Investigative Ophthalmology & Visual Science Cover Image for Volume 59, Issue 9
July 2018
Volume 59, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2018
Deep learning for prediction of AMD progression
Author Affiliations & Notes
  • Daniel B Russakoff
    Voxeleron LLC, Pleasanton, California, United States
  • Ali Lamin
    UCL Institute of Ophthalmology, London, United Kingdom
    Moorfields Eye Hospital, London, United Kingdom
  • Jonathan D Oakley
    Voxeleron LLC, Pleasanton, California, United States
  • Adam M Dubis
    UCL Institute of Ophthalmology, London, United Kingdom
    Moorfields Eye Hospital, London, United Kingdom
  • Susan Lightman
    UCL Institute of Ophthalmology, London, United Kingdom
    Moorfields Eye Hospital, London, United Kingdom
  • Sobha Sivaprasad
    UCL Institute of Ophthalmology, London, United Kingdom
    Moorfields Eye Hospital, London, United Kingdom
  • Footnotes
    Commercial Relationships   Daniel Russakoff, Voxeleron (E), Voxeleron (P); Ali Lamin, None; Jonathan Oakley, Voxeleron (E), Voxeleron (P); Adam Dubis, None; Susan Lightman, None; Sobha Sivaprasad, Allergan (F), Bayer (F), Boehringer Inglehein (F), Heidelberg Engineering (F), Novartis (F), Optos (F), Roche (F)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2018, Vol.59, 1717. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Daniel B Russakoff, Ali Lamin, Jonathan D Oakley, Adam M Dubis, Susan Lightman, Sobha Sivaprasad; Deep learning for prediction of AMD progression. Invest. Ophthalmol. Vis. Sci. 2018;59(9):1717.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Age-related macular degeneration (AMD) is a leading cause of vision loss for people in their 50’s and older. AMD proceeds in distinct stages from early, to intermediate, to advanced. In advanced AMD, blood vessel growth (choroidal neovascularization - CNV) causes irreversible damage to the photoreceptors which leads to vision loss. Patients can progress to advanced AMD without symptoms or any measurable change, underscoring the importance of determining which patients are at the highest risk for AMD progression. In the absence of compelling biomarkers enabling this, we have trained a deep learning-based classifier to help predict which patients will progress to advanced AMD based on OCT data.

Methods : 71 eyes of 71 patients were imaged using Topcon OCT (3D OCT-1000) over 2 years. These eyes were divided into 2 groups: patients who did not convert to advanced AMD (n = 40); and those who did (n = 31). The OCTs were taken before the development of CNV in Group 2. Volumes were decomposed into 9088 B-scans which were pre-processed using automated layer segmentation software (Orion, Voxeleron LLC), to identify the ILM and Bruch’s membrane. To help reduce variance, each B-scan was cropped from the ILM to a fixed offset (390 microns) below Bruch’s and resampled to a uniform size. A deep convolutional neural network (CNN) with 26 layers was trained using Theano over 500 epochs with early-stopping (patience = 8). The loss function was categorical cross-entropy. The CNN was evaluated using 5-fold cross validation such that no eye’s B-scan’s were in both the training and testing sets. To investigate the utility of the pre-processing, the same analysis using the original B-scans (resized) was performed.

Results : The area under the ROC curve for the pre-processed data was 0.78 at the B-scan level (Fig. 1). We also produced an aggregate prediction for each volume. The area under the ROC for the volume-level predictions was 0.82. For the original, unprocessed images, the ROC’s were, 0.69 and 0.70 respectively (Fig 1).

Conclusions : A deep learning CNN with layer segmentation-based preprocessing shows strong predictive power with respect to the progression of early/intermediate AMD to advanced AMD. Such adjunct analysis could be useful in, for example, setting the frequency of patient visits.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.

 

Fig 1 - ROC curves at the B-scan and volume levels for both pre-processed and unprocessed data.

Fig 1 - ROC curves at the B-scan and volume levels for both pre-processed and unprocessed 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.

×