Related Experiment Video
Updated: Jul 24, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
444
Hepatic vessel segmentation based on 3D swin-transformer with inductive biased multi-head self-attention
Mian Wu1, Yinling Qian1, Xiangyun Liao2
1Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, Shenzhen Institute of Advanced Technology, Chinese Academy of Science, Shenzhen, China.
BMC Medical Imaging
|July 8, 2023
Summary
This study introduces IBIMHAV-Net, a novel 3D deep learning model for accurate liver vessel segmentation in CT images. The network effectively combines convolutional and self-attention mechanisms for improved surgical planning.
Area of Science:
- Medical Image Analysis
- Deep Learning
- Computer Vision
Background:
- Accurate liver vessel segmentation from CT images is crucial for surgical planning.
- Existing methods using FCN, U-net, and V-net variants struggle with complex structures and low-contrast backgrounds.
- Limited locality reception fields of convolutional operators in current models lead to misclassified voxels.
Purpose of the Study:
- To propose a robust end-to-end liver vessel segmentation network, IBIMHAV-Net.
- To enhance the accuracy of 3D liver vessel segmentation in CT images.
- To improve upon existing deep learning and graph cut methods for this task.
Main Methods:
- Expanding Swin Transformer to 3D and combining convolution with self-attention.
- Introducing voxel-wise embedding for precise liver vessel localization.
- Utilizing multi-scale convolutional operators and inductive biased multi-head self-attention for enhanced feature capture.
Main Results:
- Achieved average Dice score of 74.8% and sensitivity of 77.5% on the 3DIRCADb dataset.
- Outperformed existing deep learning methods and improved graph cuts.
- Branches Detected (BD)/Tree-length Detected (TD) indexes demonstrated superior global/local feature capture ability.
Conclusions:
- IBIMHAV-Net offers automatic and accurate 3D liver vessel segmentation.
- The interleaved architecture effectively leverages both global and local spatial features in CT volumes.
- The model has potential for extension to other clinical imaging data.
