June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
Synthetic data facilitates deep-learning-based segmentation of OCT angiography images without human annotations
Author Affiliations & Notes
  • Martin Joseph Menten
    Technische Universitat Munchen, Munchen, Bayern, Germany
    Imperial College London, London, London, United Kingdom
  • Linus Kreitner
    Technische Universitat Munchen, Munchen, Bayern, Germany
  • Johannes C. C. Paetzold
    Technische Universitat Munchen, Munchen, Bayern, Germany
    Imperial College London, London, London, United Kingdom
  • Ahmed M Hagag
    NIHR Moorfields Biomedical Research Centre, London, Greater London, United Kingdom
    Boehringer Ingelheim International GmbH, Bracknell, United Kingdom
  • Sherry M. Bassily
    Watany Research & Development Center, Watany Eye Hospital, Cairo, Egypt
  • Sobha Sivaprasad
    NIHR Moorfields Biomedical Research Centre, London, Greater London, United Kingdom
    University College London Institute of Ophthalmology, London, London, United Kingdom
  • Daniel Rueckert
    Technische Universitat Munchen, Munchen, Bayern, Germany
    Imperial College London, London, London, United Kingdom
  • Alaa E Fayed
    Watany Research & Development Center, Watany Eye Hospital, Cairo, Egypt
    Department of Ophthalmology, Kasr Al-Ainy School of Medicine, Cairo University, Giza, Egypt
  • Footnotes
    Commercial Relationships   Martin Menten None; Linus Kreitner None; Johannes C. Paetzold None; Ahmed Hagag Boehringer Ingelheim International GmbH, Code E (Employment); Sherry Bassily None; Sobha Sivaprasad None; Daniel Rueckert None; Alaa Fayed None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 5450. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Martin Joseph Menten, Linus Kreitner, Johannes C. C. Paetzold, Ahmed M Hagag, Sherry M. Bassily, Sobha Sivaprasad, Daniel Rueckert, Alaa E Fayed; Synthetic data facilitates deep-learning-based segmentation of OCT angiography images without human annotations. Invest. Ophthalmol. Vis. Sci. 2023;64(8):5450.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Quantitative analysis of optical coherence tomography angiography (OCTA) images often requires segmentation of the blood vessels. So far, the application of deep learning to this task has been inhibited by a lack of training data with expert-derived labels. We propose a strategy that circumvents the need for human annotations by generating synthetic OCTA images and using these to train a segmentation neural network.

Methods : Synthetic OCTA images are created in two steps (see figure 1). Initially, the retinal vasculature is modeled using a statistical simulation. Blood vessels are iteratively grown until they supply the entire simulation volume with oxygen. Vessel growth is governed by a set of rules that enforces realistic physiological properties, such as vessel radius, density, bifurcation frequency, and branching ratio. The obtained vasculature models are converted to grayscale images. The images' contrast and noise are adapted to resemble that of real OCTA images using a generative adversarial network with cycle-consistency loss.
The synthetic images and intrinsically available ground truth labels are then used to train a nnU-Net for image segmentation. The neural network is tested on a dataset of real 3×3 mm2 OCTA images of 849 healthy eyes. We compare our method to a classical segmentation algorithm, optimal oriented flux (OOF), that relies on local vesselness filters and thresholding.

Results : The segmentation network, trained exclusively with synthetic training data, accurately localizes the blood vessels in real OCTA images (see figure 2). The approach facilitates the segmentation of images of both the superficial and deep vascular complexes.
The classical algorithm, OOF, is affected by image noise causing it to predict discontinuous vessel maps. This is reflected in the number of connected components for our and OOF’s segmentation maps (37 ± 29 vs. 394 ± 130). Furthermore, OOF fails to detect many of the smallest vessels.

Conclusions : Generating synthetic OCTA images and using them to train neural networks facilitates the accurate segmentation of real OCTA images. Our proposed method is not device-specific and does not require any manual annotations, making it a powerful tool for quantitative analysis of OCTA images.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Workflow to generate and utilize synthetic OCTA images.

Workflow to generate and utilize synthetic OCTA images.

 

Representative segmentation results of our method and the classical algorithm.

Representative segmentation results of our method and the classical algorithm.

×
×

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.

×