Fre-LKMUnet:Frequency Large Kernels in MambaUnet for efficient skin lesion segmentation

Precise segmentation of skin lesions in dermoscopic images is crucial for skin lesion segmentation. However, current deep learning-based automated skin lesion segmentation methods still face challenges in balancing computational efficiency and global contextual perception. This paper proposes a novel network model named Fre-LKMUnet to solve the above problem. The model adopts a U-shaped encoder–decoder architecture and integrates a Frequency Large Kernel Mamba (Fre-LKM) module. This module can jointly improve the efficiency of feature representation and the ability of global modeling. The Fre-LKM module uses a three-level strategy. First, input features are split into two channel groups via hard splitting. The high-frequency branch uses depthwise separable convolution to enhance texture and edges, while the low-frequency branch adopts Laplacian pyramid decomposition to mine implicit details and global structure. Second, large-kernel convolution is added to improve semi-global feature extraction. Combined with Bidirectional Mamba modeling, it can effectively capture long-distance spatial dependencies. Finally, a gated convolutional unit adaptively fuses bidirectional sensing features to generate more distinguishable feature representations. Experimental results on public datasets including ISIC2016, ISIC2017, ISIC2018 and PH2 show that the proposed model achieves competitive segmentation performance for skin lesion segmentation. Meanwhile, the model maintains a lightweight architecture and achieves a favorable balance between segmentation accuracy and computational efficiency. The code can be found at https://github.com/xiaoyaozhong/Fre-LKMUNet