基于改进SwinUNet腹部多器官分割算法研究

Research on an Improved SwinUNet Algorithm for Multi-Organ Segmentation in the Abdomen

摘要: 医学图像分割中常面临全局上下文建模不足与多尺度特征表达有限的问题。本文提出一种改进SwinUNet方法。首先，采用Focal Transformer替换原始Swin Transformer，以分层注意力机制增强局部细节与全局依赖建模；其次，在编码器末端引入空洞空间金字塔池化(ASPP)结构，扩展感受野并提升多尺度特征表达；最后，在跳跃连接中加入Tokenized Interaction Fusion (TIF)模块，实现跨层语义与空间信息的高效融合。在Synapse腹部器官数据集上的实验表明，该方法在平均Dice和Hausdorff距离等指标上均优于基线模型，验证了其在腹部多器官分割中的有效性。

Abstract: In medical image segmentation, challenges often arise from insufficient global context modeling and limited multi-scale feature representation. This paper proposes an enhanced SwinUNet approach to address these issues. First, the original Swin Transformer is replaced with a Focal Transformer to strengthen both local detail capture and global dependency modeling through a hierarchical attention mechanism. Second, an Atrous Spatial Pyramid Pooling (ASPP) module is incorporated at the end of the encoder to expand the receptive field and improve multi-scale feature extraction. Finally, a Tokenized Interaction Fusion (TIF) module is integrated into the skip connections to facilitate efficient cross-layer fusion of semantic and spatial information. Experiments conducted on the Synapse multi-organ abdominal dataset demonstrate that the proposed method outperforms baseline models in key metrics such as average Dice coefficient and Hausdorff distance, confirming its effectiveness for abdominal organ segmentation.