Related Experiment Video
Updated: Apr 5, 2026
New Thrombectomy Technique for Total Portal Vein Thrombosis in Liver Transplantation
Published on: June 27, 2025
GCN combined with snake convolution for enhanced topological perception in thrombotic hepatic portal vein
Lijuan Ma1, Weiguang Wang1, Xingshun Qi2
1School of Computer Science and Technology, Northeastern University, No.195 Innovation Road, Hunnan District, Shenyang, 110819, Liaoning Province, China.
Insights
A new 3D segmentation model, SnakeGCN, improves portal vein segmentation in liver cirrhosis patients. This advanced method enhances accuracy for better hemodynamic assessment and treatment planning.
Area of Science:
- Medical Imaging
- Computer Vision
- Hepatology
Background:
- Portal vein hemodynamics are critical for liver cirrhosis management.
- Accurate portal vein segmentation is vital for quantitative assessment but challenging due to lesions and artifacts.
- Existing methods struggle with vascular discontinuities, missing segments, and topological errors.
Purpose of the Study:
- To develop an innovative 3D segmentation model for accurate portal vein segmentation in liver cirrhosis.
- To address challenges like vascular discontinuities, missing segments, and topological inaccuracies.
- To improve the quantitative assessment of portal vein hemodynamics.
Main Methods:
- Proposed SnakeGCN, a 3D segmentation model operating on non-patch data.
- Integrated a 3D snake convolution module with nnU-Net for enhanced feature extraction.
- Incorporated a Graph Convolutional Network (GCN) in the bottleneck for global topology perception.
- Introduced a relationship loss (TLoss) to handle missing vascular segments.
Main Results:
- SnakeGCN achieved improved Dice scores by 2, 2, and 1 percentage points over nnU-Net on multi-center and public datasets.
- Connected component (CC) values decreased by 0.97, 2.23, and 8, indicating improved topological accuracy.
- Demonstrated superior performance compared to state-of-the-art deep learning methods in Dice score and topological metrics.
Conclusions:
- SnakeGCN offers a robust solution for challenging portal vein segmentation in liver cirrhosis.
- The model enhances accuracy and topological correctness, crucial for clinical applications.
- This advancement aids in better identification, treatment, and prognosis of liver cirrhosis complications.
Abstract:
The hemodynamic status of the portal vein plays a crucial role in the identification, treatment, and prognostic prediction of complications associated with liver cirrhosis. Accurate segmentation of the portal vein is essential for quantitative assessment but is highly challenging due to vascular discontinuities, missing segments, and erroneous linkages caused by cirrhotic and thrombotic lesions as well as background interference. To address these challenges, we propose SnakeGCN, an innovative 3D segmentation model that operates on non-patch data. We adopt a previously proposed 3D snake convolution module as a plug-in and align it with the feature extraction of nnU-Net to better capture curvilinear vessel structures. Additionally, we introduce a Graph Convolutional Network (GCN) in the bottleneck layer to enhance the perception of the vascular network's global topology, providing a robust structural representation and addressing incorrect linkages. To handle missing vascular segments caused by thrombi, we integrate a relationship loss (TLoss) into the network training process. The SnakeGCN model was validated on a multi-center clinical dataset and two public datasets (MSD-HepaticVessel and 3D-IRCADb). Across these datasets, our method achieved Dice score improvements of 2, 2, and 1 percentage points compared to nnU-Net. The connected component (CC) values decreased by 0.97, 2.23, and 8, respectively. These results demonstrate that our method achieves better performance than state-of-the-art deep learning methods in terms of Dice score and topological metrics for portal vein vessel segmentation.
