Point-Based Shape Representation Generation with a Correspondence-Preserving Diffusion Model.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Deep Learning

Background:

• Traditional statistical shape models utilize point correspondences, but current deep learning methods often ignore them, focusing on unordered point clouds.
• Existing deep generative models for point clouds do not generate shapes with consistent point correspondences between them.

Purpose of the Study:

• To develop a diffusion model capable of generating realistic point-based shape representations that preserve point correspondences from training data.
• To address the limitation of current deep learning methods in handling point correspondences for shape generation.

Main Methods:

• Proposed a novel diffusion model specifically designed for generating point-based shape representations with inherent point correspondences.
• Utilized shape representation data with correspondences from the Open Access Series of Imaging Studies 3 (OASIS-3) dataset.

Main Results:

• The correspondence-preserving diffusion model effectively generates highly realistic point-based hippocampal shape representations.
• Achieved superior realism and correspondence preservation compared to existing generative methods for point clouds.

Conclusions:

• The developed diffusion model successfully generates realistic 3D shapes with preserved point correspondences.
• Demonstrated the model's utility in downstream tasks like conditional generation for healthy/AD subjects and predicting disease progression via counterfactual generation.