Few-shot Semantic Segmentation of Colorectal Polyp Images Based on Vision Mamba

Few-shot semantic segmentation of intestinal polyp images aims to guide the segmentation of query images using a small number of pixel-wise annotated support images, enabling the identification and segmentation of unknown targets in endoscopic images. However, existing few-shot methods for intestinal polyp segmentation face challenges due to the diversity in polyp boundaries, sizes, irregular shapes, and varying lighting conditions in endoscopic images, which lead to inaccurate segmentation. To address these issues, this paper proposes an intestinal polyp image segmentation method based on Vision Mamba. First, a multi-scale feature aggregation module is employed to capture more features and details. Then, Vision Mamba is used to model long-range interactions while considering both local and global image features, achieving multi-scale feature fusion and enhancing the model's ability to capture features of different scales. The modeling of long-range interactions enables Vision Mamba to generalize more effectively to various image types and segmentation tasks, improving its applicability and flexibility. Experimental results demonstrate that the model performs well across four datasets—including Kvasir—which are standard benchmarks in this field and cover a wide range of polyp variations, including different sizes, shapes, and lighting conditions. These datasets provide comprehensive evaluation scenarios for validating the performance and robustness of our method. The multi-scale feature aggregation and long-range interaction modeling strategies allow the model to better adapt to complex and diverse intestinal polyp images, resulting in improved segmentation accuracy. In all challenging tasks, our method significantly outperforms traditional models like UNet and UNet++, validating its strong applicability and extensive generalization capabilities.