June 2020
Volume 61, Issue 7
ARVO Annual Meeting Abstract  |   June 2020
Using Artificial Intelligence to detect glaucoma and Age related Macula Degeneration
Author Affiliations & Notes
  • Bruno Lay
    ADCIS, Saint Contest, France
  • Ronan DANNO
    ADCIS, Saint Contest, France
  • Gwenole Quellec
    INSERM, France
  • Mathieu Lamard
    INSERM, France
  • Béatrice Cochener
    CHRU BREST, France
  • Ali Erginay
    AP-HP, France
  • Pascale Massin
    AP-HP, France
    EVOLUCARE, France
  • Footnotes
    Commercial Relationships   Bruno Lay, None; Ronan DANNO, None; Gwenole Quellec, None; Mathieu Lamard, None; Béatrice Cochener, None; Ali Erginay, None; Pascale Massin, None; ALEXANDRE LE GUILCHER, None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2020, Vol.61, 1647. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Bruno Lay, Ronan DANNO, Gwenole Quellec, Mathieu Lamard, Béatrice Cochener, Ali Erginay, Pascale Massin, ALEXANDRE LE GUILCHER; Using Artificial Intelligence to detect glaucoma and Age related Macula Degeneration . Invest. Ophthalmol. Vis. Sci. 2020;61(7):1647.

      Download citation file:

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

  • Supplements

Purpose : Based on existing image databases of diabetic patient for screening Diabetic Retinopathy (DR), and using the information about suspicion of glaucoma and Age related Macula Degeneration (AMD), this work involves the development of two algorithms to assess the presence of glaucoma and AMD.

Methods : Based on Artificial Intelligence (AI) solutions, glaucoma and AMD are detected using two distinct sets of convolutional neural networks (CNNs). The solution is a complement of an existing solution that automatically assesses the quality of the photography, and DR.
For training and evaluation purposes, 20 218 images with AMD labels and 62 260 images with glaucoma suspicion labels, with multiple ethnics and different acquisition devices, were collected in Paris Hospitals. For testing the glaucoma detector, a subset of the Chinese REFUGE database was also used. Unlike competing AI solutions, CNNs are jointly trained in such a way they are complementary with one another. Thanks to a proposed heatmap generation method, patterns that each CNN detects can be overlaid on images for pathology visualization.

Results : The algorithm for AMD detection is capable of detecting early signs of AMD such as drusen, as well as geographic atrophies and choroidal neovascularization. It also provides confidence levels for AMD. The sensitivity is 96.54% and the specificity is 95.40%. The system has an area under the ROC curve (AUC) of 0.9916. Achieved results are better than those obtained in two previous large studies.

In the case of glaucoma, the sensitivity is 92.50% and the specificity is 86.94%. The algorithm also provides a confidence level for the suspicion of glaucoma. The system has an AUC of 0.9644. Achieved results are similar to those already published.

Conclusions : The proposed jointly trained CNN set improves fully automatic detection of referable DR with the automatic determination of suspicion of glaucoma and AMD. It is a big step forward in the detection of pathologies in the retina. It provides more accurate predictions in less than a second using a fast GPU graphic processor. All algorithms are CE marked to provide a fully automated system to be used in hospitals, private practices, and mass screening networks.

This is a 2020 ARVO Annual Meeting abstract.


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.