3D Graph-Connectivity Constrained Network for Hepatic Vessel Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Graph Neural Networks

Background:

• Accurate segmentation of hepatic vessels from 3D CT images is crucial for liver cancer diagnosis and surgical planning.
• Challenges in automatic segmentation include low contrast and high noise in CT images, hindering precise vessel delineation.
• Existing methods often overlook the inherent connectivity prior of hepatic vascular networks.

Purpose of the Study:

• To develop an efficient and accurate method for segmenting hepatic vessels from 3D CT images.
• To integrate the vascular connectivity prior into a deep learning framework for improved segmentation performance.
• To enhance the accuracy and connectivity of segmented hepatic vessels without increasing inference costs.

Main Methods:

• A Graph Neural Network (GNN), specifically a Graph Attention Network (GAT), was employed to model the connectivity prior of hepatic vessels.
• The GAT was integrated into a lightweight 3D U-Net architecture in a plug-in mode for training supervision.
• The GAT was used solely during training to guide the U-Net, ensuring no added computational cost during inference.

Main Results:

• The proposed method demonstrated superior performance in accuracy and connectivity compared to existing related works.
• Experiments on two public datasets validated the effectiveness of integrating the connectivity prior.
• The plug-in mechanism allowed for efficient integration of the GAT without compromising inference speed or hardware requirements.

Conclusions:

• The novel approach effectively utilizes the vascular connectivity prior for improved hepatic vessel segmentation in 3D CT images.
• The integration of GAT with 3D U-Net offers an efficient and accurate solution for a challenging medical imaging task.
• This method holds significant potential for enhancing preoperative planning and diagnostic accuracy in liver cancer cases.