Graph Convolution Based Cross-Network Multiscale Feature Fusion for Deep Vessel Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Vessel segmentation is crucial for diagnosing vascular diseases.
• Current methods lack the accuracy needed for clinical standards due to complex 3D vessel structures, sparsity, and anisotropy.
• Accurate 3D vessel reconstruction remains a significant challenge in medical imaging.

Purpose of the Study:

• To develop a novel hybrid deep neural network for high-quality 3D vessel segmentation.
• To address the limitations of existing methods in accurately reconstructing complex vascular structures.
• To improve the accuracy of vessel segmentation for clinical diagnostic applications.

Main Methods:

• Proposed a novel hybrid deep neural network with two cascaded subnetworks for initial and refined segmentation.
• The refined segmentation subnetwork integrates a Convolutional Neural Network (CNN)-based U-Net and a Graph U-Net.
• Employed cross-network multi-scale feature fusion and end-to-end training; graph construction prioritized vessel areas and orientation to handle sparsity and anisotropy.

Main Results:

• The proposed hybrid deep neural network achieved state-of-the-art 3D vessel segmentation performance.
• Demonstrated superior accuracy on multiple public and in-house datasets compared to existing methods.
• Successfully addressed challenges posed by vessel sparsity and anisotropy through specialized graph construction.

Conclusions:

• The novel hybrid deep neural network offers a significant advancement in 3D vessel segmentation accuracy.
• This method shows great potential for improving the diagnosis and management of vascular diseases.
• The developed approach provides a robust solution for high-quality vessel reconstruction in medical imaging.