Using latent diffusion models to generate synthetic OCTA images

Abstract

Purpose : Optical coherence tomography angiography (OCTA) images in combination with deep learning methods hold great potential in the study of neurodegenerative diseases where the retina is a window to neurovascular dysfunction. However, such images are considered protected health information and as such cannot be easily and widely shared for training deep learning models. Previous solutions relied upon the combination of generative adversarial networks and rule-based simulations for synthetic OCTA image creation. We propose the use of only latent diffusion models (LDM) to generate synthetic OCTA images, without the need for additional rule-based simulations.

Methods : Participants over 50 years old with either normal cognition, mild cognitive impairment (MCI), or Alzheimer’s disease (AD) were imaged to obtain 3x3 mm² fovea centered OCTA scans of the superficial capillary plexus in one or both eyes. An autoencoder was then used to project the OCTA images into a lower-dimensional latent space that represents the original images as input for training the LDM. The LDM was then trained conditionally to account for all 3 disease states (normal cognition, MCI, and AD). To evaluate the synthetic images, we segmented the original and the synthetic images by using the optimal oriented flux filter followed by thresholding to compute retinal metrics. We conducted a comparative analysis of the distributions using a kolmogorov-smirnov test and chose vessel tortuosity (VT) across the whole field of view as an example metric to assess and compare the original and synthetic images.

Results : 371 good-quality OCTA images from cognitively normal, 227 OCTA images from MCI, and 100 OCTA images from AD participants were identified. No significant difference between the original and synthetic VT distributions were observed for the 3 disease states (normal cognition, MCI, AD). Figure 1 displays boxplots comparing the VT distributions in the original and synthetic images, while Figure 2 provides examples of both original and synthetic OCTA images.

Conclusions : Using LDMs to generate synthetic OCTA images has promising potential in sharing health-protected images to increase the availability of training data for deep learning methods.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Figure 1: Synthetic and original distributions across each disease state.

 

Figure 2: Examples of original and synthetically created OCTA images.