Investigative Ophthalmology & Visual Science Cover Image for Volume 65, Issue 9
July 2024
Volume 65, Issue 9
Open Access
ARVO Imaging in the Eye Conference Abstract  |   July 2024
Using Latent Diffusion Models to Generate Synthetic Optical Coherence Tomography Scans
Author Affiliations & Notes
  • Pranay Jain
    MITRE, Arlington, Virginia, United States
  • Salim Semy
    MITRE, Arlington, Virginia, United States
  • James Dienst
    MITRE, Arlington, Virginia, United States
  • Kerry Goetz
    National Eye Institute, Bethesda, Maryland, United States
  • Footnotes
    Commercial Relationships   Pranay Jain, None; Salim Semy, None; James Dienst, None; Kerry Goetz, None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2024, Vol.65, PB0014. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Pranay Jain, Salim Semy, James Dienst, Kerry Goetz; Using Latent Diffusion Models to Generate Synthetic Optical Coherence Tomography Scans. Invest. Ophthalmol. Vis. Sci. 2024;65(9):PB0014.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : We evaluated the ability of machine learning models to generate realistic synthetic OCT scans that can be shared without restrictions for research and clinical diagnostic applications. Synthetic OCT scans have multiple uses, including the design, development, and benchmarking of novel algorithms to detect presence of disease, monitor progression, and predict effectiveness of various treatment modalities using OCT images.

Methods : Initially, synthetic 256x256 OCT foveal b-scans were generated using a generative adversarial network, specifically StyleGan3, and a Latent Diffusion Model (LDM). Images were evaluated using Frechet Inception Distance (FID) and subject matter expert (SME) review to determine ability of synthetic OCT scans to mimic real OCT scans. We similarly generated synthetic 512x512 OCT foveal b-scans generated using a LDM. Models were trained with publicly available datasets across four disease classes: choroidal neovascularization (CNV) (40K images), diabetic macular edema (DME) (11K images), drusen (9K images), and normal (26K images). SME experiments consisted of 40 synthetic and 40 real OCT scans evenly distributed across the four disease classes, Figure 1. A SME reviewed the dataset, labeling as either synthetic or real and identifying the disease.

Results : Synthetic 256x256 OCT b-scans generated by the LDM proved to be of higher quality (FID score: 38.98; 15% correctly identified as synthetic) in comparison to those generated by StyleGan3 (FID score: 50.85; 98% correctly identified as synthetic). For the 512x512 OCT b-scans generated by the LDM, the SME correctly labeled 65% as synthetic in comparison to 60% of real OCT images as synthetic. Disease detection accuracy was 95% for synthetic OCT images in comparison to 88% for real OCT images, Figure 2.

Conclusions : LDMs are a promising approach to generate synthetic OCT scans. The SMEs ability to correctly identify diseases is comparable across the synthetic and real OCT b-scans, and it is difficult for SME to distinguish real vs synthetic. We will extend the LDM to generate synthetic volumetric (3D) OCT scans, trained on a more extensive dataset of real OCT images, and conduct an evaluation of with a panel of SME reviewers. The OCT scans will also be associated with synthetic fundus photos and synthetic clinical data generated by Synthea™ to provide complete, longitudinal synthetic health records with ophthalmic imaging.

This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.

 

 

×
×

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.

×