June 2020
Volume 61, Issue 7
Free
ARVO Annual Meeting Abstract  |   June 2020
Deep learning-based automatic segmentation of intact ellipsoid zone area on optical coherence tomography images of USH2A-related retinal degeneration
Author Affiliations & Notes
  • Sina Farsiu
    Duke University, North Carolina, United States
  • Jessica Loo
    Duke University, North Carolina, United States
  • Jacque L Duncan
    UCSF, California, United States
  • David G Birch
    Retina Foundation of the Southwest, Texas, United States
  • Glenn J Jaffe
    Duke University, North Carolina, United States
  • Footnotes
    Commercial Relationships   Sina Farsiu, Google (R); Jessica Loo, None; Jacque Duncan, 4D Therapeutics (C), AGTC (C), Biogen/Nightstar Therapeutics (C), Editas Medicine (C), Eloxx (C), Foundation Fighting Blindness (C), ProQR Therapeutics (C), SparingVision (C), Spark Therapeutics (C), Vedere Bio (C); David Birch, None; Glenn Jaffe, None
  • Footnotes
    Support  The source of the data is the FFB Consortium, but the analyses, content and conclusions presented herein are solely the responsibility of the authors and have not been reviewed or approved by the Consortium and may not reflect the view of FFB.
Investigative Ophthalmology & Visual Science June 2020, Vol.61, 485. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Sina Farsiu, Jessica Loo, Jacque L Duncan, David G Birch, Glenn J Jaffe; Deep learning-based automatic segmentation of intact ellipsoid zone area on optical coherence tomography images of USH2A-related retinal degeneration. Invest. Ophthalmol. Vis. Sci. 2020;61(7):485.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : To develop an automatic method to segment the intact ellipsoid zone (EZ) area on spectral domain optical coherence tomography (SD-OCT) images of USH2A-related retinal degeneration

Methods : USH2A is a common gene mutation resulting in retinal degeneration. The dataset consisted of SD-OCT volumes of a subset of 93 eyes with measurable EZ (from a total of 127 eyes) enrolled in an international, multi-center clinical trial (NCT03146078) characterizing the natural progression of USH2A-related retinal degeneration (Usher syndrome type 2A or non-syndromic retinitis pigmentosa). Each SD-OCT volume consisted of 121 B-scans × 1536 A-scans except two volumes with 49 B-scans × 512 A-scans and one volume with 49 B-scans × 1024 A-scans. The B-scans were manually segmented by expert Readers to obtain an en face binary map of the intact EZ area for each volume. The dataset was split into non-overlapping training, validation, and testing sets in a 4:1:1 ratio. A convolutional neural network (CNN) was developed and trained to classify each A-scan as having an intact EZ or not, using A-scan clusters randomly sampled from the training set. During testing, given an SD-OCT volume, the trained CNN was used to automatically predict an en face probability map of the intact EZ area. The probability map was thresholded and post-processed to obtain the final en face binary map of the intact EZ area. To evaluate the performance of the CNN, the Dice similarity coefficient (DSC) between the binary maps and the correlation coefficient (Pearson’s r) of the intact EZ areas were calculated between the manual and automatic segmentations.

Results : Using 6-fold cross validation, the mean DSC was 0.83 (SD: 0.21, Median: 0.92) and Pearson’s r was 0.97 across all 93 volumes. Figure 1 shows an example. Upon further qualitative assessment, many differences occurred in areas where the existence of the EZ was unclear and difficult even for expert Readers to segment.

Conclusions : Overall, there was good agreement between the manual and automatic segmentations of the intact EZ area. This algorithm will be useful to quantify the intact EZ area for diagnosing and monitoring USH2A-related retinal degeneration.

This is a 2020 ARVO Annual Meeting abstract.

 

Figure 1: Example of manual and automatic segmentations of the intact EZ area for an SD-OCT volume. The B-scans correspond to the position marked by the yellow line on the en face image.

Figure 1: Example of manual and automatic segmentations of the intact EZ area for an SD-OCT volume. The B-scans correspond to the position marked by the yellow line on the en face image.

×
×

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.

×