Angiography-free diagnosis of retinal diseases via interpretable multi-modal learning.

Abstract

Accurate diagnosis of retinal diseases often depends on fluorescein fundus angiography (FFA), an invasive imaging modality requiring intravenous dye injection. Although color fundus photography (CFP) offers a safer and more accessible alternative, it lacks the vascular detail necessary for high-precision diagnosis, creating a critical gap in clinical workflows when angiography is unavailable or contraindicated. We introduce an angiography-free, interpretable multi-modal learning framework that enables accurate retinal disease diagnosis using only CFP images at test time. Rather than synthesizing FFA images, our model learns to disentangle shared and modality-specific features from paired CFP and FFA inputs during training. These features are stored in disease-specific libraries that serve as a reference during inference, allowing the model to transparently infer angiographic information from CFP data alone. Extensive internal and external validation on 119,173 retinal images across seven diseases demonstrates high diagnostic accuracy (AUCs of 0.98 and 0.96, respectively), along with clinically aligned visual explanations that support model decisions. By combining non-invasive deployment with transparent reasoning, our framework offers a robust and generalizable alternative to angiography, suitable for real-world implementation in diverse clinical settings.