July 2019
Volume 60, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2019
Automatic laterality finding using deep learning in fundus images
Author Affiliations & Notes
  • Poojan Dave
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Katherine Makedonsky
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Niranchana Manivannan
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Patricia Sha
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Michael H. Chen
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Mary K Durbin
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Footnotes
    Commercial Relationships   Poojan Dave, Carl Zeiss Meditec Inc (E); Katherine Makedonsky, Carl Zeiss Meditec Inc (E); Niranchana Manivannan, Carl Zeiss Meditec Inc (E); Patricia Sha, Carl Zeiss Meditec Inc (E); Michael Chen, Carl Zeiss Meditec Inc (E); Mary Durbin, Carl Zeiss Meditec Inc (E)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2019, Vol.60, 1492. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Poojan Dave, Katherine Makedonsky, Niranchana Manivannan, Patricia Sha, Michael H. Chen, Mary K Durbin; Automatic laterality finding using deep learning in fundus images. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1492.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Determination of laterality is critical for fundus image analysis workflows. An automatic laterality finding algorithm would simplify the workflow and reduce errors caused by manual inputs from ophthalmic photographers or sensor detection. Such an algorithm would benefit all fundus imaging devices specifically handheld devices without physical sensors.

Methods : Training data consisted of 2374 CLARUS 500 images of pixel size 3072*3072 each with OD and OS labels (OD = 1296, OS = 1078). The test data consisted of 1756 images with OD and OS labeled by an expert grader (OD = 1010, OS = 746). Data augmentation methods were used to increase the size of the data set by two-fold. The training and test data had a mixture of healthy and diseased eyes collected using different fixations in a typical clinical setting.
As a preprocessing step, images were down-sampled to 512*512 after using an anti-aliasing filter and histogram adjustment. The Convolutional Neural Network (CNN) block consisted of three layers of CNN followed by one fully connected layer. Rectified Linear Unit (ReLu) activation was used for each CNN layer followed by Sigmoid activation for the fully connected layer. An inverted Dropout with a 0.5 probability was applied as a regularization scheme to the final layer to prevent the network from over-fitting. The binary labels generated from the network was compared with labels marked by expert graders in the test data set.

Results : The algorithm achieved an accuracy of 98.42% in detecting OD labels and 98.93% in detecting OS labels. The combined accuracy of the network is 98.63%. Figure 1 shows examples of images for correct and incorrect detections. The analysis of images with incorrect detections shows most of the images lack either an optic disc or parts of major blood vessels in the retina.

Conclusions : The laterality identifying algorithm presented here provides an accurate, fast solution for automatically finding laterality for handheld or tabletop fundus imaging devices. To our knowledge, this is the first time a machine learning algorithm is used to achieve >98% accuracy in laterality finding for widefield broad line fundus images.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Figure 1: The images in the top row are correctly classified by the algorithm for laterality and the images in the bottom row are incorrectly classified

Figure 1: The images in the top row are correctly classified by the algorithm for laterality and the images in the bottom row are incorrectly classified

×
×

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.

×