MS-LKSeg: enhancing multi-semantic synergistic learning with large kernel convolution for medical image segmentation

Abstract

Medical image segmentation is dedicated to accurately distinguish different regions, tissues or lesions in images. The introduce of multiple semantic information, such as shape, texture, and location can provide more clues and bases for segmentation algorithms. However, the effective fusion and utilization of multiple semantic information remains a major challenge. To address this problem, we proposed MS-LKSeg, which aims to explore the synergistic relationship between spatial attention and channel attention at different semantic levels, and to integrate semantic information originating from different levels with rich diversity more efficiently. Specifically, we introduce the Multi-semantic Information Synergy (MIS) block in the encoder of MS-LKSeg, which efficiently captures different semantic spatial structures by extracting features of spatial dimensions (height and width), decomposing them into sub-features, and then passing them through a shared depthwise convolutional layer. Subsequently, it explicitly models long-range dependencies among channels to achieve robust feature interactions, thereby alleviating the disparities among multi-semantic information. Additionally, in the skip connection of MS-LKSeg, we apply depthwise convolution with a large kernel to the encoder’s output to capture broader contextual information and reduce information loss during feature propagation. This approach also contributes to the model maintaining higher accuracy and robustness in segmentation tasks. The superiority of our method is demonstrated by the experimental evaluations performed on multiple publicly available datasets. Our implementation code is available at https://github.com/niuniude/MS-LKSeg.