A spatiotemporal model for future fundus image prediction with irregularly sampled sequential data

Fundus images play a crucial role in diagnosing and monitoring eye diseases, and the progression of these diseases is a gradual process. Predicting fundus images using longitudinal data can help ophthalmologists provide personalized treatment for better therapeutic effects. In this paper, a spatiotemporal prediction method for future fundus image prediction using longitudinal historical images is proposed. In clinical practice, the collected longitudinal images are usually irregularly sampled. To address this problem, a motion estimation block based on the ordinary differential equation is proposed by modeling the evolution of longitudinal fundus images over time dynamically. Additionally, the collected longitudinal images have different styles. To ensure the prediction network focuses on the intrinsic changes in the sequential input, the style transfer strategy is integrated into the prediction framework to mitigate the interference caused by style differences. Experimental results demonstrate that the image quality of the fundus images predicted by our method is higher compared with those generated by the state-of-the-art methods.