Automated Interpretation of Fundus Fluorescein Angiography with Multi-Retinal Lesion Segmentation

PurposeFundus fluorescein angiography (FFA) is essential for diagnosing and managing retinal vascular diseases, while its evaluation is time-consuming and subject to inter-observer variability. We aim to develop a deep-learning-based model for accurate multi-lesion segmentation for these diseases. MethodsA dataset comprising 428 standard 55{degrees} and 53 ultra-wide-field (UWF) FFA images was labeled for various lesions, including non-perfusion areas (NPA), microaneurysms (MA), neovascularization (NV) and laser spots. A U-net-based network was trained and validated (80%) to segment FFA lesions and then tested (20%), with performance assessed via Dice score and Intersection over Union (IoU). ResultsOur model achieved Dice scores for NPA, MA, NV, and Laser on 55{degrees} FFA images at 0.65{+/-}0.24, 0.70{+/-}0.13, 0.73{+/-}0.23 and 0.70{+/-}0.17, respectively. UWF results were slightly lower for NPA (0.48{+/-}0.21, p=0.02), MA (0.58{+/-}0.19, p=0.01), NV (0.50{+/-}0.34, p=0.14), but similar for Laser (0.74{+/-}0.03, p=0.90). Notably, NV segmentation in choroidal neovascularization achieved a high Dice score of 0.90{+/-}0.09, surpassing those in DR (0.68{+/-}0.22) and RVO (0.62{+/-}0.28), p<0.01. In RVO, NPA segmentation outperformed that in DR, scoring 0.77{+/-}0.25 versus 0.59{+/-}0.22, p<0.01, while in DR, MA segmentation was superior to that in RVO, with scores of 0.70{+/-}0.18 compared to 0.53{+/-}0.20, p<0.01. Moreover, NV segmentation was significantly stronger in venous phase (0.77{+/-}0.17) and late phase (0.75{+/-}0.24) compared to arteriovenous phase (0.50{+/-}0.32), p<0.05. ConclusionThis study has established a model for precise multi-lesion segmentation in retinal vascular diseases using 55{degrees} and UWF FFA images. This multi-lesion segmentation model has the potential to expand databases, ease grader burden and standardize FFA image interpretation, thereby improving disease management. Furthermore, it enhances interpretable AI, fostering the development of sophisticated systems and promoting cross-modal image generation for medical applications. SynopsisWe developed deep-learning models for segmenting multiple retinal lesions in both normal and ultra-field FFA images; the satisfactory performances set the foundation for quantifiable clinical biomarker assessment and building interpretable generative artificial intelligence.