Abstract
PURPOSE: Diabetic retinopathy (DR) is a serious ocular complication of diabetes, affecting about 30%-40% of patients. It primarily damages the retinal microvascular system and can result in visual impairment or even blindness. A hallmark lesion of DR is vascular leakage, which is typically observed in the late-phase images of ultra-widefield fluorescein angiography (UWFA). However, accurately segmenting leakage regions in UWFA remains a challenge because of their irregular and heterogeneous morphology, as well as the substantial computational demands associated with the high-resolution nature of UWFA images. METHODS: We propose a deep learning framework that combines multiscale sampling with two-dimensional wavelet transforms and an exponential moving average (EMA) mechanism to fuse global and local features. A cross-guided neighborhood refinement strategy is further introduced to enhance boundary accuracy. RESULTS: The experimental results demonstrate that (1) the model exhibits an optimal performance when the EMA parameter is set to 0.3; (2) the performance of the UNet-Wavelet network significantly surpasses traditional networks; and (3) using a multiscale fusion framework confers greater robustness compared with non-framework approaches. CONCLUSIONS: We validated our method on a UWFA dataset from Guangdong Provincial People's Hospital and Foshan Second People's Hospital, and the results demonstrated that our model achieves efficient and accurate segmentation of leakage regions in UWFA images. TRANSLATIONAL RELEVANCE: By addressing the irregular morphology and high-resolution complexity of UWFA images, our method enhances segmentation accuracy and computational efficiency, enabling more objective and timely clinical quantification of DR-related leakage and potentially supporting earlier intervention for affected patients.