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

TL;DR

This work targets accurate, continuous segmentation of coronary vessels in volumetric CTA data, where topology and connectivity are as critical as local texture. It introduces ViG3D-UNet, a dual-branch encoder that fuses a 3D Vision GNN (ViG3D) for vascular topology with a CNN for texture, merged by a channel attention module, and paired with a paperclip-shaped offset decoder to recover high-resolution vessel maps efficiently. The method is evaluated on ASOCA and ImageCAS datasets, achieving state-of-the-art or near-state-of-the-art connectivity and boundary accuracy, with notable gains in HD95 and precision. These results suggest ViG3D-UNet offers a robust, end-to-end framework for clinically meaningful coronary vessel segmentation, potentially improving CTA-based assessment and planning while reducing topological errors common in vascular segmentation.

Abstract

Accurate vascular segmentation is essential for coronary visualization and the diagnosis of coronary heart disease. This task involves the extraction of sparse tree-like vascular branches from the volumetric space. However, existing methods have faced significant challenges due to discontinuous vascular segmentation and missing endpoints. To address this issue, a 3D vision graph neural network framework, named ViG3D-UNet, was introduced. This method integrates 3D graph representation and aggregation within a U-shaped architecture to facilitate continuous vascular segmentation. The ViG3D module captures volumetric vascular connectivity and topology, while the convolutional module extracts fine vascular details. These two branches are combined through channel attention to form the encoder feature. Subsequently, a paperclip-shaped offset decoder minimizes redundant computations in the sparse feature space and restores the feature map size to match the original input dimensions. To evaluate the effectiveness of the proposed approach for continuous vascular segmentation, evaluations were performed on two public datasets, ASOCA and ImageCAS. The segmentation results show that the ViG3D-UNet surpassed competing methods in maintaining vascular segmentation connectivity while achieving high segmentation accuracy. Our code will be available soon.

Figures (7)

• Figure 1: Limitations of general-purpose coronary CTA segmentation: vascular discontinuity and incomplete distal branch detection.
• Figure 2: Overview of ViG3D-UNet Architecture. The encoder is composed of a 3D vision GNN module and a 3D CNN module.The two modules are combined into a parallelized encoder under the block of channel attention. The fusion feature obtained via channel attention is concatenated with the first three layers of the decoder through skip connections to form the decoder feature. The first two layers of texture features, obtained through 3D convolution, are concatenated with the last two layers of the decoder through skip connections to form the decoder feature.
• Figure 3: Illustration of the 3D Vision GNN Module which identifies the connections between vascular nodes through graph aggregation operations during network training. (a) An overview of the branch of 3D Vision GNN in parallelized encoder. (b) Two stages in 3D Vision GNN Block: 3D graph processing and feed forward network.
• Figure 4: Performance comparison with varying numbers of neighbor nodes $\mathit{K}$ on the ImagesCAS dataset, with $\mathit{K}$ values ranging from 3 to 32. Optimal model performance is observed when $\mathit{K}$ is set to 7.
• Figure 5: Box plot of HD95 values for individual cases across different methods. Outliers, or data points that deviate significantly from the HD95 mean, are marked with circles