Abstract
Purpose :
Availability of large-scale annotated datasets limits progress in medical imaging Computer Vision. The representation of training data with the same variation as seen in the patient population is an important consideration as it can impact deep learning model generalisability for unseen samples. Synthetic data offers a possible solution to these shortcomings. Thickness and volume of retinal layers are affected by ophthalmological and systemic diseases. We describe an approach for generating synthetic Optical coherence tomography angiography (OCT-A) data with potential for expanding training datasets, and for use in simulating disease states.
Methods :
Synthetic images were simulated by layering planar meshes corresponding to 10 retinal layers, Bruch’s membrane, choriocapillaris, and choroidal stroma (Figure 1). A mixed Gaussian model was used to simulate the macula and optic nerve structures. These structures were subsequently projected onto a hemisphere of radius 3.5cm, and embedded in 3D volumes using ray-casting (Open3D), with acquisition geometry and pixel intensities matched to 12x12mm OCT-A images from the PLEX Elite 9000 Zeiss machine. A neural network Physics-informed generative adversarial network (PI-GAN) (Figure 1) was used for image-to-image translation between synthetic segmentation space and Clinical OCT-A space. 5820 images were used in training and 3840 were used in testing.
Results :
Synthetic retinal layer images compared favorably with clinical OCT-A data, as shown in Figure 1. After 200 training epochs, the PI-GAN produced synthetic clinical images with retinal layer and choroidal structures resembling target images, including features such as OCT-A artefacts, foveal pit, and optic nerve, at native resolution. Frechet Inception distance (FID) of the generated images was 9.21, indicating a small distance between feature vectors for real and synthetic images.
Conclusions :
Here, we demonstrate a physics-informed approach to generating synthetic OCT-A retinal layer images. The framework shows promise in simulating diseases that instigate inflammation or thinning of the retina and choroid. Additionally, these images have a known ground truth which avoids the requirement for manual labeling. To our knowledge, this is the first instance of synthetic data, an increasingly valued research area, for 3D retinal layer generation and segmentation.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.