Test-Time Fluorescein Angiography-Free Non-Perfusion Area Segmentation from Fundus Image

Retinal non-perfusion areas (NPAs) are of critical importance for the diagnosis of certain retinal pathologies, particularly diabetic retinopathy. Fundus fluorescein angiography (FFA) is widely recognized as the clinical gold standard for identifying NPAs. However, FFA is an invasive imaging method. In this paper, we propose a color fundus photography (CFP) NPA segmentation framework that is FFA-free during testing, based on prompts for different missing-modality scenarios. Specifically, we design distinct learnable modality-conditioned prompts (LMCPs) for different missing-modality scenarios, catering to both single-modality input scenarios and dual-modality scenarios, enabling the network to perceive and learn the representations between different modalities. We also utilize cross-attention mechanisms to facilitate the acquisition of information from the FFA by the CFP. Extensive experiments are conducted on our private dataset, with the superiority of our proposed framework being successfully established.