A Generative Model of Cortical Surfaces

Motivation: High-quality cortical surface enables the study of cortical thickness, area, and folding patterns, which are associated with neurodegeneration, developmental disorders, and age-related diseases. Goal(s): To develop a generative model for geometrically plausible cortical surfaces. Approach: Using 541 sets of static velocity fields that warp a spherical mesh template to cortical surfaces, we developed a VQ-VAE model to learn a latent codebook from these SVFs. We then trained a masked RNN to model these latent codebook, which generates new codes for the VQ-VAE that decodes them into new SVFs. Results: We qualitatively compared the cortical surfaces by different model configurations to evaluate their effectiveness. Impact: This cortical surface generative model can produce a large number of cortical surfaces for training deep learning models and conducting neuroimaging studies.