MGP-Net: When Morphology-Aware Activation Meets Adaptive Noise Purification

Retinal vessel segmentation is a key step in automated fundus image analysis, because downstream vascular assessment relies on vessel maps that are both structurally complete and locally reliable. In practice, thin and low-contrast vessels are easily missed during feature extraction, while pathological regions and background artifacts may trigger false responses. To address this issue, we propose the Morphology-Guided Purification Network (MGP-Net), which enhances weak vessel responses and reduces background interference before reconstruction. Specifically, the Anisotropic Morphology-Aware Block uses directional filters to strengthen elongated vascular patterns that are difficult to capture with standard isotropic convolutions. The Adaptive Noise-Purification Gate is introduced into skip connections to suppress non-vascular interference according to spatial saliency before feature fusion. In addition, a High-Frequency Preservation Flow reduces detail loss in deep layers and supports stable signal propagation. Experiments on DRIVE, STARE, and CHASE_DB1 show that MGP-Net achieves the best Sensitivity on all three datasets while maintaining competitive F1-Score, Accuracy, and Specificity.