ViG3D-UNet: Volumetric Vascular Connectivity-Aware Segmentation via 3D Vision Graph Representation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate vascular segmentation is crucial for diagnosing coronary heart disease and visualizing coronary arteries.
• Existing segmentation methods struggle with discontinuous segmentation and missing vascular endpoints in volumetric data.

Purpose of the Study:

• To introduce ViG3D-UNet, a novel 3D vision graph neural network framework for continuous vascular segmentation.
• To address the limitations of current methods in capturing vascular connectivity and topology.

Main Methods:

• Developed ViG3D-UNet, integrating 3D graph representation and aggregation within a U-shaped architecture.
• Utilized a ViG3D module for vascular connectivity and topology, and a convolutional module for fine details.
• Employed channel attention for feature fusion and a paperclip-shaped offset decoder for efficient computation and feature map restoration.

Main Results:

• ViG3D-UNet demonstrated superior performance in maintaining vascular segmentation connectivity compared to existing methods.
• The framework achieved high segmentation accuracy on the ASOCA and ImageCAS public datasets.
• The proposed method effectively addresses challenges of discontinuous segmentation and missing endpoints.

Conclusions:

• ViG3D-UNet offers an effective solution for continuous vascular segmentation in medical imaging.
• The framework shows significant potential for improving coronary visualization and aiding in the diagnosis of coronary heart disease.