Injection of fluorescent dye is a safety concern in fluorescein angiography (FA). This has led to the cautious use of this clinical diagnostic modality in certain populations (e.g., children, allergic populations). In recent years, the development of non-invasive functional imaging of fundus blood flow by computational means has become a hot spot in ophthalmic research, such as OCT angiography. Deep learning-based color fundus to FA prediction is another emerging approach, which takes advantage of the nonlinear and high-dimensional mapping capabilities of deep neural networks to establish the relationship of these two imaging modalities explicitly. Most of such studies use a small publicly-available dataset and rely on algorithm design to improve the prediction accuracy. However, the limited performance has attracted little attention and raised doubts about its viability. Here, we show that the prediction accuracy can be significantly improved by simply expanding the training dataset by a factor of ~10 without introducing new algorithms. While this result is expected based on the nature of the data-driven model, it suggests that the development of such deep learning-based prediction requires a more diverse approach rather than focusing only on algorithmic improvements.
|