Efficient anatomical labeling of pulmonary tree structures via deep point-graph representation-based implicit fields.

Area of Science:

• Medical imaging
• Computational anatomy
• Pulmonary medicine

Background:

• Pulmonary diseases are a leading cause of death globally.
• Understanding complex 3D pulmonary structures (airways, vessels) is crucial for disease research.
• Traditional methods struggle with computational efficiency, resolution, and topological preservation.

Purpose of the Study:

• To develop a more efficient and accurate method for 3D pulmonary tree reconstruction.
• To address limitations of dense voxel-based approaches and sparsity issues in point-based methods.
• To improve the analysis of pulmonary structures for better understanding of lung diseases.

Main Methods:

• Shift from dense voxel to sparse point representation for memory efficiency.
• Utilized graph learning on skeletonized structures with differentiable feature fusion.
• Employed an implicit function for end-to-end conversion from sparse to dense representations.
• Curated a comprehensive dataset to address data scarcity.

Main Results:

• Achieved state-of-the-art performance in overall and key-location labeling accuracy.
• Demonstrated efficient inference capabilities.
• Enabled the generation of closed surface shapes.
• Validated the approach on a newly curated dataset.

Conclusions:

• The proposed graph learning method significantly enhances 3D pulmonary tree labeling accuracy and efficiency.
• This approach overcomes limitations of traditional methods, offering better topological preservation and global context.
• The method provides a valuable tool for pulmonary research and disease understanding.