CT liver tumor image segmentation based on ResUNet and Transformer
10.3969/j.issn.1005-202X.2025.11.008
- VernacularTitle:基于ResUNet和Transformer的CT肝脏肿瘤图像分割
- Author:
Haochen SHAO
1
;
Wei ZHANG
1
;
Yuzhang MA
1
Author Information
1. 甘肃中医药大学医学信息工程学院,甘肃 兰州 730000
- Publication Type:Journal Article
- Keywords:
liver tumor;
CT image;
BounDer-Net model;
ResUNet;
Transformer
- From:
Chinese Journal of Medical Physics
2025;42(11):1455-1461
- CountryChina
- Language:Chinese
-
Abstract:
To address the challenges of blurred boundaries and feature loss in the segmentation of small liver tumors in CT images,a novel model named BounDer-Net is proposed based on an improved ResUNet architecture.By constructing a Transformer-based dynamic multi-scale encoder and introducing a channel-spatial dual-path attention mechanism,the model can focus on tumor features across multiple dimensions.Additionally,the model adopts boundary-sensitive dynamic feature fusion strategy which effectively captures the heterogeneous features of tumors.BounDer-Net model firstly generates initial feature maps through low-level feature extraction,then inputs the features into a Transformer-based dynamic multi-scale encoder for extracting multi-level features,and finally restores spatial details via a decoder and improves the segmentation accuracy of small tumor boundaries using a boundary enhancement module.Experimental results on the LiTS2017 dataset show that BounDer-Net model achieves a Dice similarity coefficient of 94.64%,a mean intersection over union of 92.34%,and a Hausdorff distance of 0.35 mm,significantly outperforming existing methods.This study provides a reliable solution for the automatic diagnosis of small tumors in liver CT images.
