Translation of Fundus Image to Fundus Fluorescein Angiography Boosted by Structure Self-Supervised Representation Cycle Learning.

Introduction: Fundus fluorescein angiography captures detailed images of fundus vasculature, enabling precise disease assessment. Translating fundus images to fundus fluorescein angiography images can assist patients unable to use contrast agents due to physical constraints, facilitating disease analysis. Previous studies on this translation task were limited by the use of only 17 image pairs for training, potentially restricting model performance.

Methods: Image pairs were collected from patients through a collaborating hospital to create a larger dataset. A fundus image to fundus fluorescein angiography translation model was developed using structure self-supervised representation cycle learning. This model focuses on vascular structures for self-supervised learning, incorporates an auxiliary branch, and utilizes cycle learning to enhance the main training pipeline.

Results: Comparative evaluations on the test set demonstrate superior performance of the proposed model, with significantly improved Fréchet inception distance and kernel inception distance scores. Additionally, generalization experiments conducted on public datasets further confirm the model's advantages in various evaluation metrics.

Discussion: The enhanced performance of the proposed model can be attributed to the larger dataset and the novel structure self-supervised cycle learning approach, which effectively captures vascular details critical for accurate translation. The model's robust generalization across public datasets suggests its potential applicability in diverse clinical settings. However, challenges such as computational complexity and the need for further validation in real-world scenarios warrant additional investigation to ensure scalability and clinical reliability.

Conclusion: The proposed model effectively translates fundus images to fundus fluorescein angiography images, overcoming limitations of small datasets in previous studies. This approach demonstrates strong generalization capabilities, highlighting its potential to aid in large-scale disease analysis and patient care.