nnLoGoNet: a hybrid local-global network for retinal vessel segmentation with Skeleton Recall Loss

Retinal vessel segmentation is a fundamental task in ophthalmic image analysis, playing a critical role in disease screening and clinical diagnosis. However, accurately delineating fine-grained micro-vessels while preserving global topological integrity remains challenging due to complex vascular structures, low contrast, and fragmented appearances. To tackle these issues, we propose nnLoGoNet, a novel hybrid local–global network architecture built upon the nnUNet framework. nnLoGoNet synergistically combines the complementary strengths of Convolutional Neural Networks (CNNs) and Transformers through a parallel design: CNN-based components focus on extracting local detailed features, while Transformer-based modules capture long-range dependencies and global contextual information. In addition, to explicitly enforce the tubular connectivity of retinal vessels, we introduce a Skeleton Recall Loss, which leverages efficient CPU-based skeletonization to provide topology-aware supervision with negligible computational overhead. Extensive experiments demonstrate that nnLoGoNet achieves superior performance over existing state-of-the-art methods on multiple Optical Coherence Tomography Angiography (OCTA) datasets and state-of-the-art results on standard Color Fundus Photography (CFP) benchmarks, while significantly improving vessel connectivity and reducing model complexity. Our implementation is available at: https://github.com/Luoyuchen0704/nnLoGoNet..