文献类型：期刊文献

中文题名：语义扩散对齐的多尺度感知医学图像分割方法

英文题名：Semantic Diffusion Alignment-Based Multi-Scale Perception for Medical Image Segmentation

作者：赖晓婷[1];张静[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237

年份：2025

卷号：37

期号：8

起止页码：1394

中文期刊名：计算机辅助设计与图形学学报

外文期刊名：Journal of Computer-Aided Design & Computer Graphics

收录：;北大核心:【北大核心2023】；

基金：上海市自然科学基金(22ZR1418400).

语种：中文

中文关键词：医学图像分割;多尺度感知;语义扩散;特征对齐;注意力机制

外文关键词：medical image segmentation;multi-scale perception;semantic diffusion;feature alignment;attention mechanism

摘要：有效的多尺度特征表示对于准确地分割医学图像中大小不一的病变区域具有重要意义.针对现有方法未能充分挖掘不同目标的多尺度信息导致难以实现精确分割的问题,提出一种语义扩散对齐的多尺度感知医学图像分割方法.首先从局部和全局2个角度探索多尺度上下文信息的感知能力,构建多尺度编码器和解码器.其中,多尺度编码器利用局部多尺度自注意力机制和全局重调整来提取多尺度特征,以捕获图像中不同目标的信息,多尺度解码器则通过上采样恢复空间分辨率并保留细节信息,得到更准确的分割结果;为了进一步增强特征语义表示,提出语义扩散对齐模块,实现了低级特征与高级特征的语义对齐,获得更具鉴别性的融合特征.在Synapse和ACDC数据集上的实验结果表明,所提方法的平均Dice相似系数分别达到82.42%和92.25%,性能优于大多数主流分割方法.该方法的相关代码链接为https://github.com/MiniCoCo-be/MSPSN.
Effective multi-scale feature representation is crucial for accurately segmenting lesions of varying sizes in medical images.Addressing the challenge of existing methods failing to fully exploit multi-scale information for different targets,a novel multi-scale perception medical image segmentation method was proposed based on semantic diffusion alignment.It first explored the perceptual ability of multi-scale con-textual information from both local and global perspectives to construct multi-scale encoder and decoder.The multi-scale encoder utilized a local multi-scale self-attention mechanism and global fine-tuning to ex-tract features at multiple scales,capturing information from different targets in the image.The multi-scale decoder restored spatial resolution through upsampling while preserving detailed information to achieve more precise segmentation results.Furthermore,in order to further enhance the semantic representation of features,a semantic diffusion alignment module was proposed,which realizes the semantic alignment of low-level features and high-level features,thus obtaining more discriminative fusion features.Finally,ex-perimental results on the Synapse dataset and ACDC dataset demonstrate that the proposed method achieves average Dice similarity coefficients of 82.42%and 92.25%respectively,outperforming most existing meth-ods.Additionally,the relevant code can be found at https://github.com/MiniCoCo-be/MSPSN.