Abstract
PURPOSE: This study aimed to develop and validate a deep learning model for the accurate multi-class classification of six retinal diseases using fundus fluorescein angiography (FFA) images. METHODS: We applied a knowledge-enhanced pre-training strategy (KeepFIT) using a ResNet-50 image encoder on two large FFA corpora: a curated atlas and a clinical report dataset. The resulting visual encoder was fine-tuned to classify six conditions, including diabetic retinopathy and macular degeneration. The model's performance and generalizability were assessed on two independent test sets, one of which was sourced from an external institution. RESULTS: Our proposed deep learning model, leveraging a knowledge-enhanced pre-training strategy, demonstrated robust performance in classifying six distinct retinal diseases using fundus fluorescein angiography images. The model achieved a strong and consistent micro-average area under the curve (AUC) of 0.92 across two independent test sets. Notably, it showed excellent classification performance for critical conditions such as venous occlusion (VO) and neovascular age-related macular degeneration (nAMD), with AUC values reaching 0.95 and 0.96, respectively. CONCLUSION: The knowledge-enhanced pre-training strategy significantly improves the diagnostic accuracy and generalizability of deep learning models for FFA analysis. This approach provides a scalable and effective framework for automated retinal disease screening, holding significant potential for clinical decision support, especially in resource-limited settings.