Improved MambaUNet for lightweight cascaded segmentation in liver tumor CT image
10.3969/j.issn.1005-202X.2025.08.014
- VernacularTitle:改进MambaUNet网络对肝脏肿瘤CT图像的轻量化级联分割
- Author:
Ke LI
1
;
Wenzhong LIU
;
Jingtao QIN
Author Information
1. 四川轻化工大学计算机科学与工程学院,四川 宜宾 644002
- Publication Type:Journal Article
- Keywords:
liver tumor segmentation;
MD-MambaUNet;
CT image;
Dice similarity coefficient
- From:
Chinese Journal of Medical Physics
2025;42(8):1068-1078
- CountryChina
- Language:Chinese
-
Abstract:
To address the limitations of convolutional neural networks in global context modeling and the complexity of secondary computation in Transformer's self attention mechanism,an improved multi-direction-MambaUNet(MD-MambaUNet)based on MambaUNet is proposed.This network integrates with multi directional selective scanning module to extract spatial features from multiple directions of CT images,significantly improving the global context modeling capability.By introducing partial convolution and constructing a lightweight hybrid convolution module,the model's parameter count is significantly reduced,enabling the processing of large-scale medical image data at lower computational costs while maintaining high-level performance.Experiments are conducted on the LiTS2017 and 3DIRCADB public datasets.Compared with MambaUNet on the LiTS2017 dataset,MD-MambaUNet improves the Dice similarity coefficient for liver and tumor segmentation by 3.32%and 4.77%,reaching 95.36%and 76.93%,and increases intersection over union by 4.18%and 4.92%,reaching 91.43%and 69.74%,respectively.Compared with MambaUNet on the 3DIRCADB dataset,MD-MambaUNet improves the Dice similarity coefficient for liver and tumor segmentation by 2.08%and 2.21%,reaching 93.81%and 64.68%,and increases intersection over union by 0.79%and 3.13%,reaching 87.23%and 57.65%,respectively.Meanwhile,the parameter count is 13.71 M less than MambaUNet,making it possible for the model to be deployed in clinical scenarios.
