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
Automated Geographic Atrophy Segmentation in Infrared Reflectance Images Using Deep Convolutional Neural Networks
Author Affiliations & Notes
  • Zhihong Hu
    Retina, Doheny Eye Institute, Arcadia, California, United States
  • Ziyuan Wang
    Retina, Doheny Eye Institute, Arcadia, California, United States
  • Nizar Saleh Abdelfattah
    Retina, Doheny Eye Institute, Arcadia, California, United States
  • Jaya Sadda
    Retina, Doheny Eye Institute, Arcadia, California, United States
  • Srinivas R. Sadda
    Retina, Doheny Eye Institute, Arcadia, California, United States
  • Footnotes
    Commercial Relationships   Zhihong Hu, None; Ziyuan Wang, None; Nizar Abdelfattah, None; Jaya Sadda, None; Srinivas Sadda, Allergan (F), Allergan (C), Carl Zeiss Meditec (F), Genentech (F), Genentech (C), Iconic (C), Novartis (C), Optos (F), Optos (C), Thrombogenics (C)
  • Footnotes
    Support  BrightFocus Foundation Macular Degeneration, Macula Vision Research Foundation
Investigative Ophthalmology & Visual Science July 2018, Vol.59, 1714. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Zhihong Hu, Ziyuan Wang, Nizar Saleh Abdelfattah, Jaya Sadda, Srinivas R. Sadda; Automated Geographic Atrophy Segmentation in Infrared Reflectance Images Using Deep Convolutional Neural Networks. Invest. Ophthalmol. Vis. Sci. 2018;59(9):1714.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Geographic atrophy (GA) is an end-stage manifestation of age-related macular degeneration. Although several groups (including us) have reported semi-automated and automated GA segmentation approaches in fundus autofluorescence (FAF) images, we are not aware of any GA segmentation algorithm in infrared reflectance (IR) images. The purpose of this study is to evaluate the performance of a deep neural networks-based automated system for GA segmentation in IR images.

Methods : To segment GA lesions in IR images, we developed an automated system, using a novel artificial intelligence (AI) technique of deep convolutional neural networks (CNNs). A contrast limited adaptive histogram equalization approach was first applied on the original IR images to enhance the contrast of GA regions with the background. Blood vessels demonstrated a similar intensity level to GA in IR images, which could yield GA segmentation errors. We applied a tensor-voting technique to detect the blood vessels and applied a vessel inpainting technique to suppress the segmentation errors. To handle the large GA size variation, we applied three varying-sized deep CNNs. Seventy randomly chosen IR images (Heidelberg HRA+Spectralis) were obtained from 70 subjects with GA. The algorithm-defined GA regions were compared with manual delineations by certified graders. A two-fold cross-validation was applied to evaluate the algorithm performance.

Results : Fig.1 illustrates our IR GA segmentation performance. Quantitatively, the mean segmentation accuracy, true positive rate (i.e. sensitivity), true negative rate (i.e. specificity), and positive predictive value (i.e. precision) between the algorithm- and manually-defined GA regions, are 0.95 ± 0.03, 0.79 ± 0.05, 0.98 ± 0.00, and 0.79 ± 0.01, respectively.

Conclusions : We report a novel automated GA segmentation system in IR images, using an AI technique with deep CNNs. Quantitative comparison of automated IR segmentation results with manual delineation demonstrated high levels of agreement, indicating the feasibility of automated GA segmentation in IR images. Given that IR images are commonly obtained during the acquisition of optical coherence tomography images, and are more comfortable than FAF images, tools for automated GA segmentation from IR images may be more useful in clinical practice and research.

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

 

Fig.1. Example results of the automated GA segmentation

Fig.1. Example results of the automated GA segmentation

×
×

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.

×