An End-to-End Deep Learning Generative Framework for Refinable Shape Matching and Generation.

Area of Science:

• Computational Medicine
• Medical Imaging
• Artificial Intelligence

Background:

• In-Silico Clinical Trials (ISCTs) require realistic synthetic anatomical shapes for medical device validation.
• Generating 3D surface meshes with variable data is challenging due to lack of correspondences.
• Current AI models struggle with mesh variability (vertex counts, connectivity).

Purpose of the Study:

• To develop a novel unsupervised geometric deep-learning model for generating realistic 3D anatomical shapes.
• To establish refinable shape correspondences in a latent space and construct a population-derived atlas.
• To extend the model for joint shape generation, clustering, and multi-atlas frameworks for enhanced detail preservation.

Main Methods:

• Utilized graph representations for 3D surface meshes.
• Developed an unsupervised geometric deep-learning approach for latent space shape correspondence.
• Implemented a joint generative-clustering multi-atlas framework for improved shape synthesis.

Main Results:

• Successfully generated realistic synthetic 3D shapes from mesh data.
• Established refinable shape correspondences and a population-derived atlas.
• Demonstrated applicability to liver and left-ventricular models in computational medicine.

Conclusions:

• The proposed AI model is suitable for generating diverse anatomical shapes for In-Silico Clinical Trials.
• The approach addresses challenges of mesh variability and enhances detail preservation in synthetic models.
• This work advances computational medicine by providing a robust tool for medical device validation.