文献类型：会议论文

英文题名：SEMU-Net: A Boundary Semantic-Aware Network for Medical Image Segmentation

作者：Jiang, Zongying[1];Dai, Li[1];Jin, Yuxiong[1];Li, Jianhua[1]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai, Peoples R China

会议论文集：2025 International Joint Conference on Neural Networks-IJCNN

会议日期：JUN 30-JUL 05, 2025

会议地点：Rome, ITALY

语种：英文

外文关键词：Medical image segmentation; Diffusion; Attention mechanism; Semantic information

摘要：Due to the limitations of medical imaging devices, existing segmentation models struggle to find boundary pixels that strictly distinguish the background from regions of interest, where boundary pixels are critical for accurate segmentation. This problem is described as semantic ambiguity arising from blurred boundaries. To address this issue, we propose a novel UNet architecture with a Semantic Enhancement Module (SEMUNet). This network mainly utilizes the Latent Diffusion Model (LDM) to generate additional semantic features by exploiting LDM's ability to compress and comprehend semantic information effectively. The Semantic Enhancement Module (SEM) comprises two components: the Dual-Cross Attention Module and the Token Pruning Module. In the Dual-Cross Attention Module, semantic features are integrated into a U-Net-like framework, enabling the network to incorporate semantic information from both spatial and channel dimensions selectively. The Token Pruning technique compresses the large number of tokens generated during the tokenization process into a smaller subset of significant tokens. With the aid of these two modules, the fusion process is optimized and computational costs are minimized. Experimental results on two public datasets demonstrate that the proposed SEMU-Net has superior performance over other state-of-the-art algorithms.