Cross-fusion skip connection mechanism-based abdominal organ segmentation method
10.19745/j.1003-8868.2025193
- VernacularTitle:基于交叉融合跳跃连接机制的腹部器官分割方法研究
- Author:
Yu-xiu SUN
1
;
Peng-peng ZHANG
1
Author Information
1. 上海电机学院电子信息学院,上海 201306
- Publication Type:Journal Article
- Keywords:
Swin-Unet;
attention mechanism;
abdominal organ segmentation;
deep learning;
skip connection
- From:
Chinese Medical Equipment Journal
2025;46(11):1-9
- CountryChina
- Language:Chinese
-
Abstract:
Objective To propose an abdominal organ segmentation method based on a cross-fusion skip-connection mechanism to enhance the overall segmentation effect of abdominal organs.Methods Firstly,the Swin-Unet network was used as the baseline segmentation model,into which an enhanced dual cross attention(EDCA)module was integrated to refine features by fusing multi-scale encoder features;secondly,a decoder-guided cross-fusion attention(DCFA)module was added to reconstruct encoder information under the guidance of decoder features so as to mitigate the semantic gap between the features of the encoder and decoder;finally,a cross-fusion skip-connection mechanism was designed based on EDCA and DCFA modules to achieve effective feature transmission from encoder to decoder features.The cross-fusion skip connection mechanism-based abdominal organ segmentation model had its effectiveness validated by the experiments carried out on Synapse multi-organ CT dataset and comparison with some mainstream abdominal organ segmentation models including TransUNet,UNetR,UCTransNet,MT-UNet and Swin-Unet.Results The Swin-Unet model based on the cross-fusion skip connection mechanism achieved a Dice similarity coefficient(DSC)of 80.30%and a Hausdorff distance(HD)of 18.26 mm,outperforming the other mainstream abdominal organ segmentation models.Moreover,the model proposed had the number of parameters(27.18M)and floating-point operations per second(6.16×10?/s)both superior to those of the other mainstream abdominal organ segmentation models while close to those of Swin-Unet.Conclusion The proposed method effectively enhances the segmentation of abdominal organs,providing an effective design approach for optimizing segmentation models based on U-shaped structures.
