SMNet: Spatial Information-Aware Multi-Level Fusion Network for Medical Image Segmentation

Abstract

This paper discusses the limitations of U-Net structure-based medical image segmentation methods and the employment of traditional attention mechanisms (channel or spatial attention). To this end, we propose a novel spatial information-aware multi-level fusion network tailored for medical image segmentation, which incorporates the multi-level deep fusion (MDF) module, channel-spatial semantic interaction (CSSI) module and split detail enhancement (SDE) module. These modules are designed to harmonize feature integration across scales, unearth positional cues from high-level features by exploring their intrinsic channel and spatial relationships, and enhance fine details via variable dilation rates in low-level features, respectively. Additionally, cross-level connections are established to counter the information loss typically associated with bottom-up feature transmission, ensuring a more comprehensive fusion of location and detail information. We conduct extensive experiments on the three medical image datasets, and comparison results show that the proposed SMNet has superior performance.