LGCF-Net: Local–Global Context Modeling and Decoder-Guided Residual Correction for Retinal Vessel Segmentation

Accurate retinal vessel segmentation is important for quantitative assessment of vascular abnormalities in retinopathy of prematurity (ROP), but remains challenging because vessels vary greatly in caliber and spatial extent, while thin branches and continuous trajectories are easily disrupted by complex retinal backgrounds. We propose LGCF-Net, a U-shaped retinal vessel segmentation network with a ResNet34 encoder. A Local–Global Vessel Context (LGVC) module combines local spatial representation with global frequency-domain context through adaptive fusion, while a Decoder-Guided Residual Correction Fusion (DRCF) module uses decoder semantics to extract and selectively compensate complementary encoder details. Cross-entropy, Dice, and clDice losses are jointly employed for pixel-level, region-level, and topology-aware supervision. On the HVDROPDB-BV Neo subset, LGCF-Net achieved the highest F1 scores, improving F1 by 1.59 percentage points over the second-best method. On DRIVE, CHASE_DB1, and HRF, it also obtained the highest F1 scores among all compared methods. These results indicate that LGCF-Net provides an effective approach for retinal vessel segmentation, particularly for premature-infant retinal images.