Neural network reconstruction of the left atrium using sparse catheter paths

Area of Science:

• Medical Imaging
• Computational Anatomy
• Artificial Intelligence in Medicine

Background:

• Catheter-based radiofrequency ablation is a primary treatment for atrial fibrillation.
• Accurate left atrial mapping is crucial for pulmonary vein isolation but is time-consuming (>10 min).
• Current methods require dense surface sampling, posing challenges for certain catheter types.

Purpose of the Study:

• To develop an early left atrial visualization method for atrial fibrillation ablation procedures.
• To simplify procedural complexity and support advanced catheter workflows.
• To enable reconstruction from partial, sparsely sampled data.

Main Methods:

• Proposed a dense encoder-decoder network with a novel regularization term.
• Trained the network on a large dataset of 3D atria shapes and simulated catheter trajectories.
• Reconstructed the left atrium shape from partial point cloud data derived from catheter maneuvers.

Main Results:

• Demonstrated realistic left atrial visualization from partial data acquisition.
• Achieved reconstruction within a 3-minute timeframe in human clinical cases.
• Compared and validated the proposed network solution against other methods.

Conclusions:

• The proposed network effectively reconstructs left atrial shape for pulmonary vein isolation.
• Early visualization accelerates procedures and enhances catheter ablation efficacy.
• This AI-driven approach addresses limitations of current dense sampling techniques.