FFA-SegGPT: A Two-Stage Cross-Modal AI System to Optimize the Clinical Workflow of Diabetic Retinopathy FFA Interpretation

Purpose: The purpose of this study was to develop an interpretable, two-stage artificial intelligence (AI) system (fundus fluorescein angiography [FFA]-SegGPT) designed to optimize the clinical workflow of diabetic retinopathy (DR) FFA interpretation and reduce the risk of factual inaccuracies common in end-to-end models. Methods: We developed a cross-modal system that decouples lesion quantification from report generation. A segmentation module first explicitly segments non-perfusion areas (NPAs) and neovascularization (NV). These segmentation maps then serve as visual constraints to guide an evidence-based report generator, ensuring diagnostic text is strictly grounded in detected lesions. Results: In a simulated clinical workflow, AI-aided reporting significantly reduced mean interpretation time by 50% (from 8.13 ± 0.47 to 4.09 ± 0.21 minutes, P < 0.001) and improved median report quality (4.0 vs. 3.5, P < 0.001) compared to manual interpretation. Objectively, the model achieved high Dice Similarity Coefficients (NV = 85.35% and NPAs = 86.49%) and maintained high semantic fidelity for report generation (BERTScore = 0.91), with robust generalization demonstrated across three external cohorts. Conclusions: FFA-SegGPT enhances both the efficiency and quality of DR diagnostics. By decoupling lesion quantification from report generation, the system provides a transparent, dual-output mechanism that reduces the risk of factual inaccuracies, offering an interpretable workflow to facilitate safe clinical adoption. Translational Relevance: This interpretable AI system optimizes the clinical workflow of DR by providing verifiable lesion segmentation and automated reporting, potentially reducing clinician workload and standardizing diagnostic quality.