Recursive variational autoencoders for 3D blood vessel generative modeling.

Area of Science:

• Biomedical Engineering
• Computer Science
• Medical Imaging

Background:

• Accurate anatomical tree representation is vital for clinical diagnosis and treatment planning.
• Existing rule-based methods for synthesizing vasculature lack the diversity and complexity of real anatomical data.
• Intricate topology and geometry of vascular structures present significant representation challenges.

Purpose of the Study:

• To develop a novel method for synthesizing diverse and accurate 3D blood vessel models.
• To overcome limitations of existing rule-based vasculature synthesis techniques.
• To create realistic vascular geometries for applications in medical training and hemodynamic simulations.

Main Methods:

• Developed a Recursive variational Neural Network (RvNN) that leverages hierarchical vessel organization.
• Learned a low-dimensional manifold encoding branch connectivity and geometric features.
• Sampled the RvNN latent space to generate new vessel geometries.

Main Results:

• Generated 3D blood vessel models with high accuracy and diversity.
• Achieved high similarity to real data in vessel radii, length, and tortuosity.
• Demonstrated effectiveness across various datasets, including those with aneurysms.

Conclusions:

• The Recursive variational Neural Network (RvNN) successfully synthesizes realistic 3D blood vessel models.
• This technique offers a significant advancement over previous rule-based methods.
• The generated models are suitable for medical and surgical training, and hemodynamic simulations.