Human-airway surface mesh smoothing based on graph convolutional neural networks.

Area of Science:

• Medical imaging and computational modeling.
• Development of advanced algorithms for biomedical applications.

Background:

• Accurate airway geometry is crucial for computational fluid dynamics (CFD) simulations based on computed tomography (CT) images.
• Existing surface smoothing methods often cause shrinkage and information loss in smaller airway branches.

Purpose of the Study:

• To introduce an unsupervised airway-mesh-smoothing learning (AMSL) method.
• To preserve the original 3D airway geometry for accurate CT-image-based CFD simulations.

Main Methods:

• Jointly training two graph convolutional neural networks (GCNNs) to filter vertex positions and face normal vectors.
• Regularizing loss functions for reproducibility, smoothness, and consistency.
• Utilizing a deep mesh prior model for self-similar mesh restoration without large datasets.

Main Results:

• The AMSL method outperformed conventional and state-of-the-art deep learning methods in 18 of 20 benchmark problems.
• AMSL successfully smoothed 20 airway datasets, maintaining original airway diameters, especially in smaller branches.
• CFD simulations using AMSL-generated airways showed reduced pressure drop and wall shear stress compared to traditional methods.

Conclusions:

• The AMSL method accurately reproduces airway branch diameters without shrinkage, preserving critical details of smaller airways.
• Accurate airway geometry smoothing is essential for reliable flow property estimations in CFD simulations.