CAR-Net: A Deep Learning-Based Deformation Model for 3D/2D Coronary Artery Registration.

Area of Science:

• Medical Imaging
• Cardiovascular Interventions
• Artificial Intelligence in Medicine

Background:

• Percutaneous coronary intervention (PCI) relies on X-ray coronary angiography (XCA), which lacks 3D information due to its projective nature.
• This 3D information loss complicates PCI procedures for coronary artery disease (CAD).
• Deformable 3D/2D coronary artery registration can fuse pre-operative CT angiography (CTA) with intra-operative XCA to address this limitation.

Purpose of the Study:

• To propose a novel deep learning-based neural network for deformable 3D/2D coronary artery registration.
• To enable accurate fusion of CTA and XCA for improved 3D visualization during PCI.
• To enhance the precision and safety of interventions for coronary artery disease.

Main Methods:

• A deep learning network performing segment-by-segment registration.
• Decomposition of vessel segment centerlines into origin and spherical coordinate shape tensors.
• Fusion of multi-modal features (CTA and XCA) for predicting angular deflections.
• Incorporation of motion and length preservation constraints within the deformation field.

Main Results:

• The proposed method achieved a low average registration error of 1.13 mm on a clinical dataset.
• Demonstrated effective fusion of 3D and 2D coronary imaging modalities.
• Validated the capability of the deep learning approach for accurate vessel segment registration.

Conclusions:

• The deep learning-based 3D/2D coronary artery registration method shows significant potential for clinical application.
• Accurate 3D structural information recovery can enhance guidance during percutaneous coronary intervention.
• This technique offers a promising advancement in interventional cardiology for treating coronary artery disease.