July 2018
Volume 59, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2018
Predicting eye laterality for non-mydriatic fundus images: A comparative study of deep learning and image processing.
Author Affiliations & Notes
  • Keyur Ranipa
    CARIn, Carl Zeiss India, Bengaluru, Karnataka, India
  • Krunalkumar Ramanbhai Patel
    CARIn, Carl Zeiss India, Bengaluru, Karnataka, India
  • Footnotes
    Commercial Relationships   Keyur Ranipa, Carl Zeiss India (E); Krunalkumar Ramanbhai Patel, Carl Zeiss India (E)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2018, Vol.59, 1728. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Keyur Ranipa, Krunalkumar Ramanbhai Patel; Predicting eye laterality for non-mydriatic fundus images: A comparative study of deep learning and image processing.. Invest. Ophthalmol. Vis. Sci. 2018;59(9):1728.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Often, the optometrist has to manually select the eye laterality, i.e. right eye (OD) or left eye (OS), from camera settings at the time of acquisition of a fundus image. For remote retinal screening programs, this approach is not clearly scalable. Hence, an algorithm for predicting the eye laterality from a retinal fundus image is an essential tool to automate the large-scale remote screening programs. In this paper, we evaluate a deep learning algorithm for eye laterality prediction and compare it with a conventional image processing algorithm.

Methods : We trained a deep convolutional neural network (CNN) with Tensorflow for the task of eye laterality prediction. For model training and testing, we recorded 968 OD and 968 OS images (either Fovea or Optic Nerve Head centered) taken with VISUSCOUT 100 (Carl Zeiss Meditec AG). We used 60% randomly chosen images for training while the remaining 40% is used for testing. Our CNN consists of 7 convolution layers, 2 fully connected layers and a loss layer for the task of eye laterality prediction. The image processing algorithm predicts the eye laterality by segmenting the Optic Disc and locating the centroid of the segmented optic disc in the image. If a centroid is located in the left region, the eye laterality is predicted as OS, otherwise as OD. We evaluated both algorithms on the same test set that consist of 388 OS and 388 OD images.

Results : Results are shown in Fig. 2. using confusion matrices. The deep learning algorithm achieves an Area Under Curve ( AUC ) of 99.99%, accuracy of 99.48% for OS and 100% for OD predications indicating an overall accuracy of 99.74%. The image processing algorithm achieves the accuracy of 97.42% for OS predications and 98.19% for OD predictions and overall accuracy of 97.80%.

Conclusions : We presented a comparative analysis of the task of eye laterality prediction. The accuracy obtained with a deep neural network clearly surpasses the accuracy of an image processing algorithm. Our study reveals that automated eye laterality prediction is possible and this solution should help enable the large-scale automation of remote retinal screening programs.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.

 

Fig. 1: Example Test Images

Fig. 1: Example Test Images

 

Fig. 2: Results in confusion matrices.

Fig. 2: Results in confusion matrices.

×
×

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.

×