Enhancing Polyp Segmentation Based on Multi-Frequency Guided Attention Network

Accurate segmentation of polyps is vital for effective pathological assessment. Nonetheless, colonoscopic images frequently suffer from indistinct edges, insufficient contrast, and structural blurring, which complicate the differentiation between polyps and adjacent tissue. These challenges considerably hinder the effectiveness of conventional segmentation techniques. In response to this problem, we design a Multi-Frequency Guided Attention Network (MFGANet) specifically designed for polyp segmentation. Compared to recent advanced models, MFGANet demonstrates superior segmentation performance. The network adopts an end-to-end architecture, consisting of four core modules: an encoder, a decoder, a Multi-Frequency Guided Attention (MFGA) module, and an Iterative Content-Guided Attention (ICGA) module. The MFGA module significantly enhances sensitivity to weak boundary targets and improves the ability to capture boundary information, global structures, and local details by integrating multi-frequency feature extraction with attention mechanisms. Additionally, the ICGA module refines segmentation boundaries step-by-step, effectively improving segmentation accuracy in blurred regions. We conduct extensive experiments on five benchmark datasets. The results show that MFGANet achieves the highest average Dice segmentation accuracy and outperforms existing state-of-the-art (SOTA) methods in multiple benchmark tests, demonstrating excellent segmentation performance.