Unsupervised Machine Learning for Vascular Mesh Compression.

Area of Science:

• Biomedical engineering
• Computational cardiovascular science
• Medical imaging and simulation

Background:

• Machine learning (ML) models are crucial for cardiovascular prediction and simulation.
• A key challenge is mesh compression to maintain geometric fidelity and optimize parameter efficiency.
• Abdominal aortic aneurysm (AAA) mesh compression is vital for accurate modeling.

Purpose of the Study:

• To present innovative mesh compression approaches for AAA using statistical and deep learning methods.
• To compare the effectiveness of Principal Component Analysis (PCA) against various deep learning models.
• To evaluate methods for preserving geometric fidelity and optimizing parameter efficiency in cardiovascular simulations.

Main Methods:

• Exploration of Principal Component Analysis (PCA) for mesh compression.
• Implementation and comparison of deep learning models: autoencoder, PCA-based autoencoder, Convolutional Neural Network (CNN), and Graph Neural Network (GNN).
• Compression of human aorta meshes, reconstruction, and comparative error analysis of reconstructed meshes.

Main Results:

• Principal Component Analysis (PCA) with 64 components demonstrated superior performance compared to deep learning models.
• Deep learning models were evaluated with a comparable latent space of 64.
• The PCA-based autoencoder showed the highest effectiveness among the deep learning approaches.

Conclusions:

• PCA is a highly effective method for abdominal aortic aneurysm (AAA) mesh compression.
• PCA outperforms deep learning models in preserving geometric fidelity and optimizing parameter efficiency for cardiovascular simulations.
• PCA-based autoencoders offer a promising deep learning approach for mesh compression in this domain.