An interpretable graph reasoning framework for enhanced diagnosis in fundus fluorescein angiography

Fundus fluorescein angiography (FFA) is the gold standard for diagnosing retinal vascular diseases. However, interpretable artificial intelligence (AI) systems for FFA image analysis that align with clinicians’ reasoning processes and real-world diagnostic workflows remain scarce. In this study, we introduce an interpretable graph reasoning framework, FFA-RGNet, that explicitly models the temporal patterns of fine-grained lesions across imaging phases and anatomical locations for disease diagnosis from sequential FFA images. FFA-RGNet was trained and validated on a large, fine-grained, multi-center dataset consisting of 25,312 FFA images from 2651 patients. FFA-RGNet outperformed nine baseline models on the internal test set, achieving an F1-score of 95.38% (95% CI: 92.24%–97.88%), and demonstrated consistent performance across multi-center and prospective validation datasets. Its diagnostic performance not only surpassed that of ophthalmologists, residents and ophthalmology students, but also improved their diagnostic F1-score by 7.48% to 14.11% when used as an assistive tool. Comprehensive inference analysis further revealed its superior ability to correctly recognize fine-grained semantic entities, and to learn stable, disease-specific graph reasoning patterns aligned with clinical knowledge. These results suggest that FFA-RGNet enables accurate diagnosis with interpretable reasoning progresses, highlighting its strong potential as a reliable AI system for FFA interpretation and clinical decision support.