MLK-Net: Leveraging multi-scale and large kernel convolutions for robust skin lesion segmentation

Abstract

Accurate medical image segmentation, especially for skin lesions, is challenging due to fuzzy boundaries, diverse shapes, and varying lesion sizes. Most existing SSM-based skin lesion segmentation methods use pure SSM or simply combine CNN with SSM as the network backbone, but often fail to fully consider background information and multi-scale features. To address these issues, we propose a novel network that integrates large convolution kernels for rich background feature extraction and combines multi-head mixed convolutions with small and large kernels to capture multi-scale features. This design improves the segmentation of complex structures and diverse lesion scales. Experiments on three benchmark skin lesion segmentation datasets demonstrate that our method consistently outperforms state-of-the-art approaches across multiple evaluation metrics, showcasing its robustness and effectiveness in tackling critical segmentation challenges. For reproduction, the implementation codes can be checked out at https://github.com/yuxl2023/MLK-Net.