Investigative Ophthalmology & Visual Science Cover Image for Volume 59, Issue 9
July 2018
Volume 59, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2018
Identifying suspicious regions in optical coherence tomography angiography using convolutional neural networks
Author Affiliations & Notes
  • Conor Leahy
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Thomas Perez
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Mary K Durbin
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Amir H Kashani
    Keck School of Medicine, Los Angeles, California, United States
  • Toshinori Murata
    Department of Cardiovascular Research, Shinshu University Graduate School of Medicine, Matsumoto, Japan
  • Nathan Shemonski
    Carl Zeiss Meditec, Inc., Dublin, California, United States
  • Footnotes
    Commercial Relationships   Conor Leahy, Carl Zeiss Meditec (E); Thomas Perez, Carl Zeiss Meditec (C); Mary Durbin, Carl Zeiss Meditec (E); Amir Kashani, Carl Zeiss Meditec (F), Carl Zeiss Meditec (R); Toshinori Murata, Carl Zeiss Meditec (F), Carl Zeiss Meditec (R); Nathan Shemonski, Carl Zeiss Meditec (E)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2018, Vol.59, 1724. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Conor Leahy, Thomas Perez, Mary K Durbin, Amir H Kashani, Toshinori Murata, Nathan Shemonski; Identifying suspicious regions in optical coherence tomography angiography using convolutional neural networks. Invest. Ophthalmol. Vis. Sci. 2018;59(9):1724.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Interpretation of optical coherence tomography (OCT) Angiography data can be challenging to a non-expert. In this paper we propose a fully-automated method of identifying regions displaying pathology in OCT Angiography images, based on the use of convolutional neural networks.

Methods : A total of 242 OCT Angiography 3mm macula scans from 112 eyes and 64 patients were used for initial development and training of the automated algorithm. The scans were acquired using CIRRUS™ HD-OCT 5000 with AngioPlex® OCT Angiography (ZEISS, Dublin, CA) and consisted of a mix of healthy and diseased eyes. Included diseases were those known to affect the retinal vasculature such as diabetic retinopathy (DR) and retinal vein occlusions (RVO). The superficial slab was annotated at the pixel level by a human grader into 4 classes: 1) Normal-looking 2) Diseased-looking 3) Unsure 4) Ungradable/Poor image quality. Using these pixel-level annotations, a convolutional neural network (CNN) was trained. Data from PLEX® Elite 9000 (ZEISS, Dublin, CA) was used for testing the CNN. A pre-processing step was used to convert the OCT Angiography images from the PLEX Elite to look visually similar to the CIRRUS AngioPLEX en face images.

Results : Qualitatively, the trained CNN performed well with high specificity and moderate sensitivity. Figure 1 shows a few select results. In general, for cases of RVO, the region of the vein occlusion was very well delineated. For cases with DR, various regions in the scan were highlighted as diseased-looking, but not all of these might be considered suspicious by a human grader. For normal eyes, little or no regions were highlighted.

Conclusions : As has been seen in many other fields, convolutional neural networks are a powerful tool for performing tasks which are simple for humans, but deceivingly challenging for a step-by-step algorithm. With the inclusion of more data, the sensitivity of this technique is expected to improve.

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

 

Figure 1: Three sample image classifications. Regions highlighted in red were classified as diseased-looking and the other regions were classified as normal-looking.

Figure 1: Three sample image classifications. Regions highlighted in red were classified as diseased-looking and the other regions were classified as normal-looking.

×
×

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.

×