Synthetic Ground-truth Image-based Drusen Segmentation in OCT Images: A Novel Approach Based on Sinusoidal Image Deformations and Generative Adversarial Networks

We introduce a synthetic ground-truth approach for drusen segmentation in OCT, combining sinusoidal image deformations of healthy retinal scans with a generative adversarial network to produce paired image-mask training data without manual annotation. The method enables robust drusen detection while sidestepping the cost and variability of expert pixel-level labeling.

In Investigative Ophthalmology & Visual Science, vol. 66, no. 8, pp. 4188–4188, 2025. The Association for Research in Vision and Ophthalmology.