A Multiscale Edge-Guided Polynomial Approximation Network for Medical Image Segmentation

Fuxian Sui, Hua Wang, Fan Zhang

Abstract: As the core cornerstone of building an efficient medical care system, especially promoting accurate disease diagnosis and treatment, medical image segmentation is of great importance. However, medical segmentation faces many challenges, including complex background, shape and size changes, resulting in inaccurate or fuzzy segmentation boundaries. To meet these challenges, this paper proposes a multiscale edge-guided polynomial approximation network (AMEPANet). The well-designed edge guided bridge module in this paper uses the Laplacian operator to accurately capture and strengthen the edge information in the image, and realizes the robust preservation of edge information across multiple scales. At the same time, by building an information mixed attention mechanism, the network can further mine and use the subtle features of the boundary area to further improve the segmentation accuracy. In order to maximize the use of rich feature information at different scales and stages, this paper combines Kolmogorov-Arnold theorem to build an efficient decoder architecture, which can seamlessly integrate multi-source features to achieve comprehensive fusion and optimization of feature information. In addition, this paper also proposes an innovative \(C^1\) continuous activation function, which shows significant advantages in reducing the fluctuation of model calculation and promoting the stable convergence of the model, and further enhances the comprehensive processing ability of the model for complex medical image features. Through extensive and in-depth experiments on multiple authoritative data sets such as Synapse, the excellent performance of AMEPANet has been verified.