation software (Fig. 1a) (Orion, Voxeleron). The CNN was evaluated by 5-fold cross validation with no eye’s scans in both the training and testing sets. We compared these results to the average ganglion cell - inner plexiform complex (GCIPL) thickness recorded in the Cirrus OCT reports for each of these eyes. We also ran the same analysis using the original, resized volumes without homogenization. We performed an occlusion sensitivity analysis to investigate which areas of the volumes were most useful in discriminating the two classes.

Results : Using the average thickness of the GCIPL as reported by the Cirrus, we found an area under the ROC curve (AUC) of 0.85. The 3D CNN with data homogenization achieved better performance with an AUC of 0.92 (Fig. 1b). The same CNN, run without the homogenization, achieved an AUC of 0.73. The occlusion sensitivity analysis indicated that the areas of the volume activating the most for determining referrals were in the inferior regions of the macula (Fig. 2).

Conclusions : A novel deep learning 3D CNN with layer segmentation-based preprocessing demonstrates enhanced predictive power for glaucoma referrals using SD-OCT macula cube scans.

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

 

(a) Illustration of the homogenization process to reduce data variance; (b) ROC curves generated using gNet3D, our deep learning framework, and the average GCIPL thickness output from the Cirrus.

(a) Illustration of the homogenization process to reduce data variance; (b) ROC curves generated using gNet3D, our deep learning framework, and the average GCIPL thickness output from the Cirrus.

 

Results from the occlusion sensitivity analysis averaged over a random sampling of 10% of the data.

Results from the occlusion sensitivity analysis averaged over a random sampling of 10% of the data.

×
×

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.

×