August 2019
Volume 60, Issue 11
Open Access
ARVO Imaging in the Eye Conference Abstract  |   August 2019
Automatic detection of optic nerve head in widefield OCT using deep learning
Author Affiliations & Notes
  • Ali Fard
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Homayoun Bagherinia
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Footnotes
    Commercial Relationships   Ali Fard, Carl Zeiss Meditec, Inc. (E); Homayoun Bagherinia, Carl Zeiss Meditec, Inc. (E)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science August 2019, Vol.60, PB0100. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Ali Fard, Homayoun Bagherinia; Automatic detection of optic nerve head in widefield OCT using deep learning. Invest. Ophthalmol. Vis. Sci. 2019;60(11):PB0100.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Localization of the optic nerve head (ONH) in OCT and OCTA images is of crucial importance for accurate analysis of the peripapillary region. Automated detection of this landmark is particularly challenging in widefield OCT images of diseased eyes due to the presence of pathologies. Here we present a robust deep learning method for automatic ONH detection in OCT images of healthy and diseased eyes.

Methods : Our method employs a convolutional neural network to segment ONH regions in OCT en face images. OCT en face images were generated from outer boundary of outer plexiform layer to 0.5mm below retinal pigment epithelium in 1064 6x6mm and 12x12mm OCT volumes acquired using PLEX® Elite 9000 SS-OCT (ZEISS, Dublin, CA). The ONH center was marked in the images (821 for training and 243 for testing) by human experts. All OS eyes were flipped along the vertical axis to match the OD eye for training purposes. A 3-channel U-net architecture was used in which 5 contracting and 5 expansive convolutional layers, ReLU activation, max pooling, binary cross entropy loss, and sigmoid activation in final layer were employed. The input channels were OCT en face, vessel enhance OCT en face, and OCT contrast map (intergral of absolute axial gradient) (Fig 1). A 4mm binary mask was created as the target image around the grader-indentified ONH center. The binary mask was shifted by 1-mm temporally due to the location of ONH at the edge of the scan. Data augmentation (rotation around the center between -9° and 9° with a step of 3°) was performed to increase the number of training images. The U-net predicted the ONH area followed by a template matching using a 4mm diameter disc to find the ONH center.

Results : Figure 2 shows examples of test images with predicted ONH area, and success rate and histogram plots of the error between the ONH detected by the algorithm and human expert. As it is evident, the algorithm identified the ONH with an accuracy of 300µm and 250µm in 98% and 95% of all cases, respectively. The mean and standard deviation of the error were found to be 97µm and 76µm.

Conclusions : Our results suggest that the ONH can be robustly and accurately detected in OCT en face images using a U-net architecture in presence of pathology.

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

 

Fig.1: Example of training input (A-C) and target output (D) of the algorithm.

Fig.1: Example of training input (A-C) and target output (D) of the algorithm.

 

Fig.2: (A1-A4): Test input and output images. (B-C) Success rate plot and histogram of error in detecting the ONH center.

Fig.2: (A1-A4): Test input and output images. (B-C) Success rate plot and histogram of error in detecting the ONH center.

×
×

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.

×