Dual-Domain Structural Guidance with Grayscale-Enhanced Edge Learning for Robust Medical Image Segmentation

Colorectal polyp segmentation from colonoscopy images remains challenging due to variable shapes, blurred boundaries, and heterogeneous appearances. Existing methods suffer from insufficient multi-scale semantic modeling and limited use of cross-domain structural cues. Here, we present a dual-domain structural guidance framework that integrates grayscale-guided edge features and graph-based multi-scale fusion to improve segmentation robustness. The method employs a learnable edge extractor initialized with Sobel operators to capture complementary high-frequency information from RGB and grayscale domains, and uses graph convolution to model semantic dependencies across multi-scale branches. We also introduce an edge-guided progressive decoder to refine boundary details. Experiments on five public benchmarks show that the method achieves 93.62% Dice on Kvasir-SEG and 95.06% Dice on CVC-ClinicDB, achieving competitive or top-tier performance relative to state-of-the-art approaches, with particular advantages in boundary delineation and robustness on the cross-dataset generalization benchmarks. This work provides a reliable visual computing solution for clinical endoscopic image analysis. The code is available at https://github.com/ZCX327/ECNNet.