August 2019
Volume 60, Issue 11
Free
ARVO Imaging in the Eye Conference Abstract  |   August 2019
Automated segmentation of geographic atrophy using U-Net on custom-generated SD-OCT en face images
Author Affiliations & Notes
  • Niranchana Manivannan
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Luis de Sisternes
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Giovanni Gregori
    Department of Ophthalmology, Bascom Palmer Eye Institute, University of Miami Miller School of Medicine, Miami, Florida, United States
  • Philip J Rosenfeld
    Department of Ophthalmology, Bascom Palmer Eye Institute, University of Miami Miller School of Medicine, Miami, Florida, United States
  • Mary Durbin
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Footnotes
    Commercial Relationships   Niranchana Manivannan, Carl Zeiss Meditec, Inc. (E); Luis de Sisternes, Carl Zeiss Meditec, Inc. (E); Giovanni Gregori, Carl Zeiss Meditec, Inc. (F), Carl Zeiss Meditec, Inc. (P); Philip Rosenfeld, Carl Zeiss Meditec, Inc. (F), Carl Zeiss Meditec, Inc. (C); Mary Durbin, Carl Zeiss Meditec, Inc. (E)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science August 2019, Vol.60, PB0173. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Niranchana Manivannan, Luis de Sisternes, Giovanni Gregori, Philip J Rosenfeld, Mary Durbin; Automated segmentation of geographic atrophy using U-Net on custom-generated SD-OCT en face images. Invest. Ophthalmol. Vis. Sci. 2019;60(11):PB0173.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Geographic atrophy (GA) is a condition associated with loss of the retinal pigment epithelium (RPE) and represents the late stage of non-exudative age-related macular degeneration (AMD). This research aims to develop a fully-automated segmentation and quantification algorithm for spectral domain optical coherence tomography (SD-OCT) en face images.

Methods : This retrospective study used 250 macular cubes (512x128x1024: 58, 200x200x1024: 192) obtained from 155 patients using CIRRUS™ HD-OCT 4000 and 5000 (ZEISS,Dublin,CA). Experts manually drew the GA ground truth (GT) segmentations in the en face images. For each macular cube, a 3-channel en face GA projection image was generated by combining 1) sub-volume section of choroid; 2) slab projection surrounding RPE and 3) retinal thickness between the RPE and inner limiting (ILM) layer.
The training and testing sets of custom-generated en face images were comprised of 225 eyes (GA:187, drusen with no GA:19 and healthy:19) and 25 eyes (GA:11, drusen with no GA:5 and healthy:9). The contracting, bottleneck and expansive path of the U-Net consisted of 4 convolutional neural networks (CNN), 2 CNN with 0.5 dropout and 5 CNN blocks (fig. 1). Binary cross entropy and dice coefficient loss were used for training. ‘Icing on the Cake’ was used to fine-tune the model. Segmentations by the algorithm in the test set were compared with the GT using quantitative measurements (Bland-Altman, area and Pearson’s correlation).

Results : Fig. 2 shows the results of the proposed algorithm, Advanced RPE Analysis and the GT. The absolute and fractional area differences between GA regions generated by the proposed algorithm and the GT were 0.11±0.17mm2 and 5.51±4.7% as opposed to 0.54±0.82mm2 and 25.61±42.3% for Advanced RPE Analysis. The inference time was 1183 ms per en face image using an Intel® i7CPU. Correlations of GA areas generated by the proposed algorithm and Advanced RPE Analysis with the GT were 0.9996 (p-value<0.001) and 0.9259 (p-value<0.001). The Bland-Altman plot between the GT and the segments generated using proposed algorithm showed stronger agreement than advanced RPE analysis.

Conclusions : Quantitative and qualitative evaluations demonstrated that the proposed algorithm for segmenting GA in SD-OCT en face images showed very strong agreement with ground truth by manual grading.

This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.

 

Fig 1. Flowchart of proposed algorithm

Fig 1. Flowchart of proposed algorithm

 

Fig 2. Results of the proposed algorithm

Fig 2. Results of the proposed algorithm

×
×

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.

×