Multi-scale fusion networks for retinal vessel segmentation

Accurate segmentation of retinal vasculature constitutes an essential prerequisite for clinical diagnosis of ophthalmic pathologies. The inherent complexity of microvascular structures in fundus imaging presents significant acquisition difficulties, while artifacts inherent to conventional U-Net architectures further intensify segmentation challenges. This investigation presents a novel neural framework designated as the Multi-Scale Fusion Network (MSF-Net), specifically engineered to overcome these technical barriers in vascular segmentation. Architecturally, the encoder embeds a hierarchical attention mechanism that establishes cross-channel correlations through multi-resolution feature extraction, substantially augmenting representational capacity. A cross-hierarchical integration mechanism enables synergistic combination of heterogeneous semantic information via optimized skip-connections, producing enriched feature embeddings. The decoder incorporates a bidirectional feature amalgamation component employing spatial-channel recalibration to refine segmentation boundaries. Experimental results on three generalized datasets demonstrate that the proposed method surpasses U-Net and recent state-of-the-art approaches over the past three years across multiple evaluation metrics, which confirms its clinical applicability in retinal image analysis.