Abstract
PURPOSE: Fundus fluorescein angiography (FFA) is essential for diagnosing and managing retinal vascular diseases, yet its evaluation is time-consuming and subject to inter-observer variability. We aim to develop a deep-learning-based framework for multi-lesion segmentation in FFA images and to evaluate its primary performance on standard 55 ° images. METHODS: A dataset of 428 standard 55 ° FFA images and 53 ultra-wide-field (UWF) FFA images was annotated for non-perfusion areas (NPA), microaneurysms (MA), neovascularization (NV) and laser spots. A residual U-Net framework was trained using a data-pooling strategy. The primary analysis was performed on the 55 ° test set, whereas the UWF subset was analyzed only as a preliminary exploratory observation because of its limited sample size. Performance was assessed using Dice score, Intersection over Union (IoU) and recall. RESULTS: On 55 ° FFA images, the model achieved Dice scores of 0.65 ± 0.24 for NPA, 0.70 ± 0.13 for MA, 0.73 ± 0.23 for NV, and 0.70 ± 0.17 for laser spots. In an exploratory analysis of UWF images, overlap-based performance was lower for NPA (0.48 ± 0.21, p = 0.02) and MA (0.58 ± 0.19, p = 0.01), whereas laser spot segmentation was similar. NV segmentation in CNV achieved a Dice score of 0.90 ± 0.09. NPA segmentation was better in RVO than in DR, whereas MA segmentation was better in DR than in RVO. Moreover, NV segmentation was significantly stronger in the venous phase (0.77 ± 0.17) and late phase (0.75 ± 0.24) than in the arteriovenous phase (0.50 ± 0.32, both p < 0.05). Exploratory UWF analysis showed lower overlap-based performance, highlighting the likely impact of domain shift and the need for larger UWF-specific datasets. CONCLUSION: This study has established a highly consistent and reproducible framework for multi-lesion segmentation in FFA images. The model may help standardize lesion quantification, reduce manual grading burden, and support future development of larger multi-center FFA datasets.