GVM-Net: A GNN-Based Vessel Matching Network for 2D/3D Non-Rigid Coronary Artery Registration

Area of Science:

• Medical Imaging
• Computer Vision
• Cardiovascular Interventions

Background:

• Accurate registration of coronary artery structures between preoperative coronary computed tomography angiography (CCTA) and intraoperative coronary angiography (ICA) is crucial for guiding percutaneous coronary interventions (PCIs).
• Challenges include non-rigid deformations and dimensional/topological discrepancies between CCTA and ICA, hindering precise 2D/3D coronary artery registration.

Purpose of the Study:

• To develop an end-to-end deep learning approach for robust 2D/3D coronary artery registration.
• To establish dense correspondence between coronary artery structures from different imaging modalities.

Main Methods:

• Formulated coronary artery registration as a centerline feature matching task.
• Proposed a Graph Neural Network-based Vessel Matching Network (GVM-Net) utilizing graph nodes for centerline points and attention mechanisms for topological relationship modeling.
• Incorporated redundant rows/columns in the matching matrix and a query-based nodes grouping module to handle structural inconsistencies and explore topological relationships.

Main Results:

• GVM-Net achieved an average F1-score of 89.74% (0.48 pixels mean distance) on a synthetic dataset (276 pairs).
• On 55 clinical cases, GVM-Net attained an average F1-score of 83.35% (1.52 mm mean error).
• Performance exceeded existing feature matching methods in both synthetic and clinical evaluations.

Conclusions:

• GVM-Net effectively addresses non-rigid deformation and topological discrepancies in coronary artery registration.
• The proposed GNN-based approach enables accurate dense correspondence between CCTA and ICA, enhancing guidance for PCIs.