June 2021
Volume 62, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2021
Identification of choroidal neovascularization activity using deep learning
Author Affiliations & Notes
  • Xiaoyong Wang
    Genentech Inc, South San Francisco, California, United States
  • Simon S. Gao
    Genentech Inc, South San Francisco, California, United States
  • Michael Kawczynski
    Genentech Inc, South San Francisco, California, United States
  • Neha Anegondi
    Genentech Inc, South San Francisco, California, United States
  • Andreas Thalhammer
    Roche Pharma Research and Early Development Informatics, Basel, Switzerland
  • Thomas Bengtsson
    Genentech Inc, South San Francisco, California, United States
  • Jeffrey R Willis
    Genentech Inc, South San Francisco, California, United States
  • Jian Dai
    Genentech Inc, South San Francisco, California, United States
  • Footnotes
    Commercial Relationships   Xiaoyong Wang, Genentech, Inc. (E); Simon Gao, Genentech, Inc. (E); Michael Kawczynski, Genentech, Inc. (E); Neha Anegondi, Genentech, Inc. (E); Andreas Thalhammer, F. Hoffmann-La Roche Ltd. (E); Thomas Bengtsson, Genentech, Inc. (E); Jeffrey Willis, Genentech, Inc. (E); Jian Dai, Genentech, Inc. (E)
  • Footnotes
    Support  Yes, Genentech, Inc., South San Francisco, CA, provided support for the study and participated in the study design; conducting the study; and data collection, management, and interpretation
Investigative Ophthalmology & Visual Science June 2021, Vol.62, 2141. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Xiaoyong Wang, Simon S. Gao, Michael Kawczynski, Neha Anegondi, Andreas Thalhammer, Thomas Bengtsson, Jeffrey R Willis, Jian Dai; Identification of choroidal neovascularization activity using deep learning. Invest. Ophthalmol. Vis. Sci. 2021;62(8):2141.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : To develop an automated tool for choroidal neovascularization (CNV) detection using deep learning (DL) algorithms.

Methods : We evaluated 8527 optical coherence tomography (OCT) images from 521 patients in the as-needed arms of the HARBOR trial (NCT00891735) that evaluated ranibizumab in neovascular age-related macular degeneration (nAMD). Disease activity in the study eye was defined as fluid on OCT (eg, intraretinal fluid, subretinal fluid, subretinal pigment epithelial fluid) or if a patient's visual acuity decreased by ≥5 letters from the previous visit. Consequently, disease activity could be defined solely by the presence of retinal fluid associated with underlying CNV on OCT without the requirement of a decrease of ≥5 letters relative to the previous visit. In this subset of visits, study eye OCT scans of 1024×512×128 voxels collected from a Zeiss Cirrus machine were flattened toward the retinal pigment epithelium (RPE) layer, and cropped to 384 pixels above and 128 below the RPE. To accommodate the GPU memory constraints, the central 15 B-scans were selected as representatives. As such, the input size of each scan to the network was 512×512×15. 3618 scans from diseased eyes and 4909 from eyes without disease were split into training and test sets in a 4:1 ratio using stratified sampling. 5-fold cross-validation was applied only using the training set to optimize parameters. A Squeeze-and-Excitation (SE)-embedded MobileNet was designed to classify eyes with and without CNV. MobileNet greatly reduces model size using depth-wise separable convolutions and the SE module contributes by adaptively recalibrating channel-wise feature response. Weighted cross-entropy was used as a loss function because the number of samples in each class was unbalanced. Data augmentation, including rotation, translation, and flipping, was applied during training.

Results : The model was evaluated on a test set of 1706 images from 102 patients and achieved an area under the receiver operating curve of 0.81±0.012, with accuracy of 0.76±0.027, sensitivity of 0.66±0.028, and specificity of 0.83±0.029.

Conclusions : Our study demonstrated that a prototype DL model can accurately detect CNV disease activity based on changes in the retinal anatomy on OCT. While this algorithm needs to be validated on other datasets, it could potentially be applied for remote and automated monitoring of nAMD.

This is a 2021 ARVO Annual Meeting abstract.

 

Fig 1. Cross-validation ROC

Fig 1. Cross-validation ROC

×
×

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.

×