Some ocular imaging treatments, such as fluorescein angiography (FA), are intrusive and have the potential for negative side effects, while others, such as funduscopy, are non-invasive and safe for the patient. However, successful diagnosis of ocular disorders necessitated various data modalities and the possibility of invasive treatments. For a study, researchers presented a unique conditional generative adversarial network (GAN) capable of concurrently generating FA pictures from fundus photos and predicting retinal degeneration. The proposed approach solved the challenge of non-invasively mapping retinal vasculature while using cross-modality pictures to predict the presence of retinal abnormalities. One of the proposed work’s significant innovations was using a semi-supervised technique in training the network to address the problem of data reliance that plagues typical deep learning architectures. 

Experiment results showed that the suggested design outperformed state-of-the-art generative networks for picture creation across imaging modalities. They showed, in particular, that there was a statistically significant (P<.0001) difference in structural correctness of the translated pictures between our technique and the state-of-the-art. Furthermore, the findings showed that the suggested vision transformers generalized well on out-of-distribution data sets for retinal disease prediction, which was a challenge that many standard deep networks experience.

Reference:jov.arvojournals.org/article.aspx?articleid=2778479